-#!/usr/local/bin/perl
+#!/usr/bin/env perl
+# ====================================================================
+# [Re]written by Andy Polyakov <appro@fy.chalmers.se> 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/.
+# ====================================================================
+
+# "[Re]written" was achieved in two major overhauls. In 2004 BODY_*
+# functions were re-implemented to address P4 performance issue [see
+# commentary below], and in 2006 the rest was rewritten in order to
+# gain freedom to liberate licensing terms.
+
+# January, September 2004.
+#
# It was noted that Intel IA-32 C compiler generates code which
# performs ~30% *faster* on P4 CPU than original *hand-coded*
# SHA1 assembler implementation. To address this problem (and
# improvement on P4 outweights the loss and incorporate this
# re-tuned code to 0.9.7 and later.
# ----------------------------------------------------------------
-# Those who for any particular reason absolutely must score on
-# Pentium can replace this module with one from 0.9.6 distribution.
-# This "offer" shall be revoked the moment programming interface to
-# this module is changed, in which case this paragraph should be
-# removed.
-# ----------------------------------------------------------------
# <appro@fy.chalmers.se>
-$normal=0;
+# August 2009.
+#
+# George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as
+# '(c&d) + (b&(c^d))', which allows to accumulate partial results
+# and lighten "pressure" on scratch registers. This resulted in
+# >12% performance improvement on contemporary AMD cores (with no
+# degradation on other CPUs:-). Also, the code was revised to maximize
+# "distance" between instructions producing input to 'lea' instruction
+# and the 'lea' instruction itself, which is essential for Intel Atom
+# core and resulted in ~15% improvement.
-push(@INC,"perlasm","../../perlasm");
-require "x86asm.pl";
+# October 2010.
+#
+# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
+# is to offload message schedule denoted by Wt in NIST specification,
+# or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel,
+# and in SSE2 context was first explored by Dean Gaudet in 2004, see
+# http://arctic.org/~dean/crypto/sha1.html. Since then several things
+# have changed that made it interesting again:
+#
+# a) XMM units became faster and wider;
+# b) instruction set became more versatile;
+# c) an important observation was made by Max Locktykhin, which made
+# it possible to reduce amount of instructions required to perform
+# the operation in question, for further details see
+# http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/.
-&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
+# April 2011.
+#
+# Add AVX code path, probably most controversial... The thing is that
+# switch to AVX alone improves performance by as little as 4% in
+# comparison to SSSE3 code path. But below result doesn't look like
+# 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as
+# pair of µ-ops, and it's the additional µ-ops, two per round, that
+# make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded
+# as single µ-op by Sandy Bridge and it's replacing 'ro[rl]' with
+# equivalent 'sh[rl]d' that is responsible for the impressive 5.1
+# cycles per processed byte. But 'sh[rl]d' is not something that used
+# to be fast, nor does it appear to be fast in upcoming Bulldozer
+# [according to its optimization manual]. Which is why AVX code path
+# is guarded by *both* AVX and synthetic bit denoting Intel CPUs.
+# One can argue that it's unfair to AMD, but without 'sh[rl]d' it
+# makes no sense to keep the AVX code path. If somebody feels that
+# strongly, it's probably more appropriate to discuss possibility of
+# using vector rotate XOP on AMD...
-$A="eax";
-$B="ecx";
-$C="ebx";
-$D="edx";
-$E="edi";
-$T="esi";
-$tmp1="ebp";
+######################################################################
+# Current performance is summarized in following table. Numbers are
+# CPU clock cycles spent to process single byte (less is better).
+#
+# x86 SSSE3 AVX
+# Pentium 15.7 -
+# PIII 11.5 -
+# P4 10.6 -
+# AMD K8 7.1 -
+# Core2 7.3 6.1/+20% -
+# Atom 12.5 9.5(*)/+32% -
+# Westmere 7.3 5.6/+30% -
+# Sandy Bridge 8.8 6.2/+40% 5.1(**)/+70%
+#
+# (*) Loop is 1056 instructions long and expected result is ~8.25.
+# It remains mystery [to me] why ILP is limited to 1.7.
+#
+# (**) As per above comment, the result is for AVX *plus* sh[rl]d.
-$off=9*4;
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+push(@INC,"${dir}","${dir}../../perlasm");
+require "x86asm.pl";
-@K=(0x5a827999,0x6ed9eba1,0x8f1bbcdc,0xca62c1d6);
+&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
-&sha1_block_data("sha1_block_asm_data_order");
+$xmm=$ymm=0;
+for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
-&asm_finish();
+$ymm=1 if ($xmm &&
+ `$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
+ =~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
+ $1>=2.19); # first version supporting AVX
-sub Nn
- {
- local($p)=@_;
- local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
- return($n{$p});
- }
+$ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
+ `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
+ $1>=2.03); # first version supporting AVX
-sub Np
- {
- local($p)=@_;
- local(%n)=($A,$T,$B,$A,$C,$B,$D,$C,$E,$D,$T,$E);
- local(%n)=($A,$B,$B,$C,$C,$D,$D,$E,$E,$T,$T,$A);
- return($n{$p});
- }
-
-sub Na
- {
- local($n)=@_;
- return( (($n )&0x0f),
- (($n+ 2)&0x0f),
- (($n+ 8)&0x0f),
- (($n+13)&0x0f),
- (($n+ 1)&0x0f));
- }
+&external_label("OPENSSL_ia32cap_P") if ($xmm);
-sub X_expand
- {
- local($in)=@_;
- &comment("First, load the words onto the stack in network byte order");
- for ($i=0; $i<16; $i+=2)
- {
- &mov($A,&DWP(($i+0)*4,$in,"",0));# unless $i == 0;
- &mov($B,&DWP(($i+1)*4,$in,"",0));
- &bswap($A);
- &bswap($B);
- &mov(&swtmp($i+0),$A);
- &mov(&swtmp($i+1),$B);
- }
+$A="eax";
+$B="ebx";
+$C="ecx";
+$D="edx";
+$E="edi";
+$T="esi";
+$tmp1="ebp";
- &comment("We now have the X array on the stack");
- &comment("starting at sp-4");
- }
+@V=($A,$B,$C,$D,$E,$T);
-# Rules of engagement
-# F is always trashable at the start, the running total.
-# E becomes the next F so it can be trashed after it has been 'accumulated'
-# F becomes A in the next round. We don't need to access it much.
-# During the X update part, the result ends up in $X[$n0].
+$alt=0; # 1 denotes alternative IALU implementation, which performs
+ # 8% *worse* on P4, same on Westmere and Atom, 2% better on
+ # Sandy Bridge...
sub BODY_00_15
{
- local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
+ local($n,$a,$b,$c,$d,$e,$f)=@_;
&comment("00_15 $n");
else { &mov($a,$tmp1); }
&rotl($tmp1,5); # tmp1=ROTATE(a,5)
&xor($f,$d);
+ &add($tmp1,$e); # tmp1+=e;
+ &mov($e,&swtmp($n%16)); # e becomes volatile and is loaded
+ # with xi, also note that e becomes
+ # f in next round...
&and($f,$b);
- &add($tmp1,$e); # tmp1+=e;
- &mov($e,&swtmp($n)); # e becomes volatile and
- # is loaded with xi
- &xor($f,$d); # f holds F_00_19(b,c,d)
&rotr($b,2); # b=ROTATE(b,30)
- &lea($tmp1,&DWP($K,$tmp1,$e,1));# tmp1+=K_00_19+xi
+ &xor($f,$d); # f holds F_00_19(b,c,d)
+ &lea($tmp1,&DWP(0x5a827999,$tmp1,$e)); # tmp1+=K_00_19+xi
- if ($n==15) { &add($f,$tmp1); } # f+=tmp1
- else { &add($tmp1,$f); }
+ if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
+ &add($f,$tmp1); } # f+=tmp1
+ else { &add($tmp1,$f); } # f becomes a in next round
+ &mov($tmp1,$a) if ($alt && $n==15);
}
sub BODY_16_19
{
- local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
- local($n0,$n1,$n2,$n3,$np)=&Na($n);
+ local($n,$a,$b,$c,$d,$e,$f)=@_;
&comment("16_19 $n");
- &mov($f,&swtmp($n1)); # f to hold Xupdate(xi,xa,xb,xc,xd)
- &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d)
- &xor($f,&swtmp($n0));
- &xor($tmp1,$d);
- &xor($f,&swtmp($n2));
- &and($tmp1,$b); # tmp1 holds F_00_19(b,c,d)
- &rotr($b,2); # b=ROTATE(b,30)
- &xor($f,&swtmp($n3)); # f holds xa^xb^xc^xd
- &rotl($f,1); # f=ROATE(f,1)
+if ($alt) {
+ &xor($c,$d);
+ &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
+ &and($tmp1,$c); # tmp1 to hold F_00_19(b,c,d), b&=c^d
+ &xor($f,&swtmp(($n+8)%16));
+ &xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
+ &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
+ &rotl($f,1); # f=ROTATE(f,1)
+ &add($e,$tmp1); # e+=F_00_19(b,c,d)
+ &xor($c,$d); # restore $c
+ &mov($tmp1,$a); # b in next round
+ &rotr($b,$n==16?2:7); # b=ROTATE(b,30)
+ &mov(&swtmp($n%16),$f); # xi=f
+ &rotl($a,5); # ROTATE(a,5)
+ &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
+ &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
+ &add($f,$a); # f+=ROTATE(a,5)
+} else {
+ &mov($tmp1,$c); # tmp1 to hold F_00_19(b,c,d)
+ &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
+ &xor($tmp1,$d);
+ &xor($f,&swtmp(($n+8)%16));
+ &and($tmp1,$b);
+ &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
+ &rotl($f,1); # f=ROTATE(f,1)
&xor($tmp1,$d); # tmp1=F_00_19(b,c,d)
- &mov(&swtmp($n0),$f); # xi=f
- &lea($f,&DWP($K,$f,$e,1)); # f+=K_00_19+e
- &mov($e,$a); # e becomes volatile
- &rotl($e,5); # e=ROTATE(a,5)
- &add($f,$tmp1); # f+=F_00_19(b,c,d)
- &add($f,$e); # f+=ROTATE(a,5)
+ &add($e,$tmp1); # e+=F_00_19(b,c,d)
+ &mov($tmp1,$a);
+ &rotr($b,2); # b=ROTATE(b,30)
+ &mov(&swtmp($n%16),$f); # xi=f
+ &rotl($tmp1,5); # ROTATE(a,5)
+ &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
+ &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
+ &add($f,$tmp1); # f+=ROTATE(a,5)
+}
}
sub BODY_20_39
{
- local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
+ local($n,$a,$b,$c,$d,$e,$f)=@_;
+ local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
&comment("20_39 $n");
- local($n0,$n1,$n2,$n3,$np)=&Na($n);
+if ($alt) {
+ &xor($tmp1,$c); # tmp1 to hold F_20_39(b,c,d), b^=c
+ &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
+ &xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
+ &xor($f,&swtmp(($n+8)%16));
+ &add($e,$tmp1); # e+=F_20_39(b,c,d)
+ &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
+ &rotl($f,1); # f=ROTATE(f,1)
+ &mov($tmp1,$a); # b in next round
+ &rotr($b,7); # b=ROTATE(b,30)
+ &mov(&swtmp($n%16),$f) if($n<77);# xi=f
+ &rotl($a,5); # ROTATE(a,5)
+ &xor($b,$c) if($n==39);# warm up for BODY_40_59
+ &and($tmp1,$b) if($n==39);
+ &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
+ &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
+ &add($f,$a); # f+=ROTATE(a,5)
+ &rotr($a,5) if ($n==79);
+} else {
&mov($tmp1,$b); # tmp1 to hold F_20_39(b,c,d)
- &mov($f,&swtmp($n0)); # f to hold Xupdate(xi,xa,xb,xc,xd)
- &rotr($b,2); # b=ROTATE(b,30)
- &xor($f,&swtmp($n1));
+ &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
&xor($tmp1,$c);
- &xor($f,&swtmp($n2));
+ &xor($f,&swtmp(($n+8)%16));
&xor($tmp1,$d); # tmp1 holds F_20_39(b,c,d)
- &xor($f,&swtmp($n3)); # f holds xa^xb^xc^xd
+ &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
&rotl($f,1); # f=ROTATE(f,1)
- &add($tmp1,$e);
- &mov(&swtmp($n0),$f); # xi=f
- &mov($e,$a); # e becomes volatile
- &rotl($e,5); # e=ROTATE(a,5)
- &lea($f,&DWP($K,$f,$tmp1,1)); # f+=K_20_39+e
- &add($f,$e); # f+=ROTATE(a,5)
+ &add($e,$tmp1); # e+=F_20_39(b,c,d)
+ &rotr($b,2); # b=ROTATE(b,30)
+ &mov($tmp1,$a);
+ &rotl($tmp1,5); # ROTATE(a,5)
+ &mov(&swtmp($n%16),$f) if($n<77);# xi=f
+ &lea($f,&DWP($K,$f,$e)); # f+=e+K_XX_YY
+ &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
+ &add($f,$tmp1); # f+=ROTATE(a,5)
+}
}
sub BODY_40_59
{
- local($pos,$K,$X,$n,$a,$b,$c,$d,$e,$f)=@_;
+ local($n,$a,$b,$c,$d,$e,$f)=@_;
&comment("40_59 $n");
- local($n0,$n1,$n2,$n3,$np)=&Na($n);
-
- &mov($f,&swtmp($n0)); # f to hold Xupdate(xi,xa,xb,xc,xd)
- &mov($tmp1,&swtmp($n1));
- &xor($f,$tmp1);
- &mov($tmp1,&swtmp($n2));
- &xor($f,$tmp1);
- &mov($tmp1,&swtmp($n3));
- &xor($f,$tmp1); # f holds xa^xb^xc^xd
- &mov($tmp1,$b); # tmp1 to hold F_40_59(b,c,d)
+
+if ($alt) {
+ &add($e,$tmp1); # e+=b&(c^d)
+ &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
+ &mov($tmp1,$d);
+ &xor($f,&swtmp(($n+8)%16));
+ &xor($c,$d); # restore $c
+ &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
&rotl($f,1); # f=ROTATE(f,1)
- &or($tmp1,$c);
- &mov(&swtmp($n0),$f); # xi=f
- &and($tmp1,$d);
- &lea($f,&DWP($K,$f,$e,1)); # f+=K_40_59+e
- &mov($e,$b); # e becomes volatile and is used
- # to calculate F_40_59(b,c,d)
+ &and($tmp1,$c);
+ &rotr($b,7); # b=ROTATE(b,30)
+ &add($e,$tmp1); # e+=c&d
+ &mov($tmp1,$a); # b in next round
+ &mov(&swtmp($n%16),$f); # xi=f
+ &rotl($a,5); # ROTATE(a,5)
+ &xor($b,$c) if ($n<59);
+ &and($tmp1,$b) if ($n<59);# tmp1 to hold F_40_59(b,c,d)
+ &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
+ &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
+ &add($f,$a); # f+=ROTATE(a,5)
+} else {
+ &mov($tmp1,$c); # tmp1 to hold F_40_59(b,c,d)
+ &xor($f,&swtmp(($n+2)%16)); # f to hold Xupdate(xi,xa,xb,xc,xd)
+ &xor($tmp1,$d);
+ &xor($f,&swtmp(($n+8)%16));
+ &and($tmp1,$b);
+ &xor($f,&swtmp(($n+13)%16)); # f holds xa^xb^xc^xd
+ &rotl($f,1); # f=ROTATE(f,1)
+ &add($tmp1,$e); # b&(c^d)+=e
&rotr($b,2); # b=ROTATE(b,30)
- &and($e,$c);
- &or($tmp1,$e); # tmp1 holds F_40_59(b,c,d)
- &mov($e,$a);
- &rotl($e,5); # e=ROTATE(a,5)
- &add($f,$tmp1); # f+=tmp1;
+ &mov($e,$a); # e becomes volatile
+ &rotl($e,5); # ROTATE(a,5)
+ &mov(&swtmp($n%16),$f); # xi=f
+ &lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
+ &mov($tmp1,$c);
&add($f,$e); # f+=ROTATE(a,5)
+ &and($tmp1,$d);
+ &mov($e,&swtmp(($n+1)%16)); # pre-fetch f for next round
+ &add($f,$tmp1); # f+=c&d
+}
}
-sub BODY_60_79
- {
- &BODY_20_39(@_);
+&function_begin("sha1_block_data_order");
+if ($xmm) {
+ &static_label("ssse3_shortcut");
+ &static_label("avx_shortcut") if ($ymm);
+ &static_label("K_XX_XX");
+
+ &call (&label("pic_point")); # make it PIC!
+ &set_label("pic_point");
+ &blindpop($tmp1);
+ &picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
+ &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
+
+ &mov ($A,&DWP(0,$T));
+ &mov ($D,&DWP(4,$T));
+ &test ($D,1<<9); # check SSSE3 bit
+ &jz (&label("x86"));
+ &test ($A,1<<24); # check FXSR bit
+ &jz (&label("x86"));
+ if ($ymm) {
+ &and ($D,1<<28); # mask AVX bit
+ &and ($A,1<<30); # mask "Intel CPU" bit
+ &or ($A,$D);
+ &cmp ($A,1<<28|1<<30);
+ &je (&label("avx_shortcut"));
}
+ &jmp (&label("ssse3_shortcut"));
+ &set_label("x86",16);
+}
+ &mov($tmp1,&wparam(0)); # SHA_CTX *c
+ &mov($T,&wparam(1)); # const void *input
+ &mov($A,&wparam(2)); # size_t num
+ &stack_push(16+3); # allocate X[16]
+ &shl($A,6);
+ &add($A,$T);
+ &mov(&wparam(2),$A); # pointer beyond the end of input
+ &mov($E,&DWP(16,$tmp1));# pre-load E
+ &jmp(&label("loop"));
-sub sha1_block_host
- {
- local($name, $sclabel)=@_;
-
- &function_begin_B($name,"");
-
- # parameter 1 is the MD5_CTX structure.
- # A 0
- # B 4
- # C 8
- # D 12
- # E 16
-
- &mov("ecx", &wparam(2));
- &push("esi");
- &shl("ecx",6);
- &mov("esi", &wparam(1));
- &push("ebp");
- &add("ecx","esi"); # offset to leave on
- &push("ebx");
- &mov("ebp", &wparam(0));
- &push("edi");
- &mov($D, &DWP(12,"ebp","",0));
- &stack_push(18+9);
- &mov($E, &DWP(16,"ebp","",0));
- &mov($C, &DWP( 8,"ebp","",0));
- &mov(&swtmp(17),"ecx");
-
- &comment("First we need to setup the X array");
-
- for ($i=0; $i<16; $i+=2)
+&set_label("loop",16);
+
+ # copy input chunk to X, but reversing byte order!
+ for ($i=0; $i<16; $i+=4)
{
- &mov($A,&DWP(($i+0)*4,"esi","",0));# unless $i == 0;
- &mov($B,&DWP(($i+1)*4,"esi","",0));
+ &mov($A,&DWP(4*($i+0),$T));
+ &mov($B,&DWP(4*($i+1),$T));
+ &mov($C,&DWP(4*($i+2),$T));
+ &mov($D,&DWP(4*($i+3),$T));
+ &bswap($A);
+ &bswap($B);
+ &bswap($C);
+ &bswap($D);
&mov(&swtmp($i+0),$A);
- &mov(&swtmp($i+1),$B);
+ &mov(&swtmp($i+1),$B);
+ &mov(&swtmp($i+2),$C);
+ &mov(&swtmp($i+3),$D);
}
- &jmp($sclabel);
- &function_end_B($name);
- }
+ &mov(&wparam(1),$T); # redundant in 1st spin
+ &mov($A,&DWP(0,$tmp1)); # load SHA_CTX
+ &mov($B,&DWP(4,$tmp1));
+ &mov($C,&DWP(8,$tmp1));
+ &mov($D,&DWP(12,$tmp1));
+ # E is pre-loaded
-sub sha1_block_data
- {
- local($name)=@_;
-
- &function_begin_B($name,"");
-
- # parameter 1 is the MD5_CTX structure.
- # A 0
- # B 4
- # C 8
- # D 12
- # E 16
-
- &mov("ecx", &wparam(2));
- &push("esi");
- &shl("ecx",6);
- &mov("esi", &wparam(1));
- &push("ebp");
- &add("ecx","esi"); # offset to leave on
- &push("ebx");
- &mov("ebp", &wparam(0));
- &push("edi");
- &mov($D, &DWP(12,"ebp","",0));
- &stack_push(18+9);
- &mov($E, &DWP(16,"ebp","",0));
- &mov($C, &DWP( 8,"ebp","",0));
- &mov(&swtmp(17),"ecx");
-
- &comment("First we need to setup the X array");
-
- &set_label("start") unless $normal;
-
- &X_expand("esi");
- &mov(&wparam(1),"esi");
-
- &set_label("shortcut", 0, 1);
- &comment("");
- &comment("Start processing");
-
- # odd start
- &mov($A, &DWP( 0,"ebp","",0));
- &mov($B, &DWP( 4,"ebp","",0));
- $X="esp";
- &BODY_00_15(-2,$K[0],$X, 0,$A,$B,$C,$D,$E,$T);
- &BODY_00_15( 0,$K[0],$X, 1,$T,$A,$B,$C,$D,$E);
- &BODY_00_15( 0,$K[0],$X, 2,$E,$T,$A,$B,$C,$D);
- &BODY_00_15( 0,$K[0],$X, 3,$D,$E,$T,$A,$B,$C);
- &BODY_00_15( 0,$K[0],$X, 4,$C,$D,$E,$T,$A,$B);
- &BODY_00_15( 0,$K[0],$X, 5,$B,$C,$D,$E,$T,$A);
- &BODY_00_15( 0,$K[0],$X, 6,$A,$B,$C,$D,$E,$T);
- &BODY_00_15( 0,$K[0],$X, 7,$T,$A,$B,$C,$D,$E);
- &BODY_00_15( 0,$K[0],$X, 8,$E,$T,$A,$B,$C,$D);
- &BODY_00_15( 0,$K[0],$X, 9,$D,$E,$T,$A,$B,$C);
- &BODY_00_15( 0,$K[0],$X,10,$C,$D,$E,$T,$A,$B);
- &BODY_00_15( 0,$K[0],$X,11,$B,$C,$D,$E,$T,$A);
- &BODY_00_15( 0,$K[0],$X,12,$A,$B,$C,$D,$E,$T);
- &BODY_00_15( 0,$K[0],$X,13,$T,$A,$B,$C,$D,$E);
- &BODY_00_15( 0,$K[0],$X,14,$E,$T,$A,$B,$C,$D);
- &BODY_00_15( 1,$K[0],$X,15,$D,$E,$T,$A,$B,$C);
- &BODY_16_19(-1,$K[0],$X,16,$C,$D,$E,$T,$A,$B);
- &BODY_16_19( 0,$K[0],$X,17,$B,$C,$D,$E,$T,$A);
- &BODY_16_19( 0,$K[0],$X,18,$A,$B,$C,$D,$E,$T);
- &BODY_16_19( 1,$K[0],$X,19,$T,$A,$B,$C,$D,$E);
-
- &BODY_20_39(-1,$K[1],$X,20,$E,$T,$A,$B,$C,$D);
- &BODY_20_39( 0,$K[1],$X,21,$D,$E,$T,$A,$B,$C);
- &BODY_20_39( 0,$K[1],$X,22,$C,$D,$E,$T,$A,$B);
- &BODY_20_39( 0,$K[1],$X,23,$B,$C,$D,$E,$T,$A);
- &BODY_20_39( 0,$K[1],$X,24,$A,$B,$C,$D,$E,$T);
- &BODY_20_39( 0,$K[1],$X,25,$T,$A,$B,$C,$D,$E);
- &BODY_20_39( 0,$K[1],$X,26,$E,$T,$A,$B,$C,$D);
- &BODY_20_39( 0,$K[1],$X,27,$D,$E,$T,$A,$B,$C);
- &BODY_20_39( 0,$K[1],$X,28,$C,$D,$E,$T,$A,$B);
- &BODY_20_39( 0,$K[1],$X,29,$B,$C,$D,$E,$T,$A);
- &BODY_20_39( 0,$K[1],$X,30,$A,$B,$C,$D,$E,$T);
- &BODY_20_39( 0,$K[1],$X,31,$T,$A,$B,$C,$D,$E);
- &BODY_20_39( 0,$K[1],$X,32,$E,$T,$A,$B,$C,$D);
- &BODY_20_39( 0,$K[1],$X,33,$D,$E,$T,$A,$B,$C);
- &BODY_20_39( 0,$K[1],$X,34,$C,$D,$E,$T,$A,$B);
- &BODY_20_39( 0,$K[1],$X,35,$B,$C,$D,$E,$T,$A);
- &BODY_20_39( 0,$K[1],$X,36,$A,$B,$C,$D,$E,$T);
- &BODY_20_39( 0,$K[1],$X,37,$T,$A,$B,$C,$D,$E);
- &BODY_20_39( 0,$K[1],$X,38,$E,$T,$A,$B,$C,$D);
- &BODY_20_39( 1,$K[1],$X,39,$D,$E,$T,$A,$B,$C);
-
- &BODY_40_59(-1,$K[2],$X,40,$C,$D,$E,$T,$A,$B);
- &BODY_40_59( 0,$K[2],$X,41,$B,$C,$D,$E,$T,$A);
- &BODY_40_59( 0,$K[2],$X,42,$A,$B,$C,$D,$E,$T);
- &BODY_40_59( 0,$K[2],$X,43,$T,$A,$B,$C,$D,$E);
- &BODY_40_59( 0,$K[2],$X,44,$E,$T,$A,$B,$C,$D);
- &BODY_40_59( 0,$K[2],$X,45,$D,$E,$T,$A,$B,$C);
- &BODY_40_59( 0,$K[2],$X,46,$C,$D,$E,$T,$A,$B);
- &BODY_40_59( 0,$K[2],$X,47,$B,$C,$D,$E,$T,$A);
- &BODY_40_59( 0,$K[2],$X,48,$A,$B,$C,$D,$E,$T);
- &BODY_40_59( 0,$K[2],$X,49,$T,$A,$B,$C,$D,$E);
- &BODY_40_59( 0,$K[2],$X,50,$E,$T,$A,$B,$C,$D);
- &BODY_40_59( 0,$K[2],$X,51,$D,$E,$T,$A,$B,$C);
- &BODY_40_59( 0,$K[2],$X,52,$C,$D,$E,$T,$A,$B);
- &BODY_40_59( 0,$K[2],$X,53,$B,$C,$D,$E,$T,$A);
- &BODY_40_59( 0,$K[2],$X,54,$A,$B,$C,$D,$E,$T);
- &BODY_40_59( 0,$K[2],$X,55,$T,$A,$B,$C,$D,$E);
- &BODY_40_59( 0,$K[2],$X,56,$E,$T,$A,$B,$C,$D);
- &BODY_40_59( 0,$K[2],$X,57,$D,$E,$T,$A,$B,$C);
- &BODY_40_59( 0,$K[2],$X,58,$C,$D,$E,$T,$A,$B);
- &BODY_40_59( 1,$K[2],$X,59,$B,$C,$D,$E,$T,$A);
-
- &BODY_60_79(-1,$K[3],$X,60,$A,$B,$C,$D,$E,$T);
- &BODY_60_79( 0,$K[3],$X,61,$T,$A,$B,$C,$D,$E);
- &BODY_60_79( 0,$K[3],$X,62,$E,$T,$A,$B,$C,$D);
- &BODY_60_79( 0,$K[3],$X,63,$D,$E,$T,$A,$B,$C);
- &BODY_60_79( 0,$K[3],$X,64,$C,$D,$E,$T,$A,$B);
- &BODY_60_79( 0,$K[3],$X,65,$B,$C,$D,$E,$T,$A);
- &BODY_60_79( 0,$K[3],$X,66,$A,$B,$C,$D,$E,$T);
- &BODY_60_79( 0,$K[3],$X,67,$T,$A,$B,$C,$D,$E);
- &BODY_60_79( 0,$K[3],$X,68,$E,$T,$A,$B,$C,$D);
- &BODY_60_79( 0,$K[3],$X,69,$D,$E,$T,$A,$B,$C);
- &BODY_60_79( 0,$K[3],$X,70,$C,$D,$E,$T,$A,$B);
- &BODY_60_79( 0,$K[3],$X,71,$B,$C,$D,$E,$T,$A);
- &BODY_60_79( 0,$K[3],$X,72,$A,$B,$C,$D,$E,$T);
- &BODY_60_79( 0,$K[3],$X,73,$T,$A,$B,$C,$D,$E);
- &BODY_60_79( 0,$K[3],$X,74,$E,$T,$A,$B,$C,$D);
- &BODY_60_79( 0,$K[3],$X,75,$D,$E,$T,$A,$B,$C);
- &BODY_60_79( 0,$K[3],$X,76,$C,$D,$E,$T,$A,$B);
- &BODY_60_79( 0,$K[3],$X,77,$B,$C,$D,$E,$T,$A);
- &BODY_60_79( 0,$K[3],$X,78,$A,$B,$C,$D,$E,$T);
- &BODY_60_79( 2,$K[3],$X,79,$T,$A,$B,$C,$D,$E);
-
- &comment("End processing");
- &comment("");
- # D is the tmp value
-
- # E -> A
- # T -> B
- # A -> C
- # B -> D
- # C -> E
- # D -> T
-
- &mov($tmp1,&wparam(0));
-
- &mov($D, &DWP(12,$tmp1,"",0));
- &add($D,$B);
- &mov($B, &DWP( 4,$tmp1,"",0));
- &add($B,$T);
- &mov($T, $A);
- &mov($A, &DWP( 0,$tmp1,"",0));
- &mov(&DWP(12,$tmp1,"",0),$D);
-
- &add($A,$E);
- &mov($E, &DWP(16,$tmp1,"",0));
- &add($E,$C);
- &mov($C, &DWP( 8,$tmp1,"",0));
- &add($C,$T);
-
- &mov(&DWP( 0,$tmp1,"",0),$A);
- &mov("esi",&wparam(1));
- &mov(&DWP( 8,$tmp1,"",0),$C);
- &add("esi",64);
- &mov("eax",&swtmp(17));
- &mov(&DWP(16,$tmp1,"",0),$E);
- &cmp("esi","eax");
- &mov(&DWP( 4,$tmp1,"",0),$B);
- &jl(&label("start"));
-
- &stack_pop(18+9);
- &pop("edi");
- &pop("ebx");
- &pop("ebp");
- &pop("esi");
- &ret();
-
- # keep a note of shortcut label so it can be used outside
- # block.
- my $sclabel = &label("shortcut");
-
- &function_end_B($name);
- # Putting this here avoids problems with MASM in debugging mode
- &sha1_block_host("sha1_block_asm_host_order", $sclabel);
- }
+ for($i=0;$i<16;$i++) { &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
+ for(;$i<20;$i++) { &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
+ for(;$i<40;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
+ for(;$i<60;$i++) { &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
+ for(;$i<80;$i++) { &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
+
+ (($V[5] eq $D) and ($V[0] eq $E)) or die; # double-check
+
+ &mov($tmp1,&wparam(0)); # re-load SHA_CTX*
+ &mov($D,&wparam(1)); # D is last "T" and is discarded
+
+ &add($E,&DWP(0,$tmp1)); # E is last "A"...
+ &add($T,&DWP(4,$tmp1));
+ &add($A,&DWP(8,$tmp1));
+ &add($B,&DWP(12,$tmp1));
+ &add($C,&DWP(16,$tmp1));
+
+ &mov(&DWP(0,$tmp1),$E); # update SHA_CTX
+ &add($D,64); # advance input pointer
+ &mov(&DWP(4,$tmp1),$T);
+ &cmp($D,&wparam(2)); # have we reached the end yet?
+ &mov(&DWP(8,$tmp1),$A);
+ &mov($E,$C); # C is last "E" which needs to be "pre-loaded"
+ &mov(&DWP(12,$tmp1),$B);
+ &mov($T,$D); # input pointer
+ &mov(&DWP(16,$tmp1),$C);
+ &jb(&label("loop"));
+
+ &stack_pop(16+3);
+&function_end("sha1_block_data_order");
+
+if ($xmm) {
+######################################################################
+# The SSSE3 implementation.
+#
+# %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
+# 32 elements of the message schedule or Xupdate outputs. First 4
+# quadruples are simply byte-swapped input, next 4 are calculated
+# according to method originally suggested by Dean Gaudet (modulo
+# being implemented in SSSE3). Once 8 quadruples or 32 elements are
+# collected, it switches to routine proposed by Max Locktyukhin.
+#
+# Calculations inevitably require temporary reqisters, and there are
+# no %xmm registers left to spare. For this reason part of the ring
+# buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
+# buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
+# X[-5], and X[4] - X[-4]...
+#
+# Another notable optimization is aggressive stack frame compression
+# aiming to minimize amount of 9-byte instructions...
+#
+# Yet another notable optimization is "jumping" $B variable. It means
+# that there is no register permanently allocated for $B value. This
+# allowed to eliminate one instruction from body_20_39...
+#
+my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
+my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
+my @V=($A,$B,$C,$D,$E);
+my $j=0; # hash round
+my @T=($T,$tmp1);
+my $inp;
+
+my $_rol=sub { &rol(@_) };
+my $_ror=sub { &ror(@_) };
+
+&function_begin("_sha1_block_data_order_ssse3");
+ &call (&label("pic_point")); # make it PIC!
+ &set_label("pic_point");
+ &blindpop($tmp1);
+ &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
+&set_label("ssse3_shortcut");
+
+ &movdqa (@X[3],&QWP(0,$tmp1)); # K_00_19
+ &movdqa (@X[4],&QWP(16,$tmp1)); # K_20_39
+ &movdqa (@X[5],&QWP(32,$tmp1)); # K_40_59
+ &movdqa (@X[6],&QWP(48,$tmp1)); # K_60_79
+ &movdqa (@X[2],&QWP(64,$tmp1)); # pbswap mask
+
+ &mov ($E,&wparam(0)); # load argument block
+ &mov ($inp=@T[1],&wparam(1));
+ &mov ($D,&wparam(2));
+ &mov (@T[0],"esp");
+
+ # stack frame layout
+ #
+ # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
+ # X[4]+K X[5]+K X[6]+K X[7]+K
+ # X[8]+K X[9]+K X[10]+K X[11]+K
+ # X[12]+K X[13]+K X[14]+K X[15]+K
+ #
+ # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
+ # X[4] X[5] X[6] X[7]
+ # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
+ #
+ # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
+ # K_40_59 K_40_59 K_40_59 K_40_59
+ # K_60_79 K_60_79 K_60_79 K_60_79
+ # K_00_19 K_00_19 K_00_19 K_00_19
+ # pbswap mask
+ #
+ # +192 ctx # argument block
+ # +196 inp
+ # +200 end
+ # +204 esp
+ &sub ("esp",208);
+ &and ("esp",-64);
+
+ &movdqa (&QWP(112+0,"esp"),@X[4]); # copy constants
+ &movdqa (&QWP(112+16,"esp"),@X[5]);
+ &movdqa (&QWP(112+32,"esp"),@X[6]);
+ &shl ($D,6); # len*64
+ &movdqa (&QWP(112+48,"esp"),@X[3]);
+ &add ($D,$inp); # end of input
+ &movdqa (&QWP(112+64,"esp"),@X[2]);
+ &add ($inp,64);
+ &mov (&DWP(192+0,"esp"),$E); # save argument block
+ &mov (&DWP(192+4,"esp"),$inp);
+ &mov (&DWP(192+8,"esp"),$D);
+ &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
+
+ &mov ($A,&DWP(0,$E)); # load context
+ &mov ($B,&DWP(4,$E));
+ &mov ($C,&DWP(8,$E));
+ &mov ($D,&DWP(12,$E));
+ &mov ($E,&DWP(16,$E));
+ &mov (@T[0],$B); # magic seed
+
+ &movdqu (@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
+ &movdqu (@X[-3&7],&QWP(-48,$inp));
+ &movdqu (@X[-2&7],&QWP(-32,$inp));
+ &movdqu (@X[-1&7],&QWP(-16,$inp));
+ &pshufb (@X[-4&7],@X[2]); # byte swap
+ &pshufb (@X[-3&7],@X[2]);
+ &pshufb (@X[-2&7],@X[2]);
+ &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
+ &pshufb (@X[-1&7],@X[2]);
+ &paddd (@X[-4&7],@X[3]); # add K_00_19
+ &paddd (@X[-3&7],@X[3]);
+ &paddd (@X[-2&7],@X[3]);
+ &movdqa (&QWP(0,"esp"),@X[-4&7]); # X[]+K xfer to IALU
+ &psubd (@X[-4&7],@X[3]); # restore X[]
+ &movdqa (&QWP(0+16,"esp"),@X[-3&7]);
+ &psubd (@X[-3&7],@X[3]);
+ &movdqa (&QWP(0+32,"esp"),@X[-2&7]);
+ &psubd (@X[-2&7],@X[3]);
+ &movdqa (@X[0],@X[-3&7]);
+ &jmp (&label("loop"));
+
+######################################################################
+# SSE instruction sequence is first broken to groups of indepentent
+# instructions, independent in respect to their inputs and shifter
+# (not all architectures have more than one). Then IALU instructions
+# are "knitted in" between the SSE groups. Distance is maintained for
+# SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
+# [which allegedly also implements SSSE3]...
+#
+# Temporary registers usage. X[2] is volatile at the entry and at the
+# end is restored from backtrace ring buffer. X[3] is expected to
+# contain current K_XX_XX constant and is used to caclulate X[-1]+K
+# from previous round, it becomes volatile the moment the value is
+# saved to stack for transfer to IALU. X[4] becomes volatile whenever
+# X[-4] is accumulated and offloaded to backtrace ring buffer, at the
+# end it is loaded with next K_XX_XX [which becomes X[3] in next
+# round]...
+#
+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);
+
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &palignr(@X[0],@X[-4&7],8); # compose "X[-14]" in "X[0]"
+ &movdqa (@X[2],@X[-1&7]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &paddd (@X[3],@X[-1&7]);
+ &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &psrldq (@X[2],4); # "X[-3]", 3 dwords
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &pxor (@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &pxor (@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &pxor (@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &movdqa (@X[4],@X[0]);
+ &movdqa (@X[2],@X[0]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &pslldq (@X[4],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 (@X[2],31);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &movdqa (@X[3],@X[4]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &psrld (@X[4],30);
+ &por (@X[0],@X[2]); # "X[0]"<<<=1
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &pslld (@X[3],2);
+ &pxor (@X[0],@X[4]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &movdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &pxor (@X[0],@X[3]); # "X[0]"^=("X[0]"<<96)<<<2
+ &movdqa (@X[1],@X[-2&7]) if ($Xi<7);
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ foreach (@insns) { eval; } # remaining instructions [if any]
+
+ $Xi++; push(@X,shift(@X)); # "rotate" X[]
+}
+
+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 (@X[2],@X[-1&7]) if ($Xi==8);
+ eval(shift(@insns)); # body_20_39
+ &pxor (@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
+ &palignr(@X[2],@X[-2&7],8); # compose "X[-6]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # rol
+
+ &pxor (@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
+ &movdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns));
+ if ($Xi%5) {
+ &movdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
+ } else { # ... or load next one
+ &movdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
+ }
+ &paddd (@X[3],@X[-1&7]);
+ eval(shift(@insns)); # ror
+ eval(shift(@insns));
+
+ &pxor (@X[0],@X[2]); # "X[0]"^="X[-6]"
+ eval(shift(@insns)); # body_20_39
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # rol
+
+ &movdqa (@X[2],@X[0]);
+ &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # ror
+ eval(shift(@insns));
+
+ &pslld (@X[0],2);
+ eval(shift(@insns)); # body_20_39
+ eval(shift(@insns));
+ &psrld (@X[2],30);
+ eval(shift(@insns));
+ eval(shift(@insns)); # rol
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # ror
+ eval(shift(@insns));
+
+ &por (@X[0],@X[2]); # "X[0]"<<<=2
+ eval(shift(@insns)); # body_20_39
+ eval(shift(@insns));
+ &movdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns)); # rol
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # ror
+ &movdqa (@X[3],@X[0]) if ($Xi<19);
+ eval(shift(@insns));
+
+ foreach (@insns) { eval; } # remaining instructions
+
+ $Xi++; push(@X,shift(@X)); # "rotate" X[]
+}
+
+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 (@X[3],@X[-1&7]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &movdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
+
+ foreach (@insns) { eval; } # remaining instructions
+
+ &mov ($inp=@T[1],&DWP(192+4,"esp"));
+ &cmp ($inp,&DWP(192+8,"esp"));
+ &je (&label("done"));
+
+ &movdqa (@X[3],&QWP(112+48,"esp")); # K_00_19
+ &movdqa (@X[2],&QWP(112+64,"esp")); # pbswap mask
+ &movdqu (@X[-4&7],&QWP(0,$inp)); # load input
+ &movdqu (@X[-3&7],&QWP(16,$inp));
+ &movdqu (@X[-2&7],&QWP(32,$inp));
+ &movdqu (@X[-1&7],&QWP(48,$inp));
+ &add ($inp,64);
+ &pshufb (@X[-4&7],@X[2]); # byte swap
+ &mov (&DWP(192+4,"esp"),$inp);
+ &movdqa (&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
+
+ $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],@X[3]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &movdqa (&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]); # X[]+K xfer to IALU
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &psubd (@X[($Xi-4)&7],@X[3]);
+
+ 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,&DWP(4*($j&15),"esp"));', # X[]+K xfer
+ '&xor ($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 ($e,&DWP(4*($j++&15),"esp"));', # 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 ($e,&DWP(4*($j++&15),"esp"));', # 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));'
+ );
+}
+
+&set_label("loop",16);
+ &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);
+
+ &mov (@T[1],&DWP(192,"esp")); # update context
+ &add ($A,&DWP(0,@T[1]));
+ &add (@T[0],&DWP(4,@T[1])); # $b
+ &add ($C,&DWP(8,@T[1]));
+ &mov (&DWP(0,@T[1]),$A);
+ &add ($D,&DWP(12,@T[1]));
+ &mov (&DWP(4,@T[1]),@T[0]);
+ &add ($E,&DWP(16,@T[1]));
+ &mov (&DWP(8,@T[1]),$C);
+ &mov ($B,@T[0]);
+ &mov (&DWP(12,@T[1]),$D);
+ &mov (&DWP(16,@T[1]),$E);
+ &movdqa (@X[0],@X[-3&7]);
+
+ &jmp (&label("loop"));
+
+&set_label("done",16); $j=$saved_j; @V=@saved_V;
+
+ &Xtail_ssse3(\&body_20_39);
+ &Xtail_ssse3(\&body_20_39);
+ &Xtail_ssse3(\&body_20_39);
+
+ &mov (@T[1],&DWP(192,"esp")); # update context
+ &add ($A,&DWP(0,@T[1]));
+ &mov ("esp",&DWP(192+12,"esp")); # restore %esp
+ &add (@T[0],&DWP(4,@T[1])); # $b
+ &add ($C,&DWP(8,@T[1]));
+ &mov (&DWP(0,@T[1]),$A);
+ &add ($D,&DWP(12,@T[1]));
+ &mov (&DWP(4,@T[1]),@T[0]);
+ &add ($E,&DWP(16,@T[1]));
+ &mov (&DWP(8,@T[1]),$C);
+ &mov (&DWP(12,@T[1]),$D);
+ &mov (&DWP(16,@T[1]),$E);
+
+&function_end("_sha1_block_data_order_ssse3");
+
+if ($ymm) {
+my $Xi=4; # 4xSIMD Xupdate round, start pre-seeded
+my @X=map("xmm$_",(4..7,0..3)); # pre-seeded for $Xi=4
+my @V=($A,$B,$C,$D,$E);
+my $j=0; # hash round
+my @T=($T,$tmp1);
+my $inp;
+
+my $_rol=sub { &shld(@_[0],@_) };
+my $_ror=sub { &shrd(@_[0],@_) };
+
+&function_begin("_sha1_block_data_order_avx");
+ &call (&label("pic_point")); # make it PIC!
+ &set_label("pic_point");
+ &blindpop($tmp1);
+ &lea ($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
+&set_label("avx_shortcut");
+ &vzeroall();
+
+ &vmovdqa(@X[3],&QWP(0,$tmp1)); # K_00_19
+ &vmovdqa(@X[4],&QWP(16,$tmp1)); # K_20_39
+ &vmovdqa(@X[5],&QWP(32,$tmp1)); # K_40_59
+ &vmovdqa(@X[6],&QWP(48,$tmp1)); # K_60_79
+ &vmovdqa(@X[2],&QWP(64,$tmp1)); # pbswap mask
+
+ &mov ($E,&wparam(0)); # load argument block
+ &mov ($inp=@T[1],&wparam(1));
+ &mov ($D,&wparam(2));
+ &mov (@T[0],"esp");
+
+ # stack frame layout
+ #
+ # +0 X[0]+K X[1]+K X[2]+K X[3]+K # XMM->IALU xfer area
+ # X[4]+K X[5]+K X[6]+K X[7]+K
+ # X[8]+K X[9]+K X[10]+K X[11]+K
+ # X[12]+K X[13]+K X[14]+K X[15]+K
+ #
+ # +64 X[0] X[1] X[2] X[3] # XMM->XMM backtrace area
+ # X[4] X[5] X[6] X[7]
+ # X[8] X[9] X[10] X[11] # even borrowed for K_00_19
+ #
+ # +112 K_20_39 K_20_39 K_20_39 K_20_39 # constants
+ # K_40_59 K_40_59 K_40_59 K_40_59
+ # K_60_79 K_60_79 K_60_79 K_60_79
+ # K_00_19 K_00_19 K_00_19 K_00_19
+ # pbswap mask
+ #
+ # +192 ctx # argument block
+ # +196 inp
+ # +200 end
+ # +204 esp
+ &sub ("esp",208);
+ &and ("esp",-64);
+
+ &vmovdqa(&QWP(112+0,"esp"),@X[4]); # copy constants
+ &vmovdqa(&QWP(112+16,"esp"),@X[5]);
+ &vmovdqa(&QWP(112+32,"esp"),@X[6]);
+ &shl ($D,6); # len*64
+ &vmovdqa(&QWP(112+48,"esp"),@X[3]);
+ &add ($D,$inp); # end of input
+ &vmovdqa(&QWP(112+64,"esp"),@X[2]);
+ &add ($inp,64);
+ &mov (&DWP(192+0,"esp"),$E); # save argument block
+ &mov (&DWP(192+4,"esp"),$inp);
+ &mov (&DWP(192+8,"esp"),$D);
+ &mov (&DWP(192+12,"esp"),@T[0]); # save original %esp
+
+ &mov ($A,&DWP(0,$E)); # load context
+ &mov ($B,&DWP(4,$E));
+ &mov ($C,&DWP(8,$E));
+ &mov ($D,&DWP(12,$E));
+ &mov ($E,&DWP(16,$E));
+ &mov (@T[0],$B); # magic seed
+
+ &vmovdqu(@X[-4&7],&QWP(-64,$inp)); # load input to %xmm[0-3]
+ &vmovdqu(@X[-3&7],&QWP(-48,$inp));
+ &vmovdqu(@X[-2&7],&QWP(-32,$inp));
+ &vmovdqu(@X[-1&7],&QWP(-16,$inp));
+ &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
+ &vpshufb(@X[-3&7],@X[-3&7],@X[2]);
+ &vpshufb(@X[-2&7],@X[-2&7],@X[2]);
+ &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
+ &vpshufb(@X[-1&7],@X[-1&7],@X[2]);
+ &vpaddd (@X[0],@X[-4&7],@X[3]); # add K_00_19
+ &vpaddd (@X[1],@X[-3&7],@X[3]);
+ &vpaddd (@X[2],@X[-2&7],@X[3]);
+ &vmovdqa(&QWP(0,"esp"),@X[0]); # X[]+K xfer to IALU
+ &vmovdqa(&QWP(0+16,"esp"),@X[1]);
+ &vmovdqa(&QWP(0+32,"esp"),@X[2]);
+ &jmp (&label("loop"));
+
+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 (@X[3],@X[3],@X[-1&7]);
+ &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &vpsrldq(@X[2],@X[-1&7],4); # "X[-3]", 3 dwords
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &vpxor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vpxor (@X[2],@X[2],@X[-2&7]); # "X[-3]"^"X[-8]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer to IALU
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-3]"^"X[-8]"
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vpsrld (@X[2],@X[0],31);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vpslldq(@X[4],@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 (@X[3],@X[4],30);
+ &vpor (@X[0],@X[0],@X[2]); # "X[0]"<<<=1
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vpslld (@X[4],@X[4],2);
+ &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5); # restore X[] from backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &vpxor (@X[0],@X[0],@X[3]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vpxor (@X[0],@X[0],@X[4]); # "X[0]"^=("X[0]"<<96)<<<2
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &vmovdqa (@X[4],&QWP(112-16+16*(($Xi)/5),"esp")); # K_XX_XX
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ foreach (@insns) { eval; } # remaining instructions [if any]
+
+ $Xi++; push(@X,shift(@X)); # "rotate" X[]
+}
+
+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(@X[2],@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]"
+ &vmovdqa (&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]); # save X[] to backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns));
+ if ($Xi%5) {
+ &vmovdqa (@X[4],@X[3]); # "perpetuate" K_XX_XX...
+ } else { # ... or load next one
+ &vmovdqa (@X[4],&QWP(112-16+16*($Xi/5),"esp"));
+ }
+ &vpaddd (@X[3],@X[3],@X[-1&7]);
+ eval(shift(@insns)); # ror
+ eval(shift(@insns));
+
+ &vpxor (@X[0],@X[0],@X[2]); # "X[0]"^="X[-6]"
+ eval(shift(@insns)); # body_20_39
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # rol
+
+ &vpsrld (@X[2],@X[0],30);
+ &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # 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],@X[2]); # "X[0]"<<<=2
+ eval(shift(@insns)); # body_20_39
+ eval(shift(@insns));
+ &vmovdqa (@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19); # restore X[] from backtrace buffer
+ eval(shift(@insns));
+ eval(shift(@insns)); # rol
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns)); # ror
+ eval(shift(@insns));
+
+ foreach (@insns) { eval; } # remaining instructions
+
+ $Xi++; push(@X,shift(@X)); # "rotate" X[]
+}
+
+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 (@X[3],@X[3],@X[-1&7]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+
+ &vmovdqa (&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]); # X[]+K xfer IALU
+
+ foreach (@insns) { eval; } # remaining instructions
+
+ &mov ($inp=@T[1],&DWP(192+4,"esp"));
+ &cmp ($inp,&DWP(192+8,"esp"));
+ &je (&label("done"));
+
+ &vmovdqa(@X[3],&QWP(112+48,"esp")); # K_00_19
+ &vmovdqa(@X[2],&QWP(112+64,"esp")); # pbswap mask
+ &vmovdqu(@X[-4&7],&QWP(0,$inp)); # load input
+ &vmovdqu(@X[-3&7],&QWP(16,$inp));
+ &vmovdqu(@X[-2&7],&QWP(32,$inp));
+ &vmovdqu(@X[-1&7],&QWP(48,$inp));
+ &add ($inp,64);
+ &vpshufb(@X[-4&7],@X[-4&7],@X[2]); # byte swap
+ &mov (&DWP(192+4,"esp"),$inp);
+ &vmovdqa(&QWP(112-16,"esp"),@X[3]); # borrow last backtrace slot
+
+ $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],@X[3]);
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ eval(shift(@insns));
+ &vmovdqa (&QWP(0+16*$Xi,"esp"),@X[$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; }
+}
+
+&set_label("loop",16);
+ &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);
+
+ &mov (@T[1],&DWP(192,"esp")); # update context
+ &add ($A,&DWP(0,@T[1]));
+ &add (@T[0],&DWP(4,@T[1])); # $b
+ &add ($C,&DWP(8,@T[1]));
+ &mov (&DWP(0,@T[1]),$A);
+ &add ($D,&DWP(12,@T[1]));
+ &mov (&DWP(4,@T[1]),@T[0]);
+ &add ($E,&DWP(16,@T[1]));
+ &mov (&DWP(8,@T[1]),$C);
+ &mov ($B,@T[0]);
+ &mov (&DWP(12,@T[1]),$D);
+ &mov (&DWP(16,@T[1]),$E);
+
+ &jmp (&label("loop"));
+
+&set_label("done",16); $j=$saved_j; @V=@saved_V;
+
+ &Xtail_avx(\&body_20_39);
+ &Xtail_avx(\&body_20_39);
+ &Xtail_avx(\&body_20_39);
+
+ &vzeroall();
+
+ &mov (@T[1],&DWP(192,"esp")); # update context
+ &add ($A,&DWP(0,@T[1]));
+ &mov ("esp",&DWP(192+12,"esp")); # restore %esp
+ &add (@T[0],&DWP(4,@T[1])); # $b
+ &add ($C,&DWP(8,@T[1]));
+ &mov (&DWP(0,@T[1]),$A);
+ &add ($D,&DWP(12,@T[1]));
+ &mov (&DWP(4,@T[1]),@T[0]);
+ &add ($E,&DWP(16,@T[1]));
+ &mov (&DWP(8,@T[1]),$C);
+ &mov (&DWP(12,@T[1]),$D);
+ &mov (&DWP(16,@T[1]),$E);
+&function_end("_sha1_block_data_order_avx");
+}
+&set_label("K_XX_XX",64);
+&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999); # K_00_19
+&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1); # K_20_39
+&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc); # K_40_59
+&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6); # K_60_79
+&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f); # pbswap mask
+}
+&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
+
+&asm_finish();