crypto/rc4/asm/rc4-586.pl

   1 #!/usr/local/bin/perl
   2
   3 # At some point it became apparent that the original SSLeay RC4
   4 # assembler implementation performs suboptimal on latest IA-32
   5 # microarchitectures. After re-tuning performance has changed as
   6 # following:
   7 #
   8 # Pentium       +0%
   9 # Pentium III   +17%
  10 # AMD           +52%(*)
  11 # P4            +180%(**)
  12 #
  13 # (*)   This number is actually a trade-off:-) It's possible to
  14 #       achieve +72%, but at the cost of -48% off PIII performance.
  15 #       In other words code performing further 13% faster on AMD
  16 #       would perform almost 2 times slower on Intel PIII...
  17 #       For reference! This code delivers ~80% of rc4-amd64.pl
  18 #       performance on same Opteron machine.
  19 # (**)  This number requires compressed key schedule set up by
  20 #       RC4_set_key, see commentary section in rc4_skey.c for
  21 #       further details.
  22 #                                       <appro@fy.chalmers.se>
  23
  24 push(@INC,"perlasm","../../perlasm");
  25 require "x86asm.pl";
  26
  27 &asm_init($ARGV[0],"rc4-586.pl");
  28
  29 $x="eax";
  30 $y="ebx";
  31 $tx="ecx";
  32 $ty="edx";
  33 $in="esi";
  34 $out="edi";
  35 $d="ebp";
  36
  37 &RC4("RC4");
  38
  39 &asm_finish();
  40
  41 sub RC4_loop
  42         {
  43         local($n,$p,$char)=@_;
  44
  45         &comment("Round $n");
  46
  47         if ($char)
  48                 {
  49                 if ($p >= 0)
  50                         {
  51                          &mov($ty,      &swtmp(2));
  52                         &cmp($ty,       $in);
  53                          &jbe(&label("finished"));
  54                         &inc($in);
  55                         }
  56                 else
  57                         {
  58                         &add($ty,       8);
  59                          &inc($in);
  60                         &cmp($ty,       $in);
  61                          &jb(&label("finished"));
  62                         &mov(&swtmp(2), $ty);
  63                         }
  64                 }
  65         # Moved out
  66         # &mov( $tx,            &DWP(0,$d,$x,4)) if $p < 0;
  67
  68         &add(   &LB($y),        &LB($tx));
  69          &inc(  &LB($x));                       # NEXT ROUND
  70         &mov(   $ty,            &DWP(0,$d,$y,4));
  71          # XXX
  72         &mov(   &DWP(-4,$d,$x,4),$ty);                  # AGI
  73          &add(  $ty,            $tx);
  74         &mov(   &DWP(0,$d,$y,4),$tx);
  75          &and(  $ty,            0xff);
  76         &mov(   $tx,            &DWP(0,$d,$x,4)) if $p < 1; # NEXT ROUND
  77          &mov(  $ty,            &DWP(0,$d,$ty,4));
  78
  79         if (!$char)
  80                 {
  81                 #moved up into last round
  82                 if ($p >= 1)
  83                         {
  84                         &add(   $out,   8)
  85                         }
  86                 &movb(  &BP($n,"esp","",0),     &LB($ty));
  87                 }
  88         else
  89                 {
  90                 # Note in+=8 has occured
  91                 &movb(  &HB($ty),       &BP(-1,$in,"",0));
  92                  # XXX
  93                 &xorb(&LB($ty),         &HB($ty));
  94                  # XXX
  95                 &movb(&BP($n,$out,"",0),&LB($ty));
  96                 }
  97         }
  98
  99
 100 sub RC4
 101         {
 102         local($name)=@_;
 103
 104         &function_begin_B($name,"");
 105
 106         &mov($ty,&wparam(1));           # len
 107         &cmp($ty,0);
 108         &jne(&label("proceed"));
 109         &ret();
 110         &set_label("proceed");
 111
 112         &comment("");
 113
 114         &push("ebp");
 115          &push("ebx");
 116         &push("esi");
 117          &xor(  $x,     $x);            # avoid partial register stalls
 118         &push("edi");
 119          &xor(  $y,     $y);            # avoid partial register stalls
 120         &mov(   $d,     &wparam(0));    # key
 121          &mov(  $in,    &wparam(2));
 122
 123         &movb(  &LB($x),        &BP(0,$d,"",1));
 124          &movb( &LB($y),        &BP(4,$d,"",1));
 125
 126         &mov(   $out,   &wparam(3));
 127          &inc(  &LB($x));
 128
 129         &stack_push(3); # 3 temp variables
 130          &add(  $d,     8);
 131
 132         # detect compressed schedule, see commentary section in rc4_skey.c...
 133         &cmp(&DWP(256,$d),-1);
 134         &je(&label("RC4_CHAR"));
 135
 136          &lea(  $ty,    &DWP(-8,$ty,$in));
 137
 138         # check for 0 length input
 139
 140          &mov(  &swtmp(2),      $ty);   # this is now address to exit at
 141         &mov(   $tx,    &DWP(0,$d,$x,4));
 142
 143          &cmp(  $ty,    $in);
 144         &jb(    &label("end")); # less than 8 bytes
 145
 146         &set_label("start");
 147
 148         # filling DELAY SLOT
 149         &add(   $in,    8);
 150
 151         &RC4_loop(0,-1,0);
 152         &RC4_loop(1,0,0);
 153         &RC4_loop(2,0,0);
 154         &RC4_loop(3,0,0);
 155         &RC4_loop(4,0,0);
 156         &RC4_loop(5,0,0);
 157         &RC4_loop(6,0,0);
 158         &RC4_loop(7,1,0);
 159
 160         &comment("apply the cipher text");
 161         # xor the cipher data with input
 162
 163         #&add(  $out,   8); #moved up into last round
 164
 165         &mov(   $tx,    &swtmp(0));
 166          &mov(  $ty,    &DWP(-8,$in,"",0));
 167         &xor(   $tx,    $ty);
 168          &mov(  $ty,    &DWP(-4,$in,"",0));
 169         &mov(   &DWP(-8,$out,"",0),     $tx);
 170          &mov(  $tx,    &swtmp(1));
 171         &xor(   $tx,    $ty);
 172          &mov(  $ty,    &swtmp(2));     # load end ptr;
 173         &mov(   &DWP(-4,$out,"",0),     $tx);
 174          &mov(  $tx,            &DWP(0,$d,$x,4));
 175         &cmp($in,       $ty);
 176          &jbe(&label("start"));
 177
 178         &set_label("end");
 179
 180         # There is quite a bit of extra crap in RC4_loop() for this
 181         # first round
 182         &RC4_loop(0,-1,1);
 183         &RC4_loop(1,0,1);
 184         &RC4_loop(2,0,1);
 185         &RC4_loop(3,0,1);
 186         &RC4_loop(4,0,1);
 187         &RC4_loop(5,0,1);
 188         &RC4_loop(6,1,1);
 189
 190         &jmp(&label("finished"));
 191
 192         &align(16);
 193         # this is essentially Intel P4 specific codepath, see rc4_skey.c...
 194         &set_label("RC4_CHAR");
 195
 196         &lea    ($ty,&DWP(0,$in,$ty));
 197         &mov    (&swtmp(2),$ty);
 198
 199         # strangely enough unrolled loop performs over 20% slower...
 200         &set_label("RC4_CHAR_loop");
 201                 &movz   ($tx,&BP(0,$d,$x));
 202                 &add    (&LB($y),&LB($tx));
 203                 &movz   ($ty,&BP(0,$d,$y));
 204                 &movb   (&BP(0,$d,$y),&LB($tx));
 205                 &movb   (&BP(0,$d,$x),&LB($ty));
 206                 &add    (&LB($ty),&LB($tx));
 207                 &movz   ($ty,&BP(0,$d,$ty));
 208                 &xorb   (&LB($ty),&BP(0,$in));
 209                 &movb   (&BP(0,$out),&LB($ty));
 210                 &inc    (&LB($x));
 211                 &inc    ($in);
 212                 &inc    ($out);
 213                 &cmp    ($in,&swtmp(2));
 214         &jb     (&label("RC4_CHAR_loop"));
 215
 216         &set_label("finished");
 217         &dec(   $x);
 218          &stack_pop(3);
 219         &movb(  &BP(-4,$d,"",0),&LB($y));
 220          &movb( &BP(-8,$d,"",0),&LB($x));
 221
 222         &function_end($name);
 223         }
 224