crypto/bn/asm/alpha.works/mul_c4.works.pl

   1 #!/usr/local/bin/perl
   2 # alpha assember
   3
   4 sub mul_add_c
   5         {
   6         local($a,$b,$c0,$c1,$c2)=@_;
   7         local($l1,$h1,$t1,$t2);
   8
   9 print STDERR "count=$cnt\n"; $cnt++;
  10         &mul($a,$b,($l1)=&NR(1));
  11         &muh($a,$b,($h1)=&NR(1));
  12         &add($c0,$l1,$c0);
  13         &cmpult($c0,$l1,($t1)=&NR(1));  &FR($l1);
  14         &add($t1,$h1,$h1);              &FR($t1);
  15         &add($c1,$h1,$c1);
  16         &cmpult($c1,$h1,($t2)=&NR(1));  &FR($h1);
  17         &add($c2,$t2,$c2);              &FR($t2);
  18         }
  19
  20 sub bn_mul_comba4
  21         {
  22         local($name)=@_;
  23         local(@a,@b,$r,$c0,$c1,$c2);
  24
  25         $cnt=1;
  26         &init_pool(3);
  27
  28         $rp=&wparam(0);
  29         $ap=&wparam(1);
  30         $bp=&wparam(2);
  31
  32         &function_begin($name,"");
  33
  34         &comment("");
  35
  36         &ld(($a[0])=&NR(1),&QWPw(0,$ap));
  37         &ld(($b[0])=&NR(1),&QWPw(0,$bp));
  38         &ld(($a[1])=&NR(1),&QWPw(1,$ap));
  39         &ld(($b[1])=&NR(1),&QWPw(1,$bp));
  40         &ld(($a[2])=&NR(1),&QWPw(2,$ap));
  41         &ld(($b[2])=&NR(1),&QWPw(2,$bp));
  42         &ld(($a[3])=&NR(1),&QWPw(3,$ap));       &FR($ap);
  43         &ld(($b[3])=&NR(1),&QWPw(3,$bp));       &FR($bp);
  44
  45         ($c0,$c1,$c2)=&NR(3);
  46         &mov("zero",$c2);
  47         &mul($a[0],$b[0],$c0);
  48         &muh($a[0],$b[0],$c1);
  49         &st($c0,&QWPw(0,$rp));                  &FR($c0); ($c0)=&NR($c0);
  50         ($c0,$c1,$c2)=($c1,$c2,$c0);
  51         &mov("zero",$c2);
  52
  53         &mul_add_c($a[0],$b[1],$c0,$c1,$c2);
  54         &mul_add_c($a[1],$b[0],$c0,$c1,$c2);
  55         &st($c0,&QWPw(1,$rp));                  &FR($c0); ($c0)=&NR($c0);
  56         ($c0,$c1,$c2)=($c1,$c2,$c0);
  57         &mov("zero",$c2);
  58
  59         &mul_add_c($a[1],$b[1],$c0,$c1,$c2);
  60         &mul_add_c($a[0],$b[2],$c0,$c1,$c2);
  61         &mul_add_c($a[2],$b[0],$c0,$c1,$c2);
  62         &st($c0,&QWPw(2,$rp));                  &FR($c0); ($c0)=&NR($c0);
  63         ($c0,$c1,$c2)=($c1,$c2,$c0);
  64         &mov("zero",$c2);
  65
  66         &mul_add_c($a[0],$b[3],$c0,$c1,$c2);    &FR($a[0]);
  67         &mul_add_c($a[1],$b[2],$c0,$c1,$c2);
  68         &mul_add_c($a[2],$b[1],$c0,$c1,$c2);
  69         &mul_add_c($a[3],$b[0],$c0,$c1,$c2);    &FR($b[0]);
  70         &st($c0,&QWPw(3,$rp));                  &FR($c0); ($c0)=&NR($c0);
  71         ($c0,$c1,$c2)=($c1,$c2,$c0);
  72         &mov("zero",$c2);
  73
  74         &mul_add_c($a[1],$b[3],$c0,$c1,$c2);    &FR($a[1]);
  75         &mul_add_c($a[2],$b[2],$c0,$c1,$c2);
  76         &mul_add_c($a[3],$b[1],$c0,$c1,$c2);    &FR($b[1]);
  77         &st($c0,&QWPw(4,$rp));                  &FR($c0); ($c0)=&NR($c0);
  78         ($c0,$c1,$c2)=($c1,$c2,$c0);
  79         &mov("zero",$c2);
  80
  81         &mul_add_c($a[2],$b[3],$c0,$c1,$c2);    &FR($a[2]);
  82         &mul_add_c($a[3],$b[2],$c0,$c1,$c2);    &FR($b[2]);
  83         &st($c0,&QWPw(5,$rp));                  &FR($c0); ($c0)=&NR($c0);
  84         ($c0,$c1,$c2)=($c1,$c2,$c0);
  85         &mov("zero",$c2);
  86
  87         &mul_add_c($a[3],$b[3],$c0,$c1,$c2);    &FR($a[3],$b[3]);
  88         &st($c0,&QWPw(6,$rp));
  89         &st($c1,&QWPw(7,$rp));
  90
  91         &FR($c0,$c1,$c2);
  92
  93         &function_end($name);
  94
  95         &fin_pool;
  96         }
  97
  98 1;