crypto/whrlpool/asm/wp-mmx.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2005-2020 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the Apache License 2.0 (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. Rights for redistribution and usage in source and binary
  13 # forms are granted according to the License.
  14 # ====================================================================
  15 #
  16 # whirlpool_block_mmx implementation.
  17 #
  18 *SCALE=\(2); # 2 or 8, that is the question:-) Value of 8 results
  19 # in 16KB large table, which is tough on L1 cache, but eliminates
  20 # unaligned references to it. Value of 2 results in 4KB table, but
  21 # 7/8 of references to it are unaligned. AMD cores seem to be
  22 # allergic to the latter, while Intel ones - to former [see the
  23 # table]. I stick to value of 2 for two reasons: 1. smaller table
  24 # minimizes cache trashing and thus mitigates the hazard of side-
  25 # channel leakage similar to AES cache-timing one; 2. performance
  26 # gap among different µ-archs is smaller.
  27 #
  28 # Performance table lists rounded amounts of CPU cycles spent by
  29 # whirlpool_block_mmx routine on single 64 byte input block, i.e.
  30 # smaller is better and asymptotic throughput can be estimated by
  31 # multiplying 64 by CPU clock frequency and dividing by relevant
  32 # value from the given table:
  33 #
  34 #               $SCALE=2/8      icc8    gcc3
  35 # Intel P4      3200/4600       4600(*) 6400
  36 # Intel PIII    2900/3000       4900    5400
  37 # AMD K[78]     2500/1800       9900    8200(**)
  38 #
  39 # (*)   I've sketched even non-MMX assembler, but for the record
  40 #       I've failed to beat the Intel compiler on P4, without using
  41 #       MMX that is...
  42 # (**)  ... on AMD on the other hand non-MMX assembler was observed
  43 #       to perform significantly better, but I figured this MMX
  44 #       implementation is even faster anyway, so why bother? As for
  45 #       pre-MMX AMD core[s], the improvement coefficient is more
  46 #       than likely to vary anyway and I don't know how. But the
  47 #       least I know is that gcc-generated code compiled with
  48 #       -DL_ENDIAN and -DOPENSSL_SMALL_FOOTPRINT [see C module for
  49 #       details] and optimized for Pentium was observed to perform
  50 #       *better* on Pentium 100 than unrolled non-MMX assembler
  51 #       loop... So we just say that I don't know if maintaining
  52 #       non-MMX implementation would actually pay off, but till
  53 #       opposite is proved "unlikely" is assumed.
  54
  55 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  56 push(@INC,"${dir}","${dir}../../perlasm");
  57 require "x86asm.pl";
  58
  59 $output=pop and open STDOUT,">$output";
  60
  61 &asm_init($ARGV[0]);
  62
  63 sub L()  { &data_byte(@_); }
  64 sub LL()
  65 {       if      ($SCALE==2)     { &data_byte(@_); &data_byte(@_); }
  66         elsif   ($SCALE==8)     { for ($i=0;$i<8;$i++) {
  67                                         &data_byte(@_);
  68                                         unshift(@_,pop(@_));
  69                                   }
  70                                 }
  71         else                    { die "invalid SCALE value"; }
  72 }
  73
  74 sub scale()
  75 {       if      ($SCALE==2)     { &lea(@_[0],&DWP(0,@_[1],@_[1])); }
  76         elsif   ($SCALE==8)     { &lea(@_[0],&DWP(0,"",@_[1],8));  }
  77         else                    { die "invalid SCALE value";       }
  78 }
  79
  80 sub row()
  81 {       if      ($SCALE==2)     { ((8-shift)&7); }
  82         elsif   ($SCALE==8)     { (8*shift);     }
  83         else                    { die "invalid SCALE value"; }
  84 }
  85
  86 $tbl="ebp";
  87 @mm=("mm0","mm1","mm2","mm3","mm4","mm5","mm6","mm7");
  88
  89 &function_begin_B("whirlpool_block_mmx");
  90         &push   ("ebp");
  91         &push   ("ebx");
  92         &push   ("esi");
  93         &push   ("edi");
  94
  95         &mov    ("esi",&wparam(0));             # hash value
  96         &mov    ("edi",&wparam(1));             # input data stream
  97         &mov    ("ebp",&wparam(2));             # number of chunks in input
  98
  99         &mov    ("eax","esp");                  # copy stack pointer
 100         &sub    ("esp",128+20);                 # allocate frame
 101         &and    ("esp",-64);                    # align for cache-line
 102
 103         &lea    ("ebx",&DWP(128,"esp"));
 104         &mov    (&DWP(0,"ebx"),"esi");          # save parameter block
 105         &mov    (&DWP(4,"ebx"),"edi");
 106         &mov    (&DWP(8,"ebx"),"ebp");
 107         &mov    (&DWP(16,"ebx"),"eax");         # saved stack pointer
 108
 109         &call   (&label("pic_point"));
 110 &set_label("pic_point");
 111         &blindpop($tbl);
 112         &lea    ($tbl,&DWP(&label("table")."-".&label("pic_point"),$tbl));
 113
 114         &xor    ("ecx","ecx");
 115         &xor    ("edx","edx");
 116
 117         for($i=0;$i<8;$i++) { &movq(@mm[$i],&QWP($i*8,"esi")); }    # L=H
 118 &set_label("outerloop");
 119         for($i=0;$i<8;$i++) { &movq(&QWP($i*8,"esp"),@mm[$i]); }    # K=L
 120         for($i=0;$i<8;$i++) { &pxor(@mm[$i],&QWP($i*8,"edi")); }    # L^=inp
 121         for($i=0;$i<8;$i++) { &movq(&QWP(64+$i*8,"esp"),@mm[$i]); } # S=L
 122
 123         &xor    ("esi","esi");
 124         &mov    (&DWP(12,"ebx"),"esi");         # zero round counter
 125
 126 &set_label("round",16);
 127         &movq   (@mm[0],&QWP(2048*$SCALE,$tbl,"esi",8));        # rc[r]
 128         &mov    ("eax",&DWP(0,"esp"));
 129         &mov    ("ebx",&DWP(4,"esp"));
 130         &movz   ("ecx",&LB("eax"));
 131         &movz   ("edx",&HB("eax"));
 132 for($i=0;$i<8;$i++) {
 133     my $func = ($i==0)? \&movq : \&pxor;
 134         &shr    ("eax",16);
 135         &scale  ("esi","ecx");
 136         &movz   ("ecx",&LB("eax"));
 137         &scale  ("edi","edx");
 138         &movz   ("edx",&HB("eax"));
 139         &pxor   (@mm[0],&QWP(&row(0),$tbl,"esi",8));
 140         &$func  (@mm[1],&QWP(&row(1),$tbl,"edi",8));
 141         &mov    ("eax",&DWP(($i+1)*8,"esp"));
 142         &scale  ("esi","ecx");
 143         &movz   ("ecx",&LB("ebx"));
 144         &scale  ("edi","edx");
 145         &movz   ("edx",&HB("ebx"));
 146         &$func  (@mm[2],&QWP(&row(2),$tbl,"esi",8));
 147         &$func  (@mm[3],&QWP(&row(3),$tbl,"edi",8));
 148         &shr    ("ebx",16);
 149         &scale  ("esi","ecx");
 150         &movz   ("ecx",&LB("ebx"));
 151         &scale  ("edi","edx");
 152         &movz   ("edx",&HB("ebx"));
 153         &$func  (@mm[4],&QWP(&row(4),$tbl,"esi",8));
 154         &$func  (@mm[5],&QWP(&row(5),$tbl,"edi",8));
 155         &mov    ("ebx",&DWP(($i+1)*8+4,"esp"));
 156         &scale  ("esi","ecx");
 157         &movz   ("ecx",&LB("eax"));
 158         &scale  ("edi","edx");
 159         &movz   ("edx",&HB("eax"));
 160         &$func  (@mm[6],&QWP(&row(6),$tbl,"esi",8));
 161         &$func  (@mm[7],&QWP(&row(7),$tbl,"edi",8));
 162     push(@mm,shift(@mm));
 163 }
 164
 165         for($i=0;$i<8;$i++) { &movq(&QWP($i*8,"esp"),@mm[$i]); }    # K=L
 166
 167 for($i=0;$i<8;$i++) {
 168         &shr    ("eax",16);
 169         &scale  ("esi","ecx");
 170         &movz   ("ecx",&LB("eax"));
 171         &scale  ("edi","edx");
 172         &movz   ("edx",&HB("eax"));
 173         &pxor   (@mm[0],&QWP(&row(0),$tbl,"esi",8));
 174         &pxor   (@mm[1],&QWP(&row(1),$tbl,"edi",8));
 175         &mov    ("eax",&DWP(64+($i+1)*8,"esp"))         if ($i<7);
 176         &scale  ("esi","ecx");
 177         &movz   ("ecx",&LB("ebx"));
 178         &scale  ("edi","edx");
 179         &movz   ("edx",&HB("ebx"));
 180         &pxor   (@mm[2],&QWP(&row(2),$tbl,"esi",8));
 181         &pxor   (@mm[3],&QWP(&row(3),$tbl,"edi",8));
 182         &shr    ("ebx",16);
 183         &scale  ("esi","ecx");
 184         &movz   ("ecx",&LB("ebx"));
 185         &scale  ("edi","edx");
 186         &movz   ("edx",&HB("ebx"));
 187         &pxor   (@mm[4],&QWP(&row(4),$tbl,"esi",8));
 188         &pxor   (@mm[5],&QWP(&row(5),$tbl,"edi",8));
 189         &mov    ("ebx",&DWP(64+($i+1)*8+4,"esp"))       if ($i<7);
 190         &scale  ("esi","ecx");
 191         &movz   ("ecx",&LB("eax"));
 192         &scale  ("edi","edx");
 193         &movz   ("edx",&HB("eax"));
 194         &pxor   (@mm[6],&QWP(&row(6),$tbl,"esi",8));
 195         &pxor   (@mm[7],&QWP(&row(7),$tbl,"edi",8));
 196     push(@mm,shift(@mm));
 197 }
 198         &lea    ("ebx",&DWP(128,"esp"));
 199         &mov    ("esi",&DWP(12,"ebx"));         # pull round counter
 200         &add    ("esi",1);
 201         &cmp    ("esi",10);
 202         &je     (&label("roundsdone"));
 203
 204         &mov    (&DWP(12,"ebx"),"esi");         # update round counter
 205         for($i=0;$i<8;$i++) { &movq(&QWP(64+$i*8,"esp"),@mm[$i]); } # S=L
 206         &jmp    (&label("round"));
 207
 208 &set_label("roundsdone",16);
 209         &mov    ("esi",&DWP(0,"ebx"));          # reload argument block
 210         &mov    ("edi",&DWP(4,"ebx"));
 211         &mov    ("eax",&DWP(8,"ebx"));
 212
 213         for($i=0;$i<8;$i++) { &pxor(@mm[$i],&QWP($i*8,"edi")); }    # L^=inp
 214         for($i=0;$i<8;$i++) { &pxor(@mm[$i],&QWP($i*8,"esi")); }    # L^=H
 215         for($i=0;$i<8;$i++) { &movq(&QWP($i*8,"esi"),@mm[$i]); }    # H=L
 216
 217         &lea    ("edi",&DWP(64,"edi"));         # inp+=64
 218         &sub    ("eax",1);                      # num--
 219         &jz     (&label("alldone"));
 220         &mov    (&DWP(4,"ebx"),"edi");          # update argument block
 221         &mov    (&DWP(8,"ebx"),"eax");
 222         &jmp    (&label("outerloop"));
 223
 224 &set_label("alldone");
 225         &emms   ();
 226         &mov    ("esp",&DWP(16,"ebx"));         # restore saved stack pointer
 227         &pop    ("edi");
 228         &pop    ("esi");
 229         &pop    ("ebx");
 230         &pop    ("ebp");
 231         &ret    ();
 232
 233 &align(64);
 234 &set_label("table");
 235         &LL(0x18,0x18,0x60,0x18,0xc0,0x78,0x30,0xd8);
 236         &LL(0x23,0x23,0x8c,0x23,0x05,0xaf,0x46,0x26);
 237         &LL(0xc6,0xc6,0x3f,0xc6,0x7e,0xf9,0x91,0xb8);
 238         &LL(0xe8,0xe8,0x87,0xe8,0x13,0x6f,0xcd,0xfb);
 239         &LL(0x87,0x87,0x26,0x87,0x4c,0xa1,0x13,0xcb);
 240         &LL(0xb8,0xb8,0xda,0xb8,0xa9,0x62,0x6d,0x11);
 241         &LL(0x01,0x01,0x04,0x01,0x08,0x05,0x02,0x09);
 242         &LL(0x4f,0x4f,0x21,0x4f,0x42,0x6e,0x9e,0x0d);
 243         &LL(0x36,0x36,0xd8,0x36,0xad,0xee,0x6c,0x9b);
 244         &LL(0xa6,0xa6,0xa2,0xa6,0x59,0x04,0x51,0xff);
 245         &LL(0xd2,0xd2,0x6f,0xd2,0xde,0xbd,0xb9,0x0c);
 246         &LL(0xf5,0xf5,0xf3,0xf5,0xfb,0x06,0xf7,0x0e);
 247         &LL(0x79,0x79,0xf9,0x79,0xef,0x80,0xf2,0x96);
 248         &LL(0x6f,0x6f,0xa1,0x6f,0x5f,0xce,0xde,0x30);
 249         &LL(0x91,0x91,0x7e,0x91,0xfc,0xef,0x3f,0x6d);
 250         &LL(0x52,0x52,0x55,0x52,0xaa,0x07,0xa4,0xf8);
 251         &LL(0x60,0x60,0x9d,0x60,0x27,0xfd,0xc0,0x47);
 252         &LL(0xbc,0xbc,0xca,0xbc,0x89,0x76,0x65,0x35);
 253         &LL(0x9b,0x9b,0x56,0x9b,0xac,0xcd,0x2b,0x37);
 254         &LL(0x8e,0x8e,0x02,0x8e,0x04,0x8c,0x01,0x8a);
 255         &LL(0xa3,0xa3,0xb6,0xa3,0x71,0x15,0x5b,0xd2);
 256         &LL(0x0c,0x0c,0x30,0x0c,0x60,0x3c,0x18,0x6c);
 257         &LL(0x7b,0x7b,0xf1,0x7b,0xff,0x8a,0xf6,0x84);
 258         &LL(0x35,0x35,0xd4,0x35,0xb5,0xe1,0x6a,0x80);
 259         &LL(0x1d,0x1d,0x74,0x1d,0xe8,0x69,0x3a,0xf5);
 260         &LL(0xe0,0xe0,0xa7,0xe0,0x53,0x47,0xdd,0xb3);
 261         &LL(0xd7,0xd7,0x7b,0xd7,0xf6,0xac,0xb3,0x21);
 262         &LL(0xc2,0xc2,0x2f,0xc2,0x5e,0xed,0x99,0x9c);
 263         &LL(0x2e,0x2e,0xb8,0x2e,0x6d,0x96,0x5c,0x43);
 264         &LL(0x4b,0x4b,0x31,0x4b,0x62,0x7a,0x96,0x29);
 265         &LL(0xfe,0xfe,0xdf,0xfe,0xa3,0x21,0xe1,0x5d);
 266         &LL(0x57,0x57,0x41,0x57,0x82,0x16,0xae,0xd5);
 267         &LL(0x15,0x15,0x54,0x15,0xa8,0x41,0x2a,0xbd);
 268         &LL(0x77,0x77,0xc1,0x77,0x9f,0xb6,0xee,0xe8);
 269         &LL(0x37,0x37,0xdc,0x37,0xa5,0xeb,0x6e,0x92);
 270         &LL(0xe5,0xe5,0xb3,0xe5,0x7b,0x56,0xd7,0x9e);
 271         &LL(0x9f,0x9f,0x46,0x9f,0x8c,0xd9,0x23,0x13);
 272         &LL(0xf0,0xf0,0xe7,0xf0,0xd3,0x17,0xfd,0x23);
 273         &LL(0x4a,0x4a,0x35,0x4a,0x6a,0x7f,0x94,0x20);
 274         &LL(0xda,0xda,0x4f,0xda,0x9e,0x95,0xa9,0x44);
 275         &LL(0x58,0x58,0x7d,0x58,0xfa,0x25,0xb0,0xa2);
 276         &LL(0xc9,0xc9,0x03,0xc9,0x06,0xca,0x8f,0xcf);
 277         &LL(0x29,0x29,0xa4,0x29,0x55,0x8d,0x52,0x7c);
 278         &LL(0x0a,0x0a,0x28,0x0a,0x50,0x22,0x14,0x5a);
 279         &LL(0xb1,0xb1,0xfe,0xb1,0xe1,0x4f,0x7f,0x50);
 280         &LL(0xa0,0xa0,0xba,0xa0,0x69,0x1a,0x5d,0xc9);
 281         &LL(0x6b,0x6b,0xb1,0x6b,0x7f,0xda,0xd6,0x14);
 282         &LL(0x85,0x85,0x2e,0x85,0x5c,0xab,0x17,0xd9);
 283         &LL(0xbd,0xbd,0xce,0xbd,0x81,0x73,0x67,0x3c);
 284         &LL(0x5d,0x5d,0x69,0x5d,0xd2,0x34,0xba,0x8f);
 285         &LL(0x10,0x10,0x40,0x10,0x80,0x50,0x20,0x90);
 286         &LL(0xf4,0xf4,0xf7,0xf4,0xf3,0x03,0xf5,0x07);
 287         &LL(0xcb,0xcb,0x0b,0xcb,0x16,0xc0,0x8b,0xdd);
 288         &LL(0x3e,0x3e,0xf8,0x3e,0xed,0xc6,0x7c,0xd3);
 289         &LL(0x05,0x05,0x14,0x05,0x28,0x11,0x0a,0x2d);
 290         &LL(0x67,0x67,0x81,0x67,0x1f,0xe6,0xce,0x78);
 291         &LL(0xe4,0xe4,0xb7,0xe4,0x73,0x53,0xd5,0x97);
 292         &LL(0x27,0x27,0x9c,0x27,0x25,0xbb,0x4e,0x02);
 293         &LL(0x41,0x41,0x19,0x41,0x32,0x58,0x82,0x73);
 294         &LL(0x8b,0x8b,0x16,0x8b,0x2c,0x9d,0x0b,0xa7);
 295         &LL(0xa7,0xa7,0xa6,0xa7,0x51,0x01,0x53,0xf6);
 296         &LL(0x7d,0x7d,0xe9,0x7d,0xcf,0x94,0xfa,0xb2);
 297         &LL(0x95,0x95,0x6e,0x95,0xdc,0xfb,0x37,0x49);
 298         &LL(0xd8,0xd8,0x47,0xd8,0x8e,0x9f,0xad,0x56);
 299         &LL(0xfb,0xfb,0xcb,0xfb,0x8b,0x30,0xeb,0x70);
 300         &LL(0xee,0xee,0x9f,0xee,0x23,0x71,0xc1,0xcd);
 301         &LL(0x7c,0x7c,0xed,0x7c,0xc7,0x91,0xf8,0xbb);
 302         &LL(0x66,0x66,0x85,0x66,0x17,0xe3,0xcc,0x71);
 303         &LL(0xdd,0xdd,0x53,0xdd,0xa6,0x8e,0xa7,0x7b);
 304         &LL(0x17,0x17,0x5c,0x17,0xb8,0x4b,0x2e,0xaf);
 305         &LL(0x47,0x47,0x01,0x47,0x02,0x46,0x8e,0x45);
 306         &LL(0x9e,0x9e,0x42,0x9e,0x84,0xdc,0x21,0x1a);
 307         &LL(0xca,0xca,0x0f,0xca,0x1e,0xc5,0x89,0xd4);
 308         &LL(0x2d,0x2d,0xb4,0x2d,0x75,0x99,0x5a,0x58);
 309         &LL(0xbf,0xbf,0xc6,0xbf,0x91,0x79,0x63,0x2e);
 310         &LL(0x07,0x07,0x1c,0x07,0x38,0x1b,0x0e,0x3f);
 311         &LL(0xad,0xad,0x8e,0xad,0x01,0x23,0x47,0xac);
 312         &LL(0x5a,0x5a,0x75,0x5a,0xea,0x2f,0xb4,0xb0);
 313         &LL(0x83,0x83,0x36,0x83,0x6c,0xb5,0x1b,0xef);
 314         &LL(0x33,0x33,0xcc,0x33,0x85,0xff,0x66,0xb6);
 315         &LL(0x63,0x63,0x91,0x63,0x3f,0xf2,0xc6,0x5c);
 316         &LL(0x02,0x02,0x08,0x02,0x10,0x0a,0x04,0x12);
 317         &LL(0xaa,0xaa,0x92,0xaa,0x39,0x38,0x49,0x93);
 318         &LL(0x71,0x71,0xd9,0x71,0xaf,0xa8,0xe2,0xde);
 319         &LL(0xc8,0xc8,0x07,0xc8,0x0e,0xcf,0x8d,0xc6);
 320         &LL(0x19,0x19,0x64,0x19,0xc8,0x7d,0x32,0xd1);
 321         &LL(0x49,0x49,0x39,0x49,0x72,0x70,0x92,0x3b);
 322         &LL(0xd9,0xd9,0x43,0xd9,0x86,0x9a,0xaf,0x5f);
 323         &LL(0xf2,0xf2,0xef,0xf2,0xc3,0x1d,0xf9,0x31);
 324         &LL(0xe3,0xe3,0xab,0xe3,0x4b,0x48,0xdb,0xa8);
 325         &LL(0x5b,0x5b,0x71,0x5b,0xe2,0x2a,0xb6,0xb9);
 326         &LL(0x88,0x88,0x1a,0x88,0x34,0x92,0x0d,0xbc);
 327         &LL(0x9a,0x9a,0x52,0x9a,0xa4,0xc8,0x29,0x3e);
 328         &LL(0x26,0x26,0x98,0x26,0x2d,0xbe,0x4c,0x0b);
 329         &LL(0x32,0x32,0xc8,0x32,0x8d,0xfa,0x64,0xbf);
 330         &LL(0xb0,0xb0,0xfa,0xb0,0xe9,0x4a,0x7d,0x59);
 331         &LL(0xe9,0xe9,0x83,0xe9,0x1b,0x6a,0xcf,0xf2);
 332         &LL(0x0f,0x0f,0x3c,0x0f,0x78,0x33,0x1e,0x77);
 333         &LL(0xd5,0xd5,0x73,0xd5,0xe6,0xa6,0xb7,0x33);
 334         &LL(0x80,0x80,0x3a,0x80,0x74,0xba,0x1d,0xf4);
 335         &LL(0xbe,0xbe,0xc2,0xbe,0x99,0x7c,0x61,0x27);
 336         &LL(0xcd,0xcd,0x13,0xcd,0x26,0xde,0x87,0xeb);
 337         &LL(0x34,0x34,0xd0,0x34,0xbd,0xe4,0x68,0x89);
 338         &LL(0x48,0x48,0x3d,0x48,0x7a,0x75,0x90,0x32);
 339         &LL(0xff,0xff,0xdb,0xff,0xab,0x24,0xe3,0x54);
 340         &LL(0x7a,0x7a,0xf5,0x7a,0xf7,0x8f,0xf4,0x8d);
 341         &LL(0x90,0x90,0x7a,0x90,0xf4,0xea,0x3d,0x64);
 342         &LL(0x5f,0x5f,0x61,0x5f,0xc2,0x3e,0xbe,0x9d);
 343         &LL(0x20,0x20,0x80,0x20,0x1d,0xa0,0x40,0x3d);
 344         &LL(0x68,0x68,0xbd,0x68,0x67,0xd5,0xd0,0x0f);
 345         &LL(0x1a,0x1a,0x68,0x1a,0xd0,0x72,0x34,0xca);
 346         &LL(0xae,0xae,0x82,0xae,0x19,0x2c,0x41,0xb7);
 347         &LL(0xb4,0xb4,0xea,0xb4,0xc9,0x5e,0x75,0x7d);
 348         &LL(0x54,0x54,0x4d,0x54,0x9a,0x19,0xa8,0xce);
 349         &LL(0x93,0x93,0x76,0x93,0xec,0xe5,0x3b,0x7f);
 350         &LL(0x22,0x22,0x88,0x22,0x0d,0xaa,0x44,0x2f);
 351         &LL(0x64,0x64,0x8d,0x64,0x07,0xe9,0xc8,0x63);
 352         &LL(0xf1,0xf1,0xe3,0xf1,0xdb,0x12,0xff,0x2a);
 353         &LL(0x73,0x73,0xd1,0x73,0xbf,0xa2,0xe6,0xcc);
 354         &LL(0x12,0x12,0x48,0x12,0x90,0x5a,0x24,0x82);
 355         &LL(0x40,0x40,0x1d,0x40,0x3a,0x5d,0x80,0x7a);
 356         &LL(0x08,0x08,0x20,0x08,0x40,0x28,0x10,0x48);
 357         &LL(0xc3,0xc3,0x2b,0xc3,0x56,0xe8,0x9b,0x95);
 358         &LL(0xec,0xec,0x97,0xec,0x33,0x7b,0xc5,0xdf);
 359         &LL(0xdb,0xdb,0x4b,0xdb,0x96,0x90,0xab,0x4d);
 360         &LL(0xa1,0xa1,0xbe,0xa1,0x61,0x1f,0x5f,0xc0);
 361         &LL(0x8d,0x8d,0x0e,0x8d,0x1c,0x83,0x07,0x91);
 362         &LL(0x3d,0x3d,0xf4,0x3d,0xf5,0xc9,0x7a,0xc8);
 363         &LL(0x97,0x97,0x66,0x97,0xcc,0xf1,0x33,0x5b);
 364         &LL(0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
 365         &LL(0xcf,0xcf,0x1b,0xcf,0x36,0xd4,0x83,0xf9);
 366         &LL(0x2b,0x2b,0xac,0x2b,0x45,0x87,0x56,0x6e);
 367         &LL(0x76,0x76,0xc5,0x76,0x97,0xb3,0xec,0xe1);
 368         &LL(0x82,0x82,0x32,0x82,0x64,0xb0,0x19,0xe6);
 369         &LL(0xd6,0xd6,0x7f,0xd6,0xfe,0xa9,0xb1,0x28);
 370         &LL(0x1b,0x1b,0x6c,0x1b,0xd8,0x77,0x36,0xc3);
 371         &LL(0xb5,0xb5,0xee,0xb5,0xc1,0x5b,0x77,0x74);
 372         &LL(0xaf,0xaf,0x86,0xaf,0x11,0x29,0x43,0xbe);
 373         &LL(0x6a,0x6a,0xb5,0x6a,0x77,0xdf,0xd4,0x1d);
 374         &LL(0x50,0x50,0x5d,0x50,0xba,0x0d,0xa0,0xea);
 375         &LL(0x45,0x45,0x09,0x45,0x12,0x4c,0x8a,0x57);
 376         &LL(0xf3,0xf3,0xeb,0xf3,0xcb,0x18,0xfb,0x38);
 377         &LL(0x30,0x30,0xc0,0x30,0x9d,0xf0,0x60,0xad);
 378         &LL(0xef,0xef,0x9b,0xef,0x2b,0x74,0xc3,0xc4);
 379         &LL(0x3f,0x3f,0xfc,0x3f,0xe5,0xc3,0x7e,0xda);
 380         &LL(0x55,0x55,0x49,0x55,0x92,0x1c,0xaa,0xc7);
 381         &LL(0xa2,0xa2,0xb2,0xa2,0x79,0x10,0x59,0xdb);
 382         &LL(0xea,0xea,0x8f,0xea,0x03,0x65,0xc9,0xe9);
 383         &LL(0x65,0x65,0x89,0x65,0x0f,0xec,0xca,0x6a);
 384         &LL(0xba,0xba,0xd2,0xba,0xb9,0x68,0x69,0x03);
 385         &LL(0x2f,0x2f,0xbc,0x2f,0x65,0x93,0x5e,0x4a);
 386         &LL(0xc0,0xc0,0x27,0xc0,0x4e,0xe7,0x9d,0x8e);
 387         &LL(0xde,0xde,0x5f,0xde,0xbe,0x81,0xa1,0x60);
 388         &LL(0x1c,0x1c,0x70,0x1c,0xe0,0x6c,0x38,0xfc);
 389         &LL(0xfd,0xfd,0xd3,0xfd,0xbb,0x2e,0xe7,0x46);
 390         &LL(0x4d,0x4d,0x29,0x4d,0x52,0x64,0x9a,0x1f);
 391         &LL(0x92,0x92,0x72,0x92,0xe4,0xe0,0x39,0x76);
 392         &LL(0x75,0x75,0xc9,0x75,0x8f,0xbc,0xea,0xfa);
 393         &LL(0x06,0x06,0x18,0x06,0x30,0x1e,0x0c,0x36);
 394         &LL(0x8a,0x8a,0x12,0x8a,0x24,0x98,0x09,0xae);
 395         &LL(0xb2,0xb2,0xf2,0xb2,0xf9,0x40,0x79,0x4b);
 396         &LL(0xe6,0xe6,0xbf,0xe6,0x63,0x59,0xd1,0x85);
 397         &LL(0x0e,0x0e,0x38,0x0e,0x70,0x36,0x1c,0x7e);
 398         &LL(0x1f,0x1f,0x7c,0x1f,0xf8,0x63,0x3e,0xe7);
 399         &LL(0x62,0x62,0x95,0x62,0x37,0xf7,0xc4,0x55);
 400         &LL(0xd4,0xd4,0x77,0xd4,0xee,0xa3,0xb5,0x3a);
 401         &LL(0xa8,0xa8,0x9a,0xa8,0x29,0x32,0x4d,0x81);
 402         &LL(0x96,0x96,0x62,0x96,0xc4,0xf4,0x31,0x52);
 403         &LL(0xf9,0xf9,0xc3,0xf9,0x9b,0x3a,0xef,0x62);
 404         &LL(0xc5,0xc5,0x33,0xc5,0x66,0xf6,0x97,0xa3);
 405         &LL(0x25,0x25,0x94,0x25,0x35,0xb1,0x4a,0x10);
 406         &LL(0x59,0x59,0x79,0x59,0xf2,0x20,0xb2,0xab);
 407         &LL(0x84,0x84,0x2a,0x84,0x54,0xae,0x15,0xd0);
 408         &LL(0x72,0x72,0xd5,0x72,0xb7,0xa7,0xe4,0xc5);
 409         &LL(0x39,0x39,0xe4,0x39,0xd5,0xdd,0x72,0xec);
 410         &LL(0x4c,0x4c,0x2d,0x4c,0x5a,0x61,0x98,0x16);
 411         &LL(0x5e,0x5e,0x65,0x5e,0xca,0x3b,0xbc,0x94);
 412         &LL(0x78,0x78,0xfd,0x78,0xe7,0x85,0xf0,0x9f);
 413         &LL(0x38,0x38,0xe0,0x38,0xdd,0xd8,0x70,0xe5);
 414         &LL(0x8c,0x8c,0x0a,0x8c,0x14,0x86,0x05,0x98);
 415         &LL(0xd1,0xd1,0x63,0xd1,0xc6,0xb2,0xbf,0x17);
 416         &LL(0xa5,0xa5,0xae,0xa5,0x41,0x0b,0x57,0xe4);
 417         &LL(0xe2,0xe2,0xaf,0xe2,0x43,0x4d,0xd9,0xa1);
 418         &LL(0x61,0x61,0x99,0x61,0x2f,0xf8,0xc2,0x4e);
 419         &LL(0xb3,0xb3,0xf6,0xb3,0xf1,0x45,0x7b,0x42);
 420         &LL(0x21,0x21,0x84,0x21,0x15,0xa5,0x42,0x34);
 421         &LL(0x9c,0x9c,0x4a,0x9c,0x94,0xd6,0x25,0x08);
 422         &LL(0x1e,0x1e,0x78,0x1e,0xf0,0x66,0x3c,0xee);
 423         &LL(0x43,0x43,0x11,0x43,0x22,0x52,0x86,0x61);
 424         &LL(0xc7,0xc7,0x3b,0xc7,0x76,0xfc,0x93,0xb1);
 425         &LL(0xfc,0xfc,0xd7,0xfc,0xb3,0x2b,0xe5,0x4f);
 426         &LL(0x04,0x04,0x10,0x04,0x20,0x14,0x08,0x24);
 427         &LL(0x51,0x51,0x59,0x51,0xb2,0x08,0xa2,0xe3);
 428         &LL(0x99,0x99,0x5e,0x99,0xbc,0xc7,0x2f,0x25);
 429         &LL(0x6d,0x6d,0xa9,0x6d,0x4f,0xc4,0xda,0x22);
 430         &LL(0x0d,0x0d,0x34,0x0d,0x68,0x39,0x1a,0x65);
 431         &LL(0xfa,0xfa,0xcf,0xfa,0x83,0x35,0xe9,0x79);
 432         &LL(0xdf,0xdf,0x5b,0xdf,0xb6,0x84,0xa3,0x69);
 433         &LL(0x7e,0x7e,0xe5,0x7e,0xd7,0x9b,0xfc,0xa9);
 434         &LL(0x24,0x24,0x90,0x24,0x3d,0xb4,0x48,0x19);
 435         &LL(0x3b,0x3b,0xec,0x3b,0xc5,0xd7,0x76,0xfe);
 436         &LL(0xab,0xab,0x96,0xab,0x31,0x3d,0x4b,0x9a);
 437         &LL(0xce,0xce,0x1f,0xce,0x3e,0xd1,0x81,0xf0);
 438         &LL(0x11,0x11,0x44,0x11,0x88,0x55,0x22,0x99);
 439         &LL(0x8f,0x8f,0x06,0x8f,0x0c,0x89,0x03,0x83);
 440         &LL(0x4e,0x4e,0x25,0x4e,0x4a,0x6b,0x9c,0x04);
 441         &LL(0xb7,0xb7,0xe6,0xb7,0xd1,0x51,0x73,0x66);
 442         &LL(0xeb,0xeb,0x8b,0xeb,0x0b,0x60,0xcb,0xe0);
 443         &LL(0x3c,0x3c,0xf0,0x3c,0xfd,0xcc,0x78,0xc1);
 444         &LL(0x81,0x81,0x3e,0x81,0x7c,0xbf,0x1f,0xfd);
 445         &LL(0x94,0x94,0x6a,0x94,0xd4,0xfe,0x35,0x40);
 446         &LL(0xf7,0xf7,0xfb,0xf7,0xeb,0x0c,0xf3,0x1c);
 447         &LL(0xb9,0xb9,0xde,0xb9,0xa1,0x67,0x6f,0x18);
 448         &LL(0x13,0x13,0x4c,0x13,0x98,0x5f,0x26,0x8b);
 449         &LL(0x2c,0x2c,0xb0,0x2c,0x7d,0x9c,0x58,0x51);
 450         &LL(0xd3,0xd3,0x6b,0xd3,0xd6,0xb8,0xbb,0x05);
 451         &LL(0xe7,0xe7,0xbb,0xe7,0x6b,0x5c,0xd3,0x8c);
 452         &LL(0x6e,0x6e,0xa5,0x6e,0x57,0xcb,0xdc,0x39);
 453         &LL(0xc4,0xc4,0x37,0xc4,0x6e,0xf3,0x95,0xaa);
 454         &LL(0x03,0x03,0x0c,0x03,0x18,0x0f,0x06,0x1b);
 455         &LL(0x56,0x56,0x45,0x56,0x8a,0x13,0xac,0xdc);
 456         &LL(0x44,0x44,0x0d,0x44,0x1a,0x49,0x88,0x5e);
 457         &LL(0x7f,0x7f,0xe1,0x7f,0xdf,0x9e,0xfe,0xa0);
 458         &LL(0xa9,0xa9,0x9e,0xa9,0x21,0x37,0x4f,0x88);
 459         &LL(0x2a,0x2a,0xa8,0x2a,0x4d,0x82,0x54,0x67);
 460         &LL(0xbb,0xbb,0xd6,0xbb,0xb1,0x6d,0x6b,0x0a);
 461         &LL(0xc1,0xc1,0x23,0xc1,0x46,0xe2,0x9f,0x87);
 462         &LL(0x53,0x53,0x51,0x53,0xa2,0x02,0xa6,0xf1);
 463         &LL(0xdc,0xdc,0x57,0xdc,0xae,0x8b,0xa5,0x72);
 464         &LL(0x0b,0x0b,0x2c,0x0b,0x58,0x27,0x16,0x53);
 465         &LL(0x9d,0x9d,0x4e,0x9d,0x9c,0xd3,0x27,0x01);
 466         &LL(0x6c,0x6c,0xad,0x6c,0x47,0xc1,0xd8,0x2b);
 467         &LL(0x31,0x31,0xc4,0x31,0x95,0xf5,0x62,0xa4);
 468         &LL(0x74,0x74,0xcd,0x74,0x87,0xb9,0xe8,0xf3);
 469         &LL(0xf6,0xf6,0xff,0xf6,0xe3,0x09,0xf1,0x15);
 470         &LL(0x46,0x46,0x05,0x46,0x0a,0x43,0x8c,0x4c);
 471         &LL(0xac,0xac,0x8a,0xac,0x09,0x26,0x45,0xa5);
 472         &LL(0x89,0x89,0x1e,0x89,0x3c,0x97,0x0f,0xb5);
 473         &LL(0x14,0x14,0x50,0x14,0xa0,0x44,0x28,0xb4);
 474         &LL(0xe1,0xe1,0xa3,0xe1,0x5b,0x42,0xdf,0xba);
 475         &LL(0x16,0x16,0x58,0x16,0xb0,0x4e,0x2c,0xa6);
 476         &LL(0x3a,0x3a,0xe8,0x3a,0xcd,0xd2,0x74,0xf7);
 477         &LL(0x69,0x69,0xb9,0x69,0x6f,0xd0,0xd2,0x06);
 478         &LL(0x09,0x09,0x24,0x09,0x48,0x2d,0x12,0x41);
 479         &LL(0x70,0x70,0xdd,0x70,0xa7,0xad,0xe0,0xd7);
 480         &LL(0xb6,0xb6,0xe2,0xb6,0xd9,0x54,0x71,0x6f);
 481         &LL(0xd0,0xd0,0x67,0xd0,0xce,0xb7,0xbd,0x1e);
 482         &LL(0xed,0xed,0x93,0xed,0x3b,0x7e,0xc7,0xd6);
 483         &LL(0xcc,0xcc,0x17,0xcc,0x2e,0xdb,0x85,0xe2);
 484         &LL(0x42,0x42,0x15,0x42,0x2a,0x57,0x84,0x68);
 485         &LL(0x98,0x98,0x5a,0x98,0xb4,0xc2,0x2d,0x2c);
 486         &LL(0xa4,0xa4,0xaa,0xa4,0x49,0x0e,0x55,0xed);
 487         &LL(0x28,0x28,0xa0,0x28,0x5d,0x88,0x50,0x75);
 488         &LL(0x5c,0x5c,0x6d,0x5c,0xda,0x31,0xb8,0x86);
 489         &LL(0xf8,0xf8,0xc7,0xf8,0x93,0x3f,0xed,0x6b);
 490         &LL(0x86,0x86,0x22,0x86,0x44,0xa4,0x11,0xc2);
 491
 492         &L(0x18,0x23,0xc6,0xe8,0x87,0xb8,0x01,0x4f);    # rc[ROUNDS]
 493         &L(0x36,0xa6,0xd2,0xf5,0x79,0x6f,0x91,0x52);
 494         &L(0x60,0xbc,0x9b,0x8e,0xa3,0x0c,0x7b,0x35);
 495         &L(0x1d,0xe0,0xd7,0xc2,0x2e,0x4b,0xfe,0x57);
 496         &L(0x15,0x77,0x37,0xe5,0x9f,0xf0,0x4a,0xda);
 497         &L(0x58,0xc9,0x29,0x0a,0xb1,0xa0,0x6b,0x85);
 498         &L(0xbd,0x5d,0x10,0xf4,0xcb,0x3e,0x05,0x67);
 499         &L(0xe4,0x27,0x41,0x8b,0xa7,0x7d,0x95,0xd8);
 500         &L(0xfb,0xee,0x7c,0x66,0xdd,0x17,0x47,0x9e);
 501         &L(0xca,0x2d,0xbf,0x07,0xad,0x5a,0x83,0x33);
 502
 503 &function_end_B("whirlpool_block_mmx");
 504 &asm_finish();
 505
 506 close STDOUT or die "error closing STDOUT: $!";