From 9833757b5d8683cc2e92ab45115794bc2bd9e66c Mon Sep 17 00:00:00 2001
From: Andy Polyakov <appro@openssl.org>
Date: Mon, 14 Nov 2011 20:47:22 +0000
Subject: [PATCH] s390x assembler pack update from HEAD.

---
 crypto/aes/asm/aes-s390x.pl    | 1064 +++++++++++++++++++++++++++++---
 crypto/bn/asm/s390x-gf2m.pl    |  221 +++++++
 crypto/bn/asm/s390x-mont.pl    |  102 ++-
 crypto/rc4/asm/rc4-s390x.pl    |   49 +-
 crypto/s390xcap.c              |   12 +-
 crypto/s390xcpuid.S            |   17 +-
 crypto/sha/asm/sha1-s390x.pl   |   50 +-
 crypto/sha/asm/sha512-s390x.pl |   63 +-
 8 files changed, 1423 insertions(+), 155 deletions(-)
 create mode 100644 crypto/bn/asm/s390x-gf2m.pl

diff --git a/crypto/aes/asm/aes-s390x.pl b/crypto/aes/asm/aes-s390x.pl
index c15059cd86..687bdd9499 100644
--- a/crypto/aes/asm/aes-s390x.pl
+++ b/crypto/aes/asm/aes-s390x.pl
@@ -28,7 +28,7 @@
 
 # May 2007.
 #
-# Implement private_AES_set_[en|de]crypt_key. Key schedule setup is avoided
+# Implement AES_set_[en|de]crypt_key. Key schedule setup is avoided
 # for 128-bit keys, if hardware support is detected.
 
 # Januray 2009.
@@ -44,12 +44,57 @@
 # Unlike previous version hardware support detection takes place only
 # at the moment of key schedule setup, which is denoted in key->rounds.
 # This is done, because deferred key setup can't be made MT-safe, not
-# for key lengthes longer than 128 bits.
+# for keys longer than 128 bits.
 #
 # Add AES_cbc_encrypt, which gives incredible performance improvement,
 # it was measured to be ~6.6x. It's less than previously mentioned 8x,
 # because software implementation was optimized.
 
+# May 2010.
+#
+# Add AES_ctr32_encrypt. If hardware-assisted, it provides up to 4.3x
+# performance improvement over "generic" counter mode routine relying
+# on single-block, also hardware-assisted, AES_encrypt. "Up to" refers
+# to the fact that exact throughput value depends on current stack
+# frame alignment within 4KB page. In worst case you get ~75% of the
+# maximum, but *on average* it would be as much as ~98%. Meaning that
+# worst case is unlike, it's like hitting ravine on plateau.
+
+# 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
+# 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, 37% - at most like 512-bytes block size.
+
+$flavour = shift;
+
+if ($flavour =~ /3[12]/) {
+	$SIZE_T=4;
+	$g="";
+} else {
+	$SIZE_T=8;
+	$g="g";
+}
+
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
+open STDOUT,">$output";
+
 $softonly=0;	# allow hardware support
 
 $t0="%r0";	$mask="%r0";
@@ -69,6 +114,8 @@ $rounds="%r13";
 $ra="%r14";
 $sp="%r15";
 
+$stdframe=16*$SIZE_T+4*8;
+
 sub _data_word()
 { my $i;
     while(defined($i=shift)) { $code.=sprintf".long\t0x%08x,0x%08x\n",$i,$i; }
@@ -210,7 +257,7 @@ $code.=<<___ if (!$softonly);
 .Lesoft:
 ___
 $code.=<<___;
-	stmg	%r3,$ra,24($sp)
+	stm${g}	%r3,$ra,3*$SIZE_T($sp)
 
 	llgf	$s0,0($inp)
 	llgf	$s1,4($inp)
@@ -220,20 +267,20 @@ $code.=<<___;
 	larl	$tbl,AES_Te
 	bras	$ra,_s390x_AES_encrypt
 
-	lg	$out,24($sp)
+	l${g}	$out,3*$SIZE_T($sp)
 	st	$s0,0($out)
 	st	$s1,4($out)
 	st	$s2,8($out)
 	st	$s3,12($out)
 
-	lmg	%r6,$ra,48($sp)
+	lm${g}	%r6,$ra,6*$SIZE_T($sp)
 	br	$ra
 .size	AES_encrypt,.-AES_encrypt
 
 .type   _s390x_AES_encrypt,\@function
 .align	16
 _s390x_AES_encrypt:
-	stg	$ra,152($sp)
+	st${g}	$ra,15*$SIZE_T($sp)
 	x	$s0,0($key)
 	x	$s1,4($key)
 	x	$s2,8($key)
@@ -397,7 +444,7 @@ _s390x_AES_encrypt:
 	or	$s2,$i3
 	or	$s3,$t3
 
-	lg	$ra,152($sp)
+	l${g}	$ra,15*$SIZE_T($sp)
 	xr	$s0,$t0
 	xr	$s1,$t2
 	x	$s2,24($key)
@@ -536,7 +583,7 @@ $code.=<<___ if (!$softonly);
 .Ldsoft:
 ___
 $code.=<<___;
-	stmg	%r3,$ra,24($sp)
+	stm${g}	%r3,$ra,3*$SIZE_T($sp)
 
 	llgf	$s0,0($inp)
 	llgf	$s1,4($inp)
@@ -546,20 +593,20 @@ $code.=<<___;
 	larl	$tbl,AES_Td
 	bras	$ra,_s390x_AES_decrypt
 
-	lg	$out,24($sp)
+	l${g}	$out,3*$SIZE_T($sp)
 	st	$s0,0($out)
 	st	$s1,4($out)
 	st	$s2,8($out)
 	st	$s3,12($out)
 
-	lmg	%r6,$ra,48($sp)
+	lm${g}	%r6,$ra,6*$SIZE_T($sp)
 	br	$ra
 .size	AES_decrypt,.-AES_decrypt
 
 .type   _s390x_AES_decrypt,\@function
 .align	16
 _s390x_AES_decrypt:
-	stg	$ra,152($sp)
+	st${g}	$ra,15*$SIZE_T($sp)
 	x	$s0,0($key)
 	x	$s1,4($key)
 	x	$s2,8($key)
@@ -703,7 +750,7 @@ _s390x_AES_decrypt:
 	nr	$i1,$mask
 	nr	$i2,$mask
 
-	lg	$ra,152($sp)
+	l${g}	$ra,15*$SIZE_T($sp)
 	or	$s1,$t1
 	l	$t0,16($key)
 	l	$t1,20($key)
@@ -730,16 +777,16 @@ _s390x_AES_decrypt:
 ___
 
 $code.=<<___;
-# void private_AES_set_encrypt_key(const unsigned char *in, int bits,
+# void AES_set_encrypt_key(const unsigned char *in, int bits,
 # 		 AES_KEY *key) {
-.globl	private_AES_set_encrypt_key
-.type	private_AES_set_encrypt_key,\@function
+.globl	AES_set_encrypt_key
+.type	AES_set_encrypt_key,\@function
 .align	16
-private_AES_set_encrypt_key:
+AES_set_encrypt_key:
 	lghi	$t0,0
-	clgr	$inp,$t0
+	cl${g}r	$inp,$t0
 	je	.Lminus1
-	clgr	$key,$t0
+	cl${g}r	$key,$t0
 	je	.Lminus1
 
 	lghi	$t0,128
@@ -797,7 +844,7 @@ ___
 $code.=<<___;
 .align	16
 .Lekey_internal:
-	stmg	%r6,%r13,48($sp)	# all non-volatile regs
+	stm${g}	%r6,%r13,6*$SIZE_T($sp)	# all non-volatile regs
 
 	larl	$tbl,AES_Te+2048
 
@@ -858,7 +905,7 @@ $code.=<<___;
 	la	$t3,4($t3)		# i++
 	brct	$rounds,.L128_loop
 	lghi	%r2,0
-	lmg	%r6,%r13,48($sp)
+	lm${g}	%r6,%r13,6*$SIZE_T($sp)
 	br	$ra
 
 .align	16
@@ -906,7 +953,7 @@ $code.=<<___;
 	st	$s3,36($key)
 	brct	$rounds,.L192_continue
 	lghi	%r2,0
-	lmg	%r6,%r13,48($sp)
+	lm${g}	%r6,%r13,6*$SIZE_T($sp)
 	br	$ra
 
 .align	16
@@ -968,7 +1015,7 @@ $code.=<<___;
 	st	$s3,44($key)
 	brct	$rounds,.L256_continue
 	lghi	%r2,0
-	lmg	%r6,%r13,48($sp)
+	lm${g}	%r6,%r13,6*$SIZE_T($sp)
 	br	$ra
 
 .align	16
@@ -1011,19 +1058,19 @@ $code.=<<___;
 .Lminus1:
 	lghi	%r2,-1
 	br	$ra
-.size	private_AES_set_encrypt_key,.-private_AES_set_encrypt_key
+.size	AES_set_encrypt_key,.-AES_set_encrypt_key
 
-# void private_AES_set_decrypt_key(const unsigned char *in, int bits,
+# void AES_set_decrypt_key(const unsigned char *in, int bits,
 # 		 AES_KEY *key) {
-.globl	private_AES_set_decrypt_key
-.type	private_AES_set_decrypt_key,\@function
+.globl	AES_set_decrypt_key
+.type	AES_set_decrypt_key,\@function
 .align	16
-private_AES_set_decrypt_key:
-	stg	$key,32($sp)		# I rely on private_AES_set_encrypt_key to
-	stg	$ra,112($sp)		# save non-volatile registers!
-	bras	$ra,private_AES_set_encrypt_key
-	lg	$key,32($sp)
-	lg	$ra,112($sp)
+AES_set_decrypt_key:
+	st${g}	$key,4*$SIZE_T($sp)	# I rely on AES_set_encrypt_key to
+	st${g}	$ra,14*$SIZE_T($sp)	# save non-volatile registers!
+	bras	$ra,AES_set_encrypt_key
+	l${g}	$key,4*$SIZE_T($sp)
+	l${g}	$ra,14*$SIZE_T($sp)
 	ltgr	%r2,%r2
 	bnzr	$ra
 ___
@@ -1038,11 +1085,11 @@ $code.=<<___ if (!$softonly);
 
 .align	16
 .Ldkey_internal:
-	stg	$key,32($sp)
-	stg	$ra,40($sp)
+	st${g}	$key,4*$SIZE_T($sp)
+	st${g}	$ra,14*$SIZE_T($sp)
 	bras	$ra,.Lekey_internal
-	lg	$key,32($sp)
-	lg	$ra,40($sp)
+	l${g}	$key,4*$SIZE_T($sp)
+	l${g}	$ra,14*$SIZE_T($sp)
 ___
 $code.=<<___;
 
@@ -1123,13 +1170,14 @@ $code.=<<___;
 	la	$key,4($key)
 	brct	$rounds,.Lmix
 
-	lmg	%r6,%r13,48($sp)# as was saved by private_AES_set_encrypt_key!
+	lm${g}	%r6,%r13,6*$SIZE_T($sp)# as was saved by AES_set_encrypt_key!
 	lghi	%r2,0
 	br	$ra
-.size	private_AES_set_decrypt_key,.-private_AES_set_decrypt_key
+.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)
 {
@@ -1163,7 +1211,7 @@ $code.=<<___ if (!$softonly);
 	l	%r0,240($key)	# load kmc code
 	lghi	$key,15		# res=len%16, len-=res;
 	ngr	$key,$len
-	slgr	$len,$key
+	sl${g}r	$len,$key
 	la	%r1,16($sp)	# parameter block - ivec || key
 	jz	.Lkmc_truncated
 	.long	0xb92f0042	# kmc %r4,%r2
@@ -1181,34 +1229,34 @@ $code.=<<___ if (!$softonly);
 	tmll	%r0,0x80
 	jnz	.Lkmc_truncated_dec
 	lghi	%r1,0
-	stg	%r1,128($sp)
-	stg	%r1,136($sp)
+	stg	%r1,16*$SIZE_T($sp)
+	stg	%r1,16*$SIZE_T+8($sp)
 	bras	%r1,1f
-	mvc	128(1,$sp),0($inp)
+	mvc	16*$SIZE_T(1,$sp),0($inp)
 1:	ex	$key,0(%r1)
 	la	%r1,16($sp)	# restore parameter block
-	la	$inp,128($sp)
+	la	$inp,16*$SIZE_T($sp)
 	lghi	$len,16
 	.long	0xb92f0042	# kmc %r4,%r2
 	j	.Lkmc_done
 .align	16
 .Lkmc_truncated_dec:
-	stg	$out,64($sp)
-	la	$out,128($sp)
+	st${g}	$out,4*$SIZE_T($sp)
+	la	$out,16*$SIZE_T($sp)
 	lghi	$len,16
 	.long	0xb92f0042	# kmc %r4,%r2
-	lg	$out,64($sp)
+	l${g}	$out,4*$SIZE_T($sp)
 	bras	%r1,2f
-	mvc	0(1,$out),128($sp)
+	mvc	0(1,$out),16*$SIZE_T($sp)
 2:	ex	$key,0(%r1)
 	j	.Lkmc_done
 .align	16
 .Lcbc_software:
 ___
 $code.=<<___;
-	stmg	$key,$ra,40($sp)
+	stm${g}	$key,$ra,5*$SIZE_T($sp)
 	lhi	%r0,0
-	cl	%r0,164($sp)
+	cl	%r0,`$stdframe+$SIZE_T-4`($sp)
 	je	.Lcbc_decrypt
 
 	larl	$tbl,AES_Te
@@ -1219,10 +1267,10 @@ $code.=<<___;
 	llgf	$s3,12($ivp)
 
 	lghi	$t0,16
-	slgr	$len,$t0
+	sl${g}r	$len,$t0
 	brc	4,.Lcbc_enc_tail	# if borrow
 .Lcbc_enc_loop:
-	stmg	$inp,$out,16($sp)
+	stm${g}	$inp,$out,2*$SIZE_T($sp)
 	x	$s0,0($inp)
 	x	$s1,4($inp)
 	x	$s2,8($inp)
@@ -1231,7 +1279,7 @@ $code.=<<___;
 
 	bras	$ra,_s390x_AES_encrypt
 
-	lmg	$inp,$key,16($sp)
+	lm${g}	$inp,$key,2*$SIZE_T($sp)
 	st	$s0,0($out)
 	st	$s1,4($out)
 	st	$s2,8($out)
@@ -1240,33 +1288,33 @@ $code.=<<___;
 	la	$inp,16($inp)
 	la	$out,16($out)
 	lghi	$t0,16
-	ltgr	$len,$len
+	lt${g}r	$len,$len
 	jz	.Lcbc_enc_done
-	slgr	$len,$t0
+	sl${g}r	$len,$t0
 	brc	4,.Lcbc_enc_tail	# if borrow
 	j	.Lcbc_enc_loop
 .align	16
 .Lcbc_enc_done:
-	lg	$ivp,48($sp)
+	l${g}	$ivp,6*$SIZE_T($sp)
 	st	$s0,0($ivp)
 	st	$s1,4($ivp)	
 	st	$s2,8($ivp)
 	st	$s3,12($ivp)
 
-	lmg	%r7,$ra,56($sp)
+	lm${g}	%r7,$ra,7*$SIZE_T($sp)
 	br	$ra
 
 .align	16
 .Lcbc_enc_tail:
 	aghi	$len,15
 	lghi	$t0,0
-	stg	$t0,128($sp)
-	stg	$t0,136($sp)
+	stg	$t0,16*$SIZE_T($sp)
+	stg	$t0,16*$SIZE_T+8($sp)
 	bras	$t1,3f
-	mvc	128(1,$sp),0($inp)
+	mvc	16*$SIZE_T(1,$sp),0($inp)
 3:	ex	$len,0($t1)
 	lghi	$len,0
-	la	$inp,128($sp)
+	la	$inp,16*$SIZE_T($sp)
 	j	.Lcbc_enc_loop
 
 .align	16
@@ -1275,10 +1323,10 @@ $code.=<<___;
 
 	lg	$t0,0($ivp)
 	lg	$t1,8($ivp)
-	stmg	$t0,$t1,128($sp)
+	stmg	$t0,$t1,16*$SIZE_T($sp)
 
 .Lcbc_dec_loop:
-	stmg	$inp,$out,16($sp)
+	stm${g}	$inp,$out,2*$SIZE_T($sp)
 	llgf	$s0,0($inp)
 	llgf	$s1,4($inp)
 	llgf	$s2,8($inp)
@@ -1287,7 +1335,7 @@ $code.=<<___;
 
 	bras	$ra,_s390x_AES_decrypt
 
-	lmg	$inp,$key,16($sp)
+	lm${g}	$inp,$key,2*$SIZE_T($sp)
 	sllg	$s0,$s0,32
 	sllg	$s2,$s2,32
 	lr	$s0,$s1
@@ -1295,15 +1343,15 @@ $code.=<<___;
 
 	lg	$t0,0($inp)
 	lg	$t1,8($inp)
-	xg	$s0,128($sp)
-	xg	$s2,136($sp)
+	xg	$s0,16*$SIZE_T($sp)
+	xg	$s2,16*$SIZE_T+8($sp)
 	lghi	$s1,16
-	slgr	$len,$s1
+	sl${g}r	$len,$s1
 	brc	4,.Lcbc_dec_tail	# if borrow
 	brc	2,.Lcbc_dec_done	# if zero
 	stg	$s0,0($out)
 	stg	$s2,8($out)
-	stmg	$t0,$t1,128($sp)
+	stmg	$t0,$t1,16*$SIZE_T($sp)
 
 	la	$inp,16($inp)
 	la	$out,16($out)
@@ -1313,7 +1361,7 @@ $code.=<<___;
 	stg	$s0,0($out)
 	stg	$s2,8($out)
 .Lcbc_dec_exit:
-	lmg	$ivp,$ra,48($sp)
+	lm${g}	%r6,$ra,6*$SIZE_T($sp)
 	stmg	$t0,$t1,0($ivp)
 
 	br	$ra
@@ -1321,19 +1369,889 @@ $code.=<<___;
 .align	16
 .Lcbc_dec_tail:
 	aghi	$len,15
-	stg	$s0,128($sp)
-	stg	$s2,136($sp)
+	stg	$s0,16*$SIZE_T($sp)
+	stg	$s2,16*$SIZE_T+8($sp)
 	bras	$s1,4f
-	mvc	0(1,$out),128($sp)
+	mvc	0(1,$out),16*$SIZE_T($sp)
 4:	ex	$len,0($s1)
 	j	.Lcbc_dec_exit
 .size	AES_cbc_encrypt,.-AES_cbc_encrypt
-.comm  OPENSSL_s390xcap_P,8,8
+___
+}
+########################################################################
+# 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="%r4";	# blocks and out are swapped
+my $len="%r3";
+my $key="%r5";	my $iv0="%r5";
+my $ivp="%r6";
+my $fp ="%r7";
+
+$code.=<<___;
+.globl	AES_ctr32_encrypt
+.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);
+	l	%r0,240($key)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lctr32_software
+
+	stm${g}	%r6,$s3,6*$SIZE_T($sp)
+
+	slgr	$out,$inp
+	la	%r1,0($key)	# %r1 is permanent copy of $key
+	lg	$iv0,0($ivp)	# load ivec
+	lg	$ivp,8($ivp)
+
+	# 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
+	srlg	$fp,$fp,4	# convert bytes to blocks, minimum 16
+	st${g}	$s2,0($sp)	# back-chain
+	st${g}	$fp,$SIZE_T($sp)
+
+	slgr	$len,$fp
+	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($sp)
+
+.Lctr32_hw_switch:
+___
+$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_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_km_prepare
+
+	la	$s0,16($sp)	# inp
+	sllg	$s1,$fp,4	# len
+	la	$s2,16($sp)	# out
+	.long	0xb92e00a8	# km %r10,%r8
+	brc	1,.-4		# pay attention to "partial completion"
+
+	la	$s2,16($sp)
+	lgr	$s3,$fp
+	slgr	$s2,$inp
+.Lctr32_km_xor:
+	lg	$s0,0($inp)
+	lg	$s1,8($inp)
+	xg	$s0,0($s2,$inp)
+	xg	$s1,8($s2,$inp)
+	stg	$s0,0($out,$inp)
+	stg	$s1,8($out,$inp)
+	la	$inp,16($inp)
+	brct	$s3,.Lctr32_km_xor
+
+	slgr	$len,$fp
+	brc	1,.Lctr32_km_loop	# not zero, no borrow
+	algr	$fp,$len
+	lghi	$len,0
+	brc	4+1,.Lctr32_km_loop	# not zero
+
+	l${g}	$s0,0($sp)
+	l${g}	$s1,$SIZE_T($sp)
+	la	$s2,16($sp)
+.Lctr32_km_zap:
+	stg	$s0,0($s2)
+	stg	$s0,8($s2)
+	la	$s2,16($s2)
+	brct	$s1,.Lctr32_km_zap
+
+	la	$sp,0($s0)
+	lm${g}	%r6,$s3,6*$SIZE_T($sp)
+	br	$ra
+.align	16
+.Lctr32_software:
+___
+$code.=<<___;
+	stm${g}	$key,$ra,5*$SIZE_T($sp)
+	sl${g}r	$inp,$out
+	larl	$tbl,AES_Te
+	llgf	$t1,12($ivp)
+
+.Lctr32_loop:
+	stm${g}	$inp,$out,2*$SIZE_T($sp)
+	llgf	$s0,0($ivp)
+	llgf	$s1,4($ivp)
+	llgf	$s2,8($ivp)
+	lgr	$s3,$t1
+	st	$t1,16*$SIZE_T($sp)
+	lgr	%r4,$key
+
+	bras	$ra,_s390x_AES_encrypt
+
+	lm${g}	$inp,$ivp,2*$SIZE_T($sp)
+	llgf	$t1,16*$SIZE_T($sp)
+	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,
+#	const unsigned char iv[16]);
+#
+{
+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(1);
+	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
+
+	# generate the tweak value
+	l${g}	$s3,$stdframe($sp)	# pointer to iv
+	la	$s2,$tweak($sp)
+	lmg	$s0,$s1,0($s3)
+	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
+	l${g}	$s3,$stdframe($sp)	# pointer to iv
+	la	$s2,$tweak($sp)
+	lmg	$s0,$s1,0($s3)
+	lghi	$s3,16
+	stmg	$s0,$s1,0($s2)
+	la	%r1,0($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)
+	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_xts_decrypt,.-AES_xts_decrypt
 ___
 }
 $code.=<<___;
 .string	"AES for s390x, CRYPTOGAMS by <appro\@openssl.org>"
+.comm	OPENSSL_s390xcap_P,16,8
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;
 print $code;
+close STDOUT;	# force flush
diff --git a/crypto/bn/asm/s390x-gf2m.pl b/crypto/bn/asm/s390x-gf2m.pl
new file mode 100644
index 0000000000..cd9f13eca2
--- /dev/null
+++ b/crypto/bn/asm/s390x-gf2m.pl
@@ -0,0 +1,221 @@
+#!/usr/bin/env perl
+#
+# ====================================================================
+# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
+#
+# May 2011
+#
+# The module implements bn_GF2m_mul_2x2 polynomial multiplication used
+# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for
+# the time being... gcc 4.3 appeared to generate poor code, therefore
+# the effort. And indeed, the module delivers 55%-90%(*) improvement
+# on haviest ECDSA verify and ECDH benchmarks for 163- and 571-bit
+# key lengths on z990, 30%-55%(*) - on z10, and 70%-110%(*) - on z196.
+# This is for 64-bit build. In 32-bit "highgprs" case improvement is
+# even higher, for example on z990 it was measured 80%-150%. ECDSA
+# sign is modest 9%-12% faster. Keep in mind that these coefficients
+# are not ones for bn_GF2m_mul_2x2 itself, as not all CPU time is
+# burnt in it...
+#
+# (*)	gcc 4.1 was observed to deliver better results than gcc 4.3,
+#	so that improvement coefficients can vary from one specific
+#	setup to another.
+
+$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			# a1·b1
+	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			# a0·b0
+	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;
diff --git a/crypto/bn/asm/s390x-mont.pl b/crypto/bn/asm/s390x-mont.pl
index f61246f5b6..9fd64e81ee 100644
--- a/crypto/bn/asm/s390x-mont.pl
+++ b/crypto/bn/asm/s390x-mont.pl
@@ -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
@@ -207,19 +254,24 @@ 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
+.Lcopy:	lg	$alo,0($j,$ap)		# copy or in-place refresh
+	_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;
diff --git a/crypto/rc4/asm/rc4-s390x.pl b/crypto/rc4/asm/rc4-s390x.pl
index c7ed59510b..1aa754820c 100644
--- a/crypto/rc4/asm/rc4-s390x.pl
+++ b/crypto/rc4/asm/rc4-s390x.pl
@@ -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
diff --git a/crypto/s390xcap.c b/crypto/s390xcap.c
index ffbe0235f9..f2e94ef47e 100644
--- a/crypto/s390xcap.c
+++ b/crypto/s390xcap.c
@@ -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);
diff --git a/crypto/s390xcpuid.S b/crypto/s390xcpuid.S
index b053c6a281..06815347e6 100644
--- a/crypto/s390xcpuid.S
+++ b/crypto/s390xcpuid.S
@@ -5,10 +5,14 @@
 .align	16
 OPENSSL_s390x_facilities:
 	lghi	%r0,0
-	.long	0xb2b0f010	# stfle	16(%r15)
-	lg	%r2,16(%r15)
-	larl	%r1,OPENSSL_s390xcap_P
-	stg	%r2,0(%r1)
+	larl	%r2,OPENSSL_s390xcap_P
+	stg	%r0,8(%r2)
+	.long	0xb2b02000	# stfle	0(%r2)
+	brc	8,.Ldone
+	lghi	%r0,1
+	.long	0xb2b02000	# stfle 0(%r2)
+.Ldone:
+	lg	%r2,0(%r2)
 	br	%r14
 .size	OPENSSL_s390x_facilities,.-OPENSSL_s390x_facilities
 
@@ -58,6 +62,9 @@ OPENSSL_wipe_cpu:
 .type	OPENSSL_cleanse,@function
 .align	16
 OPENSSL_cleanse:
+#if !defined(__s390x__) && !defined(__s390x)
+	llgfr	%r3,%r3
+#endif
 	lghi	%r4,15
 	lghi	%r0,0
 	clgr	%r3,%r4
@@ -89,4 +96,4 @@ OPENSSL_cleanse:
 .section	.init
 	brasl	%r14,OPENSSL_cpuid_setup
 
-.comm	OPENSSL_s390xcap_P,8,8
+.comm	OPENSSL_s390xcap_P,16,8
diff --git a/crypto/sha/asm/sha1-s390x.pl b/crypto/sha/asm/sha1-s390x.pl
index 4b17848287..9193dda45e 100644
--- a/crypto/sha/asm/sha1-s390x.pl
+++ b/crypto/sha/asm/sha1-s390x.pl
@@ -21,9 +21,28 @@
 # instructions to favour dual-issue z10 pipeline. On z10 hardware is
 # "only" ~2.3x faster than software.
 
+# 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.
+
 $kimdfunc=1;	# magic function code for kimd instruction
 
-$output=shift;
+$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";
 
 $K_00_39="%r0"; $K=$K_00_39;
@@ -42,13 +61,14 @@ $t1="%r11";
 @X=("%r12","%r13","%r14");
 $sp="%r15";
 
-$frame=160+16*4;
+$stdframe=16*$SIZE_T+4*8;
+$frame=$stdframe+16*4;
 
 sub Xupdate {
 my $i=shift;
 
 $code.=<<___ if ($i==15);
-	lg	$prefetch,160($sp)	### Xupdate(16) warm-up
+	lg	$prefetch,$stdframe($sp)	### Xupdate(16) warm-up
 	lr	$X[0],$X[2]
 ___
 return if ($i&1);	# Xupdate is vectorized and executed every 2nd cycle
@@ -58,8 +78,8 @@ $code.=<<___ if ($i<16);
 ___
 $code.=<<___ if ($i>=16);
 	xgr	$X[0],$prefetch		### Xupdate($i)
-	lg	$prefetch,`160+4*(($i+2)%16)`($sp)
-	xg	$X[0],`160+4*(($i+8)%16)`($sp)
+	lg	$prefetch,`$stdframe+4*(($i+2)%16)`($sp)
+	xg	$X[0],`$stdframe+4*(($i+8)%16)`($sp)
 	xgr	$X[0],$prefetch
 	rll	$X[0],$X[0],1
 	rllg	$X[1],$X[0],32
@@ -68,7 +88,7 @@ $code.=<<___ if ($i>=16);
 	lr	$X[2],$X[1]		# feedback
 ___
 $code.=<<___ if ($i<=70);
-	stg	$X[0],`160+4*($i%16)`($sp)
+	stg	$X[0],`$stdframe+4*($i%16)`($sp)
 ___
 unshift(@X,pop(@X));
 }
@@ -148,9 +168,9 @@ $code.=<<___ if ($kimdfunc);
 	tmhl	%r0,0x4000	# check for message-security assist
 	jz	.Lsoftware
 	lghi	%r0,0
-	la	%r1,16($sp)
+	la	%r1,`2*$SIZE_T`($sp)
 	.long	0xb93e0002	# kimd %r0,%r2
-	lg	%r0,16($sp)
+	lg	%r0,`2*$SIZE_T`($sp)
 	tmhh	%r0,`0x8000>>$kimdfunc`
 	jz	.Lsoftware
 	lghi	%r0,$kimdfunc
@@ -165,11 +185,11 @@ $code.=<<___ if ($kimdfunc);
 ___
 $code.=<<___;
 	lghi	%r1,-$frame
-	stg	$ctx,16($sp)
-	stmg	%r6,%r15,48($sp)
+	st${g}	$ctx,`2*$SIZE_T`($sp)
+	stm${g}	%r6,%r15,`6*$SIZE_T`($sp)
 	lgr	%r0,$sp
 	la	$sp,0(%r1,$sp)
-	stg	%r0,0($sp)
+	st${g}	%r0,0($sp)
 
 	larl	$t0,Ktable
 	llgf	$A,0($ctx)
@@ -199,7 +219,7 @@ ___
 for (;$i<80;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
 $code.=<<___;
 
-	lg	$ctx,`$frame+16`($sp)
+	l${g}	$ctx,`$frame+2*$SIZE_T`($sp)
 	la	$inp,64($inp)
 	al	$A,0($ctx)
 	al	$B,4($ctx)
@@ -211,13 +231,13 @@ $code.=<<___;
 	st	$C,8($ctx)
 	st	$D,12($ctx)
 	st	$E,16($ctx)
-	brct	$len,.Lloop
+	brct${g} $len,.Lloop
 
-	lmg	%r6,%r15,`$frame+48`($sp)
+	lm${g}	%r6,%r15,`$frame+6*$SIZE_T`($sp)
 	br	%r14
 .size	sha1_block_data_order,.-sha1_block_data_order
 .string	"SHA1 block transform for s390x, CRYPTOGAMS by <appro\@openssl.org>"
-.comm	OPENSSL_s390xcap_P,8,8
+.comm	OPENSSL_s390xcap_P,16,8
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;
diff --git a/crypto/sha/asm/sha512-s390x.pl b/crypto/sha/asm/sha512-s390x.pl
index e7ef2d5a9f..079a3fc78a 100644
--- a/crypto/sha/asm/sha512-s390x.pl
+++ b/crypto/sha/asm/sha512-s390x.pl
@@ -26,6 +26,26 @@
 # favour dual-issue z10 pipeline. Hardware SHA256/512 is ~4.7x faster
 # than software.
 
+# 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 z900 SHA256 was measured to
+# perform 2.4x and SHA512 - 13x better than code generated by gcc 4.3.
+
+$flavour = shift;
+
+if ($flavour =~ /3[12]/) {
+	$SIZE_T=4;
+	$g="";
+} else {
+	$SIZE_T=8;
+	$g="g";
+}
+
 $t0="%r0";
 $t1="%r1";
 $ctx="%r2";	$t2="%r2";
@@ -44,7 +64,7 @@ $tbl="%r13";
 $T1="%r14";
 $sp="%r15";
 
-$output=shift;
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
 if ($output =~ /512/) {
@@ -78,7 +98,8 @@ if ($output =~ /512/) {
 }
 $Func="sha${label}_block_data_order";
 $Table="K${label}";
-$frame=160+16*$SZ;
+$stdframe=16*$SIZE_T+4*8;
+$frame=$stdframe+16*$SZ;
 
 sub BODY_00_15 {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
@@ -93,9 +114,9 @@ $code.=<<___;
 	xgr	$t0,$t1
 	$ROT	$t1,$t1,`$Sigma1[2]-$Sigma1[1]`
 	 xgr	$t2,$g
-	$ST	$T1,`160+$SZ*($i%16)`($sp)
+	$ST	$T1,`$stdframe+$SZ*($i%16)`($sp)
 	xgr	$t0,$t1			# Sigma1(e)
-	la	$T1,0($T1,$h)		# T1+=h
+	algr	$T1,$h			# T1+=h
 	 ngr	$t2,$e
 	 lgr	$t1,$a
 	algr	$T1,$t0			# T1+=Sigma1(e)
@@ -113,7 +134,7 @@ $code.=<<___;
 	 ngr	$t2,$b
 	algr	$h,$T1			# h+=T1
 	 ogr	$t2,$t1			# Maj(a,b,c)
-	la	$d,0($d,$T1)		# d+=T1
+	algr	$d,$T1			# d+=T1
 	algr	$h,$t2			# h+=Maj(a,b,c)
 ___
 }
@@ -122,19 +143,19 @@ sub BODY_16_XX {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 
 $code.=<<___;
-	$LD	$T1,`160+$SZ*(($i+1)%16)`($sp)	### $i
-	$LD	$t1,`160+$SZ*(($i+14)%16)`($sp)
+	$LD	$T1,`$stdframe+$SZ*(($i+1)%16)`($sp)	### $i
+	$LD	$t1,`$stdframe+$SZ*(($i+14)%16)`($sp)
 	$ROT	$t0,$T1,$sigma0[0]
 	$SHR	$T1,$sigma0[2]
 	$ROT	$t2,$t0,`$sigma0[1]-$sigma0[0]`
 	xgr	$T1,$t0
 	$ROT	$t0,$t1,$sigma1[0]
-	xgr	$T1,$t2				# sigma0(X[i+1])
+	xgr	$T1,$t2					# sigma0(X[i+1])
 	$SHR	$t1,$sigma1[2]
-	$ADD	$T1,`160+$SZ*($i%16)`($sp)	# +=X[i]
+	$ADD	$T1,`$stdframe+$SZ*($i%16)`($sp)	# +=X[i]
 	xgr	$t1,$t0
 	$ROT	$t0,$t0,`$sigma1[1]-$sigma1[0]`
-	$ADD	$T1,`160+$SZ*(($i+9)%16)`($sp)	# +=X[i+9]
+	$ADD	$T1,`$stdframe+$SZ*(($i+9)%16)`($sp)	# +=X[i+9]
 	xgr	$t1,$t0				# sigma1(X[i+14])
 	algr	$T1,$t1				# +=sigma1(X[i+14])
 ___
@@ -212,6 +233,7 @@ $code.=<<___;
 .globl	$Func
 .type	$Func,\@function
 $Func:
+	sllg	$len,$len,`log(16*$SZ)/log(2)`
 ___
 $code.=<<___ if ($kimdfunc);
 	larl	%r1,OPENSSL_s390xcap_P
@@ -219,15 +241,15 @@ $code.=<<___ if ($kimdfunc);
 	tmhl	%r0,0x4000	# check for message-security assist
 	jz	.Lsoftware
 	lghi	%r0,0
-	la	%r1,16($sp)
+	la	%r1,`2*$SIZE_T`($sp)
 	.long	0xb93e0002	# kimd %r0,%r2
-	lg	%r0,16($sp)
+	lg	%r0,`2*$SIZE_T`($sp)
 	tmhh	%r0,`0x8000>>$kimdfunc`
 	jz	.Lsoftware
 	lghi	%r0,$kimdfunc
 	lgr	%r1,$ctx
 	lgr	%r2,$inp
-	sllg	%r3,$len,`log(16*$SZ)/log(2)`
+	lgr	%r3,$len
 	.long	0xb93e0002	# kimd %r0,%r2
 	brc	1,.-4		# pay attention to "partial completion"
 	br	%r14
@@ -235,13 +257,12 @@ $code.=<<___ if ($kimdfunc);
 .Lsoftware:
 ___
 $code.=<<___;
-	sllg	$len,$len,`log(16*$SZ)/log(2)`
 	lghi	%r1,-$frame
-	agr	$len,$inp
-	stmg	$ctx,%r15,16($sp)
+	la	$len,0($len,$inp)
+	stm${g}	$ctx,%r15,`2*$SIZE_T`($sp)
 	lgr	%r0,$sp
 	la	$sp,0(%r1,$sp)
-	stg	%r0,0($sp)
+	st${g}	%r0,0($sp)
 
 	larl	$tbl,$Table
 	$LD	$A,`0*$SZ`($ctx)
@@ -265,7 +286,7 @@ $code.=<<___;
 	clgr	$len,$t0
 	jne	.Lrounds_16_xx
 
-	lg	$ctx,`$frame+16`($sp)
+	l${g}	$ctx,`$frame+2*$SIZE_T`($sp)
 	la	$inp,`16*$SZ`($inp)
 	$ADD	$A,`0*$SZ`($ctx)
 	$ADD	$B,`1*$SZ`($ctx)
@@ -283,14 +304,14 @@ $code.=<<___;
 	$ST	$F,`5*$SZ`($ctx)
 	$ST	$G,`6*$SZ`($ctx)
 	$ST	$H,`7*$SZ`($ctx)
-	clg	$inp,`$frame+32`($sp)
+	cl${g}	$inp,`$frame+4*$SIZE_T`($sp)
 	jne	.Lloop
 
-	lmg	%r6,%r15,`$frame+48`($sp)	
+	lm${g}	%r6,%r15,`$frame+6*$SIZE_T`($sp)	
 	br	%r14
 .size	$Func,.-$Func
 .string	"SHA${label} block transform for s390x, CRYPTOGAMS by <appro\@openssl.org>"
-.comm	OPENSSL_s390xcap_P,8,8
+.comm	OPENSSL_s390xcap_P,16,8
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;
-- 
2.25.1