crypto/cast/asm/cast-586.pl

   1 #! /usr/bin/env perl
   2 # Copyright 1995-2020 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 # This flag makes the inner loop one cycle longer, but generates
  11 # code that runs %30 faster on the pentium pro/II, 44% faster
  12 # of PIII, while only %7 slower on the pentium.
  13 # By default, this flag is on.
  14 $ppro=1;
  15
  16 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  17 push(@INC,"${dir}","${dir}../../perlasm");
  18 require "x86asm.pl";
  19 require "cbc.pl";
  20
  21 $output=pop and open STDOUT,">$output";
  22
  23 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
  24
  25 $CAST_ROUNDS=16;
  26 $L="edi";
  27 $R="esi";
  28 $K="ebp";
  29 $tmp1="ecx";
  30 $tmp2="ebx";
  31 $tmp3="eax";
  32 $tmp4="edx";
  33 $S1="CAST_S_table0";
  34 $S2="CAST_S_table1";
  35 $S3="CAST_S_table2";
  36 $S4="CAST_S_table3";
  37
  38 @F1=("add","xor","sub");
  39 @F2=("xor","sub","add");
  40 @F3=("sub","add","xor");
  41
  42 &CAST_encrypt("CAST_encrypt",1);
  43 &CAST_encrypt("CAST_decrypt",0);
  44 &cbc("CAST_cbc_encrypt","CAST_encrypt","CAST_decrypt",1,4,5,3,-1,-1);
  45
  46 &asm_finish();
  47
  48 close STDOUT or die "error closing STDOUT: $!";
  49
  50 sub CAST_encrypt {
  51     local($name,$enc)=@_;
  52
  53     local($win_ex)=<<"EOF";
  54 EXTERN  _CAST_S_table0:DWORD
  55 EXTERN  _CAST_S_table1:DWORD
  56 EXTERN  _CAST_S_table2:DWORD
  57 EXTERN  _CAST_S_table3:DWORD
  58 EOF
  59     &main::external_label(
  60                           "CAST_S_table0",
  61                           "CAST_S_table1",
  62                           "CAST_S_table2",
  63                           "CAST_S_table3",
  64                           );
  65
  66     &function_begin_B($name,$win_ex);
  67
  68     &comment("");
  69
  70     &push("ebp");
  71     &push("ebx");
  72     &mov($tmp2,&wparam(0));
  73     &mov($K,&wparam(1));
  74     &push("esi");
  75     &push("edi");
  76
  77     &comment("Load the 2 words");
  78     &mov($L,&DWP(0,$tmp2,"",0));
  79     &mov($R,&DWP(4,$tmp2,"",0));
  80
  81     &comment('Get short key flag');
  82     &mov($tmp3,&DWP(128,$K,"",0));
  83     if($enc) {
  84         &push($tmp3);
  85     } else {
  86         &or($tmp3,$tmp3);
  87         &jnz(&label('cast_dec_skip'));
  88     }
  89
  90     &xor($tmp3, $tmp3);
  91
  92     # encrypting part
  93
  94     if ($enc) {
  95         &E_CAST( 0,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
  96         &E_CAST( 1,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
  97         &E_CAST( 2,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
  98         &E_CAST( 3,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
  99         &E_CAST( 4,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 100         &E_CAST( 5,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 101         &E_CAST( 6,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 102         &E_CAST( 7,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 103         &E_CAST( 8,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 104         &E_CAST( 9,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 105         &E_CAST(10,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 106         &E_CAST(11,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 107         &comment('test short key flag');
 108         &pop($tmp4);
 109         &or($tmp4,$tmp4);
 110         &jnz(&label('cast_enc_done'));
 111         &E_CAST(12,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 112         &E_CAST(13,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 113         &E_CAST(14,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 114         &E_CAST(15,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 115     } else {
 116         &E_CAST(15,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 117         &E_CAST(14,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 118         &E_CAST(13,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 119         &E_CAST(12,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 120         &set_label('cast_dec_skip');
 121         &E_CAST(11,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 122         &E_CAST(10,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 123         &E_CAST( 9,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 124         &E_CAST( 8,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 125         &E_CAST( 7,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 126         &E_CAST( 6,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 127         &E_CAST( 5,$S,$L,$R,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 128         &E_CAST( 4,$S,$R,$L,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 129         &E_CAST( 3,$S,$L,$R,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 130         &E_CAST( 2,$S,$R,$L,$K,@F3,$tmp1,$tmp2,$tmp3,$tmp4);
 131         &E_CAST( 1,$S,$L,$R,$K,@F2,$tmp1,$tmp2,$tmp3,$tmp4);
 132         &E_CAST( 0,$S,$R,$L,$K,@F1,$tmp1,$tmp2,$tmp3,$tmp4);
 133     }
 134
 135     &set_label('cast_enc_done') if $enc;
 136 # Why the nop? - Ben 17/1/99
 137     &nop();
 138     &mov($tmp3,&wparam(0));
 139     &mov(&DWP(4,$tmp3,"",0),$L);
 140     &mov(&DWP(0,$tmp3,"",0),$R);
 141     &function_end($name);
 142 }
 143
 144 sub E_CAST {
 145     local($i,$S,$L,$R,$K,$OP1,$OP2,$OP3,$tmp1,$tmp2,$tmp3,$tmp4)=@_;
 146     # Ri needs to have 16 pre added.
 147
 148     &comment("round $i");
 149     &mov(       $tmp4,          &DWP($i*8,$K,"",1));
 150
 151     &mov(       $tmp1,          &DWP($i*8+4,$K,"",1));
 152     &$OP1(      $tmp4,          $R);
 153
 154     &rotl(      $tmp4,          &LB($tmp1));
 155
 156     if ($ppro) {
 157         &xor(   $tmp1,          $tmp1);
 158         &mov(   $tmp2,          0xff);
 159
 160         &movb(  &LB($tmp1),     &HB($tmp4));    # A
 161         &and(   $tmp2,          $tmp4);
 162
 163         &shr(   $tmp4,          16);            #
 164         &xor(   $tmp3,          $tmp3);
 165     } else {
 166         &mov(   $tmp2,          $tmp4);         # B
 167         &movb(  &LB($tmp1),     &HB($tmp4));    # A     # BAD BAD BAD
 168
 169         &shr(   $tmp4,          16);            #
 170         &and(   $tmp2,          0xff);
 171     }
 172
 173     &movb(      &LB($tmp3),     &HB($tmp4));    # C     # BAD BAD BAD
 174     &and(       $tmp4,          0xff);          # D
 175
 176     &mov(       $tmp1,          &DWP($S1,"",$tmp1,4));
 177     &mov(       $tmp2,          &DWP($S2,"",$tmp2,4));
 178
 179     &$OP2(      $tmp1,          $tmp2);
 180     &mov(       $tmp2,          &DWP($S3,"",$tmp3,4));
 181
 182     &$OP3(      $tmp1,          $tmp2);
 183     &mov(       $tmp2,          &DWP($S4,"",$tmp4,4));
 184
 185     &$OP1(      $tmp1,          $tmp2);
 186     # XXX
 187
 188     &xor(       $L,             $tmp1);
 189     # XXX
 190 }
 191