crypto/bn/asm/x86-mont.pl

   1 #!/usr/bin/env perl
   2
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
   5 # project. The module is, however, dual licensed under OpenSSL and
   6 # CRYPTOGAMS licenses depending on where you obtain it. For further
   7 # details see http://www.openssl.org/~appro/cryptogams/.
   8 # ====================================================================
   9
  10 # October 2005
  11 #
  12 # This is a "teaser" code, as it can be improved in several ways...
  13 # First of all non-SSE2 path should be implemented (yes, for now it
  14 # performs Montgomery multiplication/convolution only on SSE2-capable
  15 # CPUs such as P4, others fall down to original code). Then inner loop
  16 # can be unrolled and modulo-scheduled to improve ILP and possibly
  17 # moved to 128-bit XMM register bank (though it would require input
  18 # rearrangement and/or increase bus bandwidth utilization). Dedicated
  19 # squaring procedure should give further performance improvement...
  20 # Yet, for being draft, the code improves rsa512 *sign* benchmark by
  21 # 110%(!), rsa1024 one - by 70% and rsa4096 - by 20%:-)
  22
  23 # December 2006
  24 #
  25 # Modulo-scheduling SSE2 loops results in further 15-20% improvement.
  26 # Integer-only code [being equipped with dedicated squaring procedure]
  27 # gives ~40% on rsa512 sign benchmark...
  28
  29 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  30 push(@INC,"${dir}","${dir}../../perlasm");
  31 require "x86asm.pl";
  32
  33 $output = pop;
  34 open STDOUT,">$output";
  35
  36 &asm_init($ARGV[0],$0);
  37
  38 $sse2=0;
  39 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  40
  41 &external_label("OPENSSL_ia32cap_P") if ($sse2);
  42
  43 &function_begin("bn_mul_mont");
  44
  45 $i="edx";
  46 $j="ecx";
  47 $ap="esi";      $tp="esi";              # overlapping variables!!!
  48 $rp="edi";      $bp="edi";              # overlapping variables!!!
  49 $np="ebp";
  50 $num="ebx";
  51
  52 $_num=&DWP(4*0,"esp");                  # stack top layout
  53 $_rp=&DWP(4*1,"esp");
  54 $_ap=&DWP(4*2,"esp");
  55 $_bp=&DWP(4*3,"esp");
  56 $_np=&DWP(4*4,"esp");
  57 $_n0=&DWP(4*5,"esp");   $_n0q=&QWP(4*5,"esp");
  58 $_sp=&DWP(4*6,"esp");
  59 $_bpend=&DWP(4*7,"esp");
  60 $frame=32;                              # size of above frame rounded up to 16n
  61
  62         &xor    ("eax","eax");
  63         &mov    ("edi",&wparam(5));     # int num
  64         &cmp    ("edi",4);
  65         &jl     (&label("just_leave"));
  66
  67         &lea    ("esi",&wparam(0));     # put aside pointer to argument block
  68         &lea    ("edx",&wparam(1));     # load ap
  69         &mov    ("ebp","esp");          # saved stack pointer!
  70         &add    ("edi",2);              # extra two words on top of tp
  71         &neg    ("edi");
  72         &lea    ("esp",&DWP(-$frame,"esp","edi",4));    # alloca($frame+4*(num+2))
  73         &neg    ("edi");
  74
  75         # minimize cache contention by arraning 2K window between stack
  76         # pointer and ap argument [np is also position sensitive vector,
  77         # but it's assumed to be near ap, as it's allocated at ~same
  78         # time].
  79         &mov    ("eax","esp");
  80         &sub    ("eax","edx");
  81         &and    ("eax",2047);
  82         &sub    ("esp","eax");          # this aligns sp and ap modulo 2048
  83
  84         &xor    ("edx","esp");
  85         &and    ("edx",2048);
  86         &xor    ("edx",2048);
  87         &sub    ("esp","edx");          # this splits them apart modulo 4096
  88
  89         &and    ("esp",-64);            # align to cache line
  90
  91         # Some OSes (Windows) insist on stack being "wired" to
  92         # physical memory in strictly sequential manner, i.e. if stack
  93         # allocation spans two pages, then reference to farmost one can
  94         # be punishable by SEGV. But page walking can do good even on
  95         # other OSes, because it guarantees that villain thread hits
  96         # the guard page before it can make damage to innocent one...
  97         &mov    ("eax","ebp");
  98         &sub    ("eax","esp");
  99         &and    ("eax",-4096);
 100 &set_label("page_walk");
 101         &mov    ("edx",&DWP(0,"esp","eax"));
 102         &sub    ("eax",4096);
 103         &data_byte(0x2e);
 104         &jnc    (&label("page_walk"));
 105
 106         ################################# load argument block...
 107         &mov    ("eax",&DWP(0*4,"esi"));# BN_ULONG *rp
 108         &mov    ("ebx",&DWP(1*4,"esi"));# const BN_ULONG *ap
 109         &mov    ("ecx",&DWP(2*4,"esi"));# const BN_ULONG *bp
 110         &mov    ("edx",&DWP(3*4,"esi"));# const BN_ULONG *np
 111         &mov    ("esi",&DWP(4*4,"esi"));# const BN_ULONG *n0
 112         #&mov   ("edi",&DWP(5*4,"esi"));# int num
 113
 114         &mov    ("esi",&DWP(0,"esi"));  # pull n0[0]
 115         &mov    ($_rp,"eax");           # ... save a copy of argument block
 116         &mov    ($_ap,"ebx");
 117         &mov    ($_bp,"ecx");
 118         &mov    ($_np,"edx");
 119         &mov    ($_n0,"esi");
 120         &lea    ($num,&DWP(-3,"edi"));  # num=num-1 to assist modulo-scheduling
 121         #&mov   ($_num,$num);           # redundant as $num is not reused
 122         &mov    ($_sp,"ebp");           # saved stack pointer!
 123 \f
 124 if($sse2) {
 125 $acc0="mm0";    # mmx register bank layout
 126 $acc1="mm1";
 127 $car0="mm2";
 128 $car1="mm3";
 129 $mul0="mm4";
 130 $mul1="mm5";
 131 $temp="mm6";
 132 $mask="mm7";
 133
 134         &picmeup("eax","OPENSSL_ia32cap_P");
 135         &bt     (&DWP(0,"eax"),26);
 136         &jnc    (&label("non_sse2"));
 137
 138         &mov    ("eax",-1);
 139         &movd   ($mask,"eax");          # mask 32 lower bits
 140
 141         &mov    ($ap,$_ap);             # load input pointers
 142         &mov    ($bp,$_bp);
 143         &mov    ($np,$_np);
 144
 145         &xor    ($i,$i);                # i=0
 146         &xor    ($j,$j);                # j=0
 147
 148         &movd   ($mul0,&DWP(0,$bp));            # bp[0]
 149         &movd   ($mul1,&DWP(0,$ap));            # ap[0]
 150         &movd   ($car1,&DWP(0,$np));            # np[0]
 151
 152         &pmuludq($mul1,$mul0);                  # ap[0]*bp[0]
 153         &movq   ($car0,$mul1);
 154         &movq   ($acc0,$mul1);                  # I wish movd worked for
 155         &pand   ($acc0,$mask);                  # inter-register transfers
 156
 157         &pmuludq($mul1,$_n0q);                  # *=n0
 158
 159         &pmuludq($car1,$mul1);                  # "t[0]"*np[0]*n0
 160         &paddq  ($car1,$acc0);
 161
 162         &movd   ($acc1,&DWP(4,$np));            # np[1]
 163         &movd   ($acc0,&DWP(4,$ap));            # ap[1]
 164
 165         &psrlq  ($car0,32);
 166         &psrlq  ($car1,32);
 167
 168         &inc    ($j);                           # j++
 169 &set_label("1st",16);
 170         &pmuludq($acc0,$mul0);                  # ap[j]*bp[0]
 171         &pmuludq($acc1,$mul1);                  # np[j]*m1
 172         &paddq  ($car0,$acc0);                  # +=c0
 173         &paddq  ($car1,$acc1);                  # +=c1
 174
 175         &movq   ($acc0,$car0);
 176         &pand   ($acc0,$mask);
 177         &movd   ($acc1,&DWP(4,$np,$j,4));       # np[j+1]
 178         &paddq  ($car1,$acc0);                  # +=ap[j]*bp[0];
 179         &movd   ($acc0,&DWP(4,$ap,$j,4));       # ap[j+1]
 180         &psrlq  ($car0,32);
 181         &movd   (&DWP($frame-4,"esp",$j,4),$car1);      # tp[j-1]=
 182         &psrlq  ($car1,32);
 183
 184         &lea    ($j,&DWP(1,$j));
 185         &cmp    ($j,$num);
 186         &jl     (&label("1st"));
 187
 188         &pmuludq($acc0,$mul0);                  # ap[num-1]*bp[0]
 189         &pmuludq($acc1,$mul1);                  # np[num-1]*m1
 190         &paddq  ($car0,$acc0);                  # +=c0
 191         &paddq  ($car1,$acc1);                  # +=c1
 192
 193         &movq   ($acc0,$car0);
 194         &pand   ($acc0,$mask);
 195         &paddq  ($car1,$acc0);                  # +=ap[num-1]*bp[0];
 196         &movd   (&DWP($frame-4,"esp",$j,4),$car1);      # tp[num-2]=
 197
 198         &psrlq  ($car0,32);
 199         &psrlq  ($car1,32);
 200
 201         &paddq  ($car1,$car0);
 202         &movq   (&QWP($frame,"esp",$num,4),$car1);      # tp[num].tp[num-1]
 203 \f
 204         &inc    ($i);                           # i++
 205 &set_label("outer");
 206         &xor    ($j,$j);                        # j=0
 207
 208         &movd   ($mul0,&DWP(0,$bp,$i,4));       # bp[i]
 209         &movd   ($mul1,&DWP(0,$ap));            # ap[0]
 210         &movd   ($temp,&DWP($frame,"esp"));     # tp[0]
 211         &movd   ($car1,&DWP(0,$np));            # np[0]
 212         &pmuludq($mul1,$mul0);                  # ap[0]*bp[i]
 213
 214         &paddq  ($mul1,$temp);                  # +=tp[0]
 215         &movq   ($acc0,$mul1);
 216         &movq   ($car0,$mul1);
 217         &pand   ($acc0,$mask);
 218
 219         &pmuludq($mul1,$_n0q);                  # *=n0
 220
 221         &pmuludq($car1,$mul1);
 222         &paddq  ($car1,$acc0);
 223
 224         &movd   ($temp,&DWP($frame+4,"esp"));   # tp[1]
 225         &movd   ($acc1,&DWP(4,$np));            # np[1]
 226         &movd   ($acc0,&DWP(4,$ap));            # ap[1]
 227
 228         &psrlq  ($car0,32);
 229         &psrlq  ($car1,32);
 230         &paddq  ($car0,$temp);                  # +=tp[1]
 231
 232         &inc    ($j);                           # j++
 233         &dec    ($num);
 234 &set_label("inner");
 235         &pmuludq($acc0,$mul0);                  # ap[j]*bp[i]
 236         &pmuludq($acc1,$mul1);                  # np[j]*m1
 237         &paddq  ($car0,$acc0);                  # +=c0
 238         &paddq  ($car1,$acc1);                  # +=c1
 239
 240         &movq   ($acc0,$car0);
 241         &movd   ($temp,&DWP($frame+4,"esp",$j,4));# tp[j+1]
 242         &pand   ($acc0,$mask);
 243         &movd   ($acc1,&DWP(4,$np,$j,4));       # np[j+1]
 244         &paddq  ($car1,$acc0);                  # +=ap[j]*bp[i]+tp[j]
 245         &movd   ($acc0,&DWP(4,$ap,$j,4));       # ap[j+1]
 246         &psrlq  ($car0,32);
 247         &movd   (&DWP($frame-4,"esp",$j,4),$car1);# tp[j-1]=
 248         &psrlq  ($car1,32);
 249         &paddq  ($car0,$temp);                  # +=tp[j+1]
 250
 251         &dec    ($num);
 252         &lea    ($j,&DWP(1,$j));                # j++
 253         &jnz    (&label("inner"));
 254
 255         &mov    ($num,$j);
 256         &pmuludq($acc0,$mul0);                  # ap[num-1]*bp[i]
 257         &pmuludq($acc1,$mul1);                  # np[num-1]*m1
 258         &paddq  ($car0,$acc0);                  # +=c0
 259         &paddq  ($car1,$acc1);                  # +=c1
 260
 261         &movq   ($acc0,$car0);
 262         &pand   ($acc0,$mask);
 263         &paddq  ($car1,$acc0);                  # +=ap[num-1]*bp[i]+tp[num-1]
 264         &movd   (&DWP($frame-4,"esp",$j,4),$car1);      # tp[num-2]=
 265         &psrlq  ($car0,32);
 266         &psrlq  ($car1,32);
 267
 268         &movd   ($temp,&DWP($frame+4,"esp",$num,4));    # += tp[num]
 269         &paddq  ($car1,$car0);
 270         &paddq  ($car1,$temp);
 271         &movq   (&QWP($frame,"esp",$num,4),$car1);      # tp[num].tp[num-1]
 272
 273         &lea    ($i,&DWP(1,$i));                # i++
 274         &cmp    ($i,$num);
 275         &jle    (&label("outer"));
 276
 277         &emms   ();                             # done with mmx bank
 278         &jmp    (&label("common_tail"));
 279
 280 &set_label("non_sse2",16);
 281 }
 282 \f
 283 if (0) {
 284         &mov    ("esp",$_sp);
 285         &xor    ("eax","eax");  # signal "not fast enough [yet]"
 286         &jmp    (&label("just_leave"));
 287         # While the below code provides competitive performance for
 288         # all key lengthes on modern Intel cores, it's still more
 289         # than 10% slower for 4096-bit key elsewhere:-( "Competitive"
 290         # means compared to the original integer-only assembler.
 291         # 512-bit RSA sign is better by ~40%, but that's about all
 292         # one can say about all CPUs...
 293 } else {
 294 $inp="esi";     # integer path uses these registers differently
 295 $word="edi";
 296 $carry="ebp";
 297
 298         &mov    ($inp,$_ap);
 299         &lea    ($carry,&DWP(1,$num));
 300         &mov    ($word,$_bp);
 301         &xor    ($j,$j);                                # j=0
 302         &mov    ("edx",$inp);
 303         &and    ($carry,1);                             # see if num is even
 304         &sub    ("edx",$word);                          # see if ap==bp
 305         &lea    ("eax",&DWP(4,$word,$num,4));           # &bp[num]
 306         &or     ($carry,"edx");
 307         &mov    ($word,&DWP(0,$word));                  # bp[0]
 308         &jz     (&label("bn_sqr_mont"));
 309         &mov    ($_bpend,"eax");
 310         &mov    ("eax",&DWP(0,$inp));
 311         &xor    ("edx","edx");
 312
 313 &set_label("mull",16);
 314         &mov    ($carry,"edx");
 315         &mul    ($word);                                # ap[j]*bp[0]
 316         &add    ($carry,"eax");
 317         &lea    ($j,&DWP(1,$j));
 318         &adc    ("edx",0);
 319         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j+1]
 320         &cmp    ($j,$num);
 321         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 322         &jl     (&label("mull"));
 323
 324         &mov    ($carry,"edx");
 325         &mul    ($word);                                # ap[num-1]*bp[0]
 326          &mov   ($word,$_n0);
 327         &add    ("eax",$carry);
 328          &mov   ($inp,$_np);
 329         &adc    ("edx",0);
 330          &imul  ($word,&DWP($frame,"esp"));             # n0*tp[0]
 331
 332         &mov    (&DWP($frame,"esp",$num,4),"eax");      # tp[num-1]=
 333         &xor    ($j,$j);
 334         &mov    (&DWP($frame+4,"esp",$num,4),"edx");    # tp[num]=
 335         &mov    (&DWP($frame+8,"esp",$num,4),$j);       # tp[num+1]=
 336
 337         &mov    ("eax",&DWP(0,$inp));                   # np[0]
 338         &mul    ($word);                                # np[0]*m
 339         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 340         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 341         &adc    ("edx",0);
 342         &inc    ($j);
 343
 344         &jmp    (&label("2ndmadd"));
 345 \f\f
 346 &set_label("1stmadd",16);
 347         &mov    ($carry,"edx");
 348         &mul    ($word);                                # ap[j]*bp[i]
 349         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 350         &lea    ($j,&DWP(1,$j));
 351         &adc    ("edx",0);
 352         &add    ($carry,"eax");
 353         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j+1]
 354         &adc    ("edx",0);
 355         &cmp    ($j,$num);
 356         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 357         &jl     (&label("1stmadd"));
 358
 359         &mov    ($carry,"edx");
 360         &mul    ($word);                                # ap[num-1]*bp[i]
 361         &add    ("eax",&DWP($frame,"esp",$num,4));      # +=tp[num-1]
 362          &mov   ($word,$_n0);
 363         &adc    ("edx",0);
 364          &mov   ($inp,$_np);
 365         &add    ($carry,"eax");
 366         &adc    ("edx",0);
 367          &imul  ($word,&DWP($frame,"esp"));             # n0*tp[0]
 368
 369         &xor    ($j,$j);
 370         &add    ("edx",&DWP($frame+4,"esp",$num,4));    # carry+=tp[num]
 371         &mov    (&DWP($frame,"esp",$num,4),$carry);     # tp[num-1]=
 372         &adc    ($j,0);
 373          &mov   ("eax",&DWP(0,$inp));                   # np[0]
 374         &mov    (&DWP($frame+4,"esp",$num,4),"edx");    # tp[num]=
 375         &mov    (&DWP($frame+8,"esp",$num,4),$j);       # tp[num+1]=
 376
 377         &mul    ($word);                                # np[0]*m
 378         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 379         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 380         &adc    ("edx",0);
 381         &mov    ($j,1);
 382 \f
 383 &set_label("2ndmadd",16);
 384         &mov    ($carry,"edx");
 385         &mul    ($word);                                # np[j]*m
 386         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 387         &lea    ($j,&DWP(1,$j));
 388         &adc    ("edx",0);
 389         &add    ($carry,"eax");
 390         &mov    ("eax",&DWP(0,$inp,$j,4));              # np[j+1]
 391         &adc    ("edx",0);
 392         &cmp    ($j,$num);
 393         &mov    (&DWP($frame-8,"esp",$j,4),$carry);     # tp[j-1]=
 394         &jl     (&label("2ndmadd"));
 395
 396         &mov    ($carry,"edx");
 397         &mul    ($word);                                # np[j]*m
 398         &add    ($carry,&DWP($frame,"esp",$num,4));     # +=tp[num-1]
 399         &adc    ("edx",0);
 400         &add    ($carry,"eax");
 401         &adc    ("edx",0);
 402         &mov    (&DWP($frame-4,"esp",$num,4),$carry);   # tp[num-2]=
 403
 404         &xor    ("eax","eax");
 405          &mov   ($j,$_bp);                              # &bp[i]
 406         &add    ("edx",&DWP($frame+4,"esp",$num,4));    # carry+=tp[num]
 407         &adc    ("eax",&DWP($frame+8,"esp",$num,4));    # +=tp[num+1]
 408          &lea   ($j,&DWP(4,$j));
 409         &mov    (&DWP($frame,"esp",$num,4),"edx");      # tp[num-1]=
 410          &cmp   ($j,$_bpend);
 411         &mov    (&DWP($frame+4,"esp",$num,4),"eax");    # tp[num]=
 412         &je     (&label("common_tail"));
 413
 414         &mov    ($word,&DWP(0,$j));                     # bp[i+1]
 415         &mov    ($inp,$_ap);
 416         &mov    ($_bp,$j);                              # &bp[++i]
 417         &xor    ($j,$j);
 418         &xor    ("edx","edx");
 419         &mov    ("eax",&DWP(0,$inp));
 420         &jmp    (&label("1stmadd"));
 421 \f
 422 &set_label("bn_sqr_mont",16);
 423 $sbit=$num;
 424         &mov    ($_num,$num);
 425         &mov    ($_bp,$j);                              # i=0
 426
 427         &mov    ("eax",$word);                          # ap[0]
 428         &mul    ($word);                                # ap[0]*ap[0]
 429         &mov    (&DWP($frame,"esp"),"eax");             # tp[0]=
 430         &mov    ($sbit,"edx");
 431         &shr    ("edx",1);
 432         &and    ($sbit,1);
 433         &inc    ($j);
 434 &set_label("sqr",16);
 435         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j]
 436         &mov    ($carry,"edx");
 437         &mul    ($word);                                # ap[j]*ap[0]
 438         &add    ("eax",$carry);
 439         &lea    ($j,&DWP(1,$j));
 440         &adc    ("edx",0);
 441         &lea    ($carry,&DWP(0,$sbit,"eax",2));
 442         &shr    ("eax",31);
 443         &cmp    ($j,$_num);
 444         &mov    ($sbit,"eax");
 445         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 446         &jl     (&label("sqr"));
 447
 448         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[num-1]
 449         &mov    ($carry,"edx");
 450         &mul    ($word);                                # ap[num-1]*ap[0]
 451         &add    ("eax",$carry);
 452          &mov   ($word,$_n0);
 453         &adc    ("edx",0);
 454          &mov   ($inp,$_np);
 455         &lea    ($carry,&DWP(0,$sbit,"eax",2));
 456          &imul  ($word,&DWP($frame,"esp"));             # n0*tp[0]
 457         &shr    ("eax",31);
 458         &mov    (&DWP($frame,"esp",$j,4),$carry);       # tp[num-1]=
 459
 460         &lea    ($carry,&DWP(0,"eax","edx",2));
 461          &mov   ("eax",&DWP(0,$inp));                   # np[0]
 462         &shr    ("edx",31);
 463         &mov    (&DWP($frame+4,"esp",$j,4),$carry);     # tp[num]=
 464         &mov    (&DWP($frame+8,"esp",$j,4),"edx");      # tp[num+1]=
 465
 466         &mul    ($word);                                # np[0]*m
 467         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 468         &mov    ($num,$j);
 469         &adc    ("edx",0);
 470         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 471         &mov    ($j,1);
 472 \f\f
 473 &set_label("3rdmadd",16);
 474         &mov    ($carry,"edx");
 475         &mul    ($word);                                # np[j]*m
 476         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 477         &adc    ("edx",0);
 478         &add    ($carry,"eax");
 479         &mov    ("eax",&DWP(4,$inp,$j,4));              # np[j+1]
 480         &adc    ("edx",0);
 481         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j-1]=
 482
 483         &mov    ($carry,"edx");
 484         &mul    ($word);                                # np[j+1]*m
 485         &add    ($carry,&DWP($frame+4,"esp",$j,4));     # +=tp[j+1]
 486         &lea    ($j,&DWP(2,$j));
 487         &adc    ("edx",0);
 488         &add    ($carry,"eax");
 489         &mov    ("eax",&DWP(0,$inp,$j,4));              # np[j+2]
 490         &adc    ("edx",0);
 491         &cmp    ($j,$num);
 492         &mov    (&DWP($frame-8,"esp",$j,4),$carry);     # tp[j]=
 493         &jl     (&label("3rdmadd"));
 494
 495         &mov    ($carry,"edx");
 496         &mul    ($word);                                # np[j]*m
 497         &add    ($carry,&DWP($frame,"esp",$num,4));     # +=tp[num-1]
 498         &adc    ("edx",0);
 499         &add    ($carry,"eax");
 500         &adc    ("edx",0);
 501         &mov    (&DWP($frame-4,"esp",$num,4),$carry);   # tp[num-2]=
 502
 503         &mov    ($j,$_bp);                              # i
 504         &xor    ("eax","eax");
 505         &mov    ($inp,$_ap);
 506         &add    ("edx",&DWP($frame+4,"esp",$num,4));    # carry+=tp[num]
 507         &adc    ("eax",&DWP($frame+8,"esp",$num,4));    # +=tp[num+1]
 508         &mov    (&DWP($frame,"esp",$num,4),"edx");      # tp[num-1]=
 509         &cmp    ($j,$num);
 510         &mov    (&DWP($frame+4,"esp",$num,4),"eax");    # tp[num]=
 511         &je     (&label("common_tail"));
 512 \f
 513         &mov    ($word,&DWP(4,$inp,$j,4));              # ap[i]
 514         &lea    ($j,&DWP(1,$j));
 515         &mov    ("eax",$word);
 516         &mov    ($_bp,$j);                              # ++i
 517         &mul    ($word);                                # ap[i]*ap[i]
 518         &add    ("eax",&DWP($frame,"esp",$j,4));        # +=tp[i]
 519         &adc    ("edx",0);
 520         &mov    (&DWP($frame,"esp",$j,4),"eax");        # tp[i]=
 521         &xor    ($carry,$carry);
 522         &cmp    ($j,$num);
 523         &lea    ($j,&DWP(1,$j));
 524         &je     (&label("sqrlast"));
 525
 526         &mov    ($sbit,"edx");                          # zaps $num
 527         &shr    ("edx",1);
 528         &and    ($sbit,1);
 529 &set_label("sqradd",16);
 530         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j]
 531         &mov    ($carry,"edx");
 532         &mul    ($word);                                # ap[j]*ap[i]
 533         &add    ("eax",$carry);
 534         &lea    ($carry,&DWP(0,"eax","eax"));
 535         &adc    ("edx",0);
 536         &shr    ("eax",31);
 537         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 538         &lea    ($j,&DWP(1,$j));
 539         &adc    ("eax",0);
 540         &add    ($carry,$sbit);
 541         &adc    ("eax",0);
 542         &cmp    ($j,$_num);
 543         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 544         &mov    ($sbit,"eax");
 545         &jle    (&label("sqradd"));
 546
 547         &mov    ($carry,"edx");
 548         &add    ("edx","edx");
 549         &shr    ($carry,31);
 550         &add    ("edx",$sbit);
 551         &adc    ($carry,0);
 552 &set_label("sqrlast");
 553         &mov    ($word,$_n0);
 554         &mov    ($inp,$_np);
 555         &imul   ($word,&DWP($frame,"esp"));             # n0*tp[0]
 556
 557         &add    ("edx",&DWP($frame,"esp",$j,4));        # +=tp[num]
 558         &mov    ("eax",&DWP(0,$inp));                   # np[0]
 559         &adc    ($carry,0);
 560         &mov    (&DWP($frame,"esp",$j,4),"edx");        # tp[num]=
 561         &mov    (&DWP($frame+4,"esp",$j,4),$carry);     # tp[num+1]=
 562
 563         &mul    ($word);                                # np[0]*m
 564         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 565         &lea    ($num,&DWP(-1,$j));
 566         &adc    ("edx",0);
 567         &mov    ($j,1);
 568         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 569
 570         &jmp    (&label("3rdmadd"));
 571 }
 572 \f
 573 &set_label("common_tail",16);
 574         &mov    ($np,$_np);                     # load modulus pointer
 575         &mov    ($rp,$_rp);                     # load result pointer
 576         &lea    ($tp,&DWP($frame,"esp"));       # [$ap and $bp are zapped]
 577
 578         &mov    ("eax",&DWP(0,$tp));            # tp[0]
 579         &mov    ($j,$num);                      # j=num-1
 580         &xor    ($i,$i);                        # i=0 and clear CF!
 581
 582 &set_label("sub",16);
 583         &sbb    ("eax",&DWP(0,$np,$i,4));
 584         &mov    (&DWP(0,$rp,$i,4),"eax");       # rp[i]=tp[i]-np[i]
 585         &dec    ($j);                           # doesn't affect CF!
 586         &mov    ("eax",&DWP(4,$tp,$i,4));       # tp[i+1]
 587         &lea    ($i,&DWP(1,$i));                # i++
 588         &jge    (&label("sub"));
 589
 590         &sbb    ("eax",0);                      # handle upmost overflow bit
 591         &and    ($tp,"eax");
 592         &not    ("eax");
 593         &mov    ($np,$rp);
 594         &and    ($np,"eax");
 595         &or     ($tp,$np);                      # tp=carry?tp:rp
 596
 597 &set_label("copy",16);                          # copy or in-place refresh
 598         &mov    ("eax",&DWP(0,$tp,$num,4));
 599         &mov    (&DWP(0,$rp,$num,4),"eax");     # rp[i]=tp[i]
 600         &mov    (&DWP($frame,"esp",$num,4),$j); # zap temporary vector
 601         &dec    ($num);
 602         &jge    (&label("copy"));
 603
 604         &mov    ("esp",$_sp);           # pull saved stack pointer
 605         &mov    ("eax",1);
 606 &set_label("just_leave");
 607 &function_end("bn_mul_mont");
 608
 609 &asciz("Montgomery Multiplication for x86, CRYPTOGAMS by <appro\@openssl.org>");
 610
 611 &asm_finish();
 612
 613 close STDOUT;