lib/libcrypto/sha/asm/sha512-x86_64.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
   5 # project. Rights for redistribution and usage in source and binary
   6 # forms are granted according to the OpenSSL license.
   7 # ====================================================================
   8 #
   9 # sha256/512_block procedure for x86_64.
  10 #
  11 # 40% improvement over compiler-generated code on Opteron. On EM64T
  12 # sha256 was observed to run >80% faster and sha512 - >40%. No magical
  13 # tricks, just straight implementation... I really wonder why gcc
  14 # [being armed with inline assembler] fails to generate as fast code.
  15 # The only thing which is cool about this module is that it's very
  16 # same instruction sequence used for both SHA-256 and SHA-512. In
  17 # former case the instructions operate on 32-bit operands, while in
  18 # latter - on 64-bit ones. All I had to do is to get one flavor right,
  19 # the other one passed the test right away:-)
  20 #
  21 # sha256_block runs in ~1005 cycles on Opteron, which gives you
  22 # asymptotic performance of 64*1000/1005=63.7MBps times CPU clock
  23 # frequency in GHz. sha512_block runs in ~1275 cycles, which results
  24 # in 128*1000/1275=100MBps per GHz. Is there room for improvement?
  25 # Well, if you compare it to IA-64 implementation, which maintains
  26 # X[16] in register bank[!], tends to 4 instructions per CPU clock
  27 # cycle and runs in 1003 cycles, 1275 is very good result for 3-way
  28 # issue Opteron pipeline and X[16] maintained in memory. So that *if*
  29 # there is a way to improve it, *then* the only way would be to try to
  30 # offload X[16] updates to SSE unit, but that would require "deeper"
  31 # loop unroll, which in turn would naturally cause size blow-up, not
  32 # to mention increased complexity! And once again, only *if* it's
  33 # actually possible to noticeably improve overall ILP, instruction
  34 # level parallelism, on a given CPU implementation in this case.
  35 #
  36 # Special note on Intel EM64T. While Opteron CPU exhibits perfect
  37 # perfromance ratio of 1.5 between 64- and 32-bit flavors [see above],
  38 # [currently available] EM64T CPUs apparently are far from it. On the
  39 # contrary, 64-bit version, sha512_block, is ~30% *slower* than 32-bit
  40 # sha256_block:-( This is presumably because 64-bit shifts/rotates
  41 # apparently are not atomic instructions, but implemented in microcode.
  42
  43 $flavour = shift;
  44 $output  = shift;
  45 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
  46
  47 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  48 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
  49 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
  50 die "can't locate x86_64-xlate.pl";
  51
  52 open OUT,"| \"$^X\" $xlate $flavour $output";
  53 *STDOUT=*OUT;
  54
  55 if ($output =~ /512/) {
  56         $func="sha512_block_data_order";
  57         $TABLE="K512";
  58         $SZ=8;
  59         @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
  60                                         "%r8", "%r9", "%r10","%r11");
  61         ($T1,$a0,$a1,$a2)=("%r12","%r13","%r14","%r15");
  62         @Sigma0=(28,34,39);
  63         @Sigma1=(14,18,41);
  64         @sigma0=(1,  8, 7);
  65         @sigma1=(19,61, 6);
  66         $rounds=80;
  67 } else {
  68         $func="sha256_block_data_order";
  69         $TABLE="K256";
  70         $SZ=4;
  71         @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
  72                                         "%r8d","%r9d","%r10d","%r11d");
  73         ($T1,$a0,$a1,$a2)=("%r12d","%r13d","%r14d","%r15d");
  74         @Sigma0=( 2,13,22);
  75         @Sigma1=( 6,11,25);
  76         @sigma0=( 7,18, 3);
  77         @sigma1=(17,19,10);
  78         $rounds=64;
  79 }
  80
  81 $ctx="%rdi";    # 1st arg
  82 $round="%rdi";  # zaps $ctx
  83 $inp="%rsi";    # 2nd arg
  84 $Tbl="%rbp";
  85
  86 $_ctx="16*$SZ+0*8(%rsp)";
  87 $_inp="16*$SZ+1*8(%rsp)";
  88 $_end="16*$SZ+2*8(%rsp)";
  89 $_rsp="16*$SZ+3*8(%rsp)";
  90 $framesz="16*$SZ+4*8";
  91
  92
  93 sub ROUND_00_15()
  94 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
  95
  96 $code.=<<___;
  97         ror     \$`$Sigma1[2]-$Sigma1[1]`,$a0
  98         mov     $f,$a2
  99         mov     $T1,`$SZ*($i&0xf)`(%rsp)
 100
 101         ror     \$`$Sigma0[2]-$Sigma0[1]`,$a1
 102         xor     $e,$a0
 103         xor     $g,$a2                  # f^g
 104
 105         ror     \$`$Sigma1[1]-$Sigma1[0]`,$a0
 106         add     $h,$T1                  # T1+=h
 107         xor     $a,$a1
 108
 109         add     ($Tbl,$round,$SZ),$T1   # T1+=K[round]
 110         and     $e,$a2                  # (f^g)&e
 111         mov     $b,$h
 112
 113         ror     \$`$Sigma0[1]-$Sigma0[0]`,$a1
 114         xor     $e,$a0
 115         xor     $g,$a2                  # Ch(e,f,g)=((f^g)&e)^g
 116
 117         xor     $c,$h                   # b^c
 118         xor     $a,$a1
 119         add     $a2,$T1                 # T1+=Ch(e,f,g)
 120         mov     $b,$a2
 121
 122         ror     \$$Sigma1[0],$a0        # Sigma1(e)
 123         and     $a,$h                   # h=(b^c)&a
 124         and     $c,$a2                  # b&c
 125
 126         ror     \$$Sigma0[0],$a1        # Sigma0(a)
 127         add     $a0,$T1                 # T1+=Sigma1(e)
 128         add     $a2,$h                  # h+=b&c (completes +=Maj(a,b,c)
 129
 130         add     $T1,$d                  # d+=T1
 131         add     $T1,$h                  # h+=T1
 132         lea     1($round),$round        # round++
 133         add     $a1,$h                  # h+=Sigma0(a)
 134
 135 ___
 136 }
 137
 138 sub ROUND_16_XX()
 139 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 140
 141 $code.=<<___;
 142         mov     `$SZ*(($i+1)&0xf)`(%rsp),$a0
 143         mov     `$SZ*(($i+14)&0xf)`(%rsp),$a1
 144         mov     $a0,$T1
 145         mov     $a1,$a2
 146
 147         ror     \$`$sigma0[1]-$sigma0[0]`,$T1
 148         xor     $a0,$T1
 149         shr     \$$sigma0[2],$a0
 150
 151         ror     \$$sigma0[0],$T1
 152         xor     $T1,$a0                 # sigma0(X[(i+1)&0xf])
 153         mov     `$SZ*(($i+9)&0xf)`(%rsp),$T1
 154
 155         ror     \$`$sigma1[1]-$sigma1[0]`,$a2
 156         xor     $a1,$a2
 157         shr     \$$sigma1[2],$a1
 158
 159         ror     \$$sigma1[0],$a2
 160         add     $a0,$T1
 161         xor     $a2,$a1                 # sigma1(X[(i+14)&0xf])
 162
 163         add     `$SZ*($i&0xf)`(%rsp),$T1
 164         mov     $e,$a0
 165         add     $a1,$T1
 166         mov     $a,$a1
 167 ___
 168         &ROUND_00_15(@_);
 169 }
 170
 171 $code=<<___;
 172 .text
 173
 174 .globl  $func
 175 .type   $func,\@function,4
 176 .align  16
 177 $func:
 178         push    %rbx
 179         push    %rbp
 180         push    %r12
 181         push    %r13
 182         push    %r14
 183         push    %r15
 184         mov     %rsp,%r11               # copy %rsp
 185         shl     \$4,%rdx                # num*16
 186         sub     \$$framesz,%rsp
 187         lea     ($inp,%rdx,$SZ),%rdx    # inp+num*16*$SZ
 188         and     \$-64,%rsp              # align stack frame
 189         mov     $ctx,$_ctx              # save ctx, 1st arg
 190         mov     $inp,$_inp              # save inp, 2nd arh
 191         mov     %rdx,$_end              # save end pointer, "3rd" arg
 192         mov     %r11,$_rsp              # save copy of %rsp
 193 .Lprologue:
 194
 195         lea     $TABLE(%rip),$Tbl
 196
 197         mov     $SZ*0($ctx),$A
 198         mov     $SZ*1($ctx),$B
 199         mov     $SZ*2($ctx),$C
 200         mov     $SZ*3($ctx),$D
 201         mov     $SZ*4($ctx),$E
 202         mov     $SZ*5($ctx),$F
 203         mov     $SZ*6($ctx),$G
 204         mov     $SZ*7($ctx),$H
 205         jmp     .Lloop
 206
 207 .align  16
 208 .Lloop:
 209         xor     $round,$round
 210 ___
 211         for($i=0;$i<16;$i++) {
 212                 $code.="        mov     $SZ*$i($inp),$T1\n";
 213                 $code.="        mov     @ROT[4],$a0\n";
 214                 $code.="        mov     @ROT[0],$a1\n";
 215                 $code.="        bswap   $T1\n";
 216                 &ROUND_00_15($i,@ROT);
 217                 unshift(@ROT,pop(@ROT));
 218         }
 219 $code.=<<___;
 220         jmp     .Lrounds_16_xx
 221 .align  16
 222 .Lrounds_16_xx:
 223 ___
 224         for(;$i<32;$i++) {
 225                 &ROUND_16_XX($i,@ROT);
 226                 unshift(@ROT,pop(@ROT));
 227         }
 228
 229 $code.=<<___;
 230         cmp     \$$rounds,$round
 231         jb      .Lrounds_16_xx
 232
 233         mov     $_ctx,$ctx
 234         lea     16*$SZ($inp),$inp
 235
 236         add     $SZ*0($ctx),$A
 237         add     $SZ*1($ctx),$B
 238         add     $SZ*2($ctx),$C
 239         add     $SZ*3($ctx),$D
 240         add     $SZ*4($ctx),$E
 241         add     $SZ*5($ctx),$F
 242         add     $SZ*6($ctx),$G
 243         add     $SZ*7($ctx),$H
 244
 245         cmp     $_end,$inp
 246
 247         mov     $A,$SZ*0($ctx)
 248         mov     $B,$SZ*1($ctx)
 249         mov     $C,$SZ*2($ctx)
 250         mov     $D,$SZ*3($ctx)
 251         mov     $E,$SZ*4($ctx)
 252         mov     $F,$SZ*5($ctx)
 253         mov     $G,$SZ*6($ctx)
 254         mov     $H,$SZ*7($ctx)
 255         jb      .Lloop
 256
 257         mov     $_rsp,%rsi
 258         mov     (%rsi),%r15
 259         mov     8(%rsi),%r14
 260         mov     16(%rsi),%r13
 261         mov     24(%rsi),%r12
 262         mov     32(%rsi),%rbp
 263         mov     40(%rsi),%rbx
 264         lea     48(%rsi),%rsp
 265 .Lepilogue:
 266         ret
 267 .size   $func,.-$func
 268 ___
 269
 270 if ($SZ==4) {
 271 $code.=<<___;
 272 .align  64
 273 .type   $TABLE,\@object
 274 $TABLE:
 275         .long   0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
 276         .long   0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
 277         .long   0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
 278         .long   0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
 279         .long   0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
 280         .long   0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
 281         .long   0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
 282         .long   0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
 283         .long   0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
 284         .long   0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
 285         .long   0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
 286         .long   0xd192e819,0xd6990624,0xf40e3585,0x106aa070
 287         .long   0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
 288         .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
 289         .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
 290         .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
 291 ___
 292 } else {
 293 $code.=<<___;
 294 .align  64
 295 .type   $TABLE,\@object
 296 $TABLE:
 297         .quad   0x428a2f98d728ae22,0x7137449123ef65cd
 298         .quad   0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
 299         .quad   0x3956c25bf348b538,0x59f111f1b605d019
 300         .quad   0x923f82a4af194f9b,0xab1c5ed5da6d8118
 301         .quad   0xd807aa98a3030242,0x12835b0145706fbe
 302         .quad   0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
 303         .quad   0x72be5d74f27b896f,0x80deb1fe3b1696b1
 304         .quad   0x9bdc06a725c71235,0xc19bf174cf692694
 305         .quad   0xe49b69c19ef14ad2,0xefbe4786384f25e3
 306         .quad   0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
 307         .quad   0x2de92c6f592b0275,0x4a7484aa6ea6e483
 308         .quad   0x5cb0a9dcbd41fbd4,0x76f988da831153b5
 309         .quad   0x983e5152ee66dfab,0xa831c66d2db43210
 310         .quad   0xb00327c898fb213f,0xbf597fc7beef0ee4
 311         .quad   0xc6e00bf33da88fc2,0xd5a79147930aa725
 312         .quad   0x06ca6351e003826f,0x142929670a0e6e70
 313         .quad   0x27b70a8546d22ffc,0x2e1b21385c26c926
 314         .quad   0x4d2c6dfc5ac42aed,0x53380d139d95b3df
 315         .quad   0x650a73548baf63de,0x766a0abb3c77b2a8
 316         .quad   0x81c2c92e47edaee6,0x92722c851482353b
 317         .quad   0xa2bfe8a14cf10364,0xa81a664bbc423001
 318         .quad   0xc24b8b70d0f89791,0xc76c51a30654be30
 319         .quad   0xd192e819d6ef5218,0xd69906245565a910
 320         .quad   0xf40e35855771202a,0x106aa07032bbd1b8
 321         .quad   0x19a4c116b8d2d0c8,0x1e376c085141ab53
 322         .quad   0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
 323         .quad   0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
 324         .quad   0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
 325         .quad   0x748f82ee5defb2fc,0x78a5636f43172f60
 326         .quad   0x84c87814a1f0ab72,0x8cc702081a6439ec
 327         .quad   0x90befffa23631e28,0xa4506cebde82bde9
 328         .quad   0xbef9a3f7b2c67915,0xc67178f2e372532b
 329         .quad   0xca273eceea26619c,0xd186b8c721c0c207
 330         .quad   0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
 331         .quad   0x06f067aa72176fba,0x0a637dc5a2c898a6
 332         .quad   0x113f9804bef90dae,0x1b710b35131c471b
 333         .quad   0x28db77f523047d84,0x32caab7b40c72493
 334         .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
 335         .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
 336         .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
 337 ___
 338 }
 339
 340 $code =~ s/\`([^\`]*)\`/eval $1/gem;
 341 print $code;
 342 close STDOUT;