On OpenBSD/armv7 we deliberately trap unaligned access. Unfortunately

the assembly code in libcrypto assumes unaligned access is allowed for
ARMv7.  Make these paths conditional on __STRICT_ALIGNMENT not being
defined and define __STRICT_ALIGNMENT in arm_arch.h for OpenBSD.

ok tom@

5 files changed, 408 insertions, 177 deletions

diff --git a/lib/libcrypto/aes/asm/aes-armv4.pl b/lib/libcrypto/aes/asm/aes-armv4.pl
index 717cc1ed7f0..1cb9586d4b9 100644
--- a/lib/libcrypto/aes/asm/aes-armv4.pl
+++ b/lib/libcrypto/aes/asm/aes-armv4.pl
@@ -172,7 +172,7 @@ AES_encrypt:
 	mov	$rounds,r0		@ inp
 	mov	$key,r2
 	sub	$tbl,r3,#AES_encrypt-AES_Te	@ Te
-#if __ARM_ARCH__<7
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
 	ldrb	$s0,[$rounds,#3]	@ load input data in endian-neutral
 	ldrb	$t1,[$rounds,#2]	@ manner...
 	ldrb	$t2,[$rounds,#1]
@@ -216,7 +216,7 @@ AES_encrypt:
 	bl	_armv4_AES_encrypt
 
 	ldr	$rounds,[sp],#4		@ pop out
-#if __ARM_ARCH__>=7
+#if __ARM_ARCH__>=7 && !defined(__STRICT_ALIGNMENT)
 #ifdef __ARMEL__
 	rev	$s0,$s0
 	rev	$s1,$s1
@@ -432,7 +432,7 @@ _armv4_AES_set_encrypt_key:
 	mov	lr,r1			@ bits
 	mov	$key,r2			@ key
 
-#if __ARM_ARCH__<7
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
 	ldrb	$s0,[$rounds,#3]	@ load input data in endian-neutral
 	ldrb	$t1,[$rounds,#2]	@ manner...
 	ldrb	$t2,[$rounds,#1]
@@ -517,7 +517,7 @@ _armv4_AES_set_encrypt_key:
 	b	.Ldone
 
 .Lnot128:
-#if __ARM_ARCH__<7
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
 	ldrb	$i2,[$rounds,#19]
 	ldrb	$t1,[$rounds,#18]
 	ldrb	$t2,[$rounds,#17]
@@ -588,7 +588,7 @@ _armv4_AES_set_encrypt_key:
 	b	.L192_loop
 
 .Lnot192:
-#if __ARM_ARCH__<7
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
 	ldrb	$i2,[$rounds,#27]
 	ldrb	$t1,[$rounds,#26]
 	ldrb	$t2,[$rounds,#25]
@@ -888,7 +888,7 @@ AES_decrypt:
 	mov	$rounds,r0		@ inp
 	mov	$key,r2
 	sub	$tbl,r3,#AES_decrypt-AES_Td		@ Td
-#if __ARM_ARCH__<7
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
 	ldrb	$s0,[$rounds,#3]	@ load input data in endian-neutral
 	ldrb	$t1,[$rounds,#2]	@ manner...
 	ldrb	$t2,[$rounds,#1]
@@ -932,7 +932,7 @@ AES_decrypt:
 	bl	_armv4_AES_decrypt
 
 	ldr	$rounds,[sp],#4		@ pop out
-#if __ARM_ARCH__>=7
+#if __ARM_ARCH__>=7 && !defined(__STRICT_ALIGNMENT)
 #ifdef __ARMEL__
 	rev	$s0,$s0
 	rev	$s1,$s1
diff --git a/lib/libcrypto/arm_arch.h b/lib/libcrypto/arm_arch.h
index 3304be81ab1..8b8a05b5f71 100644
--- a/lib/libcrypto/arm_arch.h
+++ b/lib/libcrypto/arm_arch.h
@@ -1,4 +1,4 @@
-/* $OpenBSD: arm_arch.h,v 1.7 2015/06/29 06:40:06 jsg Exp $ */
+/* $OpenBSD: arm_arch.h,v 1.8 2018/01/07 12:35:52 kettenis Exp $ */
 #ifndef __ARM_ARCH_H__
 #define __ARM_ARCH_H__
 
@@ -44,4 +44,8 @@ extern unsigned int OPENSSL_armcap_P;
 #define ARMV7_NEON      (1<<0)
 #endif
 
+#if defined(__OpenBSD__)
+#define __STRICT_ALIGNMENT
+#endif
+
 #endif
diff --git a/lib/libcrypto/sha/asm/sha1-armv4-large.pl b/lib/libcrypto/sha/asm/sha1-armv4-large.pl
index 88861af6411..8f0cdaf83c8 100644
--- a/lib/libcrypto/sha/asm/sha1-armv4-large.pl
+++ b/lib/libcrypto/sha/asm/sha1-armv4-large.pl
@@ -37,9 +37,22 @@
 #	modes are limited. As result it takes more instructions to do
 #	the same job in Thumb, therefore the code is never twice as
 #	small and always slower.
-# [***]	which is also ~35% better than compiler generated code.
+# [***]	which is also ~35% better than compiler generated code. Dual-
+#	issue Cortex A8 core was measured to process input block in
+#	~990 cycles.
 
-$output=shift;
+# August 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 13% improvement on
+# Cortex A8 core and in absolute terms ~870 cycles per input block
+# [or 13.6 cycles per byte].
+
+# February 2011.
+#
+# Profiler-assisted and platform-specific optimization resulted in 10%
+# improvement on Cortex A8 core and 12.2 cycles per byte.
+
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
 $ctx="r0";
@@ -58,63 +71,62 @@ $t3="r12";
 $Xi="r14";
 @V=($a,$b,$c,$d,$e);
 
-# One can optimize this for aligned access on big-endian architecture,
-# but code's endian neutrality makes it too pretty:-)
-sub Xload {
-my ($a,$b,$c,$d,$e)=@_;
-$code.=<<___;
-	ldrb	$t0,[$inp],#4
-	ldrb	$t1,[$inp,#-3]
-	ldrb	$t2,[$inp,#-2]
-	ldrb	$t3,[$inp,#-1]
-	add	$e,$K,$e,ror#2			@ E+=K_00_19
-	orr	$t0,$t1,$t0,lsl#8
-	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
-	orr	$t0,$t2,$t0,lsl#8
-	eor	$t1,$c,$d			@ F_xx_xx
-	orr	$t0,$t3,$t0,lsl#8
-	add	$e,$e,$t0			@ E+=X[i]
-	str	$t0,[$Xi,#-4]!
-___
-}
 sub Xupdate {
-my ($a,$b,$c,$d,$e,$flag)=@_;
+my ($a,$b,$c,$d,$e,$opt1,$opt2)=@_;
 $code.=<<___;
 	ldr	$t0,[$Xi,#15*4]
 	ldr	$t1,[$Xi,#13*4]
 	ldr	$t2,[$Xi,#7*4]
-	ldr	$t3,[$Xi,#2*4]
 	add	$e,$K,$e,ror#2			@ E+=K_xx_xx
+	ldr	$t3,[$Xi,#2*4]
 	eor	$t0,$t0,$t1
-	eor	$t0,$t0,$t2
-	eor	$t0,$t0,$t3
-	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
-___
-$code.=<<___ if (!defined($flag));
-	eor	$t1,$c,$d			@ F_xx_xx, but not in 40_59
-___
-$code.=<<___;
+	eor	$t2,$t2,$t3			@ 1 cycle stall
+	eor	$t1,$c,$d			@ F_xx_xx
 	mov	$t0,$t0,ror#31
-	add	$e,$e,$t0			@ E+=X[i]
+	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
+	eor	$t0,$t0,$t2,ror#31
 	str	$t0,[$Xi,#-4]!
+	$opt1					@ F_xx_xx
+	$opt2					@ F_xx_xx
+	add	$e,$e,$t0			@ E+=X[i]
 ___
 }
 
 sub BODY_00_15 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xload(@_);
 $code.=<<___;
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
+	ldrb	$t1,[$inp,#2]
+	ldrb	$t0,[$inp,#3]
+	ldrb	$t2,[$inp,#1]
+	add	$e,$K,$e,ror#2			@ E+=K_00_19
+	ldrb	$t3,[$inp],#4
+	orr	$t0,$t0,$t1,lsl#8
+	eor	$t1,$c,$d			@ F_xx_xx
+	orr	$t0,$t0,$t2,lsl#16
+	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
+	orr	$t0,$t0,$t3,lsl#24
+#else
+	ldr	$t0,[$inp],#4			@ handles unaligned
+	add	$e,$K,$e,ror#2			@ E+=K_00_19
+	eor	$t1,$c,$d			@ F_xx_xx
+	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
+#ifdef __ARMEL__
+	rev	$t0,$t0				@ byte swap
+#endif
+#endif
 	and	$t1,$b,$t1,ror#2
+	add	$e,$e,$t0			@ E+=X[i]
 	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
+	str	$t0,[$Xi,#-4]!
 	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
 ___
 }
 
 sub BODY_16_19 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xupdate(@_);
+	&Xupdate(@_,"and $t1,$b,$t1,ror#2");
 $code.=<<___;
-	and	$t1,$b,$t1,ror#2
 	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
 	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
 ___
@@ -122,26 +134,24 @@ ___
 
 sub BODY_20_39 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xupdate(@_);
+	&Xupdate(@_,"eor $t1,$b,$t1,ror#2");
 $code.=<<___;
-	eor	$t1,$b,$t1,ror#2		@ F_20_39(B,C,D)
 	add	$e,$e,$t1			@ E+=F_20_39(B,C,D)
 ___
 }
 
 sub BODY_40_59 {
 my ($a,$b,$c,$d,$e)=@_;
-	&Xupdate(@_,1);
+	&Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d");
 $code.=<<___;
-	and	$t1,$b,$c,ror#2
-	orr	$t2,$b,$c,ror#2
-	and	$t2,$t2,$d,ror#2
-	orr	$t1,$t1,$t2			@ F_40_59(B,C,D)
 	add	$e,$e,$t1			@ E+=F_40_59(B,C,D)
+	add	$e,$e,$t2,ror#2
 ___
 }
 
 $code=<<___;
+#include "arm_arch.h"
+
 .text
 
 .global	sha1_block_data_order
@@ -167,6 +177,7 @@ for($i=0;$i<5;$i++) {
 $code.=<<___;
 	teq	$Xi,sp
 	bne	.L_00_15		@ [((11+4)*5+2)*3]
+	sub	sp,sp,#25*4
 ___
 	&BODY_00_15(@V);	unshift(@V,pop(@V));
 	&BODY_16_19(@V);	unshift(@V,pop(@V));
@@ -176,7 +187,6 @@ ___
 $code.=<<___;
 
 	ldr	$K,.LK_20_39		@ [+15+16*4]
-	sub	sp,sp,#25*4
 	cmn	sp,#0			@ [+3], clear carry to denote 20_39
 .L_20_39_or_60_79:
 ___
@@ -215,10 +225,14 @@ $code.=<<___;
 	teq	$inp,$len
 	bne	.Lloop			@ [+18], total 1307
 
+#if __ARM_ARCH__>=5
+	ldmia	sp!,{r4-r12,pc}
+#else
 	ldmia	sp!,{r4-r12,lr}
 	tst	lr,#1
 	moveq	pc,lr			@ be binary compatible with V4, yet
 	bx	lr			@ interoperable with Thumb ISA:-)
+#endif
 .align	2
 .LK_00_19:	.word	0x5a827999
 .LK_20_39:	.word	0x6ed9eba1
diff --git a/lib/libcrypto/sha/asm/sha256-armv4.pl b/lib/libcrypto/sha/asm/sha256-armv4.pl
index 48d846deec3..292520731cd 100644
--- a/lib/libcrypto/sha/asm/sha256-armv4.pl
+++ b/lib/libcrypto/sha/asm/sha256-armv4.pl
@@ -11,13 +11,23 @@
 
 # Performance is ~2x better than gcc 3.4 generated code and in "abso-
 # lute" terms is ~2250 cycles per 64-byte block or ~35 cycles per
-# byte.
+# byte [on single-issue Xscale PXA250 core].
 
-$output=shift;
+# July 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 22% improvement on
+# Cortex A8 core and ~20 cycles per processed byte.
+
+# February 2011.
+#
+# Profiler-assisted and platform-specific optimization resulted in 16%
+# improvement on Cortex A8 core and ~17 cycles per processed byte.
+
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
 $ctx="r0";	$t0="r0";
-$inp="r1";
+$inp="r1";	$t3="r1";
 $len="r2";	$t1="r2";
 $T1="r3";
 $A="r4";
@@ -41,6 +51,9 @@ sub BODY_00_15 {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 
 $code.=<<___ if ($i<16);
+#if __ARM_ARCH__>=7 && !defined(__STRICT_ALIGNMENT)
+	ldr	$T1,[$inp],#4
+#else
 	ldrb	$T1,[$inp,#3]			@ $i
 	ldrb	$t2,[$inp,#2]
 	ldrb	$t1,[$inp,#1]
@@ -48,31 +61,42 @@ $code.=<<___ if ($i<16);
 	orr	$T1,$T1,$t2,lsl#8
 	orr	$T1,$T1,$t1,lsl#16
 	orr	$T1,$T1,$t0,lsl#24
-	`"str	$inp,[sp,#17*4]"	if ($i==15)`
+#endif
 ___
 $code.=<<___;
-	ldr	$t2,[$Ktbl],#4			@ *K256++
-	str	$T1,[sp,#`$i%16`*4]
 	mov	$t0,$e,ror#$Sigma1[0]
+	ldr	$t2,[$Ktbl],#4			@ *K256++
 	eor	$t0,$t0,$e,ror#$Sigma1[1]
-	eor	$t0,$t0,$e,ror#$Sigma1[2]	@ Sigma1(e)
-	add	$T1,$T1,$t0
 	eor	$t1,$f,$g
+#if $i>=16
+	add	$T1,$T1,$t3			@ from BODY_16_xx
+#elif __ARM_ARCH__>=7 && defined(__ARMEL__) && !defined(__STRICT_ALIGNMENT)
+	rev	$T1,$T1
+#endif
+#if $i==15
+	str	$inp,[sp,#17*4]			@ leave room for $t3
+#endif
+	eor	$t0,$t0,$e,ror#$Sigma1[2]	@ Sigma1(e)
 	and	$t1,$t1,$e
+	str	$T1,[sp,#`$i%16`*4]
+	add	$T1,$T1,$t0
 	eor	$t1,$t1,$g			@ Ch(e,f,g)
-	add	$T1,$T1,$t1
 	add	$T1,$T1,$h
-	add	$T1,$T1,$t2
 	mov	$h,$a,ror#$Sigma0[0]
+	add	$T1,$T1,$t1
 	eor	$h,$h,$a,ror#$Sigma0[1]
+	add	$T1,$T1,$t2
 	eor	$h,$h,$a,ror#$Sigma0[2]		@ Sigma0(a)
+#if $i>=15
+	ldr	$t3,[sp,#`($i+2)%16`*4]		@ from BODY_16_xx
+#endif
 	orr	$t0,$a,$b
-	and	$t0,$t0,$c
 	and	$t1,$a,$b
+	and	$t0,$t0,$c
+	add	$h,$h,$T1
 	orr	$t0,$t0,$t1			@ Maj(a,b,c)
-	add	$h,$h,$t0
 	add	$d,$d,$T1
-	add	$h,$h,$T1
+	add	$h,$h,$t0
 ___
 }
 
@@ -80,24 +104,26 @@ sub BODY_16_XX {
 my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
 
 $code.=<<___;
-	ldr	$t1,[sp,#`($i+1)%16`*4]	@ $i
+	@ ldr	$t3,[sp,#`($i+1)%16`*4]		@ $i
 	ldr	$t2,[sp,#`($i+14)%16`*4]
+	mov	$t0,$t3,ror#$sigma0[0]
 	ldr	$T1,[sp,#`($i+0)%16`*4]
-	ldr	$inp,[sp,#`($i+9)%16`*4]
-	mov	$t0,$t1,ror#$sigma0[0]
-	eor	$t0,$t0,$t1,ror#$sigma0[1]
-	eor	$t0,$t0,$t1,lsr#$sigma0[2]	@ sigma0(X[i+1])
-	mov	$t1,$t2,ror#$sigma1[0]
-	eor	$t1,$t1,$t2,ror#$sigma1[1]
-	eor	$t1,$t1,$t2,lsr#$sigma1[2]	@ sigma1(X[i+14])
+	eor	$t0,$t0,$t3,ror#$sigma0[1]
+	ldr	$t1,[sp,#`($i+9)%16`*4]
+	eor	$t0,$t0,$t3,lsr#$sigma0[2]	@ sigma0(X[i+1])
+	mov	$t3,$t2,ror#$sigma1[0]
 	add	$T1,$T1,$t0
+	eor	$t3,$t3,$t2,ror#$sigma1[1]
 	add	$T1,$T1,$t1
-	add	$T1,$T1,$inp
+	eor	$t3,$t3,$t2,lsr#$sigma1[2]	@ sigma1(X[i+14])
+	@ add	$T1,$T1,$t3
 ___
 	&BODY_00_15(@_);
 }
 
 $code=<<___;
+#include "arm_arch.h"
+
 .text
 .code	32
 
@@ -127,7 +153,7 @@ K256:
 sha256_block_data_order:
 	sub	r3,pc,#8		@ sha256_block_data_order
 	add	$len,$inp,$len,lsl#6	@ len to point at the end of inp
-	stmdb	sp!,{$ctx,$inp,$len,r4-r12,lr}
+	stmdb	sp!,{$ctx,$inp,$len,r4-r11,lr}
 	ldmia	$ctx,{$A,$B,$C,$D,$E,$F,$G,$H}
 	sub	$Ktbl,r3,#256		@ K256
 	sub	sp,sp,#16*4		@ alloca(X[16])
@@ -166,10 +192,14 @@ $code.=<<___;
 	bne	.Loop
 
 	add	sp,sp,#`16+3`*4	@ destroy frame
-	ldmia	sp!,{r4-r12,lr}
+#if __ARM_ARCH__>=5
+	ldmia	sp!,{r4-r11,pc}
+#else
+	ldmia	sp!,{r4-r11,lr}
 	tst	lr,#1
 	moveq	pc,lr			@ be binary compatible with V4, yet
 	bx	lr			@ interoperable with Thumb ISA:-)
+#endif
 .size   sha256_block_data_order,.-sha256_block_data_order
 .asciz  "SHA256 block transform for ARMv4, CRYPTOGAMS by <appro\@openssl.org>"
 .align	2
diff --git a/lib/libcrypto/sha/asm/sha512-armv4.pl b/lib/libcrypto/sha/asm/sha512-armv4.pl
index 4fbb94a914f..28ae155f4b4 100644
--- a/lib/libcrypto/sha/asm/sha512-armv4.pl
+++ b/lib/libcrypto/sha/asm/sha512-armv4.pl
@@ -10,24 +10,41 @@
 # SHA512 block procedure for ARMv4. September 2007.
 
 # This code is ~4.5 (four and a half) times faster than code generated
-# by gcc 3.4 and it spends ~72 clock cycles per byte. 
+# by gcc 3.4 and it spends ~72 clock cycles per byte [on single-issue
+# Xscale PXA250 core].
+#
+# July 2010.
+#
+# Rescheduling for dual-issue pipeline resulted in 6% improvement on
+# Cortex A8 core and ~40 cycles per processed byte.
+
+# February 2011.
+#
+# Profiler-assisted and platform-specific optimization resulted in 7%
+# improvement on Coxtex A8 core and ~38 cycles per byte.
+
+# March 2011.
+#
+# Add NEON implementation. On Cortex A8 it was measured to process
+# one byte in 25.5 cycles or 47% faster than integer-only code.
 
 # Byte order [in]dependence. =========================================
 #
-# Caller is expected to maintain specific *dword* order in h[0-7],
-# namely with most significant dword at *lower* address, which is
-# reflected in below two parameters. *Byte* order within these dwords
-# in turn is whatever *native* byte order on current platform.
-$hi=0;
-$lo=4;
+# Originally caller was expected to maintain specific *dword* order in
+# h[0-7], namely with most significant dword at *lower* address, which
+# was reflected in below two parameters as 0 and 4. Now caller is
+# expected to maintain native byte order for whole 64-bit values.
+$hi="HI";
+$lo="LO";
 # ====================================================================
 
-$output=shift;
+while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
 open STDOUT,">$output";
 
-$ctx="r0";
+$ctx="r0";	# parameter block
 $inp="r1";
 $len="r2";
+
 $Tlo="r3";
 $Thi="r4";
 $Alo="r5";
@@ -55,15 +72,17 @@ $Xoff=8*8;
 sub BODY_00_15() {
 my $magic = shift;
 $code.=<<___;
-	ldr	$t2,[sp,#$Hoff+0]	@ h.lo
-	ldr	$t3,[sp,#$Hoff+4]	@ h.hi
 	@ Sigma1(x)	(ROTR((x),14) ^ ROTR((x),18)  ^ ROTR((x),41))
 	@ LO		lo>>14^hi<<18 ^ lo>>18^hi<<14 ^ hi>>9^lo<<23
 	@ HI		hi>>14^lo<<18 ^ hi>>18^lo<<14 ^ lo>>9^hi<<23
 	mov	$t0,$Elo,lsr#14
+	str	$Tlo,[sp,#$Xoff+0]
 	mov	$t1,$Ehi,lsr#14
+	str	$Thi,[sp,#$Xoff+4]
 	eor	$t0,$t0,$Ehi,lsl#18
+	ldr	$t2,[sp,#$Hoff+0]	@ h.lo
 	eor	$t1,$t1,$Elo,lsl#18
+	ldr	$t3,[sp,#$Hoff+4]	@ h.hi
 	eor	$t0,$t0,$Elo,lsr#18
 	eor	$t1,$t1,$Ehi,lsr#18
 	eor	$t0,$t0,$Ehi,lsl#14
@@ -73,44 +92,41 @@ $code.=<<___;
 	eor	$t0,$t0,$Elo,lsl#23
 	eor	$t1,$t1,$Ehi,lsl#23	@ Sigma1(e)
 	adds	$Tlo,$Tlo,$t0
-	adc	$Thi,$Thi,$t1		@ T += Sigma1(e)
-	adds	$Tlo,$Tlo,$t2
-	adc	$Thi,$Thi,$t3		@ T += h
-
 	ldr	$t0,[sp,#$Foff+0]	@ f.lo
+	adc	$Thi,$Thi,$t1		@ T += Sigma1(e)
 	ldr	$t1,[sp,#$Foff+4]	@ f.hi
+	adds	$Tlo,$Tlo,$t2
 	ldr	$t2,[sp,#$Goff+0]	@ g.lo
+	adc	$Thi,$Thi,$t3		@ T += h
 	ldr	$t3,[sp,#$Goff+4]	@ g.hi
-	str	$Elo,[sp,#$Eoff+0]
-	str	$Ehi,[sp,#$Eoff+4]
-	str	$Alo,[sp,#$Aoff+0]
-	str	$Ahi,[sp,#$Aoff+4]
 
 	eor	$t0,$t0,$t2
+	str	$Elo,[sp,#$Eoff+0]
 	eor	$t1,$t1,$t3
+	str	$Ehi,[sp,#$Eoff+4]
 	and	$t0,$t0,$Elo
+	str	$Alo,[sp,#$Aoff+0]
 	and	$t1,$t1,$Ehi
+	str	$Ahi,[sp,#$Aoff+4]
 	eor	$t0,$t0,$t2
+	ldr	$t2,[$Ktbl,#$lo]	@ K[i].lo
 	eor	$t1,$t1,$t3		@ Ch(e,f,g)
-
-	ldr	$t2,[$Ktbl,#4]		@ K[i].lo
-	ldr	$t3,[$Ktbl,#0]		@ K[i].hi
-	ldr	$Elo,[sp,#$Doff+0]	@ d.lo
-	ldr	$Ehi,[sp,#$Doff+4]	@ d.hi
+	ldr	$t3,[$Ktbl,#$hi]	@ K[i].hi
 
 	adds	$Tlo,$Tlo,$t0
+	ldr	$Elo,[sp,#$Doff+0]	@ d.lo
 	adc	$Thi,$Thi,$t1		@ T += Ch(e,f,g)
+	ldr	$Ehi,[sp,#$Doff+4]	@ d.hi
 	adds	$Tlo,$Tlo,$t2
+	and	$t0,$t2,#0xff
 	adc	$Thi,$Thi,$t3		@ T += K[i]
 	adds	$Elo,$Elo,$Tlo
+	ldr	$t2,[sp,#$Boff+0]	@ b.lo
 	adc	$Ehi,$Ehi,$Thi		@ d += T
-
-	and	$t0,$t2,#0xff
 	teq	$t0,#$magic
-	orreq	$Ktbl,$Ktbl,#1
 
-	ldr	$t2,[sp,#$Boff+0]	@ b.lo
 	ldr	$t3,[sp,#$Coff+0]	@ c.lo
+	orreq	$Ktbl,$Ktbl,#1
 	@ Sigma0(x)	(ROTR((x),28) ^ ROTR((x),34) ^ ROTR((x),39))
 	@ LO		lo>>28^hi<<4  ^ hi>>2^lo<<30 ^ hi>>7^lo<<25
 	@ HI		hi>>28^lo<<4  ^ lo>>2^hi<<30 ^ lo>>7^hi<<25
@@ -127,80 +143,100 @@ $code.=<<___;
 	eor	$t0,$t0,$Alo,lsl#25
 	eor	$t1,$t1,$Ahi,lsl#25	@ Sigma0(a)
 	adds	$Tlo,$Tlo,$t0
+	and	$t0,$Alo,$t2
 	adc	$Thi,$Thi,$t1		@ T += Sigma0(a)
 
-	and	$t0,$Alo,$t2
-	orr	$Alo,$Alo,$t2
 	ldr	$t1,[sp,#$Boff+4]	@ b.hi
+	orr	$Alo,$Alo,$t2
 	ldr	$t2,[sp,#$Coff+4]	@ c.hi
 	and	$Alo,$Alo,$t3
-	orr	$Alo,$Alo,$t0		@ Maj(a,b,c).lo
 	and	$t3,$Ahi,$t1
 	orr	$Ahi,$Ahi,$t1
+	orr	$Alo,$Alo,$t0		@ Maj(a,b,c).lo
 	and	$Ahi,$Ahi,$t2
-	orr	$Ahi,$Ahi,$t3		@ Maj(a,b,c).hi
 	adds	$Alo,$Alo,$Tlo
-	adc	$Ahi,$Ahi,$Thi		@ h += T
-
+	orr	$Ahi,$Ahi,$t3		@ Maj(a,b,c).hi
 	sub	sp,sp,#8
+	adc	$Ahi,$Ahi,$Thi		@ h += T
+	tst	$Ktbl,#1
 	add	$Ktbl,$Ktbl,#8
 ___
 }
 $code=<<___;
+#include "arm_arch.h"
+#ifdef __ARMEL__
+# define LO 0
+# define HI 4
+# define WORD64(hi0,lo0,hi1,lo1)	.word	lo0,hi0, lo1,hi1
+#else
+# define HI 0
+# define LO 4
+# define WORD64(hi0,lo0,hi1,lo1)	.word	hi0,lo0, hi1,lo1
+#endif
+
 .text
 .code	32
 .type	K512,%object
 .align	5
 K512:
-.word	0x428a2f98,0xd728ae22, 0x71374491,0x23ef65cd
-.word	0xb5c0fbcf,0xec4d3b2f, 0xe9b5dba5,0x8189dbbc
-.word	0x3956c25b,0xf348b538, 0x59f111f1,0xb605d019
-.word	0x923f82a4,0xaf194f9b, 0xab1c5ed5,0xda6d8118
-.word	0xd807aa98,0xa3030242, 0x12835b01,0x45706fbe
-.word	0x243185be,0x4ee4b28c, 0x550c7dc3,0xd5ffb4e2
-.word	0x72be5d74,0xf27b896f, 0x80deb1fe,0x3b1696b1
-.word	0x9bdc06a7,0x25c71235, 0xc19bf174,0xcf692694
-.word	0xe49b69c1,0x9ef14ad2, 0xefbe4786,0x384f25e3
-.word	0x0fc19dc6,0x8b8cd5b5, 0x240ca1cc,0x77ac9c65
-.word	0x2de92c6f,0x592b0275, 0x4a7484aa,0x6ea6e483
-.word	0x5cb0a9dc,0xbd41fbd4, 0x76f988da,0x831153b5
-.word	0x983e5152,0xee66dfab, 0xa831c66d,0x2db43210
-.word	0xb00327c8,0x98fb213f, 0xbf597fc7,0xbeef0ee4
-.word	0xc6e00bf3,0x3da88fc2, 0xd5a79147,0x930aa725
-.word	0x06ca6351,0xe003826f, 0x14292967,0x0a0e6e70
-.word	0x27b70a85,0x46d22ffc, 0x2e1b2138,0x5c26c926
-.word	0x4d2c6dfc,0x5ac42aed, 0x53380d13,0x9d95b3df
-.word	0x650a7354,0x8baf63de, 0x766a0abb,0x3c77b2a8
-.word	0x81c2c92e,0x47edaee6, 0x92722c85,0x1482353b
-.word	0xa2bfe8a1,0x4cf10364, 0xa81a664b,0xbc423001
-.word	0xc24b8b70,0xd0f89791, 0xc76c51a3,0x0654be30
-.word	0xd192e819,0xd6ef5218, 0xd6990624,0x5565a910
-.word	0xf40e3585,0x5771202a, 0x106aa070,0x32bbd1b8
-.word	0x19a4c116,0xb8d2d0c8, 0x1e376c08,0x5141ab53
-.word	0x2748774c,0xdf8eeb99, 0x34b0bcb5,0xe19b48a8
-.word	0x391c0cb3,0xc5c95a63, 0x4ed8aa4a,0xe3418acb
-.word	0x5b9cca4f,0x7763e373, 0x682e6ff3,0xd6b2b8a3
-.word	0x748f82ee,0x5defb2fc, 0x78a5636f,0x43172f60
-.word	0x84c87814,0xa1f0ab72, 0x8cc70208,0x1a6439ec
-.word	0x90befffa,0x23631e28, 0xa4506ceb,0xde82bde9
-.word	0xbef9a3f7,0xb2c67915, 0xc67178f2,0xe372532b
-.word	0xca273ece,0xea26619c, 0xd186b8c7,0x21c0c207
-.word	0xeada7dd6,0xcde0eb1e, 0xf57d4f7f,0xee6ed178
-.word	0x06f067aa,0x72176fba, 0x0a637dc5,0xa2c898a6
-.word	0x113f9804,0xbef90dae, 0x1b710b35,0x131c471b
-.word	0x28db77f5,0x23047d84, 0x32caab7b,0x40c72493
-.word	0x3c9ebe0a,0x15c9bebc, 0x431d67c4,0x9c100d4c
-.word	0x4cc5d4be,0xcb3e42b6, 0x597f299c,0xfc657e2a
-.word	0x5fcb6fab,0x3ad6faec, 0x6c44198c,0x4a475817
+WORD64(0x428a2f98,0xd728ae22, 0x71374491,0x23ef65cd)
+WORD64(0xb5c0fbcf,0xec4d3b2f, 0xe9b5dba5,0x8189dbbc)
+WORD64(0x3956c25b,0xf348b538, 0x59f111f1,0xb605d019)
+WORD64(0x923f82a4,0xaf194f9b, 0xab1c5ed5,0xda6d8118)
+WORD64(0xd807aa98,0xa3030242, 0x12835b01,0x45706fbe)
+WORD64(0x243185be,0x4ee4b28c, 0x550c7dc3,0xd5ffb4e2)
+WORD64(0x72be5d74,0xf27b896f, 0x80deb1fe,0x3b1696b1)
+WORD64(0x9bdc06a7,0x25c71235, 0xc19bf174,0xcf692694)
+WORD64(0xe49b69c1,0x9ef14ad2, 0xefbe4786,0x384f25e3)
+WORD64(0x0fc19dc6,0x8b8cd5b5, 0x240ca1cc,0x77ac9c65)
+WORD64(0x2de92c6f,0x592b0275, 0x4a7484aa,0x6ea6e483)
+WORD64(0x5cb0a9dc,0xbd41fbd4, 0x76f988da,0x831153b5)
+WORD64(0x983e5152,0xee66dfab, 0xa831c66d,0x2db43210)
+WORD64(0xb00327c8,0x98fb213f, 0xbf597fc7,0xbeef0ee4)
+WORD64(0xc6e00bf3,0x3da88fc2, 0xd5a79147,0x930aa725)
+WORD64(0x06ca6351,0xe003826f, 0x14292967,0x0a0e6e70)
+WORD64(0x27b70a85,0x46d22ffc, 0x2e1b2138,0x5c26c926)
+WORD64(0x4d2c6dfc,0x5ac42aed, 0x53380d13,0x9d95b3df)
+WORD64(0x650a7354,0x8baf63de, 0x766a0abb,0x3c77b2a8)
+WORD64(0x81c2c92e,0x47edaee6, 0x92722c85,0x1482353b)
+WORD64(0xa2bfe8a1,0x4cf10364, 0xa81a664b,0xbc423001)
+WORD64(0xc24b8b70,0xd0f89791, 0xc76c51a3,0x0654be30)
+WORD64(0xd192e819,0xd6ef5218, 0xd6990624,0x5565a910)
+WORD64(0xf40e3585,0x5771202a, 0x106aa070,0x32bbd1b8)
+WORD64(0x19a4c116,0xb8d2d0c8, 0x1e376c08,0x5141ab53)
+WORD64(0x2748774c,0xdf8eeb99, 0x34b0bcb5,0xe19b48a8)
+WORD64(0x391c0cb3,0xc5c95a63, 0x4ed8aa4a,0xe3418acb)
+WORD64(0x5b9cca4f,0x7763e373, 0x682e6ff3,0xd6b2b8a3)
+WORD64(0x748f82ee,0x5defb2fc, 0x78a5636f,0x43172f60)
+WORD64(0x84c87814,0xa1f0ab72, 0x8cc70208,0x1a6439ec)
+WORD64(0x90befffa,0x23631e28, 0xa4506ceb,0xde82bde9)
+WORD64(0xbef9a3f7,0xb2c67915, 0xc67178f2,0xe372532b)
+WORD64(0xca273ece,0xea26619c, 0xd186b8c7,0x21c0c207)
+WORD64(0xeada7dd6,0xcde0eb1e, 0xf57d4f7f,0xee6ed178)
+WORD64(0x06f067aa,0x72176fba, 0x0a637dc5,0xa2c898a6)
+WORD64(0x113f9804,0xbef90dae, 0x1b710b35,0x131c471b)
+WORD64(0x28db77f5,0x23047d84, 0x32caab7b,0x40c72493)
+WORD64(0x3c9ebe0a,0x15c9bebc, 0x431d67c4,0x9c100d4c)
+WORD64(0x4cc5d4be,0xcb3e42b6, 0x597f299c,0xfc657e2a)
+WORD64(0x5fcb6fab,0x3ad6faec, 0x6c44198c,0x4a475817)
 .size	K512,.-K512
+.LOPENSSL_armcap:
+.word	OPENSSL_armcap_P-sha512_block_data_order
+.skip	32-4
 
 .global	sha512_block_data_order
 .type	sha512_block_data_order,%function
 sha512_block_data_order:
 	sub	r3,pc,#8		@ sha512_block_data_order
 	add	$len,$inp,$len,lsl#7	@ len to point at the end of inp
+#if __ARM_ARCH__>=7
+	ldr	r12,.LOPENSSL_armcap
+	ldr	r12,[r3,r12]		@ OPENSSL_armcap_P
+	tst	r12,#1
+	bne	.LNEON
+#endif
 	stmdb	sp!,{r4-r12,lr}
-	sub	$Ktbl,r3,#640		@ K512
+	sub	$Ktbl,r3,#672		@ K512
 	sub	sp,sp,#9*8
 
 	ldr	$Elo,[$ctx,#$Eoff+$lo]
@@ -234,6 +270,7 @@ sha512_block_data_order:
 	str	$Thi,[sp,#$Foff+4]
 
 .L00_15:
+#if __ARM_ARCH__<7 || defined(__STRICT_ALIGNMENT)
 	ldrb	$Tlo,[$inp,#7]
 	ldrb	$t0, [$inp,#6]
 	ldrb	$t1, [$inp,#5]
@@ -248,26 +285,30 @@ sha512_block_data_order:
 	orr	$Thi,$Thi,$t3,lsl#8
 	orr	$Thi,$Thi,$t0,lsl#16
 	orr	$Thi,$Thi,$t1,lsl#24
-	str	$Tlo,[sp,#$Xoff+0]
-	str	$Thi,[sp,#$Xoff+4]
+#else
+	ldr	$Tlo,[$inp,#4]
+	ldr	$Thi,[$inp],#8
+#ifdef __ARMEL__
+	rev	$Tlo,$Tlo
+	rev	$Thi,$Thi
+#endif
+#endif
 ___
 	&BODY_00_15(0x94);
 $code.=<<___;
 	tst	$Ktbl,#1
 	beq	.L00_15
-	bic	$Ktbl,$Ktbl,#1
-
-.L16_79:
 	ldr	$t0,[sp,#`$Xoff+8*(16-1)`+0]
 	ldr	$t1,[sp,#`$Xoff+8*(16-1)`+4]
-	ldr	$t2,[sp,#`$Xoff+8*(16-14)`+0]
-	ldr	$t3,[sp,#`$Xoff+8*(16-14)`+4]
-
+	bic	$Ktbl,$Ktbl,#1
+.L16_79:
 	@ sigma0(x)	(ROTR((x),1)  ^ ROTR((x),8)  ^ ((x)>>7))
 	@ LO		lo>>1^hi<<31  ^ lo>>8^hi<<24 ^ lo>>7^hi<<25
 	@ HI		hi>>1^lo<<31  ^ hi>>8^lo<<24 ^ hi>>7
 	mov	$Tlo,$t0,lsr#1
+	ldr	$t2,[sp,#`$Xoff+8*(16-14)`+0]
 	mov	$Thi,$t1,lsr#1
+	ldr	$t3,[sp,#`$Xoff+8*(16-14)`+4]
 	eor	$Tlo,$Tlo,$t1,lsl#31
 	eor	$Thi,$Thi,$t0,lsl#31
 	eor	$Tlo,$Tlo,$t0,lsr#8
@@ -291,25 +332,24 @@ $code.=<<___;
 	eor	$t1,$t1,$t3,lsl#3
 	eor	$t0,$t0,$t2,lsr#6
 	eor	$t1,$t1,$t3,lsr#6
+	ldr	$t2,[sp,#`$Xoff+8*(16-9)`+0]
 	eor	$t0,$t0,$t3,lsl#26
 
-	ldr	$t2,[sp,#`$Xoff+8*(16-9)`+0]
 	ldr	$t3,[sp,#`$Xoff+8*(16-9)`+4]
 	adds	$Tlo,$Tlo,$t0
+	ldr	$t0,[sp,#`$Xoff+8*16`+0]
 	adc	$Thi,$Thi,$t1
 
-	ldr	$t0,[sp,#`$Xoff+8*16`+0]
 	ldr	$t1,[sp,#`$Xoff+8*16`+4]
 	adds	$Tlo,$Tlo,$t2
 	adc	$Thi,$Thi,$t3
 	adds	$Tlo,$Tlo,$t0
 	adc	$Thi,$Thi,$t1
-	str	$Tlo,[sp,#$Xoff+0]
-	str	$Thi,[sp,#$Xoff+4]
 ___
 	&BODY_00_15(0x17);
 $code.=<<___;
-	tst	$Ktbl,#1
+	ldreq	$t0,[sp,#`$Xoff+8*(16-1)`+0]
+	ldreq	$t1,[sp,#`$Xoff+8*(16-1)`+4]
 	beq	.L16_79
 	bic	$Ktbl,$Ktbl,#1
 
@@ -320,12 +360,12 @@ $code.=<<___;
 	ldr	$t2, [$ctx,#$Boff+$lo]
 	ldr	$t3, [$ctx,#$Boff+$hi]
 	adds	$t0,$Alo,$t0
-	adc	$t1,$Ahi,$t1
-	adds	$t2,$Tlo,$t2
-	adc	$t3,$Thi,$t3
 	str	$t0, [$ctx,#$Aoff+$lo]
+	adc	$t1,$Ahi,$t1
 	str	$t1, [$ctx,#$Aoff+$hi]
+	adds	$t2,$Tlo,$t2
 	str	$t2, [$ctx,#$Boff+$lo]
+	adc	$t3,$Thi,$t3
 	str	$t3, [$ctx,#$Boff+$hi]
 
 	ldr	$Alo,[sp,#$Coff+0]
@@ -337,12 +377,12 @@ $code.=<<___;
 	ldr	$t2, [$ctx,#$Doff+$lo]
 	ldr	$t3, [$ctx,#$Doff+$hi]
 	adds	$t0,$Alo,$t0
-	adc	$t1,$Ahi,$t1
-	adds	$t2,$Tlo,$t2
-	adc	$t3,$Thi,$t3
 	str	$t0, [$ctx,#$Coff+$lo]
+	adc	$t1,$Ahi,$t1
 	str	$t1, [$ctx,#$Coff+$hi]
+	adds	$t2,$Tlo,$t2
 	str	$t2, [$ctx,#$Doff+$lo]
+	adc	$t3,$Thi,$t3
 	str	$t3, [$ctx,#$Doff+$hi]
 
 	ldr	$Tlo,[sp,#$Foff+0]
@@ -352,12 +392,12 @@ $code.=<<___;
 	ldr	$t2, [$ctx,#$Foff+$lo]
 	ldr	$t3, [$ctx,#$Foff+$hi]
 	adds	$Elo,$Elo,$t0
-	adc	$Ehi,$Ehi,$t1
-	adds	$t2,$Tlo,$t2
-	adc	$t3,$Thi,$t3
 	str	$Elo,[$ctx,#$Eoff+$lo]
+	adc	$Ehi,$Ehi,$t1
 	str	$Ehi,[$ctx,#$Eoff+$hi]
+	adds	$t2,$Tlo,$t2
 	str	$t2, [$ctx,#$Foff+$lo]
+	adc	$t3,$Thi,$t3
 	str	$t3, [$ctx,#$Foff+$hi]
 
 	ldr	$Alo,[sp,#$Goff+0]
@@ -369,12 +409,12 @@ $code.=<<___;
 	ldr	$t2, [$ctx,#$Hoff+$lo]
 	ldr	$t3, [$ctx,#$Hoff+$hi]
 	adds	$t0,$Alo,$t0
-	adc	$t1,$Ahi,$t1
-	adds	$t2,$Tlo,$t2
-	adc	$t3,$Thi,$t3
 	str	$t0, [$ctx,#$Goff+$lo]
+	adc	$t1,$Ahi,$t1
 	str	$t1, [$ctx,#$Goff+$hi]
+	adds	$t2,$Tlo,$t2
 	str	$t2, [$ctx,#$Hoff+$lo]
+	adc	$t3,$Thi,$t3
 	str	$t3, [$ctx,#$Hoff+$hi]
 
 	add	sp,sp,#640
@@ -384,13 +424,156 @@ $code.=<<___;
 	bne	.Loop
 
 	add	sp,sp,#8*9		@ destroy frame
+#if __ARM_ARCH__>=5
+	ldmia	sp!,{r4-r12,pc}
+#else
 	ldmia	sp!,{r4-r12,lr}
 	tst	lr,#1
 	moveq	pc,lr			@ be binary compatible with V4, yet
 	bx	lr			@ interoperable with Thumb ISA:-)
-.size   sha512_block_data_order,.-sha512_block_data_order
-.asciz  "SHA512 block transform for ARMv4, CRYPTOGAMS by <appro\@openssl.org>"
+#endif
+___
+
+{
+my @Sigma0=(28,34,39);
+my @Sigma1=(14,18,41);
+my @sigma0=(1, 8, 7);
+my @sigma1=(19,61,6);
+
+my $Ktbl="r3";
+my $cnt="r12";	# volatile register known as ip, intra-procedure-call scratch
+
+my @X=map("d$_",(0..15));
+my @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("d$_",(16..23));
+
+sub NEON_00_15() {
+my $i=shift;
+my ($a,$b,$c,$d,$e,$f,$g,$h)=@_;
+my ($t0,$t1,$t2,$T1,$K,$Ch,$Maj)=map("d$_",(24..31));	# temps
+
+$code.=<<___ if ($i<16 || $i&1);
+	vshr.u64	$t0,$e,#@Sigma1[0]	@ $i
+#if $i<16
+	vld1.64		{@X[$i%16]},[$inp]!	@ handles unaligned
+#endif
+	vshr.u64	$t1,$e,#@Sigma1[1]
+	vshr.u64	$t2,$e,#@Sigma1[2]
+___
+$code.=<<___;
+	vld1.64		{$K},[$Ktbl,:64]!	@ K[i++]
+	vsli.64		$t0,$e,#`64-@Sigma1[0]`
+	vsli.64		$t1,$e,#`64-@Sigma1[1]`
+	vsli.64		$t2,$e,#`64-@Sigma1[2]`
+#if $i<16 && defined(__ARMEL__)
+	vrev64.8	@X[$i],@X[$i]
+#endif
+	vadd.i64	$T1,$K,$h
+	veor		$Ch,$f,$g
+	veor		$t0,$t1
+	vand		$Ch,$e
+	veor		$t0,$t2			@ Sigma1(e)
+	veor		$Ch,$g			@ Ch(e,f,g)
+	vadd.i64	$T1,$t0
+	vshr.u64	$t0,$a,#@Sigma0[0]
+	vadd.i64	$T1,$Ch
+	vshr.u64	$t1,$a,#@Sigma0[1]
+	vshr.u64	$t2,$a,#@Sigma0[2]
+	vsli.64		$t0,$a,#`64-@Sigma0[0]`
+	vsli.64		$t1,$a,#`64-@Sigma0[1]`
+	vsli.64		$t2,$a,#`64-@Sigma0[2]`
+	vadd.i64	$T1,@X[$i%16]
+	vorr		$Maj,$a,$c
+	vand		$Ch,$a,$c
+	veor		$h,$t0,$t1
+	vand		$Maj,$b
+	veor		$h,$t2			@ Sigma0(a)
+	vorr		$Maj,$Ch		@ Maj(a,b,c)
+	vadd.i64	$h,$T1
+	vadd.i64	$d,$T1
+	vadd.i64	$h,$Maj
+___
+}
+
+sub NEON_16_79() {
+my $i=shift;
+
+if ($i&1)	{ &NEON_00_15($i,@_); return; }
+
+# 2x-vectorized, therefore runs every 2nd round
+my @X=map("q$_",(0..7));			# view @X as 128-bit vector
+my ($t0,$t1,$s0,$s1) = map("q$_",(12..15));	# temps
+my ($d0,$d1,$d2) = map("d$_",(24..26));		# temps from NEON_00_15
+my $e=@_[4];					# $e from NEON_00_15
+$i /= 2;
+$code.=<<___;
+	vshr.u64	$t0,@X[($i+7)%8],#@sigma1[0]
+	vshr.u64	$t1,@X[($i+7)%8],#@sigma1[1]
+	vshr.u64	$s1,@X[($i+7)%8],#@sigma1[2]
+	vsli.64		$t0,@X[($i+7)%8],#`64-@sigma1[0]`
+	vext.8		$s0,@X[$i%8],@X[($i+1)%8],#8	@ X[i+1]
+	vsli.64		$t1,@X[($i+7)%8],#`64-@sigma1[1]`
+	veor		$s1,$t0
+	vshr.u64	$t0,$s0,#@sigma0[0]
+	veor		$s1,$t1				@ sigma1(X[i+14])
+	vshr.u64	$t1,$s0,#@sigma0[1]
+	vadd.i64	@X[$i%8],$s1
+	vshr.u64	$s1,$s0,#@sigma0[2]
+	vsli.64		$t0,$s0,#`64-@sigma0[0]`
+	vsli.64		$t1,$s0,#`64-@sigma0[1]`
+	vext.8		$s0,@X[($i+4)%8],@X[($i+5)%8],#8	@ X[i+9]
+	veor		$s1,$t0
+	vshr.u64	$d0,$e,#@Sigma1[0]		@ from NEON_00_15
+	vadd.i64	@X[$i%8],$s0
+	vshr.u64	$d1,$e,#@Sigma1[1]		@ from NEON_00_15
+	veor		$s1,$t1				@ sigma0(X[i+1])
+	vshr.u64	$d2,$e,#@Sigma1[2]		@ from NEON_00_15
+	vadd.i64	@X[$i%8],$s1
+___
+	&NEON_00_15(2*$i,@_);
+}
+
+$code.=<<___;
+#if __ARM_ARCH__>=7
+.fpu	neon
+
+.align	4
+.LNEON:
+	dmb				@ errata #451034 on early Cortex A8
+	vstmdb	sp!,{d8-d15}		@ ABI specification says so
+	sub	$Ktbl,r3,#672		@ K512
+	vldmia	$ctx,{$A-$H}		@ load context
+.Loop_neon:
+___
+for($i=0;$i<16;$i++)	{ &NEON_00_15($i,@V); unshift(@V,pop(@V)); }
+$code.=<<___;
+	mov		$cnt,#4
+.L16_79_neon:
+	subs		$cnt,#1
+___
+for(;$i<32;$i++)	{ &NEON_16_79($i,@V); unshift(@V,pop(@V)); }
+$code.=<<___;
+	bne		.L16_79_neon
+
+	vldmia		$ctx,{d24-d31}	@ load context to temp
+	vadd.i64	q8,q12		@ vectorized accumulate
+	vadd.i64	q9,q13
+	vadd.i64	q10,q14
+	vadd.i64	q11,q15
+	vstmia		$ctx,{$A-$H}	@ save context
+	teq		$inp,$len
+	sub		$Ktbl,#640	@ rewind K512
+	bne		.Loop_neon
+
+	vldmia	sp!,{d8-d15}		@ epilogue
+	bx	lr
+#endif
+___
+}
+$code.=<<___;
+.size	sha512_block_data_order,.-sha512_block_data_order
+.asciz	"SHA512 block transform for ARMv4/NEON, CRYPTOGAMS by <appro\@openssl.org>"
 .align	2
+.comm	OPENSSL_armcap_P,4,4
 ___
 
 $code =~ s/\`([^\`]*)\`/eval $1/gem;