crypto/modes/asm/ghashp8-ppc.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@openssl.org> 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 # GHASH for for PowerISA v2.07.
  11 #
  12 # July 2014
  13 #
  14 # Initial version is 2.27x slower than hardware-assisted AES-128-CTR,
  15 # 11x faster than "4-bit" integer-only compiler-generated 64-bit code.
  16
  17 $flavour=shift;
  18 $output =shift;
  19
  20 if ($flavour =~ /64/) {
  21         $SIZE_T=8;
  22         $LRSAVE=2*$SIZE_T;
  23         $STU="stdu";
  24         $POP="ld";
  25         $PUSH="std";
  26 } elsif ($flavour =~ /32/) {
  27         $SIZE_T=4;
  28         $LRSAVE=$SIZE_T;
  29         $STU="stwu";
  30         $POP="lwz";
  31         $PUSH="stw";
  32 } else { die "nonsense $flavour"; }
  33
  34 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  35 ( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
  36 ( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or
  37 die "can't locate ppc-xlate.pl";
  38
  39 open STDOUT,"| $^X $xlate $flavour $output" || die "can't call $xlate: $!";
  40
  41 my ($Xip,$Htbl,$inp,$len)=map("r$_",(3..6));    # argument block
  42
  43 my ($Xl,$Xm,$Xh,$IN)=map("v$_",(0..3));
  44 my ($zero,$t0,$t1,$t2,$xC2,$H,$Hh,$Hl,$lemask)=map("v$_",(4..12));
  45 my $vrsave="r12";
  46
  47 $code=<<___;
  48 .machine        "any"
  49
  50 .text
  51
  52 .globl  .gcm_init_p8
  53 .align  5
  54 .gcm_init_p8:
  55         lis             r0,0xfff0
  56         li              r8,0x10
  57         mfspr           $vrsave,256
  58         li              r9,0x20
  59         mtspr           256,r0
  60         li              r10,0x30
  61         lvx_u           $H,0,r4                 # load H
  62
  63         vspltisb        $xC2,-16                # 0xf0
  64         vspltisb        $t0,1                   # one
  65         vaddubm         $xC2,$xC2,$xC2          # 0xe0
  66         vxor            $zero,$zero,$zero
  67         vor             $xC2,$xC2,$t0           # 0xe1
  68         vsldoi          $xC2,$xC2,$zero,15      # 0xe1...
  69         vsldoi          $t1,$zero,$t0,1         # ...1
  70         vaddubm         $xC2,$xC2,$xC2          # 0xc2...
  71         vspltisb        $t2,7
  72         vor             $xC2,$xC2,$t1           # 0xc2....01
  73         vspltb          $t1,$H,0                # most significant byte
  74         vsl             $H,$H,$t0               # H<<=1
  75         vsrab           $t1,$t1,$t2             # broadcast carry bit
  76         vand            $t1,$t1,$xC2
  77         vxor            $H,$H,$t1               # twisted H
  78
  79         vsldoi          $H,$H,$H,8              # twist even more ...
  80         vsldoi          $xC2,$zero,$xC2,8       # 0xc2.0
  81         vsldoi          $Hl,$zero,$H,8          # ... and split
  82         vsldoi          $Hh,$H,$zero,8
  83
  84         stvx_u          $xC2,0,r3               # save pre-computed table
  85         stvx_u          $Hl,r8,r3
  86         stvx_u          $H, r9,r3
  87         stvx_u          $Hh,r10,r3
  88
  89         mtspr           256,$vrsave
  90         blr
  91         .long           0
  92         .byte           0,12,0x14,0,0,0,2,0
  93         .long           0
  94 .size   .gcm_init_p8,.-.gcm_init_p8
  95
  96 .globl  .gcm_gmult_p8
  97 .align  5
  98 .gcm_gmult_p8:
  99         lis             r0,0xfff8
 100         li              r8,0x10
 101         mfspr           $vrsave,256
 102         li              r9,0x20
 103         mtspr           256,r0
 104         li              r10,0x30
 105         lvx_u           $IN,0,$Xip              # load Xi
 106
 107         lvx_u           $Hl,r8,$Htbl            # load pre-computed table
 108          le?lvsl        $lemask,r0,r0
 109         lvx_u           $H, r9,$Htbl
 110          le?vspltisb    $t0,0x07
 111         lvx_u           $Hh,r10,$Htbl
 112          le?vxor        $lemask,$lemask,$t0
 113         lvx_u           $xC2,0,$Htbl
 114          le?vperm       $IN,$IN,$IN,$lemask
 115         vxor            $zero,$zero,$zero
 116
 117         vpmsumd         $Xl,$IN,$Hl             # H.lo·Xi.lo
 118         vpmsumd         $Xm,$IN,$H              # H.hi·Xi.lo+H.lo·Xi.hi
 119         vpmsumd         $Xh,$IN,$Hh             # H.hi·Xi.hi
 120
 121         vpmsumd         $t2,$Xl,$xC2            # 1st phase
 122
 123         vsldoi          $t0,$Xm,$zero,8
 124         vsldoi          $t1,$zero,$Xm,8
 125         vxor            $Xl,$Xl,$t0
 126         vxor            $Xh,$Xh,$t1
 127
 128         vsldoi          $Xl,$Xl,$Xl,8
 129         vxor            $Xl,$Xl,$t2
 130
 131         vsldoi          $t1,$Xl,$Xl,8           # 2nd phase
 132         vpmsumd         $Xl,$Xl,$xC2
 133         vxor            $t1,$t1,$Xh
 134         vxor            $Xl,$Xl,$t1
 135
 136         le?vperm        $Xl,$Xl,$Xl,$lemask
 137         stvx_u          $Xl,0,$Xip              # write out Xi
 138
 139         mtspr           256,$vrsave
 140         blr
 141         .long           0
 142         .byte           0,12,0x14,0,0,0,2,0
 143         .long           0
 144 .size   .gcm_gmult_p8,.-.gcm_gmult_p8
 145
 146 .globl  .gcm_ghash_p8
 147 .align  5
 148 .gcm_ghash_p8:
 149         lis             r0,0xfff8
 150         li              r8,0x10
 151         mfspr           $vrsave,256
 152         li              r9,0x20
 153         mtspr           256,r0
 154         li              r10,0x30
 155         lvx_u           $Xl,0,$Xip              # load Xi
 156
 157         lvx_u           $Hl,r8,$Htbl            # load pre-computed table
 158          le?lvsl        $lemask,r0,r0
 159         lvx_u           $H, r9,$Htbl
 160          le?vspltisb    $t0,0x07
 161         lvx_u           $Hh,r10,$Htbl
 162          le?vxor        $lemask,$lemask,$t0
 163         lvx_u           $xC2,0,$Htbl
 164          le?vperm       $Xl,$Xl,$Xl,$lemask
 165         vxor            $zero,$zero,$zero
 166
 167         lvx_u           $IN,0,$inp
 168         addi            $inp,$inp,16
 169         subi            $len,$len,16
 170          le?vperm       $IN,$IN,$IN,$lemask
 171         b               Loop
 172
 173 .align  5
 174 Loop:
 175         vxor            $IN,$IN,$Xl
 176          subic          $len,$len,16
 177
 178         vpmsumd         $Xl,$IN,$Hl             # H.lo·Xi.lo
 179          subfe.         r0,r0,r0                # borrow?-1:0
 180         vpmsumd         $Xm,$IN,$H              # H.hi·Xi.lo+H.lo·Xi.hi
 181          and            r0,r0,$len
 182         vpmsumd         $Xh,$IN,$Hh             # H.hi·Xi.hi
 183          add            $inp,$inp,r0
 184
 185         vpmsumd         $t2,$Xl,$xC2            # 1st phase
 186
 187         vsldoi          $t0,$Xm,$zero,8
 188         vsldoi          $t1,$zero,$Xm,8
 189         vxor            $Xl,$Xl,$t0
 190         vxor            $Xh,$Xh,$t1
 191
 192         vsldoi          $Xl,$Xl,$Xl,8
 193         vxor            $Xl,$Xl,$t2
 194          lvx_u          $IN,0,$inp
 195          addi           $inp,$inp,16
 196
 197         vsldoi          $t1,$Xl,$Xl,8           # 2nd phase
 198         vpmsumd         $Xl,$Xl,$xC2
 199         vxor            $t1,$t1,$Xh
 200          le?vperm       $IN,$IN,$IN,$lemask
 201         vxor            $Xl,$Xl,$t1
 202         beq             Loop                    # did $len-=16 borrow?
 203
 204         le?vperm        $Xl,$Xl,$Xl,$lemask
 205         stvx_u          $Xl,0,$Xip              # write out Xi
 206
 207         mtspr           256,$vrsave
 208         blr
 209         .long           0
 210         .byte           0,12,0x14,0,0,0,4,0
 211         .long           0
 212 .size   .gcm_ghash_p8,.-.gcm_ghash_p8
 213
 214 .asciz  "GHASH for PowerISA 2.07, CRYPTOGAMS by <appro\@openssl.org>"
 215 .align  2
 216 ___
 217
 218 foreach (split("\n",$code)) {
 219         if ($flavour =~ /le$/o) {       # little-endian
 220             s/le\?//o           or
 221             s/be\?/#be#/o;
 222         } else {
 223             s/le\?/#le#/o       or
 224             s/be\?//o;
 225         }
 226         print $_,"\n";
 227 }
 228
 229 close STDOUT; # enforce flush