# This code is taken from the OpenSSL project but the author (Andy Polyakov) # has relicensed it under the GPLv2. Therefore this program is free software; # you can redistribute it and/or modify it under the terms of the GNU General # Public License version 2 as published by the Free Software Foundation. # # The original headers, including the original license headers, are # included below for completeness.
# Copyright 2014-2016 The OpenSSL Project Authors. All Rights Reserved. # # Licensed under the OpenSSL license (the "License"). You may not use # this file except in compliance with the License. You can obtain a copy # in the file LICENSE in the source distribution or at # https://www.openssl.org/source/license.html
# ==================================================================== # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL # project. The module is, however, dual licensed under OpenSSL and # CRYPTOGAMS licenses depending on where you obtain it. For further # details see http://www.openssl.org/~appro/cryptogams/. # ==================================================================== # # SHA256/512 for ARMv8. # # Performance in cycles per processed byte and improvement coefficient # over code generated with "default" compiler: # # SHA256-hw SHA256(*) SHA512 # Apple A7 1.97 10.5 (+33%) 6.73 (-1%(**)) # Cortex-A53 2.38 15.5 (+115%) 10.0 (+150%(***)) # Cortex-A57 2.31 11.6 (+86%) 7.51 (+260%(***)) # Denver 2.01 10.5 (+26%) 6.70 (+8%) # X-Gene 20.0 (+100%) 12.8 (+300%(***)) # Mongoose 2.36 13.0 (+50%) 8.36 (+33%) # # (*) Software SHA256 results are of lesser relevance, presented # mostly for informational purposes. # (**) The result is a trade-off: it's possible to improve it by # 10% (or by 1 cycle per round), but at the cost of 20% loss # on Cortex-A53 (or by 4 cycles per round). # (***) Super-impressive coefficients over gcc-generated code are # indication of some compiler "pathology", most notably code # generated with -mgeneral-regs-only is significantly faster # and the gap is only 40-90%. # # October 2016. # # Originally it was reckoned that it makes no sense to implement NEON # version of SHA256 for 64-bit processors. This is because performance # improvement on most wide-spread Cortex-A5x processors was observed # to be marginal, same on Cortex-A53 and ~10% on A57. But then it was # observed that 32-bit NEON SHA256 performs significantly better than # 64-bit scalar version on *some* of the more recent processors. As # result 64-bit NEON version of SHA256 was added to provide best # all-round performance. For example it executes ~30% faster on X-Gene # and Mongoose. [For reference, NEON version of SHA512 is bound to # deliver much less improvement, likely *negative* on Cortex-A5x. # Which is why NEON support is limited to SHA256.]
$output=pop;
$flavour=pop;
if ($flavour && $flavour ne "void") {
$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
die "can't locate arm-xlate.pl";
open OUT,"| \"$^X\" $xlate $flavour $output";
*STDOUT=*OUT;
} else {
open STDOUT,">$output";
}
sub BODY_00_xx {
my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
my $j=($i+1)&15;
my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
$T0=@X[$i+3] if ($i<11);
$code.=<<___ if ($i<16); #ifndef __AARCH64EB__
rev @X[$i],@X[$i] // $i #endif
___
$code.=<<___ if ($i<13 && ($i&1));
ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ
___
$code.=<<___ if ($i==13);
ldp @X[14],@X[15],[$inp]
___
$code.=<<___ if ($i>=14);
ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
___
$code.=<<___ if ($i>0 && $i<16);
add $a,$a,$t1 // h+=Sigma0(a)
___
$code.=<<___ if ($i>=11);
str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
___ # While ARMv8 specifies merged rotate-n-logical operation such as # 'eor x,y,z,ror#n', it was found to negatively affect performance # on Apple A7. The reason seems to be that it requires even 'y' to # be available earlier. This means that such merged instruction is # not necessarily best choice on critical path... On the other hand # Cortex-A5x handles merged instructions much better than disjoint # rotate and logical... See (**) footnote above.
$code.=<<___ if ($i<15);
ror $t0,$e,#$Sigma1[0]
add $h,$h,$t2 // h+=K[i]
eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
and $t1,$f,$e
bic $t2,$g,$e
add $h,$h,@X[$i&15] // h+=X[i]
orr $t1,$t1,$t2 // Ch(e,f,g)
eor $t2,$a,$b // a^b, b^c in next round
eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e)
ror $T0,$a,#$Sigma0[0]
add $h,$h,$t1 // h+=Ch(e,f,g)
eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
add $h,$h,$t0 // h+=Sigma1(e)
and $t3,$t3,$t2 // (b^c)&=(a^b)
add $d,$d,$h // d+=h
eor $t3,$t3,$b // Maj(a,b,c)
eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a)
add $h,$h,$t3 // h+=Maj(a,b,c)
ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round
//add $h,$h,$t1 // h+=Sigma0(a)
___
$code.=<<___ if ($i>=15);
ror $t0,$e,#$Sigma1[0]
add $h,$h,$t2 // h+=K[i]
ror $T1,@X[($j+1)&15],#$sigma0[0]
and $t1,$f,$e
ror $T2,@X[($j+14)&15],#$sigma1[0]
bic $t2,$g,$e
ror $T0,$a,#$Sigma0[0]
add $h,$h,@X[$i&15] // h+=X[i]
eor $t0,$t0,$e,ror#$Sigma1[1]
eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
orr $t1,$t1,$t2 // Ch(e,f,g)
eor $t2,$a,$b // a^b, b^c in next round
eor $t0,$t0,$e,ror#$Sigma1[2] // Sigma1(e)
eor $T0,$T0,$a,ror#$Sigma0[1]
add $h,$h,$t1 // h+=Ch(e,f,g)
and $t3,$t3,$t2 // (b^c)&=(a^b)
eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1])
add $h,$h,$t0 // h+=Sigma1(e)
eor $t3,$t3,$b // Maj(a,b,c)
eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a)
eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14])
add @X[$j],@X[$j],@X[($j+9)&15]
add $d,$d,$h // d+=h
add $h,$h,$t3 // h+=Maj(a,b,c)
ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round
add @X[$j],@X[$j],$T1
add $h,$h,$t1 // h+=Sigma0(a)
add @X[$j],@X[$j],$T2
___
($t2,$t3)=($t3,$t2);
}
$code.=<<___; #ifndef __KERNEL__ # include "arm_arch.h" #endif
my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
my @MSG=map("v$_.16b",(4..7));
my ($W0,$W1)=("v16.4s","v17.4s");
my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
ldr x29,[sp],#16
ret
.size sha256_block_armv8,.-sha256_block_armv8 #endif
___
}
if ($SZ==4) { ######################################### NEON stuff # # You'll surely note a lot of similarities with sha256-armv4 module, # and of course it's not a coincidence. sha256-armv4 was used as # initial template, but was adapted for ARMv8 instruction set and # extensively re-tuned for all-round performance.
my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
my $Ktbl="x16";
my $Xfer="x17";
my @X = map("q$_",(0..3));
my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
my $j=0;
sub AUTOLOAD() # thunk [simplified] x86-style perlasm
{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
my $arg = pop;
$arg = "#$arg" if ($arg*1 eq $arg);
$code .= "\t$opcode\t".join(',',@_,$arg)."\n";
}
sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
sub Dlo { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
sub Dhi { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
sub Xupdate()
{ use integer;
my $body = shift;
my @insns = (&$body,&$body,&$body,&$body);
my ($a,$b,$c,$d,$e,$f,$g,$h);
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.