diff options
author | Andy Polyakov <appro@openssl.org> | 2011-05-29 12:39:48 +0000 |
---|---|---|
committer | Andy Polyakov <appro@openssl.org> | 2011-05-29 12:39:48 +0000 |
commit | 0c149802a299ec7f448278ed90e53ac6c2dac88e (patch) | |
tree | 353c82d297d0bc06672d0a28fb534ce05b6870c7 /crypto/sha/asm/sha1-x86_64.pl | |
parent | c33066900ca7639797d8722ec9c7cc716ced5d0e (diff) | |
download | openssl-0c149802a299ec7f448278ed90e53ac6c2dac88e.zip openssl-0c149802a299ec7f448278ed90e53ac6c2dac88e.tar.gz openssl-0c149802a299ec7f448278ed90e53ac6c2dac88e.tar.bz2 |
sha1-586|x86_64.pl: add SSSE3 and AVX code paths.
Diffstat (limited to 'crypto/sha/asm/sha1-x86_64.pl')
-rwxr-xr-x | crypto/sha/asm/sha1-x86_64.pl | 919 |
1 files changed, 912 insertions, 7 deletions
diff --git a/crypto/sha/asm/sha1-x86_64.pl b/crypto/sha/asm/sha1-x86_64.pl index 35ab0e7..5414489 100755 --- a/crypto/sha/asm/sha1-x86_64.pl +++ b/crypto/sha/asm/sha1-x86_64.pl @@ -36,6 +36,31 @@ # instruction and the 'lea' instruction itself, which is essential # for Intel Atom core. +# October 2010. +# +# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it +# is to offload message schedule denoted by Wt in NIST specification, +# or Xupdate in OpenSSL source, to SIMD unit. See sha1-586.pl module +# for background and implementation details. The only difference from +# 32-bit code is that 64-bit code doesn't have to spill @X[] elements +# to free temporary registers. + +# April 2011. +# +# Add AVX code path. See sha1-586.pl for further information. + +###################################################################### +# Current performance is summarized in following table. Numbers are +# CPU clock cycles spent to process single byte (less is better). +# +# x86_64 SSSE3 AVX +# P4 9.8 - +# Opteron 6.6 - +# Core2 6.7 6.1/+10% - +# Atom 11.0 9.7/+13% - +# Westmere 7.1 5.6/+27% - +# Sandy Bridge 7.9 6.3/+25% 5.2/+51% + $flavour = shift; $output = shift; if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } @@ -47,6 +72,13 @@ $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or die "can't locate x86_64-xlate.pl"; +$avx=1 if (`$ENV{CC} -Wa,-V -c -o /dev/null -x assembler /dev/null 2>&1` + =~ /GNU assembler version ([2-9]\.[0-9]+)/ && + $1>=2.19); +$avx=1 if (!$avx && $flavour =~ /nasm/ && + `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ && + $1>=2.03); + open STDOUT,"| $^X $xlate $flavour $output"; $ctx="%rdi"; # 1st arg @@ -179,11 +211,29 @@ unshift(@xi,pop(@xi)); $code.=<<___; .text +.extern OPENSSL_ia32cap_P .globl sha1_block_data_order .type sha1_block_data_order,\@function,3 .align 16 sha1_block_data_order: + mov OPENSSL_ia32cap_P+0(%rip),%r9d + mov OPENSSL_ia32cap_P+4(%rip),%r8d + test \$`1<<9`,%r8d # check SSSE3 bit + jz .Lialu +___ +$code.=<<___ if ($avx); + and \$`1<<28`,%r8d # mask AVX bit + and \$`1<<30`,%r9d # mask "Intel CPU" bit + or %r9d,%r8d + cmp \$`1<<28|1<<30`,%r8d + je _avx_shortcut +___ +$code.=<<___; + jmp _ssse3_shortcut + +.align 16 +.Lialu: push %rbx push %rbp push %r12 @@ -202,8 +252,9 @@ sha1_block_data_order: mov 8($ctx),$C mov 12($ctx),$D mov 16($ctx),$E + jmp .Lloop -.align 4 +.align 16 .Lloop: ___ for($i=0;$i<20;$i++) { &BODY_00_19($i,@V); unshift(@V,pop(@V)); } @@ -235,9 +286,798 @@ $code.=<<___; .Lepilogue: ret .size sha1_block_data_order,.-sha1_block_data_order +___ +{{{ +my $Xi=4; +my @X=map("%xmm$_",(4..7,0..3)); +my @Tx=map("%xmm$_",(8..10)); +my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization +my @T=("%esi","%edi"); +my $j=0; +my $K_XX_XX="%r11"; -.asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>" +my $_rol=sub { &rol(@_) }; +my $_ror=sub { &ror(@_) }; + +$code.=<<___; +.type sha1_block_data_order_ssse3,\@function,3 .align 16 +sha1_block_data_order_ssse3: +_ssse3_shortcut: + push %rbx + push %rbp + push %r12 + lea `-64-($win64?5*16:0)`(%rsp),%rsp +___ +$code.=<<___ if ($win64); + movaps %xmm6,64+0(%rsp) + movaps %xmm7,64+16(%rsp) + movaps %xmm8,64+32(%rsp) + movaps %xmm9,64+48(%rsp) + movaps %xmm10,64+64(%rsp) +.Lprologue_ssse3: +___ +$code.=<<___; + mov %rdi,$ctx # reassigned argument + mov %rsi,$inp # reassigned argument + mov %rdx,$num # reassigned argument + + shl \$6,$num + add $inp,$num + lea K_XX_XX(%rip),$K_XX_XX + + mov 0($ctx),$A # load context + mov 4($ctx),$B + mov 8($ctx),$C + mov 12($ctx),$D + mov $B,@T[0] # magic seed + mov 16($ctx),$E + + movdqa 64($K_XX_XX),@X[2] # pbswap mask + movdqa 0($K_XX_XX),@Tx[1] # K_00_19 + movdqu 0($inp),@X[-4&7] # load input to %xmm[0-3] + movdqu 16($inp),@X[-3&7] + movdqu 32($inp),@X[-2&7] + movdqu 48($inp),@X[-1&7] + pshufb @X[2],@X[-4&7] # byte swap + add \$64,$inp + pshufb @X[2],@X[-3&7] + pshufb @X[2],@X[-2&7] + pshufb @X[2],@X[-1&7] + paddd @Tx[1],@X[-4&7] # add K_00_19 + paddd @Tx[1],@X[-3&7] + paddd @Tx[1],@X[-2&7] + movdqa @X[-4&7],0(%rsp) # X[]+K xfer to IALU + psubd @Tx[1],@X[-4&7] # restore X[] + movdqa @X[-3&7],16(%rsp) + psubd @Tx[1],@X[-3&7] + movdqa @X[-2&7],32(%rsp) + psubd @Tx[1],@X[-2&7] + jmp .Loop_ssse3 +___ + +sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm +{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; + my $arg = pop; + $arg = "\$$arg" if ($arg*1 eq $arg); + $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n"; +} + +sub Xupdate_ssse3_16_31() # recall that $Xi starts wtih 4 +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 40 instructions + my ($a,$b,$c,$d,$e); + + &movdqa (@X[0],@X[-3&7]); + eval(shift(@insns)); + eval(shift(@insns)); + &movdqa (@Tx[0],@X[-1&7]); + &palignr(@X[0],@X[-4&7],8); # compose "X[-14]" in "X[0]" + eval(shift(@insns)); + eval(shift(@insns)); + + &paddd (@Tx[1],@X[-1&7]); + eval(shift(@insns)); + eval(shift(@insns)); + &psrldq (@Tx[0],4); # "X[-3]", 3 dwords + eval(shift(@insns)); + eval(shift(@insns)); + &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]" + eval(shift(@insns)); + eval(shift(@insns)); + + &pxor (@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" + eval(shift(@insns)); + eval(shift(@insns)); + &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU + eval(shift(@insns)); + eval(shift(@insns)); + + &movdqa (@Tx[2],@X[0]); + &movdqa (@Tx[0],@X[0]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &pslldq (@Tx[2],12); # "X[0]"<<96, extract one dword + &paddd (@X[0],@X[0]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &psrld (@Tx[0],31); + eval(shift(@insns)); + eval(shift(@insns)); + &movdqa (@Tx[1],@Tx[2]); + eval(shift(@insns)); + eval(shift(@insns)); + + &psrld (@Tx[2],30); + &por (@X[0],@Tx[0]); # "X[0]"<<<=1 + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &pslld (@Tx[1],2); + &pxor (@X[0],@Tx[2]); + eval(shift(@insns)); + eval(shift(@insns)); + &movdqa (@Tx[2],eval(16*(($Xi)/5))."($K_XX_XX)"); # K_XX_XX + eval(shift(@insns)); + eval(shift(@insns)); + + &pxor (@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2 + + foreach (@insns) { eval; } # remaining instructions [if any] + + $Xi++; push(@X,shift(@X)); # "rotate" X[] + push(@Tx,shift(@Tx)); +} + +sub Xupdate_ssse3_32_79() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 to 48 instructions + my ($a,$b,$c,$d,$e); + + &movdqa (@Tx[0],@X[-1&7]) if ($Xi==8); + eval(shift(@insns)); # body_20_39 + &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" + &palignr(@Tx[0],@X[-2&7],8); # compose "X[-6]" + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + + &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]" + eval(shift(@insns)); + eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/); + if ($Xi%5) { + &movdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX... + } else { # ... or load next one + &movdqa (@Tx[2],eval(16*($Xi/5))."($K_XX_XX)"); + } + &paddd (@Tx[1],@X[-1&7]); + eval(shift(@insns)); # ror + eval(shift(@insns)); + + &pxor (@X[0],@Tx[0]); # "X[0]"^="X[-6]" + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + + &movdqa (@Tx[0],@X[0]); + &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # ror + eval(shift(@insns)); + + &pslld (@X[0],2); + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + &psrld (@Tx[0],30); + eval(shift(@insns)); + eval(shift(@insns)); # rol + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # ror + eval(shift(@insns)); + + &por (@X[0],@Tx[0]); # "X[0]"<<<=2 + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + &movdqa (@Tx[1],@X[0]) if ($Xi<19); + eval(shift(@insns)); + eval(shift(@insns)); # rol + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + eval(shift(@insns)); + + foreach (@insns) { eval; } # remaining instructions + + $Xi++; push(@X,shift(@X)); # "rotate" X[] + push(@Tx,shift(@Tx)); +} + +sub Xuplast_ssse3_80() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 instructions + my ($a,$b,$c,$d,$e); + + eval(shift(@insns)); + &paddd (@Tx[1],@X[-1&7]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU + + foreach (@insns) { eval; } # remaining instructions + + &cmp ($inp,$num); + &je (".Ldone_ssse3"); + + unshift(@Tx,pop(@Tx)); + + &movdqa (@X[2],"64($K_XX_XX)"); # pbswap mask + &movdqa (@Tx[1],"0($K_XX_XX)"); # K_00_19 + &movdqu (@X[-4&7],"0($inp)"); # load input + &movdqu (@X[-3&7],"16($inp)"); + &movdqu (@X[-2&7],"32($inp)"); + &movdqu (@X[-1&7],"48($inp)"); + &pshufb (@X[-4&7],@X[2]); # byte swap + &add ($inp,64); + + $Xi=0; +} + +sub Xloop_ssse3() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 instructions + my ($a,$b,$c,$d,$e); + + eval(shift(@insns)); + eval(shift(@insns)); + &pshufb (@X[($Xi-3)&7],@X[2]); + eval(shift(@insns)); + eval(shift(@insns)); + &paddd (@X[($Xi-4)&7],@Tx[1]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + &movdqa (eval(16*$Xi)."(%rsp)",@X[($Xi-4)&7]); # X[]+K xfer to IALU + eval(shift(@insns)); + eval(shift(@insns)); + &psubd (@X[($Xi-4)&7],@Tx[1]); + + foreach (@insns) { eval; } + $Xi++; +} + +sub Xtail_ssse3() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 instructions + my ($a,$b,$c,$d,$e); + + foreach (@insns) { eval; } +} + +sub body_00_19 () { + ( + '($a,$b,$c,$d,$e)=@V;'. + '&add ($e,eval(4*($j&15))."(%rsp)");', # X[]+K xfer + '&xor ($c,$d);', + '&mov (@T[1],$a);', # $b in next round + '&$_rol ($a,5);', + '&and (@T[0],$c);', # ($b&($c^$d)) + '&xor ($c,$d);', # restore $c + '&xor (@T[0],$d);', + '&add ($e,$a);', + '&$_ror ($b,$j?7:2);', # $b>>>2 + '&add ($e,@T[0]);' .'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));' + ); +} + +sub body_20_39 () { + ( + '($a,$b,$c,$d,$e)=@V;'. + '&add ($e,eval(4*($j++&15))."(%rsp)");', # X[]+K xfer + '&xor (@T[0],$d);', # ($b^$d) + '&mov (@T[1],$a);', # $b in next round + '&$_rol ($a,5);', + '&xor (@T[0],$c);', # ($b^$d^$c) + '&add ($e,$a);', + '&$_ror ($b,7);', # $b>>>2 + '&add ($e,@T[0]);' .'unshift(@V,pop(@V)); unshift(@T,pop(@T));' + ); +} + +sub body_40_59 () { + ( + '($a,$b,$c,$d,$e)=@V;'. + '&mov (@T[1],$c);', + '&xor ($c,$d);', + '&add ($e,eval(4*($j++&15))."(%rsp)");', # X[]+K xfer + '&and (@T[1],$d);', + '&and (@T[0],$c);', # ($b&($c^$d)) + '&$_ror ($b,7);', # $b>>>2 + '&add ($e,@T[1]);', + '&mov (@T[1],$a);', # $b in next round + '&$_rol ($a,5);', + '&add ($e,@T[0]);', + '&xor ($c,$d);', # restore $c + '&add ($e,$a);' .'unshift(@V,pop(@V)); unshift(@T,pop(@T));' + ); +} +$code.=<<___; +.align 16 +.Loop_ssse3: +___ + &Xupdate_ssse3_16_31(\&body_00_19); + &Xupdate_ssse3_16_31(\&body_00_19); + &Xupdate_ssse3_16_31(\&body_00_19); + &Xupdate_ssse3_16_31(\&body_00_19); + &Xupdate_ssse3_32_79(\&body_00_19); + &Xupdate_ssse3_32_79(\&body_20_39); + &Xupdate_ssse3_32_79(\&body_20_39); + &Xupdate_ssse3_32_79(\&body_20_39); + &Xupdate_ssse3_32_79(\&body_20_39); + &Xupdate_ssse3_32_79(\&body_20_39); + &Xupdate_ssse3_32_79(\&body_40_59); + &Xupdate_ssse3_32_79(\&body_40_59); + &Xupdate_ssse3_32_79(\&body_40_59); + &Xupdate_ssse3_32_79(\&body_40_59); + &Xupdate_ssse3_32_79(\&body_40_59); + &Xupdate_ssse3_32_79(\&body_20_39); + &Xuplast_ssse3_80(\&body_20_39); # can jump to "done" + + $saved_j=$j; @saved_V=@V; + + &Xloop_ssse3(\&body_20_39); + &Xloop_ssse3(\&body_20_39); + &Xloop_ssse3(\&body_20_39); + +$code.=<<___; + add 0($ctx),$A # update context + add 4($ctx),@T[0] + add 8($ctx),$C + add 12($ctx),$D + mov $A,0($ctx) + add 16($ctx),$E + mov @T[0],4($ctx) + mov @T[0],$B # magic seed + mov $C,8($ctx) + mov $D,12($ctx) + mov $E,16($ctx) + jmp .Loop_ssse3 + +.align 16 +.Ldone_ssse3: +___ + $j=$saved_j; @V=@saved_V; + + &Xtail_ssse3(\&body_20_39); + &Xtail_ssse3(\&body_20_39); + &Xtail_ssse3(\&body_20_39); + +$code.=<<___; + add 0($ctx),$A # update context + add 4($ctx),@T[0] + add 8($ctx),$C + mov $A,0($ctx) + add 12($ctx),$D + mov @T[0],4($ctx) + add 16($ctx),$E + mov $C,8($ctx) + mov $D,12($ctx) + mov $E,16($ctx) +___ +$code.=<<___ if ($win64); + movaps 64+0(%rsp),%xmm6 + movaps 64+16(%rsp),%xmm7 + movaps 64+32(%rsp),%xmm8 + movaps 64+48(%rsp),%xmm9 + movaps 64+64(%rsp),%xmm10 +___ +$code.=<<___; + lea `64+($win64?6*16:0)`(%rsp),%rsi + mov 0(%rsi),%r12 + mov 8(%rsi),%rbp + mov 16(%rsi),%rbx + lea 24(%rsi),%rsp +.Lepilogue_ssse3: + ret +.size sha1_block_data_order_ssse3,.-sha1_block_data_order_ssse3 +___ + +if ($avx) { +my $Xi=4; +my @X=map("%xmm$_",(4..7,0..3)); +my @Tx=map("%xmm$_",(8..10)); +my @V=($A,$B,$C,$D,$E)=("%eax","%ebx","%ecx","%edx","%ebp"); # size optimization +my @T=("%esi","%edi"); +my $j=0; +my $K_XX_XX="%r11"; + +my $_rol=sub { &shld(@_[0],@_) }; +my $_ror=sub { &shrd(@_[0],@_) }; + +$code.=<<___; +.type sha1_block_data_order_avx,\@function,3 +.align 16 +sha1_block_data_order_avx: +_avx_shortcut: + push %rbx + push %rbp + push %r12 + lea `-64-($win64?5*16:0)`(%rsp),%rsp +___ +$code.=<<___ if ($win64); + movaps %xmm6,64+0(%rsp) + movaps %xmm7,64+16(%rsp) + movaps %xmm8,64+32(%rsp) + movaps %xmm9,64+48(%rsp) + movaps %xmm10,64+64(%rsp) +.Lprologue_avx: +___ +$code.=<<___; + mov %rdi,$ctx # reassigned argument + mov %rsi,$inp # reassigned argument + mov %rdx,$num # reassigned argument + vzeroall + + shl \$6,$num + add $inp,$num + lea K_XX_XX(%rip),$K_XX_XX + + mov 0($ctx),$A # load context + mov 4($ctx),$B + mov 8($ctx),$C + mov 12($ctx),$D + mov $B,@T[0] # magic seed + mov 16($ctx),$E + + vmovdqa 64($K_XX_XX),@X[2] # pbswap mask + vmovdqa 0($K_XX_XX),@Tx[1] # K_00_19 + vmovdqu 0($inp),@X[-4&7] # load input to %xmm[0-3] + vmovdqu 16($inp),@X[-3&7] + vmovdqu 32($inp),@X[-2&7] + vmovdqu 48($inp),@X[-1&7] + vpshufb @X[2],@X[-4&7],@X[-4&7] # byte swap + add \$64,$inp + vpshufb @X[2],@X[-3&7],@X[-3&7] + vpshufb @X[2],@X[-2&7],@X[-2&7] + vpshufb @X[2],@X[-1&7],@X[-1&7] + vpaddd @Tx[1],@X[-4&7],@X[0] # add K_00_19 + vpaddd @Tx[1],@X[-3&7],@X[1] + vpaddd @Tx[1],@X[-2&7],@X[2] + vmovdqa @X[0],0(%rsp) # X[]+K xfer to IALU + vmovdqa @X[1],16(%rsp) + vmovdqa @X[2],32(%rsp) + jmp .Loop_avx +___ + +sub Xupdate_avx_16_31() # recall that $Xi starts wtih 4 +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 40 instructions + my ($a,$b,$c,$d,$e); + + eval(shift(@insns)); + eval(shift(@insns)); + &vpalignr(@X[0],@X[-3&7],@X[-4&7],8); # compose "X[-14]" in "X[0]" + eval(shift(@insns)); + eval(shift(@insns)); + + &vpaddd (@Tx[1],@Tx[1],@X[-1&7]); + eval(shift(@insns)); + eval(shift(@insns)); + &vpsrldq(@Tx[0],@X[-1&7],4); # "X[-3]", 3 dwords + eval(shift(@insns)); + eval(shift(@insns)); + &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]" + eval(shift(@insns)); + eval(shift(@insns)); + + &vpxor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]" + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-3]"^"X[-8]" + eval(shift(@insns)); + eval(shift(@insns)); + &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU + eval(shift(@insns)); + eval(shift(@insns)); + + &vpsrld (@Tx[0],@X[0],31); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &vpslldq(@Tx[2],@X[0],12); # "X[0]"<<96, extract one dword + &vpaddd (@X[0],@X[0],@X[0]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &vpsrld (@Tx[1],@Tx[2],30); + &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=1 + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &vpslld (@Tx[2],@Tx[2],2); + &vpxor (@X[0],@X[0],@Tx[1]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &vpxor (@X[0],@X[0],@Tx[2]); # "X[0]"^=("X[0]">>96)<<<2 + eval(shift(@insns)); + eval(shift(@insns)); + &vmovdqa (@Tx[2],eval(16*(($Xi)/5))."($K_XX_XX)"); # K_XX_XX + eval(shift(@insns)); + eval(shift(@insns)); + + + foreach (@insns) { eval; } # remaining instructions [if any] + + $Xi++; push(@X,shift(@X)); # "rotate" X[] + push(@Tx,shift(@Tx)); +} + +sub Xupdate_avx_32_79() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 to 48 instructions + my ($a,$b,$c,$d,$e); + + &vpalignr(@Tx[0],@X[-1&7],@X[-2&7],8); # compose "X[-6]" + &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]" + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + + &vpxor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]" + eval(shift(@insns)); + eval(shift(@insns)) if (@insns[0] !~ /&ro[rl]/); + if ($Xi%5) { + &vmovdqa (@Tx[2],@Tx[1]);# "perpetuate" K_XX_XX... + } else { # ... or load next one + &vmovdqa (@Tx[2],eval(16*($Xi/5))."($K_XX_XX)"); + } + &vpaddd (@Tx[1],@Tx[1],@X[-1&7]); + eval(shift(@insns)); # ror + eval(shift(@insns)); + + &vpxor (@X[0],@X[0],@Tx[0]); # "X[0]"^="X[-6]" + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + + &vpsrld (@Tx[0],@X[0],30); + &vmovdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer to IALU + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # ror + eval(shift(@insns)); + + &vpslld (@X[0],@X[0],2); + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # ror + eval(shift(@insns)); + + &vpor (@X[0],@X[0],@Tx[0]); # "X[0]"<<<=2 + eval(shift(@insns)); # body_20_39 + eval(shift(@insns)); + &vmovdqa (@Tx[1],@X[0]) if ($Xi<19); + eval(shift(@insns)); + eval(shift(@insns)); # rol + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); # rol + eval(shift(@insns)); + + foreach (@insns) { eval; } # remaining instructions + + $Xi++; push(@X,shift(@X)); # "rotate" X[] + push(@Tx,shift(@Tx)); +} + +sub Xuplast_avx_80() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 instructions + my ($a,$b,$c,$d,$e); + + eval(shift(@insns)); + &vpaddd (@Tx[1],@Tx[1],@X[-1&7]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + + &movdqa (eval(16*(($Xi-1)&3))."(%rsp)",@Tx[1]); # X[]+K xfer IALU + + foreach (@insns) { eval; } # remaining instructions + + &cmp ($inp,$num); + &je (".Ldone_avx"); + + unshift(@Tx,pop(@Tx)); + + &vmovdqa(@X[2],"64($K_XX_XX)"); # pbswap mask + &vmovdqa(@Tx[1],"0($K_XX_XX)"); # K_00_19 + &vmovdqu(@X[-4&7],"0($inp)"); # load input + &vmovdqu(@X[-3&7],"16($inp)"); + &vmovdqu(@X[-2&7],"32($inp)"); + &vmovdqu(@X[-1&7],"48($inp)"); + &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap + &add ($inp,64); + + $Xi=0; +} + +sub Xloop_avx() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 instructions + my ($a,$b,$c,$d,$e); + + eval(shift(@insns)); + eval(shift(@insns)); + &vpshufb(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]); + eval(shift(@insns)); + eval(shift(@insns)); + &vpaddd (@X[$Xi&7],@X[($Xi-4)&7],@Tx[1]); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + eval(shift(@insns)); + &vmovdqa(eval(16*$Xi)."(%rsp)",@X[$Xi&7]); # X[]+K xfer to IALU + eval(shift(@insns)); + eval(shift(@insns)); + + foreach (@insns) { eval; } + $Xi++; +} + +sub Xtail_avx() +{ use integer; + my $body = shift; + my @insns = (&$body,&$body,&$body,&$body); # 32 instructions + my ($a,$b,$c,$d,$e); + + foreach (@insns) { eval; } +} + +$code.=<<___; +.align 16 +.Loop_avx: +___ + &Xupdate_avx_16_31(\&body_00_19); + &Xupdate_avx_16_31(\&body_00_19); + &Xupdate_avx_16_31(\&body_00_19); + &Xupdate_avx_16_31(\&body_00_19); + &Xupdate_avx_32_79(\&body_00_19); + &Xupdate_avx_32_79(\&body_20_39); + &Xupdate_avx_32_79(\&body_20_39); + &Xupdate_avx_32_79(\&body_20_39); + &Xupdate_avx_32_79(\&body_20_39); + &Xupdate_avx_32_79(\&body_20_39); + &Xupdate_avx_32_79(\&body_40_59); + &Xupdate_avx_32_79(\&body_40_59); + &Xupdate_avx_32_79(\&body_40_59); + &Xupdate_avx_32_79(\&body_40_59); + &Xupdate_avx_32_79(\&body_40_59); + &Xupdate_avx_32_79(\&body_20_39); + &Xuplast_avx_80(\&body_20_39); # can jump to "done" + + $saved_j=$j; @saved_V=@V; + + &Xloop_avx(\&body_20_39); + &Xloop_avx(\&body_20_39); + &Xloop_avx(\&body_20_39); + +$code.=<<___; + add 0($ctx),$A # update context + add 4($ctx),@T[0] + add 8($ctx),$C + add 12($ctx),$D + mov $A,0($ctx) + add 16($ctx),$E + mov @T[0],4($ctx) + mov @T[0],$B # magic seed + mov $C,8($ctx) + mov $D,12($ctx) + mov $E,16($ctx) + jmp .Loop_avx + +.align 16 +.Ldone_avx: +___ + $j=$saved_j; @V=@saved_V; + + &Xtail_avx(\&body_20_39); + &Xtail_avx(\&body_20_39); + &Xtail_avx(\&body_20_39); + +$code.=<<___; + vzeroall + + add 0($ctx),$A # update context + add 4($ctx),@T[0] + add 8($ctx),$C + mov $A,0($ctx) + add 12($ctx),$D + mov @T[0],4($ctx) + add 16($ctx),$E + mov $C,8($ctx) + mov $D,12($ctx) + mov $E,16($ctx) +___ +$code.=<<___ if ($win64); + movaps 64+0(%rsp),%xmm6 + movaps 64+16(%rsp),%xmm7 + movaps 64+32(%rsp),%xmm8 + movaps 64+48(%rsp),%xmm9 + movaps 64+64(%rsp),%xmm10 +___ +$code.=<<___; + lea `64+($win64?6*16:0)`(%rsp),%rsi + mov 0(%rsi),%r12 + mov 8(%rsi),%rbp + mov 16(%rsi),%rbx + lea 24(%rsi),%rsp +.Lepilogue_avx: + ret +.size sha1_block_data_order_avx,.-sha1_block_data_order_avx +___ +} +$code.=<<___; +.align 64 +K_XX_XX: +.long 0x5a827999,0x5a827999,0x5a827999,0x5a827999 # K_00_19 +.long 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1 # K_20_39 +.long 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc # K_40_59 +.long 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6 # K_60_79 +.long 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f # pbswap mask +___ +}}} +$code.=<<___; +.asciz "SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>" +.align 64 ___ # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, @@ -269,13 +1109,13 @@ se_handler: lea .Lprologue(%rip),%r10 cmp %r10,%rbx # context->Rip<.Lprologue - jb .Lin_prologue + jb .Lcommon_seh_tail mov 152($context),%rax # pull context->Rsp lea .Lepilogue(%rip),%r10 cmp %r10,%rbx # context->Rip>=.Lepilogue - jae .Lin_prologue + jae .Lcommon_seh_tail mov `16*4`(%rax),%rax # pull saved stack pointer lea 32(%rax),%rax @@ -289,7 +1129,53 @@ se_handler: mov %r12,216($context) # restore context->R12 mov %r13,224($context) # restore context->R13 -.Lin_prologue: + jmp .Lcommon_seh_tail +.size se_handler,.-se_handler + +.type ssse3_handler,\@abi-omnipotent +.align 16 +ssse3_handler: + push %rsi + push %rdi + push %rbx + push %rbp + push %r12 + push %r13 + push %r14 + push %r15 + pushfq + sub \$64,%rsp + + mov 120($context),%rax # pull context->Rax + mov 248($context),%rbx # pull context->Rip + + mov 8($disp),%rsi # disp->ImageBase + mov 56($disp),%r11 # disp->HandlerData + + mov 0(%r11),%r10d # HandlerData[0] + lea (%rsi,%r10),%r10 # prologue label + cmp %r10,%rbx # context->Rip<prologue label + jb .Lcommon_seh_tail + + mov 152($context),%rax # pull context->Rsp + + mov 4(%r11),%r10d # HandlerData[1] + lea (%rsi,%r10),%r10 # epilogue label + cmp %r10,%rbx # context->Rip>=epilogue label + jae .Lcommon_seh_tail + + lea 64(%rax),%rsi + lea 512($context),%rdi # &context.Xmm6 + mov \$10,%ecx + .long 0xa548f3fc # cld; rep movsq + lea 24+5*16(%rax),%rax # adjust stack pointer + + mov -8(%rax),%rbx + mov -16(%rax),%rbp + mov %rbx,144($context) # restore context->Rbx + mov %rbp,160($context) # restore context->Rbp + +.Lcommon_seh_tail: mov 8(%rax),%rdi mov 16(%rax),%rsi mov %rax,152($context) # restore context->Rsp @@ -327,19 +1213,38 @@ se_handler: pop %rdi pop %rsi ret -.size se_handler,.-se_handler +.size ssse3_handler,.-ssse3_handler .section .pdata .align 4 .rva .LSEH_begin_sha1_block_data_order .rva .LSEH_end_sha1_block_data_order .rva .LSEH_info_sha1_block_data_order - + .rva .LSEH_begin_sha1_block_data_order_ssse3 + .rva .LSEH_end_sha1_block_data_order_ssse3 + .rva .LSEH_info_sha1_block_data_order_ssse3 +___ +$code.=<<___ if ($avx); + .rva .LSEH_begin_sha1_block_data_order_avx + .rva .LSEH_end_sha1_block_data_order_avx + .rva .LSEH_info_sha1_block_data_order_avx +___ +$code.=<<___; .section .xdata .align 8 .LSEH_info_sha1_block_data_order: .byte 9,0,0,0 .rva se_handler +.LSEH_info_sha1_block_data_order_ssse3: + .byte 9,0,0,0 + .rva ssse3_handler + .rva .Lprologue_ssse3,.Lepilogue_ssse3 # HandlerData[] +___ +$code.=<<___ if ($avx); +.LSEH_info_sha1_block_data_order_avx: + .byte 9,0,0,0 + .rva ssse3_handler + .rva .Lprologue_avx,.Lepilogue_avx # HandlerData[] ___ } |