From: Andy Polyakov <appro@openssl.org>
Date: Fri, 4 Mar 2011 13:09:16 +0000 (+0000)
Subject: s390x assembler pack: tune-up and support for new z196 hardware.
X-Git-Tag: OpenSSL-fips-2_0-rc1~696
X-Git-Url: https://git.librecmc.org/?a=commitdiff_plain;h=0ab8fd58e1c3077c1e5e9057b3d73b73b99a843e;p=oweals%2Fopenssl.git

s390x assembler pack: tune-up and support for new z196 hardware.
---

diff --git a/crypto/aes/asm/aes-s390x.pl b/crypto/aes/asm/aes-s390x.pl
index db963c9df0..d5d514dc2d 100644
--- a/crypto/aes/asm/aes-s390x.pl
+++ b/crypto/aes/asm/aes-s390x.pl
@@ -70,6 +70,18 @@
 # remains z/Architecture specific. On z990 it was measured to perform
 # 2x better than code generated by gcc 4.3.
 
+# December 2010.
+#
+# Add support for z196 "cipher message with counter" instruction.
+# Note however that it's disengaged, because it was measured to
+# perform ~12% worse than vanilla km-based code...
+
+# February 2011.
+#
+# Add AES_xts_[en|de]crypt. This includes support for z196
+# km-xts-aes instructions, which deliver ~70% improvement at 8KB
+# block size over vanilla km-based code.
+
 $flavour = shift;
 
 if ($flavour =~ /3[12]/) {
@@ -268,7 +280,7 @@ $code.=<<___;
 .type   _s390x_AES_encrypt,\@function
 .align	16
 _s390x_AES_encrypt:
-	st${g}	$ra,`$stdframe-$SIZE_T`($sp)
+	st${g}	$ra,15*$SIZE_T($sp)
 	x	$s0,0($key)
 	x	$s1,4($key)
 	x	$s2,8($key)
@@ -432,7 +444,7 @@ _s390x_AES_encrypt:
 	or	$s2,$i3
 	or	$s3,$t3
 
-	l${g}	$ra,`$stdframe-$SIZE_T`($sp)
+	l${g}	$ra,15*$SIZE_T($sp)
 	xr	$s0,$t0
 	xr	$s1,$t2
 	x	$s2,24($key)
@@ -594,7 +606,7 @@ $code.=<<___;
 .type   _s390x_AES_decrypt,\@function
 .align	16
 _s390x_AES_decrypt:
-	st${g}	$ra,`$stdframe-$SIZE_T`($sp)
+	st${g}	$ra,15*$SIZE_T($sp)
 	x	$s0,0($key)
 	x	$s1,4($key)
 	x	$s2,8($key)
@@ -738,7 +750,7 @@ _s390x_AES_decrypt:
 	nr	$i1,$mask
 	nr	$i2,$mask
 
-	l${g}	$ra,`$stdframe-$SIZE_T`($sp)
+	l${g}	$ra,15*$SIZE_T($sp)
 	or	$s1,$t1
 	l	$t0,16($key)
 	l	$t1,20($key)
@@ -1164,7 +1176,8 @@ $code.=<<___;
 .size	AES_set_decrypt_key,.-AES_set_decrypt_key
 ___
 
-#void AES_cbc_encrypt(const unsigned char *in, unsigned char *out,
+########################################################################
+# void AES_cbc_encrypt(const unsigned char *in, unsigned char *out,
 #                     size_t length, const AES_KEY *key,
 #                     unsigned char *ivec, const int enc)
 {
@@ -1365,13 +1378,14 @@ $code.=<<___;
 .size	AES_cbc_encrypt,.-AES_cbc_encrypt
 ___
 }
-#void AES_ctr32_encrypt(const unsigned char *in, unsigned char *out,
+########################################################################
+# void AES_ctr32_encrypt(const unsigned char *in, unsigned char *out,
 #                     size_t blocks, const AES_KEY *key,
 #                     const unsigned char *ivec)
 {
 my $inp="%r2";
-my $out="%r3";
-my $len="%r4";
+my $out="%r4";	# blocks and out are swapped
+my $len="%r3";
 my $key="%r5";	my $iv0="%r5";
 my $ivp="%r6";
 my $fp ="%r7";
@@ -1381,6 +1395,9 @@ $code.=<<___;
 .type	AES_ctr32_encrypt,\@function
 .align	16
 AES_ctr32_encrypt:
+	xgr	%r3,%r4		# flip %r3 and %r4, $out and $len
+	xgr	%r4,%r3
+	xgr	%r3,%r4
 	llgfr	$len,$len	# safe in ctr32 subroutine even in 64-bit case
 ___
 $code.=<<___ if (!$softonly);
@@ -1415,20 +1432,75 @@ $code.=<<___ if (!$softonly);
 	st${g}	$fp,$SIZE_T($sp)
 
 	slgr	$len,$fp
-	brc	1,.Lctr32_hw_loop	# not zero, no borrow
+	brc	1,.Lctr32_hw_switch	# not zero, no borrow
 	algr	$fp,$len	# input is shorter than allocated buffer
 	lghi	$len,0
 	st${g}	$fp,$SIZE_T($sp)
 
-.Lctr32_hw_loop:
+.Lctr32_hw_switch:
+___
+$code.=<<___ if (0);	######### kmctr code was measured to be ~12% slower
+	larl	$s0,OPENSSL_s390xcap_P
+	lg	$s0,8($s0)
+	tmhh	$s0,0x0004	# check for message_security-assist-4
+	jz	.Lctr32_km_loop
+
+	llgfr	$s0,%r0
+	lgr	$s1,%r1
+	lghi	%r0,0
+	la	%r1,16($sp)
+	.long	0xb92d2042	# kmctr %r4,%r2,%r2
+
+	llihh	%r0,0x8000	# check if kmctr supports the function code
+	srlg	%r0,%r0,0($s0)
+	ng	%r0,16($sp)
+	lgr	%r0,$s0
+	lgr	%r1,$s1
+	jz	.Lctr32_km_loop
+
+####### kmctr code
+	algr	$out,$inp	# restore $out
+	lgr	$s1,$len	# $s1 undertakes $len
+	j	.Lctr32_kmctr_loop
+.align	16
+.Lctr32_kmctr_loop:
+	la	$s2,16($sp)
+	lgr	$s3,$fp
+.Lctr32_kmctr_prepare:
+	stg	$iv0,0($s2)
+	stg	$ivp,8($s2)
+	la	$s2,16($s2)
+	ahi	$ivp,1		# 32-bit increment, preserves upper half
+	brct	$s3,.Lctr32_kmctr_prepare
+
+	#la	$inp,0($inp)	# inp
+	sllg	$len,$fp,4	# len
+	#la	$out,0($out)	# out
+	la	$s2,16($sp)	# iv
+	.long	0xb92da042	# kmctr $out,$s2,$inp
+	brc	1,.-4		# pay attention to "partial completion"
+
+	slgr	$s1,$fp
+	brc	1,.Lctr32_kmctr_loop	# not zero, no borrow
+	algr	$fp,$s1
+	lghi	$s1,0
+	brc	4+1,.Lctr32_kmctr_loop	# not zero
+
+	l${g}	$sp,0($sp)
+	lm${g}	%r6,$s3,6*$SIZE_T($sp)
+	br	$ra
+.align	16
+___
+$code.=<<___;
+.Lctr32_km_loop:
 	la	$s2,16($sp)
 	lgr	$s3,$fp
-.Lctr32_hw_prepare:
+.Lctr32_km_prepare:
 	stg	$iv0,0($s2)
 	stg	$ivp,8($s2)
 	la	$s2,16($s2)
 	ahi	$ivp,1		# 32-bit increment, preserves upper half
-	brct	$s3,.Lctr32_hw_prepare
+	brct	$s3,.Lctr32_km_prepare
 
 	la	$s0,16($sp)	# inp
 	sllg	$s1,$fp,4	# len
@@ -1439,7 +1511,7 @@ $code.=<<___ if (!$softonly);
 	la	$s2,16($sp)
 	lgr	$s3,$fp
 	slgr	$s2,$inp
-.Lctr32_hw_xor:
+.Lctr32_km_xor:
 	lg	$s0,0($inp)
 	lg	$s1,8($inp)
 	xg	$s0,0($s2,$inp)
@@ -1447,22 +1519,22 @@ $code.=<<___ if (!$softonly);
 	stg	$s0,0($out,$inp)
 	stg	$s1,8($out,$inp)
 	la	$inp,16($inp)
-	brct	$s3,.Lctr32_hw_xor
+	brct	$s3,.Lctr32_km_xor
 
 	slgr	$len,$fp
-	brc	1,.Lctr32_hw_loop	# not zero, no borrow
+	brc	1,.Lctr32_km_loop	# not zero, no borrow
 	algr	$fp,$len
 	lghi	$len,0
-	brc	4+1,.Lctr32_hw_loop	# not zero
+	brc	4+1,.Lctr32_km_loop	# not zero
 
 	l${g}	$s0,0($sp)
 	l${g}	$s1,$SIZE_T($sp)
 	la	$s2,16($sp)
-.Lctr32_hw_zap:
+.Lctr32_km_zap:
 	stg	$s0,0($s2)
 	stg	$s0,8($s2)
 	la	$s2,16($s2)
-	brct	$s1,.Lctr32_hw_zap
+	brct	$s1,.Lctr32_km_zap
 
 	la	$sp,0($s0)
 	lm${g}	%r6,$s3,6*$SIZE_T($sp)
@@ -1472,12 +1544,12 @@ $code.=<<___ if (!$softonly);
 ___
 $code.=<<___;
 	stm${g}	$key,$ra,5*$SIZE_T($sp)
-	sl${g}r	$out,$inp
+	sl${g}r	$inp,$out
 	larl	$tbl,AES_Te
 	llgf	$t1,12($ivp)
 
 .Lctr32_loop:
-	stm${g}	$inp,$len,2*$SIZE_T($sp)
+	stm${g}	$inp,$out,2*$SIZE_T($sp)
 	llgf	$s0,0($ivp)
 	llgf	$s1,4($ivp)
 	llgf	$s2,8($ivp)
@@ -1489,27 +1561,694 @@ $code.=<<___;
 
 	lm${g}	$inp,$ivp,2*$SIZE_T($sp)
 	llgf	$t1,16*$SIZE_T($sp)
-	x	$s0,0($inp)
+	x	$s0,0($inp,$out)
+	x	$s1,4($inp,$out)
+	x	$s2,8($inp,$out)
+	x	$s3,12($inp,$out)
+	stm	$s0,$s3,0($out)
+
+	la	$out,16($out)
+	ahi	$t1,1		# 32-bit increment
+	brct	$len,.Lctr32_loop
+
+	lm${g}	%r6,$ra,6*$SIZE_T($sp)
+	br	$ra
+.size	AES_ctr32_encrypt,.-AES_ctr32_encrypt
+___
+}
+
+########################################################################
+# void AES_xts_encrypt(const char *inp,char *out,size_t len,
+#	const AES_KEY *key1, const AES_KEY *key2,u64 secno);
+#
+{
+my $inp="%r2";
+my $out="%r4";	# len and out are swapped
+my $len="%r3";
+my $key1="%r5";	# $i1
+my $key2="%r6";	# $i2
+my $fp="%r7";	# $i3
+my $tweak=16*$SIZE_T+16;	# or $stdframe-16, bottom of the frame...
+
+$code.=<<___;
+.type	_s390x_xts_km,\@function
+.align	16
+_s390x_xts_km:
+___
+$code.=<<___ if(0);
+	llgfr	$s0,%r0			# put aside the function code
+	lghi	$s1,0x7f
+	nr	$s1,%r0
+	lghi	%r0,0			# query capability vector
+	la	%r1,2*$SIZE_T($sp)
+	.long	0xb92e0042		# km %r4,%r2
+	llihh	%r1,0x8000
+	srlg	%r1,%r1,32($s1)		# check for 32+function code
+	ng	%r1,2*$SIZE_T($sp)
+	lgr	%r0,$s0			# restore the function code
+	la	%r1,0($key1)		# restore $key1
+	jz	.Lxts_km_vanilla
+
+	lmg	$i2,$i3,$tweak($sp)	# put aside the tweak value
+	algr	$out,$inp
+
+	oill	%r0,32			# switch to xts function code
+	aghi	$s1,-18			#
+	sllg	$s1,$s1,3		# (function code - 18)*8, 0 or 16
+	la	%r1,$tweak-16($sp)
+	slgr	%r1,$s1			# parameter block position
+	lmg	$s0,$s3,0($key1)	# load 256 bits of key material,
+	stmg	$s0,$s3,0(%r1)		# and copy it to parameter block.
+					# yes, it contains junk and overlaps
+					# with the tweak in 128-bit case.
+					# it's done to avoid conditional
+					# branch.
+	stmg	$i2,$i3,$tweak($sp)	# "re-seat" the tweak value
+
+	.long	0xb92e0042		# km %r4,%r2
+	brc	1,.-4			# pay attention to "partial completion"
+
+	lrvg	$s0,$tweak+0($sp)	# load the last tweak
+	lrvg	$s1,$tweak+8($sp)
+	stmg	%r0,%r3,$tweak-32(%r1)	# wipe copy of the key
+
+	nill	%r0,0xffdf		# switch back to original function code
+	la	%r1,0($key1)		# restore pointer to $key1
+	slgr	$out,$inp
+
+	llgc	$len,2*$SIZE_T-1($sp)
+	nill	$len,0x0f		# $len%=16
+	br	$ra
+	
+.align	16
+.Lxts_km_vanilla:
+___
+$code.=<<___;
+	# prepare and allocate stack frame at the top of 4K page
+	# with 1K reserved for eventual signal handling
+	lghi	$s0,-1024-256-16# guarantee at least 256-bytes buffer
+	lghi	$s1,-4096
+	algr	$s0,$sp
+	lgr	$fp,$sp
+	ngr	$s0,$s1		# align at page boundary
+	slgr	$fp,$s0		# total buffer size
+	lgr	$s2,$sp
+	lghi	$s1,1024+16	# sl[g]fi is extended-immediate facility
+	slgr	$fp,$s1		# deduct reservation to get usable buffer size
+	# buffer size is at lest 256 and at most 3072+256-16
+
+	la	$sp,1024($s0)	# alloca
+	nill	$fp,0xfff0	# round to 16*n
+	st${g}	$s2,0($sp)	# back-chain
+	nill	$len,0xfff0	# redundant
+	st${g}	$fp,$SIZE_T($sp)
+
+	slgr	$len,$fp
+	brc	1,.Lxts_km_go	# not zero, no borrow
+	algr	$fp,$len	# input is shorter than allocated buffer
+	lghi	$len,0
+	st${g}	$fp,$SIZE_T($sp)
+
+.Lxts_km_go:
+	lrvg	$s0,$tweak+0($s2)	# load the tweak value in little-endian
+	lrvg	$s1,$tweak+8($s2)
+
+	la	$s2,16($sp)		# vector of ascending tweak values
+	slgr	$s2,$inp
+	srlg	$s3,$fp,4
+	j	.Lxts_km_start
+
+.Lxts_km_loop:
+	la	$s2,16($sp)
+	slgr	$s2,$inp
+	srlg	$s3,$fp,4
+.Lxts_km_prepare:
+	lghi	$i1,0x87
+	srag	$i2,$s1,63		# broadcast upper bit
+	ngr	$i1,$i2			# rem
+	srlg	$i2,$s0,63		# carry bit from lower half
+	sllg	$s0,$s0,1
+	sllg	$s1,$s1,1
+	xgr	$s0,$i1
+	ogr	$s1,$i2
+.Lxts_km_start:
+	lrvgr	$i1,$s0			# flip byte order
+	lrvgr	$i2,$s1
+	stg	$i1,0($s2,$inp)
+	stg	$i2,8($s2,$inp)
+	xg	$i1,0($inp)
+	xg	$i2,8($inp)
+	stg	$i1,0($out,$inp)
+	stg	$i2,8($out,$inp)
+	la	$inp,16($inp)
+	brct	$s3,.Lxts_km_prepare
+
+	slgr	$inp,$fp		# rewind $inp
+	la	$s2,0($out,$inp)
+	lgr	$s3,$fp
+	.long	0xb92e00aa		# km $s2,$s2
+	brc	1,.-4			# pay attention to "partial completion"
+
+	la	$s2,16($sp)
+	slgr	$s2,$inp
+	srlg	$s3,$fp,4
+.Lxts_km_xor:
+	lg	$i1,0($out,$inp)
+	lg	$i2,8($out,$inp)
+	xg	$i1,0($s2,$inp)
+	xg	$i2,8($s2,$inp)
+	stg	$i1,0($out,$inp)
+	stg	$i2,8($out,$inp)
+	la	$inp,16($inp)
+	brct	$s3,.Lxts_km_xor
+
+	slgr	$len,$fp
+	brc	1,.Lxts_km_loop		# not zero, no borrow
+	algr	$fp,$len
+	lghi	$len,0
+	brc	4+1,.Lxts_km_loop	# not zero
+
+	l${g}	$i1,0($sp)		# back-chain
+	llgf	$fp,`2*$SIZE_T-4`($sp)	# bytes used
+	la	$i2,16($sp)
+	srlg	$fp,$fp,4
+.Lxts_km_zap:
+	stg	$i1,0($i2)
+	stg	$i1,8($i2)
+	la	$i2,16($i2)
+	brct	$fp,.Lxts_km_zap
+
+	la	$sp,0($i1)
+	llgc	$len,2*$SIZE_T-1($i1)
+	nill	$len,0x0f		# $len%=16
+	bzr	$ra
+
+	# generate one more tweak...
+	lghi	$i1,0x87
+	srag	$i2,$s1,63		# broadcast upper bit
+	ngr	$i1,$i2			# rem
+	srlg	$i2,$s0,63		# carry bit from lower half
+	sllg	$s0,$s0,1
+	sllg	$s1,$s1,1
+	xgr	$s0,$i1
+	ogr	$s1,$i2
+
+	ltr	$len,$len		# clear zero flag
+	br	$ra
+.size	_s390x_xts_km,.-_s390x_xts_km
+
+.globl	AES_xts_encrypt
+.type	AES_xts_encrypt,\@function
+.align	16
+AES_xts_encrypt:
+	xgr	%r3,%r4			# flip %r3 and %r4, $out and $len
+	xgr	%r4,%r3
+	xgr	%r3,%r4
+___
+$code.=<<___ if ($SIZE_T==4);
+	llgfr	$len,$len
+___
+$code.=<<___;
+	st${g}	$len,1*$SIZE_T($sp)	# save copy of $len
+	srag	$len,$len,4		# formally wrong, because it expands
+					# sign byte, but who can afford asking
+					# to process more than 2^63-1 bytes?
+					# I use it, because it sets condition
+					# code...
+	bcr	8,$ra			# abort if zero (i.e. less than 16)
+___
+$code.=<<___ if (!$softonly);
+	llgf	%r0,240($key2)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lxts_enc_software
+
+	stm${g}	%r6,$s3,6*$SIZE_T($sp)
+	st${g}	$ra,14*$SIZE_T($sp)
+
+	sllg	$len,$len,4		# $len&=~15
+	slgr	$out,$inp
+
+	lrvg	$s0,$stdframe($sp)	# load secno
+	lghi	$s1,0
+	la	$s2,$tweak($sp)
+	lghi	$s3,16
+	stmg	$s0,$s1,0($s2)
+	la	%r1,0($key2)		# $key2 is not needed anymore
+	.long	0xb92e00aa		# km $s2,$s2, generate the tweak
+	brc	1,.-4			# can this happen?
+
+	l	%r0,240($key1)
+	la	%r1,0($key1)		# $key1 is not needed anymore
+	bras	$ra,_s390x_xts_km
+	jz	.Lxts_enc_km_done
+
+	aghi	$inp,-16		# take one step back
+	la	$i3,0($out,$inp)	# put aside real $out
+.Lxts_enc_km_steal:
+	llgc	$i1,16($inp)
+	llgc	$i2,0($out,$inp)
+	stc	$i1,0($out,$inp)
+	stc	$i2,16($out,$inp)
+	la	$inp,1($inp)
+	brct	$len,.Lxts_enc_km_steal
+
+	la	$s2,0($i3)
+	lghi	$s3,16
+	lrvgr	$i1,$s0			# flip byte order
+	lrvgr	$i2,$s1
+	xg	$i1,0($s2)
+	xg	$i2,8($s2)
+	stg	$i1,0($s2)
+	stg	$i2,8($s2)
+	.long	0xb92e00aa		# km $s2,$s2
+	brc	1,.-4			# can this happen?
+	lrvgr	$i1,$s0			# flip byte order
+	lrvgr	$i2,$s1
+	xg	$i1,0($i3)
+	xg	$i2,8($i3)
+	stg	$i1,0($i3)
+	stg	$i2,8($i3)
+
+.Lxts_enc_km_done:
+	l${g}	$ra,14*$SIZE_T($sp)
+	st${g}	$sp,$tweak($sp)		# wipe tweak
+	st${g}	$sp,$tweak($sp)
+	lm${g}	%r6,$s3,6*$SIZE_T($sp)
+	br	$ra
+.align	16
+.Lxts_enc_software:
+___
+$code.=<<___;
+	stm${g}	%r6,$ra,6*$SIZE_T($sp)
+
+	slgr	$out,$inp
+
+	xgr	$s0,$s0			# clear upper half
+	xgr	$s1,$s1
+	lrv	$s0,$stdframe+4($sp)	# load secno
+	lrv	$s1,$stdframe+0($sp)
+	xgr	$s2,$s2
+	xgr	$s3,$s3
+	stm${g}	%r2,%r5,2*$SIZE_T($sp)
+	la	$key,0($key2)
+	larl	$tbl,AES_Te
+	bras	$ra,_s390x_AES_encrypt	# generate the tweak
+	lm${g}	%r2,%r5,2*$SIZE_T($sp)
+	stm	$s0,$s3,$tweak($sp)	# save the tweak
+	j	.Lxts_enc_enter
+
+.align	16
+.Lxts_enc_loop:
+	lrvg	$s1,$tweak+0($sp)	# load the tweak in little-endian
+	lrvg	$s3,$tweak+8($sp)
+	lghi	%r1,0x87
+	srag	%r0,$s3,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	srlg	%r0,$s1,63		# carry bit from lower half
+	sllg	$s1,$s1,1
+	sllg	$s3,$s3,1
+	xgr	$s1,%r1
+	ogr	$s3,%r0
+	lrvgr	$s1,$s1			# flip byte order
+	lrvgr	$s3,$s3
+	srlg	$s0,$s1,32		# smash the tweak to 4x32-bits 
+	stg	$s1,$tweak+0($sp)	# save the tweak
+	llgfr	$s1,$s1
+	srlg	$s2,$s3,32
+	stg	$s3,$tweak+8($sp)
+	llgfr	$s3,$s3
+	la	$inp,16($inp)		# $inp+=16
+.Lxts_enc_enter:
+	x	$s0,0($inp)		# ^=*($inp)
 	x	$s1,4($inp)
 	x	$s2,8($inp)
 	x	$s3,12($inp)
+	stm${g}	%r2,%r3,2*$SIZE_T($sp)	# only two registers are changing
+	la	$key,0($key1)
+	bras	$ra,_s390x_AES_encrypt
+	lm${g}	%r2,%r5,2*$SIZE_T($sp)
+	x	$s0,$tweak+0($sp)	# ^=tweak
+	x	$s1,$tweak+4($sp)
+	x	$s2,$tweak+8($sp)
+	x	$s3,$tweak+12($sp)
 	st	$s0,0($out,$inp)
 	st	$s1,4($out,$inp)
 	st	$s2,8($out,$inp)
 	st	$s3,12($out,$inp)
+	brct${g}	$len,.Lxts_enc_loop
+
+	llgc	$len,`2*$SIZE_T-1`($sp)
+	nill	$len,0x0f		# $len%16
+	jz	.Lxts_enc_done
+
+	la	$i3,0($inp,$out)	# put aside real $out
+.Lxts_enc_steal:
+	llgc	%r0,16($inp)
+	llgc	%r1,0($out,$inp)
+	stc	%r0,0($out,$inp)
+	stc	%r1,16($out,$inp)
+	la	$inp,1($inp)
+	brct	$len,.Lxts_enc_steal
+	la	$out,0($i3)		# restore real $out
+
+	# generate last tweak...
+	lrvg	$s1,$tweak+0($sp)	# load the tweak in little-endian
+	lrvg	$s3,$tweak+8($sp)
+	lghi	%r1,0x87
+	srag	%r0,$s3,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	srlg	%r0,$s1,63		# carry bit from lower half
+	sllg	$s1,$s1,1
+	sllg	$s3,$s3,1
+	xgr	$s1,%r1
+	ogr	$s3,%r0
+	lrvgr	$s1,$s1			# flip byte order
+	lrvgr	$s3,$s3
+	srlg	$s0,$s1,32		# smash the tweak to 4x32-bits 
+	stg	$s1,$tweak+0($sp)	# save the tweak
+	llgfr	$s1,$s1
+	srlg	$s2,$s3,32
+	stg	$s3,$tweak+8($sp)
+	llgfr	$s3,$s3
+
+	x	$s0,0($out)		# ^=*(inp)|stolen cipther-text
+	x	$s1,4($out)
+	x	$s2,8($out)
+	x	$s3,12($out)
+	st${g}	$out,4*$SIZE_T($sp)
+	la	$key,0($key1)
+	bras	$ra,_s390x_AES_encrypt
+	l${g}	$out,4*$SIZE_T($sp)
+	x	$s0,`$tweak+0`($sp)	# ^=tweak
+	x	$s1,`$tweak+4`($sp)
+	x	$s2,`$tweak+8`($sp)
+	x	$s3,`$tweak+12`($sp)
+	st	$s0,0($out)
+	st	$s1,4($out)
+	st	$s2,8($out)
+	st	$s3,12($out)
+
+.Lxts_enc_done:
+	stg	$sp,$tweak+0($sp)	# wipe tweak
+	stg	$sp,$twesk+8($sp)
+	lm${g}	%r6,$ra,6*$SIZE_T($sp)
+	br	$ra
+.size	AES_xts_encrypt,.-AES_xts_encrypt
+___
+# void AES_xts_decrypt(const char *inp,char *out,size_t len,
+#	const AES_KEY *key1, const AES_KEY *key2,u64 secno);
+#
+$code.=<<___;
+.globl	AES_xts_decrypt
+.type	AES_xts_decrypt,\@function
+.align	16
+AES_xts_decrypt:
+	xgr	%r3,%r4			# flip %r3 and %r4, $out and $len
+	xgr	%r4,%r3
+	xgr	%r3,%r4
+___
+$code.=<<___ if ($SIZE_T==4);
+	llgfr	$len,$len
+___
+$code.=<<___;
+	st${g}	$len,1*$SIZE_T($sp)	# save copy of $len
+	aghi	$len,-16
+	bcr	4,$ra			# abort if less than zero. formally
+					# wrong, because $len is unsigned,
+					# but who can afford asking to
+					# process more than 2^63-1 bytes?
+	tmll	$len,0x0f
+	jnz	.Lxts_dec_proceed
+	aghi	$len,16
+.Lxts_dec_proceed:
+___
+$code.=<<___ if (!$softonly);
+	llgf	%r0,240($key2)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lxts_dec_software
 
+	stm${g}	%r6,$s3,6*$SIZE_T($sp)
+	st${g}	$ra,14*$SIZE_T($sp)
+
+	nill	$len,0xfff0		# $len&=~15
+	slgr	$out,$inp
+
+	# generate the tweak value
+	lrvg	$s0,$stdframe($sp)	# load secno
+	lghi	$s1,0
+	la	$s2,$tweak($sp)
+	lghi	$s3,16
+	stg	$s0,0($s2)
+	stg	$s1,8($s2)
+	la	%r1,0($key2)		# $key2 is not needed past this point
+	.long	0xb92e00aa		# km $s2,$s2, generate the tweak
+	brc	1,.-4			# can this happen?
+
+	l	%r0,240($key1)
+	la	%r1,0($key1)		# $key1 is not needed anymore
+
+	ltgr	$len,$len
+	jz	.Lxts_dec_km_short
+	bras	$ra,_s390x_xts_km
+	jz	.Lxts_dec_km_done
+
+	lrvgr	$s2,$s0			# make copy in reverse byte order
+	lrvgr	$s3,$s1
+	j	.Lxts_dec_km_2ndtweak
+
+.Lxts_dec_km_short:
+	llgc	$len,`2*$SIZE_T-1`($sp)
+	nill	$len,0x0f		# $len%=16
+	lrvg	$s0,$tweak+0($sp)	# load the tweak
+	lrvg	$s1,$tweak+8($sp)
+	lrvgr	$s2,$s0			# make copy in reverse byte order
+	lrvgr	$s3,$s1
+
+.Lxts_dec_km_2ndtweak:
+	lghi	$i1,0x87
+	srag	$i2,$s1,63		# broadcast upper bit
+	ngr	$i1,$i2			# rem
+	srlg	$i2,$s0,63		# carry bit from lower half
+	sllg	$s0,$s0,1
+	sllg	$s1,$s1,1
+	xgr	$s0,$i1
+	ogr	$s1,$i2
+	lrvgr	$i1,$s0			# flip byte order
+	lrvgr	$i2,$s1
+
+	xg	$i1,0($inp)
+	xg	$i2,8($inp)
+	stg	$i1,0($out,$inp)
+	stg	$i2,8($out,$inp)
+	la	$i2,0($out,$inp)
+	lghi	$i3,16
+	.long	0xb92e0066		# km $i2,$i2
+	brc	1,.-4			# can this happen?
+	lrvgr	$i1,$s0
+	lrvgr	$i2,$s1
+	xg	$i1,0($out,$inp)
+	xg	$i2,8($out,$inp)
+	stg	$i1,0($out,$inp)
+	stg	$i2,8($out,$inp)
+
+	la	$i3,0($out,$inp)	# put aside real $out
+.Lxts_dec_km_steal:
+	llgc	$i1,16($inp)
+	llgc	$i2,0($out,$inp)
+	stc	$i1,0($out,$inp)
+	stc	$i2,16($out,$inp)
+	la	$inp,1($inp)
+	brct	$len,.Lxts_dec_km_steal
+
+	lgr	$s0,$s2
+	lgr	$s1,$s3
+	xg	$s0,0($i3)
+	xg	$s1,8($i3)
+	stg	$s0,0($i3)
+	stg	$s1,8($i3)
+	la	$s0,0($i3)
+	lghi	$s1,16
+	.long	0xb92e0088		# km $s0,$s0
+	brc	1,.-4			# can this happen?
+	xg	$s2,0($i3)
+	xg	$s3,8($i3)
+	stg	$s2,0($i3)
+	stg	$s3,8($i3)
+.Lxts_dec_km_done:
+	l${g}	$ra,14*$SIZE_T($sp)
+	st${g}	$sp,$tweak($sp)		# wipe tweak
+	st${g}	$sp,$tweak($sp)
+	lm${g}	%r6,$s3,6*$SIZE_T($sp)
+	br	$ra
+.align	16
+.Lxts_dec_software:
+___
+$code.=<<___;
+	stm${g}	%r6,$ra,6*$SIZE_T($sp)
+
+	srlg	$len,$len,4
+	slgr	$out,$inp
+
+	xgr	$s0,$s0			# clear upper half
+	xgr	$s1,$s1
+	lrv	$s0,$stdframe+4($sp)	# load secno
+	lrv	$s1,$stdframe+0($sp)
+	xgr	$s2,$s2
+	xgr	$s3,$s3
+	stm${g}	%r2,%r5,2*$SIZE_T($sp)
+	la	$key,0($key2)
+	larl	$tbl,AES_Te
+	bras	$ra,_s390x_AES_encrypt	# generate the tweak
+	lm${g}	%r2,%r5,2*$SIZE_T($sp)
+	larl	$tbl,AES_Td
+	lt${g}r	$len,$len
+	stm	$s0,$s3,$tweak($sp)	# save the tweak
+	jz	.Lxts_dec_short
+	j	.Lxts_dec_enter
+
+.align	16
+.Lxts_dec_loop:
+	lrvg	$s1,$tweak+0($sp)	# load the tweak in little-endian
+	lrvg	$s3,$tweak+8($sp)
+	lghi	%r1,0x87
+	srag	%r0,$s3,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	srlg	%r0,$s1,63		# carry bit from lower half
+	sllg	$s1,$s1,1
+	sllg	$s3,$s3,1
+	xgr	$s1,%r1
+	ogr	$s3,%r0
+	lrvgr	$s1,$s1			# flip byte order
+	lrvgr	$s3,$s3
+	srlg	$s0,$s1,32		# smash the tweak to 4x32-bits 
+	stg	$s1,$tweak+0($sp)	# save the tweak
+	llgfr	$s1,$s1
+	srlg	$s2,$s3,32
+	stg	$s3,$tweak+8($sp)
+	llgfr	$s3,$s3
+.Lxts_dec_enter:
+	x	$s0,0($inp)		# tweak^=*(inp)
+	x	$s1,4($inp)
+	x	$s2,8($inp)
+	x	$s3,12($inp)
+	stm${g}	%r2,%r3,2*$SIZE_T($sp)	# only two registers are changing
+	la	$key,0($key1)
+	bras	$ra,_s390x_AES_decrypt
+	lm${g}	%r2,%r5,2*$SIZE_T($sp)
+	x	$s0,$tweak+0($sp)	# ^=tweak
+	x	$s1,$tweak+4($sp)
+	x	$s2,$tweak+8($sp)
+	x	$s3,$tweak+12($sp)
+	st	$s0,0($out,$inp)
+	st	$s1,4($out,$inp)
+	st	$s2,8($out,$inp)
+	st	$s3,12($out,$inp)
 	la	$inp,16($inp)
-	ahi	$t1,1		# 32-bit increment
-	brct	$len,.Lctr32_loop
+	brct${g}	$len,.Lxts_dec_loop
+
+	llgc	$len,`2*$SIZE_T-1`($sp)
+	nill	$len,0x0f		# $len%16
+	jz	.Lxts_dec_done
+
+	# generate pair of tweaks...
+	lrvg	$s1,$tweak+0($sp)	# load the tweak in little-endian
+	lrvg	$s3,$tweak+8($sp)
+	lghi	%r1,0x87
+	srag	%r0,$s3,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	srlg	%r0,$s1,63		# carry bit from lower half
+	sllg	$s1,$s1,1
+	sllg	$s3,$s3,1
+	xgr	$s1,%r1
+	ogr	$s3,%r0
+	lrvgr	$i2,$s1			# flip byte order
+	lrvgr	$i3,$s3
+	stmg	$i2,$i3,$tweak($sp)	# save the 1st tweak
+	j	.Lxts_dec_2ndtweak
+
+.align	16
+.Lxts_dec_short:
+	llgc	$len,`2*$SIZE_T-1`($sp)
+	nill	$len,0x0f		# $len%16
+	lrvg	$s1,$tweak+0($sp)	# load the tweak in little-endian
+	lrvg	$s3,$tweak+8($sp)
+.Lxts_dec_2ndtweak:
+	lghi	%r1,0x87
+	srag	%r0,$s3,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	srlg	%r0,$s1,63		# carry bit from lower half
+	sllg	$s1,$s1,1
+	sllg	$s3,$s3,1
+	xgr	$s1,%r1
+	ogr	$s3,%r0
+	lrvgr	$s1,$s1			# flip byte order
+	lrvgr	$s3,$s3
+	srlg	$s0,$s1,32		# smash the tweak to 4x32-bits
+	stg	$s1,$tweak-16+0($sp)	# save the 2nd tweak
+	llgfr	$s1,$s1
+	srlg	$s2,$s3,32
+	stg	$s3,$tweak-16+8($sp)
+	llgfr	$s3,$s3
+
+	x	$s0,0($inp)		# tweak_the_2nd^=*(inp)
+	x	$s1,4($inp)
+	x	$s2,8($inp)
+	x	$s3,12($inp)
+	stm${g}	%r2,%r3,2*$SIZE_T($sp)
+	la	$key,0($key1)
+	bras	$ra,_s390x_AES_decrypt
+	lm${g}	%r2,%r5,2*$SIZE_T($sp)
+	x	$s0,$tweak-16+0($sp)	# ^=tweak_the_2nd
+	x	$s1,$tweak-16+4($sp)
+	x	$s2,$tweak-16+8($sp)
+	x	$s3,$tweak-16+12($sp)
+	st	$s0,0($out,$inp)
+	st	$s1,4($out,$inp)
+	st	$s2,8($out,$inp)
+	st	$s3,12($out,$inp)
 
+	la	$i3,0($out,$inp)	# put aside real $out
+.Lxts_dec_steal:
+	llgc	%r0,16($inp)
+	llgc	%r1,0($out,$inp)
+	stc	%r0,0($out,$inp)
+	stc	%r1,16($out,$inp)
+	la	$inp,1($inp)
+	brct	$len,.Lxts_dec_steal
+	la	$out,0($i3)		# restore real $out
+
+	lm	$s0,$s3,$tweak($sp)	# load the 1st tweak
+	x	$s0,0($out)		# tweak^=*(inp)|stolen cipher-text
+	x	$s1,4($out)
+	x	$s2,8($out)
+	x	$s3,12($out)
+	st${g}	$out,4*$SIZE_T($sp)
+	la	$key,0($key1)
+	bras	$ra,_s390x_AES_decrypt
+	l${g}	$out,4*$SIZE_T($sp)
+	x	$s0,$tweak+0($sp)	# ^=tweak
+	x	$s1,$tweak+4($sp)
+	x	$s2,$tweak+8($sp)
+	x	$s3,$tweak+12($sp)
+	st	$s0,0($out)
+	st	$s1,4($out)
+	st	$s2,8($out)
+	st	$s3,12($out)
+	stg	$sp,$tweak-16+0($sp)	# wipe 2nd tweak
+	stg	$sp,$tweak-16+8($sp)
+.Lxts_dec_done:
+	stg	$sp,$tweak+0($sp)	# wipe tweak
+	stg	$sp,$twesk+8($sp)
 	lm${g}	%r6,$ra,6*$SIZE_T($sp)
 	br	$ra
-.size	AES_ctr32_encrypt,.-AES_ctr32_encrypt
+.size	AES_xts_decrypt,.-AES_xts_decrypt
 ___
 }
 $code.=<<___;
-.comm  OPENSSL_s390xcap_P,16,8
 .string	"AES for s390x, CRYPTOGAMS by <appro\@openssl.org>"
+.comm	OPENSSL_s390xcap_P,16,8
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;
diff --git a/crypto/bn/asm/s390x-mont.pl b/crypto/bn/asm/s390x-mont.pl
index 0c5f0638e1..9fd64e81ee 100644
--- a/crypto/bn/asm/s390x-mont.pl
+++ b/crypto/bn/asm/s390x-mont.pl
@@ -41,8 +41,8 @@
 # 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, less for
-# longer keys...
+# On z990 it was measured to perform 2.6-2.2 times better than
+# compiler-generated code, less for longer keys...
 
 $flavour = shift;
 
@@ -102,8 +102,8 @@ $code.=<<___ if ($flavour =~ /3[12]/);
 	bnzr	%r14		# if ($num&1) return 0;
 ___
 $code.=<<___ if ($flavour !~ /3[12]/);
-	cghi	$num,128	#
-	bhr	%r14		# if($num>128) return 0;
+	cghi	$num,96		#
+	bhr	%r14		# if($num>96) return 0;
 ___
 $code.=<<___;
 	stm${g}	%r3,%r15,3*$SIZE_T($sp)
diff --git a/crypto/modes/asm/ghash-s390x.pl b/crypto/modes/asm/ghash-s390x.pl
index 16ad034fc1..48cb08d338 100644
--- a/crypto/modes/asm/ghash-s390x.pl
+++ b/crypto/modes/asm/ghash-s390x.pl
@@ -28,6 +28,15 @@
 # remains z/Architecture specific. On z990 it was measured to perform
 # 2.8x better than 32-bit code generated by gcc 4.3.
 
+# March 2011.
+#
+# Support for hardware KIMD-GHASH is verified to produce correct
+# result and therefore is engaged. On z196 it was measured to process
+# 8KB buffer ~7 faster than software implementation. It's not as
+# impressive for smaller buffer sizes and for smallest 16-bytes buffer
+# it's actually almost 2 times slower. Which is the reason why
+# KIMD-GHASH is not used in gcm_gmult_4bit.
+
 $flavour = shift;
 
 if ($flavour =~ /3[12]/) {
@@ -41,7 +50,7 @@ if ($flavour =~ /3[12]/) {
 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
-$softonly=1;	# disable hardware support for now
+$softonly=0;
 
 $Zhi="%r0";
 $Zlo="%r1";
@@ -70,7 +79,7 @@ $code.=<<___;
 .align	32
 gcm_gmult_4bit:
 ___
-$code.=<<___ if(!$softonly);
+$code.=<<___ if(!$softonly && 0);	# hardware is slow for single block...
 	larl	%r1,OPENSSL_s390xcap_P
 	lg	%r0,0(%r1)
 	tmhl	%r0,0x4000	# check for message-security-assist