sha1-x86* assembler update: F_40_59 and Atom-specific optimizations.

2 files changed, 189 insertions, 180 deletions

diff --git a/crypto/sha/asm/sha1-586.pl b/crypto/sha/asm/sha1-586.pl
index a1f876281a..dc855ef557 100644
--- a/crypto/sha/asm/sha1-586.pl
+++ b/crypto/sha/asm/sha1-586.pl
@@ -12,6 +12,8 @@
 # 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
@@ -31,6 +33,17 @@
 # ----------------------------------------------------------------
 #					<appro@fy.chalmers.se>
 
+# 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.
+
 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
 push(@INC,"${dir}","${dir}../../perlasm");
 require "x86asm.pl";
@@ -59,15 +72,16 @@ sub BODY_00_15
 	&rotl($tmp1,5);			# tmp1=ROTATE(a,5)
 	 &xor($f,$d);
 	&add($tmp1,$e);			# tmp1+=e;
-	 &and($f,$b);
-	&mov($e,&swtmp($n%16));		# e becomes volatile and is loaded
+	 &mov($e,&swtmp($n%16));	# e becomes volatile and is loaded
 	 				# with xi, also note that e becomes
 					# f in next round...
-	 &xor($f,$d);			# f holds F_00_19(b,c,d)
+	&and($f,$b);
 	&rotr($b,2);			# b=ROTATE(b,30)
-	 &lea($tmp1,&DWP(0x5a827999,$tmp1,$e));	# 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,$tmp1); }	# f+=tmp1
+	if ($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
 	}
 
@@ -77,22 +91,22 @@ sub BODY_16_19
 
 	&comment("16_19 $n");
 
-	&mov($f,&swtmp($n%16));		# f to hold Xupdate(xi,xa,xb,xc,xd)
-	 &mov($tmp1,$c);		# tmp1 to hold F_00_19(b,c,d)
-	&xor($f,&swtmp(($n+2)%16));
-	 &xor($tmp1,$d);
-	&xor($f,&swtmp(($n+8)%16));
-	 &and($tmp1,$b);		# tmp1 holds F_00_19(b,c,d)
-	&rotr($b,2);			# b=ROTATE(b,30)
+	&mov($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($n%16),$f);		# xi=f
-	&lea($f,&DWP(0x5a827999,$f,$e));# f+=K_00_19+e
-	 &mov($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
@@ -103,20 +117,20 @@ sub BODY_20_39
 	&comment("20_39 $n");
 
 	&mov($tmp1,$b);			# tmp1 to hold F_20_39(b,c,d)
-	 &mov($f,&swtmp($n%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
-	&rotr($b,2);			# b=ROTATE(b,30)
-	 &xor($f,&swtmp(($n+2)%16));
+	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
 	&xor($tmp1,$c);
 	 &xor($f,&swtmp(($n+8)%16));
 	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
 	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
 	&rotl($f,1);			# f=ROTATE(f,1)
-	 &add($tmp1,$e);
-	&mov(&swtmp($n%16),$f);		# xi=f
-	 &mov($e,$a);			# e becomes volatile
-	&rotl($e,5);			# e=ROTATE(a,5)
-	 &lea($f,&DWP($K,$f,$tmp1));	# f+=K_20_39+e
-	&add($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
@@ -125,28 +139,24 @@ sub BODY_40_59
 
 	&comment("40_59 $n");
 
-	&mov($f,&swtmp($n%16));		# f to hold Xupdate(xi,xa,xb,xc,xd)
-	 &mov($tmp1,&swtmp(($n+2)%16));
-	&xor($f,$tmp1);
-	 &mov($tmp1,&swtmp(($n+8)%16));
-	&xor($f,$tmp1);
-	 &mov($tmp1,&swtmp(($n+13)%16));
-	&xor($f,$tmp1);			# f holds xa^xb^xc^xd
-	 &mov($tmp1,$b);		# tmp1 to hold F_40_59(b,c,d)
+	&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($f,1);			# f=ROTATE(f,1)
-	 &or($tmp1,$c);
-	&mov(&swtmp($n%16),$f);		# xi=f
-	 &and($tmp1,$d);
-	&lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e
-	 &mov($e,$b);			# e becomes volatile and is used
-					# to calculate F_40_59(b,c,d)
+	 &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
 	}
 
 &function_begin("sha1_block_data_order");
diff --git a/crypto/sha/asm/sha1-x86_64.pl b/crypto/sha/asm/sha1-x86_64.pl
index 4edc5ea9ad..35ab0e7e78 100755
--- a/crypto/sha/asm/sha1-x86_64.pl
+++ b/crypto/sha/asm/sha1-x86_64.pl
@@ -16,7 +16,7 @@
 # There was suggestion to mechanically translate 32-bit code, but I
 # dismissed it, reasoning that x86_64 offers enough register bank
 # capacity to fully utilize SHA-1 parallelism. Therefore this fresh
-# implementation:-) However! While 64-bit code does performs better
+# implementation:-) However! While 64-bit code does perform better
 # on Opteron, I failed to beat 32-bit assembler on EM64T core. Well,
 # x86_64 does offer larger *addressable* bank, but out-of-order core
 # reaches for even more registers through dynamic aliasing, and EM64T
@@ -29,6 +29,13 @@
 # Xeon P4	+65%		+0%		9.9
 # Core2		+60%		+10%		7.0
 
+# August 2009.
+#
+# The code was revised to minimize code size and to maximize
+# "distance" between instructions producing input to 'lea'
+# instruction and the 'lea' instruction itself, which is essential
+# for Intel Atom core.
+
 $flavour = shift;
 $output  = shift;
 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
@@ -51,194 +58,184 @@ $ctx="%r8";
 $inp="%r9";
 $num="%r10";
 
-$xi="%eax";
-$t0="%ebx";
-$t1="%ecx";
-$A="%edx";
-$B="%esi";
-$C="%edi";
-$D="%ebp";
-$E="%r11d";
-$T="%r12d";
+$t0="%eax";
+$t1="%ebx";
+$t2="%ecx";
+@xi=("%edx","%ebp");
+$A="%esi";
+$B="%edi";
+$C="%r11d";
+$D="%r12d";
+$E="%r13d";
 
-@V=($A,$B,$C,$D,$E,$T);
-
-sub PROLOGUE {
-my $func=shift;
-$code.=<<___;
-.globl	$func
-.type	$func,\@function,3
-.align	16
-$func:
-	push	%rbx
-	push	%rbp
-	push	%r12
-	mov	%rsp,%r11
-	mov	%rdi,$ctx	# reassigned argument
-	sub	\$`8+16*4`,%rsp
-	mov	%rsi,$inp	# reassigned argument
-	and	\$-64,%rsp
-	mov	%rdx,$num	# reassigned argument
-	mov	%r11,`16*4`(%rsp)
-.Lprologue:
-
-	mov	0($ctx),$A
-	mov	4($ctx),$B
-	mov	8($ctx),$C
-	mov	12($ctx),$D
-	mov	16($ctx),$E
-___
-}
-
-sub EPILOGUE {
-my $func=shift;
-$code.=<<___;
-	mov	`16*4`(%rsp),%rsi
-	mov	(%rsi),%r12
-	mov	8(%rsi),%rbp
-	mov	16(%rsi),%rbx
-	lea	24(%rsi),%rsp
-.Lepilogue:
-	ret
-.size	$func,.-$func
-___
-}
+@V=($A,$B,$C,$D,$E);
 
 sub BODY_00_19 {
-my ($i,$a,$b,$c,$d,$e,$f,$host)=@_;
+my ($i,$a,$b,$c,$d,$e)=@_;
 my $j=$i+1;
 $code.=<<___ if ($i==0);
-	mov	`4*$i`($inp),$xi	
-	`"bswap	$xi"	if(!defined($host))`
-	mov	$xi,`4*$i`(%rsp)
+	mov	`4*$i`($inp),$xi[0]
+	bswap	$xi[0]
+	mov	$xi[0],`4*$i`(%rsp)
 ___
 $code.=<<___ if ($i<15);
-	lea	0x5a827999($xi,$e),$f
 	mov	$c,$t0
-	mov	`4*$j`($inp),$xi
-	mov	$a,$e
+	mov	`4*$j`($inp),$xi[1]
+	mov	$a,$t2
 	xor	$d,$t0
-	`"bswap	$xi"	if(!defined($host))`	
-	rol	\$5,$e
+	bswap	$xi[1]
+	rol	\$5,$t2
+	lea	0x5a827999($xi[0],$e),$e
 	and	$b,$t0
-	mov	$xi,`4*$j`(%rsp)
-	add	$e,$f
+	mov	$xi[1],`4*$j`(%rsp)
+	add	$t2,$e
 	xor	$d,$t0
 	rol	\$30,$b
-	add	$t0,$f
+	add	$t0,$e
 ___
 $code.=<<___ if ($i>=15);
-	lea	0x5a827999($xi,$e),$f
-	mov	`4*($j%16)`(%rsp),$xi
+	mov	`4*($j%16)`(%rsp),$xi[1]
 	mov	$c,$t0
-	mov	$a,$e
-	xor	`4*(($j+2)%16)`(%rsp),$xi
+	mov	$a,$t2
+	xor	`4*(($j+2)%16)`(%rsp),$xi[1]
 	xor	$d,$t0
-	rol	\$5,$e
-	xor	`4*(($j+8)%16)`(%rsp),$xi
+	rol	\$5,$t2
+	xor	`4*(($j+8)%16)`(%rsp),$xi[1]
 	and	$b,$t0
-	add	$e,$f
-	xor	`4*(($j+13)%16)`(%rsp),$xi
+	lea	0x5a827999($xi[0],$e),$e
+	xor	`4*(($j+13)%16)`(%rsp),$xi[1]
 	xor	$d,$t0
+	rol	\$1,$xi[1]
+	add	$t2,$e
 	rol	\$30,$b
-	add	$t0,$f
-	rol	\$1,$xi
-	mov	$xi,`4*($j%16)`(%rsp)
+	mov	$xi[1],`4*($j%16)`(%rsp)
+	add	$t0,$e
 ___
+unshift(@xi,pop(@xi));
 }
 
 sub BODY_20_39 {
-my ($i,$a,$b,$c,$d,$e,$f)=@_;
+my ($i,$a,$b,$c,$d,$e)=@_;
 my $j=$i+1;
 my $K=($i<40)?0x6ed9eba1:0xca62c1d6;
 $code.=<<___ if ($i<79);
-	lea	$K($xi,$e),$f
-	mov	`4*($j%16)`(%rsp),$xi
+	mov	`4*($j%16)`(%rsp),$xi[1]
 	mov	$c,$t0
-	mov	$a,$e
-	xor	`4*(($j+2)%16)`(%rsp),$xi
+	mov	$a,$t2
+	xor	`4*(($j+2)%16)`(%rsp),$xi[1]
 	xor	$b,$t0
-	rol	\$5,$e
-	xor	`4*(($j+8)%16)`(%rsp),$xi
+	rol	\$5,$t2
+	lea	$K($xi[0],$e),$e
+	xor	`4*(($j+8)%16)`(%rsp),$xi[1]
 	xor	$d,$t0
-	add	$e,$f
-	xor	`4*(($j+13)%16)`(%rsp),$xi
+	add	$t2,$e
+	xor	`4*(($j+13)%16)`(%rsp),$xi[1]
 	rol	\$30,$b
-	add	$t0,$f
-	rol	\$1,$xi
+	add	$t0,$e
+	rol	\$1,$xi[1]
 ___
 $code.=<<___ if ($i<76);
-	mov	$xi,`4*($j%16)`(%rsp)
+	mov	$xi[1],`4*($j%16)`(%rsp)
 ___
 $code.=<<___ if ($i==79);
-	lea	$K($xi,$e),$f
 	mov	$c,$t0
-	mov	$a,$e
+	mov	$a,$t2
 	xor	$b,$t0
-	rol	\$5,$e
+	lea	$K($xi[0],$e),$e
+	rol	\$5,$t2
 	xor	$d,$t0
-	add	$e,$f
+	add	$t2,$e
 	rol	\$30,$b
-	add	$t0,$f
+	add	$t0,$e
 ___
+unshift(@xi,pop(@xi));
 }
 
 sub BODY_40_59 {
-my ($i,$a,$b,$c,$d,$e,$f)=@_;
+my ($i,$a,$b,$c,$d,$e)=@_;
 my $j=$i+1;
 $code.=<<___;
-	lea	0x8f1bbcdc($xi,$e),$f
-	mov	`4*($j%16)`(%rsp),$xi
-	mov	$b,$t0
-	mov	$b,$t1
-	xor	`4*(($j+2)%16)`(%rsp),$xi
-	mov	$a,$e
-	and	$c,$t0
-	xor	`4*(($j+8)%16)`(%rsp),$xi
-	or	$c,$t1
-	rol	\$5,$e
-	xor	`4*(($j+13)%16)`(%rsp),$xi
-	and	$d,$t1
-	add	$e,$f
-	rol	\$1,$xi
-	or	$t1,$t0
+	mov	`4*($j%16)`(%rsp),$xi[1]
+	mov	$c,$t0
+	mov	$c,$t1
+	xor	`4*(($j+2)%16)`(%rsp),$xi[1]
+	and	$d,$t0
+	mov	$a,$t2
+	xor	`4*(($j+8)%16)`(%rsp),$xi[1]
+	xor	$d,$t1
+	lea	0x8f1bbcdc($xi[0],$e),$e
+	rol	\$5,$t2
+	xor	`4*(($j+13)%16)`(%rsp),$xi[1]
+	add	$t0,$e
+	and	$b,$t1
+	rol	\$1,$xi[1]
+	add	$t1,$e
 	rol	\$30,$b
-	mov	$xi,`4*($j%16)`(%rsp)
-	add	$t0,$f
+	mov	$xi[1],`4*($j%16)`(%rsp)
+	add	$t2,$e
 ___
+unshift(@xi,pop(@xi));
 }
 
-$code=".text\n";
+$code.=<<___;
+.text
 
-&PROLOGUE("sha1_block_data_order");
-$code.=".align	4\n.Lloop:\n";
+.globl	sha1_block_data_order
+.type	sha1_block_data_order,\@function,3
+.align	16
+sha1_block_data_order:
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	mov	%rsp,%r11
+	mov	%rdi,$ctx	# reassigned argument
+	sub	\$`8+16*4`,%rsp
+	mov	%rsi,$inp	# reassigned argument
+	and	\$-64,%rsp
+	mov	%rdx,$num	# reassigned argument
+	mov	%r11,`16*4`(%rsp)
+.Lprologue:
+
+	mov	0($ctx),$A
+	mov	4($ctx),$B
+	mov	8($ctx),$C
+	mov	12($ctx),$D
+	mov	16($ctx),$E
+
+.align	4
+.Lloop:
+___
 for($i=0;$i<20;$i++)	{ &BODY_00_19($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)); }
 $code.=<<___;
-	add	0($ctx),$E
-	add	4($ctx),$T
-	add	8($ctx),$A
-	add	12($ctx),$B
-	add	16($ctx),$C
-	mov	$E,0($ctx)
-	mov	$T,4($ctx)
-	mov	$A,8($ctx)
-	mov	$B,12($ctx)
-	mov	$C,16($ctx)
-
-	xchg	$E,$A	# mov	$E,$A
-	xchg	$T,$B	# mov	$T,$B
-	xchg	$E,$C	# mov	$A,$C
-	xchg	$T,$D	# mov	$B,$D
-			# mov	$C,$E
-	lea	`16*4`($inp),$inp
+	add	0($ctx),$A
+	add	4($ctx),$B
+	add	8($ctx),$C
+	add	12($ctx),$D
+	add	16($ctx),$E
+	mov	$A,0($ctx)
+	mov	$B,4($ctx)
+	mov	$C,8($ctx)
+	mov	$D,12($ctx)
+	mov	$E,16($ctx)
+
 	sub	\$1,$num
+	lea	`16*4`($inp),$inp
 	jnz	.Lloop
-___
-&EPILOGUE("sha1_block_data_order");
-$code.=<<___;
+
+	mov	`16*4`(%rsp),%rsi
+	mov	(%rsi),%r13
+	mov	8(%rsi),%r12
+	mov	16(%rsi),%rbp
+	mov	24(%rsi),%rbx
+	lea	32(%rsi),%rsp
+.Lepilogue:
+	ret
+.size	sha1_block_data_order,.-sha1_block_data_order
+
 .asciz	"SHA1 block transform for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
 .align	16
 ___
@@ -281,14 +278,16 @@ se_handler:
 	jae	.Lin_prologue
 
 	mov	`16*4`(%rax),%rax	# pull saved stack pointer
-	lea	24(%rax),%rax
+	lea	32(%rax),%rax
 
 	mov	-8(%rax),%rbx
 	mov	-16(%rax),%rbp
 	mov	-24(%rax),%r12
+	mov	-32(%rax),%r13
 	mov	%rbx,144($context)	# restore context->Rbx
 	mov	%rbp,160($context)	# restore context->Rbp
 	mov	%r12,216($context)	# restore context->R12
+	mov	%r13,224($context)	# restore context->R13
 
 .Lin_prologue:
 	mov	8(%rax),%rdi