crypto/bn/asm/alpha/sqr_c4.pl

   1 #!/usr/local/bin/perl
   2 # alpha assember
   3
   4 sub sqr_add_c
   5         {
   6         local($a,$c0,$c1,$c2)=@_;
   7         local($l1,$h1,$t1,$t2);
   8
   9         &mul($a,$a,($l1)=&NR(1));
  10         &muh($a,$a,($h1)=&NR(1));
  11         &add($c0,$l1,$c0);
  12         &add($c1,$h1,$c1);
  13         &cmpult($c0,$l1,($t1)=&NR(1));  &FR($l1);
  14         &cmpult($c1,$h1,($t2)=&NR(1));  &FR($h1);
  15         &add($c1,$t1,$c1);              &FR($t1);
  16         &add($c2,$t2,$c2);              &FR($t2);
  17         }
  18
  19 sub sqr_add_c2
  20         {
  21         local($a,$b,$c0,$c1,$c2)=@_;
  22         local($l1,$h1,$t1,$t2);
  23
  24         &mul($a,$b,($l1)=&NR(1));
  25         &muh($a,$b,($h1)=&NR(1));
  26         &cmplt($l1,"zero",($lc1)=&NR(1));
  27         &cmplt($h1,"zero",($hc1)=&NR(1));
  28         &add($l1,$l1,$l1);
  29         &add($h1,$h1,$h1);
  30         &add($h1,$lc1,$h1);             &FR($lc1);
  31         &add($c2,$hc1,$c2);             &FR($hc1);
  32
  33         &add($c0,$l1,$c0);
  34         &add($c1,$h1,$c1);
  35         &cmpult($c0,$l1,($lc1)=&NR(1)); &FR($l1);
  36         &cmpult($c1,$h1,($hc1)=&NR(1)); &FR($h1);
  37
  38         &add($c1,$lc1,$c1);             &FR($lc1);
  39         &add($c2,$hc1,$c2);             &FR($hc1);
  40         }
  41
  42
  43 sub bn_sqr_comba4
  44         {
  45         local($name)=@_;
  46         local(@a,@b,$r,$c0,$c1,$c2);
  47
  48         $cnt=1;
  49         &init_pool(2);
  50
  51         $rp=&wparam(0);
  52         $ap=&wparam(1);
  53
  54         &function_begin($name,"");
  55
  56         &comment("");
  57
  58         &ld(($a[0])=&NR(1),&QWPw(0,$ap));
  59         &ld(($a[1])=&NR(1),&QWPw(1,$ap));
  60         &ld(($a[2])=&NR(1),&QWPw(2,$ap));
  61         &ld(($a[3])=&NR(1),&QWPw(3,$ap)); &FR($ap);
  62
  63         ($c0,$c1,$c2)=&NR(3);
  64
  65         &mov("zero",$c2);
  66         &mul($a[0],$a[0],$c0);
  67         &muh($a[0],$a[0],$c1);
  68         &st($c0,&QWPw(0,$rp));
  69         ($c0,$c1,$c2)=($c1,$c2,$c0);
  70         &mov("zero",$c2);
  71
  72         &sqr_add_c2($a[0],$a[1],$c0,$c1,$c2);
  73         &st($c0,&QWPw(1,$rp));
  74         ($c0,$c1,$c2)=($c1,$c2,$c0);
  75         &mov("zero",$c2);
  76
  77         &sqr_add_c($a[1],$c0,$c1,$c2);
  78         &sqr_add_c2($a[2],$a[0],$c0,$c1,$c2);
  79         &st($c0,&QWPw(2,$rp));
  80         ($c0,$c1,$c2)=($c1,$c2,$c0);
  81         &mov("zero",$c2);
  82
  83         &sqr_add_c2($a[3],$a[0],$c0,$c1,$c2);
  84         &sqr_add_c2($a[2],$a[1],$c0,$c1,$c2);
  85         &st($c0,&QWPw(3,$rp));
  86         ($c0,$c1,$c2)=($c1,$c2,$c0);
  87         &mov("zero",$c2);
  88
  89         &sqr_add_c($a[2],$c0,$c1,$c2);
  90         &sqr_add_c2($a[3],$a[1],$c0,$c1,$c2);
  91         &st($c0,&QWPw(4,$rp));
  92         ($c0,$c1,$c2)=($c1,$c2,$c0);
  93         &mov("zero",$c2);
  94
  95         &sqr_add_c2($a[3],$a[2],$c0,$c1,$c2);
  96         &st($c0,&QWPw(5,$rp));
  97         ($c0,$c1,$c2)=($c1,$c2,$c0);
  98         &mov("zero",$c2);
  99
 100         &sqr_add_c($a[3],$c0,$c1,$c2);
 101         &st($c0,&QWPw(6,$rp));
 102         &st($c1,&QWPw(7,$rp));
 103
 104         &function_end($name);
 105
 106         &fin_pool;
 107         }
 108
 109 1;