crypto/bn/asm/x86_64-mont.pl

   1 #!/usr/bin/env perl
   2
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
   5 # project. The module is, however, dual licensed under OpenSSL and
   6 # CRYPTOGAMS licenses depending on where you obtain it. For further
   7 # details see http://www.openssl.org/~appro/cryptogams/.
   8 # ====================================================================
   9
  10 # October 2005.
  11 #
  12 # Montgomery multiplication routine for x86_64. While it gives modest
  13 # 9% improvement of rsa4096 sign on Opteron, rsa512 sign runs more
  14 # than twice, >2x, as fast. Most common rsa1024 sign is improved by
  15 # respectful 50%. It remains to be seen if loop unrolling and
  16 # dedicated squaring routine can provide further improvement...
  17
  18 $output=shift;
  19
  20 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  21 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  22 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  23 die "can't locate x86_64-xlate.pl";
  24
  25 open STDOUT,"| $^X $xlate $output";
  26
  27 # int bn_mul_mont(
  28 $rp="%rdi";     # BN_ULONG *rp,
  29 $ap="%rsi";     # const BN_ULONG *ap,
  30 $bp="%rdx";     # const BN_ULONG *bp,
  31 $np="%rcx";     # const BN_ULONG *np,
  32 $n0="%r8";      # const BN_ULONG *n0,
  33 $num="%r9";     # int num);
  34 $lo0="%r10";
  35 $hi0="%r11";
  36 $bp="%r12";     # reassign $bp
  37 $hi1="%r13";
  38 $i="%r14";
  39 $j="%r15";
  40 $m0="%rbx";
  41 $m1="%rbp";
  42
  43 $code=<<___;
  44 .text
  45
  46 .globl  bn_mul_mont
  47 .type   bn_mul_mont,\@function,6
  48 .align  16
  49 bn_mul_mont:
  50         push    %rbx
  51         push    %rbp
  52         push    %r12
  53         push    %r13
  54         push    %r14
  55         push    %r15
  56
  57         lea     2($num),%rax
  58         mov     %rsp,%rbp
  59         neg     %rax
  60         lea     (%rsp,%rax,8),%rsp      # tp=alloca(8*(num+2))
  61         and     \$-1024,%rsp            # minimize TLB usage
  62
  63         mov     %rbp,8(%rsp,$num,8)     # tp[num+1]=%rsp
  64         mov     %rdx,$bp                # $bp reassigned, remember?
  65
  66         mov     ($n0),$n0               # pull n0[0] value
  67
  68         xor     $i,$i                   # i=0
  69         xor     $j,$j                   # j=0
  70
  71         mov     ($bp),$m0               # m0=bp[0]
  72         mov     ($ap),%rax
  73         mulq    $m0                     # ap[0]*bp[0]
  74         mov     %rax,$lo0
  75         mov     %rdx,$hi0
  76
  77         imulq   $n0,%rax                # "tp[0]"*n0
  78         mov     %rax,$m1
  79
  80         mulq    ($np)                   # np[0]*m1
  81         add     $lo0,%rax               # discarded
  82         adc     \$0,%rdx
  83         mov     %rdx,$hi1
  84
  85         lea     1($j),$j                # j++
  86 .L1st:
  87         mov     ($ap,$j,8),%rax
  88         mulq    $m0                     # ap[j]*bp[0]
  89         add     $hi0,%rax
  90         adc     \$0,%rdx
  91         mov     %rax,$lo0
  92         mov     ($np,$j,8),%rax
  93         mov     %rdx,$hi0
  94
  95         mulq    $m1                     # np[j]*m1
  96         add     $hi1,%rax
  97         lea     1($j),$j                # j++
  98         adc     \$0,%rdx
  99         add     $lo0,%rax               # np[j]*m1+ap[j]*bp[0]
 100         adc     \$0,%rdx
 101         mov     %rax,-16(%rsp,$j,8)     # tp[j-1]
 102         cmp     $num,$j
 103         mov     %rdx,$hi1
 104         jl      .L1st
 105
 106         xor     %rdx,%rdx
 107         add     $hi0,$hi1
 108         adc     \$0,%rdx
 109         mov     $hi1,-8(%rsp,$num,8)
 110         mov     %rdx,(%rsp,$num,8)      # store upmost overflow bit
 111
 112         lea     1($i),$i                # i++
 113 .align  4
 114 .Louter:
 115         xor     $j,$j                   # j=0
 116
 117         mov     ($bp,$i,8),$m0          # m0=bp[i]
 118         mov     ($ap),%rax              # ap[0]
 119         mulq    $m0                     # ap[0]*bp[i]
 120         add     (%rsp),%rax             # ap[0]*bp[i]+tp[0]
 121         adc     \$0,%rdx
 122         mov     %rax,$lo0
 123         mov     %rdx,$hi0
 124
 125         imulq   $n0,%rax                # tp[0]*n0
 126         mov     %rax,$m1
 127
 128         mulq    ($np,$j,8)              # np[0]*m1
 129         add     $lo0,%rax               # discarded
 130         mov     8(%rsp),$lo0            # tp[1]
 131         adc     \$0,%rdx
 132         mov     %rdx,$hi1
 133
 134         lea     1($j),$j                # j++
 135 .align  4
 136 .Linner:
 137         mov     ($ap,$j,8),%rax
 138         mulq    $m0                     # ap[j]*bp[i]
 139         add     $hi0,%rax
 140         adc     \$0,%rdx
 141         add     %rax,$lo0               # ap[j]*bp[i]+tp[j]
 142         mov     ($np,$j,8),%rax
 143         adc     \$0,%rdx
 144         mov     %rdx,$hi0
 145
 146         mulq    $m1                     # np[j]*m1
 147         add     $hi1,%rax
 148         lea     1($j),$j                # j++
 149         adc     \$0,%rdx
 150         add     $lo0,%rax               # np[j]*m1+ap[j]*bp[i]+tp[j]
 151         adc     \$0,%rdx
 152         mov     (%rsp,$j,8),$lo0
 153         cmp     $num,$j
 154         mov     %rax,-16(%rsp,$j,8)     # tp[j-1]
 155         mov     %rdx,$hi1
 156         jl      .Linner
 157
 158         xor     %rdx,%rdx
 159         add     $hi0,$hi1
 160         adc     \$0,%rdx
 161         add     $lo0,$hi1               # pull upmost overflow bit
 162         adc     \$0,%rdx
 163         mov     $hi1,-8(%rsp,$num,8)
 164         mov     %rdx,(%rsp,$num,8)      # store upmost overflow bit
 165
 166         lea     1($i),$i                # i++
 167         cmp     $num,$i
 168         jl      .Louter
 169
 170         mov     -8($np,$num,8),%rax     # np[num-1]
 171         lea     (%rsp),$ap              # borrow ap for tp
 172         shr     \$62,%rax               # check for boundary condition
 173         lea     -1($num),$j             # j=num-1
 174         jz      .Lcopy
 175
 176         mov     ($ap),%rax              # tp[0]
 177         xor     $i,$i                   # i=0 and clear CF!
 178         jmp     .Lsub
 179 .align  16
 180 .Lsub:  sbb     ($np,$i,8),%rax
 181         mov     %rax,($rp,$i,8)         # rp[i]=tp[i]-np[i]
 182         dec     $j                      # doesn't affect CF!
 183         mov     8($ap,$i,8),%rax        # tp[i+1]
 184         lea     1($i),$i                # i++
 185         jge     .Lsub
 186
 187         sbb     \$0,%rax                # handle upmost overflow bit
 188         and     %rax,$ap
 189         not     %rax
 190         mov     $rp,$np
 191         and     %rax,$np
 192         lea     -1($num),$j
 193         or      $np,$ap                 # ap=borrow?tp:rp
 194 .align  16
 195 .Lcopy:                                 # copy or in-place refresh
 196         mov     ($ap,$j,8),%rax
 197         mov     %rax,($rp,$j,8)         # rp[i]=tp[i]
 198         mov     $i,(%rsp,$j,8)          # zap temporary vector
 199         dec     $j
 200         jge     .Lcopy
 201
 202         mov     8(%rsp,$num,8),%rsp     # restore %rsp
 203         mov     \$1,%rax
 204         pop     %r15
 205         pop     %r14
 206         pop     %r13
 207         pop     %r12
 208         pop     %rbp
 209         pop     %rbx
 210         ret
 211 .size   bn_mul_mont,.-bn_mul_mont
 212 .asciz  "Montgomery Multiplication for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 213 ___
 214
 215 print $code;
 216 close STDOUT;