crypto/bn/asm/alpha/sqr_c8.pl

   1 #!/usr/local/bin/perl
   2 # alpha assember
   3
   4 sub bn_sqr_comba8
   5         {
   6         local($name)=@_;
   7         local(@a,@b,$r,$c0,$c1,$c2);
   8
   9         $cnt=1;
  10         &init_pool(2);
  11
  12         $rp=&wparam(0);
  13         $ap=&wparam(1);
  14
  15         &function_begin($name,"");
  16
  17         &comment("");
  18
  19         &ld(($a[0])=&NR(1),&QWPw(0,$ap));
  20         &ld(($a[1])=&NR(1),&QWPw(1,$ap));
  21         &ld(($a[2])=&NR(1),&QWPw(2,$ap));
  22         &ld(($a[3])=&NR(1),&QWPw(3,$ap));
  23         &ld(($a[4])=&NR(1),&QWPw(4,$ap));
  24         &ld(($a[5])=&NR(1),&QWPw(5,$ap));
  25         &ld(($a[6])=&NR(1),&QWPw(6,$ap));
  26         &ld(($a[7])=&NR(1),&QWPw(7,$ap)); &FR($ap);
  27
  28         ($c0,$c1,$c2)=&NR(3);
  29
  30         &mov("zero",$c2);
  31         &mul($a[0],$a[0],$c0);
  32         &muh($a[0],$a[0],$c1);
  33         &st($c0,&QWPw(0,$rp));
  34         ($c0,$c1,$c2)=($c1,$c2,$c0);
  35         &mov("zero",$c2);
  36
  37         &sqr_add_c2($a[1],$a[0],$c0,$c1,$c2);
  38         &st($c0,&QWPw(1,$rp));
  39         ($c0,$c1,$c2)=($c1,$c2,$c0);
  40         &mov("zero",$c2);
  41
  42         &sqr_add_c($a[1],$c0,$c1,$c2);
  43         &sqr_add_c2($a[2],$a[0],$c0,$c1,$c2);
  44         &st($c0,&QWPw(2,$rp));
  45         ($c0,$c1,$c2)=($c1,$c2,$c0);
  46         &mov("zero",$c2);
  47
  48         &sqr_add_c2($a[2],$a[1],$c0,$c1,$c2);
  49         &sqr_add_c2($a[3],$a[0],$c0,$c1,$c2);
  50         &st($c0,&QWPw(3,$rp));
  51         ($c0,$c1,$c2)=($c1,$c2,$c0);
  52         &mov("zero",$c2);
  53
  54         &sqr_add_c($a[2],$c0,$c1,$c2);
  55         &sqr_add_c2($a[3],$a[1],$c0,$c1,$c2);
  56         &sqr_add_c2($a[4],$a[0],$c0,$c1,$c2);
  57         &st($c0,&QWPw(4,$rp));
  58         ($c0,$c1,$c2)=($c1,$c2,$c0);
  59         &mov("zero",$c2);
  60
  61         &sqr_add_c2($a[3],$a[2],$c0,$c1,$c2);
  62         &sqr_add_c2($a[4],$a[1],$c0,$c1,$c2);
  63         &sqr_add_c2($a[5],$a[0],$c0,$c1,$c2);
  64         &st($c0,&QWPw(5,$rp));
  65         ($c0,$c1,$c2)=($c1,$c2,$c0);
  66         &mov("zero",$c2);
  67
  68         &sqr_add_c($a[3],$c0,$c1,$c2);
  69         &sqr_add_c2($a[4],$a[2],$c0,$c1,$c2);
  70         &sqr_add_c2($a[5],$a[1],$c0,$c1,$c2);
  71         &sqr_add_c2($a[6],$a[0],$c0,$c1,$c2);
  72         &st($c0,&QWPw(6,$rp));
  73         ($c0,$c1,$c2)=($c1,$c2,$c0);
  74         &mov("zero",$c2);
  75
  76         &sqr_add_c2($a[4],$a[3],$c0,$c1,$c2);
  77         &sqr_add_c2($a[5],$a[2],$c0,$c1,$c2);
  78         &sqr_add_c2($a[6],$a[1],$c0,$c1,$c2);
  79         &sqr_add_c2($a[7],$a[0],$c0,$c1,$c2);
  80         &st($c0,&QWPw(7,$rp));
  81         ($c0,$c1,$c2)=($c1,$c2,$c0);
  82         &mov("zero",$c2);
  83
  84         &sqr_add_c($a[4],$c0,$c1,$c2);
  85         &sqr_add_c2($a[5],$a[3],$c0,$c1,$c2);
  86         &sqr_add_c2($a[6],$a[2],$c0,$c1,$c2);
  87         &sqr_add_c2($a[7],$a[1],$c0,$c1,$c2);
  88         &st($c0,&QWPw(8,$rp));
  89         ($c0,$c1,$c2)=($c1,$c2,$c0);
  90         &mov("zero",$c2);
  91
  92         &sqr_add_c2($a[5],$a[4],$c0,$c1,$c2);
  93         &sqr_add_c2($a[6],$a[3],$c0,$c1,$c2);
  94         &sqr_add_c2($a[7],$a[2],$c0,$c1,$c2);
  95         &st($c0,&QWPw(9,$rp));
  96         ($c0,$c1,$c2)=($c1,$c2,$c0);
  97         &mov("zero",$c2);
  98
  99         &sqr_add_c($a[5],$c0,$c1,$c2);
 100         &sqr_add_c2($a[6],$a[4],$c0,$c1,$c2);
 101         &sqr_add_c2($a[7],$a[3],$c0,$c1,$c2);
 102         &st($c0,&QWPw(10,$rp));
 103         ($c0,$c1,$c2)=($c1,$c2,$c0);
 104         &mov("zero",$c2);
 105
 106         &sqr_add_c2($a[6],$a[5],$c0,$c1,$c2);
 107         &sqr_add_c2($a[7],$a[4],$c0,$c1,$c2);
 108         &st($c0,&QWPw(11,$rp));
 109         ($c0,$c1,$c2)=($c1,$c2,$c0);
 110         &mov("zero",$c2);
 111
 112         &sqr_add_c($a[6],$c0,$c1,$c2);
 113         &sqr_add_c2($a[7],$a[5],$c0,$c1,$c2);
 114         &st($c0,&QWPw(12,$rp));
 115         ($c0,$c1,$c2)=($c1,$c2,$c0);
 116         &mov("zero",$c2);
 117
 118         &sqr_add_c2($a[7],$a[6],$c0,$c1,$c2);
 119         &st($c0,&QWPw(13,$rp));
 120         ($c0,$c1,$c2)=($c1,$c2,$c0);
 121         &mov("zero",$c2);
 122
 123         &sqr_add_c($a[7],$c0,$c1,$c2);
 124         &st($c0,&QWPw(14,$rp));
 125         &st($c1,&QWPw(15,$rp));
 126
 127         &function_end($name);
 128
 129         &fin_pool;
 130         }
 131
 132 1;