crypto/bn/asm/bn86unix.cpp

   1 /* Run the C pre-processor over this file with one of the following defined
   2  * ELF - elf object files,
   3  * OUT - a.out object files,
   4  * BSDI - BSDI style a.out object files
   5  * SOL - Solaris style elf
   6  */
   7
   8 #define TYPE(a,b)       .type   a,b
   9 #define SIZE(a,b)       .size   a,b
  10
  11 #if defined(OUT) || defined(BSDI)
  12 #define bn_mul_add_words _bn_mul_add_words
  13 #define bn_mul_words _bn_mul_words
  14 #define bn_sqr_words _bn_sqr_words
  15 #define bn_div_words _bn_div_words
  16 #define bn_add_words _bn_add_words
  17 #define bn_sub_words _bn_sub_words
  18 #define bn_mul_comba8 _bn_mul_comba8
  19 #define bn_mul_comba4 _bn_mul_comba4
  20 #define bn_sqr_comba8 _bn_sqr_comba8
  21 #define bn_sqr_comba4 _bn_sqr_comba4
  22
  23 #endif
  24
  25 #ifdef OUT
  26 #define OK      1
  27 #define ALIGN   4
  28 #endif
  29
  30 #ifdef BSDI
  31 #define OK              1
  32 #define ALIGN           4
  33 #undef SIZE
  34 #undef TYPE
  35 #define SIZE(a,b)
  36 #define TYPE(a,b)
  37 #endif
  38
  39 #if defined(ELF) || defined(SOL)
  40 #define OK              1
  41 #define ALIGN           16
  42 #endif
  43
  44 #ifndef OK
  45 You need to define one of
  46 ELF - elf systems - linux-elf, NetBSD and DG-UX
  47 OUT - a.out systems - linux-a.out and FreeBSD
  48 SOL - solaris systems, which are elf with strange comment lines
  49 BSDI - a.out with a very primative version of as.
  50 #endif
  51
  52 /* Let the Assembler begin :-) */
  53         /* Don't even think of reading this code */
  54         /* It was automatically generated by bn-586.pl */
  55         /* Which is a perl program used to generate the x86 assember for */
  56         /* any of elf, a.out, BSDI,Win32, or Solaris */
  57         /* eric <eay@cryptsoft.com> */
  58
  59         .file   "bn-586.s"
  60         .version        "01.01"
  61 gcc2_compiled.:
  62 .text
  63         .align ALIGN
  64 .globl bn_mul_add_words
  65         TYPE(bn_mul_add_words,@function)
  66 bn_mul_add_words:
  67         pushl   %ebp
  68         pushl   %ebx
  69         pushl   %esi
  70         pushl   %edi
  71
  72
  73         xorl    %esi,           %esi
  74         movl    20(%esp),       %edi
  75         movl    28(%esp),       %ecx
  76         movl    24(%esp),       %ebx
  77         andl    $4294967288,    %ecx
  78         movl    32(%esp),       %ebp
  79         pushl   %ecx
  80         jz      .L000maw_finish
  81 .L001maw_loop:
  82         movl    %ecx,           (%esp)
  83         /* Round 0 */
  84         movl    (%ebx),         %eax
  85         mull    %ebp
  86         addl    %esi,           %eax
  87         movl    (%edi),         %esi
  88         adcl    $0,             %edx
  89         addl    %esi,           %eax
  90         adcl    $0,             %edx
  91         movl    %eax,           (%edi)
  92         movl    %edx,           %esi
  93         /* Round 4 */
  94         movl    4(%ebx),        %eax
  95         mull    %ebp
  96         addl    %esi,           %eax
  97         movl    4(%edi),        %esi
  98         adcl    $0,             %edx
  99         addl    %esi,           %eax
 100         adcl    $0,             %edx
 101         movl    %eax,           4(%edi)
 102         movl    %edx,           %esi
 103         /* Round 8 */
 104         movl    8(%ebx),        %eax
 105         mull    %ebp
 106         addl    %esi,           %eax
 107         movl    8(%edi),        %esi
 108         adcl    $0,             %edx
 109         addl    %esi,           %eax
 110         adcl    $0,             %edx
 111         movl    %eax,           8(%edi)
 112         movl    %edx,           %esi
 113         /* Round 12 */
 114         movl    12(%ebx),       %eax
 115         mull    %ebp
 116         addl    %esi,           %eax
 117         movl    12(%edi),       %esi
 118         adcl    $0,             %edx
 119         addl    %esi,           %eax
 120         adcl    $0,             %edx
 121         movl    %eax,           12(%edi)
 122         movl    %edx,           %esi
 123         /* Round 16 */
 124         movl    16(%ebx),       %eax
 125         mull    %ebp
 126         addl    %esi,           %eax
 127         movl    16(%edi),       %esi
 128         adcl    $0,             %edx
 129         addl    %esi,           %eax
 130         adcl    $0,             %edx
 131         movl    %eax,           16(%edi)
 132         movl    %edx,           %esi
 133         /* Round 20 */
 134         movl    20(%ebx),       %eax
 135         mull    %ebp
 136         addl    %esi,           %eax
 137         movl    20(%edi),       %esi
 138         adcl    $0,             %edx
 139         addl    %esi,           %eax
 140         adcl    $0,             %edx
 141         movl    %eax,           20(%edi)
 142         movl    %edx,           %esi
 143         /* Round 24 */
 144         movl    24(%ebx),       %eax
 145         mull    %ebp
 146         addl    %esi,           %eax
 147         movl    24(%edi),       %esi
 148         adcl    $0,             %edx
 149         addl    %esi,           %eax
 150         adcl    $0,             %edx
 151         movl    %eax,           24(%edi)
 152         movl    %edx,           %esi
 153         /* Round 28 */
 154         movl    28(%ebx),       %eax
 155         mull    %ebp
 156         addl    %esi,           %eax
 157         movl    28(%edi),       %esi
 158         adcl    $0,             %edx
 159         addl    %esi,           %eax
 160         adcl    $0,             %edx
 161         movl    %eax,           28(%edi)
 162         movl    %edx,           %esi
 163
 164         movl    (%esp),         %ecx
 165         addl    $32,            %ebx
 166         addl    $32,            %edi
 167         subl    $8,             %ecx
 168         jnz     .L001maw_loop
 169 .L000maw_finish:
 170         movl    32(%esp),       %ecx
 171         andl    $7,             %ecx
 172         jnz     .L002maw_finish2
 173         jmp     .L003maw_end
 174 .align ALIGN
 175 .L002maw_finish2:
 176         /* Tail Round 0 */
 177         movl    (%ebx),         %eax
 178         mull    %ebp
 179         addl    %esi,           %eax
 180         movl    (%edi),         %esi
 181         adcl    $0,             %edx
 182         addl    %esi,           %eax
 183         adcl    $0,             %edx
 184         decl    %ecx
 185         movl    %eax,           (%edi)
 186         movl    %edx,           %esi
 187         jz      .L003maw_end
 188         /* Tail Round 1 */
 189         movl    4(%ebx),        %eax
 190         mull    %ebp
 191         addl    %esi,           %eax
 192         movl    4(%edi),        %esi
 193         adcl    $0,             %edx
 194         addl    %esi,           %eax
 195         adcl    $0,             %edx
 196         decl    %ecx
 197         movl    %eax,           4(%edi)
 198         movl    %edx,           %esi
 199         jz      .L003maw_end
 200         /* Tail Round 2 */
 201         movl    8(%ebx),        %eax
 202         mull    %ebp
 203         addl    %esi,           %eax
 204         movl    8(%edi),        %esi
 205         adcl    $0,             %edx
 206         addl    %esi,           %eax
 207         adcl    $0,             %edx
 208         decl    %ecx
 209         movl    %eax,           8(%edi)
 210         movl    %edx,           %esi
 211         jz      .L003maw_end
 212         /* Tail Round 3 */
 213         movl    12(%ebx),       %eax
 214         mull    %ebp
 215         addl    %esi,           %eax
 216         movl    12(%edi),       %esi
 217         adcl    $0,             %edx
 218         addl    %esi,           %eax
 219         adcl    $0,             %edx
 220         decl    %ecx
 221         movl    %eax,           12(%edi)
 222         movl    %edx,           %esi
 223         jz      .L003maw_end
 224         /* Tail Round 4 */
 225         movl    16(%ebx),       %eax
 226         mull    %ebp
 227         addl    %esi,           %eax
 228         movl    16(%edi),       %esi
 229         adcl    $0,             %edx
 230         addl    %esi,           %eax
 231         adcl    $0,             %edx
 232         decl    %ecx
 233         movl    %eax,           16(%edi)
 234         movl    %edx,           %esi
 235         jz      .L003maw_end
 236         /* Tail Round 5 */
 237         movl    20(%ebx),       %eax
 238         mull    %ebp
 239         addl    %esi,           %eax
 240         movl    20(%edi),       %esi
 241         adcl    $0,             %edx
 242         addl    %esi,           %eax
 243         adcl    $0,             %edx
 244         decl    %ecx
 245         movl    %eax,           20(%edi)
 246         movl    %edx,           %esi
 247         jz      .L003maw_end
 248         /* Tail Round 6 */
 249         movl    24(%ebx),       %eax
 250         mull    %ebp
 251         addl    %esi,           %eax
 252         movl    24(%edi),       %esi
 253         adcl    $0,             %edx
 254         addl    %esi,           %eax
 255         adcl    $0,             %edx
 256         movl    %eax,           24(%edi)
 257         movl    %edx,           %esi
 258 .L003maw_end:
 259         movl    %esi,           %eax
 260         popl    %ecx
 261         popl    %edi
 262         popl    %esi
 263         popl    %ebx
 264         popl    %ebp
 265         ret
 266 .bn_mul_add_words_end:
 267         SIZE(bn_mul_add_words,.bn_mul_add_words_end-bn_mul_add_words)
 268 .ident  "bn_mul_add_words"
 269 .text
 270         .align ALIGN
 271 .globl bn_mul_words
 272         TYPE(bn_mul_words,@function)
 273 bn_mul_words:
 274         pushl   %ebp
 275         pushl   %ebx
 276         pushl   %esi
 277         pushl   %edi
 278
 279
 280         xorl    %esi,           %esi
 281         movl    20(%esp),       %edi
 282         movl    24(%esp),       %ebx
 283         movl    28(%esp),       %ebp
 284         movl    32(%esp),       %ecx
 285         andl    $4294967288,    %ebp
 286         jz      .L004mw_finish
 287 .L005mw_loop:
 288         /* Round 0 */
 289         movl    (%ebx),         %eax
 290         mull    %ecx
 291         addl    %esi,           %eax
 292         adcl    $0,             %edx
 293         movl    %eax,           (%edi)
 294         movl    %edx,           %esi
 295         /* Round 4 */
 296         movl    4(%ebx),        %eax
 297         mull    %ecx
 298         addl    %esi,           %eax
 299         adcl    $0,             %edx
 300         movl    %eax,           4(%edi)
 301         movl    %edx,           %esi
 302         /* Round 8 */
 303         movl    8(%ebx),        %eax
 304         mull    %ecx
 305         addl    %esi,           %eax
 306         adcl    $0,             %edx
 307         movl    %eax,           8(%edi)
 308         movl    %edx,           %esi
 309         /* Round 12 */
 310         movl    12(%ebx),       %eax
 311         mull    %ecx
 312         addl    %esi,           %eax
 313         adcl    $0,             %edx
 314         movl    %eax,           12(%edi)
 315         movl    %edx,           %esi
 316         /* Round 16 */
 317         movl    16(%ebx),       %eax
 318         mull    %ecx
 319         addl    %esi,           %eax
 320         adcl    $0,             %edx
 321         movl    %eax,           16(%edi)
 322         movl    %edx,           %esi
 323         /* Round 20 */
 324         movl    20(%ebx),       %eax
 325         mull    %ecx
 326         addl    %esi,           %eax
 327         adcl    $0,             %edx
 328         movl    %eax,           20(%edi)
 329         movl    %edx,           %esi
 330         /* Round 24 */
 331         movl    24(%ebx),       %eax
 332         mull    %ecx
 333         addl    %esi,           %eax
 334         adcl    $0,             %edx
 335         movl    %eax,           24(%edi)
 336         movl    %edx,           %esi
 337         /* Round 28 */
 338         movl    28(%ebx),       %eax
 339         mull    %ecx
 340         addl    %esi,           %eax
 341         adcl    $0,             %edx
 342         movl    %eax,           28(%edi)
 343         movl    %edx,           %esi
 344
 345         addl    $32,            %ebx
 346         addl    $32,            %edi
 347         subl    $8,             %ebp
 348         jz      .L004mw_finish
 349         jmp     .L005mw_loop
 350 .L004mw_finish:
 351         movl    28(%esp),       %ebp
 352         andl    $7,             %ebp
 353         jnz     .L006mw_finish2
 354         jmp     .L007mw_end
 355 .align ALIGN
 356 .L006mw_finish2:
 357         /* Tail Round 0 */
 358         movl    (%ebx),         %eax
 359         mull    %ecx
 360         addl    %esi,           %eax
 361         adcl    $0,             %edx
 362         movl    %eax,           (%edi)
 363         movl    %edx,           %esi
 364         decl    %ebp
 365         jz      .L007mw_end
 366         /* Tail Round 1 */
 367         movl    4(%ebx),        %eax
 368         mull    %ecx
 369         addl    %esi,           %eax
 370         adcl    $0,             %edx
 371         movl    %eax,           4(%edi)
 372         movl    %edx,           %esi
 373         decl    %ebp
 374         jz      .L007mw_end
 375         /* Tail Round 2 */
 376         movl    8(%ebx),        %eax
 377         mull    %ecx
 378         addl    %esi,           %eax
 379         adcl    $0,             %edx
 380         movl    %eax,           8(%edi)
 381         movl    %edx,           %esi
 382         decl    %ebp
 383         jz      .L007mw_end
 384         /* Tail Round 3 */
 385         movl    12(%ebx),       %eax
 386         mull    %ecx
 387         addl    %esi,           %eax
 388         adcl    $0,             %edx
 389         movl    %eax,           12(%edi)
 390         movl    %edx,           %esi
 391         decl    %ebp
 392         jz      .L007mw_end
 393         /* Tail Round 4 */
 394         movl    16(%ebx),       %eax
 395         mull    %ecx
 396         addl    %esi,           %eax
 397         adcl    $0,             %edx
 398         movl    %eax,           16(%edi)
 399         movl    %edx,           %esi
 400         decl    %ebp
 401         jz      .L007mw_end
 402         /* Tail Round 5 */
 403         movl    20(%ebx),       %eax
 404         mull    %ecx
 405         addl    %esi,           %eax
 406         adcl    $0,             %edx
 407         movl    %eax,           20(%edi)
 408         movl    %edx,           %esi
 409         decl    %ebp
 410         jz      .L007mw_end
 411         /* Tail Round 6 */
 412         movl    24(%ebx),       %eax
 413         mull    %ecx
 414         addl    %esi,           %eax
 415         adcl    $0,             %edx
 416         movl    %eax,           24(%edi)
 417         movl    %edx,           %esi
 418 .L007mw_end:
 419         movl    %esi,           %eax
 420         popl    %edi
 421         popl    %esi
 422         popl    %ebx
 423         popl    %ebp
 424         ret
 425 .bn_mul_words_end:
 426         SIZE(bn_mul_words,.bn_mul_words_end-bn_mul_words)
 427 .ident  "bn_mul_words"
 428 .text
 429         .align ALIGN
 430 .globl bn_sqr_words
 431         TYPE(bn_sqr_words,@function)
 432 bn_sqr_words:
 433         pushl   %ebp
 434         pushl   %ebx
 435         pushl   %esi
 436         pushl   %edi
 437
 438
 439         movl    20(%esp),       %esi
 440         movl    24(%esp),       %edi
 441         movl    28(%esp),       %ebx
 442         andl    $4294967288,    %ebx
 443         jz      .L008sw_finish
 444 .L009sw_loop:
 445         /* Round 0 */
 446         movl    (%edi),         %eax
 447         mull    %eax
 448         movl    %eax,           (%esi)
 449         movl    %edx,           4(%esi)
 450         /* Round 4 */
 451         movl    4(%edi),        %eax
 452         mull    %eax
 453         movl    %eax,           8(%esi)
 454         movl    %edx,           12(%esi)
 455         /* Round 8 */
 456         movl    8(%edi),        %eax
 457         mull    %eax
 458         movl    %eax,           16(%esi)
 459         movl    %edx,           20(%esi)
 460         /* Round 12 */
 461         movl    12(%edi),       %eax
 462         mull    %eax
 463         movl    %eax,           24(%esi)
 464         movl    %edx,           28(%esi)
 465         /* Round 16 */
 466         movl    16(%edi),       %eax
 467         mull    %eax
 468         movl    %eax,           32(%esi)
 469         movl    %edx,           36(%esi)
 470         /* Round 20 */
 471         movl    20(%edi),       %eax
 472         mull    %eax
 473         movl    %eax,           40(%esi)
 474         movl    %edx,           44(%esi)
 475         /* Round 24 */
 476         movl    24(%edi),       %eax
 477         mull    %eax
 478         movl    %eax,           48(%esi)
 479         movl    %edx,           52(%esi)
 480         /* Round 28 */
 481         movl    28(%edi),       %eax
 482         mull    %eax
 483         movl    %eax,           56(%esi)
 484         movl    %edx,           60(%esi)
 485
 486         addl    $32,            %edi
 487         addl    $64,            %esi
 488         subl    $8,             %ebx
 489         jnz     .L009sw_loop
 490 .L008sw_finish:
 491         movl    28(%esp),       %ebx
 492         andl    $7,             %ebx
 493         jz      .L010sw_end
 494         /* Tail Round 0 */
 495         movl    (%edi),         %eax
 496         mull    %eax
 497         movl    %eax,           (%esi)
 498         decl    %ebx
 499         movl    %edx,           4(%esi)
 500         jz      .L010sw_end
 501         /* Tail Round 1 */
 502         movl    4(%edi),        %eax
 503         mull    %eax
 504         movl    %eax,           8(%esi)
 505         decl    %ebx
 506         movl    %edx,           12(%esi)
 507         jz      .L010sw_end
 508         /* Tail Round 2 */
 509         movl    8(%edi),        %eax
 510         mull    %eax
 511         movl    %eax,           16(%esi)
 512         decl    %ebx
 513         movl    %edx,           20(%esi)
 514         jz      .L010sw_end
 515         /* Tail Round 3 */
 516         movl    12(%edi),       %eax
 517         mull    %eax
 518         movl    %eax,           24(%esi)
 519         decl    %ebx
 520         movl    %edx,           28(%esi)
 521         jz      .L010sw_end
 522         /* Tail Round 4 */
 523         movl    16(%edi),       %eax
 524         mull    %eax
 525         movl    %eax,           32(%esi)
 526         decl    %ebx
 527         movl    %edx,           36(%esi)
 528         jz      .L010sw_end
 529         /* Tail Round 5 */
 530         movl    20(%edi),       %eax
 531         mull    %eax
 532         movl    %eax,           40(%esi)
 533         decl    %ebx
 534         movl    %edx,           44(%esi)
 535         jz      .L010sw_end
 536         /* Tail Round 6 */
 537         movl    24(%edi),       %eax
 538         mull    %eax
 539         movl    %eax,           48(%esi)
 540         movl    %edx,           52(%esi)
 541 .L010sw_end:
 542         popl    %edi
 543         popl    %esi
 544         popl    %ebx
 545         popl    %ebp
 546         ret
 547 .bn_sqr_words_end:
 548         SIZE(bn_sqr_words,.bn_sqr_words_end-bn_sqr_words)
 549 .ident  "bn_sqr_words"
 550 .text
 551         .align ALIGN
 552 .globl bn_div_words
 553         TYPE(bn_div_words,@function)
 554 bn_div_words:
 555         pushl   %ebp
 556         pushl   %ebx
 557         pushl   %esi
 558         pushl   %edi
 559
 560         movl    20(%esp),       %edx
 561         movl    24(%esp),       %eax
 562         movl    28(%esp),       %ebx
 563         divl    %ebx
 564         popl    %edi
 565         popl    %esi
 566         popl    %ebx
 567         popl    %ebp
 568         ret
 569 .bn_div_words_end:
 570         SIZE(bn_div_words,.bn_div_words_end-bn_div_words)
 571 .ident  "bn_div_words"
 572 .text
 573         .align ALIGN
 574 .globl bn_add_words
 575         TYPE(bn_add_words,@function)
 576 bn_add_words:
 577         pushl   %ebp
 578         pushl   %ebx
 579         pushl   %esi
 580         pushl   %edi
 581
 582
 583         movl    20(%esp),       %ebx
 584         movl    24(%esp),       %esi
 585         movl    28(%esp),       %edi
 586         movl    32(%esp),       %ebp
 587         xorl    %eax,           %eax
 588         andl    $4294967288,    %ebp
 589         jz      .L011aw_finish
 590 .L012aw_loop:
 591         /* Round 0 */
 592         movl    (%esi),         %ecx
 593         movl    (%edi),         %edx
 594         addl    %eax,           %ecx
 595         movl    $0,             %eax
 596         adcl    %eax,           %eax
 597         addl    %edx,           %ecx
 598         adcl    $0,             %eax
 599         movl    %ecx,           (%ebx)
 600         /* Round 1 */
 601         movl    4(%esi),        %ecx
 602         movl    4(%edi),        %edx
 603         addl    %eax,           %ecx
 604         movl    $0,             %eax
 605         adcl    %eax,           %eax
 606         addl    %edx,           %ecx
 607         adcl    $0,             %eax
 608         movl    %ecx,           4(%ebx)
 609         /* Round 2 */
 610         movl    8(%esi),        %ecx
 611         movl    8(%edi),        %edx
 612         addl    %eax,           %ecx
 613         movl    $0,             %eax
 614         adcl    %eax,           %eax
 615         addl    %edx,           %ecx
 616         adcl    $0,             %eax
 617         movl    %ecx,           8(%ebx)
 618         /* Round 3 */
 619         movl    12(%esi),       %ecx
 620         movl    12(%edi),       %edx
 621         addl    %eax,           %ecx
 622         movl    $0,             %eax
 623         adcl    %eax,           %eax
 624         addl    %edx,           %ecx
 625         adcl    $0,             %eax
 626         movl    %ecx,           12(%ebx)
 627         /* Round 4 */
 628         movl    16(%esi),       %ecx
 629         movl    16(%edi),       %edx
 630         addl    %eax,           %ecx
 631         movl    $0,             %eax
 632         adcl    %eax,           %eax
 633         addl    %edx,           %ecx
 634         adcl    $0,             %eax
 635         movl    %ecx,           16(%ebx)
 636         /* Round 5 */
 637         movl    20(%esi),       %ecx
 638         movl    20(%edi),       %edx
 639         addl    %eax,           %ecx
 640         movl    $0,             %eax
 641         adcl    %eax,           %eax
 642         addl    %edx,           %ecx
 643         adcl    $0,             %eax
 644         movl    %ecx,           20(%ebx)
 645         /* Round 6 */
 646         movl    24(%esi),       %ecx
 647         movl    24(%edi),       %edx
 648         addl    %eax,           %ecx
 649         movl    $0,             %eax
 650         adcl    %eax,           %eax
 651         addl    %edx,           %ecx
 652         adcl    $0,             %eax
 653         movl    %ecx,           24(%ebx)
 654         /* Round 7 */
 655         movl    28(%esi),       %ecx
 656         movl    28(%edi),       %edx
 657         addl    %eax,           %ecx
 658         movl    $0,             %eax
 659         adcl    %eax,           %eax
 660         addl    %edx,           %ecx
 661         adcl    $0,             %eax
 662         movl    %ecx,           28(%ebx)
 663
 664         addl    $32,            %esi
 665         addl    $32,            %edi
 666         addl    $32,            %ebx
 667         subl    $8,             %ebp
 668         jnz     .L012aw_loop
 669 .L011aw_finish:
 670         movl    32(%esp),       %ebp
 671         andl    $7,             %ebp
 672         jz      .L013aw_end
 673         /* Tail Round 0 */
 674         movl    (%esi),         %ecx
 675         movl    (%edi),         %edx
 676         addl    %eax,           %ecx
 677         movl    $0,             %eax
 678         adcl    %eax,           %eax
 679         addl    %edx,           %ecx
 680         adcl    $0,             %eax
 681         decl    %ebp
 682         movl    %ecx,           (%ebx)
 683         jz      .L013aw_end
 684         /* Tail Round 1 */
 685         movl    4(%esi),        %ecx
 686         movl    4(%edi),        %edx
 687         addl    %eax,           %ecx
 688         movl    $0,             %eax
 689         adcl    %eax,           %eax
 690         addl    %edx,           %ecx
 691         adcl    $0,             %eax
 692         decl    %ebp
 693         movl    %ecx,           4(%ebx)
 694         jz      .L013aw_end
 695         /* Tail Round 2 */
 696         movl    8(%esi),        %ecx
 697         movl    8(%edi),        %edx
 698         addl    %eax,           %ecx
 699         movl    $0,             %eax
 700         adcl    %eax,           %eax
 701         addl    %edx,           %ecx
 702         adcl    $0,             %eax
 703         decl    %ebp
 704         movl    %ecx,           8(%ebx)
 705         jz      .L013aw_end
 706         /* Tail Round 3 */
 707         movl    12(%esi),       %ecx
 708         movl    12(%edi),       %edx
 709         addl    %eax,           %ecx
 710         movl    $0,             %eax
 711         adcl    %eax,           %eax
 712         addl    %edx,           %ecx
 713         adcl    $0,             %eax
 714         decl    %ebp
 715         movl    %ecx,           12(%ebx)
 716         jz      .L013aw_end
 717         /* Tail Round 4 */
 718         movl    16(%esi),       %ecx
 719         movl    16(%edi),       %edx
 720         addl    %eax,           %ecx
 721         movl    $0,             %eax
 722         adcl    %eax,           %eax
 723         addl    %edx,           %ecx
 724         adcl    $0,             %eax
 725         decl    %ebp
 726         movl    %ecx,           16(%ebx)
 727         jz      .L013aw_end
 728         /* Tail Round 5 */
 729         movl    20(%esi),       %ecx
 730         movl    20(%edi),       %edx
 731         addl    %eax,           %ecx
 732         movl    $0,             %eax
 733         adcl    %eax,           %eax
 734         addl    %edx,           %ecx
 735         adcl    $0,             %eax
 736         decl    %ebp
 737         movl    %ecx,           20(%ebx)
 738         jz      .L013aw_end
 739         /* Tail Round 6 */
 740         movl    24(%esi),       %ecx
 741         movl    24(%edi),       %edx
 742         addl    %eax,           %ecx
 743         movl    $0,             %eax
 744         adcl    %eax,           %eax
 745         addl    %edx,           %ecx
 746         adcl    $0,             %eax
 747         movl    %ecx,           24(%ebx)
 748 .L013aw_end:
 749         popl    %edi
 750         popl    %esi
 751         popl    %ebx
 752         popl    %ebp
 753         ret
 754 .bn_add_words_end:
 755         SIZE(bn_add_words,.bn_add_words_end-bn_add_words)
 756 .ident  "bn_add_words"
 757 .text
 758         .align ALIGN
 759 .globl bn_sub_words
 760         TYPE(bn_sub_words,@function)
 761 bn_sub_words:
 762         pushl   %ebp
 763         pushl   %ebx
 764         pushl   %esi
 765         pushl   %edi
 766
 767
 768         movl    20(%esp),       %ebx
 769         movl    24(%esp),       %esi
 770         movl    28(%esp),       %edi
 771         movl    32(%esp),       %ebp
 772         xorl    %eax,           %eax
 773         andl    $4294967288,    %ebp
 774         jz      .L014aw_finish
 775 .L015aw_loop:
 776         /* Round 0 */
 777         movl    (%esi),         %ecx
 778         movl    (%edi),         %edx
 779         subl    %eax,           %ecx
 780         movl    $0,             %eax
 781         adcl    %eax,           %eax
 782         subl    %edx,           %ecx
 783         adcl    $0,             %eax
 784         movl    %ecx,           (%ebx)
 785         /* Round 1 */
 786         movl    4(%esi),        %ecx
 787         movl    4(%edi),        %edx
 788         subl    %eax,           %ecx
 789         movl    $0,             %eax
 790         adcl    %eax,           %eax
 791         subl    %edx,           %ecx
 792         adcl    $0,             %eax
 793         movl    %ecx,           4(%ebx)
 794         /* Round 2 */
 795         movl    8(%esi),        %ecx
 796         movl    8(%edi),        %edx
 797         subl    %eax,           %ecx
 798         movl    $0,             %eax
 799         adcl    %eax,           %eax
 800         subl    %edx,           %ecx
 801         adcl    $0,             %eax
 802         movl    %ecx,           8(%ebx)
 803         /* Round 3 */
 804         movl    12(%esi),       %ecx
 805         movl    12(%edi),       %edx
 806         subl    %eax,           %ecx
 807         movl    $0,             %eax
 808         adcl    %eax,           %eax
 809         subl    %edx,           %ecx
 810         adcl    $0,             %eax
 811         movl    %ecx,           12(%ebx)
 812         /* Round 4 */
 813         movl    16(%esi),       %ecx
 814         movl    16(%edi),       %edx
 815         subl    %eax,           %ecx
 816         movl    $0,             %eax
 817         adcl    %eax,           %eax
 818         subl    %edx,           %ecx
 819         adcl    $0,             %eax
 820         movl    %ecx,           16(%ebx)
 821         /* Round 5 */
 822         movl    20(%esi),       %ecx
 823         movl    20(%edi),       %edx
 824         subl    %eax,           %ecx
 825         movl    $0,             %eax
 826         adcl    %eax,           %eax
 827         subl    %edx,           %ecx
 828         adcl    $0,             %eax
 829         movl    %ecx,           20(%ebx)
 830         /* Round 6 */
 831         movl    24(%esi),       %ecx
 832         movl    24(%edi),       %edx
 833         subl    %eax,           %ecx
 834         movl    $0,             %eax
 835         adcl    %eax,           %eax
 836         subl    %edx,           %ecx
 837         adcl    $0,             %eax
 838         movl    %ecx,           24(%ebx)
 839         /* Round 7 */
 840         movl    28(%esi),       %ecx
 841         movl    28(%edi),       %edx
 842         subl    %eax,           %ecx
 843         movl    $0,             %eax
 844         adcl    %eax,           %eax
 845         subl    %edx,           %ecx
 846         adcl    $0,             %eax
 847         movl    %ecx,           28(%ebx)
 848
 849         addl    $32,            %esi
 850         addl    $32,            %edi
 851         addl    $32,            %ebx
 852         subl    $8,             %ebp
 853         jnz     .L015aw_loop
 854 .L014aw_finish:
 855         movl    32(%esp),       %ebp
 856         andl    $7,             %ebp
 857         jz      .L016aw_end
 858         /* Tail Round 0 */
 859         movl    (%esi),         %ecx
 860         movl    (%edi),         %edx
 861         subl    %eax,           %ecx
 862         movl    $0,             %eax
 863         adcl    %eax,           %eax
 864         subl    %edx,           %ecx
 865         adcl    $0,             %eax
 866         decl    %ebp
 867         movl    %ecx,           (%ebx)
 868         jz      .L016aw_end
 869         /* Tail Round 1 */
 870         movl    4(%esi),        %ecx
 871         movl    4(%edi),        %edx
 872         subl    %eax,           %ecx
 873         movl    $0,             %eax
 874         adcl    %eax,           %eax
 875         subl    %edx,           %ecx
 876         adcl    $0,             %eax
 877         decl    %ebp
 878         movl    %ecx,           4(%ebx)
 879         jz      .L016aw_end
 880         /* Tail Round 2 */
 881         movl    8(%esi),        %ecx
 882         movl    8(%edi),        %edx
 883         subl    %eax,           %ecx
 884         movl    $0,             %eax
 885         adcl    %eax,           %eax
 886         subl    %edx,           %ecx
 887         adcl    $0,             %eax
 888         decl    %ebp
 889         movl    %ecx,           8(%ebx)
 890         jz      .L016aw_end
 891         /* Tail Round 3 */
 892         movl    12(%esi),       %ecx
 893         movl    12(%edi),       %edx
 894         subl    %eax,           %ecx
 895         movl    $0,             %eax
 896         adcl    %eax,           %eax
 897         subl    %edx,           %ecx
 898         adcl    $0,             %eax
 899         decl    %ebp
 900         movl    %ecx,           12(%ebx)
 901         jz      .L016aw_end
 902         /* Tail Round 4 */
 903         movl    16(%esi),       %ecx
 904         movl    16(%edi),       %edx
 905         subl    %eax,           %ecx
 906         movl    $0,             %eax
 907         adcl    %eax,           %eax
 908         subl    %edx,           %ecx
 909         adcl    $0,             %eax
 910         decl    %ebp
 911         movl    %ecx,           16(%ebx)
 912         jz      .L016aw_end
 913         /* Tail Round 5 */
 914         movl    20(%esi),       %ecx
 915         movl    20(%edi),       %edx
 916         subl    %eax,           %ecx
 917         movl    $0,             %eax
 918         adcl    %eax,           %eax
 919         subl    %edx,           %ecx
 920         adcl    $0,             %eax
 921         decl    %ebp
 922         movl    %ecx,           20(%ebx)
 923         jz      .L016aw_end
 924         /* Tail Round 6 */
 925         movl    24(%esi),       %ecx
 926         movl    24(%edi),       %edx
 927         subl    %eax,           %ecx
 928         movl    $0,             %eax
 929         adcl    %eax,           %eax
 930         subl    %edx,           %ecx
 931         adcl    $0,             %eax
 932         movl    %ecx,           24(%ebx)
 933 .L016aw_end:
 934         popl    %edi
 935         popl    %esi
 936         popl    %ebx
 937         popl    %ebp
 938         ret
 939 .bn_sub_words_end:
 940         SIZE(bn_sub_words,.bn_sub_words_end-bn_sub_words)
 941 .ident  "bn_sub_words"
 942 .text
 943         .align ALIGN
 944 .globl bn_mul_comba8
 945         TYPE(bn_mul_comba8,@function)
 946 bn_mul_comba8:
 947         pushl   %esi
 948         movl    12(%esp),       %esi
 949         pushl   %edi
 950         movl    20(%esp),       %edi
 951         pushl   %ebp
 952         pushl   %ebx
 953         xorl    %ebx,           %ebx
 954         movl    (%esi),         %eax
 955         xorl    %ecx,           %ecx
 956         movl    (%edi),         %edx
 957         /* ################## Calculate word 0 */
 958         xorl    %ebp,           %ebp
 959         /* mul a[0]*b[0] */
 960         mull    %edx
 961         addl    %eax,           %ebx
 962         movl    20(%esp),       %eax
 963         adcl    %edx,           %ecx
 964         movl    (%edi),         %edx
 965         adcl    $0,             %ebp
 966         movl    %ebx,           (%eax)
 967         movl    4(%esi),        %eax
 968         /* saved r[0] */
 969         /* ################## Calculate word 1 */
 970         xorl    %ebx,           %ebx
 971         /* mul a[1]*b[0] */
 972         mull    %edx
 973         addl    %eax,           %ecx
 974         movl    (%esi),         %eax
 975         adcl    %edx,           %ebp
 976         movl    4(%edi),        %edx
 977         adcl    $0,             %ebx
 978         /* mul a[0]*b[1] */
 979         mull    %edx
 980         addl    %eax,           %ecx
 981         movl    20(%esp),       %eax
 982         adcl    %edx,           %ebp
 983         movl    (%edi),         %edx
 984         adcl    $0,             %ebx
 985         movl    %ecx,           4(%eax)
 986         movl    8(%esi),        %eax
 987         /* saved r[1] */
 988         /* ################## Calculate word 2 */
 989         xorl    %ecx,           %ecx
 990         /* mul a[2]*b[0] */
 991         mull    %edx
 992         addl    %eax,           %ebp
 993         movl    4(%esi),        %eax
 994         adcl    %edx,           %ebx
 995         movl    4(%edi),        %edx
 996         adcl    $0,             %ecx
 997         /* mul a[1]*b[1] */
 998         mull    %edx
 999         addl    %eax,           %ebp
1000         movl    (%esi),         %eax
1001         adcl    %edx,           %ebx
1002         movl    8(%edi),        %edx
1003         adcl    $0,             %ecx
1004         /* mul a[0]*b[2] */
1005         mull    %edx
1006         addl    %eax,           %ebp
1007         movl    20(%esp),       %eax
1008         adcl    %edx,           %ebx
1009         movl    (%edi),         %edx
1010         adcl    $0,             %ecx
1011         movl    %ebp,           8(%eax)
1012         movl    12(%esi),       %eax
1013         /* saved r[2] */
1014         /* ################## Calculate word 3 */
1015         xorl    %ebp,           %ebp
1016         /* mul a[3]*b[0] */
1017         mull    %edx
1018         addl    %eax,           %ebx
1019         movl    8(%esi),        %eax
1020         adcl    %edx,           %ecx
1021         movl    4(%edi),        %edx
1022         adcl    $0,             %ebp
1023         /* mul a[2]*b[1] */
1024         mull    %edx
1025         addl    %eax,           %ebx
1026         movl    4(%esi),        %eax
1027         adcl    %edx,           %ecx
1028         movl    8(%edi),        %edx
1029         adcl    $0,             %ebp
1030         /* mul a[1]*b[2] */
1031         mull    %edx
1032         addl    %eax,           %ebx
1033         movl    (%esi),         %eax
1034         adcl    %edx,           %ecx
1035         movl    12(%edi),       %edx
1036         adcl    $0,             %ebp
1037         /* mul a[0]*b[3] */
1038         mull    %edx
1039         addl    %eax,           %ebx
1040         movl    20(%esp),       %eax
1041         adcl    %edx,           %ecx
1042         movl    (%edi),         %edx
1043         adcl    $0,             %ebp
1044         movl    %ebx,           12(%eax)
1045         movl    16(%esi),       %eax
1046         /* saved r[3] */
1047         /* ################## Calculate word 4 */
1048         xorl    %ebx,           %ebx
1049         /* mul a[4]*b[0] */
1050         mull    %edx
1051         addl    %eax,           %ecx
1052         movl    12(%esi),       %eax
1053         adcl    %edx,           %ebp
1054         movl    4(%edi),        %edx
1055         adcl    $0,             %ebx
1056         /* mul a[3]*b[1] */
1057         mull    %edx
1058         addl    %eax,           %ecx
1059         movl    8(%esi),        %eax
1060         adcl    %edx,           %ebp
1061         movl    8(%edi),        %edx
1062         adcl    $0,             %ebx
1063         /* mul a[2]*b[2] */
1064         mull    %edx
1065         addl    %eax,           %ecx
1066         movl    4(%esi),        %eax
1067         adcl    %edx,           %ebp
1068         movl    12(%edi),       %edx
1069         adcl    $0,             %ebx
1070         /* mul a[1]*b[3] */
1071         mull    %edx
1072         addl    %eax,           %ecx
1073         movl    (%esi),         %eax
1074         adcl    %edx,           %ebp
1075         movl    16(%edi),       %edx
1076         adcl    $0,             %ebx
1077         /* mul a[0]*b[4] */
1078         mull    %edx
1079         addl    %eax,           %ecx
1080         movl    20(%esp),       %eax
1081         adcl    %edx,           %ebp
1082         movl    (%edi),         %edx
1083         adcl    $0,             %ebx
1084         movl    %ecx,           16(%eax)
1085         movl    20(%esi),       %eax
1086         /* saved r[4] */
1087         /* ################## Calculate word 5 */
1088         xorl    %ecx,           %ecx
1089         /* mul a[5]*b[0] */
1090         mull    %edx
1091         addl    %eax,           %ebp
1092         movl    16(%esi),       %eax
1093         adcl    %edx,           %ebx
1094         movl    4(%edi),        %edx
1095         adcl    $0,             %ecx
1096         /* mul a[4]*b[1] */
1097         mull    %edx
1098         addl    %eax,           %ebp
1099         movl    12(%esi),       %eax
1100         adcl    %edx,           %ebx
1101         movl    8(%edi),        %edx
1102         adcl    $0,             %ecx
1103         /* mul a[3]*b[2] */
1104         mull    %edx
1105         addl    %eax,           %ebp
1106         movl    8(%esi),        %eax
1107         adcl    %edx,           %ebx
1108         movl    12(%edi),       %edx
1109         adcl    $0,             %ecx
1110         /* mul a[2]*b[3] */
1111         mull    %edx
1112         addl    %eax,           %ebp
1113         movl    4(%esi),        %eax
1114         adcl    %edx,           %ebx
1115         movl    16(%edi),       %edx
1116         adcl    $0,             %ecx
1117         /* mul a[1]*b[4] */
1118         mull    %edx
1119         addl    %eax,           %ebp
1120         movl    (%esi),         %eax
1121         adcl    %edx,           %ebx
1122         movl    20(%edi),       %edx
1123         adcl    $0,             %ecx
1124         /* mul a[0]*b[5] */
1125         mull    %edx
1126         addl    %eax,           %ebp
1127         movl    20(%esp),       %eax
1128         adcl    %edx,           %ebx
1129         movl    (%edi),         %edx
1130         adcl    $0,             %ecx
1131         movl    %ebp,           20(%eax)
1132         movl    24(%esi),       %eax
1133         /* saved r[5] */
1134         /* ################## Calculate word 6 */
1135         xorl    %ebp,           %ebp
1136         /* mul a[6]*b[0] */
1137         mull    %edx
1138         addl    %eax,           %ebx
1139         movl    20(%esi),       %eax
1140         adcl    %edx,           %ecx
1141         movl    4(%edi),        %edx
1142         adcl    $0,             %ebp
1143         /* mul a[5]*b[1] */
1144         mull    %edx
1145         addl    %eax,           %ebx
1146         movl    16(%esi),       %eax
1147         adcl    %edx,           %ecx
1148         movl    8(%edi),        %edx
1149         adcl    $0,             %ebp
1150         /* mul a[4]*b[2] */
1151         mull    %edx
1152         addl    %eax,           %ebx
1153         movl    12(%esi),       %eax
1154         adcl    %edx,           %ecx
1155         movl    12(%edi),       %edx
1156         adcl    $0,             %ebp
1157         /* mul a[3]*b[3] */
1158         mull    %edx
1159         addl    %eax,           %ebx
1160         movl    8(%esi),        %eax
1161         adcl    %edx,           %ecx
1162         movl    16(%edi),       %edx
1163         adcl    $0,             %ebp
1164         /* mul a[2]*b[4] */
1165         mull    %edx
1166         addl    %eax,           %ebx
1167         movl    4(%esi),        %eax
1168         adcl    %edx,           %ecx
1169         movl    20(%edi),       %edx
1170         adcl    $0,             %ebp
1171         /* mul a[1]*b[5] */
1172         mull    %edx
1173         addl    %eax,           %ebx
1174         movl    (%esi),         %eax
1175         adcl    %edx,           %ecx
1176         movl    24(%edi),       %edx
1177         adcl    $0,             %ebp
1178         /* mul a[0]*b[6] */
1179         mull    %edx
1180         addl    %eax,           %ebx
1181         movl    20(%esp),       %eax
1182         adcl    %edx,           %ecx
1183         movl    (%edi),         %edx
1184         adcl    $0,             %ebp
1185         movl    %ebx,           24(%eax)
1186         movl    28(%esi),       %eax
1187         /* saved r[6] */
1188         /* ################## Calculate word 7 */
1189         xorl    %ebx,           %ebx
1190         /* mul a[7]*b[0] */
1191         mull    %edx
1192         addl    %eax,           %ecx
1193         movl    24(%esi),       %eax
1194         adcl    %edx,           %ebp
1195         movl    4(%edi),        %edx
1196         adcl    $0,             %ebx
1197         /* mul a[6]*b[1] */
1198         mull    %edx
1199         addl    %eax,           %ecx
1200         movl    20(%esi),       %eax
1201         adcl    %edx,           %ebp
1202         movl    8(%edi),        %edx
1203         adcl    $0,             %ebx
1204         /* mul a[5]*b[2] */
1205         mull    %edx
1206         addl    %eax,           %ecx
1207         movl    16(%esi),       %eax
1208         adcl    %edx,           %ebp
1209         movl    12(%edi),       %edx
1210         adcl    $0,             %ebx
1211         /* mul a[4]*b[3] */
1212         mull    %edx
1213         addl    %eax,           %ecx
1214         movl    12(%esi),       %eax
1215         adcl    %edx,           %ebp
1216         movl    16(%edi),       %edx
1217         adcl    $0,             %ebx
1218         /* mul a[3]*b[4] */
1219         mull    %edx
1220         addl    %eax,           %ecx
1221         movl    8(%esi),        %eax
1222         adcl    %edx,           %ebp
1223         movl    20(%edi),       %edx
1224         adcl    $0,             %ebx
1225         /* mul a[2]*b[5] */
1226         mull    %edx
1227         addl    %eax,           %ecx
1228         movl    4(%esi),        %eax
1229         adcl    %edx,           %ebp
1230         movl    24(%edi),       %edx
1231         adcl    $0,             %ebx
1232         /* mul a[1]*b[6] */
1233         mull    %edx
1234         addl    %eax,           %ecx
1235         movl    (%esi),         %eax
1236         adcl    %edx,           %ebp
1237         movl    28(%edi),       %edx
1238         adcl    $0,             %ebx
1239         /* mul a[0]*b[7] */
1240         mull    %edx
1241         addl    %eax,           %ecx
1242         movl    20(%esp),       %eax
1243         adcl    %edx,           %ebp
1244         movl    4(%edi),        %edx
1245         adcl    $0,             %ebx
1246         movl    %ecx,           28(%eax)
1247         movl    28(%esi),       %eax
1248         /* saved r[7] */
1249         /* ################## Calculate word 8 */
1250         xorl    %ecx,           %ecx
1251         /* mul a[7]*b[1] */
1252         mull    %edx
1253         addl    %eax,           %ebp
1254         movl    24(%esi),       %eax
1255         adcl    %edx,           %ebx
1256         movl    8(%edi),        %edx
1257         adcl    $0,             %ecx
1258         /* mul a[6]*b[2] */
1259         mull    %edx
1260         addl    %eax,           %ebp
1261         movl    20(%esi),       %eax
1262         adcl    %edx,           %ebx
1263         movl    12(%edi),       %edx
1264         adcl    $0,             %ecx
1265         /* mul a[5]*b[3] */
1266         mull    %edx
1267         addl    %eax,           %ebp
1268         movl    16(%esi),       %eax
1269         adcl    %edx,           %ebx
1270         movl    16(%edi),       %edx
1271         adcl    $0,             %ecx
1272         /* mul a[4]*b[4] */
1273         mull    %edx
1274         addl    %eax,           %ebp
1275         movl    12(%esi),       %eax
1276         adcl    %edx,           %ebx
1277         movl    20(%edi),       %edx
1278         adcl    $0,             %ecx
1279         /* mul a[3]*b[5] */
1280         mull    %edx
1281         addl    %eax,           %ebp
1282         movl    8(%esi),        %eax
1283         adcl    %edx,           %ebx
1284         movl    24(%edi),       %edx
1285         adcl    $0,             %ecx
1286         /* mul a[2]*b[6] */
1287         mull    %edx
1288         addl    %eax,           %ebp
1289         movl    4(%esi),        %eax
1290         adcl    %edx,           %ebx
1291         movl    28(%edi),       %edx
1292         adcl    $0,             %ecx
1293         /* mul a[1]*b[7] */
1294         mull    %edx
1295         addl    %eax,           %ebp
1296         movl    20(%esp),       %eax
1297         adcl    %edx,           %ebx
1298         movl    8(%edi),        %edx
1299         adcl    $0,             %ecx
1300         movl    %ebp,           32(%eax)
1301         movl    28(%esi),       %eax
1302         /* saved r[8] */
1303         /* ################## Calculate word 9 */
1304         xorl    %ebp,           %ebp
1305         /* mul a[7]*b[2] */
1306         mull    %edx
1307         addl    %eax,           %ebx
1308         movl    24(%esi),       %eax
1309         adcl    %edx,           %ecx
1310         movl    12(%edi),       %edx
1311         adcl    $0,             %ebp
1312         /* mul a[6]*b[3] */
1313         mull    %edx
1314         addl    %eax,           %ebx
1315         movl    20(%esi),       %eax
1316         adcl    %edx,           %ecx
1317         movl    16(%edi),       %edx
1318         adcl    $0,             %ebp
1319         /* mul a[5]*b[4] */
1320         mull    %edx
1321         addl    %eax,           %ebx
1322         movl    16(%esi),       %eax
1323         adcl    %edx,           %ecx
1324         movl    20(%edi),       %edx
1325         adcl    $0,             %ebp
1326         /* mul a[4]*b[5] */
1327         mull    %edx
1328         addl    %eax,           %ebx
1329         movl    12(%esi),       %eax
1330         adcl    %edx,           %ecx
1331         movl    24(%edi),       %edx
1332         adcl    $0,             %ebp
1333         /* mul a[3]*b[6] */
1334         mull    %edx
1335         addl    %eax,           %ebx
1336         movl    8(%esi),        %eax
1337         adcl    %edx,           %ecx
1338         movl    28(%edi),       %edx
1339         adcl    $0,             %ebp
1340         /* mul a[2]*b[7] */
1341         mull    %edx
1342         addl    %eax,           %ebx
1343         movl    20(%esp),       %eax
1344         adcl    %edx,           %ecx
1345         movl    12(%edi),       %edx
1346         adcl    $0,             %ebp
1347         movl    %ebx,           36(%eax)
1348         movl    28(%esi),       %eax
1349         /* saved r[9] */
1350         /* ################## Calculate word 10 */
1351         xorl    %ebx,           %ebx
1352         /* mul a[7]*b[3] */
1353         mull    %edx
1354         addl    %eax,           %ecx
1355         movl    24(%esi),       %eax
1356         adcl    %edx,           %ebp
1357         movl    16(%edi),       %edx
1358         adcl    $0,             %ebx
1359         /* mul a[6]*b[4] */
1360         mull    %edx
1361         addl    %eax,           %ecx
1362         movl    20(%esi),       %eax
1363         adcl    %edx,           %ebp
1364         movl    20(%edi),       %edx
1365         adcl    $0,             %ebx
1366         /* mul a[5]*b[5] */
1367         mull    %edx
1368         addl    %eax,           %ecx
1369         movl    16(%esi),       %eax
1370         adcl    %edx,           %ebp
1371         movl    24(%edi),       %edx
1372         adcl    $0,             %ebx
1373         /* mul a[4]*b[6] */
1374         mull    %edx
1375         addl    %eax,           %ecx
1376         movl    12(%esi),       %eax
1377         adcl    %edx,           %ebp
1378         movl    28(%edi),       %edx
1379         adcl    $0,             %ebx
1380         /* mul a[3]*b[7] */
1381         mull    %edx
1382         addl    %eax,           %ecx
1383         movl    20(%esp),       %eax
1384         adcl    %edx,           %ebp
1385         movl    16(%edi),       %edx
1386         adcl    $0,             %ebx
1387         movl    %ecx,           40(%eax)
1388         movl    28(%esi),       %eax
1389         /* saved r[10] */
1390         /* ################## Calculate word 11 */
1391         xorl    %ecx,           %ecx
1392         /* mul a[7]*b[4] */
1393         mull    %edx
1394         addl    %eax,           %ebp
1395         movl    24(%esi),       %eax
1396         adcl    %edx,           %ebx
1397         movl    20(%edi),       %edx
1398         adcl    $0,             %ecx
1399         /* mul a[6]*b[5] */
1400         mull    %edx
1401         addl    %eax,           %ebp
1402         movl    20(%esi),       %eax
1403         adcl    %edx,           %ebx
1404         movl    24(%edi),       %edx
1405         adcl    $0,             %ecx
1406         /* mul a[5]*b[6] */
1407         mull    %edx
1408         addl    %eax,           %ebp
1409         movl    16(%esi),       %eax
1410         adcl    %edx,           %ebx
1411         movl    28(%edi),       %edx
1412         adcl    $0,             %ecx
1413         /* mul a[4]*b[7] */
1414         mull    %edx
1415         addl    %eax,           %ebp
1416         movl    20(%esp),       %eax
1417         adcl    %edx,           %ebx
1418         movl    20(%edi),       %edx
1419         adcl    $0,             %ecx
1420         movl    %ebp,           44(%eax)
1421         movl    28(%esi),       %eax
1422         /* saved r[11] */
1423         /* ################## Calculate word 12 */
1424         xorl    %ebp,           %ebp
1425         /* mul a[7]*b[5] */
1426         mull    %edx
1427         addl    %eax,           %ebx
1428         movl    24(%esi),       %eax
1429         adcl    %edx,           %ecx
1430         movl    24(%edi),       %edx
1431         adcl    $0,             %ebp
1432         /* mul a[6]*b[6] */
1433         mull    %edx
1434         addl    %eax,           %ebx
1435         movl    20(%esi),       %eax
1436         adcl    %edx,           %ecx
1437         movl    28(%edi),       %edx
1438         adcl    $0,             %ebp
1439         /* mul a[5]*b[7] */
1440         mull    %edx
1441         addl    %eax,           %ebx
1442         movl    20(%esp),       %eax
1443         adcl    %edx,           %ecx
1444         movl    24(%edi),       %edx
1445         adcl    $0,             %ebp
1446         movl    %ebx,           48(%eax)
1447         movl    28(%esi),       %eax
1448         /* saved r[12] */
1449         /* ################## Calculate word 13 */
1450         xorl    %ebx,           %ebx
1451         /* mul a[7]*b[6] */
1452         mull    %edx
1453         addl    %eax,           %ecx
1454         movl    24(%esi),       %eax
1455         adcl    %edx,           %ebp
1456         movl    28(%edi),       %edx
1457         adcl    $0,             %ebx
1458         /* mul a[6]*b[7] */
1459         mull    %edx
1460         addl    %eax,           %ecx
1461         movl    20(%esp),       %eax
1462         adcl    %edx,           %ebp
1463         movl    28(%edi),       %edx
1464         adcl    $0,             %ebx
1465         movl    %ecx,           52(%eax)
1466         movl    28(%esi),       %eax
1467         /* saved r[13] */
1468         /* ################## Calculate word 14 */
1469         xorl    %ecx,           %ecx
1470         /* mul a[7]*b[7] */
1471         mull    %edx
1472         addl    %eax,           %ebp
1473         movl    20(%esp),       %eax
1474         adcl    %edx,           %ebx
1475         adcl    $0,             %ecx
1476         movl    %ebp,           56(%eax)
1477         /* saved r[14] */
1478         /* save r[15] */
1479         movl    %ebx,           60(%eax)
1480         popl    %ebx
1481         popl    %ebp
1482         popl    %edi
1483         popl    %esi
1484         ret
1485 .bn_mul_comba8_end:
1486         SIZE(bn_mul_comba8,.bn_mul_comba8_end-bn_mul_comba8)
1487 .ident  "desasm.pl"
1488 .text
1489         .align ALIGN
1490 .globl bn_mul_comba4
1491         TYPE(bn_mul_comba4,@function)
1492 bn_mul_comba4:
1493         pushl   %esi
1494         movl    12(%esp),       %esi
1495         pushl   %edi
1496         movl    20(%esp),       %edi
1497         pushl   %ebp
1498         pushl   %ebx
1499         xorl    %ebx,           %ebx
1500         movl    (%esi),         %eax
1501         xorl    %ecx,           %ecx
1502         movl    (%edi),         %edx
1503         /* ################## Calculate word 0 */
1504         xorl    %ebp,           %ebp
1505         /* mul a[0]*b[0] */
1506         mull    %edx
1507         addl    %eax,           %ebx
1508         movl    20(%esp),       %eax
1509         adcl    %edx,           %ecx
1510         movl    (%edi),         %edx
1511         adcl    $0,             %ebp
1512         movl    %ebx,           (%eax)
1513         movl    4(%esi),        %eax
1514         /* saved r[0] */
1515         /* ################## Calculate word 1 */
1516         xorl    %ebx,           %ebx
1517         /* mul a[1]*b[0] */
1518         mull    %edx
1519         addl    %eax,           %ecx
1520         movl    (%esi),         %eax
1521         adcl    %edx,           %ebp
1522         movl    4(%edi),        %edx
1523         adcl    $0,             %ebx
1524         /* mul a[0]*b[1] */
1525         mull    %edx
1526         addl    %eax,           %ecx
1527         movl    20(%esp),       %eax
1528         adcl    %edx,           %ebp
1529         movl    (%edi),         %edx
1530         adcl    $0,             %ebx
1531         movl    %ecx,           4(%eax)
1532         movl    8(%esi),        %eax
1533         /* saved r[1] */
1534         /* ################## Calculate word 2 */
1535         xorl    %ecx,           %ecx
1536         /* mul a[2]*b[0] */
1537         mull    %edx
1538         addl    %eax,           %ebp
1539         movl    4(%esi),        %eax
1540         adcl    %edx,           %ebx
1541         movl    4(%edi),        %edx
1542         adcl    $0,             %ecx
1543         /* mul a[1]*b[1] */
1544         mull    %edx
1545         addl    %eax,           %ebp
1546         movl    (%esi),         %eax
1547         adcl    %edx,           %ebx
1548         movl    8(%edi),        %edx
1549         adcl    $0,             %ecx
1550         /* mul a[0]*b[2] */
1551         mull    %edx
1552         addl    %eax,           %ebp
1553         movl    20(%esp),       %eax
1554         adcl    %edx,           %ebx
1555         movl    (%edi),         %edx
1556         adcl    $0,             %ecx
1557         movl    %ebp,           8(%eax)
1558         movl    12(%esi),       %eax
1559         /* saved r[2] */
1560         /* ################## Calculate word 3 */
1561         xorl    %ebp,           %ebp
1562         /* mul a[3]*b[0] */
1563         mull    %edx
1564         addl    %eax,           %ebx
1565         movl    8(%esi),        %eax
1566         adcl    %edx,           %ecx
1567         movl    4(%edi),        %edx
1568         adcl    $0,             %ebp
1569         /* mul a[2]*b[1] */
1570         mull    %edx
1571         addl    %eax,           %ebx
1572         movl    4(%esi),        %eax
1573         adcl    %edx,           %ecx
1574         movl    8(%edi),        %edx
1575         adcl    $0,             %ebp
1576         /* mul a[1]*b[2] */
1577         mull    %edx
1578         addl    %eax,           %ebx
1579         movl    (%esi),         %eax
1580         adcl    %edx,           %ecx
1581         movl    12(%edi),       %edx
1582         adcl    $0,             %ebp
1583         /* mul a[0]*b[3] */
1584         mull    %edx
1585         addl    %eax,           %ebx
1586         movl    20(%esp),       %eax
1587         adcl    %edx,           %ecx
1588         movl    4(%edi),        %edx
1589         adcl    $0,             %ebp
1590         movl    %ebx,           12(%eax)
1591         movl    12(%esi),       %eax
1592         /* saved r[3] */
1593         /* ################## Calculate word 4 */
1594         xorl    %ebx,           %ebx
1595         /* mul a[3]*b[1] */
1596         mull    %edx
1597         addl    %eax,           %ecx
1598         movl    8(%esi),        %eax
1599         adcl    %edx,           %ebp
1600         movl    8(%edi),        %edx
1601         adcl    $0,             %ebx
1602         /* mul a[2]*b[2] */
1603         mull    %edx
1604         addl    %eax,           %ecx
1605         movl    4(%esi),        %eax
1606         adcl    %edx,           %ebp
1607         movl    12(%edi),       %edx
1608         adcl    $0,             %ebx
1609         /* mul a[1]*b[3] */
1610         mull    %edx
1611         addl    %eax,           %ecx
1612         movl    20(%esp),       %eax
1613         adcl    %edx,           %ebp
1614         movl    8(%edi),        %edx
1615         adcl    $0,             %ebx
1616         movl    %ecx,           16(%eax)
1617         movl    12(%esi),       %eax
1618         /* saved r[4] */
1619         /* ################## Calculate word 5 */
1620         xorl    %ecx,           %ecx
1621         /* mul a[3]*b[2] */
1622         mull    %edx
1623         addl    %eax,           %ebp
1624         movl    8(%esi),        %eax
1625         adcl    %edx,           %ebx
1626         movl    12(%edi),       %edx
1627         adcl    $0,             %ecx
1628         /* mul a[2]*b[3] */
1629         mull    %edx
1630         addl    %eax,           %ebp
1631         movl    20(%esp),       %eax
1632         adcl    %edx,           %ebx
1633         movl    12(%edi),       %edx
1634         adcl    $0,             %ecx
1635         movl    %ebp,           20(%eax)
1636         movl    12(%esi),       %eax
1637         /* saved r[5] */
1638         /* ################## Calculate word 6 */
1639         xorl    %ebp,           %ebp
1640         /* mul a[3]*b[3] */
1641         mull    %edx
1642         addl    %eax,           %ebx
1643         movl    20(%esp),       %eax
1644         adcl    %edx,           %ecx
1645         adcl    $0,             %ebp
1646         movl    %ebx,           24(%eax)
1647         /* saved r[6] */
1648         /* save r[7] */
1649         movl    %ecx,           28(%eax)
1650         popl    %ebx
1651         popl    %ebp
1652         popl    %edi
1653         popl    %esi
1654         ret
1655 .bn_mul_comba4_end:
1656         SIZE(bn_mul_comba4,.bn_mul_comba4_end-bn_mul_comba4)
1657 .ident  "desasm.pl"
1658 .text
1659         .align ALIGN
1660 .globl bn_sqr_comba8
1661         TYPE(bn_sqr_comba8,@function)
1662 bn_sqr_comba8:
1663         pushl   %esi
1664         pushl   %edi
1665         pushl   %ebp
1666         pushl   %ebx
1667         movl    20(%esp),       %edi
1668         movl    24(%esp),       %esi
1669         xorl    %ebx,           %ebx
1670         xorl    %ecx,           %ecx
1671         movl    (%esi),         %eax
1672         /* ############### Calculate word 0 */
1673         xorl    %ebp,           %ebp
1674         /* sqr a[0]*a[0] */
1675         mull    %eax
1676         addl    %eax,           %ebx
1677         adcl    %edx,           %ecx
1678         movl    (%esi),         %edx
1679         adcl    $0,             %ebp
1680         movl    %ebx,           (%edi)
1681         movl    4(%esi),        %eax
1682         /* saved r[0] */
1683         /* ############### Calculate word 1 */
1684         xorl    %ebx,           %ebx
1685         /* sqr a[1]*a[0] */
1686         mull    %edx
1687         addl    %eax,           %eax
1688         adcl    %edx,           %edx
1689         adcl    $0,             %ebx
1690         addl    %eax,           %ecx
1691         adcl    %edx,           %ebp
1692         movl    8(%esi),        %eax
1693         adcl    $0,             %ebx
1694         movl    %ecx,           4(%edi)
1695         movl    (%esi),         %edx
1696         /* saved r[1] */
1697         /* ############### Calculate word 2 */
1698         xorl    %ecx,           %ecx
1699         /* sqr a[2]*a[0] */
1700         mull    %edx
1701         addl    %eax,           %eax
1702         adcl    %edx,           %edx
1703         adcl    $0,             %ecx
1704         addl    %eax,           %ebp
1705         adcl    %edx,           %ebx
1706         movl    4(%esi),        %eax
1707         adcl    $0,             %ecx
1708         /* sqr a[1]*a[1] */
1709         mull    %eax
1710         addl    %eax,           %ebp
1711         adcl    %edx,           %ebx
1712         movl    (%esi),         %edx
1713         adcl    $0,             %ecx
1714         movl    %ebp,           8(%edi)
1715         movl    12(%esi),       %eax
1716         /* saved r[2] */
1717         /* ############### Calculate word 3 */
1718         xorl    %ebp,           %ebp
1719         /* sqr a[3]*a[0] */
1720         mull    %edx
1721         addl    %eax,           %eax
1722         adcl    %edx,           %edx
1723         adcl    $0,             %ebp
1724         addl    %eax,           %ebx
1725         adcl    %edx,           %ecx
1726         movl    8(%esi),        %eax
1727         adcl    $0,             %ebp
1728         movl    4(%esi),        %edx
1729         /* sqr a[2]*a[1] */
1730         mull    %edx
1731         addl    %eax,           %eax
1732         adcl    %edx,           %edx
1733         adcl    $0,             %ebp
1734         addl    %eax,           %ebx
1735         adcl    %edx,           %ecx
1736         movl    16(%esi),       %eax
1737         adcl    $0,             %ebp
1738         movl    %ebx,           12(%edi)
1739         movl    (%esi),         %edx
1740         /* saved r[3] */
1741         /* ############### Calculate word 4 */
1742         xorl    %ebx,           %ebx
1743         /* sqr a[4]*a[0] */
1744         mull    %edx
1745         addl    %eax,           %eax
1746         adcl    %edx,           %edx
1747         adcl    $0,             %ebx
1748         addl    %eax,           %ecx
1749         adcl    %edx,           %ebp
1750         movl    12(%esi),       %eax
1751         adcl    $0,             %ebx
1752         movl    4(%esi),        %edx
1753         /* sqr a[3]*a[1] */
1754         mull    %edx
1755         addl    %eax,           %eax
1756         adcl    %edx,           %edx
1757         adcl    $0,             %ebx
1758         addl    %eax,           %ecx
1759         adcl    %edx,           %ebp
1760         movl    8(%esi),        %eax
1761         adcl    $0,             %ebx
1762         /* sqr a[2]*a[2] */
1763         mull    %eax
1764         addl    %eax,           %ecx
1765         adcl    %edx,           %ebp
1766         movl    (%esi),         %edx
1767         adcl    $0,             %ebx
1768         movl    %ecx,           16(%edi)
1769         movl    20(%esi),       %eax
1770         /* saved r[4] */
1771         /* ############### Calculate word 5 */
1772         xorl    %ecx,           %ecx
1773         /* sqr a[5]*a[0] */
1774         mull    %edx
1775         addl    %eax,           %eax
1776         adcl    %edx,           %edx
1777         adcl    $0,             %ecx
1778         addl    %eax,           %ebp
1779         adcl    %edx,           %ebx
1780         movl    16(%esi),       %eax
1781         adcl    $0,             %ecx
1782         movl    4(%esi),        %edx
1783         /* sqr a[4]*a[1] */
1784         mull    %edx
1785         addl    %eax,           %eax
1786         adcl    %edx,           %edx
1787         adcl    $0,             %ecx
1788         addl    %eax,           %ebp
1789         adcl    %edx,           %ebx
1790         movl    12(%esi),       %eax
1791         adcl    $0,             %ecx
1792         movl    8(%esi),        %edx
1793         /* sqr a[3]*a[2] */
1794         mull    %edx
1795         addl    %eax,           %eax
1796         adcl    %edx,           %edx
1797         adcl    $0,             %ecx
1798         addl    %eax,           %ebp
1799         adcl    %edx,           %ebx
1800         movl    24(%esi),       %eax
1801         adcl    $0,             %ecx
1802         movl    %ebp,           20(%edi)
1803         movl    (%esi),         %edx
1804         /* saved r[5] */
1805         /* ############### Calculate word 6 */
1806         xorl    %ebp,           %ebp
1807         /* sqr a[6]*a[0] */
1808         mull    %edx
1809         addl    %eax,           %eax
1810         adcl    %edx,           %edx
1811         adcl    $0,             %ebp
1812         addl    %eax,           %ebx
1813         adcl    %edx,           %ecx
1814         movl    20(%esi),       %eax
1815         adcl    $0,             %ebp
1816         movl    4(%esi),        %edx
1817         /* sqr a[5]*a[1] */
1818         mull    %edx
1819         addl    %eax,           %eax
1820         adcl    %edx,           %edx
1821         adcl    $0,             %ebp
1822         addl    %eax,           %ebx
1823         adcl    %edx,           %ecx
1824         movl    16(%esi),       %eax
1825         adcl    $0,             %ebp
1826         movl    8(%esi),        %edx
1827         /* sqr a[4]*a[2] */
1828         mull    %edx
1829         addl    %eax,           %eax
1830         adcl    %edx,           %edx
1831         adcl    $0,             %ebp
1832         addl    %eax,           %ebx
1833         adcl    %edx,           %ecx
1834         movl    12(%esi),       %eax
1835         adcl    $0,             %ebp
1836         /* sqr a[3]*a[3] */
1837         mull    %eax
1838         addl    %eax,           %ebx
1839         adcl    %edx,           %ecx
1840         movl    (%esi),         %edx
1841         adcl    $0,             %ebp
1842         movl    %ebx,           24(%edi)
1843         movl    28(%esi),       %eax
1844         /* saved r[6] */
1845         /* ############### Calculate word 7 */
1846         xorl    %ebx,           %ebx
1847         /* sqr a[7]*a[0] */
1848         mull    %edx
1849         addl    %eax,           %eax
1850         adcl    %edx,           %edx
1851         adcl    $0,             %ebx
1852         addl    %eax,           %ecx
1853         adcl    %edx,           %ebp
1854         movl    24(%esi),       %eax
1855         adcl    $0,             %ebx
1856         movl    4(%esi),        %edx
1857         /* sqr a[6]*a[1] */
1858         mull    %edx
1859         addl    %eax,           %eax
1860         adcl    %edx,           %edx
1861         adcl    $0,             %ebx
1862         addl    %eax,           %ecx
1863         adcl    %edx,           %ebp
1864         movl    20(%esi),       %eax
1865         adcl    $0,             %ebx
1866         movl    8(%esi),        %edx
1867         /* sqr a[5]*a[2] */
1868         mull    %edx
1869         addl    %eax,           %eax
1870         adcl    %edx,           %edx
1871         adcl    $0,             %ebx
1872         addl    %eax,           %ecx
1873         adcl    %edx,           %ebp
1874         movl    16(%esi),       %eax
1875         adcl    $0,             %ebx
1876         movl    12(%esi),       %edx
1877         /* sqr a[4]*a[3] */
1878         mull    %edx
1879         addl    %eax,           %eax
1880         adcl    %edx,           %edx
1881         adcl    $0,             %ebx
1882         addl    %eax,           %ecx
1883         adcl    %edx,           %ebp
1884         movl    28(%esi),       %eax
1885         adcl    $0,             %ebx
1886         movl    %ecx,           28(%edi)
1887         movl    4(%esi),        %edx
1888         /* saved r[7] */
1889         /* ############### Calculate word 8 */
1890         xorl    %ecx,           %ecx
1891         /* sqr a[7]*a[1] */
1892         mull    %edx
1893         addl    %eax,           %eax
1894         adcl    %edx,           %edx
1895         adcl    $0,             %ecx
1896         addl    %eax,           %ebp
1897         adcl    %edx,           %ebx
1898         movl    24(%esi),       %eax
1899         adcl    $0,             %ecx
1900         movl    8(%esi),        %edx
1901         /* sqr a[6]*a[2] */
1902         mull    %edx
1903         addl    %eax,           %eax
1904         adcl    %edx,           %edx
1905         adcl    $0,             %ecx
1906         addl    %eax,           %ebp
1907         adcl    %edx,           %ebx
1908         movl    20(%esi),       %eax
1909         adcl    $0,             %ecx
1910         movl    12(%esi),       %edx
1911         /* sqr a[5]*a[3] */
1912         mull    %edx
1913         addl    %eax,           %eax
1914         adcl    %edx,           %edx
1915         adcl    $0,             %ecx
1916         addl    %eax,           %ebp
1917         adcl    %edx,           %ebx
1918         movl    16(%esi),       %eax
1919         adcl    $0,             %ecx
1920         /* sqr a[4]*a[4] */
1921         mull    %eax
1922         addl    %eax,           %ebp
1923         adcl    %edx,           %ebx
1924         movl    8(%esi),        %edx
1925         adcl    $0,             %ecx
1926         movl    %ebp,           32(%edi)
1927         movl    28(%esi),       %eax
1928         /* saved r[8] */
1929         /* ############### Calculate word 9 */
1930         xorl    %ebp,           %ebp
1931         /* sqr a[7]*a[2] */
1932         mull    %edx
1933         addl    %eax,           %eax
1934         adcl    %edx,           %edx
1935         adcl    $0,             %ebp
1936         addl    %eax,           %ebx
1937         adcl    %edx,           %ecx
1938         movl    24(%esi),       %eax
1939         adcl    $0,             %ebp
1940         movl    12(%esi),       %edx
1941         /* sqr a[6]*a[3] */
1942         mull    %edx
1943         addl    %eax,           %eax
1944         adcl    %edx,           %edx
1945         adcl    $0,             %ebp
1946         addl    %eax,           %ebx
1947         adcl    %edx,           %ecx
1948         movl    20(%esi),       %eax
1949         adcl    $0,             %ebp
1950         movl    16(%esi),       %edx
1951         /* sqr a[5]*a[4] */
1952         mull    %edx
1953         addl    %eax,           %eax
1954         adcl    %edx,           %edx
1955         adcl    $0,             %ebp
1956         addl    %eax,           %ebx
1957         adcl    %edx,           %ecx
1958         movl    28(%esi),       %eax
1959         adcl    $0,             %ebp
1960         movl    %ebx,           36(%edi)
1961         movl    12(%esi),       %edx
1962         /* saved r[9] */
1963         /* ############### Calculate word 10 */
1964         xorl    %ebx,           %ebx
1965         /* sqr a[7]*a[3] */
1966         mull    %edx
1967         addl    %eax,           %eax
1968         adcl    %edx,           %edx
1969         adcl    $0,             %ebx
1970         addl    %eax,           %ecx
1971         adcl    %edx,           %ebp
1972         movl    24(%esi),       %eax
1973         adcl    $0,             %ebx
1974         movl    16(%esi),       %edx
1975         /* sqr a[6]*a[4] */
1976         mull    %edx
1977         addl    %eax,           %eax
1978         adcl    %edx,           %edx
1979         adcl    $0,             %ebx
1980         addl    %eax,           %ecx
1981         adcl    %edx,           %ebp
1982         movl    20(%esi),       %eax
1983         adcl    $0,             %ebx
1984         /* sqr a[5]*a[5] */
1985         mull    %eax
1986         addl    %eax,           %ecx
1987         adcl    %edx,           %ebp
1988         movl    16(%esi),       %edx
1989         adcl    $0,             %ebx
1990         movl    %ecx,           40(%edi)
1991         movl    28(%esi),       %eax
1992         /* saved r[10] */
1993         /* ############### Calculate word 11 */
1994         xorl    %ecx,           %ecx
1995         /* sqr a[7]*a[4] */
1996         mull    %edx
1997         addl    %eax,           %eax
1998         adcl    %edx,           %edx
1999         adcl    $0,             %ecx
2000         addl    %eax,           %ebp
2001         adcl    %edx,           %ebx
2002         movl    24(%esi),       %eax
2003         adcl    $0,             %ecx
2004         movl    20(%esi),       %edx
2005         /* sqr a[6]*a[5] */
2006         mull    %edx
2007         addl    %eax,           %eax
2008         adcl    %edx,           %edx
2009         adcl    $0,             %ecx
2010         addl    %eax,           %ebp
2011         adcl    %edx,           %ebx
2012         movl    28(%esi),       %eax
2013         adcl    $0,             %ecx
2014         movl    %ebp,           44(%edi)
2015         movl    20(%esi),       %edx
2016         /* saved r[11] */
2017         /* ############### Calculate word 12 */
2018         xorl    %ebp,           %ebp
2019         /* sqr a[7]*a[5] */
2020         mull    %edx
2021         addl    %eax,           %eax
2022         adcl    %edx,           %edx
2023         adcl    $0,             %ebp
2024         addl    %eax,           %ebx
2025         adcl    %edx,           %ecx
2026         movl    24(%esi),       %eax
2027         adcl    $0,             %ebp
2028         /* sqr a[6]*a[6] */
2029         mull    %eax
2030         addl    %eax,           %ebx
2031         adcl    %edx,           %ecx
2032         movl    24(%esi),       %edx
2033         adcl    $0,             %ebp
2034         movl    %ebx,           48(%edi)
2035         movl    28(%esi),       %eax
2036         /* saved r[12] */
2037         /* ############### Calculate word 13 */
2038         xorl    %ebx,           %ebx
2039         /* sqr a[7]*a[6] */
2040         mull    %edx
2041         addl    %eax,           %eax
2042         adcl    %edx,           %edx
2043         adcl    $0,             %ebx
2044         addl    %eax,           %ecx
2045         adcl    %edx,           %ebp
2046         movl    28(%esi),       %eax
2047         adcl    $0,             %ebx
2048         movl    %ecx,           52(%edi)
2049         /* saved r[13] */
2050         /* ############### Calculate word 14 */
2051         xorl    %ecx,           %ecx
2052         /* sqr a[7]*a[7] */
2053         mull    %eax
2054         addl    %eax,           %ebp
2055         adcl    %edx,           %ebx
2056         adcl    $0,             %ecx
2057         movl    %ebp,           56(%edi)
2058         /* saved r[14] */
2059         movl    %ebx,           60(%edi)
2060         popl    %ebx
2061         popl    %ebp
2062         popl    %edi
2063         popl    %esi
2064         ret
2065 .bn_sqr_comba8_end:
2066         SIZE(bn_sqr_comba8,.bn_sqr_comba8_end-bn_sqr_comba8)
2067 .ident  "desasm.pl"
2068 .text
2069         .align ALIGN
2070 .globl bn_sqr_comba4
2071         TYPE(bn_sqr_comba4,@function)
2072 bn_sqr_comba4:
2073         pushl   %esi
2074         pushl   %edi
2075         pushl   %ebp
2076         pushl   %ebx
2077         movl    20(%esp),       %edi
2078         movl    24(%esp),       %esi
2079         xorl    %ebx,           %ebx
2080         xorl    %ecx,           %ecx
2081         movl    (%esi),         %eax
2082         /* ############### Calculate word 0 */
2083         xorl    %ebp,           %ebp
2084         /* sqr a[0]*a[0] */
2085         mull    %eax
2086         addl    %eax,           %ebx
2087         adcl    %edx,           %ecx
2088         movl    (%esi),         %edx
2089         adcl    $0,             %ebp
2090         movl    %ebx,           (%edi)
2091         movl    4(%esi),        %eax
2092         /* saved r[0] */
2093         /* ############### Calculate word 1 */
2094         xorl    %ebx,           %ebx
2095         /* sqr a[1]*a[0] */
2096         mull    %edx
2097         addl    %eax,           %eax
2098         adcl    %edx,           %edx
2099         adcl    $0,             %ebx
2100         addl    %eax,           %ecx
2101         adcl    %edx,           %ebp
2102         movl    8(%esi),        %eax
2103         adcl    $0,             %ebx
2104         movl    %ecx,           4(%edi)
2105         movl    (%esi),         %edx
2106         /* saved r[1] */
2107         /* ############### Calculate word 2 */
2108         xorl    %ecx,           %ecx
2109         /* sqr a[2]*a[0] */
2110         mull    %edx
2111         addl    %eax,           %eax
2112         adcl    %edx,           %edx
2113         adcl    $0,             %ecx
2114         addl    %eax,           %ebp
2115         adcl    %edx,           %ebx
2116         movl    4(%esi),        %eax
2117         adcl    $0,             %ecx
2118         /* sqr a[1]*a[1] */
2119         mull    %eax
2120         addl    %eax,           %ebp
2121         adcl    %edx,           %ebx
2122         movl    (%esi),         %edx
2123         adcl    $0,             %ecx
2124         movl    %ebp,           8(%edi)
2125         movl    12(%esi),       %eax
2126         /* saved r[2] */
2127         /* ############### Calculate word 3 */
2128         xorl    %ebp,           %ebp
2129         /* sqr a[3]*a[0] */
2130         mull    %edx
2131         addl    %eax,           %eax
2132         adcl    %edx,           %edx
2133         adcl    $0,             %ebp
2134         addl    %eax,           %ebx
2135         adcl    %edx,           %ecx
2136         movl    8(%esi),        %eax
2137         adcl    $0,             %ebp
2138         movl    4(%esi),        %edx
2139         /* sqr a[2]*a[1] */
2140         mull    %edx
2141         addl    %eax,           %eax
2142         adcl    %edx,           %edx
2143         adcl    $0,             %ebp
2144         addl    %eax,           %ebx
2145         adcl    %edx,           %ecx
2146         movl    12(%esi),       %eax
2147         adcl    $0,             %ebp
2148         movl    %ebx,           12(%edi)
2149         movl    4(%esi),        %edx
2150         /* saved r[3] */
2151         /* ############### Calculate word 4 */
2152         xorl    %ebx,           %ebx
2153         /* sqr a[3]*a[1] */
2154         mull    %edx
2155         addl    %eax,           %eax
2156         adcl    %edx,           %edx
2157         adcl    $0,             %ebx
2158         addl    %eax,           %ecx
2159         adcl    %edx,           %ebp
2160         movl    8(%esi),        %eax
2161         adcl    $0,             %ebx
2162         /* sqr a[2]*a[2] */
2163         mull    %eax
2164         addl    %eax,           %ecx
2165         adcl    %edx,           %ebp
2166         movl    8(%esi),        %edx
2167         adcl    $0,             %ebx
2168         movl    %ecx,           16(%edi)
2169         movl    12(%esi),       %eax
2170         /* saved r[4] */
2171         /* ############### Calculate word 5 */
2172         xorl    %ecx,           %ecx
2173         /* sqr a[3]*a[2] */
2174         mull    %edx
2175         addl    %eax,           %eax
2176         adcl    %edx,           %edx
2177         adcl    $0,             %ecx
2178         addl    %eax,           %ebp
2179         adcl    %edx,           %ebx
2180         movl    12(%esi),       %eax
2181         adcl    $0,             %ecx
2182         movl    %ebp,           20(%edi)
2183         /* saved r[5] */
2184         /* ############### Calculate word 6 */
2185         xorl    %ebp,           %ebp
2186         /* sqr a[3]*a[3] */
2187         mull    %eax
2188         addl    %eax,           %ebx
2189         adcl    %edx,           %ecx
2190         adcl    $0,             %ebp
2191         movl    %ebx,           24(%edi)
2192         /* saved r[6] */
2193         movl    %ecx,           28(%edi)
2194         popl    %ebx
2195         popl    %ebp
2196         popl    %edi
2197         popl    %esi
2198         ret
2199 .bn_sqr_comba4_end:
2200         SIZE(bn_sqr_comba4,.bn_sqr_comba4_end-bn_sqr_comba4)
2201 .ident  "desasm.pl"