Commit 9833757b authored by Andy Polyakov's avatar Andy Polyakov
Browse files

s390x assembler pack update from HEAD.

parent 4195343c
Loading
Loading
Loading
Loading
+991 −73

File changed.

Preview size limit exceeded, changes collapsed.

+221 −0
Original line number Diff line number Diff line
#!/usr/bin/env perl
#
# ====================================================================
# 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/.
# ====================================================================
#
# May 2011
#
# The module implements bn_GF2m_mul_2x2 polynomial multiplication used
# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for
# the time being... gcc 4.3 appeared to generate poor code, therefore
# the effort. And indeed, the module delivers 55%-90%(*) improvement
# on haviest ECDSA verify and ECDH benchmarks for 163- and 571-bit
# key lengths on z990, 30%-55%(*) - on z10, and 70%-110%(*) - on z196.
# This is for 64-bit build. In 32-bit "highgprs" case improvement is
# even higher, for example on z990 it was measured 80%-150%. ECDSA
# sign is modest 9%-12% faster. Keep in mind that these coefficients
# are not ones for bn_GF2m_mul_2x2 itself, as not all CPU time is
# burnt in it...
#
# (*)	gcc 4.1 was observed to deliver better results than gcc 4.3,
#	so that improvement coefficients can vary from one specific
#	setup to another.

$flavour = shift;

if ($flavour =~ /3[12]/) {
        $SIZE_T=4;
        $g="";
} else {
        $SIZE_T=8;
        $g="g";
}

while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
open STDOUT,">$output";

$stdframe=16*$SIZE_T+4*8;

$rp="%r2";
$a1="%r3";
$a0="%r4";
$b1="%r5";
$b0="%r6";

$ra="%r14";
$sp="%r15";

@T=("%r0","%r1");
@i=("%r12","%r13");

($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(6..11));
($lo,$hi,$b)=map("%r$_",(3..5)); $a=$lo; $mask=$a8;

$code.=<<___;
.text

.type	_mul_1x1,\@function
.align	16
_mul_1x1:
	lgr	$a1,$a
	sllg	$a2,$a,1
	sllg	$a4,$a,2
	sllg	$a8,$a,3

	srag	$lo,$a1,63			# broadcast 63rd bit
	nihh	$a1,0x1fff
	srag	@i[0],$a2,63			# broadcast 62nd bit
	nihh	$a2,0x3fff
	srag	@i[1],$a4,63			# broadcast 61st bit
	nihh	$a4,0x7fff
	ngr	$lo,$b
	ngr	@i[0],$b
	ngr	@i[1],$b

	lghi	@T[0],0
	lgr	$a12,$a1
	stg	@T[0],`$stdframe+0*8`($sp)	# tab[0]=0
	xgr	$a12,$a2
	stg	$a1,`$stdframe+1*8`($sp)	# tab[1]=a1
	 lgr	$a48,$a4
	stg	$a2,`$stdframe+2*8`($sp)	# tab[2]=a2
	 xgr	$a48,$a8
	stg	$a12,`$stdframe+3*8`($sp)	# tab[3]=a1^a2
	 xgr	$a1,$a4

	stg	$a4,`$stdframe+4*8`($sp)	# tab[4]=a4
	xgr	$a2,$a4
	stg	$a1,`$stdframe+5*8`($sp)	# tab[5]=a1^a4
	xgr	$a12,$a4
	stg	$a2,`$stdframe+6*8`($sp)	# tab[6]=a2^a4
	 xgr	$a1,$a48
	stg	$a12,`$stdframe+7*8`($sp)	# tab[7]=a1^a2^a4
	 xgr	$a2,$a48

	stg	$a8,`$stdframe+8*8`($sp)	# tab[8]=a8
	xgr	$a12,$a48
	stg	$a1,`$stdframe+9*8`($sp)	# tab[9]=a1^a8
	 xgr	$a1,$a4
	stg	$a2,`$stdframe+10*8`($sp)	# tab[10]=a2^a8
	 xgr	$a2,$a4
	stg	$a12,`$stdframe+11*8`($sp)	# tab[11]=a1^a2^a8

	xgr	$a12,$a4
	stg	$a48,`$stdframe+12*8`($sp)	# tab[12]=a4^a8
	 srlg	$hi,$lo,1
	stg	$a1,`$stdframe+13*8`($sp)	# tab[13]=a1^a4^a8
	 sllg	$lo,$lo,63
	stg	$a2,`$stdframe+14*8`($sp)	# tab[14]=a2^a4^a8
	 srlg	@T[0],@i[0],2
	stg	$a12,`$stdframe+15*8`($sp)	# tab[15]=a1^a2^a4^a8

	lghi	$mask,`0xf<<3`
	sllg	$a1,@i[0],62
	 sllg	@i[0],$b,3
	srlg	@T[1],@i[1],3
	 ngr	@i[0],$mask
	sllg	$a2,@i[1],61
	 srlg	@i[1],$b,4-3
	xgr	$hi,@T[0]
	 ngr	@i[1],$mask
	xgr	$lo,$a1
	xgr	$hi,@T[1]
	xgr	$lo,$a2

	xg	$lo,$stdframe(@i[0],$sp)
	srlg	@i[0],$b,8-3
	ngr	@i[0],$mask
___
for($n=1;$n<14;$n++) {
$code.=<<___;
	lg	@T[1],$stdframe(@i[1],$sp)
	srlg	@i[1],$b,`($n+2)*4`-3
	sllg	@T[0],@T[1],`$n*4`
	ngr	@i[1],$mask
	srlg	@T[1],@T[1],`64-$n*4`
	xgr	$lo,@T[0]
	xgr	$hi,@T[1]
___
	push(@i,shift(@i)); push(@T,shift(@T));
}
$code.=<<___;
	lg	@T[1],$stdframe(@i[1],$sp)
	sllg	@T[0],@T[1],`$n*4`
	srlg	@T[1],@T[1],`64-$n*4`
	xgr	$lo,@T[0]
	xgr	$hi,@T[1]

	lg	@T[0],$stdframe(@i[0],$sp)
	sllg	@T[1],@T[0],`($n+1)*4`
	srlg	@T[0],@T[0],`64-($n+1)*4`
	xgr	$lo,@T[1]
	xgr	$hi,@T[0]

	br	$ra
.size	_mul_1x1,.-_mul_1x1

.globl	bn_GF2m_mul_2x2
.type	bn_GF2m_mul_2x2,\@function
.align	16
bn_GF2m_mul_2x2:
	stm${g}	%r3,%r15,3*$SIZE_T($sp)

	lghi	%r1,-$stdframe-128
	la	%r0,0($sp)
	la	$sp,0(%r1,$sp)			# alloca
	st${g}	%r0,0($sp)			# back chain
___
if ($SIZE_T==8) {
my @r=map("%r$_",(6..9));
$code.=<<___;
	bras	$ra,_mul_1x1			# a1b1
	stmg	$lo,$hi,16($rp)

	lg	$a,`$stdframe+128+4*$SIZE_T`($sp)
	lg	$b,`$stdframe+128+6*$SIZE_T`($sp)
	bras	$ra,_mul_1x1			# a0b0
	stmg	$lo,$hi,0($rp)

	lg	$a,`$stdframe+128+3*$SIZE_T`($sp)
	lg	$b,`$stdframe+128+5*$SIZE_T`($sp)
	xg	$a,`$stdframe+128+4*$SIZE_T`($sp)
	xg	$b,`$stdframe+128+6*$SIZE_T`($sp)
	bras	$ra,_mul_1x1			# (a0+a1)(b0+b1)
	lmg	@r[0],@r[3],0($rp)

	xgr	$lo,$hi
	xgr	$hi,@r[1]
	xgr	$lo,@r[0]
	xgr	$hi,@r[2]
	xgr	$lo,@r[3]	
	xgr	$hi,@r[3]
	xgr	$lo,$hi
	stg	$hi,16($rp)
	stg	$lo,8($rp)
___
} else {
$code.=<<___;
	sllg	%r3,%r3,32
	sllg	%r5,%r5,32
	or	%r3,%r4
	or	%r5,%r6
	bras	$ra,_mul_1x1
	rllg	$lo,$lo,32
	rllg	$hi,$hi,32
	stmg	$lo,$hi,0($rp)
___
}
$code.=<<___;
	lm${g}	%r6,%r15,`$stdframe+128+6*$SIZE_T`($sp)
	br	$ra
.size	bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2
.string	"GF(2^m) Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>"
___

$code =~ s/\`([^\`]*)\`/eval($1)/gem;
print $code;
close STDOUT;
+77 −25
Original line number Diff line number Diff line
@@ -32,6 +32,33 @@
# Reschedule to minimize/avoid Address Generation Interlock hazard,
# make inner loops counter-based.

# November 2010.
#
# Adapt for -m31 build. If kernel supports what's called "highgprs"
# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
# instructions and achieve "64-bit" performance even in 31-bit legacy
# application context. The feature is not specific to any particular
# processor, as long as it's "z-CPU". Latter implies that the code
# remains z/Architecture specific. Compatibility with 32-bit BN_ULONG
# is achieved by swapping words after 64-bit loads, follow _dswap-s.
# On z990 it was measured to perform 2.6-2.2 times better than
# compiler-generated code, less for longer keys...

$flavour = shift;

if ($flavour =~ /3[12]/) {
	$SIZE_T=4;
	$g="";
} else {
	$SIZE_T=8;
	$g="g";
}

while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
open STDOUT,">$output";

$stdframe=16*$SIZE_T+4*8;

$mn0="%r0";
$num="%r1";

@@ -60,34 +87,44 @@ $code.=<<___;
.globl	bn_mul_mont
.type	bn_mul_mont,\@function
bn_mul_mont:
	lgf	$num,164($sp)	# pull $num
	sla	$num,3		# $num to enumerate bytes
	lgf	$num,`$stdframe+$SIZE_T-4`($sp)	# pull $num
	sla	$num,`log($SIZE_T)/log(2)`	# $num to enumerate bytes
	la	$bp,0($num,$bp)

	stg	%r2,16($sp)
	st${g}	%r2,2*$SIZE_T($sp)

	cghi	$num,16		#
	lghi	%r2,0		#
	blr	%r14		# if($num<16) return 0;
___
$code.=<<___ if ($flavour =~ /3[12]/);
	tmll	$num,4
	bnzr	%r14		# if ($num&1) return 0;
___
$code.=<<___ if ($flavour !~ /3[12]/);
	cghi	$num,96		#
	bhr	%r14		# if($num>96) return 0;
___
$code.=<<___;
	stm${g}	%r3,%r15,3*$SIZE_T($sp)

	stmg	%r3,%r15,24($sp)

	lghi	$rp,-160-8	# leave room for carry bit
	lghi	$rp,-$stdframe-8	# leave room for carry bit
	lcgr	$j,$num		# -$num
	lgr	%r0,$sp
	la	$rp,0($rp,$sp)
	la	$sp,0($j,$rp)	# alloca
	stg	%r0,0($sp)	# back chain
	st${g}	%r0,0($sp)	# back chain

	sra	$num,3		# restore $num
	la	$bp,0($j,$bp)	# restore $bp
	ahi	$num,-1		# adjust $num for inner loop
	lg	$n0,0($n0)	# pull n0
	_dswap	$n0

	lg	$bi,0($bp)
	_dswap	$bi
	lg	$alo,0($ap)
	_dswap	$alo
	mlgr	$ahi,$bi	# ap[0]*bp[0]
	lgr	$AHI,$ahi

@@ -95,6 +132,7 @@ bn_mul_mont:
	msgr	$mn0,$n0

	lg	$nlo,0($np)	#
	_dswap	$nlo
	mlgr	$nhi,$mn0	# np[0]*m1
	algr	$nlo,$alo	# +="tp[0]"
	lghi	$NHI,0
@@ -106,12 +144,14 @@ bn_mul_mont:
.align	16
.L1st:
	lg	$alo,0($j,$ap)
	_dswap	$alo
	mlgr	$ahi,$bi	# ap[j]*bp[0]
	algr	$alo,$AHI
	lghi	$AHI,0
	alcgr	$AHI,$ahi

	lg	$nlo,0($j,$np)
	_dswap	$nlo
	mlgr	$nhi,$mn0	# np[j]*m1
	algr	$nlo,$NHI
	lghi	$NHI,0
@@ -119,22 +159,24 @@ bn_mul_mont:
	algr	$nlo,$alo
	alcgr	$NHI,$nhi

	stg	$nlo,160-8($j,$sp)	# tp[j-1]=
	stg	$nlo,$stdframe-8($j,$sp)	# tp[j-1]=
	la	$j,8($j)	# j++
	brct	$count,.L1st

	algr	$NHI,$AHI
	lghi	$AHI,0
	alcgr	$AHI,$AHI	# upmost overflow bit
	stg	$NHI,160-8($j,$sp)
	stg	$AHI,160($j,$sp)
	stg	$NHI,$stdframe-8($j,$sp)
	stg	$AHI,$stdframe($j,$sp)
	la	$bp,8($bp)	# bp++

.Louter:
	lg	$bi,0($bp)	# bp[i]
	_dswap	$bi
	lg	$alo,0($ap)
	_dswap	$alo
	mlgr	$ahi,$bi	# ap[0]*bp[i]
	alg	$alo,160($sp)	# +=tp[0]
	alg	$alo,$stdframe($sp)	# +=tp[0]
	lghi	$AHI,0
	alcgr	$AHI,$ahi

@@ -142,6 +184,7 @@ bn_mul_mont:
	msgr	$mn0,$n0	# tp[0]*n0

	lg	$nlo,0($np)	# np[0]
	_dswap	$nlo
	mlgr	$nhi,$mn0	# np[0]*m1
	algr	$nlo,$alo	# +="tp[0]"
	lghi	$NHI,0
@@ -153,14 +196,16 @@ bn_mul_mont:
.align	16
.Linner:
	lg	$alo,0($j,$ap)
	_dswap	$alo
	mlgr	$ahi,$bi	# ap[j]*bp[i]
	algr	$alo,$AHI
	lghi	$AHI,0
	alcgr	$ahi,$AHI
	alg	$alo,160($j,$sp)# +=tp[j]
	alg	$alo,$stdframe($j,$sp)# +=tp[j]
	alcgr	$AHI,$ahi

	lg	$nlo,0($j,$np)
	_dswap	$nlo
	mlgr	$nhi,$mn0	# np[j]*m1
	algr	$nlo,$NHI
	lghi	$NHI,0
@@ -168,31 +213,33 @@ bn_mul_mont:
	algr	$nlo,$alo	# +="tp[j]"
	alcgr	$NHI,$nhi

	stg	$nlo,160-8($j,$sp)	# tp[j-1]=
	stg	$nlo,$stdframe-8($j,$sp)	# tp[j-1]=
	la	$j,8($j)	# j++
	brct	$count,.Linner

	algr	$NHI,$AHI
	lghi	$AHI,0
	alcgr	$AHI,$AHI
	alg	$NHI,160($j,$sp)# accumulate previous upmost overflow bit
	alg	$NHI,$stdframe($j,$sp)# accumulate previous upmost overflow bit
	lghi	$ahi,0
	alcgr	$AHI,$ahi	# new upmost overflow bit
	stg	$NHI,160-8($j,$sp)
	stg	$AHI,160($j,$sp)
	stg	$NHI,$stdframe-8($j,$sp)
	stg	$AHI,$stdframe($j,$sp)

	la	$bp,8($bp)	# bp++
	clg	$bp,160+8+32($j,$sp)	# compare to &bp[num]
	cl${g}	$bp,`$stdframe+8+4*$SIZE_T`($j,$sp)	# compare to &bp[num]
	jne	.Louter

	lg	$rp,160+8+16($j,$sp)	# reincarnate rp
	la	$ap,160($sp)
	l${g}	$rp,`$stdframe+8+2*$SIZE_T`($j,$sp)	# reincarnate rp
	la	$ap,$stdframe($sp)
	ahi	$num,1		# restore $num, incidentally clears "borrow"

	la	$j,0(%r0)
	lr	$count,$num
.Lsub:	lg	$alo,0($j,$ap)
	slbg	$alo,0($j,$np)
	lg	$nlo,0($j,$np)
	_dswap	$nlo
	slbgr	$alo,$nlo
	stg	$alo,0($j,$rp)
	la	$j,8($j)
	brct	$count,.Lsub
@@ -208,18 +255,23 @@ bn_mul_mont:
	la	$j,0(%r0)
	lgr	$count,$num
.Lcopy:	lg	$alo,0($j,$ap)		# copy or in-place refresh
	stg	$j,160($j,$sp)	# zap tp
	_dswap	$alo
	stg	$j,$stdframe($j,$sp)	# zap tp
	stg	$alo,0($j,$rp)
	la	$j,8($j)
	brct	$count,.Lcopy

	la	%r1,160+8+48($j,$sp)
	lmg	%r6,%r15,0(%r1)
	la	%r1,`$stdframe+8+6*$SIZE_T`($j,$sp)
	lm${g}	%r6,%r15,0(%r1)
	lghi	%r2,1		# signal "processed"
	br	%r14
.size	bn_mul_mont,.-bn_mul_mont
.string	"Montgomery Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>"
___

print $code;
foreach (split("\n",$code)) {
	s/\`([^\`]*)\`/eval $1/ge;
	s/_dswap\s+(%r[0-9]+)/sprintf("rllg\t%s,%s,32",$1,$1) if($SIZE_T==4)/e;
	print $_,"\n";
}
close STDOUT;
+39 −10
Original line number Diff line number Diff line
@@ -13,6 +13,29 @@
# "cluster" Address Generation Interlocks, so that one pipeline stall
# resolves several dependencies.

# November 2010.
#
# Adapt for -m31 build. If kernel supports what's called "highgprs"
# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit
# instructions and achieve "64-bit" performance even in 31-bit legacy
# application context. The feature is not specific to any particular
# processor, as long as it's "z-CPU". Latter implies that the code
# remains z/Architecture specific. On z990 it was measured to perform
# 50% better than code generated by gcc 4.3.

$flavour = shift;

if ($flavour =~ /3[12]/) {
	$SIZE_T=4;
	$g="";
} else {
	$SIZE_T=8;
	$g="g";
}

while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
open STDOUT,">$output";

$rp="%r14";
$sp="%r15";
$code=<<___;
@@ -39,7 +62,12 @@ $code.=<<___;
.type	RC4,\@function
.align	64
RC4:
	stmg	%r6,%r11,48($sp)
	stm${g}	%r6,%r11,6*$SIZE_T($sp)
___
$code.=<<___ if ($flavour =~ /3[12]/);
	llgfr	$len,$len
___
$code.=<<___;
	llgc	$XX[0],0($key)
	llgc	$YY,1($key)
	la	$XX[0],1($XX[0])
@@ -90,7 +118,7 @@ $code.=<<___;
	xgr	$acc,$TX[1]
	stg	$acc,0($out)
	la	$out,8($out)
	brct	$cnt,.Loop8
	brctg	$cnt,.Loop8

.Lshort:
	lghi	$acc,7
@@ -122,7 +150,7 @@ $code.=<<___;
	ahi	$XX[0],-1
	stc	$XX[0],0($key)
	stc	$YY,1($key)
	lmg	%r6,%r11,48($sp)
	lm${g}	%r6,%r11,6*$SIZE_T($sp)
	br	$rp
.size	RC4,.-RC4
.string	"RC4 for s390x, CRYPTOGAMS by <appro\@openssl.org>"
@@ -130,7 +158,7 @@ $code.=<<___;
___
}

# void private_RC4_set_key(RC4_KEY *key,unsigned int len,const void *inp)
# void RC4_set_key(RC4_KEY *key,unsigned int len,const void *inp)
{
$cnt="%r0";
$idx="%r1";
@@ -143,11 +171,11 @@ $ikey="%r7";
$iinp="%r8";

$code.=<<___;
.globl	private_RC4_set_key
.type	private_RC4_set_key,\@function
.globl	RC4_set_key
.type	RC4_set_key,\@function
.align	64
private_RC4_set_key:
	stmg	%r6,%r8,48($sp)
RC4_set_key:
	stm${g}	%r6,%r8,6*$SIZE_T($sp)
	lhi	$cnt,256
	la	$idx,0(%r0)
	sth	$idx,0($key)
@@ -180,9 +208,9 @@ private_RC4_set_key:
	la	$iinp,0(%r0)
	j	.L2ndloop
.Ldone:
	lmg	%r6,%r8,48($sp)
	lm${g}	%r6,%r8,6*$SIZE_T($sp)
	br	$rp
.size	private_RC4_set_key,.-private_RC4_set_key
.size	RC4_set_key,.-RC4_set_key

___
}
@@ -203,3 +231,4 @@ RC4_options:
___

print $code;
close STDOUT;	# force flush
+6 −6
Original line number Diff line number Diff line
@@ -4,7 +4,7 @@
#include <setjmp.h>
#include <signal.h>

extern unsigned long OPENSSL_s390xcap_P;
extern unsigned long OPENSSL_s390xcap_P[];

static sigjmp_buf ill_jmp;
static void ill_handler (int sig) { siglongjmp(ill_jmp,sig); }
@@ -16,7 +16,9 @@ void OPENSSL_cpuid_setup(void)
	sigset_t oset;
	struct sigaction ill_act,oact;

	if (OPENSSL_s390xcap_P) return;
	if (OPENSSL_s390xcap_P[0]) return;

	OPENSSL_s390xcap_P[0] = 1UL<<(8*sizeof(unsigned long)-1);

	memset(&ill_act,0,sizeof(ill_act));
	ill_act.sa_handler = ill_handler;
@@ -27,10 +29,8 @@ void OPENSSL_cpuid_setup(void)
	sigaction (SIGILL,&ill_act,&oact);

	/* protection against missing store-facility-list-extended */
	if (sigsetjmp(ill_jmp,0) == 0)
		OPENSSL_s390xcap_P = OPENSSL_s390x_facilities();
	else
		OPENSSL_s390xcap_P = 1UL<<63;
	if (sigsetjmp(ill_jmp,1) == 0)
		OPENSSL_s390x_facilities();

	sigaction (SIGILL,&oact,NULL);
	sigprocmask(SIG_SETMASK,&oset,NULL);
Loading