crypto/ec/asm/ecp_nistz256-x86.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2015-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16 #
  17 # ECP_NISTZ256 module for x86/SSE2.
  18 #
  19 # October 2014.
  20 #
  21 # Original ECP_NISTZ256 submission targeting x86_64 is detailed in
  22 # http://eprint.iacr.org/2013/816. In the process of adaptation
  23 # original .c module was made 32-bit savvy in order to make this
  24 # implementation possible.
  25 #
  26 #               with/without -DECP_NISTZ256_ASM
  27 # Pentium       +66-163%
  28 # PIII          +72-172%
  29 # P4            +65-132%
  30 # Core2         +90-215%
  31 # Sandy Bridge  +105-265% (contemporary i[57]-* are all close to this)
  32 # Atom          +65-155%
  33 # Opteron       +54-110%
  34 # Bulldozer     +99-240%
  35 # VIA Nano      +93-290%
  36 #
  37 # Ranges denote minimum and maximum improvement coefficients depending
  38 # on benchmark. Lower coefficients are for ECDSA sign, server-side
  39 # operation. Keep in mind that +200% means 3x improvement.
  40
  41 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  42 push(@INC,"${dir}","${dir}../../perlasm");
  43 require "x86asm.pl";
  44
  45 $output=pop;
  46 open STDOUT,">$output";
  47
  48 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
  49
  50 $sse2=0;
  51 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
  52
  53 &external_label("OPENSSL_ia32cap_P") if ($sse2);
  54
  55
  56 ########################################################################
  57 # Convert ecp_nistz256_table.c to layout expected by ecp_nistz_gather_w7
  58 #
  59 open TABLE,"<ecp_nistz256_table.c"              or
  60 open TABLE,"<${dir}../ecp_nistz256_table.c"     or
  61 die "failed to open ecp_nistz256_table.c:",$!;
  62
  63 use integer;
  64
  65 foreach(<TABLE>) {
  66         s/TOBN\(\s*(0x[0-9a-f]+),\s*(0x[0-9a-f]+)\s*\)/push @arr,hex($2),hex($1)/geo;
  67 }
  68 close TABLE;
  69
  70 # See ecp_nistz256_table.c for explanation for why it's 64*16*37.
  71 # 64*16*37-1 is because $#arr returns last valid index or @arr, not
  72 # amount of elements.
  73 die "insane number of elements" if ($#arr != 64*16*37-1);
  74
  75 &public_label("ecp_nistz256_precomputed");
  76 &align(4096);
  77 &set_label("ecp_nistz256_precomputed");
  78
  79 ########################################################################
  80 # this conversion smashes P256_POINT_AFFINE by individual bytes with
  81 # 64 byte interval, similar to
  82 #       1111222233334444
  83 #       1234123412341234
  84 for(1..37) {
  85         @tbl = splice(@arr,0,64*16);
  86         for($i=0;$i<64;$i++) {
  87                 undef @line;
  88                 for($j=0;$j<64;$j++) {
  89                         push @line,(@tbl[$j*16+$i/4]>>(($i%4)*8))&0xff;
  90                 }
  91                 &data_byte(join(',',map { sprintf "0x%02x",$_} @line));
  92         }
  93 }
  94
  95 ########################################################################
  96 # Keep in mind that constants are stored least to most significant word
  97 &static_label("RR");
  98 &set_label("RR",64);
  99 &data_word(3,0,-1,-5,-2,-1,-3,4);       # 2^512 mod P-256
 100
 101 &static_label("ONE_mont");
 102 &set_label("ONE_mont");
 103 &data_word(1,0,0,-1,-1,-1,-2,0);
 104
 105 &static_label("ONE");
 106 &set_label("ONE");
 107 &data_word(1,0,0,0,0,0,0,0);
 108 &asciz("ECP_NISZ256 for x86/SSE2, CRYPTOGAMS by <appro\@openssl.org>");
 109 &align(64);
 110
 111 ########################################################################
 112 # void ecp_nistz256_mul_by_2(BN_ULONG edi[8],const BN_ULONG esi[8]);
 113 &function_begin("ecp_nistz256_mul_by_2");
 114         &mov    ("esi",&wparam(1));
 115         &mov    ("edi",&wparam(0));
 116         &mov    ("ebp","esi");
 117 ########################################################################
 118 # common pattern for internal functions is that %edi is result pointer,
 119 # %esi and %ebp are input ones, %ebp being optional. %edi is preserved.
 120         &call   ("_ecp_nistz256_add");
 121 &function_end("ecp_nistz256_mul_by_2");
 122
 123 ########################################################################
 124 # void ecp_nistz256_mul_by_3(BN_ULONG edi[8],const BN_ULONG esi[8]);
 125 &function_begin("ecp_nistz256_mul_by_3");
 126         &mov    ("esi",&wparam(1));
 127                                         # multiplication by 3 is performed
 128                                         # as 2*n+n, but we can't use output
 129                                         # to store 2*n, because if output
 130                                         # pointer equals to input, then
 131                                         # we'll get 2*n+2*n.
 132         &stack_push(8);                 # therefore we need to allocate
 133                                         # 256-bit intermediate buffer.
 134         &mov    ("edi","esp");
 135         &mov    ("ebp","esi");
 136         &call   ("_ecp_nistz256_add");
 137         &lea    ("esi",&DWP(0,"edi"));
 138         &mov    ("ebp",&wparam(1));
 139         &mov    ("edi",&wparam(0));
 140         &call   ("_ecp_nistz256_add");
 141         &stack_pop(8);
 142 &function_end("ecp_nistz256_mul_by_3");
 143
 144 ########################################################################
 145 # void ecp_nistz256_div_by_2(BN_ULONG edi[8],const BN_ULONG esi[8]);
 146 &function_begin("ecp_nistz256_div_by_2");
 147         &mov    ("esi",&wparam(1));
 148         &mov    ("edi",&wparam(0));
 149         &call   ("_ecp_nistz256_div_by_2");
 150 &function_end("ecp_nistz256_div_by_2");
 151
 152 &function_begin_B("_ecp_nistz256_div_by_2");
 153         # tmp = a is odd ? a+mod : a
 154         #
 155         # note that because mod has special form, i.e. consists of
 156         # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
 157         # assigning least significant bit of input to one register,
 158         # %ebp, and its negative to another, %edx.
 159
 160         &mov    ("ebp",&DWP(0,"esi"));
 161         &xor    ("edx","edx");
 162         &mov    ("ebx",&DWP(4,"esi"));
 163         &mov    ("eax","ebp");
 164         &and    ("ebp",1);
 165         &mov    ("ecx",&DWP(8,"esi"));
 166         &sub    ("edx","ebp");
 167
 168         &add    ("eax","edx");
 169         &adc    ("ebx","edx");
 170         &mov    (&DWP(0,"edi"),"eax");
 171         &adc    ("ecx","edx");
 172         &mov    (&DWP(4,"edi"),"ebx");
 173         &mov    (&DWP(8,"edi"),"ecx");
 174
 175         &mov    ("eax",&DWP(12,"esi"));
 176         &mov    ("ebx",&DWP(16,"esi"));
 177         &adc    ("eax",0);
 178         &mov    ("ecx",&DWP(20,"esi"));
 179         &adc    ("ebx",0);
 180         &mov    (&DWP(12,"edi"),"eax");
 181         &adc    ("ecx",0);
 182         &mov    (&DWP(16,"edi"),"ebx");
 183         &mov    (&DWP(20,"edi"),"ecx");
 184
 185         &mov    ("eax",&DWP(24,"esi"));
 186         &mov    ("ebx",&DWP(28,"esi"));
 187         &adc    ("eax","ebp");
 188         &adc    ("ebx","edx");
 189         &mov    (&DWP(24,"edi"),"eax");
 190         &sbb    ("esi","esi");                  # broadcast carry bit
 191         &mov    (&DWP(28,"edi"),"ebx");
 192
 193         # ret = tmp >> 1
 194
 195         &mov    ("eax",&DWP(0,"edi"));
 196         &mov    ("ebx",&DWP(4,"edi"));
 197         &mov    ("ecx",&DWP(8,"edi"));
 198         &mov    ("edx",&DWP(12,"edi"));
 199
 200         &shr    ("eax",1);
 201         &mov    ("ebp","ebx");
 202         &shl    ("ebx",31);
 203         &or     ("eax","ebx");
 204
 205         &shr    ("ebp",1);
 206         &mov    ("ebx","ecx");
 207         &shl    ("ecx",31);
 208         &mov    (&DWP(0,"edi"),"eax");
 209         &or     ("ebp","ecx");
 210         &mov    ("eax",&DWP(16,"edi"));
 211
 212         &shr    ("ebx",1);
 213         &mov    ("ecx","edx");
 214         &shl    ("edx",31);
 215         &mov    (&DWP(4,"edi"),"ebp");
 216         &or     ("ebx","edx");
 217         &mov    ("ebp",&DWP(20,"edi"));
 218
 219         &shr    ("ecx",1);
 220         &mov    ("edx","eax");
 221         &shl    ("eax",31);
 222         &mov    (&DWP(8,"edi"),"ebx");
 223         &or     ("ecx","eax");
 224         &mov    ("ebx",&DWP(24,"edi"));
 225
 226         &shr    ("edx",1);
 227         &mov    ("eax","ebp");
 228         &shl    ("ebp",31);
 229         &mov    (&DWP(12,"edi"),"ecx");
 230         &or     ("edx","ebp");
 231         &mov    ("ecx",&DWP(28,"edi"));
 232
 233         &shr    ("eax",1);
 234         &mov    ("ebp","ebx");
 235         &shl    ("ebx",31);
 236         &mov    (&DWP(16,"edi"),"edx");
 237         &or     ("eax","ebx");
 238
 239         &shr    ("ebp",1);
 240         &mov    ("ebx","ecx");
 241         &shl    ("ecx",31);
 242         &mov    (&DWP(20,"edi"),"eax");
 243         &or     ("ebp","ecx");
 244
 245         &shr    ("ebx",1);
 246         &shl    ("esi",31);
 247         &mov    (&DWP(24,"edi"),"ebp");
 248         &or     ("ebx","esi");                  # handle top-most carry bit
 249         &mov    (&DWP(28,"edi"),"ebx");
 250
 251         &ret    ();
 252 &function_end_B("_ecp_nistz256_div_by_2");
 253
 254 ########################################################################
 255 # void ecp_nistz256_add(BN_ULONG edi[8],const BN_ULONG esi[8],
 256 #                                       const BN_ULONG ebp[8]);
 257 &function_begin("ecp_nistz256_add");
 258         &mov    ("esi",&wparam(1));
 259         &mov    ("ebp",&wparam(2));
 260         &mov    ("edi",&wparam(0));
 261         &call   ("_ecp_nistz256_add");
 262 &function_end("ecp_nistz256_add");
 263
 264 &function_begin_B("_ecp_nistz256_add");
 265         &mov    ("eax",&DWP(0,"esi"));
 266         &mov    ("ebx",&DWP(4,"esi"));
 267         &mov    ("ecx",&DWP(8,"esi"));
 268         &add    ("eax",&DWP(0,"ebp"));
 269         &mov    ("edx",&DWP(12,"esi"));
 270         &adc    ("ebx",&DWP(4,"ebp"));
 271         &mov    (&DWP(0,"edi"),"eax");
 272         &adc    ("ecx",&DWP(8,"ebp"));
 273         &mov    (&DWP(4,"edi"),"ebx");
 274         &adc    ("edx",&DWP(12,"ebp"));
 275         &mov    (&DWP(8,"edi"),"ecx");
 276         &mov    (&DWP(12,"edi"),"edx");
 277
 278         &mov    ("eax",&DWP(16,"esi"));
 279         &mov    ("ebx",&DWP(20,"esi"));
 280         &mov    ("ecx",&DWP(24,"esi"));
 281         &adc    ("eax",&DWP(16,"ebp"));
 282         &mov    ("edx",&DWP(28,"esi"));
 283         &adc    ("ebx",&DWP(20,"ebp"));
 284         &mov    (&DWP(16,"edi"),"eax");
 285         &adc    ("ecx",&DWP(24,"ebp"));
 286         &mov    (&DWP(20,"edi"),"ebx");
 287         &mov    ("esi",0);
 288         &adc    ("edx",&DWP(28,"ebp"));
 289         &mov    (&DWP(24,"edi"),"ecx");
 290         &adc    ("esi",0);
 291         &mov    (&DWP(28,"edi"),"edx");
 292
 293         # if a+b >= modulus, subtract modulus.
 294         #
 295         # But since comparison implies subtraction, we subtract modulus
 296         # to see if it borrows, and then subtract it for real if
 297         # subtraction didn't borrow.
 298
 299         &mov    ("eax",&DWP(0,"edi"));
 300         &mov    ("ebx",&DWP(4,"edi"));
 301         &mov    ("ecx",&DWP(8,"edi"));
 302         &sub    ("eax",-1);
 303         &mov    ("edx",&DWP(12,"edi"));
 304         &sbb    ("ebx",-1);
 305         &mov    ("eax",&DWP(16,"edi"));
 306         &sbb    ("ecx",-1);
 307         &mov    ("ebx",&DWP(20,"edi"));
 308         &sbb    ("edx",0);
 309         &mov    ("ecx",&DWP(24,"edi"));
 310         &sbb    ("eax",0);
 311         &mov    ("edx",&DWP(28,"edi"));
 312         &sbb    ("ebx",0);
 313         &sbb    ("ecx",1);
 314         &sbb    ("edx",-1);
 315         &sbb    ("esi",0);
 316
 317         # Note that because mod has special form, i.e. consists of
 318         # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
 319         # by using borrow.
 320
 321         &not    ("esi");
 322         &mov    ("eax",&DWP(0,"edi"));
 323         &mov    ("ebp","esi");
 324         &mov    ("ebx",&DWP(4,"edi"));
 325         &shr    ("ebp",31);
 326         &mov    ("ecx",&DWP(8,"edi"));
 327         &sub    ("eax","esi");
 328         &mov    ("edx",&DWP(12,"edi"));
 329         &sbb    ("ebx","esi");
 330         &mov    (&DWP(0,"edi"),"eax");
 331         &sbb    ("ecx","esi");
 332         &mov    (&DWP(4,"edi"),"ebx");
 333         &sbb    ("edx",0);
 334         &mov    (&DWP(8,"edi"),"ecx");
 335         &mov    (&DWP(12,"edi"),"edx");
 336
 337         &mov    ("eax",&DWP(16,"edi"));
 338         &mov    ("ebx",&DWP(20,"edi"));
 339         &mov    ("ecx",&DWP(24,"edi"));
 340         &sbb    ("eax",0);
 341         &mov    ("edx",&DWP(28,"edi"));
 342         &sbb    ("ebx",0);
 343         &mov    (&DWP(16,"edi"),"eax");
 344         &sbb    ("ecx","ebp");
 345         &mov    (&DWP(20,"edi"),"ebx");
 346         &sbb    ("edx","esi");
 347         &mov    (&DWP(24,"edi"),"ecx");
 348         &mov    (&DWP(28,"edi"),"edx");
 349
 350         &ret    ();
 351 &function_end_B("_ecp_nistz256_add");
 352
 353 ########################################################################
 354 # void ecp_nistz256_sub(BN_ULONG edi[8],const BN_ULONG esi[8],
 355 #                                       const BN_ULONG ebp[8]);
 356 &function_begin("ecp_nistz256_sub");
 357         &mov    ("esi",&wparam(1));
 358         &mov    ("ebp",&wparam(2));
 359         &mov    ("edi",&wparam(0));
 360         &call   ("_ecp_nistz256_sub");
 361 &function_end("ecp_nistz256_sub");
 362
 363 &function_begin_B("_ecp_nistz256_sub");
 364         &mov    ("eax",&DWP(0,"esi"));
 365         &mov    ("ebx",&DWP(4,"esi"));
 366         &mov    ("ecx",&DWP(8,"esi"));
 367         &sub    ("eax",&DWP(0,"ebp"));
 368         &mov    ("edx",&DWP(12,"esi"));
 369         &sbb    ("ebx",&DWP(4,"ebp"));
 370         &mov    (&DWP(0,"edi"),"eax");
 371         &sbb    ("ecx",&DWP(8,"ebp"));
 372         &mov    (&DWP(4,"edi"),"ebx");
 373         &sbb    ("edx",&DWP(12,"ebp"));
 374         &mov    (&DWP(8,"edi"),"ecx");
 375         &mov    (&DWP(12,"edi"),"edx");
 376
 377         &mov    ("eax",&DWP(16,"esi"));
 378         &mov    ("ebx",&DWP(20,"esi"));
 379         &mov    ("ecx",&DWP(24,"esi"));
 380         &sbb    ("eax",&DWP(16,"ebp"));
 381         &mov    ("edx",&DWP(28,"esi"));
 382         &sbb    ("ebx",&DWP(20,"ebp"));
 383         &sbb    ("ecx",&DWP(24,"ebp"));
 384         &mov    (&DWP(16,"edi"),"eax");
 385         &sbb    ("edx",&DWP(28,"ebp"));
 386         &mov    (&DWP(20,"edi"),"ebx");
 387         &sbb    ("esi","esi");                  # broadcast borrow bit
 388         &mov    (&DWP(24,"edi"),"ecx");
 389         &mov    (&DWP(28,"edi"),"edx");
 390
 391         # if a-b borrows, add modulus.
 392         #
 393         # Note that because mod has special form, i.e. consists of
 394         # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
 395         # assigning borrow bit to one register, %ebp, and its negative
 396         # to another, %esi. But we started by calculating %esi...
 397
 398         &mov    ("eax",&DWP(0,"edi"));
 399         &mov    ("ebp","esi");
 400         &mov    ("ebx",&DWP(4,"edi"));
 401         &shr    ("ebp",31);
 402         &mov    ("ecx",&DWP(8,"edi"));
 403         &add    ("eax","esi");
 404         &mov    ("edx",&DWP(12,"edi"));
 405         &adc    ("ebx","esi");
 406         &mov    (&DWP(0,"edi"),"eax");
 407         &adc    ("ecx","esi");
 408         &mov    (&DWP(4,"edi"),"ebx");
 409         &adc    ("edx",0);
 410         &mov    (&DWP(8,"edi"),"ecx");
 411         &mov    (&DWP(12,"edi"),"edx");
 412
 413         &mov    ("eax",&DWP(16,"edi"));
 414         &mov    ("ebx",&DWP(20,"edi"));
 415         &mov    ("ecx",&DWP(24,"edi"));
 416         &adc    ("eax",0);
 417         &mov    ("edx",&DWP(28,"edi"));
 418         &adc    ("ebx",0);
 419         &mov    (&DWP(16,"edi"),"eax");
 420         &adc    ("ecx","ebp");
 421         &mov    (&DWP(20,"edi"),"ebx");
 422         &adc    ("edx","esi");
 423         &mov    (&DWP(24,"edi"),"ecx");
 424         &mov    (&DWP(28,"edi"),"edx");
 425
 426         &ret    ();
 427 &function_end_B("_ecp_nistz256_sub");
 428
 429 ########################################################################
 430 # void ecp_nistz256_neg(BN_ULONG edi[8],const BN_ULONG esi[8]);
 431 &function_begin("ecp_nistz256_neg");
 432         &mov    ("ebp",&wparam(1));
 433         &mov    ("edi",&wparam(0));
 434
 435         &xor    ("eax","eax");
 436         &stack_push(8);
 437         &mov    (&DWP(0,"esp"),"eax");
 438         &mov    ("esi","esp");
 439         &mov    (&DWP(4,"esp"),"eax");
 440         &mov    (&DWP(8,"esp"),"eax");
 441         &mov    (&DWP(12,"esp"),"eax");
 442         &mov    (&DWP(16,"esp"),"eax");
 443         &mov    (&DWP(20,"esp"),"eax");
 444         &mov    (&DWP(24,"esp"),"eax");
 445         &mov    (&DWP(28,"esp"),"eax");
 446
 447         &call   ("_ecp_nistz256_sub");
 448
 449         &stack_pop(8);
 450 &function_end("ecp_nistz256_neg");
 451
 452 &function_begin_B("_picup_eax");
 453         &mov    ("eax",&DWP(0,"esp"));
 454         &ret    ();
 455 &function_end_B("_picup_eax");
 456
 457 ########################################################################
 458 # void ecp_nistz256_to_mont(BN_ULONG edi[8],const BN_ULONG esi[8]);
 459 &function_begin("ecp_nistz256_to_mont");
 460         &mov    ("esi",&wparam(1));
 461         &call   ("_picup_eax");
 462     &set_label("pic");
 463         &lea    ("ebp",&DWP(&label("RR")."-".&label("pic"),"eax"));
 464                                                 if ($sse2) {
 465         &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
 466         &mov    ("eax",&DWP(0,"eax"));          }
 467         &mov    ("edi",&wparam(0));
 468         &call   ("_ecp_nistz256_mul_mont");
 469 &function_end("ecp_nistz256_to_mont");
 470
 471 ########################################################################
 472 # void ecp_nistz256_from_mont(BN_ULONG edi[8],const BN_ULONG esi[8]);
 473 &function_begin("ecp_nistz256_from_mont");
 474         &mov    ("esi",&wparam(1));
 475         &call   ("_picup_eax");
 476     &set_label("pic");
 477         &lea    ("ebp",&DWP(&label("ONE")."-".&label("pic"),"eax"));
 478                                                 if ($sse2) {
 479         &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
 480         &mov    ("eax",&DWP(0,"eax"));          }
 481         &mov    ("edi",&wparam(0));
 482         &call   ("_ecp_nistz256_mul_mont");
 483 &function_end("ecp_nistz256_from_mont");
 484
 485 ########################################################################
 486 # void ecp_nistz256_mul_mont(BN_ULONG edi[8],const BN_ULONG esi[8],
 487 #                                            const BN_ULONG ebp[8]);
 488 &function_begin("ecp_nistz256_mul_mont");
 489         &mov    ("esi",&wparam(1));
 490         &mov    ("ebp",&wparam(2));
 491                                                 if ($sse2) {
 492         &call   ("_picup_eax");
 493     &set_label("pic");
 494         &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
 495         &mov    ("eax",&DWP(0,"eax"));          }
 496         &mov    ("edi",&wparam(0));
 497         &call   ("_ecp_nistz256_mul_mont");
 498 &function_end("ecp_nistz256_mul_mont");
 499
 500 ########################################################################
 501 # void ecp_nistz256_sqr_mont(BN_ULONG edi[8],const BN_ULONG esi[8]);
 502 &function_begin("ecp_nistz256_sqr_mont");
 503         &mov    ("esi",&wparam(1));
 504                                                 if ($sse2) {
 505         &call   ("_picup_eax");
 506     &set_label("pic");
 507         &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
 508         &mov    ("eax",&DWP(0,"eax"));          }
 509         &mov    ("edi",&wparam(0));
 510         &mov    ("ebp","esi");
 511         &call   ("_ecp_nistz256_mul_mont");
 512 &function_end("ecp_nistz256_sqr_mont");
 513
 514 &function_begin_B("_ecp_nistz256_mul_mont");
 515                                                 if ($sse2) {
 516         &and    ("eax",1<<24|1<<26);
 517         &cmp    ("eax",1<<24|1<<26);            # see if XMM+SSE2 is on
 518         &jne    (&label("mul_mont_ialu"));
 519
 520         ########################################
 521         # SSE2 code path featuring 32x16-bit
 522         # multiplications is ~2x faster than
 523         # IALU counterpart (except on Atom)...
 524         ########################################
 525         # stack layout:
 526         # +------------------------------------+< %esp
 527         # | 7 16-byte temporary XMM words,     |
 528         # | "sliding" toward lower address     |
 529         # .                                    .
 530         # +------------------------------------+
 531         # | unused XMM word                    |
 532         # +------------------------------------+< +128,%ebx
 533         # | 8 16-byte XMM words holding copies |
 534         # | of a[i]<<64|a[i]                   |
 535         # .                                    .
 536         # .                                    .
 537         # +------------------------------------+< +256
 538         &mov    ("edx","esp");
 539         &sub    ("esp",0x100);
 540
 541         &movd   ("xmm7",&DWP(0,"ebp"));         # b[0] -> 0000.00xy
 542         &lea    ("ebp",&DWP(4,"ebp"));
 543         &pcmpeqd("xmm6","xmm6");
 544         &psrlq  ("xmm6",48);                    # compose 0xffff<<64|0xffff
 545
 546         &pshuflw("xmm7","xmm7",0b11011100);     # 0000.00xy -> 0000.0x0y
 547         &and    ("esp",-64);
 548         &pshufd ("xmm7","xmm7",0b11011100);     # 0000.0x0y -> 000x.000y
 549         &lea    ("ebx",&DWP(0x80,"esp"));
 550
 551         &movd   ("xmm0",&DWP(4*0,"esi"));       # a[0] -> 0000.00xy
 552         &pshufd ("xmm0","xmm0",0b11001100);     # 0000.00xy -> 00xy.00xy
 553         &movd   ("xmm1",&DWP(4*1,"esi"));       # a[1] -> ...
 554         &movdqa (&QWP(0x00,"ebx"),"xmm0");      # offload converted a[0]
 555         &pmuludq("xmm0","xmm7");                # a[0]*b[0]
 556
 557         &movd   ("xmm2",&DWP(4*2,"esi"));
 558         &pshufd ("xmm1","xmm1",0b11001100);
 559         &movdqa (&QWP(0x10,"ebx"),"xmm1");
 560         &pmuludq("xmm1","xmm7");                # a[1]*b[0]
 561
 562          &movq  ("xmm4","xmm0");                # clear upper 64 bits
 563          &pslldq("xmm4",6);
 564          &paddq ("xmm4","xmm0");
 565          &movdqa("xmm5","xmm4");
 566          &psrldq("xmm4",10);                    # upper 32 bits of a[0]*b[0]
 567          &pand  ("xmm5","xmm6");                # lower 32 bits of a[0]*b[0]
 568
 569         # Upper half of a[0]*b[i] is carried into next multiplication
 570         # iteration, while lower one "participates" in actual reduction.
 571         # Normally latter is done by accumulating result of multiplication
 572         # of modulus by "magic" digit, but thanks to special form of modulus
 573         # and "magic" digit it can be performed only with additions and
 574         # subtractions (see note in IALU section below). Note that we are
 575         # not bothered with carry bits, they are accumulated in "flatten"
 576         # phase after all multiplications and reductions.
 577
 578         &movd   ("xmm3",&DWP(4*3,"esi"));
 579         &pshufd ("xmm2","xmm2",0b11001100);
 580         &movdqa (&QWP(0x20,"ebx"),"xmm2");
 581         &pmuludq("xmm2","xmm7");                # a[2]*b[0]
 582          &paddq ("xmm1","xmm4");                # a[1]*b[0]+hw(a[0]*b[0]), carry
 583         &movdqa (&QWP(0x00,"esp"),"xmm1");      # t[0]
 584
 585         &movd   ("xmm0",&DWP(4*4,"esi"));
 586         &pshufd ("xmm3","xmm3",0b11001100);
 587         &movdqa (&QWP(0x30,"ebx"),"xmm3");
 588         &pmuludq("xmm3","xmm7");                # a[3]*b[0]
 589         &movdqa (&QWP(0x10,"esp"),"xmm2");
 590
 591         &movd   ("xmm1",&DWP(4*5,"esi"));
 592         &pshufd ("xmm0","xmm0",0b11001100);
 593         &movdqa (&QWP(0x40,"ebx"),"xmm0");
 594         &pmuludq("xmm0","xmm7");                # a[4]*b[0]
 595          &paddq ("xmm3","xmm5");                # a[3]*b[0]+lw(a[0]*b[0]), reduction step
 596         &movdqa (&QWP(0x20,"esp"),"xmm3");
 597
 598         &movd   ("xmm2",&DWP(4*6,"esi"));
 599         &pshufd ("xmm1","xmm1",0b11001100);
 600         &movdqa (&QWP(0x50,"ebx"),"xmm1");
 601         &pmuludq("xmm1","xmm7");                # a[5]*b[0]
 602         &movdqa (&QWP(0x30,"esp"),"xmm0");
 603          &pshufd("xmm4","xmm5",0b10110001);     # xmm4 = xmm5<<32, reduction step
 604
 605         &movd   ("xmm3",&DWP(4*7,"esi"));
 606         &pshufd ("xmm2","xmm2",0b11001100);
 607         &movdqa (&QWP(0x60,"ebx"),"xmm2");
 608         &pmuludq("xmm2","xmm7");                # a[6]*b[0]
 609         &movdqa (&QWP(0x40,"esp"),"xmm1");
 610          &psubq ("xmm4","xmm5");                # xmm4 = xmm5*0xffffffff, reduction step
 611
 612         &movd   ("xmm0",&DWP(0,"ebp"));         # b[1] -> 0000.00xy
 613         &pshufd ("xmm3","xmm3",0b11001100);
 614         &movdqa (&QWP(0x70,"ebx"),"xmm3");
 615         &pmuludq("xmm3","xmm7");                # a[7]*b[0]
 616
 617         &pshuflw("xmm7","xmm0",0b11011100);     # 0000.00xy -> 0000.0x0y
 618         &movdqa ("xmm0",&QWP(0x00,"ebx"));      # pre-load converted a[0]
 619         &pshufd ("xmm7","xmm7",0b11011100);     # 0000.0x0y -> 000x.000y
 620
 621         &mov    ("ecx",6);
 622         &lea    ("ebp",&DWP(4,"ebp"));
 623         &jmp    (&label("madd_sse2"));
 624
 625 &set_label("madd_sse2",16);
 626          &paddq ("xmm2","xmm5");                # a[6]*b[i-1]+lw(a[0]*b[i-1]), reduction step [modulo-scheduled]
 627          &paddq ("xmm3","xmm4");                # a[7]*b[i-1]+lw(a[0]*b[i-1])*0xffffffff, reduction step [modulo-scheduled]
 628         &movdqa ("xmm1",&QWP(0x10,"ebx"));
 629         &pmuludq("xmm0","xmm7");                # a[0]*b[i]
 630          &movdqa(&QWP(0x50,"esp"),"xmm2");
 631
 632         &movdqa ("xmm2",&QWP(0x20,"ebx"));
 633         &pmuludq("xmm1","xmm7");                # a[1]*b[i]
 634          &movdqa(&QWP(0x60,"esp"),"xmm3");
 635         &paddq  ("xmm0",&QWP(0x00,"esp"));
 636
 637         &movdqa ("xmm3",&QWP(0x30,"ebx"));
 638         &pmuludq("xmm2","xmm7");                # a[2]*b[i]
 639          &movq  ("xmm4","xmm0");                # clear upper 64 bits
 640          &pslldq("xmm4",6);
 641         &paddq  ("xmm1",&QWP(0x10,"esp"));
 642          &paddq ("xmm4","xmm0");
 643          &movdqa("xmm5","xmm4");
 644          &psrldq("xmm4",10);                    # upper 33 bits of a[0]*b[i]+t[0]
 645
 646         &movdqa ("xmm0",&QWP(0x40,"ebx"));
 647         &pmuludq("xmm3","xmm7");                # a[3]*b[i]
 648          &paddq ("xmm1","xmm4");                # a[1]*b[i]+hw(a[0]*b[i]), carry
 649         &paddq  ("xmm2",&QWP(0x20,"esp"));
 650         &movdqa (&QWP(0x00,"esp"),"xmm1");
 651
 652         &movdqa ("xmm1",&QWP(0x50,"ebx"));
 653         &pmuludq("xmm0","xmm7");                # a[4]*b[i]
 654         &paddq  ("xmm3",&QWP(0x30,"esp"));
 655         &movdqa (&QWP(0x10,"esp"),"xmm2");
 656          &pand  ("xmm5","xmm6");                # lower 32 bits of a[0]*b[i]
 657
 658         &movdqa ("xmm2",&QWP(0x60,"ebx"));
 659         &pmuludq("xmm1","xmm7");                # a[5]*b[i]
 660          &paddq ("xmm3","xmm5");                # a[3]*b[i]+lw(a[0]*b[i]), reduction step
 661         &paddq  ("xmm0",&QWP(0x40,"esp"));
 662         &movdqa (&QWP(0x20,"esp"),"xmm3");
 663          &pshufd("xmm4","xmm5",0b10110001);     # xmm4 = xmm5<<32, reduction step
 664
 665         &movdqa ("xmm3","xmm7");
 666         &pmuludq("xmm2","xmm7");                # a[6]*b[i]
 667          &movd  ("xmm7",&DWP(0,"ebp"));         # b[i++] -> 0000.00xy
 668          &lea   ("ebp",&DWP(4,"ebp"));
 669         &paddq  ("xmm1",&QWP(0x50,"esp"));
 670          &psubq ("xmm4","xmm5");                # xmm4 = xmm5*0xffffffff, reduction step
 671         &movdqa (&QWP(0x30,"esp"),"xmm0");
 672          &pshuflw("xmm7","xmm7",0b11011100);    # 0000.00xy -> 0000.0x0y
 673
 674         &pmuludq("xmm3",&QWP(0x70,"ebx"));      # a[7]*b[i]
 675          &pshufd("xmm7","xmm7",0b11011100);     # 0000.0x0y -> 000x.000y
 676          &movdqa("xmm0",&QWP(0x00,"ebx"));      # pre-load converted a[0]
 677         &movdqa (&QWP(0x40,"esp"),"xmm1");
 678         &paddq  ("xmm2",&QWP(0x60,"esp"));
 679
 680         &dec    ("ecx");
 681         &jnz    (&label("madd_sse2"));
 682
 683          &paddq ("xmm2","xmm5");                # a[6]*b[6]+lw(a[0]*b[6]), reduction step [modulo-scheduled]
 684          &paddq ("xmm3","xmm4");                # a[7]*b[6]+lw(a[0]*b[6])*0xffffffff, reduction step [modulo-scheduled]
 685         &movdqa ("xmm1",&QWP(0x10,"ebx"));
 686         &pmuludq("xmm0","xmm7");                # a[0]*b[7]
 687          &movdqa(&QWP(0x50,"esp"),"xmm2");
 688
 689         &movdqa ("xmm2",&QWP(0x20,"ebx"));
 690         &pmuludq("xmm1","xmm7");                # a[1]*b[7]
 691          &movdqa(&QWP(0x60,"esp"),"xmm3");
 692         &paddq  ("xmm0",&QWP(0x00,"esp"));
 693
 694         &movdqa ("xmm3",&QWP(0x30,"ebx"));
 695         &pmuludq("xmm2","xmm7");                # a[2]*b[7]
 696          &movq  ("xmm4","xmm0");                # clear upper 64 bits
 697          &pslldq("xmm4",6);
 698         &paddq  ("xmm1",&QWP(0x10,"esp"));
 699          &paddq ("xmm4","xmm0");
 700          &movdqa("xmm5","xmm4");
 701          &psrldq("xmm4",10);                    # upper 33 bits of a[0]*b[i]+t[0]
 702
 703         &movdqa ("xmm0",&QWP(0x40,"ebx"));
 704         &pmuludq("xmm3","xmm7");                # a[3]*b[7]
 705          &paddq ("xmm1","xmm4");                # a[1]*b[7]+hw(a[0]*b[7]), carry
 706         &paddq  ("xmm2",&QWP(0x20,"esp"));
 707         &movdqa (&QWP(0x00,"esp"),"xmm1");
 708
 709         &movdqa ("xmm1",&QWP(0x50,"ebx"));
 710         &pmuludq("xmm0","xmm7");                # a[4]*b[7]
 711         &paddq  ("xmm3",&QWP(0x30,"esp"));
 712         &movdqa (&QWP(0x10,"esp"),"xmm2");
 713          &pand  ("xmm5","xmm6");                # lower 32 bits of a[0]*b[i]
 714
 715         &movdqa ("xmm2",&QWP(0x60,"ebx"));
 716         &pmuludq("xmm1","xmm7");                # a[5]*b[7]
 717          &paddq ("xmm3","xmm5");                # reduction step
 718         &paddq  ("xmm0",&QWP(0x40,"esp"));
 719         &movdqa (&QWP(0x20,"esp"),"xmm3");
 720          &pshufd("xmm4","xmm5",0b10110001);     # xmm4 = xmm5<<32, reduction step
 721
 722         &movdqa ("xmm3",&QWP(0x70,"ebx"));
 723         &pmuludq("xmm2","xmm7");                # a[6]*b[7]
 724         &paddq  ("xmm1",&QWP(0x50,"esp"));
 725          &psubq ("xmm4","xmm5");                # xmm4 = xmm5*0xffffffff, reduction step
 726         &movdqa (&QWP(0x30,"esp"),"xmm0");
 727
 728         &pmuludq("xmm3","xmm7");                # a[7]*b[7]
 729         &pcmpeqd("xmm7","xmm7");
 730         &movdqa ("xmm0",&QWP(0x00,"esp"));
 731         &pslldq ("xmm7",8);
 732         &movdqa (&QWP(0x40,"esp"),"xmm1");
 733         &paddq  ("xmm2",&QWP(0x60,"esp"));
 734
 735          &paddq ("xmm2","xmm5");                # a[6]*b[7]+lw(a[0]*b[7]), reduction step
 736          &paddq ("xmm3","xmm4");                # a[6]*b[7]+lw(a[0]*b[7])*0xffffffff, reduction step
 737          &movdqa(&QWP(0x50,"esp"),"xmm2");
 738          &movdqa(&QWP(0x60,"esp"),"xmm3");
 739
 740         &movdqa ("xmm1",&QWP(0x10,"esp"));
 741         &movdqa ("xmm2",&QWP(0x20,"esp"));
 742         &movdqa ("xmm3",&QWP(0x30,"esp"));
 743
 744         &movq   ("xmm4","xmm0");                # "flatten"
 745         &pand   ("xmm0","xmm7");
 746         &xor    ("ebp","ebp");
 747         &pslldq ("xmm4",6);
 748          &movq  ("xmm5","xmm1");
 749         &paddq  ("xmm0","xmm4");
 750          &pand  ("xmm1","xmm7");
 751         &psrldq ("xmm0",6);
 752         &movd   ("eax","xmm0");
 753         &psrldq ("xmm0",4);
 754
 755         &paddq  ("xmm5","xmm0");
 756         &movdqa ("xmm0",&QWP(0x40,"esp"));
 757         &sub    ("eax",-1);                     # start subtracting modulus,
 758                                                 # this is used to determine
 759                                                 # if result is larger/smaller
 760                                                 # than modulus (see below)
 761         &pslldq ("xmm5",6);
 762          &movq  ("xmm4","xmm2");
 763         &paddq  ("xmm1","xmm5");
 764          &pand  ("xmm2","xmm7");
 765         &psrldq ("xmm1",6);
 766         &mov    (&DWP(4*0,"edi"),"eax");
 767         &movd   ("eax","xmm1");
 768         &psrldq ("xmm1",4);
 769
 770         &paddq  ("xmm4","xmm1");
 771         &movdqa ("xmm1",&QWP(0x50,"esp"));
 772         &sbb    ("eax",-1);
 773         &pslldq ("xmm4",6);
 774          &movq  ("xmm5","xmm3");
 775         &paddq  ("xmm2","xmm4");
 776          &pand  ("xmm3","xmm7");
 777         &psrldq ("xmm2",6);
 778         &mov    (&DWP(4*1,"edi"),"eax");
 779         &movd   ("eax","xmm2");
 780         &psrldq ("xmm2",4);
 781
 782         &paddq  ("xmm5","xmm2");
 783         &movdqa ("xmm2",&QWP(0x60,"esp"));
 784         &sbb    ("eax",-1);
 785         &pslldq ("xmm5",6);
 786          &movq  ("xmm4","xmm0");
 787         &paddq  ("xmm3","xmm5");
 788          &pand  ("xmm0","xmm7");
 789         &psrldq ("xmm3",6);
 790         &mov    (&DWP(4*2,"edi"),"eax");
 791         &movd   ("eax","xmm3");
 792         &psrldq ("xmm3",4);
 793
 794         &paddq  ("xmm4","xmm3");
 795         &sbb    ("eax",0);
 796         &pslldq ("xmm4",6);
 797          &movq  ("xmm5","xmm1");
 798         &paddq  ("xmm0","xmm4");
 799          &pand  ("xmm1","xmm7");
 800         &psrldq ("xmm0",6);
 801         &mov    (&DWP(4*3,"edi"),"eax");
 802         &movd   ("eax","xmm0");
 803         &psrldq ("xmm0",4);
 804
 805         &paddq  ("xmm5","xmm0");
 806         &sbb    ("eax",0);
 807         &pslldq ("xmm5",6);
 808          &movq  ("xmm4","xmm2");
 809         &paddq  ("xmm1","xmm5");
 810          &pand  ("xmm2","xmm7");
 811         &psrldq ("xmm1",6);
 812         &movd   ("ebx","xmm1");
 813         &psrldq ("xmm1",4);
 814         &mov    ("esp","edx");
 815
 816         &paddq  ("xmm4","xmm1");
 817         &pslldq ("xmm4",6);
 818         &paddq  ("xmm2","xmm4");
 819         &psrldq ("xmm2",6);
 820         &movd   ("ecx","xmm2");
 821         &psrldq ("xmm2",4);
 822         &sbb    ("ebx",0);
 823         &movd   ("edx","xmm2");
 824         &pextrw ("esi","xmm2",2);               # top-most overflow bit
 825         &sbb    ("ecx",1);
 826         &sbb    ("edx",-1);
 827         &sbb    ("esi",0);                      # borrow from subtraction
 828
 829         # Final step is "if result > mod, subtract mod", and at this point
 830         # we have result - mod written to output buffer, as well as borrow
 831         # bit from this subtraction, and if borrow bit is set, we add
 832         # modulus back.
 833         #
 834         # Note that because mod has special form, i.e. consists of
 835         # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
 836         # assigning borrow bit to one register, %ebp, and its negative
 837         # to another, %esi. But we started by calculating %esi...
 838
 839         &sub    ("ebp","esi");
 840         &add    (&DWP(4*0,"edi"),"esi");        # add modulus or zero
 841         &adc    (&DWP(4*1,"edi"),"esi");
 842         &adc    (&DWP(4*2,"edi"),"esi");
 843         &adc    (&DWP(4*3,"edi"),0);
 844         &adc    ("eax",0);
 845         &adc    ("ebx",0);
 846         &mov    (&DWP(4*4,"edi"),"eax");
 847         &adc    ("ecx","ebp");
 848         &mov    (&DWP(4*5,"edi"),"ebx");
 849         &adc    ("edx","esi");
 850         &mov    (&DWP(4*6,"edi"),"ecx");
 851         &mov    (&DWP(4*7,"edi"),"edx");
 852
 853         &ret    ();
 854
 855 &set_label("mul_mont_ialu",16);                 }
 856
 857         ########################################
 858         # IALU code path suitable for all CPUs.
 859         ########################################
 860         # stack layout:
 861         # +------------------------------------+< %esp
 862         # | 8 32-bit temporary words, accessed |
 863         # | as circular buffer                 |
 864         # .                                    .
 865         # .                                    .
 866         # +------------------------------------+< +32
 867         # | offloaded destination pointer      |
 868         # +------------------------------------+
 869         # | unused                             |
 870         # +------------------------------------+< +40
 871         &sub    ("esp",10*4);
 872
 873         &mov    ("eax",&DWP(0*4,"esi"));                # a[0]
 874         &mov    ("ebx",&DWP(0*4,"ebp"));                # b[0]
 875         &mov    (&DWP(8*4,"esp"),"edi");                # off-load dst ptr
 876
 877         &mul    ("ebx");                                # a[0]*b[0]
 878         &mov    (&DWP(0*4,"esp"),"eax");                # t[0]
 879         &mov    ("eax",&DWP(1*4,"esi"));
 880         &mov    ("ecx","edx")
 881
 882         &mul    ("ebx");                                # a[1]*b[0]
 883         &add    ("ecx","eax");
 884         &mov    ("eax",&DWP(2*4,"esi"));
 885         &adc    ("edx",0);
 886         &mov    (&DWP(1*4,"esp"),"ecx");                # t[1]
 887         &mov    ("ecx","edx");
 888
 889         &mul    ("ebx");                                # a[2]*b[0]
 890         &add    ("ecx","eax");
 891         &mov    ("eax",&DWP(3*4,"esi"));
 892         &adc    ("edx",0);
 893         &mov    (&DWP(2*4,"esp"),"ecx");                # t[2]
 894         &mov    ("ecx","edx");
 895
 896         &mul    ("ebx");                                # a[3]*b[0]
 897         &add    ("ecx","eax");
 898         &mov    ("eax",&DWP(4*4,"esi"));
 899         &adc    ("edx",0);
 900         &mov    (&DWP(3*4,"esp"),"ecx");                # t[3]
 901         &mov    ("ecx","edx");
 902
 903         &mul    ("ebx");                                # a[4]*b[0]
 904         &add    ("ecx","eax");
 905         &mov    ("eax",&DWP(5*4,"esi"));
 906         &adc    ("edx",0);
 907         &mov    (&DWP(4*4,"esp"),"ecx");                # t[4]
 908         &mov    ("ecx","edx");
 909
 910         &mul    ("ebx");                                # a[5]*b[0]
 911         &add    ("ecx","eax");
 912         &mov    ("eax",&DWP(6*4,"esi"));
 913         &adc    ("edx",0);
 914         &mov    (&DWP(5*4,"esp"),"ecx");                # t[5]
 915         &mov    ("ecx","edx");
 916
 917         &mul    ("ebx");                                # a[6]*b[0]
 918         &add    ("ecx","eax");
 919         &mov    ("eax",&DWP(7*4,"esi"));
 920         &adc    ("edx",0);
 921         &mov    (&DWP(6*4,"esp"),"ecx");                # t[6]
 922         &mov    ("ecx","edx");
 923
 924         &xor    ("edi","edi");                          # initial top-most carry
 925         &mul    ("ebx");                                # a[7]*b[0]
 926         &add    ("ecx","eax");                          # t[7]
 927         &mov    ("eax",&DWP(0*4,"esp"));                # t[0]
 928         &adc    ("edx",0);                              # t[8]
 929
 930 for ($i=0;$i<7;$i++) {
 931         my $j=$i+1;
 932
 933         # Reduction iteration is normally performed by accumulating
 934         # result of multiplication of modulus by "magic" digit [and
 935         # omitting least significant word, which is guaranteed to
 936         # be 0], but thanks to special form of modulus and "magic"
 937         # digit being equal to least significant word, it can be
 938         # performed with additions and subtractions alone. Indeed:
 939         #
 940         #        ffff.0001.0000.0000.0000.ffff.ffff.ffff
 941         # *                                         abcd
 942         # + xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.abcd
 943         #
 944         # Now observing that ff..ff*x = (2^n-1)*x = 2^n*x-x, we
 945         # rewrite above as:
 946         #
 947         #   xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.abcd
 948         # + abcd.0000.abcd.0000.0000.abcd.0000.0000.0000
 949         # -      abcd.0000.0000.0000.0000.0000.0000.abcd
 950         #
 951         # or marking redundant operations:
 952         #
 953         #   xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.----
 954         # + abcd.0000.abcd.0000.0000.abcd.----.----.----
 955         # -      abcd.----.----.----.----.----.----.----
 956
 957         &add    (&DWP((($i+3)%8)*4,"esp"),"eax");       # t[3]+=t[0]
 958         &adc    (&DWP((($i+4)%8)*4,"esp"),0);           # t[4]+=0
 959         &adc    (&DWP((($i+5)%8)*4,"esp"),0);           # t[5]+=0
 960         &adc    (&DWP((($i+6)%8)*4,"esp"),"eax");       # t[6]+=t[0]
 961         &adc    ("ecx",0);                              # t[7]+=0
 962         &adc    ("edx","eax");                          # t[8]+=t[0]
 963         &adc    ("edi",0);                              # top-most carry
 964          &mov   ("ebx",&DWP($j*4,"ebp"));               # b[i]
 965         &sub    ("ecx","eax");                          # t[7]-=t[0]
 966          &mov   ("eax",&DWP(0*4,"esi"));                # a[0]
 967         &sbb    ("edx",0);                              # t[8]-=0
 968         &mov    (&DWP((($i+7)%8)*4,"esp"),"ecx");
 969         &sbb    ("edi",0);                              # top-most carry,
 970                                                         # keep in mind that
 971                                                         # netto result is
 972                                                         # *addition* of value
 973                                                         # with (abcd<<32)-abcd
 974                                                         # on top, so that
 975                                                         # underflow is
 976                                                         # impossible, because
 977                                                         # (abcd<<32)-abcd
 978                                                         # doesn't underflow
 979         &mov    (&DWP((($i+8)%8)*4,"esp"),"edx");
 980
 981         &mul    ("ebx");                                # a[0]*b[i]
 982         &add    ("eax",&DWP((($j+0)%8)*4,"esp"));
 983         &adc    ("edx",0);
 984         &mov    (&DWP((($j+0)%8)*4,"esp"),"eax");
 985         &mov    ("eax",&DWP(1*4,"esi"));
 986         &mov    ("ecx","edx")
 987
 988         &mul    ("ebx");                                # a[1]*b[i]
 989         &add    ("ecx",&DWP((($j+1)%8)*4,"esp"));
 990         &adc    ("edx",0);
 991         &add    ("ecx","eax");
 992         &adc    ("edx",0);
 993         &mov    ("eax",&DWP(2*4,"esi"));
 994         &mov    (&DWP((($j+1)%8)*4,"esp"),"ecx");
 995         &mov    ("ecx","edx");
 996
 997         &mul    ("ebx");                                # a[2]*b[i]
 998         &add    ("ecx",&DWP((($j+2)%8)*4,"esp"));
 999         &adc    ("edx",0);
1000         &add    ("ecx","eax");
1001         &adc    ("edx",0);
1002         &mov    ("eax",&DWP(3*4,"esi"));
1003         &mov    (&DWP((($j+2)%8)*4,"esp"),"ecx");
1004         &mov    ("ecx","edx");
1005
1006         &mul    ("ebx");                                # a[3]*b[i]
1007         &add    ("ecx",&DWP((($j+3)%8)*4,"esp"));
1008         &adc    ("edx",0);
1009         &add    ("ecx","eax");
1010         &adc    ("edx",0);
1011         &mov    ("eax",&DWP(4*4,"esi"));
1012         &mov    (&DWP((($j+3)%8)*4,"esp"),"ecx");
1013         &mov    ("ecx","edx");
1014
1015         &mul    ("ebx");                                # a[4]*b[i]
1016         &add    ("ecx",&DWP((($j+4)%8)*4,"esp"));
1017         &adc    ("edx",0);
1018         &add    ("ecx","eax");
1019         &adc    ("edx",0);
1020         &mov    ("eax",&DWP(5*4,"esi"));
1021         &mov    (&DWP((($j+4)%8)*4,"esp"),"ecx");
1022         &mov    ("ecx","edx");
1023
1024         &mul    ("ebx");                                # a[5]*b[i]
1025         &add    ("ecx",&DWP((($j+5)%8)*4,"esp"));
1026         &adc    ("edx",0);
1027         &add    ("ecx","eax");
1028         &adc    ("edx",0);
1029         &mov    ("eax",&DWP(6*4,"esi"));
1030         &mov    (&DWP((($j+5)%8)*4,"esp"),"ecx");
1031         &mov    ("ecx","edx");
1032
1033         &mul    ("ebx");                                # a[6]*b[i]
1034         &add    ("ecx",&DWP((($j+6)%8)*4,"esp"));
1035         &adc    ("edx",0);
1036         &add    ("ecx","eax");
1037         &adc    ("edx",0);
1038         &mov    ("eax",&DWP(7*4,"esi"));
1039         &mov    (&DWP((($j+6)%8)*4,"esp"),"ecx");
1040         &mov    ("ecx","edx");
1041
1042         &mul    ("ebx");                                # a[7]*b[i]
1043         &add    ("ecx",&DWP((($j+7)%8)*4,"esp"));
1044         &adc    ("edx",0);
1045         &add    ("ecx","eax");                          # t[7]
1046         &mov    ("eax",&DWP((($j+0)%8)*4,"esp"));       # t[0]
1047         &adc    ("edx","edi");                          # t[8]
1048         &mov    ("edi",0);
1049         &adc    ("edi",0);                              # top-most carry
1050 }
1051         &mov    ("ebp",&DWP(8*4,"esp"));                # restore dst ptr
1052         &xor    ("esi","esi");
1053         my $j=$i+1;
1054
1055         # last multiplication-less reduction
1056         &add    (&DWP((($i+3)%8)*4,"esp"),"eax");       # t[3]+=t[0]
1057         &adc    (&DWP((($i+4)%8)*4,"esp"),0);           # t[4]+=0
1058         &adc    (&DWP((($i+5)%8)*4,"esp"),0);           # t[5]+=0
1059         &adc    (&DWP((($i+6)%8)*4,"esp"),"eax");       # t[6]+=t[0]
1060         &adc    ("ecx",0);                              # t[7]+=0
1061         &adc    ("edx","eax");                          # t[8]+=t[0]
1062         &adc    ("edi",0);                              # top-most carry
1063          &mov   ("ebx",&DWP((($j+1)%8)*4,"esp"));
1064         &sub    ("ecx","eax");                          # t[7]-=t[0]
1065          &mov   ("eax",&DWP((($j+0)%8)*4,"esp"));
1066         &sbb    ("edx",0);                              # t[8]-=0
1067         &mov    (&DWP((($i+7)%8)*4,"esp"),"ecx");
1068         &sbb    ("edi",0);                              # top-most carry
1069         &mov    (&DWP((($i+8)%8)*4,"esp"),"edx");
1070
1071         # Final step is "if result > mod, subtract mod", but we do it
1072         # "other way around", namely write result - mod to output buffer
1073         # and if subtraction borrowed, add modulus back.
1074
1075         &mov    ("ecx",&DWP((($j+2)%8)*4,"esp"));
1076         &sub    ("eax",-1);
1077         &mov    ("edx",&DWP((($j+3)%8)*4,"esp"));
1078         &sbb    ("ebx",-1);
1079         &mov    (&DWP(0*4,"ebp"),"eax");
1080         &sbb    ("ecx",-1);
1081         &mov    (&DWP(1*4,"ebp"),"ebx");
1082         &sbb    ("edx",0);
1083         &mov    (&DWP(2*4,"ebp"),"ecx");
1084         &mov    (&DWP(3*4,"ebp"),"edx");
1085
1086         &mov    ("eax",&DWP((($j+4)%8)*4,"esp"));
1087         &mov    ("ebx",&DWP((($j+5)%8)*4,"esp"));
1088         &mov    ("ecx",&DWP((($j+6)%8)*4,"esp"));
1089         &sbb    ("eax",0);
1090         &mov    ("edx",&DWP((($j+7)%8)*4,"esp"));
1091         &sbb    ("ebx",0);
1092         &sbb    ("ecx",1);
1093         &sbb    ("edx",-1);
1094         &sbb    ("edi",0);
1095
1096         # Note that because mod has special form, i.e. consists of
1097         # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
1098         # assigning borrow bit to one register, %ebp, and its negative
1099         # to another, %esi. But we started by calculating %esi...
1100
1101         &sub    ("esi","edi");
1102         &add    (&DWP(0*4,"ebp"),"edi");                # add modulus or zero
1103         &adc    (&DWP(1*4,"ebp"),"edi");
1104         &adc    (&DWP(2*4,"ebp"),"edi");
1105         &adc    (&DWP(3*4,"ebp"),0);
1106         &adc    ("eax",0);
1107         &adc    ("ebx",0);
1108         &mov    (&DWP(4*4,"ebp"),"eax");
1109         &adc    ("ecx","esi");
1110         &mov    (&DWP(5*4,"ebp"),"ebx");
1111         &adc    ("edx","edi");
1112         &mov    (&DWP(6*4,"ebp"),"ecx");
1113         &mov    ("edi","ebp");                          # fulfill contract
1114         &mov    (&DWP(7*4,"ebp"),"edx");
1115
1116         &add    ("esp",10*4);
1117         &ret    ();
1118 &function_end_B("_ecp_nistz256_mul_mont");
1119
1120 ########################################################################
1121 # void ecp_nistz256_scatter_w5(void *edi,const P256_POINT *esi,
1122 #                                        int ebp);
1123 &function_begin("ecp_nistz256_scatter_w5");
1124         &mov    ("edi",&wparam(0));
1125         &mov    ("esi",&wparam(1));
1126         &mov    ("ebp",&wparam(2));
1127
1128         &lea    ("edi",&DWP(128-4,"edi","ebp",4));
1129         &mov    ("ebp",96/16);
1130 &set_label("scatter_w5_loop");
1131         &mov    ("eax",&DWP(0,"esi"));
1132         &mov    ("ebx",&DWP(4,"esi"));
1133         &mov    ("ecx",&DWP(8,"esi"));
1134         &mov    ("edx",&DWP(12,"esi"));
1135         &lea    ("esi",&DWP(16,"esi"));
1136         &mov    (&DWP(64*0-128,"edi"),"eax");
1137         &mov    (&DWP(64*1-128,"edi"),"ebx");
1138         &mov    (&DWP(64*2-128,"edi"),"ecx");
1139         &mov    (&DWP(64*3-128,"edi"),"edx");
1140         &lea    ("edi",&DWP(64*4,"edi"));
1141         &dec    ("ebp");
1142         &jnz    (&label("scatter_w5_loop"));
1143 &function_end("ecp_nistz256_scatter_w5");
1144
1145 ########################################################################
1146 # void ecp_nistz256_gather_w5(P256_POINT *edi,const void *esi,
1147 #                                             int ebp);
1148 &function_begin("ecp_nistz256_gather_w5");
1149         &mov    ("esi",&wparam(1));
1150         &mov    ("ebp",&wparam(2));
1151
1152         &lea    ("esi",&DWP(0,"esi","ebp",4));
1153         &neg    ("ebp");
1154         &sar    ("ebp",31);
1155         &mov    ("edi",&wparam(0));
1156         &lea    ("esi",&DWP(0,"esi","ebp",4));
1157
1158     for($i=0;$i<24;$i+=4) {
1159         &mov    ("eax",&DWP(64*($i+0),"esi"));
1160         &mov    ("ebx",&DWP(64*($i+1),"esi"));
1161         &mov    ("ecx",&DWP(64*($i+2),"esi"));
1162         &mov    ("edx",&DWP(64*($i+3),"esi"));
1163         &and    ("eax","ebp");
1164         &and    ("ebx","ebp");
1165         &and    ("ecx","ebp");
1166         &and    ("edx","ebp");
1167         &mov    (&DWP(4*($i+0),"edi"),"eax");
1168         &mov    (&DWP(4*($i+1),"edi"),"ebx");
1169         &mov    (&DWP(4*($i+2),"edi"),"ecx");
1170         &mov    (&DWP(4*($i+3),"edi"),"edx");
1171     }
1172 &function_end("ecp_nistz256_gather_w5");
1173
1174 ########################################################################
1175 # void ecp_nistz256_scatter_w7(void *edi,const P256_POINT_AFFINE *esi,
1176 #                                        int ebp);
1177 &function_begin("ecp_nistz256_scatter_w7");
1178         &mov    ("edi",&wparam(0));
1179         &mov    ("esi",&wparam(1));
1180         &mov    ("ebp",&wparam(2));
1181
1182         &lea    ("edi",&DWP(0,"edi","ebp"));
1183         &mov    ("ebp",64/4);
1184 &set_label("scatter_w7_loop");
1185         &mov    ("eax",&DWP(0,"esi"));
1186         &lea    ("esi",&DWP(4,"esi"));
1187         &mov    (&BP(64*0,"edi"),"al");
1188         &mov    (&BP(64*1,"edi"),"ah");
1189         &shr    ("eax",16);
1190         &mov    (&BP(64*2,"edi"),"al");
1191         &mov    (&BP(64*3,"edi"),"ah");
1192         &lea    ("edi",&DWP(64*4,"edi"));
1193         &dec    ("ebp");
1194         &jnz    (&label("scatter_w7_loop"));
1195 &function_end("ecp_nistz256_scatter_w7");
1196
1197 ########################################################################
1198 # void ecp_nistz256_gather_w7(P256_POINT_AFFINE *edi,const void *esi,
1199 #                                                    int ebp);
1200 &function_begin("ecp_nistz256_gather_w7");
1201         &mov    ("esi",&wparam(1));
1202         &mov    ("ebp",&wparam(2));
1203
1204         &add    ("esi","ebp");
1205         &neg    ("ebp"),
1206         &sar    ("ebp",31);
1207         &mov    ("edi",&wparam(0));
1208         &lea    ("esi",&DWP(0,"esi","ebp"));
1209
1210     for($i=0;$i<64;$i+=4) {
1211         &movz   ("eax",&BP(64*($i+0),"esi"));
1212         &movz   ("ebx",&BP(64*($i+1),"esi"));
1213         &movz   ("ecx",&BP(64*($i+2),"esi"));
1214         &and    ("eax","ebp");
1215         &movz   ("edx",&BP(64*($i+3),"esi"));
1216         &and    ("ebx","ebp");
1217         &mov    (&BP($i+0,"edi"),"al");
1218         &and    ("ecx","ebp");
1219         &mov    (&BP($i+1,"edi"),"bl");
1220         &and    ("edx","ebp");
1221         &mov    (&BP($i+2,"edi"),"cl");
1222         &mov    (&BP($i+3,"edi"),"dl");
1223     }
1224 &function_end("ecp_nistz256_gather_w7");
1225
1226 ########################################################################
1227 # following subroutines are "literal" implementation of those found in
1228 # ecp_nistz256.c
1229 #
1230 ########################################################################
1231 # void ecp_nistz256_point_double(P256_POINT *out,const P256_POINT *inp);
1232 #
1233 &static_label("point_double_shortcut");
1234 &function_begin("ecp_nistz256_point_double");
1235 {   my ($S,$M,$Zsqr,$in_x,$tmp0)=map(32*$_,(0..4));
1236
1237         &mov    ("esi",&wparam(1));
1238
1239         # above map() describes stack layout with 5 temporary
1240         # 256-bit vectors on top, then we take extra word for
1241         # OPENSSL_ia32cap_P copy.
1242         &stack_push(8*5+1);
1243                                                 if ($sse2) {
1244         &call   ("_picup_eax");
1245     &set_label("pic");
1246         &picmeup("edx","OPENSSL_ia32cap_P","eax",&label("pic"));
1247         &mov    ("ebp",&DWP(0,"edx"));          }
1248
1249 &set_label("point_double_shortcut");
1250         &mov    ("eax",&DWP(0,"esi"));          # copy in_x
1251         &mov    ("ebx",&DWP(4,"esi"));
1252         &mov    ("ecx",&DWP(8,"esi"));
1253         &mov    ("edx",&DWP(12,"esi"));
1254         &mov    (&DWP($in_x+0,"esp"),"eax");
1255         &mov    (&DWP($in_x+4,"esp"),"ebx");
1256         &mov    (&DWP($in_x+8,"esp"),"ecx");
1257         &mov    (&DWP($in_x+12,"esp"),"edx");
1258         &mov    ("eax",&DWP(16,"esi"));
1259         &mov    ("ebx",&DWP(20,"esi"));
1260         &mov    ("ecx",&DWP(24,"esi"));
1261         &mov    ("edx",&DWP(28,"esi"));
1262         &mov    (&DWP($in_x+16,"esp"),"eax");
1263         &mov    (&DWP($in_x+20,"esp"),"ebx");
1264         &mov    (&DWP($in_x+24,"esp"),"ecx");
1265         &mov    (&DWP($in_x+28,"esp"),"edx");
1266         &mov    (&DWP(32*5,"esp"),"ebp");       # OPENSSL_ia32cap_P copy
1267
1268         &lea    ("ebp",&DWP(32,"esi"));
1269         &lea    ("esi",&DWP(32,"esi"));
1270         &lea    ("edi",&DWP($S,"esp"));
1271         &call   ("_ecp_nistz256_add");          # p256_mul_by_2(S, in_y);
1272
1273         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1274         &mov    ("esi",64);
1275         &add    ("esi",&wparam(1));
1276         &lea    ("edi",&DWP($Zsqr,"esp"));
1277         &mov    ("ebp","esi");
1278         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Zsqr, in_z);
1279
1280         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1281         &lea    ("esi",&DWP($S,"esp"));
1282         &lea    ("ebp",&DWP($S,"esp"));
1283         &lea    ("edi",&DWP($S,"esp"));
1284         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(S, S);
1285
1286         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1287         &mov    ("ebp",&wparam(1));
1288         &lea    ("esi",&DWP(32,"ebp"));
1289         &lea    ("ebp",&DWP(64,"ebp"));
1290         &lea    ("edi",&DWP($tmp0,"esp"));
1291         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(tmp0, in_z, in_y);
1292
1293         &lea    ("esi",&DWP($in_x,"esp"));
1294         &lea    ("ebp",&DWP($Zsqr,"esp"));
1295         &lea    ("edi",&DWP($M,"esp"));
1296         &call   ("_ecp_nistz256_add");          # p256_add(M, in_x, Zsqr);
1297
1298         &mov    ("edi",64);
1299         &lea    ("esi",&DWP($tmp0,"esp"));
1300         &lea    ("ebp",&DWP($tmp0,"esp"));
1301         &add    ("edi",&wparam(0));
1302         &call   ("_ecp_nistz256_add");          # p256_mul_by_2(res_z, tmp0);
1303
1304         &lea    ("esi",&DWP($in_x,"esp"));
1305         &lea    ("ebp",&DWP($Zsqr,"esp"));
1306         &lea    ("edi",&DWP($Zsqr,"esp"));
1307         &call   ("_ecp_nistz256_sub");          # p256_sub(Zsqr, in_x, Zsqr);
1308
1309         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1310         &lea    ("esi",&DWP($S,"esp"));
1311         &lea    ("ebp",&DWP($S,"esp"));
1312         &lea    ("edi",&DWP($tmp0,"esp"));
1313         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(tmp0, S);
1314
1315         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1316         &lea    ("esi",&DWP($M,"esp"));
1317         &lea    ("ebp",&DWP($Zsqr,"esp"));
1318         &lea    ("edi",&DWP($M,"esp"));
1319         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(M, M, Zsqr);
1320
1321         &mov    ("edi",32);
1322         &lea    ("esi",&DWP($tmp0,"esp"));
1323         &add    ("edi",&wparam(0));
1324         &call   ("_ecp_nistz256_div_by_2");     # p256_div_by_2(res_y, tmp0);
1325
1326         &lea    ("esi",&DWP($M,"esp"));
1327         &lea    ("ebp",&DWP($M,"esp"));
1328         &lea    ("edi",&DWP($tmp0,"esp"));
1329         &call   ("_ecp_nistz256_add");          # 1/2 p256_mul_by_3(M, M);
1330
1331         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1332         &lea    ("esi",&DWP($in_x,"esp"));
1333         &lea    ("ebp",&DWP($S,"esp"));
1334         &lea    ("edi",&DWP($S,"esp"));
1335         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S, S, in_x);
1336
1337         &lea    ("esi",&DWP($tmp0,"esp"));
1338         &lea    ("ebp",&DWP($M,"esp"));
1339         &lea    ("edi",&DWP($M,"esp"));
1340         &call   ("_ecp_nistz256_add");          # 2/2 p256_mul_by_3(M, M);
1341
1342         &lea    ("esi",&DWP($S,"esp"));
1343         &lea    ("ebp",&DWP($S,"esp"));
1344         &lea    ("edi",&DWP($tmp0,"esp"));
1345         &call   ("_ecp_nistz256_add");          # p256_mul_by_2(tmp0, S);
1346
1347         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1348         &lea    ("esi",&DWP($M,"esp"));
1349         &lea    ("ebp",&DWP($M,"esp"));
1350         &mov    ("edi",&wparam(0));
1351         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(res_x, M);
1352
1353         &mov    ("esi","edi");                  # %edi is still res_x here
1354         &lea    ("ebp",&DWP($tmp0,"esp"));
1355         &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, res_x, tmp0);
1356
1357         &lea    ("esi",&DWP($S,"esp"));
1358         &mov    ("ebp","edi");                  # %edi is still res_x
1359         &lea    ("edi",&DWP($S,"esp"));
1360         &call   ("_ecp_nistz256_sub");          # p256_sub(S, S, res_x);
1361
1362         &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
1363         &mov    ("esi","edi");                  # %edi is still &S
1364         &lea    ("ebp",&DWP($M,"esp"));
1365         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S, S, M);
1366
1367         &mov    ("ebp",32);
1368         &lea    ("esi",&DWP($S,"esp"));
1369         &add    ("ebp",&wparam(0));
1370         &mov    ("edi","ebp");
1371         &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, S, res_y);
1372
1373         &stack_pop(8*5+1);
1374 } &function_end("ecp_nistz256_point_double");
1375
1376 ########################################################################
1377 # void ecp_nistz256_point_add(P256_POINT *out,const P256_POINT *in1,
1378 #                                             const P256_POINT *in2);
1379 &function_begin("ecp_nistz256_point_add");
1380 {   my ($res_x,$res_y,$res_z,
1381         $in1_x,$in1_y,$in1_z,
1382         $in2_x,$in2_y,$in2_z,
1383         $H,$Hsqr,$R,$Rsqr,$Hcub,
1384         $U1,$U2,$S1,$S2)=map(32*$_,(0..17));
1385     my ($Z1sqr, $Z2sqr) = ($Hsqr, $Rsqr);
1386
1387         &mov    ("esi",&wparam(2));
1388
1389         # above map() describes stack layout with 18 temporary
1390         # 256-bit vectors on top, then we take extra words for
1391         # !in1infty, !in2infty, result of check for zero and
1392         # OPENSSL_ia32cap_P copy. [one unused word for padding]
1393         &stack_push(8*18+5);
1394                                                 if ($sse2) {
1395         &call   ("_picup_eax");
1396     &set_label("pic");
1397         &picmeup("edx","OPENSSL_ia32cap_P","eax",&label("pic"));
1398         &mov    ("ebp",&DWP(0,"edx"));          }
1399
1400         &lea    ("edi",&DWP($in2_x,"esp"));
1401     for($i=0;$i<96;$i+=16) {
1402         &mov    ("eax",&DWP($i+0,"esi"));       # copy in2
1403         &mov    ("ebx",&DWP($i+4,"esi"));
1404         &mov    ("ecx",&DWP($i+8,"esi"));
1405         &mov    ("edx",&DWP($i+12,"esi"));
1406         &mov    (&DWP($i+0,"edi"),"eax");
1407         &mov    (&DWP(32*18+12,"esp"),"ebp")    if ($i==0);
1408         &mov    ("ebp","eax")                   if ($i==64);
1409         &or     ("ebp","eax")                   if ($i>64);
1410         &mov    (&DWP($i+4,"edi"),"ebx");
1411         &or     ("ebp","ebx")                   if ($i>=64);
1412         &mov    (&DWP($i+8,"edi"),"ecx");
1413         &or     ("ebp","ecx")                   if ($i>=64);
1414         &mov    (&DWP($i+12,"edi"),"edx");
1415         &or     ("ebp","edx")                   if ($i>=64);
1416     }
1417         &xor    ("eax","eax");
1418         &mov    ("esi",&wparam(1));
1419         &sub    ("eax","ebp");
1420         &or     ("ebp","eax");
1421         &sar    ("ebp",31);
1422         &mov    (&DWP(32*18+4,"esp"),"ebp");    # !in2infty
1423
1424         &lea    ("edi",&DWP($in1_x,"esp"));
1425     for($i=0;$i<96;$i+=16) {
1426         &mov    ("eax",&DWP($i+0,"esi"));       # copy in1
1427         &mov    ("ebx",&DWP($i+4,"esi"));
1428         &mov    ("ecx",&DWP($i+8,"esi"));
1429         &mov    ("edx",&DWP($i+12,"esi"));
1430         &mov    (&DWP($i+0,"edi"),"eax");
1431         &mov    ("ebp","eax")                   if ($i==64);
1432         &or     ("ebp","eax")                   if ($i>64);
1433         &mov    (&DWP($i+4,"edi"),"ebx");
1434         &or     ("ebp","ebx")                   if ($i>=64);
1435         &mov    (&DWP($i+8,"edi"),"ecx");
1436         &or     ("ebp","ecx")                   if ($i>=64);
1437         &mov    (&DWP($i+12,"edi"),"edx");
1438         &or     ("ebp","edx")                   if ($i>=64);
1439     }
1440         &xor    ("eax","eax");
1441         &sub    ("eax","ebp");
1442         &or     ("ebp","eax");
1443         &sar    ("ebp",31);
1444         &mov    (&DWP(32*18+0,"esp"),"ebp");    # !in1infty
1445
1446         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1447         &lea    ("esi",&DWP($in2_z,"esp"));
1448         &lea    ("ebp",&DWP($in2_z,"esp"));
1449         &lea    ("edi",&DWP($Z2sqr,"esp"));
1450         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Z2sqr, in2_z);
1451
1452         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1453         &lea    ("esi",&DWP($in1_z,"esp"));
1454         &lea    ("ebp",&DWP($in1_z,"esp"));
1455         &lea    ("edi",&DWP($Z1sqr,"esp"));
1456         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Z1sqr, in1_z);
1457
1458         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1459         &lea    ("esi",&DWP($Z2sqr,"esp"));
1460         &lea    ("ebp",&DWP($in2_z,"esp"));
1461         &lea    ("edi",&DWP($S1,"esp"));
1462         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S1, Z2sqr, in2_z);
1463
1464         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1465         &lea    ("esi",&DWP($Z1sqr,"esp"));
1466         &lea    ("ebp",&DWP($in1_z,"esp"));
1467         &lea    ("edi",&DWP($S2,"esp"));
1468         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, Z1sqr, in1_z);
1469
1470         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1471         &lea    ("esi",&DWP($in1_y,"esp"));
1472         &lea    ("ebp",&DWP($S1,"esp"));
1473         &lea    ("edi",&DWP($S1,"esp"));
1474         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S1, S1, in1_y);
1475
1476         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1477         &lea    ("esi",&DWP($in2_y,"esp"));
1478         &lea    ("ebp",&DWP($S2,"esp"));
1479         &lea    ("edi",&DWP($S2,"esp"));
1480         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, S2, in2_y);
1481
1482         &lea    ("esi",&DWP($S2,"esp"));
1483         &lea    ("ebp",&DWP($S1,"esp"));
1484         &lea    ("edi",&DWP($R,"esp"));
1485         &call   ("_ecp_nistz256_sub");          # p256_sub(R, S2, S1);
1486
1487         &or     ("ebx","eax");                  # see if result is zero
1488         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1489         &or     ("ebx","ecx");
1490         &or     ("ebx","edx");
1491         &or     ("ebx",&DWP(0,"edi"));
1492         &or     ("ebx",&DWP(4,"edi"));
1493          &lea   ("esi",&DWP($in1_x,"esp"));
1494         &or     ("ebx",&DWP(8,"edi"));
1495          &lea   ("ebp",&DWP($Z2sqr,"esp"));
1496         &or     ("ebx",&DWP(12,"edi"));
1497          &lea   ("edi",&DWP($U1,"esp"));
1498         &mov    (&DWP(32*18+8,"esp"),"ebx");
1499
1500         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U1, in1_x, Z2sqr);
1501
1502         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1503         &lea    ("esi",&DWP($in2_x,"esp"));
1504         &lea    ("ebp",&DWP($Z1sqr,"esp"));
1505         &lea    ("edi",&DWP($U2,"esp"));
1506         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, in2_x, Z1sqr);
1507
1508         &lea    ("esi",&DWP($U2,"esp"));
1509         &lea    ("ebp",&DWP($U1,"esp"));
1510         &lea    ("edi",&DWP($H,"esp"));
1511         &call   ("_ecp_nistz256_sub");          # p256_sub(H, U2, U1);
1512
1513         &or     ("eax","ebx");                  # see if result is zero
1514         &or     ("eax","ecx");
1515         &or     ("eax","edx");
1516         &or     ("eax",&DWP(0,"edi"));
1517         &or     ("eax",&DWP(4,"edi"));
1518         &or     ("eax",&DWP(8,"edi"));
1519         &or     ("eax",&DWP(12,"edi"));
1520
1521         &data_byte(0x3e);                       # predict taken
1522         &jnz    (&label("add_proceed"));        # is_equal(U1,U2)?
1523
1524         &mov    ("eax",&DWP(32*18+0,"esp"));
1525         &and    ("eax",&DWP(32*18+4,"esp"));
1526         &mov    ("ebx",&DWP(32*18+8,"esp"));
1527         &jz     (&label("add_proceed"));        # (in1infty || in2infty)?
1528         &test   ("ebx","ebx");
1529         &jz     (&label("add_double"));         # is_equal(S1,S2)?
1530
1531         &mov    ("edi",&wparam(0));
1532         &xor    ("eax","eax");
1533         &mov    ("ecx",96/4);
1534         &data_byte(0xfc,0xf3,0xab);             # cld; stosd
1535         &jmp    (&label("add_done"));
1536
1537 &set_label("add_double",16);
1538         &mov    ("esi",&wparam(1));
1539         &mov    ("ebp",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1540         &add    ("esp",4*((8*18+5)-(8*5+1)));   # difference in frame sizes
1541         &jmp    (&label("point_double_shortcut"));
1542
1543 &set_label("add_proceed",16);
1544         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1545         &lea    ("esi",&DWP($R,"esp"));
1546         &lea    ("ebp",&DWP($R,"esp"));
1547         &lea    ("edi",&DWP($Rsqr,"esp"));
1548         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Rsqr, R);
1549
1550         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1551         &lea    ("esi",&DWP($H,"esp"));
1552         &lea    ("ebp",&DWP($in1_z,"esp"));
1553         &lea    ("edi",&DWP($res_z,"esp"));
1554         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_z, H, in1_z);
1555
1556         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1557         &lea    ("esi",&DWP($H,"esp"));
1558         &lea    ("ebp",&DWP($H,"esp"));
1559         &lea    ("edi",&DWP($Hsqr,"esp"));
1560         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Hsqr, H);
1561
1562         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1563         &lea    ("esi",&DWP($in2_z,"esp"));
1564         &lea    ("ebp",&DWP($res_z,"esp"));
1565         &lea    ("edi",&DWP($res_z,"esp"));
1566         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_z, res_z, in2_z);
1567
1568         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1569         &lea    ("esi",&DWP($Hsqr,"esp"));
1570         &lea    ("ebp",&DWP($U1,"esp"));
1571         &lea    ("edi",&DWP($U2,"esp"));
1572         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, U1, Hsqr);
1573
1574         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1575         &lea    ("esi",&DWP($H,"esp"));
1576         &lea    ("ebp",&DWP($Hsqr,"esp"));
1577         &lea    ("edi",&DWP($Hcub,"esp"));
1578         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(Hcub, Hsqr, H);
1579
1580         &lea    ("esi",&DWP($U2,"esp"));
1581         &lea    ("ebp",&DWP($U2,"esp"));
1582         &lea    ("edi",&DWP($Hsqr,"esp"));
1583         &call   ("_ecp_nistz256_add");          # p256_mul_by_2(Hsqr, U2);
1584
1585         &lea    ("esi",&DWP($Rsqr,"esp"));
1586         &lea    ("ebp",&DWP($Hsqr,"esp"));
1587         &lea    ("edi",&DWP($res_x,"esp"));
1588         &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, Rsqr, Hsqr);
1589
1590         &lea    ("esi",&DWP($res_x,"esp"));
1591         &lea    ("ebp",&DWP($Hcub,"esp"));
1592         &lea    ("edi",&DWP($res_x,"esp"));
1593         &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, res_x, Hcub);
1594
1595         &lea    ("esi",&DWP($U2,"esp"));
1596         &lea    ("ebp",&DWP($res_x,"esp"));
1597         &lea    ("edi",&DWP($res_y,"esp"));
1598         &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, U2, res_x);
1599
1600         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1601         &lea    ("esi",&DWP($Hcub,"esp"));
1602         &lea    ("ebp",&DWP($S1,"esp"));
1603         &lea    ("edi",&DWP($S2,"esp"));
1604         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, S1, Hcub);
1605
1606         &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
1607         &lea    ("esi",&DWP($R,"esp"));
1608         &lea    ("ebp",&DWP($res_y,"esp"));
1609         &lea    ("edi",&DWP($res_y,"esp"));
1610         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_y, R, res_y);
1611
1612         &lea    ("esi",&DWP($res_y,"esp"));
1613         &lea    ("ebp",&DWP($S2,"esp"));
1614         &lea    ("edi",&DWP($res_y,"esp"));
1615         &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, res_y, S2);
1616
1617         &mov    ("ebp",&DWP(32*18+0,"esp"));    # !in1infty
1618         &mov    ("esi",&DWP(32*18+4,"esp"));    # !in2infty
1619         &mov    ("edi",&wparam(0));
1620         &mov    ("edx","ebp");
1621         &not    ("ebp");
1622         &and    ("edx","esi");
1623         &and    ("ebp","esi");
1624         &not    ("esi");
1625
1626         ########################################
1627         # conditional moves
1628     for($i=64;$i<96;$i+=4) {
1629         &mov    ("eax","edx");
1630         &and    ("eax",&DWP($res_x+$i,"esp"));
1631         &mov    ("ebx","ebp");
1632         &and    ("ebx",&DWP($in2_x+$i,"esp"));
1633         &mov    ("ecx","esi");
1634         &and    ("ecx",&DWP($in1_x+$i,"esp"));
1635         &or     ("eax","ebx");
1636         &or     ("eax","ecx");
1637         &mov    (&DWP($i,"edi"),"eax");
1638     }
1639     for($i=0;$i<64;$i+=4) {
1640         &mov    ("eax","edx");
1641         &and    ("eax",&DWP($res_x+$i,"esp"));
1642         &mov    ("ebx","ebp");
1643         &and    ("ebx",&DWP($in2_x+$i,"esp"));
1644         &mov    ("ecx","esi");
1645         &and    ("ecx",&DWP($in1_x+$i,"esp"));
1646         &or     ("eax","ebx");
1647         &or     ("eax","ecx");
1648         &mov    (&DWP($i,"edi"),"eax");
1649     }
1650     &set_label("add_done");
1651         &stack_pop(8*18+5);
1652 } &function_end("ecp_nistz256_point_add");
1653
1654 ########################################################################
1655 # void ecp_nistz256_point_add_affine(P256_POINT *out,
1656 #                                    const P256_POINT *in1,
1657 #                                    const P256_POINT_AFFINE *in2);
1658 &function_begin("ecp_nistz256_point_add_affine");
1659 {
1660     my ($res_x,$res_y,$res_z,
1661         $in1_x,$in1_y,$in1_z,
1662         $in2_x,$in2_y,
1663         $U2,$S2,$H,$R,$Hsqr,$Hcub,$Rsqr)=map(32*$_,(0..14));
1664     my $Z1sqr = $S2;
1665     my @ONE_mont=(1,0,0,-1,-1,-1,-2,0);
1666
1667         &mov    ("esi",&wparam(1));
1668
1669         # above map() describes stack layout with 15 temporary
1670         # 256-bit vectors on top, then we take extra words for
1671         # !in1infty, !in2infty, and OPENSSL_ia32cap_P copy.
1672         &stack_push(8*15+3);
1673                                                 if ($sse2) {
1674         &call   ("_picup_eax");
1675     &set_label("pic");
1676         &picmeup("edx","OPENSSL_ia32cap_P","eax",&label("pic"));
1677         &mov    ("ebp",&DWP(0,"edx"));          }
1678
1679         &lea    ("edi",&DWP($in1_x,"esp"));
1680     for($i=0;$i<96;$i+=16) {
1681         &mov    ("eax",&DWP($i+0,"esi"));       # copy in1
1682         &mov    ("ebx",&DWP($i+4,"esi"));
1683         &mov    ("ecx",&DWP($i+8,"esi"));
1684         &mov    ("edx",&DWP($i+12,"esi"));
1685         &mov    (&DWP($i+0,"edi"),"eax");
1686         &mov    (&DWP(32*15+8,"esp"),"ebp")     if ($i==0);
1687         &mov    ("ebp","eax")                   if ($i==64);
1688         &or     ("ebp","eax")                   if ($i>64);
1689         &mov    (&DWP($i+4,"edi"),"ebx");
1690         &or     ("ebp","ebx")                   if ($i>=64);
1691         &mov    (&DWP($i+8,"edi"),"ecx");
1692         &or     ("ebp","ecx")                   if ($i>=64);
1693         &mov    (&DWP($i+12,"edi"),"edx");
1694         &or     ("ebp","edx")                   if ($i>=64);
1695     }
1696         &xor    ("eax","eax");
1697         &mov    ("esi",&wparam(2));
1698         &sub    ("eax","ebp");
1699         &or     ("ebp","eax");
1700         &sar    ("ebp",31);
1701         &mov    (&DWP(32*15+0,"esp"),"ebp");    # !in1infty
1702
1703         &lea    ("edi",&DWP($in2_x,"esp"));
1704     for($i=0;$i<64;$i+=16) {
1705         &mov    ("eax",&DWP($i+0,"esi"));       # copy in2
1706         &mov    ("ebx",&DWP($i+4,"esi"));
1707         &mov    ("ecx",&DWP($i+8,"esi"));
1708         &mov    ("edx",&DWP($i+12,"esi"));
1709         &mov    (&DWP($i+0,"edi"),"eax");
1710         &mov    ("ebp","eax")                   if ($i==0);
1711         &or     ("ebp","eax")                   if ($i!=0);
1712         &mov    (&DWP($i+4,"edi"),"ebx");
1713         &or     ("ebp","ebx");
1714         &mov    (&DWP($i+8,"edi"),"ecx");
1715         &or     ("ebp","ecx");
1716         &mov    (&DWP($i+12,"edi"),"edx");
1717         &or     ("ebp","edx");
1718     }
1719         &xor    ("ebx","ebx");
1720         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1721         &sub    ("ebx","ebp");
1722          &lea   ("esi",&DWP($in1_z,"esp"));
1723         &or     ("ebx","ebp");
1724          &lea   ("ebp",&DWP($in1_z,"esp"));
1725         &sar    ("ebx",31);
1726          &lea   ("edi",&DWP($Z1sqr,"esp"));
1727         &mov    (&DWP(32*15+4,"esp"),"ebx");    # !in2infty
1728
1729         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Z1sqr, in1_z);
1730
1731         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1732         &lea    ("esi",&DWP($in2_x,"esp"));
1733         &mov    ("ebp","edi");                  # %esi is stull &Z1sqr
1734         &lea    ("edi",&DWP($U2,"esp"));
1735         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, Z1sqr, in2_x);
1736
1737         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1738         &lea    ("esi",&DWP($in1_z,"esp"));
1739         &lea    ("ebp",&DWP($Z1sqr,"esp"));
1740         &lea    ("edi",&DWP($S2,"esp"));
1741         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, Z1sqr, in1_z);
1742
1743         &lea    ("esi",&DWP($U2,"esp"));
1744         &lea    ("ebp",&DWP($in1_x,"esp"));
1745         &lea    ("edi",&DWP($H,"esp"));
1746         &call   ("_ecp_nistz256_sub");          # p256_sub(H, U2, in1_x);
1747
1748         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1749         &lea    ("esi",&DWP($in2_y,"esp"));
1750         &lea    ("ebp",&DWP($S2,"esp"));
1751         &lea    ("edi",&DWP($S2,"esp"));
1752         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, S2, in2_y);
1753
1754         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1755         &lea    ("esi",&DWP($in1_z,"esp"));
1756         &lea    ("ebp",&DWP($H,"esp"));
1757         &lea    ("edi",&DWP($res_z,"esp"));
1758         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_z, H, in1_z);
1759
1760         &lea    ("esi",&DWP($S2,"esp"));
1761         &lea    ("ebp",&DWP($in1_y,"esp"));
1762         &lea    ("edi",&DWP($R,"esp"));
1763         &call   ("_ecp_nistz256_sub");          # p256_sub(R, S2, in1_y);
1764
1765         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1766         &lea    ("esi",&DWP($H,"esp"));
1767         &lea    ("ebp",&DWP($H,"esp"));
1768         &lea    ("edi",&DWP($Hsqr,"esp"));
1769         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Hsqr, H);
1770
1771         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1772         &lea    ("esi",&DWP($R,"esp"));
1773         &lea    ("ebp",&DWP($R,"esp"));
1774         &lea    ("edi",&DWP($Rsqr,"esp"));
1775         &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Rsqr, R);
1776
1777         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1778         &lea    ("esi",&DWP($in1_x,"esp"));
1779         &lea    ("ebp",&DWP($Hsqr,"esp"));
1780         &lea    ("edi",&DWP($U2,"esp"));
1781         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, in1_x, Hsqr);
1782
1783         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1784         &lea    ("esi",&DWP($H,"esp"));
1785         &lea    ("ebp",&DWP($Hsqr,"esp"));
1786         &lea    ("edi",&DWP($Hcub,"esp"));
1787         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(Hcub, Hsqr, H);
1788
1789         &lea    ("esi",&DWP($U2,"esp"));
1790         &lea    ("ebp",&DWP($U2,"esp"));
1791         &lea    ("edi",&DWP($Hsqr,"esp"));
1792         &call   ("_ecp_nistz256_add");          # p256_mul_by_2(Hsqr, U2);
1793
1794         &lea    ("esi",&DWP($Rsqr,"esp"));
1795         &lea    ("ebp",&DWP($Hsqr,"esp"));
1796         &lea    ("edi",&DWP($res_x,"esp"));
1797         &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, Rsqr, Hsqr);
1798
1799         &lea    ("esi",&DWP($res_x,"esp"));
1800         &lea    ("ebp",&DWP($Hcub,"esp"));
1801         &lea    ("edi",&DWP($res_x,"esp"));
1802         &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, res_x, Hcub);
1803
1804         &lea    ("esi",&DWP($U2,"esp"));
1805         &lea    ("ebp",&DWP($res_x,"esp"));
1806         &lea    ("edi",&DWP($res_y,"esp"));
1807         &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, U2, res_x);
1808
1809         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1810         &lea    ("esi",&DWP($Hcub,"esp"));
1811         &lea    ("ebp",&DWP($in1_y,"esp"));
1812         &lea    ("edi",&DWP($S2,"esp"));
1813         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, Hcub, in1_y);
1814
1815         &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
1816         &lea    ("esi",&DWP($R,"esp"));
1817         &lea    ("ebp",&DWP($res_y,"esp"));
1818         &lea    ("edi",&DWP($res_y,"esp"));
1819         &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_y, res_y, R);
1820
1821         &lea    ("esi",&DWP($res_y,"esp"));
1822         &lea    ("ebp",&DWP($S2,"esp"));
1823         &lea    ("edi",&DWP($res_y,"esp"));
1824         &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, res_y, S2);
1825
1826         &mov    ("ebp",&DWP(32*15+0,"esp"));    # !in1infty
1827         &mov    ("esi",&DWP(32*15+4,"esp"));    # !in2infty
1828         &mov    ("edi",&wparam(0));
1829         &mov    ("edx","ebp");
1830         &not    ("ebp");
1831         &and    ("edx","esi");
1832         &and    ("ebp","esi");
1833         &not    ("esi");
1834
1835         ########################################
1836         # conditional moves
1837     for($i=64;$i<96;$i+=4) {
1838         my $one=@ONE_mont[($i-64)/4];
1839
1840         &mov    ("eax","edx");
1841         &and    ("eax",&DWP($res_x+$i,"esp"));
1842         &mov    ("ebx","ebp")                   if ($one && $one!=-1);
1843         &and    ("ebx",$one)                    if ($one && $one!=-1);
1844         &mov    ("ecx","esi");
1845         &and    ("ecx",&DWP($in1_x+$i,"esp"));
1846         &or     ("eax",$one==-1?"ebp":"ebx")    if ($one);
1847         &or     ("eax","ecx");
1848         &mov    (&DWP($i,"edi"),"eax");
1849     }
1850     for($i=0;$i<64;$i+=4) {
1851         &mov    ("eax","edx");
1852         &and    ("eax",&DWP($res_x+$i,"esp"));
1853         &mov    ("ebx","ebp");
1854         &and    ("ebx",&DWP($in2_x+$i,"esp"));
1855         &mov    ("ecx","esi");
1856         &and    ("ecx",&DWP($in1_x+$i,"esp"));
1857         &or     ("eax","ebx");
1858         &or     ("eax","ecx");
1859         &mov    (&DWP($i,"edi"),"eax");
1860     }
1861         &stack_pop(8*15+3);
1862 } &function_end("ecp_nistz256_point_add_affine");
1863
1864 &asm_finish();
1865
1866 close STDOUT;