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         # An OS-agnostic version of __chkstk.
  92         #
  93         # Some OSes (Windows) insist on stack being "wired" to
  94         # physical memory in strictly sequential manner, i.e. if stack
  95         # allocation spans two pages, then reference to farmost one can
  96         # be punishable by SEGV. But page walking can do good even on
  97         # other OSes, because it guarantees that villain thread hits
  98         # the guard page before it can make damage to innocent one...
  99         &mov    ("eax","ebp");
 100         &sub    ("eax","esp");
 101         &and    ("eax",-4096);
 102 &set_label("page_walk");
 103         &mov    ("edx",&DWP(0,"esp","eax"));
 104         &sub    ("eax",4096);
 105         &data_byte(0x2e);
 106         &jnc    (&label("page_walk"));
 107
 108         ################################# load argument block...
 109         &mov    ("eax",&DWP(0*4,"esi"));# BN_ULONG *rp
 110         &mov    ("ebx",&DWP(1*4,"esi"));# const BN_ULONG *ap
 111         &mov    ("ecx",&DWP(2*4,"esi"));# const BN_ULONG *bp
 112         &mov    ("edx",&DWP(3*4,"esi"));# const BN_ULONG *np
 113         &mov    ("esi",&DWP(4*4,"esi"));# const BN_ULONG *n0
 114         #&mov   ("edi",&DWP(5*4,"esi"));# int num
 115
 116         &mov    ("esi",&DWP(0,"esi"));  # pull n0[0]
 117         &mov    ($_rp,"eax");           # ... save a copy of argument block
 118         &mov    ($_ap,"ebx");
 119         &mov    ($_bp,"ecx");
 120         &mov    ($_np,"edx");
 121         &mov    ($_n0,"esi");
 122         &lea    ($num,&DWP(-3,"edi"));  # num=num-1 to assist modulo-scheduling
 123         #&mov   ($_num,$num);           # redundant as $num is not reused
 124         &mov    ($_sp,"ebp");           # saved stack pointer!
 125 \f
 126 if($sse2) {
 127 $acc0="mm0";    # mmx register bank layout
 128 $acc1="mm1";
 129 $car0="mm2";
 130 $car1="mm3";
 131 $mul0="mm4";
 132 $mul1="mm5";
 133 $temp="mm6";
 134 $mask="mm7";
 135
 136         &picmeup("eax","OPENSSL_ia32cap_P");
 137         &bt     (&DWP(0,"eax"),26);
 138         &jnc    (&label("non_sse2"));
 139
 140         &mov    ("eax",-1);
 141         &movd   ($mask,"eax");          # mask 32 lower bits
 142
 143         &mov    ($ap,$_ap);             # load input pointers
 144         &mov    ($bp,$_bp);
 145         &mov    ($np,$_np);
 146
 147         &xor    ($i,$i);                # i=0
 148         &xor    ($j,$j);                # j=0
 149
 150         &movd   ($mul0,&DWP(0,$bp));            # bp[0]
 151         &movd   ($mul1,&DWP(0,$ap));            # ap[0]
 152         &movd   ($car1,&DWP(0,$np));            # np[0]
 153
 154         &pmuludq($mul1,$mul0);                  # ap[0]*bp[0]
 155         &movq   ($car0,$mul1);
 156         &movq   ($acc0,$mul1);                  # I wish movd worked for
 157         &pand   ($acc0,$mask);                  # inter-register transfers
 158
 159         &pmuludq($mul1,$_n0q);                  # *=n0
 160
 161         &pmuludq($car1,$mul1);                  # "t[0]"*np[0]*n0
 162         &paddq  ($car1,$acc0);
 163
 164         &movd   ($acc1,&DWP(4,$np));            # np[1]
 165         &movd   ($acc0,&DWP(4,$ap));            # ap[1]
 166
 167         &psrlq  ($car0,32);
 168         &psrlq  ($car1,32);
 169
 170         &inc    ($j);                           # j++
 171 &set_label("1st",16);
 172         &pmuludq($acc0,$mul0);                  # ap[j]*bp[0]
 173         &pmuludq($acc1,$mul1);                  # np[j]*m1
 174         &paddq  ($car0,$acc0);                  # +=c0
 175         &paddq  ($car1,$acc1);                  # +=c1
 176
 177         &movq   ($acc0,$car0);
 178         &pand   ($acc0,$mask);
 179         &movd   ($acc1,&DWP(4,$np,$j,4));       # np[j+1]
 180         &paddq  ($car1,$acc0);                  # +=ap[j]*bp[0];
 181         &movd   ($acc0,&DWP(4,$ap,$j,4));       # ap[j+1]
 182         &psrlq  ($car0,32);
 183         &movd   (&DWP($frame-4,"esp",$j,4),$car1);      # tp[j-1]=
 184         &psrlq  ($car1,32);
 185
 186         &lea    ($j,&DWP(1,$j));
 187         &cmp    ($j,$num);
 188         &jl     (&label("1st"));
 189
 190         &pmuludq($acc0,$mul0);                  # ap[num-1]*bp[0]
 191         &pmuludq($acc1,$mul1);                  # np[num-1]*m1
 192         &paddq  ($car0,$acc0);                  # +=c0
 193         &paddq  ($car1,$acc1);                  # +=c1
 194
 195         &movq   ($acc0,$car0);
 196         &pand   ($acc0,$mask);
 197         &paddq  ($car1,$acc0);                  # +=ap[num-1]*bp[0];
 198         &movd   (&DWP($frame-4,"esp",$j,4),$car1);      # tp[num-2]=
 199
 200         &psrlq  ($car0,32);
 201         &psrlq  ($car1,32);
 202
 203         &paddq  ($car1,$car0);
 204         &movq   (&QWP($frame,"esp",$num,4),$car1);      # tp[num].tp[num-1]
 205 \f
 206         &inc    ($i);                           # i++
 207 &set_label("outer");
 208         &xor    ($j,$j);                        # j=0
 209
 210         &movd   ($mul0,&DWP(0,$bp,$i,4));       # bp[i]
 211         &movd   ($mul1,&DWP(0,$ap));            # ap[0]
 212         &movd   ($temp,&DWP($frame,"esp"));     # tp[0]
 213         &movd   ($car1,&DWP(0,$np));            # np[0]
 214         &pmuludq($mul1,$mul0);                  # ap[0]*bp[i]
 215
 216         &paddq  ($mul1,$temp);                  # +=tp[0]
 217         &movq   ($acc0,$mul1);
 218         &movq   ($car0,$mul1);
 219         &pand   ($acc0,$mask);
 220
 221         &pmuludq($mul1,$_n0q);                  # *=n0
 222
 223         &pmuludq($car1,$mul1);
 224         &paddq  ($car1,$acc0);
 225
 226         &movd   ($temp,&DWP($frame+4,"esp"));   # tp[1]
 227         &movd   ($acc1,&DWP(4,$np));            # np[1]
 228         &movd   ($acc0,&DWP(4,$ap));            # ap[1]
 229
 230         &psrlq  ($car0,32);
 231         &psrlq  ($car1,32);
 232         &paddq  ($car0,$temp);                  # +=tp[1]
 233
 234         &inc    ($j);                           # j++
 235         &dec    ($num);
 236 &set_label("inner");
 237         &pmuludq($acc0,$mul0);                  # ap[j]*bp[i]
 238         &pmuludq($acc1,$mul1);                  # np[j]*m1
 239         &paddq  ($car0,$acc0);                  # +=c0
 240         &paddq  ($car1,$acc1);                  # +=c1
 241
 242         &movq   ($acc0,$car0);
 243         &movd   ($temp,&DWP($frame+4,"esp",$j,4));# tp[j+1]
 244         &pand   ($acc0,$mask);
 245         &movd   ($acc1,&DWP(4,$np,$j,4));       # np[j+1]
 246         &paddq  ($car1,$acc0);                  # +=ap[j]*bp[i]+tp[j]
 247         &movd   ($acc0,&DWP(4,$ap,$j,4));       # ap[j+1]
 248         &psrlq  ($car0,32);
 249         &movd   (&DWP($frame-4,"esp",$j,4),$car1);# tp[j-1]=
 250         &psrlq  ($car1,32);
 251         &paddq  ($car0,$temp);                  # +=tp[j+1]
 252
 253         &dec    ($num);
 254         &lea    ($j,&DWP(1,$j));                # j++
 255         &jnz    (&label("inner"));
 256
 257         &mov    ($num,$j);
 258         &pmuludq($acc0,$mul0);                  # ap[num-1]*bp[i]
 259         &pmuludq($acc1,$mul1);                  # np[num-1]*m1
 260         &paddq  ($car0,$acc0);                  # +=c0
 261         &paddq  ($car1,$acc1);                  # +=c1
 262
 263         &movq   ($acc0,$car0);
 264         &pand   ($acc0,$mask);
 265         &paddq  ($car1,$acc0);                  # +=ap[num-1]*bp[i]+tp[num-1]
 266         &movd   (&DWP($frame-4,"esp",$j,4),$car1);      # tp[num-2]=
 267         &psrlq  ($car0,32);
 268         &psrlq  ($car1,32);
 269
 270         &movd   ($temp,&DWP($frame+4,"esp",$num,4));    # += tp[num]
 271         &paddq  ($car1,$car0);
 272         &paddq  ($car1,$temp);
 273         &movq   (&QWP($frame,"esp",$num,4),$car1);      # tp[num].tp[num-1]
 274
 275         &lea    ($i,&DWP(1,$i));                # i++
 276         &cmp    ($i,$num);
 277         &jle    (&label("outer"));
 278
 279         &emms   ();                             # done with mmx bank
 280         &jmp    (&label("common_tail"));
 281
 282 &set_label("non_sse2",16);
 283 }
 284 \f
 285 if (0) {
 286         &mov    ("esp",$_sp);
 287         &xor    ("eax","eax");  # signal "not fast enough [yet]"
 288         &jmp    (&label("just_leave"));
 289         # While the below code provides competitive performance for
 290         # all key lengthes on modern Intel cores, it's still more
 291         # than 10% slower for 4096-bit key elsewhere:-( "Competitive"
 292         # means compared to the original integer-only assembler.
 293         # 512-bit RSA sign is better by ~40%, but that's about all
 294         # one can say about all CPUs...
 295 } else {
 296 $inp="esi";     # integer path uses these registers differently
 297 $word="edi";
 298 $carry="ebp";
 299
 300         &mov    ($inp,$_ap);
 301         &lea    ($carry,&DWP(1,$num));
 302         &mov    ($word,$_bp);
 303         &xor    ($j,$j);                                # j=0
 304         &mov    ("edx",$inp);
 305         &and    ($carry,1);                             # see if num is even
 306         &sub    ("edx",$word);                          # see if ap==bp
 307         &lea    ("eax",&DWP(4,$word,$num,4));           # &bp[num]
 308         &or     ($carry,"edx");
 309         &mov    ($word,&DWP(0,$word));                  # bp[0]
 310         &jz     (&label("bn_sqr_mont"));
 311         &mov    ($_bpend,"eax");
 312         &mov    ("eax",&DWP(0,$inp));
 313         &xor    ("edx","edx");
 314
 315 &set_label("mull",16);
 316         &mov    ($carry,"edx");
 317         &mul    ($word);                                # ap[j]*bp[0]
 318         &add    ($carry,"eax");
 319         &lea    ($j,&DWP(1,$j));
 320         &adc    ("edx",0);
 321         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j+1]
 322         &cmp    ($j,$num);
 323         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 324         &jl     (&label("mull"));
 325
 326         &mov    ($carry,"edx");
 327         &mul    ($word);                                # ap[num-1]*bp[0]
 328          &mov   ($word,$_n0);
 329         &add    ("eax",$carry);
 330          &mov   ($inp,$_np);
 331         &adc    ("edx",0);
 332          &imul  ($word,&DWP($frame,"esp"));             # n0*tp[0]
 333
 334         &mov    (&DWP($frame,"esp",$num,4),"eax");      # tp[num-1]=
 335         &xor    ($j,$j);
 336         &mov    (&DWP($frame+4,"esp",$num,4),"edx");    # tp[num]=
 337         &mov    (&DWP($frame+8,"esp",$num,4),$j);       # tp[num+1]=
 338
 339         &mov    ("eax",&DWP(0,$inp));                   # np[0]
 340         &mul    ($word);                                # np[0]*m
 341         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 342         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 343         &adc    ("edx",0);
 344         &inc    ($j);
 345
 346         &jmp    (&label("2ndmadd"));
 347 \f\f
 348 &set_label("1stmadd",16);
 349         &mov    ($carry,"edx");
 350         &mul    ($word);                                # ap[j]*bp[i]
 351         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 352         &lea    ($j,&DWP(1,$j));
 353         &adc    ("edx",0);
 354         &add    ($carry,"eax");
 355         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j+1]
 356         &adc    ("edx",0);
 357         &cmp    ($j,$num);
 358         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 359         &jl     (&label("1stmadd"));
 360
 361         &mov    ($carry,"edx");
 362         &mul    ($word);                                # ap[num-1]*bp[i]
 363         &add    ("eax",&DWP($frame,"esp",$num,4));      # +=tp[num-1]
 364          &mov   ($word,$_n0);
 365         &adc    ("edx",0);
 366          &mov   ($inp,$_np);
 367         &add    ($carry,"eax");
 368         &adc    ("edx",0);
 369          &imul  ($word,&DWP($frame,"esp"));             # n0*tp[0]
 370
 371         &xor    ($j,$j);
 372         &add    ("edx",&DWP($frame+4,"esp",$num,4));    # carry+=tp[num]
 373         &mov    (&DWP($frame,"esp",$num,4),$carry);     # tp[num-1]=
 374         &adc    ($j,0);
 375          &mov   ("eax",&DWP(0,$inp));                   # np[0]
 376         &mov    (&DWP($frame+4,"esp",$num,4),"edx");    # tp[num]=
 377         &mov    (&DWP($frame+8,"esp",$num,4),$j);       # tp[num+1]=
 378
 379         &mul    ($word);                                # np[0]*m
 380         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 381         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 382         &adc    ("edx",0);
 383         &mov    ($j,1);
 384 \f
 385 &set_label("2ndmadd",16);
 386         &mov    ($carry,"edx");
 387         &mul    ($word);                                # np[j]*m
 388         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 389         &lea    ($j,&DWP(1,$j));
 390         &adc    ("edx",0);
 391         &add    ($carry,"eax");
 392         &mov    ("eax",&DWP(0,$inp,$j,4));              # np[j+1]
 393         &adc    ("edx",0);
 394         &cmp    ($j,$num);
 395         &mov    (&DWP($frame-8,"esp",$j,4),$carry);     # tp[j-1]=
 396         &jl     (&label("2ndmadd"));
 397
 398         &mov    ($carry,"edx");
 399         &mul    ($word);                                # np[j]*m
 400         &add    ($carry,&DWP($frame,"esp",$num,4));     # +=tp[num-1]
 401         &adc    ("edx",0);
 402         &add    ($carry,"eax");
 403         &adc    ("edx",0);
 404         &mov    (&DWP($frame-4,"esp",$num,4),$carry);   # tp[num-2]=
 405
 406         &xor    ("eax","eax");
 407          &mov   ($j,$_bp);                              # &bp[i]
 408         &add    ("edx",&DWP($frame+4,"esp",$num,4));    # carry+=tp[num]
 409         &adc    ("eax",&DWP($frame+8,"esp",$num,4));    # +=tp[num+1]
 410          &lea   ($j,&DWP(4,$j));
 411         &mov    (&DWP($frame,"esp",$num,4),"edx");      # tp[num-1]=
 412          &cmp   ($j,$_bpend);
 413         &mov    (&DWP($frame+4,"esp",$num,4),"eax");    # tp[num]=
 414         &je     (&label("common_tail"));
 415
 416         &mov    ($word,&DWP(0,$j));                     # bp[i+1]
 417         &mov    ($inp,$_ap);
 418         &mov    ($_bp,$j);                              # &bp[++i]
 419         &xor    ($j,$j);
 420         &xor    ("edx","edx");
 421         &mov    ("eax",&DWP(0,$inp));
 422         &jmp    (&label("1stmadd"));
 423 \f
 424 &set_label("bn_sqr_mont",16);
 425 $sbit=$num;
 426         &mov    ($_num,$num);
 427         &mov    ($_bp,$j);                              # i=0
 428
 429         &mov    ("eax",$word);                          # ap[0]
 430         &mul    ($word);                                # ap[0]*ap[0]
 431         &mov    (&DWP($frame,"esp"),"eax");             # tp[0]=
 432         &mov    ($sbit,"edx");
 433         &shr    ("edx",1);
 434         &and    ($sbit,1);
 435         &inc    ($j);
 436 &set_label("sqr",16);
 437         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j]
 438         &mov    ($carry,"edx");
 439         &mul    ($word);                                # ap[j]*ap[0]
 440         &add    ("eax",$carry);
 441         &lea    ($j,&DWP(1,$j));
 442         &adc    ("edx",0);
 443         &lea    ($carry,&DWP(0,$sbit,"eax",2));
 444         &shr    ("eax",31);
 445         &cmp    ($j,$_num);
 446         &mov    ($sbit,"eax");
 447         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 448         &jl     (&label("sqr"));
 449
 450         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[num-1]
 451         &mov    ($carry,"edx");
 452         &mul    ($word);                                # ap[num-1]*ap[0]
 453         &add    ("eax",$carry);
 454          &mov   ($word,$_n0);
 455         &adc    ("edx",0);
 456          &mov   ($inp,$_np);
 457         &lea    ($carry,&DWP(0,$sbit,"eax",2));
 458          &imul  ($word,&DWP($frame,"esp"));             # n0*tp[0]
 459         &shr    ("eax",31);
 460         &mov    (&DWP($frame,"esp",$j,4),$carry);       # tp[num-1]=
 461
 462         &lea    ($carry,&DWP(0,"eax","edx",2));
 463          &mov   ("eax",&DWP(0,$inp));                   # np[0]
 464         &shr    ("edx",31);
 465         &mov    (&DWP($frame+4,"esp",$j,4),$carry);     # tp[num]=
 466         &mov    (&DWP($frame+8,"esp",$j,4),"edx");      # tp[num+1]=
 467
 468         &mul    ($word);                                # np[0]*m
 469         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 470         &mov    ($num,$j);
 471         &adc    ("edx",0);
 472         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 473         &mov    ($j,1);
 474 \f\f
 475 &set_label("3rdmadd",16);
 476         &mov    ($carry,"edx");
 477         &mul    ($word);                                # np[j]*m
 478         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 479         &adc    ("edx",0);
 480         &add    ($carry,"eax");
 481         &mov    ("eax",&DWP(4,$inp,$j,4));              # np[j+1]
 482         &adc    ("edx",0);
 483         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j-1]=
 484
 485         &mov    ($carry,"edx");
 486         &mul    ($word);                                # np[j+1]*m
 487         &add    ($carry,&DWP($frame+4,"esp",$j,4));     # +=tp[j+1]
 488         &lea    ($j,&DWP(2,$j));
 489         &adc    ("edx",0);
 490         &add    ($carry,"eax");
 491         &mov    ("eax",&DWP(0,$inp,$j,4));              # np[j+2]
 492         &adc    ("edx",0);
 493         &cmp    ($j,$num);
 494         &mov    (&DWP($frame-8,"esp",$j,4),$carry);     # tp[j]=
 495         &jl     (&label("3rdmadd"));
 496
 497         &mov    ($carry,"edx");
 498         &mul    ($word);                                # np[j]*m
 499         &add    ($carry,&DWP($frame,"esp",$num,4));     # +=tp[num-1]
 500         &adc    ("edx",0);
 501         &add    ($carry,"eax");
 502         &adc    ("edx",0);
 503         &mov    (&DWP($frame-4,"esp",$num,4),$carry);   # tp[num-2]=
 504
 505         &mov    ($j,$_bp);                              # i
 506         &xor    ("eax","eax");
 507         &mov    ($inp,$_ap);
 508         &add    ("edx",&DWP($frame+4,"esp",$num,4));    # carry+=tp[num]
 509         &adc    ("eax",&DWP($frame+8,"esp",$num,4));    # +=tp[num+1]
 510         &mov    (&DWP($frame,"esp",$num,4),"edx");      # tp[num-1]=
 511         &cmp    ($j,$num);
 512         &mov    (&DWP($frame+4,"esp",$num,4),"eax");    # tp[num]=
 513         &je     (&label("common_tail"));
 514 \f
 515         &mov    ($word,&DWP(4,$inp,$j,4));              # ap[i]
 516         &lea    ($j,&DWP(1,$j));
 517         &mov    ("eax",$word);
 518         &mov    ($_bp,$j);                              # ++i
 519         &mul    ($word);                                # ap[i]*ap[i]
 520         &add    ("eax",&DWP($frame,"esp",$j,4));        # +=tp[i]
 521         &adc    ("edx",0);
 522         &mov    (&DWP($frame,"esp",$j,4),"eax");        # tp[i]=
 523         &xor    ($carry,$carry);
 524         &cmp    ($j,$num);
 525         &lea    ($j,&DWP(1,$j));
 526         &je     (&label("sqrlast"));
 527
 528         &mov    ($sbit,"edx");                          # zaps $num
 529         &shr    ("edx",1);
 530         &and    ($sbit,1);
 531 &set_label("sqradd",16);
 532         &mov    ("eax",&DWP(0,$inp,$j,4));              # ap[j]
 533         &mov    ($carry,"edx");
 534         &mul    ($word);                                # ap[j]*ap[i]
 535         &add    ("eax",$carry);
 536         &lea    ($carry,&DWP(0,"eax","eax"));
 537         &adc    ("edx",0);
 538         &shr    ("eax",31);
 539         &add    ($carry,&DWP($frame,"esp",$j,4));       # +=tp[j]
 540         &lea    ($j,&DWP(1,$j));
 541         &adc    ("eax",0);
 542         &add    ($carry,$sbit);
 543         &adc    ("eax",0);
 544         &cmp    ($j,$_num);
 545         &mov    (&DWP($frame-4,"esp",$j,4),$carry);     # tp[j]=
 546         &mov    ($sbit,"eax");
 547         &jle    (&label("sqradd"));
 548
 549         &mov    ($carry,"edx");
 550         &add    ("edx","edx");
 551         &shr    ($carry,31);
 552         &add    ("edx",$sbit);
 553         &adc    ($carry,0);
 554 &set_label("sqrlast");
 555         &mov    ($word,$_n0);
 556         &mov    ($inp,$_np);
 557         &imul   ($word,&DWP($frame,"esp"));             # n0*tp[0]
 558
 559         &add    ("edx",&DWP($frame,"esp",$j,4));        # +=tp[num]
 560         &mov    ("eax",&DWP(0,$inp));                   # np[0]
 561         &adc    ($carry,0);
 562         &mov    (&DWP($frame,"esp",$j,4),"edx");        # tp[num]=
 563         &mov    (&DWP($frame+4,"esp",$j,4),$carry);     # tp[num+1]=
 564
 565         &mul    ($word);                                # np[0]*m
 566         &add    ("eax",&DWP($frame,"esp"));             # +=tp[0]
 567         &lea    ($num,&DWP(-1,$j));
 568         &adc    ("edx",0);
 569         &mov    ($j,1);
 570         &mov    ("eax",&DWP(4,$inp));                   # np[1]
 571
 572         &jmp    (&label("3rdmadd"));
 573 }
 574 \f
 575 &set_label("common_tail",16);
 576         &mov    ($np,$_np);                     # load modulus pointer
 577         &mov    ($rp,$_rp);                     # load result pointer
 578         &lea    ($tp,&DWP($frame,"esp"));       # [$ap and $bp are zapped]
 579
 580         &mov    ("eax",&DWP(0,$tp));            # tp[0]
 581         &mov    ($j,$num);                      # j=num-1
 582         &xor    ($i,$i);                        # i=0 and clear CF!
 583
 584 &set_label("sub",16);
 585         &sbb    ("eax",&DWP(0,$np,$i,4));
 586         &mov    (&DWP(0,$rp,$i,4),"eax");       # rp[i]=tp[i]-np[i]
 587         &dec    ($j);                           # doesn't affect CF!
 588         &mov    ("eax",&DWP(4,$tp,$i,4));       # tp[i+1]
 589         &lea    ($i,&DWP(1,$i));                # i++
 590         &jge    (&label("sub"));
 591
 592         &sbb    ("eax",0);                      # handle upmost overflow bit
 593         &and    ($tp,"eax");
 594         &not    ("eax");
 595         &mov    ($np,$rp);
 596         &and    ($np,"eax");
 597         &or     ($tp,$np);                      # tp=carry?tp:rp
 598
 599 &set_label("copy",16);                          # copy or in-place refresh
 600         &mov    ("eax",&DWP(0,$tp,$num,4));
 601         &mov    (&DWP(0,$rp,$num,4),"eax");     # rp[i]=tp[i]
 602         &mov    (&DWP($frame,"esp",$num,4),$j); # zap temporary vector
 603         &dec    ($num);
 604         &jge    (&label("copy"));
 605
 606         &mov    ("esp",$_sp);           # pull saved stack pointer
 607         &mov    ("eax",1);
 608 &set_label("just_leave");
 609 &function_end("bn_mul_mont");
 610
 611 &asciz("Montgomery Multiplication for x86, CRYPTOGAMS by <appro\@openssl.org>");
 612
 613 &asm_finish();
 614
 615 close STDOUT;