crypto/sha/asm/sha256-586.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 # SHA256 block transform for x86. September 2007.
  11 #
  12 # Performance improvement over compiler generated code varies from
  13 # 10% to 40% [see below]. Not very impressive on some µ-archs, but
  14 # it's 5 times smaller and optimizies amount of writes.
  15 #
  16 # May 2012.
  17 #
  18 # Optimization including one of Pavel Semjanov's ideas resulted in
  19 # ~5% improvement on AMD and Sandy Bridge, and ~15% on Atom and P4.
  20 # Pavel also suggested full unroll. While his code runs ~20%/13%/6%
  21 # faster on K8/Core2/Sandy Bridge, it's 9.6x larger and ~14%/23%/24%
  22 # slower on P4/Atom/Pentium...
  23 #
  24 # Performance in clock cycles per processed byte (less is better):
  25 #
  26 #               Pentium PIII    P4      AMD K8  Core2   SB(**)  Atom
  27 # gcc           46      36      41      27      26
  28 # icc           57      33      38      25      23
  29 # x86 asm       39      31      29      19      18      19(**)  30
  30 # x86_64 asm(*) -       -       21      16      16      18      25
  31 #
  32 # (*)   x86_64 assembler performance is presented for reference
  33 #       purposes.
  34 # (**)  Sandy Bridge results can be improved by ~20% by replacing
  35 #       ror with equivalent shrd.
  36
  37 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  38 push(@INC,"${dir}","${dir}../../perlasm");
  39 require "x86asm.pl";
  40
  41 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
  42
  43 $A="eax";
  44 $E="edx";
  45 $T="ebx";
  46 $Aoff=&DWP(4,"esp");
  47 $Boff=&DWP(8,"esp");
  48 $Coff=&DWP(12,"esp");
  49 $Doff=&DWP(16,"esp");
  50 $Eoff=&DWP(20,"esp");
  51 $Foff=&DWP(24,"esp");
  52 $Goff=&DWP(28,"esp");
  53 $Hoff=&DWP(32,"esp");
  54 $Xoff=&DWP(36,"esp");
  55 $K256="ebp";
  56
  57 # *ror = sub { &shrd(@_[0],@_) };
  58
  59 sub BODY_00_15() {
  60     my $in_16_63=shift;
  61
  62         &mov    ("ecx",$E);
  63          &mov   ("esi",$Foff);
  64         &ror    ("ecx",25-11);
  65          &add   ($T,"edi")                      if ($in_16_63); # T += sigma1(X[-2])
  66          &mov   ("edi",$Goff);
  67         &xor    ("ecx",$E);
  68          &xor   ("esi","edi");
  69          &mov   (&DWP(4*(9+15),"esp"),$T)       if ($in_16_63); # save X[0]
  70         &ror    ("ecx",11-6);
  71          &and   ("esi",$E);
  72          &mov   ($Eoff,$E);     # modulo-scheduled
  73         &xor    ($E,"ecx");
  74          &xor   ("esi","edi");  # Ch(e,f,g)
  75          &add   ($T,$Hoff);     # T += h
  76         &ror    ($E,6);         # Sigma1(e)
  77          &mov   ("ecx",$A);
  78          &add   ($T,"esi");     # T += Ch(e,f,g)
  79
  80         &ror    ("ecx",22-13);
  81          &add   ($T,$E);        # T += Sigma1(e)
  82          &mov   ("edi",$Boff);
  83         &xor    ("ecx",$A);
  84          &mov   ($Aoff,$A);     # modulo-scheduled
  85          &lea   ("esp",&DWP(-4,"esp"));
  86         &ror    ("ecx",13-2);
  87          &mov   ("esi",&DWP(0,$K256));
  88         &xor    ("ecx",$A);
  89          &mov   ($E,$Eoff);     # e becomes d, which is e in next iteration
  90          &xor   ($A,"edi");     # a ^= b
  91         &ror    ("ecx",2);      # Sigma0(a)
  92
  93          &add   ($T,"esi");     # T+= K[i]
  94          &mov   (&DWP(0,"esp"),$A);             # (b^c) in next round
  95         &add    ($E,$T);        # d += T
  96          &and   ($A,&DWP(4,"esp"));     # a &= (b^c)
  97         &add    ($T,"ecx");     # T += Sigma0(a)
  98          &xor   ($A,"edi");     # h = Maj(a,b,c) = Ch(a^b,c,b)
  99
 100         &add    ($K256,4);
 101          &add   ($A,$T);        # h += T
 102         &mov    ($T,&DWP(4*(9+15+16-1),"esp"))  if ($in_16_63); # preload T
 103 }
 104
 105 &function_begin("sha256_block_data_order");
 106         &mov    ("esi",wparam(0));      # ctx
 107         &mov    ("edi",wparam(1));      # inp
 108         &mov    ("eax",wparam(2));      # num
 109         &mov    ("ebx","esp");          # saved sp
 110
 111         &call   (&label("pic_point"));  # make it PIC!
 112 &set_label("pic_point");
 113         &blindpop($K256);
 114         &lea    ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
 115
 116         &sub    ("esp",16);
 117         &and    ("esp",-64);
 118
 119         &shl    ("eax",6);
 120         &add    ("eax","edi");
 121         &mov    (&DWP(0,"esp"),"esi");  # ctx
 122         &mov    (&DWP(4,"esp"),"edi");  # inp
 123         &mov    (&DWP(8,"esp"),"eax");  # inp+num*128
 124         &mov    (&DWP(12,"esp"),"ebx"); # saved sp
 125
 126 &set_label("loop",16);
 127     # copy input block to stack reversing byte and dword order
 128     for($i=0;$i<4;$i++) {
 129         &mov    ("eax",&DWP($i*16+0,"edi"));
 130         &mov    ("ebx",&DWP($i*16+4,"edi"));
 131         &mov    ("ecx",&DWP($i*16+8,"edi"));
 132         &bswap  ("eax");
 133         &mov    ("edx",&DWP($i*16+12,"edi"));
 134         &bswap  ("ebx");
 135         &push   ("eax");
 136         &bswap  ("ecx");
 137         &push   ("ebx");
 138         &bswap  ("edx");
 139         &push   ("ecx");
 140         &push   ("edx");
 141     }
 142         &add    ("edi",64);
 143         &lea    ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
 144         &mov    (&DWP(4*(9+16)+4,"esp"),"edi");
 145
 146         # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
 147         &mov    ($A,&DWP(0,"esi"));
 148         &mov    ("ebx",&DWP(4,"esi"));
 149         &mov    ("ecx",&DWP(8,"esi"));
 150         &mov    ("edi",&DWP(12,"esi"));
 151         # &mov  ($Aoff,$A);
 152         &mov    ($Boff,"ebx");
 153         &xor    ("ebx","ecx");
 154         &mov    ($Coff,"ecx");
 155         &mov    ($Doff,"edi");
 156         &mov    (&DWP(0,"esp"),"ebx");  # magic
 157         &mov    ($E,&DWP(16,"esi"));
 158         &mov    ("ebx",&DWP(20,"esi"));
 159         &mov    ("ecx",&DWP(24,"esi"));
 160         &mov    ("edi",&DWP(28,"esi"));
 161         # &mov  ($Eoff,$E);
 162         &mov    ($Foff,"ebx");
 163         &mov    ($Goff,"ecx");
 164         &mov    ($Hoff,"edi");
 165
 166 &set_label("00_15",16);
 167         &mov    ($T,&DWP(4*(9+15),"esp"));
 168
 169         &BODY_00_15();
 170
 171         &cmp    ("esi",0xc19bf174);
 172         &jne    (&label("00_15"));
 173
 174         &mov    ($T,&DWP(4*(9+15+16-1),"esp")); # preloaded in BODY_00_15(1)
 175 &set_label("16_63",16);
 176         &mov    ("esi",$T);
 177          &mov   ("ecx",&DWP(4*(9+15+16-14),"esp"));
 178         &ror    ("esi",18-7);
 179          &mov   ("edi","ecx");
 180         &ror    ("ecx",19-17);
 181          &xor   ("esi",$T);
 182         &shr    ($T,3);
 183          &xor   ("ecx","edi");
 184         &ror    ("esi",7);
 185          &xor   ($T,"esi");                     # T = sigma0(X[-15])
 186         &ror    ("ecx",17);
 187          &add   ($T,&DWP(4*(9+15+16),"esp"));   # T += X[-16]
 188         &shr    ("edi",10);
 189          &add   ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
 190         &xor    ("edi","ecx");                  # sigma1(X[-2])
 191         # &add  ($T,"edi");                     # T += sigma1(X[-2])
 192         # &mov  (&DWP(4*(9+15),"esp"),$T);      # save X[0]
 193
 194         &BODY_00_15(1);
 195
 196         &cmp    ("esi",0xc67178f2);
 197         &jne    (&label("16_63"));
 198
 199         &mov    ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
 200         # &mov  ($A,$Aoff);
 201         &mov    ("ebx",$Boff);
 202         &mov    ("ecx",$Coff);
 203         &mov    ("edi",$Doff);
 204         &add    ($A,&DWP(0,"esi"));
 205         &add    ("ebx",&DWP(4,"esi"));
 206         &add    ("ecx",&DWP(8,"esi"));
 207         &add    ("edi",&DWP(12,"esi"));
 208         &mov    (&DWP(0,"esi"),$A);
 209         &mov    (&DWP(4,"esi"),"ebx");
 210         &mov    (&DWP(8,"esi"),"ecx");
 211         &mov    (&DWP(12,"esi"),"edi");
 212         # &mov  ($E,$Eoff);
 213         &mov    ("eax",$Foff);
 214         &mov    ("ebx",$Goff);
 215         &mov    ("ecx",$Hoff);
 216         &mov    ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
 217         &add    ($E,&DWP(16,"esi"));
 218         &add    ("eax",&DWP(20,"esi"));
 219         &add    ("ebx",&DWP(24,"esi"));
 220         &add    ("ecx",&DWP(28,"esi"));
 221         &mov    (&DWP(16,"esi"),$E);
 222         &mov    (&DWP(20,"esi"),"eax");
 223         &mov    (&DWP(24,"esi"),"ebx");
 224         &mov    (&DWP(28,"esi"),"ecx");
 225
 226         &lea    ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
 227         &sub    ($K256,4*64);                   # rewind K
 228
 229         &cmp    ("edi",&DWP(8,"esp"));          # are we done yet?
 230         &jb     (&label("loop"));
 231
 232         &mov    ("esp",&DWP(12,"esp"));         # restore sp
 233 &function_end_A();
 234
 235 &set_label("K256",64);  # Yes! I keep it in the code segment!
 236         &data_word(0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5);
 237         &data_word(0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5);
 238         &data_word(0xd807aa98,0x12835b01,0x243185be,0x550c7dc3);
 239         &data_word(0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174);
 240         &data_word(0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc);
 241         &data_word(0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da);
 242         &data_word(0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7);
 243         &data_word(0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967);
 244         &data_word(0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13);
 245         &data_word(0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85);
 246         &data_word(0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3);
 247         &data_word(0xd192e819,0xd6990624,0xf40e3585,0x106aa070);
 248         &data_word(0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5);
 249         &data_word(0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3);
 250         &data_word(0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208);
 251         &data_word(0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2);
 252 &function_end_B("sha256_block_data_order");
 253 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
 254
 255 &asm_finish();