deps: update archs files for OpenSSL-1.1.1a

`cd deps/openssl/config; make` updates all archs dependant files.

PR-URL: https://github.com/nodejs/node/pull/25381
Reviewed-By: Daniel Bevenius <daniel.bevenius@gmail.com>
Reviewed-By: Shigeki Ohtsu <ohtsu@ohtsu.org>

15 files changed, 8170 insertions, 0 deletions

diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/aes/aes-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/aes/aes-s390x.S
new file mode 100644
index 0000000000..e00f81ecd4
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/aes/aes-s390x.S
@@ -0,0 +1,2365 @@
+#include "s390x_arch.h"
+
+.text
+
+.type	AES_Te,@object
+.align	256
+AES_Te:
+.long	0xc66363a5,0xc66363a5
+.long	0xf87c7c84,0xf87c7c84
+.long	0xee777799,0xee777799
+.long	0xf67b7b8d,0xf67b7b8d
+.long	0xfff2f20d,0xfff2f20d
+.long	0xd66b6bbd,0xd66b6bbd
+.long	0xde6f6fb1,0xde6f6fb1
+.long	0x91c5c554,0x91c5c554
+.long	0x60303050,0x60303050
+.long	0x02010103,0x02010103
+.long	0xce6767a9,0xce6767a9
+.long	0x562b2b7d,0x562b2b7d
+.long	0xe7fefe19,0xe7fefe19
+.long	0xb5d7d762,0xb5d7d762
+.long	0x4dababe6,0x4dababe6
+.long	0xec76769a,0xec76769a
+.long	0x8fcaca45,0x8fcaca45
+.long	0x1f82829d,0x1f82829d
+.long	0x89c9c940,0x89c9c940
+.long	0xfa7d7d87,0xfa7d7d87
+.long	0xeffafa15,0xeffafa15
+.long	0xb25959eb,0xb25959eb
+.long	0x8e4747c9,0x8e4747c9
+.long	0xfbf0f00b,0xfbf0f00b
+.long	0x41adadec,0x41adadec
+.long	0xb3d4d467,0xb3d4d467
+.long	0x5fa2a2fd,0x5fa2a2fd
+.long	0x45afafea,0x45afafea
+.long	0x239c9cbf,0x239c9cbf
+.long	0x53a4a4f7,0x53a4a4f7
+.long	0xe4727296,0xe4727296
+.long	0x9bc0c05b,0x9bc0c05b
+.long	0x75b7b7c2,0x75b7b7c2
+.long	0xe1fdfd1c,0xe1fdfd1c
+.long	0x3d9393ae,0x3d9393ae
+.long	0x4c26266a,0x4c26266a
+.long	0x6c36365a,0x6c36365a
+.long	0x7e3f3f41,0x7e3f3f41
+.long	0xf5f7f702,0xf5f7f702
+.long	0x83cccc4f,0x83cccc4f
+.long	0x6834345c,0x6834345c
+.long	0x51a5a5f4,0x51a5a5f4
+.long	0xd1e5e534,0xd1e5e534
+.long	0xf9f1f108,0xf9f1f108
+.long	0xe2717193,0xe2717193
+.long	0xabd8d873,0xabd8d873
+.long	0x62313153,0x62313153
+.long	0x2a15153f,0x2a15153f
+.long	0x0804040c,0x0804040c
+.long	0x95c7c752,0x95c7c752
+.long	0x46232365,0x46232365
+.long	0x9dc3c35e,0x9dc3c35e
+.long	0x30181828,0x30181828
+.long	0x379696a1,0x379696a1
+.long	0x0a05050f,0x0a05050f
+.long	0x2f9a9ab5,0x2f9a9ab5
+.long	0x0e070709,0x0e070709
+.long	0x24121236,0x24121236
+.long	0x1b80809b,0x1b80809b
+.long	0xdfe2e23d,0xdfe2e23d
+.long	0xcdebeb26,0xcdebeb26
+.long	0x4e272769,0x4e272769
+.long	0x7fb2b2cd,0x7fb2b2cd
+.long	0xea75759f,0xea75759f
+.long	0x1209091b,0x1209091b
+.long	0x1d83839e,0x1d83839e
+.long	0x582c2c74,0x582c2c74
+.long	0x341a1a2e,0x341a1a2e
+.long	0x361b1b2d,0x361b1b2d
+.long	0xdc6e6eb2,0xdc6e6eb2
+.long	0xb45a5aee,0xb45a5aee
+.long	0x5ba0a0fb,0x5ba0a0fb
+.long	0xa45252f6,0xa45252f6
+.long	0x763b3b4d,0x763b3b4d
+.long	0xb7d6d661,0xb7d6d661
+.long	0x7db3b3ce,0x7db3b3ce
+.long	0x5229297b,0x5229297b
+.long	0xdde3e33e,0xdde3e33e
+.long	0x5e2f2f71,0x5e2f2f71
+.long	0x13848497,0x13848497
+.long	0xa65353f5,0xa65353f5
+.long	0xb9d1d168,0xb9d1d168
+.long	0x00000000,0x00000000
+.long	0xc1eded2c,0xc1eded2c
+.long	0x40202060,0x40202060
+.long	0xe3fcfc1f,0xe3fcfc1f
+.long	0x79b1b1c8,0x79b1b1c8
+.long	0xb65b5bed,0xb65b5bed
+.long	0xd46a6abe,0xd46a6abe
+.long	0x8dcbcb46,0x8dcbcb46
+.long	0x67bebed9,0x67bebed9
+.long	0x7239394b,0x7239394b
+.long	0x944a4ade,0x944a4ade
+.long	0x984c4cd4,0x984c4cd4
+.long	0xb05858e8,0xb05858e8
+.long	0x85cfcf4a,0x85cfcf4a
+.long	0xbbd0d06b,0xbbd0d06b
+.long	0xc5efef2a,0xc5efef2a
+.long	0x4faaaae5,0x4faaaae5
+.long	0xedfbfb16,0xedfbfb16
+.long	0x864343c5,0x864343c5
+.long	0x9a4d4dd7,0x9a4d4dd7
+.long	0x66333355,0x66333355
+.long	0x11858594,0x11858594
+.long	0x8a4545cf,0x8a4545cf
+.long	0xe9f9f910,0xe9f9f910
+.long	0x04020206,0x04020206
+.long	0xfe7f7f81,0xfe7f7f81
+.long	0xa05050f0,0xa05050f0
+.long	0x783c3c44,0x783c3c44
+.long	0x259f9fba,0x259f9fba
+.long	0x4ba8a8e3,0x4ba8a8e3
+.long	0xa25151f3,0xa25151f3
+.long	0x5da3a3fe,0x5da3a3fe
+.long	0x804040c0,0x804040c0
+.long	0x058f8f8a,0x058f8f8a
+.long	0x3f9292ad,0x3f9292ad
+.long	0x219d9dbc,0x219d9dbc
+.long	0x70383848,0x70383848
+.long	0xf1f5f504,0xf1f5f504
+.long	0x63bcbcdf,0x63bcbcdf
+.long	0x77b6b6c1,0x77b6b6c1
+.long	0xafdada75,0xafdada75
+.long	0x42212163,0x42212163
+.long	0x20101030,0x20101030
+.long	0xe5ffff1a,0xe5ffff1a
+.long	0xfdf3f30e,0xfdf3f30e
+.long	0xbfd2d26d,0xbfd2d26d
+.long	0x81cdcd4c,0x81cdcd4c
+.long	0x180c0c14,0x180c0c14
+.long	0x26131335,0x26131335
+.long	0xc3ecec2f,0xc3ecec2f
+.long	0xbe5f5fe1,0xbe5f5fe1
+.long	0x359797a2,0x359797a2
+.long	0x884444cc,0x884444cc
+.long	0x2e171739,0x2e171739
+.long	0x93c4c457,0x93c4c457
+.long	0x55a7a7f2,0x55a7a7f2
+.long	0xfc7e7e82,0xfc7e7e82
+.long	0x7a3d3d47,0x7a3d3d47
+.long	0xc86464ac,0xc86464ac
+.long	0xba5d5de7,0xba5d5de7
+.long	0x3219192b,0x3219192b
+.long	0xe6737395,0xe6737395
+.long	0xc06060a0,0xc06060a0
+.long	0x19818198,0x19818198
+.long	0x9e4f4fd1,0x9e4f4fd1
+.long	0xa3dcdc7f,0xa3dcdc7f
+.long	0x44222266,0x44222266
+.long	0x542a2a7e,0x542a2a7e
+.long	0x3b9090ab,0x3b9090ab
+.long	0x0b888883,0x0b888883
+.long	0x8c4646ca,0x8c4646ca
+.long	0xc7eeee29,0xc7eeee29
+.long	0x6bb8b8d3,0x6bb8b8d3
+.long	0x2814143c,0x2814143c
+.long	0xa7dede79,0xa7dede79
+.long	0xbc5e5ee2,0xbc5e5ee2
+.long	0x160b0b1d,0x160b0b1d
+.long	0xaddbdb76,0xaddbdb76
+.long	0xdbe0e03b,0xdbe0e03b
+.long	0x64323256,0x64323256
+.long	0x743a3a4e,0x743a3a4e
+.long	0x140a0a1e,0x140a0a1e
+.long	0x924949db,0x924949db
+.long	0x0c06060a,0x0c06060a
+.long	0x4824246c,0x4824246c
+.long	0xb85c5ce4,0xb85c5ce4
+.long	0x9fc2c25d,0x9fc2c25d
+.long	0xbdd3d36e,0xbdd3d36e
+.long	0x43acacef,0x43acacef
+.long	0xc46262a6,0xc46262a6
+.long	0x399191a8,0x399191a8
+.long	0x319595a4,0x319595a4
+.long	0xd3e4e437,0xd3e4e437
+.long	0xf279798b,0xf279798b
+.long	0xd5e7e732,0xd5e7e732
+.long	0x8bc8c843,0x8bc8c843
+.long	0x6e373759,0x6e373759
+.long	0xda6d6db7,0xda6d6db7
+.long	0x018d8d8c,0x018d8d8c
+.long	0xb1d5d564,0xb1d5d564
+.long	0x9c4e4ed2,0x9c4e4ed2
+.long	0x49a9a9e0,0x49a9a9e0
+.long	0xd86c6cb4,0xd86c6cb4
+.long	0xac5656fa,0xac5656fa
+.long	0xf3f4f407,0xf3f4f407
+.long	0xcfeaea25,0xcfeaea25
+.long	0xca6565af,0xca6565af
+.long	0xf47a7a8e,0xf47a7a8e
+.long	0x47aeaee9,0x47aeaee9
+.long	0x10080818,0x10080818
+.long	0x6fbabad5,0x6fbabad5
+.long	0xf0787888,0xf0787888
+.long	0x4a25256f,0x4a25256f
+.long	0x5c2e2e72,0x5c2e2e72
+.long	0x381c1c24,0x381c1c24
+.long	0x57a6a6f1,0x57a6a6f1
+.long	0x73b4b4c7,0x73b4b4c7
+.long	0x97c6c651,0x97c6c651
+.long	0xcbe8e823,0xcbe8e823
+.long	0xa1dddd7c,0xa1dddd7c
+.long	0xe874749c,0xe874749c
+.long	0x3e1f1f21,0x3e1f1f21
+.long	0x964b4bdd,0x964b4bdd
+.long	0x61bdbddc,0x61bdbddc
+.long	0x0d8b8b86,0x0d8b8b86
+.long	0x0f8a8a85,0x0f8a8a85
+.long	0xe0707090,0xe0707090
+.long	0x7c3e3e42,0x7c3e3e42
+.long	0x71b5b5c4,0x71b5b5c4
+.long	0xcc6666aa,0xcc6666aa
+.long	0x904848d8,0x904848d8
+.long	0x06030305,0x06030305
+.long	0xf7f6f601,0xf7f6f601
+.long	0x1c0e0e12,0x1c0e0e12
+.long	0xc26161a3,0xc26161a3
+.long	0x6a35355f,0x6a35355f
+.long	0xae5757f9,0xae5757f9
+.long	0x69b9b9d0,0x69b9b9d0
+.long	0x17868691,0x17868691
+.long	0x99c1c158,0x99c1c158
+.long	0x3a1d1d27,0x3a1d1d27
+.long	0x279e9eb9,0x279e9eb9
+.long	0xd9e1e138,0xd9e1e138
+.long	0xebf8f813,0xebf8f813
+.long	0x2b9898b3,0x2b9898b3
+.long	0x22111133,0x22111133
+.long	0xd26969bb,0xd26969bb
+.long	0xa9d9d970,0xa9d9d970
+.long	0x078e8e89,0x078e8e89
+.long	0x339494a7,0x339494a7
+.long	0x2d9b9bb6,0x2d9b9bb6
+.long	0x3c1e1e22,0x3c1e1e22
+.long	0x15878792,0x15878792
+.long	0xc9e9e920,0xc9e9e920
+.long	0x87cece49,0x87cece49
+.long	0xaa5555ff,0xaa5555ff
+.long	0x50282878,0x50282878
+.long	0xa5dfdf7a,0xa5dfdf7a
+.long	0x038c8c8f,0x038c8c8f
+.long	0x59a1a1f8,0x59a1a1f8
+.long	0x09898980,0x09898980
+.long	0x1a0d0d17,0x1a0d0d17
+.long	0x65bfbfda,0x65bfbfda
+.long	0xd7e6e631,0xd7e6e631
+.long	0x844242c6,0x844242c6
+.long	0xd06868b8,0xd06868b8
+.long	0x824141c3,0x824141c3
+.long	0x299999b0,0x299999b0
+.long	0x5a2d2d77,0x5a2d2d77
+.long	0x1e0f0f11,0x1e0f0f11
+.long	0x7bb0b0cb,0x7bb0b0cb
+.long	0xa85454fc,0xa85454fc
+.long	0x6dbbbbd6,0x6dbbbbd6
+.long	0x2c16163a,0x2c16163a
+# Te4[256]
+.byte	0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
+.byte	0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
+.byte	0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
+.byte	0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
+.byte	0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
+.byte	0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
+.byte	0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
+.byte	0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
+.byte	0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
+.byte	0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
+.byte	0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
+.byte	0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
+.byte	0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
+.byte	0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
+.byte	0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
+.byte	0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
+.byte	0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
+.byte	0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
+.byte	0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
+.byte	0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
+.byte	0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
+.byte	0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
+.byte	0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
+.byte	0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
+.byte	0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
+.byte	0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
+.byte	0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
+.byte	0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
+.byte	0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
+.byte	0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
+.byte	0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
+.byte	0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
+# rcon[]
+.long	0x01000000, 0x02000000, 0x04000000, 0x08000000
+.long	0x10000000, 0x20000000, 0x40000000, 0x80000000
+.long	0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
+.align	256
+.size	AES_Te,.-AES_Te
+
+# void AES_encrypt(const unsigned char *inp, unsigned char *out,
+# 		 const AES_KEY *key) {
+.globl	AES_encrypt
+.type	AES_encrypt,@function
+AES_encrypt:
+	l	%r0,240(%r4)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lesoft
+
+	la	%r1,0(%r4)
+	#la	%r2,0(%r2)
+	la	%r4,0(%r3)
+	lghi	%r3,16		# single block length
+	.long	0xb92e0042	# km %r4,%r2
+	brc	1,.-4		# can this happen?
+	br	%r14
+.align	64
+.Lesoft:
+	stmg	%r3,%r14,3*8(%r15)
+
+	llgf	%r8,0(%r2)
+	llgf	%r9,4(%r2)
+	llgf	%r10,8(%r2)
+	llgf	%r11,12(%r2)
+
+	larl	%r12,AES_Te
+	bras	%r14,_s390x_AES_encrypt
+
+	lg	%r3,3*8(%r15)
+	st	%r8,0(%r3)
+	st	%r9,4(%r3)
+	st	%r10,8(%r3)
+	st	%r11,12(%r3)
+
+	lmg	%r6,%r14,6*8(%r15)
+	br	%r14
+.size	AES_encrypt,.-AES_encrypt
+
+.type   _s390x_AES_encrypt,@function
+.align	16
+_s390x_AES_encrypt:
+	stg	%r14,15*8(%r15)
+	x	%r8,0(%r4)
+	x	%r9,4(%r4)
+	x	%r10,8(%r4)
+	x	%r11,12(%r4)
+	l	%r13,240(%r4)
+	llill	%r0,2040
+	aghi	%r13,-1
+	j	.Lenc_loop
+.align	16
+.Lenc_loop:
+	sllg	%r1,%r8,3
+	srlg	%r2,%r8,5
+	srlg	%r3,%r8,13
+	srl	%r8,21
+	nr	%r8,%r0
+	ngr	%r1,%r0
+	nr	%r2,%r0
+	nr	%r3,%r0
+
+	srlg	%r5,%r9,13	# i0
+	sllg	%r6,%r9,3
+	srlg	%r7,%r9,5
+	srl	%r9,21
+	nr	%r5,%r0
+	nr	%r9,%r0
+	ngr	%r6,%r0
+	nr	%r7,%r0
+
+	l	%r8,0(%r8,%r12)	# Te0[s0>>24]
+	l	%r1,1(%r1,%r12)	# Te3[s0>>0]
+	l	%r2,2(%r2,%r12) # Te2[s0>>8]
+	l	%r3,3(%r3,%r12)	# Te1[s0>>16]
+
+	x	%r8,3(%r5,%r12)	# Te1[s1>>16]
+	l	%r9,0(%r9,%r12)	# Te0[s1>>24]
+	x	%r2,1(%r6,%r12)	# Te3[s1>>0]
+	x	%r3,2(%r7,%r12)	# Te2[s1>>8]
+
+	srlg	%r5,%r10,5	# i0
+	srlg	%r6,%r10,13	# i1
+	nr	%r5,%r0
+	nr	%r6,%r0
+	sllg	%r7,%r10,3
+	srl	%r10,21
+	nr	%r10,%r0
+	ngr	%r7,%r0
+
+	xr	%r9,%r1
+	srlg	%r14,%r11,5	# i1
+	sllg	%r1,%r11,3	# i0
+	nr	%r14,%r0
+	la	%r4,16(%r4)
+	ngr	%r1,%r0
+
+	x	%r8,2(%r5,%r12)	# Te2[s2>>8]
+	x	%r9,3(%r6,%r12)	# Te1[s2>>16]
+	l	%r10,0(%r10,%r12)	# Te0[s2>>24]
+	x	%r3,1(%r7,%r12)	# Te3[s2>>0]
+
+	srlg	%r7,%r11,13	# i2
+	xr	%r10,%r2
+	srl	%r11,21
+	nr	%r7,%r0
+	nr	%r11,%r0
+
+	x	%r8,0(%r4)
+	x	%r9,4(%r4)
+	x	%r10,8(%r4)
+	x	%r3,12(%r4)
+
+	x	%r8,1(%r1,%r12)	# Te3[s3>>0]
+	x	%r9,2(%r14,%r12)	# Te2[s3>>8]
+	x	%r10,3(%r7,%r12)	# Te1[s3>>16]
+	l	%r11,0(%r11,%r12)	# Te0[s3>>24]
+	xr	%r11,%r3
+
+	brct	%r13,.Lenc_loop
+	.align	16
+
+	sllg	%r1,%r8,3
+	srlg	%r2,%r8,5
+	ngr	%r1,%r0
+	srlg	%r3,%r8,13
+	srl	%r8,21
+	nr	%r8,%r0
+	nr	%r2,%r0
+	nr	%r3,%r0
+
+	srlg	%r5,%r9,13	# i0
+	sllg	%r6,%r9,3
+	ngr	%r6,%r0
+	srlg	%r7,%r9,5
+	srl	%r9,21
+	nr	%r5,%r0
+	nr	%r9,%r0
+	nr	%r7,%r0
+
+	llgc	%r8,2(%r8,%r12)	# Te4[s0>>24]
+	llgc	%r1,2(%r1,%r12)	# Te4[s0>>0]
+	sll	%r8,24
+	llgc	%r2,2(%r2,%r12)	# Te4[s0>>8]
+	llgc	%r3,2(%r3,%r12)	# Te4[s0>>16]
+	sll	%r2,8
+	sll	%r3,16
+
+	llgc	%r5,2(%r5,%r12)	# Te4[s1>>16]
+	llgc	%r9,2(%r9,%r12)	# Te4[s1>>24]
+	llgc	%r6,2(%r6,%r12)	# Te4[s1>>0]
+	llgc	%r7,2(%r7,%r12)	# Te4[s1>>8]
+	sll	%r5,16
+	sll	%r9,24
+	sll	%r7,8
+	or	%r8,%r5
+	or	%r9,%r1
+	or	%r2,%r6
+	or	%r3,%r7
+
+	srlg	%r5,%r10,5	# i0
+	srlg	%r6,%r10,13	# i1
+	nr	%r5,%r0
+	nr	%r6,%r0
+	sllg	%r7,%r10,3
+	srl	%r10,21
+	ngr	%r7,%r0
+	nr	%r10,%r0
+
+	sllg	%r1,%r11,3	# i0
+	srlg	%r14,%r11,5	# i1
+	ngr	%r1,%r0
+
+	llgc	%r5,2(%r5,%r12)	# Te4[s2>>8]
+	llgc	%r6,2(%r6,%r12)	# Te4[s2>>16]
+	sll	%r5,8
+	llgc	%r10,2(%r10,%r12)	# Te4[s2>>24]
+	llgc	%r7,2(%r7,%r12)	# Te4[s2>>0]
+	sll	%r6,16
+	nr	%r14,%r0
+	sll	%r10,24
+	or	%r8,%r5
+	or	%r9,%r6
+	or	%r10,%r2
+	or	%r3,%r7
+
+	srlg	%r7,%r11,13	# i2
+	srl	%r11,21
+	nr	%r7,%r0
+	nr	%r11,%r0
+
+	l	%r0,16(%r4)
+	l	%r2,20(%r4)
+
+	llgc	%r5,2(%r1,%r12)	# Te4[s3>>0]
+	llgc	%r6,2(%r14,%r12)	# Te4[s3>>8]
+	llgc	%r7,2(%r7,%r12)	# Te4[s3>>16]
+	llgc	%r11,2(%r11,%r12)	# Te4[s3>>24]
+	sll	%r6,8
+	sll	%r7,16
+	sll	%r11,24
+	or	%r8,%r5
+	or	%r9,%r6
+	or	%r10,%r7
+	or	%r11,%r3
+
+	lg	%r14,15*8(%r15)
+	xr	%r8,%r0
+	xr	%r9,%r2
+	x	%r10,24(%r4)
+	x	%r11,28(%r4)
+
+	br	%r14
+.size	_s390x_AES_encrypt,.-_s390x_AES_encrypt
+.type	AES_Td,@object
+.align	256
+AES_Td:
+.long	0x51f4a750,0x51f4a750
+.long	0x7e416553,0x7e416553
+.long	0x1a17a4c3,0x1a17a4c3
+.long	0x3a275e96,0x3a275e96
+.long	0x3bab6bcb,0x3bab6bcb
+.long	0x1f9d45f1,0x1f9d45f1
+.long	0xacfa58ab,0xacfa58ab
+.long	0x4be30393,0x4be30393
+.long	0x2030fa55,0x2030fa55
+.long	0xad766df6,0xad766df6
+.long	0x88cc7691,0x88cc7691
+.long	0xf5024c25,0xf5024c25
+.long	0x4fe5d7fc,0x4fe5d7fc
+.long	0xc52acbd7,0xc52acbd7
+.long	0x26354480,0x26354480
+.long	0xb562a38f,0xb562a38f
+.long	0xdeb15a49,0xdeb15a49
+.long	0x25ba1b67,0x25ba1b67
+.long	0x45ea0e98,0x45ea0e98
+.long	0x5dfec0e1,0x5dfec0e1
+.long	0xc32f7502,0xc32f7502
+.long	0x814cf012,0x814cf012
+.long	0x8d4697a3,0x8d4697a3
+.long	0x6bd3f9c6,0x6bd3f9c6
+.long	0x038f5fe7,0x038f5fe7
+.long	0x15929c95,0x15929c95
+.long	0xbf6d7aeb,0xbf6d7aeb
+.long	0x955259da,0x955259da
+.long	0xd4be832d,0xd4be832d
+.long	0x587421d3,0x587421d3
+.long	0x49e06929,0x49e06929
+.long	0x8ec9c844,0x8ec9c844
+.long	0x75c2896a,0x75c2896a
+.long	0xf48e7978,0xf48e7978
+.long	0x99583e6b,0x99583e6b
+.long	0x27b971dd,0x27b971dd
+.long	0xbee14fb6,0xbee14fb6
+.long	0xf088ad17,0xf088ad17
+.long	0xc920ac66,0xc920ac66
+.long	0x7dce3ab4,0x7dce3ab4
+.long	0x63df4a18,0x63df4a18
+.long	0xe51a3182,0xe51a3182
+.long	0x97513360,0x97513360
+.long	0x62537f45,0x62537f45
+.long	0xb16477e0,0xb16477e0
+.long	0xbb6bae84,0xbb6bae84
+.long	0xfe81a01c,0xfe81a01c
+.long	0xf9082b94,0xf9082b94
+.long	0x70486858,0x70486858
+.long	0x8f45fd19,0x8f45fd19
+.long	0x94de6c87,0x94de6c87
+.long	0x527bf8b7,0x527bf8b7
+.long	0xab73d323,0xab73d323
+.long	0x724b02e2,0x724b02e2
+.long	0xe31f8f57,0xe31f8f57
+.long	0x6655ab2a,0x6655ab2a
+.long	0xb2eb2807,0xb2eb2807
+.long	0x2fb5c203,0x2fb5c203
+.long	0x86c57b9a,0x86c57b9a
+.long	0xd33708a5,0xd33708a5
+.long	0x302887f2,0x302887f2
+.long	0x23bfa5b2,0x23bfa5b2
+.long	0x02036aba,0x02036aba
+.long	0xed16825c,0xed16825c
+.long	0x8acf1c2b,0x8acf1c2b
+.long	0xa779b492,0xa779b492
+.long	0xf307f2f0,0xf307f2f0
+.long	0x4e69e2a1,0x4e69e2a1
+.long	0x65daf4cd,0x65daf4cd
+.long	0x0605bed5,0x0605bed5
+.long	0xd134621f,0xd134621f
+.long	0xc4a6fe8a,0xc4a6fe8a
+.long	0x342e539d,0x342e539d
+.long	0xa2f355a0,0xa2f355a0
+.long	0x058ae132,0x058ae132
+.long	0xa4f6eb75,0xa4f6eb75
+.long	0x0b83ec39,0x0b83ec39
+.long	0x4060efaa,0x4060efaa
+.long	0x5e719f06,0x5e719f06
+.long	0xbd6e1051,0xbd6e1051
+.long	0x3e218af9,0x3e218af9
+.long	0x96dd063d,0x96dd063d
+.long	0xdd3e05ae,0xdd3e05ae
+.long	0x4de6bd46,0x4de6bd46
+.long	0x91548db5,0x91548db5
+.long	0x71c45d05,0x71c45d05
+.long	0x0406d46f,0x0406d46f
+.long	0x605015ff,0x605015ff
+.long	0x1998fb24,0x1998fb24
+.long	0xd6bde997,0xd6bde997
+.long	0x894043cc,0x894043cc
+.long	0x67d99e77,0x67d99e77
+.long	0xb0e842bd,0xb0e842bd
+.long	0x07898b88,0x07898b88
+.long	0xe7195b38,0xe7195b38
+.long	0x79c8eedb,0x79c8eedb
+.long	0xa17c0a47,0xa17c0a47
+.long	0x7c420fe9,0x7c420fe9
+.long	0xf8841ec9,0xf8841ec9
+.long	0x00000000,0x00000000
+.long	0x09808683,0x09808683
+.long	0x322bed48,0x322bed48
+.long	0x1e1170ac,0x1e1170ac
+.long	0x6c5a724e,0x6c5a724e
+.long	0xfd0efffb,0xfd0efffb
+.long	0x0f853856,0x0f853856
+.long	0x3daed51e,0x3daed51e
+.long	0x362d3927,0x362d3927
+.long	0x0a0fd964,0x0a0fd964
+.long	0x685ca621,0x685ca621
+.long	0x9b5b54d1,0x9b5b54d1
+.long	0x24362e3a,0x24362e3a
+.long	0x0c0a67b1,0x0c0a67b1
+.long	0x9357e70f,0x9357e70f
+.long	0xb4ee96d2,0xb4ee96d2
+.long	0x1b9b919e,0x1b9b919e
+.long	0x80c0c54f,0x80c0c54f
+.long	0x61dc20a2,0x61dc20a2
+.long	0x5a774b69,0x5a774b69
+.long	0x1c121a16,0x1c121a16
+.long	0xe293ba0a,0xe293ba0a
+.long	0xc0a02ae5,0xc0a02ae5
+.long	0x3c22e043,0x3c22e043
+.long	0x121b171d,0x121b171d
+.long	0x0e090d0b,0x0e090d0b
+.long	0xf28bc7ad,0xf28bc7ad
+.long	0x2db6a8b9,0x2db6a8b9
+.long	0x141ea9c8,0x141ea9c8
+.long	0x57f11985,0x57f11985
+.long	0xaf75074c,0xaf75074c
+.long	0xee99ddbb,0xee99ddbb
+.long	0xa37f60fd,0xa37f60fd
+.long	0xf701269f,0xf701269f
+.long	0x5c72f5bc,0x5c72f5bc
+.long	0x44663bc5,0x44663bc5
+.long	0x5bfb7e34,0x5bfb7e34
+.long	0x8b432976,0x8b432976
+.long	0xcb23c6dc,0xcb23c6dc
+.long	0xb6edfc68,0xb6edfc68
+.long	0xb8e4f163,0xb8e4f163
+.long	0xd731dcca,0xd731dcca
+.long	0x42638510,0x42638510
+.long	0x13972240,0x13972240
+.long	0x84c61120,0x84c61120
+.long	0x854a247d,0x854a247d
+.long	0xd2bb3df8,0xd2bb3df8
+.long	0xaef93211,0xaef93211
+.long	0xc729a16d,0xc729a16d
+.long	0x1d9e2f4b,0x1d9e2f4b
+.long	0xdcb230f3,0xdcb230f3
+.long	0x0d8652ec,0x0d8652ec
+.long	0x77c1e3d0,0x77c1e3d0
+.long	0x2bb3166c,0x2bb3166c
+.long	0xa970b999,0xa970b999
+.long	0x119448fa,0x119448fa
+.long	0x47e96422,0x47e96422
+.long	0xa8fc8cc4,0xa8fc8cc4
+.long	0xa0f03f1a,0xa0f03f1a
+.long	0x567d2cd8,0x567d2cd8
+.long	0x223390ef,0x223390ef
+.long	0x87494ec7,0x87494ec7
+.long	0xd938d1c1,0xd938d1c1
+.long	0x8ccaa2fe,0x8ccaa2fe
+.long	0x98d40b36,0x98d40b36
+.long	0xa6f581cf,0xa6f581cf
+.long	0xa57ade28,0xa57ade28
+.long	0xdab78e26,0xdab78e26
+.long	0x3fadbfa4,0x3fadbfa4
+.long	0x2c3a9de4,0x2c3a9de4
+.long	0x5078920d,0x5078920d
+.long	0x6a5fcc9b,0x6a5fcc9b
+.long	0x547e4662,0x547e4662
+.long	0xf68d13c2,0xf68d13c2
+.long	0x90d8b8e8,0x90d8b8e8
+.long	0x2e39f75e,0x2e39f75e
+.long	0x82c3aff5,0x82c3aff5
+.long	0x9f5d80be,0x9f5d80be
+.long	0x69d0937c,0x69d0937c
+.long	0x6fd52da9,0x6fd52da9
+.long	0xcf2512b3,0xcf2512b3
+.long	0xc8ac993b,0xc8ac993b
+.long	0x10187da7,0x10187da7
+.long	0xe89c636e,0xe89c636e
+.long	0xdb3bbb7b,0xdb3bbb7b
+.long	0xcd267809,0xcd267809
+.long	0x6e5918f4,0x6e5918f4
+.long	0xec9ab701,0xec9ab701
+.long	0x834f9aa8,0x834f9aa8
+.long	0xe6956e65,0xe6956e65
+.long	0xaaffe67e,0xaaffe67e
+.long	0x21bccf08,0x21bccf08
+.long	0xef15e8e6,0xef15e8e6
+.long	0xbae79bd9,0xbae79bd9
+.long	0x4a6f36ce,0x4a6f36ce
+.long	0xea9f09d4,0xea9f09d4
+.long	0x29b07cd6,0x29b07cd6
+.long	0x31a4b2af,0x31a4b2af
+.long	0x2a3f2331,0x2a3f2331
+.long	0xc6a59430,0xc6a59430
+.long	0x35a266c0,0x35a266c0
+.long	0x744ebc37,0x744ebc37
+.long	0xfc82caa6,0xfc82caa6
+.long	0xe090d0b0,0xe090d0b0
+.long	0x33a7d815,0x33a7d815
+.long	0xf104984a,0xf104984a
+.long	0x41ecdaf7,0x41ecdaf7
+.long	0x7fcd500e,0x7fcd500e
+.long	0x1791f62f,0x1791f62f
+.long	0x764dd68d,0x764dd68d
+.long	0x43efb04d,0x43efb04d
+.long	0xccaa4d54,0xccaa4d54
+.long	0xe49604df,0xe49604df
+.long	0x9ed1b5e3,0x9ed1b5e3
+.long	0x4c6a881b,0x4c6a881b
+.long	0xc12c1fb8,0xc12c1fb8
+.long	0x4665517f,0x4665517f
+.long	0x9d5eea04,0x9d5eea04
+.long	0x018c355d,0x018c355d
+.long	0xfa877473,0xfa877473
+.long	0xfb0b412e,0xfb0b412e
+.long	0xb3671d5a,0xb3671d5a
+.long	0x92dbd252,0x92dbd252
+.long	0xe9105633,0xe9105633
+.long	0x6dd64713,0x6dd64713
+.long	0x9ad7618c,0x9ad7618c
+.long	0x37a10c7a,0x37a10c7a
+.long	0x59f8148e,0x59f8148e
+.long	0xeb133c89,0xeb133c89
+.long	0xcea927ee,0xcea927ee
+.long	0xb761c935,0xb761c935
+.long	0xe11ce5ed,0xe11ce5ed
+.long	0x7a47b13c,0x7a47b13c
+.long	0x9cd2df59,0x9cd2df59
+.long	0x55f2733f,0x55f2733f
+.long	0x1814ce79,0x1814ce79
+.long	0x73c737bf,0x73c737bf
+.long	0x53f7cdea,0x53f7cdea
+.long	0x5ffdaa5b,0x5ffdaa5b
+.long	0xdf3d6f14,0xdf3d6f14
+.long	0x7844db86,0x7844db86
+.long	0xcaaff381,0xcaaff381
+.long	0xb968c43e,0xb968c43e
+.long	0x3824342c,0x3824342c
+.long	0xc2a3405f,0xc2a3405f
+.long	0x161dc372,0x161dc372
+.long	0xbce2250c,0xbce2250c
+.long	0x283c498b,0x283c498b
+.long	0xff0d9541,0xff0d9541
+.long	0x39a80171,0x39a80171
+.long	0x080cb3de,0x080cb3de
+.long	0xd8b4e49c,0xd8b4e49c
+.long	0x6456c190,0x6456c190
+.long	0x7bcb8461,0x7bcb8461
+.long	0xd532b670,0xd532b670
+.long	0x486c5c74,0x486c5c74
+.long	0xd0b85742,0xd0b85742
+# Td4[256]
+.byte	0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
+.byte	0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
+.byte	0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
+.byte	0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
+.byte	0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
+.byte	0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
+.byte	0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
+.byte	0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
+.byte	0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
+.byte	0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
+.byte	0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
+.byte	0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
+.byte	0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
+.byte	0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
+.byte	0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
+.byte	0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
+.byte	0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
+.byte	0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
+.byte	0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
+.byte	0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
+.byte	0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
+.byte	0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
+.byte	0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
+.byte	0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
+.byte	0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
+.byte	0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
+.byte	0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
+.byte	0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
+.byte	0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
+.byte	0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
+.byte	0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
+.byte	0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
+.size	AES_Td,.-AES_Td
+
+# void AES_decrypt(const unsigned char *inp, unsigned char *out,
+# 		 const AES_KEY *key) {
+.globl	AES_decrypt
+.type	AES_decrypt,@function
+AES_decrypt:
+	l	%r0,240(%r4)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Ldsoft
+
+	la	%r1,0(%r4)
+	#la	%r2,0(%r2)
+	la	%r4,0(%r3)
+	lghi	%r3,16		# single block length
+	.long	0xb92e0042	# km %r4,%r2
+	brc	1,.-4		# can this happen?
+	br	%r14
+.align	64
+.Ldsoft:
+	stmg	%r3,%r14,3*8(%r15)
+
+	llgf	%r8,0(%r2)
+	llgf	%r9,4(%r2)
+	llgf	%r10,8(%r2)
+	llgf	%r11,12(%r2)
+
+	larl	%r12,AES_Td
+	bras	%r14,_s390x_AES_decrypt
+
+	lg	%r3,3*8(%r15)
+	st	%r8,0(%r3)
+	st	%r9,4(%r3)
+	st	%r10,8(%r3)
+	st	%r11,12(%r3)
+
+	lmg	%r6,%r14,6*8(%r15)
+	br	%r14
+.size	AES_decrypt,.-AES_decrypt
+
+.type   _s390x_AES_decrypt,@function
+.align	16
+_s390x_AES_decrypt:
+	stg	%r14,15*8(%r15)
+	x	%r8,0(%r4)
+	x	%r9,4(%r4)
+	x	%r10,8(%r4)
+	x	%r11,12(%r4)
+	l	%r13,240(%r4)
+	llill	%r0,2040
+	aghi	%r13,-1
+	j	.Ldec_loop
+.align	16
+.Ldec_loop:
+	srlg	%r1,%r8,13
+	srlg	%r2,%r8,5
+	sllg	%r3,%r8,3
+	srl	%r8,21
+	nr	%r8,%r0
+	nr	%r1,%r0
+	nr	%r2,%r0
+	ngr	%r3,%r0
+
+	sllg	%r5,%r9,3	# i0
+	srlg	%r6,%r9,13
+	srlg	%r7,%r9,5
+	srl	%r9,21
+	ngr	%r5,%r0
+	nr	%r9,%r0
+	nr	%r6,%r0
+	nr	%r7,%r0
+
+	l	%r8,0(%r8,%r12)	# Td0[s0>>24]
+	l	%r1,3(%r1,%r12)	# Td1[s0>>16]
+	l	%r2,2(%r2,%r12)	# Td2[s0>>8]
+	l	%r3,1(%r3,%r12)	# Td3[s0>>0]
+
+	x	%r8,1(%r5,%r12)	# Td3[s1>>0]
+	l	%r9,0(%r9,%r12)	# Td0[s1>>24]
+	x	%r2,3(%r6,%r12)	# Td1[s1>>16]
+	x	%r3,2(%r7,%r12)	# Td2[s1>>8]
+
+	srlg	%r5,%r10,5	# i0
+	sllg	%r6,%r10,3	# i1
+	srlg	%r7,%r10,13
+	srl	%r10,21
+	nr	%r5,%r0
+	ngr	%r6,%r0
+	nr	%r10,%r0
+	nr	%r7,%r0
+
+	xr	%r9,%r1
+	srlg	%r14,%r11,5	# i1
+	srlg	%r1,%r11,13	# i0
+	nr	%r14,%r0
+	la	%r4,16(%r4)
+	nr	%r1,%r0
+
+	x	%r8,2(%r5,%r12)	# Td2[s2>>8]
+	x	%r9,1(%r6,%r12)	# Td3[s2>>0]
+	l	%r10,0(%r10,%r12)	# Td0[s2>>24]
+	x	%r3,3(%r7,%r12)	# Td1[s2>>16]
+
+	sllg	%r7,%r11,3	# i2
+	srl	%r11,21
+	ngr	%r7,%r0
+	nr	%r11,%r0
+
+	xr	%r10,%r2
+	x	%r8,0(%r4)
+	x	%r9,4(%r4)
+	x	%r10,8(%r4)
+	x	%r3,12(%r4)
+
+	x	%r8,3(%r1,%r12)	# Td1[s3>>16]
+	x	%r9,2(%r14,%r12)	# Td2[s3>>8]
+	x	%r10,1(%r7,%r12)	# Td3[s3>>0]
+	l	%r11,0(%r11,%r12)	# Td0[s3>>24]
+	xr	%r11,%r3
+
+	brct	%r13,.Ldec_loop
+	.align	16
+
+	l	%r1,2048(%r12)	# prefetch Td4
+	l	%r2,2112(%r12)
+	l	%r3,2176(%r12)
+	l	%r5,2240(%r12)
+	llill	%r0,0xff
+
+	srlg	%r7,%r8,24	# i0
+	srlg	%r1,%r8,16
+	srlg	%r2,%r8,8
+	nr	%r8,%r0	# i3
+	nr	%r1,%r0
+
+	srlg	%r5,%r9,24
+	nr	%r2,%r0
+	srlg	%r6,%r9,16
+	srlg	%r14,%r9,8
+	nr	%r9,%r0	# i0
+	nr	%r6,%r0
+	nr	%r14,%r0
+
+	llgc	%r7,2048(%r7,%r12)	# Td4[s0>>24]
+	llgc	%r1,2048(%r1,%r12)	# Td4[s0>>16]
+	llgc	%r2,2048(%r2,%r12)	# Td4[s0>>8]
+	sll	%r1,16
+	llgc	%r3,2048(%r8,%r12)	# Td4[s0>>0]
+	sllg	%r8,%r7,24
+	sll	%r2,8
+
+	llgc	%r9,2048(%r9,%r12)	# Td4[s1>>0]
+	llgc	%r5,2048(%r5,%r12)	# Td4[s1>>24]
+	llgc	%r6,2048(%r6,%r12)	# Td4[s1>>16]
+	sll	%r5,24
+	llgc	%r7,2048(%r14,%r12)	# Td4[s1>>8]
+	sll	%r6,16
+	sll	%r7,8
+	or	%r8,%r9
+	or	%r1,%r5
+	or	%r2,%r6
+	or	%r3,%r7
+
+	srlg	%r5,%r10,8	# i0
+	srlg	%r6,%r10,24
+	srlg	%r7,%r10,16
+	nr	%r10,%r0	# i1
+	nr	%r5,%r0
+	nr	%r7,%r0
+	llgc	%r5,2048(%r5,%r12)	# Td4[s2>>8]
+	llgc	%r9,2048(%r10,%r12)	# Td4[s2>>0]
+	llgc	%r6,2048(%r6,%r12)	# Td4[s2>>24]
+	llgc	%r7,2048(%r7,%r12)	# Td4[s2>>16]
+	sll	%r5,8
+	sll	%r6,24
+	or	%r8,%r5
+	sll	%r7,16
+	or	%r2,%r6
+	or	%r3,%r7
+
+	srlg	%r5,%r11,16	# i0
+	srlg	%r6,%r11,8	# i1
+	srlg	%r7,%r11,24
+	nr	%r11,%r0	# i2
+	nr	%r5,%r0
+	nr	%r6,%r0
+
+	lg	%r14,15*8(%r15)
+	or	%r9,%r1
+	l	%r0,16(%r4)
+	l	%r1,20(%r4)
+
+	llgc	%r5,2048(%r5,%r12)	# Td4[s3>>16]
+	llgc	%r6,2048(%r6,%r12)	# Td4[s3>>8]
+	sll	%r5,16
+	llgc	%r10,2048(%r11,%r12)	# Td4[s3>>0]
+	llgc	%r11,2048(%r7,%r12)	# Td4[s3>>24]
+	sll	%r6,8
+	sll	%r11,24
+	or	%r8,%r5
+	or	%r9,%r6
+	or	%r10,%r2
+	or	%r11,%r3
+
+	xr	%r8,%r0
+	xr	%r9,%r1
+	x	%r10,24(%r4)
+	x	%r11,28(%r4)
+
+	br	%r14
+.size	_s390x_AES_decrypt,.-_s390x_AES_decrypt
+# void AES_set_encrypt_key(const unsigned char *in, int bits,
+# 		 AES_KEY *key) {
+.globl	AES_set_encrypt_key
+.type	AES_set_encrypt_key,@function
+.align	16
+AES_set_encrypt_key:
+_s390x_AES_set_encrypt_key:
+	lghi	%r0,0
+	clgr	%r2,%r0
+	je	.Lminus1
+	clgr	%r4,%r0
+	je	.Lminus1
+
+	lghi	%r0,128
+	clr	%r3,%r0
+	je	.Lproceed
+	lghi	%r0,192
+	clr	%r3,%r0
+	je	.Lproceed
+	lghi	%r0,256
+	clr	%r3,%r0
+	je	.Lproceed
+	lghi	%r2,-2
+	br	%r14
+
+.align	16
+.Lproceed:
+	# convert bits to km(c) code, [128,192,256]->[18,19,20]
+	lhi	%r5,-128
+	lhi	%r0,18
+	ar	%r5,%r3
+	srl	%r5,6
+	ar	%r5,%r0
+
+	larl	%r1,OPENSSL_s390xcap_P
+	llihh	%r0,0x8000
+	srlg	%r0,%r0,0(%r5)
+	ng	%r0,S390X_KM(%r1)  # check availability of both km...
+	ng	%r0,S390X_KMC(%r1) # ...and kmc support for given key length
+	jz	.Lekey_internal
+
+	lmg	%r0,%r1,0(%r2)	# just copy 128 bits...
+	stmg	%r0,%r1,0(%r4)
+	lhi	%r0,192
+	cr	%r3,%r0
+	jl	1f
+	lg	%r1,16(%r2)
+	stg	%r1,16(%r4)
+	je	1f
+	lg	%r1,24(%r2)
+	stg	%r1,24(%r4)
+1:	st	%r3,236(%r4)	# save bits [for debugging purposes]
+	lgr	%r0,%r5
+	st	%r5,240(%r4)	# save km(c) code
+	lghi	%r2,0
+	br	%r14
+.align	16
+.Lekey_internal:
+	stmg	%r4,%r13,4*8(%r15)	# all non-volatile regs and %r4
+
+	larl	%r12,AES_Te+2048
+
+	llgf	%r8,0(%r2)
+	llgf	%r9,4(%r2)
+	llgf	%r10,8(%r2)
+	llgf	%r11,12(%r2)
+	st	%r8,0(%r4)
+	st	%r9,4(%r4)
+	st	%r10,8(%r4)
+	st	%r11,12(%r4)
+	lghi	%r0,128
+	cr	%r3,%r0
+	jne	.Lnot128
+
+	llill	%r0,0xff
+	lghi	%r3,0			# i=0
+	lghi	%r13,10
+	st	%r13,240(%r4)
+
+	llgfr	%r2,%r11			# temp=rk[3]
+	srlg	%r5,%r11,8
+	srlg	%r6,%r11,16
+	srlg	%r7,%r11,24
+	nr	%r2,%r0
+	nr	%r5,%r0
+	nr	%r6,%r0
+
+.align	16
+.L128_loop:
+	la	%r2,0(%r2,%r12)
+	la	%r5,0(%r5,%r12)
+	la	%r6,0(%r6,%r12)
+	la	%r7,0(%r7,%r12)
+	icm	%r2,2,0(%r2)		# Te4[rk[3]>>0]<<8
+	icm	%r2,4,0(%r5)		# Te4[rk[3]>>8]<<16
+	icm	%r2,8,0(%r6)		# Te4[rk[3]>>16]<<24
+	icm	%r2,1,0(%r7)		# Te4[rk[3]>>24]
+	x	%r2,256(%r3,%r12)	# rcon[i]
+	xr	%r8,%r2			# rk[4]=rk[0]^...
+	xr	%r9,%r8			# rk[5]=rk[1]^rk[4]
+	xr	%r10,%r9			# rk[6]=rk[2]^rk[5]
+	xr	%r11,%r10			# rk[7]=rk[3]^rk[6]
+
+	llgfr	%r2,%r11			# temp=rk[3]
+	srlg	%r5,%r11,8
+	srlg	%r6,%r11,16
+	nr	%r2,%r0
+	nr	%r5,%r0
+	srlg	%r7,%r11,24
+	nr	%r6,%r0
+
+	st	%r8,16(%r4)
+	st	%r9,20(%r4)
+	st	%r10,24(%r4)
+	st	%r11,28(%r4)
+	la	%r4,16(%r4)		# key+=4
+	la	%r3,4(%r3)		# i++
+	brct	%r13,.L128_loop
+	lghi	%r0,10
+	lghi	%r2,0
+	lmg	%r4,%r13,4*8(%r15)
+	br	%r14
+
+.align	16
+.Lnot128:
+	llgf	%r0,16(%r2)
+	llgf	%r1,20(%r2)
+	st	%r0,16(%r4)
+	st	%r1,20(%r4)
+	lghi	%r0,192
+	cr	%r3,%r0
+	jne	.Lnot192
+
+	llill	%r0,0xff
+	lghi	%r3,0			# i=0
+	lghi	%r13,12
+	st	%r13,240(%r4)
+	lghi	%r13,8
+
+	srlg	%r5,%r1,8
+	srlg	%r6,%r1,16
+	srlg	%r7,%r1,24
+	nr	%r1,%r0
+	nr	%r5,%r0
+	nr	%r6,%r0
+
+.align	16
+.L192_loop:
+	la	%r1,0(%r1,%r12)
+	la	%r5,0(%r5,%r12)
+	la	%r6,0(%r6,%r12)
+	la	%r7,0(%r7,%r12)
+	icm	%r1,2,0(%r1)		# Te4[rk[5]>>0]<<8
+	icm	%r1,4,0(%r5)		# Te4[rk[5]>>8]<<16
+	icm	%r1,8,0(%r6)		# Te4[rk[5]>>16]<<24
+	icm	%r1,1,0(%r7)		# Te4[rk[5]>>24]
+	x	%r1,256(%r3,%r12)	# rcon[i]
+	xr	%r8,%r1			# rk[6]=rk[0]^...
+	xr	%r9,%r8			# rk[7]=rk[1]^rk[6]
+	xr	%r10,%r9			# rk[8]=rk[2]^rk[7]
+	xr	%r11,%r10			# rk[9]=rk[3]^rk[8]
+
+	st	%r8,24(%r4)
+	st	%r9,28(%r4)
+	st	%r10,32(%r4)
+	st	%r11,36(%r4)
+	brct	%r13,.L192_continue
+	lghi	%r0,12
+	lghi	%r2,0
+	lmg	%r4,%r13,4*8(%r15)
+	br	%r14
+
+.align	16
+.L192_continue:
+	lgr	%r1,%r11
+	x	%r1,16(%r4)		# rk[10]=rk[4]^rk[9]
+	st	%r1,40(%r4)
+	x	%r1,20(%r4)		# rk[11]=rk[5]^rk[10]
+	st	%r1,44(%r4)
+
+	srlg	%r5,%r1,8
+	srlg	%r6,%r1,16
+	srlg	%r7,%r1,24
+	nr	%r1,%r0
+	nr	%r5,%r0
+	nr	%r6,%r0
+
+	la	%r4,24(%r4)		# key+=6
+	la	%r3,4(%r3)		# i++
+	j	.L192_loop
+
+.align	16
+.Lnot192:
+	llgf	%r0,24(%r2)
+	llgf	%r1,28(%r2)
+	st	%r0,24(%r4)
+	st	%r1,28(%r4)
+	llill	%r0,0xff
+	lghi	%r3,0			# i=0
+	lghi	%r13,14
+	st	%r13,240(%r4)
+	lghi	%r13,7
+
+	srlg	%r5,%r1,8
+	srlg	%r6,%r1,16
+	srlg	%r7,%r1,24
+	nr	%r1,%r0
+	nr	%r5,%r0
+	nr	%r6,%r0
+
+.align	16
+.L256_loop:
+	la	%r1,0(%r1,%r12)
+	la	%r5,0(%r5,%r12)
+	la	%r6,0(%r6,%r12)
+	la	%r7,0(%r7,%r12)
+	icm	%r1,2,0(%r1)		# Te4[rk[7]>>0]<<8
+	icm	%r1,4,0(%r5)		# Te4[rk[7]>>8]<<16
+	icm	%r1,8,0(%r6)		# Te4[rk[7]>>16]<<24
+	icm	%r1,1,0(%r7)		# Te4[rk[7]>>24]
+	x	%r1,256(%r3,%r12)	# rcon[i]
+	xr	%r8,%r1			# rk[8]=rk[0]^...
+	xr	%r9,%r8			# rk[9]=rk[1]^rk[8]
+	xr	%r10,%r9			# rk[10]=rk[2]^rk[9]
+	xr	%r11,%r10			# rk[11]=rk[3]^rk[10]
+	st	%r8,32(%r4)
+	st	%r9,36(%r4)
+	st	%r10,40(%r4)
+	st	%r11,44(%r4)
+	brct	%r13,.L256_continue
+	lghi	%r0,14
+	lghi	%r2,0
+	lmg	%r4,%r13,4*8(%r15)
+	br	%r14
+
+.align	16
+.L256_continue:
+	lgr	%r1,%r11			# temp=rk[11]
+	srlg	%r5,%r11,8
+	srlg	%r6,%r11,16
+	srlg	%r7,%r11,24
+	nr	%r1,%r0
+	nr	%r5,%r0
+	nr	%r6,%r0
+	la	%r1,0(%r1,%r12)
+	la	%r5,0(%r5,%r12)
+	la	%r6,0(%r6,%r12)
+	la	%r7,0(%r7,%r12)
+	llgc	%r1,0(%r1)		# Te4[rk[11]>>0]
+	icm	%r1,2,0(%r5)		# Te4[rk[11]>>8]<<8
+	icm	%r1,4,0(%r6)		# Te4[rk[11]>>16]<<16
+	icm	%r1,8,0(%r7)		# Te4[rk[11]>>24]<<24
+	x	%r1,16(%r4)		# rk[12]=rk[4]^...
+	st	%r1,48(%r4)
+	x	%r1,20(%r4)		# rk[13]=rk[5]^rk[12]
+	st	%r1,52(%r4)
+	x	%r1,24(%r4)		# rk[14]=rk[6]^rk[13]
+	st	%r1,56(%r4)
+	x	%r1,28(%r4)		# rk[15]=rk[7]^rk[14]
+	st	%r1,60(%r4)
+
+	srlg	%r5,%r1,8
+	srlg	%r6,%r1,16
+	srlg	%r7,%r1,24
+	nr	%r1,%r0
+	nr	%r5,%r0
+	nr	%r6,%r0
+
+	la	%r4,32(%r4)		# key+=8
+	la	%r3,4(%r3)		# i++
+	j	.L256_loop
+
+.Lminus1:
+	lghi	%r2,-1
+	br	%r14
+.size	AES_set_encrypt_key,.-AES_set_encrypt_key
+
+# void AES_set_decrypt_key(const unsigned char *in, int bits,
+# 		 AES_KEY *key) {
+.globl	AES_set_decrypt_key
+.type	AES_set_decrypt_key,@function
+.align	16
+AES_set_decrypt_key:
+	#stg	%r4,4*8(%r15)	# I rely on AES_set_encrypt_key to
+	stg	%r14,14*8(%r15)	# save non-volatile registers and %r4!
+	bras	%r14,_s390x_AES_set_encrypt_key
+	#lg	%r4,4*8(%r15)
+	lg	%r14,14*8(%r15)
+	ltgr	%r2,%r2
+	bnzr	%r14
+	#l	%r0,240(%r4)
+	lhi	%r1,16
+	cr	%r0,%r1
+	jl	.Lgo
+	oill	%r0,S390X_DECRYPT	# set "decrypt" bit
+	st	%r0,240(%r4)
+	br	%r14
+.align	16
+.Lgo:	lgr	%r13,%r0	#llgf	%r13,240(%r4)
+	la	%r5,0(%r4)
+	sllg	%r6,%r13,4
+	la	%r6,0(%r6,%r4)
+	srl	%r13,1
+	lghi	%r1,-16
+
+.align	16
+.Linv:	lmg	%r8,%r9,0(%r5)
+	lmg	%r10,%r11,0(%r6)
+	stmg	%r8,%r9,0(%r6)
+	stmg	%r10,%r11,0(%r5)
+	la	%r5,16(%r5)
+	la	%r6,0(%r1,%r6)
+	brct	%r13,.Linv
+	llgf	%r13,240(%r4)
+	aghi	%r13,-1
+	sll	%r13,2	# (rounds-1)*4
+	llilh	%r5,0x8080
+	llilh	%r6,0x1b1b
+	llilh	%r7,0xfefe
+	oill	%r5,0x8080
+	oill	%r6,0x1b1b
+	oill	%r7,0xfefe
+
+.align	16
+.Lmix:	l	%r8,16(%r4)	# tp1
+	lr	%r9,%r8
+	ngr	%r9,%r5
+	srlg	%r1,%r9,7
+	slr	%r9,%r1
+	nr	%r9,%r6
+	sllg	%r1,%r8,1
+	nr	%r1,%r7
+	xr	%r9,%r1		# tp2
+
+	lr	%r10,%r9
+	ngr	%r10,%r5
+	srlg	%r1,%r10,7
+	slr	%r10,%r1
+	nr	%r10,%r6
+	sllg	%r1,%r9,1
+	nr	%r1,%r7
+	xr	%r10,%r1		# tp4
+
+	lr	%r11,%r10
+	ngr	%r11,%r5
+	srlg	%r1,%r11,7
+	slr	%r11,%r1
+	nr	%r11,%r6
+	sllg	%r1,%r10,1
+	nr	%r1,%r7
+	xr	%r11,%r1		# tp8
+
+	xr	%r9,%r8		# tp2^tp1
+	xr	%r10,%r8		# tp4^tp1
+	rll	%r8,%r8,24	# = ROTATE(tp1,8)
+	xr	%r10,%r11		# ^=tp8
+	xr	%r8,%r9		# ^=tp2^tp1
+	xr	%r9,%r11		# tp2^tp1^tp8
+	xr	%r8,%r10		# ^=tp4^tp1^tp8
+	rll	%r9,%r9,8
+	rll	%r10,%r10,16
+	xr	%r8,%r9		# ^= ROTATE(tp8^tp2^tp1,24)
+	rll	%r11,%r11,24
+	xr	%r8,%r10    	# ^= ROTATE(tp8^tp4^tp1,16)
+	xr	%r8,%r11		# ^= ROTATE(tp8,8)
+
+	st	%r8,16(%r4)
+	la	%r4,4(%r4)
+	brct	%r13,.Lmix
+
+	lmg	%r6,%r13,6*8(%r15)# as was saved by AES_set_encrypt_key!
+	lghi	%r2,0
+	br	%r14
+.size	AES_set_decrypt_key,.-AES_set_decrypt_key
+.globl	AES_cbc_encrypt
+.type	AES_cbc_encrypt,@function
+.align	16
+AES_cbc_encrypt:
+	xgr	%r3,%r4		# flip %r3 and %r4, out and len
+	xgr	%r4,%r3
+	xgr	%r3,%r4
+	lhi	%r0,16
+	cl	%r0,240(%r5)
+	jh	.Lcbc_software
+
+	lg	%r0,0(%r6)	# copy ivec
+	lg	%r1,8(%r6)
+	stmg	%r0,%r1,16(%r15)
+	lmg	%r0,%r1,0(%r5)	# copy key, cover 256 bit
+	stmg	%r0,%r1,32(%r15)
+	lmg	%r0,%r1,16(%r5)
+	stmg	%r0,%r1,48(%r15)
+	l	%r0,240(%r5)	# load kmc code
+	lghi	%r5,15		# res=len%16, len-=res;
+	ngr	%r5,%r3
+	slgr	%r3,%r5
+	la	%r1,16(%r15)	# parameter block - ivec || key
+	jz	.Lkmc_truncated
+	.long	0xb92f0042	# kmc %r4,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+	ltr	%r5,%r5
+	jnz	.Lkmc_truncated
+.Lkmc_done:
+	lmg	%r0,%r1,16(%r15)	# copy ivec to caller
+	stg	%r0,0(%r6)
+	stg	%r1,8(%r6)
+	br	%r14
+.align	16
+.Lkmc_truncated:
+	ahi	%r5,-1		# it's the way it's encoded in mvc
+	tmll	%r0,S390X_DECRYPT
+	jnz	.Lkmc_truncated_dec
+	lghi	%r1,0
+	stg	%r1,16*8(%r15)
+	stg	%r1,16*8+8(%r15)
+	bras	%r1,1f
+	mvc	16*8(1,%r15),0(%r2)
+1:	ex	%r5,0(%r1)
+	la	%r1,16(%r15)	# restore parameter block
+	la	%r2,16*8(%r15)
+	lghi	%r3,16
+	.long	0xb92f0042	# kmc %r4,%r2
+	j	.Lkmc_done
+.align	16
+.Lkmc_truncated_dec:
+	stg	%r4,4*8(%r15)
+	la	%r4,16*8(%r15)
+	lghi	%r3,16
+	.long	0xb92f0042	# kmc %r4,%r2
+	lg	%r4,4*8(%r15)
+	bras	%r1,2f
+	mvc	0(1,%r4),16*8(%r15)
+2:	ex	%r5,0(%r1)
+	j	.Lkmc_done
+.align	16
+.Lcbc_software:
+	stmg	%r5,%r14,5*8(%r15)
+	lhi	%r0,0
+	cl	%r0,164(%r15)
+	je	.Lcbc_decrypt
+
+	larl	%r12,AES_Te
+
+	llgf	%r8,0(%r6)
+	llgf	%r9,4(%r6)
+	llgf	%r10,8(%r6)
+	llgf	%r11,12(%r6)
+
+	lghi	%r0,16
+	slgr	%r3,%r0
+	brc	4,.Lcbc_enc_tail	# if borrow
+.Lcbc_enc_loop:
+	stmg	%r2,%r4,2*8(%r15)
+	x	%r8,0(%r2)
+	x	%r9,4(%r2)
+	x	%r10,8(%r2)
+	x	%r11,12(%r2)
+	lgr	%r4,%r5
+
+	bras	%r14,_s390x_AES_encrypt
+
+	lmg	%r2,%r5,2*8(%r15)
+	st	%r8,0(%r4)
+	st	%r9,4(%r4)
+	st	%r10,8(%r4)
+	st	%r11,12(%r4)
+
+	la	%r2,16(%r2)
+	la	%r4,16(%r4)
+	lghi	%r0,16
+	ltgr	%r3,%r3
+	jz	.Lcbc_enc_done
+	slgr	%r3,%r0
+	brc	4,.Lcbc_enc_tail	# if borrow
+	j	.Lcbc_enc_loop
+.align	16
+.Lcbc_enc_done:
+	lg	%r6,6*8(%r15)
+	st	%r8,0(%r6)
+	st	%r9,4(%r6)
+	st	%r10,8(%r6)
+	st	%r11,12(%r6)
+
+	lmg	%r7,%r14,7*8(%r15)
+	br	%r14
+
+.align	16
+.Lcbc_enc_tail:
+	aghi	%r3,15
+	lghi	%r0,0
+	stg	%r0,16*8(%r15)
+	stg	%r0,16*8+8(%r15)
+	bras	%r1,3f
+	mvc	16*8(1,%r15),0(%r2)
+3:	ex	%r3,0(%r1)
+	lghi	%r3,0
+	la	%r2,16*8(%r15)
+	j	.Lcbc_enc_loop
+
+.align	16
+.Lcbc_decrypt:
+	larl	%r12,AES_Td
+
+	lg	%r0,0(%r6)
+	lg	%r1,8(%r6)
+	stmg	%r0,%r1,16*8(%r15)
+
+.Lcbc_dec_loop:
+	stmg	%r2,%r4,2*8(%r15)
+	llgf	%r8,0(%r2)
+	llgf	%r9,4(%r2)
+	llgf	%r10,8(%r2)
+	llgf	%r11,12(%r2)
+	lgr	%r4,%r5
+
+	bras	%r14,_s390x_AES_decrypt
+
+	lmg	%r2,%r5,2*8(%r15)
+	sllg	%r8,%r8,32
+	sllg	%r10,%r10,32
+	lr	%r8,%r9
+	lr	%r10,%r11
+
+	lg	%r0,0(%r2)
+	lg	%r1,8(%r2)
+	xg	%r8,16*8(%r15)
+	xg	%r10,16*8+8(%r15)
+	lghi	%r9,16
+	slgr	%r3,%r9
+	brc	4,.Lcbc_dec_tail	# if borrow
+	brc	2,.Lcbc_dec_done	# if zero
+	stg	%r8,0(%r4)
+	stg	%r10,8(%r4)
+	stmg	%r0,%r1,16*8(%r15)
+
+	la	%r2,16(%r2)
+	la	%r4,16(%r4)
+	j	.Lcbc_dec_loop
+
+.Lcbc_dec_done:
+	stg	%r8,0(%r4)
+	stg	%r10,8(%r4)
+.Lcbc_dec_exit:
+	lmg	%r6,%r14,6*8(%r15)
+	stmg	%r0,%r1,0(%r6)
+
+	br	%r14
+
+.align	16
+.Lcbc_dec_tail:
+	aghi	%r3,15
+	stg	%r8,16*8(%r15)
+	stg	%r10,16*8+8(%r15)
+	bras	%r9,4f
+	mvc	0(1,%r4),16*8(%r15)
+4:	ex	%r3,0(%r9)
+	j	.Lcbc_dec_exit
+.size	AES_cbc_encrypt,.-AES_cbc_encrypt
+.globl	AES_ctr32_encrypt
+.type	AES_ctr32_encrypt,@function
+.align	16
+AES_ctr32_encrypt:
+	xgr	%r3,%r4		# flip %r3 and %r4, %r4 and %r3
+	xgr	%r4,%r3
+	xgr	%r3,%r4
+	llgfr	%r3,%r3	# safe in ctr32 subroutine even in 64-bit case
+	l	%r0,240(%r5)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lctr32_software
+
+	stg	%r10,10*8(%r15)
+	stg	%r11,11*8(%r15)
+
+	clr	%r3,%r1		# does work even in 64-bit mode
+	jle	.Lctr32_nokma		# kma is slower for <= 16 blocks
+
+	larl	%r1,OPENSSL_s390xcap_P
+	lr	%r10,%r0
+	llihh	%r11,0x8000
+	srlg	%r11,%r11,0(%r10)
+	ng	%r11,S390X_KMA(%r1)		# check kma capability vector
+	jz	.Lctr32_nokma
+
+	lghi	%r1,-160-112
+	lgr	%r11,%r15
+	la	%r15,0(%r1,%r15)			# prepare parameter block
+
+	lhi	%r1,0x0600
+	sllg	%r3,%r3,4
+	or	%r0,%r1				# set HS and LAAD flags
+
+	stg	%r11,0(%r15)			# backchain
+	la	%r1,160(%r15)
+
+	lmg	%r10,%r11,0(%r5)			# copy key
+	stg	%r10,160+80(%r15)
+	stg	%r11,160+88(%r15)
+	lmg	%r10,%r11,16(%r5)
+	stg	%r10,160+96(%r15)
+	stg	%r11,160+104(%r15)
+
+	lmg	%r10,%r11,0(%r6)			# copy iv
+	stg	%r10,160+64(%r15)
+	ahi	%r11,-1				# kma requires counter-1
+	stg	%r11,160+72(%r15)
+	st	%r11,160+12(%r15)		# copy counter
+
+	lghi	%r10,0				# no AAD
+	lghi	%r11,0
+
+	.long	0xb929a042	# kma %r4,%r10,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	stg	%r0,160+80(%r15)		# wipe key
+	stg	%r0,160+88(%r15)
+	stg	%r0,160+96(%r15)
+	stg	%r0,160+104(%r15)
+	la	%r15,160+112(%r15)
+
+	lmg	%r10,%r11,10*8(%r15)
+	br	%r14
+
+.align	16
+.Lctr32_nokma:
+	stmg	%r6,%r9,6*8(%r15)
+
+	slgr	%r4,%r2
+	la	%r1,0(%r5)	# %r1 is permanent copy of %r5
+	lg	%r5,0(%r6)	# load ivec
+	lg	%r6,8(%r6)
+
+	# prepare and allocate stack frame at the top of 4K page
+	# with 1K reserved for eventual signal handling
+	lghi	%r8,-1024-256-16# guarantee at least 256-bytes buffer
+	lghi	%r9,-4096
+	algr	%r8,%r15
+	lgr	%r7,%r15
+	ngr	%r8,%r9		# align at page boundary
+	slgr	%r7,%r8		# total buffer size
+	lgr	%r10,%r15
+	lghi	%r9,1024+16	# sl[g]fi is extended-immediate facility
+	slgr	%r7,%r9		# deduct reservation to get usable buffer size
+	# buffer size is at lest 256 and at most 3072+256-16
+
+	la	%r15,1024(%r8)	# alloca
+	srlg	%r7,%r7,4	# convert bytes to blocks, minimum 16
+	stg	%r10,0(%r15)	# back-chain
+	stg	%r7,8(%r15)
+
+	slgr	%r3,%r7
+	brc	1,.Lctr32_hw_switch	# not zero, no borrow
+	algr	%r7,%r3	# input is shorter than allocated buffer
+	lghi	%r3,0
+	stg	%r7,8(%r15)
+
+.Lctr32_hw_switch:
+.Lctr32_km_loop:
+	la	%r10,16(%r15)
+	lgr	%r11,%r7
+.Lctr32_km_prepare:
+	stg	%r5,0(%r10)
+	stg	%r6,8(%r10)
+	la	%r10,16(%r10)
+	ahi	%r6,1		# 32-bit increment, preserves upper half
+	brct	%r11,.Lctr32_km_prepare
+
+	la	%r8,16(%r15)	# inp
+	sllg	%r9,%r7,4	# len
+	la	%r10,16(%r15)	# out
+	.long	0xb92e00a8	# km %r10,%r8
+	brc	1,.-4		# pay attention to "partial completion"
+
+	la	%r10,16(%r15)
+	lgr	%r11,%r7
+	slgr	%r10,%r2
+.Lctr32_km_xor:
+	lg	%r8,0(%r2)
+	lg	%r9,8(%r2)
+	xg	%r8,0(%r10,%r2)
+	xg	%r9,8(%r10,%r2)
+	stg	%r8,0(%r4,%r2)
+	stg	%r9,8(%r4,%r2)
+	la	%r2,16(%r2)
+	brct	%r11,.Lctr32_km_xor
+
+	slgr	%r3,%r7
+	brc	1,.Lctr32_km_loop	# not zero, no borrow
+	algr	%r7,%r3
+	lghi	%r3,0
+	brc	4+1,.Lctr32_km_loop	# not zero
+
+	lg	%r8,0(%r15)
+	lg	%r9,8(%r15)
+	la	%r10,16(%r15)
+.Lctr32_km_zap:
+	stg	%r8,0(%r10)
+	stg	%r8,8(%r10)
+	la	%r10,16(%r10)
+	brct	%r9,.Lctr32_km_zap
+
+	la	%r15,0(%r8)
+	lmg	%r6,%r11,6*8(%r15)
+	br	%r14
+.align	16
+.Lctr32_software:
+	stmg	%r5,%r14,5*8(%r15)
+	slgr	%r2,%r4
+	larl	%r12,AES_Te
+	llgf	%r1,12(%r6)
+
+.Lctr32_loop:
+	stmg	%r2,%r4,2*8(%r15)
+	llgf	%r8,0(%r6)
+	llgf	%r9,4(%r6)
+	llgf	%r10,8(%r6)
+	lgr	%r11,%r1
+	st	%r1,16*8(%r15)
+	lgr	%r4,%r5
+
+	bras	%r14,_s390x_AES_encrypt
+
+	lmg	%r2,%r6,2*8(%r15)
+	llgf	%r1,16*8(%r15)
+	x	%r8,0(%r2,%r4)
+	x	%r9,4(%r2,%r4)
+	x	%r10,8(%r2,%r4)
+	x	%r11,12(%r2,%r4)
+	stm	%r8,%r11,0(%r4)
+
+	la	%r4,16(%r4)
+	ahi	%r1,1		# 32-bit increment
+	brct	%r3,.Lctr32_loop
+
+	lmg	%r6,%r14,6*8(%r15)
+	br	%r14
+.size	AES_ctr32_encrypt,.-AES_ctr32_encrypt
+.type	_s390x_xts_km,@function
+.align	16
+_s390x_xts_km:
+	llgfr	%r8,%r0			# put aside the function code
+	lghi	%r9,0x7f
+	nr	%r9,%r0
+	larl	%r1,OPENSSL_s390xcap_P
+	llihh	%r0,0x8000
+	srlg	%r0,%r0,32(%r9)		# check for 32+function code
+	ng	%r0,S390X_KM(%r1)	# check km capability vector
+	lgr	%r0,%r8			# restore the function code
+	la	%r1,0(%r5)		# restore %r5
+	jz	.Lxts_km_vanilla
+
+	lmg	%r6,%r7,144(%r15)	# put aside the tweak value
+	algr	%r4,%r2
+
+	oill	%r0,32			# switch to xts function code
+	aghi	%r9,-18			#
+	sllg	%r9,%r9,3		# (function code - 18)*8, 0 or 16
+	la	%r1,144-16(%r15)
+	slgr	%r1,%r9			# parameter block position
+	lmg	%r8,%r11,0(%r5)	# load 256 bits of key material,
+	stmg	%r8,%r11,0(%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	%r6,%r7,144(%r15)	# "re-seat" the tweak value
+
+	.long	0xb92e0042		# km %r4,%r2
+	brc	1,.-4			# pay attention to "partial completion"
+
+	lrvg	%r8,144+0(%r15)	# load the last tweak
+	lrvg	%r9,144+8(%r15)
+	stmg	%r0,%r3,144-32(%r15)	# wipe copy of the key
+
+	nill	%r0,0xffdf		# switch back to original function code
+	la	%r1,0(%r5)		# restore pointer to %r5
+	slgr	%r4,%r2
+
+	llgc	%r3,2*8-1(%r15)
+	nill	%r3,0x0f		# %r3%=16
+	br	%r14
+
+.align	16
+.Lxts_km_vanilla:
+	# prepare and allocate stack frame at the top of 4K page
+	# with 1K reserved for eventual signal handling
+	lghi	%r8,-1024-256-16# guarantee at least 256-bytes buffer
+	lghi	%r9,-4096
+	algr	%r8,%r15
+	lgr	%r7,%r15
+	ngr	%r8,%r9		# align at page boundary
+	slgr	%r7,%r8		# total buffer size
+	lgr	%r10,%r15
+	lghi	%r9,1024+16	# sl[g]fi is extended-immediate facility
+	slgr	%r7,%r9		# deduct reservation to get usable buffer size
+	# buffer size is at lest 256 and at most 3072+256-16
+
+	la	%r15,1024(%r8)	# alloca
+	nill	%r7,0xfff0	# round to 16*n
+	stg	%r10,0(%r15)	# back-chain
+	nill	%r3,0xfff0	# redundant
+	stg	%r7,8(%r15)
+
+	slgr	%r3,%r7
+	brc	1,.Lxts_km_go	# not zero, no borrow
+	algr	%r7,%r3	# input is shorter than allocated buffer
+	lghi	%r3,0
+	stg	%r7,8(%r15)
+
+.Lxts_km_go:
+	lrvg	%r8,144+0(%r10)	# load the tweak value in little-endian
+	lrvg	%r9,144+8(%r10)
+
+	la	%r10,16(%r15)		# vector of ascending tweak values
+	slgr	%r10,%r2
+	srlg	%r11,%r7,4
+	j	.Lxts_km_start
+
+.Lxts_km_loop:
+	la	%r10,16(%r15)
+	slgr	%r10,%r2
+	srlg	%r11,%r7,4
+.Lxts_km_prepare:
+	lghi	%r5,0x87
+	srag	%r6,%r9,63		# broadcast upper bit
+	ngr	%r5,%r6			# rem
+	algr	%r8,%r8
+	alcgr	%r9,%r9
+	xgr	%r8,%r5
+.Lxts_km_start:
+	lrvgr	%r5,%r8			# flip byte order
+	lrvgr	%r6,%r9
+	stg	%r5,0(%r10,%r2)
+	stg	%r6,8(%r10,%r2)
+	xg	%r5,0(%r2)
+	xg	%r6,8(%r2)
+	stg	%r5,0(%r4,%r2)
+	stg	%r6,8(%r4,%r2)
+	la	%r2,16(%r2)
+	brct	%r11,.Lxts_km_prepare
+
+	slgr	%r2,%r7		# rewind %r2
+	la	%r10,0(%r4,%r2)
+	lgr	%r11,%r7
+	.long	0xb92e00aa		# km %r10,%r10
+	brc	1,.-4			# pay attention to "partial completion"
+
+	la	%r10,16(%r15)
+	slgr	%r10,%r2
+	srlg	%r11,%r7,4
+.Lxts_km_xor:
+	lg	%r5,0(%r4,%r2)
+	lg	%r6,8(%r4,%r2)
+	xg	%r5,0(%r10,%r2)
+	xg	%r6,8(%r10,%r2)
+	stg	%r5,0(%r4,%r2)
+	stg	%r6,8(%r4,%r2)
+	la	%r2,16(%r2)
+	brct	%r11,.Lxts_km_xor
+
+	slgr	%r3,%r7
+	brc	1,.Lxts_km_loop		# not zero, no borrow
+	algr	%r7,%r3
+	lghi	%r3,0
+	brc	4+1,.Lxts_km_loop	# not zero
+
+	lg	%r5,0(%r15)		# back-chain
+	llgf	%r7,12(%r15)	# bytes used
+	la	%r6,16(%r15)
+	srlg	%r7,%r7,4
+.Lxts_km_zap:
+	stg	%r5,0(%r6)
+	stg	%r5,8(%r6)
+	la	%r6,16(%r6)
+	brct	%r7,.Lxts_km_zap
+
+	la	%r15,0(%r5)
+	llgc	%r3,2*8-1(%r5)
+	nill	%r3,0x0f		# %r3%=16
+	bzr	%r14
+
+	# generate one more tweak...
+	lghi	%r5,0x87
+	srag	%r6,%r9,63		# broadcast upper bit
+	ngr	%r5,%r6			# rem
+	algr	%r8,%r8
+	alcgr	%r9,%r9
+	xgr	%r8,%r5
+
+	ltr	%r3,%r3		# clear zero flag
+	br	%r14
+.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, %r4 and %r3
+	xgr	%r4,%r3
+	xgr	%r3,%r4
+	stg	%r3,1*8(%r15)	# save copy of %r3
+	srag	%r3,%r3,4		# 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,%r14			# abort if zero (i.e. less than 16)
+	llgf	%r0,240(%r6)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lxts_enc_software
+
+	stg	%r14,5*8(%r15)
+	stmg	%r6,%r11,6*8(%r15)
+
+	sllg	%r3,%r3,4		# %r3&=~15
+	slgr	%r4,%r2
+
+	# generate the tweak value
+	lg	%r11,160(%r15)	# pointer to iv
+	la	%r10,144(%r15)
+	lmg	%r8,%r9,0(%r11)
+	lghi	%r11,16
+	stmg	%r8,%r9,0(%r10)
+	la	%r1,0(%r6)		# %r6 is not needed anymore
+	.long	0xb92e00aa		# km %r10,%r10, generate the tweak
+	brc	1,.-4			# can this happen?
+
+	l	%r0,240(%r5)
+	la	%r1,0(%r5)		# %r5 is not needed anymore
+	bras	%r14,_s390x_xts_km
+	jz	.Lxts_enc_km_done
+
+	aghi	%r2,-16		# take one step back
+	la	%r7,0(%r4,%r2)	# put aside real %r4
+.Lxts_enc_km_steal:
+	llgc	%r5,16(%r2)
+	llgc	%r6,0(%r4,%r2)
+	stc	%r5,0(%r4,%r2)
+	stc	%r6,16(%r4,%r2)
+	la	%r2,1(%r2)
+	brct	%r3,.Lxts_enc_km_steal
+
+	la	%r10,0(%r7)
+	lghi	%r11,16
+	lrvgr	%r5,%r8			# flip byte order
+	lrvgr	%r6,%r9
+	xg	%r5,0(%r10)
+	xg	%r6,8(%r10)
+	stg	%r5,0(%r10)
+	stg	%r6,8(%r10)
+	.long	0xb92e00aa		# km %r10,%r10
+	brc	1,.-4			# can this happen?
+	lrvgr	%r5,%r8			# flip byte order
+	lrvgr	%r6,%r9
+	xg	%r5,0(%r7)
+	xg	%r6,8(%r7)
+	stg	%r5,0(%r7)
+	stg	%r6,8(%r7)
+
+.Lxts_enc_km_done:
+	stg	%r15,144+0(%r15)	# wipe tweak
+	stg	%r15,144+8(%r15)
+	lg	%r14,5*8(%r15)
+	lmg	%r6,%r11,6*8(%r15)
+	br	%r14
+.align	16
+.Lxts_enc_software:
+	stmg	%r6,%r14,6*8(%r15)
+
+	slgr	%r4,%r2
+
+	lg	%r11,160(%r15)	# ivp
+	llgf	%r8,0(%r11)		# load iv
+	llgf	%r9,4(%r11)
+	llgf	%r10,8(%r11)
+	llgf	%r11,12(%r11)
+	stmg	%r2,%r5,2*8(%r15)
+	la	%r4,0(%r6)
+	larl	%r12,AES_Te
+	bras	%r14,_s390x_AES_encrypt	# generate the tweak
+	lmg	%r2,%r5,2*8(%r15)
+	stm	%r8,%r11,144(%r15)	# save the tweak
+	j	.Lxts_enc_enter
+
+.align	16
+.Lxts_enc_loop:
+	lrvg	%r9,144+0(%r15)	# load the tweak in little-endian
+	lrvg	%r11,144+8(%r15)
+	lghi	%r1,0x87
+	srag	%r0,%r11,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	algr	%r9,%r9
+	alcgr	%r11,%r11
+	xgr	%r9,%r1
+	lrvgr	%r9,%r9			# flip byte order
+	lrvgr	%r11,%r11
+	srlg	%r8,%r9,32		# smash the tweak to 4x32-bits
+	stg	%r9,144+0(%r15)	# save the tweak
+	llgfr	%r9,%r9
+	srlg	%r10,%r11,32
+	stg	%r11,144+8(%r15)
+	llgfr	%r11,%r11
+	la	%r2,16(%r2)		# %r2+=16
+.Lxts_enc_enter:
+	x	%r8,0(%r2)		# ^=*(%r2)
+	x	%r9,4(%r2)
+	x	%r10,8(%r2)
+	x	%r11,12(%r2)
+	stmg	%r2,%r3,2*8(%r15)	# only two registers are changing
+	la	%r4,0(%r5)
+	bras	%r14,_s390x_AES_encrypt
+	lmg	%r2,%r5,2*8(%r15)
+	x	%r8,144+0(%r15)	# ^=tweak
+	x	%r9,144+4(%r15)
+	x	%r10,144+8(%r15)
+	x	%r11,144+12(%r15)
+	st	%r8,0(%r4,%r2)
+	st	%r9,4(%r4,%r2)
+	st	%r10,8(%r4,%r2)
+	st	%r11,12(%r4,%r2)
+	brctg	%r3,.Lxts_enc_loop
+
+	llgc	%r3,15(%r15)
+	nill	%r3,0x0f		# %r3%16
+	jz	.Lxts_enc_done
+
+	la	%r7,0(%r2,%r4)	# put aside real %r4
+.Lxts_enc_steal:
+	llgc	%r0,16(%r2)
+	llgc	%r1,0(%r4,%r2)
+	stc	%r0,0(%r4,%r2)
+	stc	%r1,16(%r4,%r2)
+	la	%r2,1(%r2)
+	brct	%r3,.Lxts_enc_steal
+	la	%r4,0(%r7)		# restore real %r4
+
+	# generate last tweak...
+	lrvg	%r9,144+0(%r15)	# load the tweak in little-endian
+	lrvg	%r11,144+8(%r15)
+	lghi	%r1,0x87
+	srag	%r0,%r11,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	algr	%r9,%r9
+	alcgr	%r11,%r11
+	xgr	%r9,%r1
+	lrvgr	%r9,%r9			# flip byte order
+	lrvgr	%r11,%r11
+	srlg	%r8,%r9,32		# smash the tweak to 4x32-bits
+	stg	%r9,144+0(%r15)	# save the tweak
+	llgfr	%r9,%r9
+	srlg	%r10,%r11,32
+	stg	%r11,144+8(%r15)
+	llgfr	%r11,%r11
+
+	x	%r8,0(%r4)		# ^=*(inp)|stolen cipther-text
+	x	%r9,4(%r4)
+	x	%r10,8(%r4)
+	x	%r11,12(%r4)
+	stg	%r4,4*8(%r15)
+	la	%r4,0(%r5)
+	bras	%r14,_s390x_AES_encrypt
+	lg	%r4,4*8(%r15)
+	x	%r8,144(%r15)	# ^=tweak
+	x	%r9,148(%r15)
+	x	%r10,152(%r15)
+	x	%r11,156(%r15)
+	st	%r8,0(%r4)
+	st	%r9,4(%r4)
+	st	%r10,8(%r4)
+	st	%r11,12(%r4)
+
+.Lxts_enc_done:
+	stg	%r15,144+0(%r15)	# wipe tweak
+	stg	%r15,+8(%r15)
+	lmg	%r6,%r14,6*8(%r15)
+	br	%r14
+.size	AES_xts_encrypt,.-AES_xts_encrypt
+.globl	AES_xts_decrypt
+.type	AES_xts_decrypt,@function
+.align	16
+AES_xts_decrypt:
+	xgr	%r3,%r4			# flip %r3 and %r4, %r4 and %r3
+	xgr	%r4,%r3
+	xgr	%r3,%r4
+	stg	%r3,1*8(%r15)	# save copy of %r3
+	aghi	%r3,-16
+	bcr	4,%r14			# abort if less than zero. formally
+					# wrong, because %r3 is unsigned,
+					# but who can afford asking to
+					# process more than 2^63-1 bytes?
+	tmll	%r3,0x0f
+	jnz	.Lxts_dec_proceed
+	aghi	%r3,16
+.Lxts_dec_proceed:
+	llgf	%r0,240(%r6)
+	lhi	%r1,16
+	clr	%r0,%r1
+	jl	.Lxts_dec_software
+
+	stg	%r14,5*8(%r15)
+	stmg	%r6,%r11,6*8(%r15)
+
+	nill	%r3,0xfff0		# %r3&=~15
+	slgr	%r4,%r2
+
+	# generate the tweak value
+	lg	%r11,160(%r15)	# pointer to iv
+	la	%r10,144(%r15)
+	lmg	%r8,%r9,0(%r11)
+	lghi	%r11,16
+	stmg	%r8,%r9,0(%r10)
+	la	%r1,0(%r6)		# %r6 is not needed past this point
+	.long	0xb92e00aa		# km %r10,%r10, generate the tweak
+	brc	1,.-4			# can this happen?
+
+	l	%r0,240(%r5)
+	la	%r1,0(%r5)		# %r5 is not needed anymore
+
+	ltgr	%r3,%r3
+	jz	.Lxts_dec_km_short
+	bras	%r14,_s390x_xts_km
+	jz	.Lxts_dec_km_done
+
+	lrvgr	%r10,%r8			# make copy in reverse byte order
+	lrvgr	%r11,%r9
+	j	.Lxts_dec_km_2ndtweak
+
+.Lxts_dec_km_short:
+	llgc	%r3,15(%r15)
+	nill	%r3,0x0f		# %r3%=16
+	lrvg	%r8,144+0(%r15)	# load the tweak
+	lrvg	%r9,144+8(%r15)
+	lrvgr	%r10,%r8			# make copy in reverse byte order
+	lrvgr	%r11,%r9
+
+.Lxts_dec_km_2ndtweak:
+	lghi	%r5,0x87
+	srag	%r6,%r9,63		# broadcast upper bit
+	ngr	%r5,%r6			# rem
+	algr	%r8,%r8
+	alcgr	%r9,%r9
+	xgr	%r8,%r5
+	lrvgr	%r5,%r8			# flip byte order
+	lrvgr	%r6,%r9
+
+	xg	%r5,0(%r2)
+	xg	%r6,8(%r2)
+	stg	%r5,0(%r4,%r2)
+	stg	%r6,8(%r4,%r2)
+	la	%r6,0(%r4,%r2)
+	lghi	%r7,16
+	.long	0xb92e0066		# km %r6,%r6
+	brc	1,.-4			# can this happen?
+	lrvgr	%r5,%r8
+	lrvgr	%r6,%r9
+	xg	%r5,0(%r4,%r2)
+	xg	%r6,8(%r4,%r2)
+	stg	%r5,0(%r4,%r2)
+	stg	%r6,8(%r4,%r2)
+
+	la	%r7,0(%r4,%r2)	# put aside real %r4
+.Lxts_dec_km_steal:
+	llgc	%r5,16(%r2)
+	llgc	%r6,0(%r4,%r2)
+	stc	%r5,0(%r4,%r2)
+	stc	%r6,16(%r4,%r2)
+	la	%r2,1(%r2)
+	brct	%r3,.Lxts_dec_km_steal
+
+	lgr	%r8,%r10
+	lgr	%r9,%r11
+	xg	%r8,0(%r7)
+	xg	%r9,8(%r7)
+	stg	%r8,0(%r7)
+	stg	%r9,8(%r7)
+	la	%r8,0(%r7)
+	lghi	%r9,16
+	.long	0xb92e0088		# km %r8,%r8
+	brc	1,.-4			# can this happen?
+	xg	%r10,0(%r7)
+	xg	%r11,8(%r7)
+	stg	%r10,0(%r7)
+	stg	%r11,8(%r7)
+.Lxts_dec_km_done:
+	stg	%r15,144+0(%r15)	# wipe tweak
+	stg	%r15,144+8(%r15)
+	lg	%r14,5*8(%r15)
+	lmg	%r6,%r11,6*8(%r15)
+	br	%r14
+.align	16
+.Lxts_dec_software:
+	stmg	%r6,%r14,6*8(%r15)
+
+	srlg	%r3,%r3,4
+	slgr	%r4,%r2
+
+	lg	%r11,160(%r15)	# ivp
+	llgf	%r8,0(%r11)		# load iv
+	llgf	%r9,4(%r11)
+	llgf	%r10,8(%r11)
+	llgf	%r11,12(%r11)
+	stmg	%r2,%r5,2*8(%r15)
+	la	%r4,0(%r6)
+	larl	%r12,AES_Te
+	bras	%r14,_s390x_AES_encrypt	# generate the tweak
+	lmg	%r2,%r5,2*8(%r15)
+	larl	%r12,AES_Td
+	ltgr	%r3,%r3
+	stm	%r8,%r11,144(%r15)	# save the tweak
+	jz	.Lxts_dec_short
+	j	.Lxts_dec_enter
+
+.align	16
+.Lxts_dec_loop:
+	lrvg	%r9,144+0(%r15)	# load the tweak in little-endian
+	lrvg	%r11,144+8(%r15)
+	lghi	%r1,0x87
+	srag	%r0,%r11,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	algr	%r9,%r9
+	alcgr	%r11,%r11
+	xgr	%r9,%r1
+	lrvgr	%r9,%r9			# flip byte order
+	lrvgr	%r11,%r11
+	srlg	%r8,%r9,32		# smash the tweak to 4x32-bits
+	stg	%r9,144+0(%r15)	# save the tweak
+	llgfr	%r9,%r9
+	srlg	%r10,%r11,32
+	stg	%r11,144+8(%r15)
+	llgfr	%r11,%r11
+.Lxts_dec_enter:
+	x	%r8,0(%r2)		# tweak^=*(inp)
+	x	%r9,4(%r2)
+	x	%r10,8(%r2)
+	x	%r11,12(%r2)
+	stmg	%r2,%r3,2*8(%r15)	# only two registers are changing
+	la	%r4,0(%r5)
+	bras	%r14,_s390x_AES_decrypt
+	lmg	%r2,%r5,2*8(%r15)
+	x	%r8,144+0(%r15)	# ^=tweak
+	x	%r9,144+4(%r15)
+	x	%r10,144+8(%r15)
+	x	%r11,144+12(%r15)
+	st	%r8,0(%r4,%r2)
+	st	%r9,4(%r4,%r2)
+	st	%r10,8(%r4,%r2)
+	st	%r11,12(%r4,%r2)
+	la	%r2,16(%r2)
+	brctg	%r3,.Lxts_dec_loop
+
+	llgc	%r3,15(%r15)
+	nill	%r3,0x0f		# %r3%16
+	jz	.Lxts_dec_done
+
+	# generate pair of tweaks...
+	lrvg	%r9,144+0(%r15)	# load the tweak in little-endian
+	lrvg	%r11,144+8(%r15)
+	lghi	%r1,0x87
+	srag	%r0,%r11,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	algr	%r9,%r9
+	alcgr	%r11,%r11
+	xgr	%r9,%r1
+	lrvgr	%r6,%r9			# flip byte order
+	lrvgr	%r7,%r11
+	stmg	%r6,%r7,144(%r15)	# save the 1st tweak
+	j	.Lxts_dec_2ndtweak
+
+.align	16
+.Lxts_dec_short:
+	llgc	%r3,15(%r15)
+	nill	%r3,0x0f		# %r3%16
+	lrvg	%r9,144+0(%r15)	# load the tweak in little-endian
+	lrvg	%r11,144+8(%r15)
+.Lxts_dec_2ndtweak:
+	lghi	%r1,0x87
+	srag	%r0,%r11,63		# broadcast upper bit
+	ngr	%r1,%r0			# rem
+	algr	%r9,%r9
+	alcgr	%r11,%r11
+	xgr	%r9,%r1
+	lrvgr	%r9,%r9			# flip byte order
+	lrvgr	%r11,%r11
+	srlg	%r8,%r9,32		# smash the tweak to 4x32-bits
+	stg	%r9,144-16+0(%r15)	# save the 2nd tweak
+	llgfr	%r9,%r9
+	srlg	%r10,%r11,32
+	stg	%r11,144-16+8(%r15)
+	llgfr	%r11,%r11
+
+	x	%r8,0(%r2)		# tweak_the_2nd^=*(inp)
+	x	%r9,4(%r2)
+	x	%r10,8(%r2)
+	x	%r11,12(%r2)
+	stmg	%r2,%r3,2*8(%r15)
+	la	%r4,0(%r5)
+	bras	%r14,_s390x_AES_decrypt
+	lmg	%r2,%r5,2*8(%r15)
+	x	%r8,144-16+0(%r15)	# ^=tweak_the_2nd
+	x	%r9,144-16+4(%r15)
+	x	%r10,144-16+8(%r15)
+	x	%r11,144-16+12(%r15)
+	st	%r8,0(%r4,%r2)
+	st	%r9,4(%r4,%r2)
+	st	%r10,8(%r4,%r2)
+	st	%r11,12(%r4,%r2)
+
+	la	%r7,0(%r4,%r2)	# put aside real %r4
+.Lxts_dec_steal:
+	llgc	%r0,16(%r2)
+	llgc	%r1,0(%r4,%r2)
+	stc	%r0,0(%r4,%r2)
+	stc	%r1,16(%r4,%r2)
+	la	%r2,1(%r2)
+	brct	%r3,.Lxts_dec_steal
+	la	%r4,0(%r7)		# restore real %r4
+
+	lm	%r8,%r11,144(%r15)	# load the 1st tweak
+	x	%r8,0(%r4)		# tweak^=*(inp)|stolen cipher-text
+	x	%r9,4(%r4)
+	x	%r10,8(%r4)
+	x	%r11,12(%r4)
+	stg	%r4,4*8(%r15)
+	la	%r4,0(%r5)
+	bras	%r14,_s390x_AES_decrypt
+	lg	%r4,4*8(%r15)
+	x	%r8,144+0(%r15)	# ^=tweak
+	x	%r9,144+4(%r15)
+	x	%r10,144+8(%r15)
+	x	%r11,144+12(%r15)
+	st	%r8,0(%r4)
+	st	%r9,4(%r4)
+	st	%r10,8(%r4)
+	st	%r11,12(%r4)
+	stg	%r15,144-16+0(%r15)	# wipe 2nd tweak
+	stg	%r15,144-16+8(%r15)
+.Lxts_dec_done:
+	stg	%r15,144+0(%r15)	# wipe tweak
+	stg	%r15,+8(%r15)
+	lmg	%r6,%r14,6*8(%r15)
+	br	%r14
+.size	AES_xts_decrypt,.-AES_xts_decrypt
+.string	"AES for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/bn/s390x-gf2m.s b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/bn/s390x-gf2m.s
new file mode 100644
index 0000000000..e0b0822cae
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/bn/s390x-gf2m.s
@@ -0,0 +1,217 @@
+.text
+
+.type	_mul_1x1,@function
+.align	16
+_mul_1x1:
+	lgr	%r6,%r3
+	sllg	%r7,%r3,1
+	sllg	%r8,%r3,2
+	sllg	%r9,%r3,3
+
+	srag	%r3,%r6,63			# broadcast 63rd bit
+	nihh	%r6,0x1fff
+	srag	%r12,%r7,63			# broadcast 62nd bit
+	nihh	%r7,0x3fff
+	srag	%r13,%r8,63			# broadcast 61st bit
+	nihh	%r8,0x7fff
+	ngr	%r3,%r5
+	ngr	%r12,%r5
+	ngr	%r13,%r5
+
+	lghi	%r0,0
+	lgr	%r10,%r6
+	stg	%r0,160(%r15)	# tab[0]=0
+	xgr	%r10,%r7
+	stg	%r6,168(%r15)	# tab[1]=a1
+	 lgr	%r11,%r8
+	stg	%r7,176(%r15)	# tab[2]=a2
+	 xgr	%r11,%r9
+	stg	%r10,184(%r15)	# tab[3]=a1^a2
+	 xgr	%r6,%r8
+
+	stg	%r8,192(%r15)	# tab[4]=a4
+	xgr	%r7,%r8
+	stg	%r6,200(%r15)	# tab[5]=a1^a4
+	xgr	%r10,%r8
+	stg	%r7,208(%r15)	# tab[6]=a2^a4
+	 xgr	%r6,%r11
+	stg	%r10,216(%r15)	# tab[7]=a1^a2^a4
+	 xgr	%r7,%r11
+
+	stg	%r9,224(%r15)	# tab[8]=a8
+	xgr	%r10,%r11
+	stg	%r6,232(%r15)	# tab[9]=a1^a8
+	 xgr	%r6,%r8
+	stg	%r7,240(%r15)	# tab[10]=a2^a8
+	 xgr	%r7,%r8
+	stg	%r10,248(%r15)	# tab[11]=a1^a2^a8
+
+	xgr	%r10,%r8
+	stg	%r11,256(%r15)	# tab[12]=a4^a8
+	 srlg	%r4,%r3,1
+	stg	%r6,264(%r15)	# tab[13]=a1^a4^a8
+	 sllg	%r3,%r3,63
+	stg	%r7,272(%r15)	# tab[14]=a2^a4^a8
+	 srlg	%r0,%r12,2
+	stg	%r10,280(%r15)	# tab[15]=a1^a2^a4^a8
+
+	lghi	%r9,120
+	sllg	%r6,%r12,62
+	 sllg	%r12,%r5,3
+	srlg	%r1,%r13,3
+	 ngr	%r12,%r9
+	sllg	%r7,%r13,61
+	 srlg	%r13,%r5,4-3
+	xgr	%r4,%r0
+	 ngr	%r13,%r9
+	xgr	%r3,%r6
+	xgr	%r4,%r1
+	xgr	%r3,%r7
+
+	xg	%r3,160(%r12,%r15)
+	srlg	%r12,%r5,8-3
+	ngr	%r12,%r9
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,12-3
+	sllg	%r0,%r1,4
+	ngr	%r13,%r9
+	srlg	%r1,%r1,60
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	srlg	%r12,%r5,16-3
+	sllg	%r1,%r0,8
+	ngr	%r12,%r9
+	srlg	%r0,%r0,56
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,20-3
+	sllg	%r0,%r1,12
+	ngr	%r13,%r9
+	srlg	%r1,%r1,52
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	srlg	%r12,%r5,24-3
+	sllg	%r1,%r0,16
+	ngr	%r12,%r9
+	srlg	%r0,%r0,48
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,28-3
+	sllg	%r0,%r1,20
+	ngr	%r13,%r9
+	srlg	%r1,%r1,44
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	srlg	%r12,%r5,32-3
+	sllg	%r1,%r0,24
+	ngr	%r12,%r9
+	srlg	%r0,%r0,40
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,36-3
+	sllg	%r0,%r1,28
+	ngr	%r13,%r9
+	srlg	%r1,%r1,36
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	srlg	%r12,%r5,40-3
+	sllg	%r1,%r0,32
+	ngr	%r12,%r9
+	srlg	%r0,%r0,32
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,44-3
+	sllg	%r0,%r1,36
+	ngr	%r13,%r9
+	srlg	%r1,%r1,28
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	srlg	%r12,%r5,48-3
+	sllg	%r1,%r0,40
+	ngr	%r12,%r9
+	srlg	%r0,%r0,24
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,52-3
+	sllg	%r0,%r1,44
+	ngr	%r13,%r9
+	srlg	%r1,%r1,20
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	srlg	%r12,%r5,56-3
+	sllg	%r1,%r0,48
+	ngr	%r12,%r9
+	srlg	%r0,%r0,16
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+	lg	%r1,160(%r13,%r15)
+	srlg	%r13,%r5,60-3
+	sllg	%r0,%r1,52
+	ngr	%r13,%r9
+	srlg	%r1,%r1,12
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+	lg	%r0,160(%r12,%r15)
+	sllg	%r1,%r0,56
+	srlg	%r0,%r0,8
+	xgr	%r3,%r1
+	xgr	%r4,%r0
+
+	lg	%r1,160(%r13,%r15)
+	sllg	%r0,%r1,60
+	srlg	%r1,%r1,4
+	xgr	%r3,%r0
+	xgr	%r4,%r1
+
+	br	%r14
+.size	_mul_1x1,.-_mul_1x1
+
+.globl	bn_GF2m_mul_2x2
+.type	bn_GF2m_mul_2x2,@function
+.align	16
+bn_GF2m_mul_2x2:
+	stmg	%r3,%r15,3*8(%r15)
+
+	lghi	%r1,-160-128
+	la	%r0,0(%r15)
+	la	%r15,0(%r1,%r15)			# alloca
+	stg	%r0,0(%r15)			# back chain
+	bras	%r14,_mul_1x1			# a1·b1
+	stmg	%r3,%r4,16(%r2)
+
+	lg	%r3,320(%r15)
+	lg	%r5,336(%r15)
+	bras	%r14,_mul_1x1			# a0·b0
+	stmg	%r3,%r4,0(%r2)
+
+	lg	%r3,312(%r15)
+	lg	%r5,328(%r15)
+	xg	%r3,320(%r15)
+	xg	%r5,336(%r15)
+	bras	%r14,_mul_1x1			# (a0+a1)·(b0+b1)
+	lmg	%r6,%r9,0(%r2)
+
+	xgr	%r3,%r4
+	xgr	%r4,%r7
+	xgr	%r3,%r6
+	xgr	%r4,%r8
+	xgr	%r3,%r9
+	xgr	%r4,%r9
+	xgr	%r3,%r4
+	stg	%r4,16(%r2)
+	stg	%r3,8(%r2)
+	lmg	%r6,%r15,336(%r15)
+	br	%r14
+.size	bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2
+.string	"GF(2^m) Multiplication for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/bn/s390x-mont.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/bn/s390x-mont.S
new file mode 100644
index 0000000000..b8dea0a66f
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/bn/s390x-mont.S
@@ -0,0 +1,176 @@
+.text
+.globl	bn_mul_mont
+.type	bn_mul_mont,@function
+bn_mul_mont:
+	lgf	%r1,164(%r15)	# pull %r1
+	sla	%r1,3	# %r1 to enumerate bytes
+	la	%r4,0(%r1,%r4)
+
+	stg	%r2,2*8(%r15)
+
+	cghi	%r1,16		#
+	lghi	%r2,0		#
+	blr	%r14		# if(%r1<16) return 0;
+	cghi	%r1,96		#
+	bhr	%r14		# if(%r1>96) return 0;
+	stmg	%r3,%r15,3*8(%r15)
+
+	lghi	%r2,-160-8	# leave room for carry bit
+	lcgr	%r7,%r1		# -%r1
+	lgr	%r0,%r15
+	la	%r2,0(%r2,%r15)
+	la	%r15,0(%r7,%r2)	# alloca
+	stg	%r0,0(%r15)	# back chain
+
+	sra	%r1,3		# restore %r1
+	la	%r4,0(%r7,%r4)	# restore %r4
+	ahi	%r1,-1		# adjust %r1 for inner loop
+	lg	%r6,0(%r6)	# pull n0
+	
+
+	lg	%r2,0(%r4)
+	
+	lg	%r9,0(%r3)
+	
+	mlgr	%r8,%r2	# ap[0]*bp[0]
+	lgr	%r12,%r8
+
+	lgr	%r0,%r9	# "tp[0]"*n0
+	msgr	%r0,%r6
+
+	lg	%r11,0(%r5)	#
+	
+	mlgr	%r10,%r0	# np[0]*m1
+	algr	%r11,%r9	# +="tp[0]"
+	lghi	%r13,0
+	alcgr	%r13,%r10
+
+	la	%r7,8(%r0)	# j=1
+	lr	%r14,%r1
+
+.align	16
+.L1st:
+	lg	%r9,0(%r7,%r3)
+	
+	mlgr	%r8,%r2	# ap[j]*bp[0]
+	algr	%r9,%r12
+	lghi	%r12,0
+	alcgr	%r12,%r8
+
+	lg	%r11,0(%r7,%r5)
+	
+	mlgr	%r10,%r0	# np[j]*m1
+	algr	%r11,%r13
+	lghi	%r13,0
+	alcgr	%r10,%r13	# +="tp[j]"
+	algr	%r11,%r9
+	alcgr	%r13,%r10
+
+	stg	%r11,160-8(%r7,%r15)	# tp[j-1]=
+	la	%r7,8(%r7)	# j++
+	brct	%r14,.L1st
+
+	algr	%r13,%r12
+	lghi	%r12,0
+	alcgr	%r12,%r12	# upmost overflow bit
+	stg	%r13,160-8(%r7,%r15)
+	stg	%r12,160(%r7,%r15)
+	la	%r4,8(%r4)	# bp++
+
+.Louter:
+	lg	%r2,0(%r4)	# bp[i]
+	
+	lg	%r9,0(%r3)
+	
+	mlgr	%r8,%r2	# ap[0]*bp[i]
+	alg	%r9,160(%r15)	# +=tp[0]
+	lghi	%r12,0
+	alcgr	%r12,%r8
+
+	lgr	%r0,%r9
+	msgr	%r0,%r6	# tp[0]*n0
+
+	lg	%r11,0(%r5)	# np[0]
+	
+	mlgr	%r10,%r0	# np[0]*m1
+	algr	%r11,%r9	# +="tp[0]"
+	lghi	%r13,0
+	alcgr	%r13,%r10
+
+	la	%r7,8(%r0)	# j=1
+	lr	%r14,%r1
+
+.align	16
+.Linner:
+	lg	%r9,0(%r7,%r3)
+	
+	mlgr	%r8,%r2	# ap[j]*bp[i]
+	algr	%r9,%r12
+	lghi	%r12,0
+	alcgr	%r8,%r12
+	alg	%r9,160(%r7,%r15)# +=tp[j]
+	alcgr	%r12,%r8
+
+	lg	%r11,0(%r7,%r5)
+	
+	mlgr	%r10,%r0	# np[j]*m1
+	algr	%r11,%r13
+	lghi	%r13,0
+	alcgr	%r10,%r13
+	algr	%r11,%r9	# +="tp[j]"
+	alcgr	%r13,%r10
+
+	stg	%r11,160-8(%r7,%r15)	# tp[j-1]=
+	la	%r7,8(%r7)	# j++
+	brct	%r14,.Linner
+
+	algr	%r13,%r12
+	lghi	%r12,0
+	alcgr	%r12,%r12
+	alg	%r13,160(%r7,%r15)# accumulate previous upmost overflow bit
+	lghi	%r8,0
+	alcgr	%r12,%r8	# new upmost overflow bit
+	stg	%r13,160-8(%r7,%r15)
+	stg	%r12,160(%r7,%r15)
+
+	la	%r4,8(%r4)	# bp++
+	clg	%r4,200(%r7,%r15)	# compare to &bp[num]
+	jne	.Louter
+
+	lg	%r2,184(%r7,%r15)	# reincarnate rp
+	la	%r3,160(%r15)
+	ahi	%r1,1		# restore %r1, incidentally clears "borrow"
+
+	la	%r7,0(%r0)
+	lr	%r14,%r1
+.Lsub:	lg	%r9,0(%r7,%r3)
+	lg	%r11,0(%r7,%r5)
+	
+	slbgr	%r9,%r11
+	stg	%r9,0(%r7,%r2)
+	la	%r7,8(%r7)
+	brct	%r14,.Lsub
+	lghi	%r8,0
+	slbgr	%r12,%r8	# handle upmost carry
+	lghi	%r13,-1
+	xgr	%r13,%r12
+
+	la	%r7,0(%r0)
+	lgr	%r14,%r1
+.Lcopy:	lg	%r8,160(%r7,%r15)	# conditional copy
+	lg	%r9,0(%r7,%r2)
+	ngr	%r8,%r12
+	ngr	%r9,%r13
+	ogr	%r9,%r8
+	
+	stg	%r7,160(%r7,%r15)	# zap tp
+	stg	%r9,0(%r7,%r2)
+	la	%r7,8(%r7)
+	brct	%r14,.Lcopy
+
+	la	%r1,216(%r7,%r15)
+	lmg	%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>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/buildinf.h b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/buildinf.h
new file mode 100644
index 0000000000..59a48acd96
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/buildinf.h
@@ -0,0 +1,43 @@
+/*
+ * WARNING: do not edit!
+ * Generated by util/mkbuildinf.pl
+ *
+ * Copyright 2014-2017 The OpenSSL Project Authors. All Rights Reserved.
+ *
+ * Licensed under the OpenSSL license (the "License").  You may not use
+ * this file except in compliance with the License.  You can obtain a copy
+ * in the file LICENSE in the source distribution or at
+ * https://www.openssl.org/source/license.html
+ */
+
+#define PLATFORM "platform: linux64-s390x"
+#define DATE "built on: Thu Nov 22 19:35:23 2018 UTC"
+
+/*
+ * Generate compiler_flags as an array of individual characters. This is a
+ * workaround for the situation where CFLAGS gets too long for a C90 string
+ * literal
+ */
+static const char compiler_flags[] = {
+    'c','o','m','p','i','l','e','r',':',' ','.','.','/','c','o','n',
+    'f','i','g','/','f','a','k','e','_','g','c','c','.','p','l',' ',
+    '-','f','P','I','C',' ','-','p','t','h','r','e','a','d',' ','-',
+    'm','6','4',' ','-','W','a',',','-','-','n','o','e','x','e','c',
+    's','t','a','c','k',' ','-','W','a','l','l',' ','-','O','3',' ',
+    '-','D','O','P','E','N','S','S','L','_','U','S','E','_','N','O',
+    'D','E','L','E','T','E',' ','-','D','B','_','E','N','D','I','A',
+    'N',' ','-','D','O','P','E','N','S','S','L','_','P','I','C',' ',
+    '-','D','O','P','E','N','S','S','L','_','C','P','U','I','D','_',
+    'O','B','J',' ','-','D','O','P','E','N','S','S','L','_','B','N',
+    '_','A','S','M','_','M','O','N','T',' ','-','D','O','P','E','N',
+    'S','S','L','_','B','N','_','A','S','M','_','G','F','2','m',' ',
+    '-','D','S','H','A','1','_','A','S','M',' ','-','D','S','H','A',
+    '2','5','6','_','A','S','M',' ','-','D','S','H','A','5','1','2',
+    '_','A','S','M',' ','-','D','K','E','C','C','A','K','1','6','0',
+    '0','_','A','S','M',' ','-','D','R','C','4','_','A','S','M',' ',
+    '-','D','A','E','S','_','A','S','M',' ','-','D','A','E','S','_',
+    'C','T','R','_','A','S','M',' ','-','D','A','E','S','_','X','T',
+    'S','_','A','S','M',' ','-','D','G','H','A','S','H','_','A','S',
+    'M',' ','-','D','P','O','L','Y','1','3','0','5','_','A','S','M',
+    ' ','-','D','N','D','E','B','U','G','\0'
+};
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/chacha/chacha-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/chacha/chacha-s390x.S
new file mode 100644
index 0000000000..75cf92e487
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/chacha/chacha-s390x.S
@@ -0,0 +1,271 @@
+.text
+
+.globl	ChaCha20_ctr32
+.type	ChaCha20_ctr32,@function
+.align	32
+ChaCha20_ctr32:
+	ltgr	%r4,%r4			# %r4==0?
+	bzr	%r14
+	aghi	%r4,-64
+	lghi	%r1,-240
+	stmg	%r6,%r15,48(%r15)
+	slgr	%r2,%r3			# difference
+	la	%r4,0(%r3,%r4)		# end of input minus 64
+	larl	%r7,.Lsigma
+	lgr	%r0,%r15
+	la	%r15,0(%r1,%r15)
+	stg	%r0,0(%r15)
+
+	lmg	%r8,%r11,0(%r5)		# load key
+	lmg	%r12,%r13,0(%r6)		# load counter
+	lmg	%r6,%r7,0(%r7)			# load sigma constant
+
+	la	%r14,0(%r3)
+	stg	%r2,240+3*8(%r15)
+	stg	%r4,240+4*8(%r15)
+	stmg	%r6,%r13,160(%r15)		# copy key schedule to stack
+	srlg	%r10,%r12,32			# 32-bit counter value
+	j	.Loop_outer
+
+.align	16
+.Loop_outer:
+	lm	%r0,%r7,160+4*0(%r15)		# load x[0]-x[7]
+	lm	%r8,%r9,160+4*10(%r15)		# load x[10]-x[11]
+	lm	%r11,%r13,160+4*13(%r15)	# load x[13]-x[15]
+	stm	%r8,%r9,160+4*8+4*10(%r15)	# offload x[10]-x[11]
+	lm	%r8,%r9,160+4*8(%r15)		# load x[8]-x[9]
+	st	%r10,160+4*12(%r15)		# save counter
+	stg	%r14,240+2*8(%r15)		# save input pointer
+	lhi	%r14,10
+	j	.Loop
+
+.align	4
+.Loop:
+	alr	%r0,%r4
+	alr	%r1,%r5
+	xr	%r10,%r0
+	xr	%r11,%r1
+	rll	%r10,%r10,16
+	rll	%r11,%r11,16
+	alr	%r8,%r10
+	alr	%r9,%r11
+	xr	%r4,%r8
+	xr	%r5,%r9
+	rll	%r4,%r4,12
+	rll	%r5,%r5,12
+	alr	%r0,%r4
+	alr	%r1,%r5
+	xr	%r10,%r0
+	xr	%r11,%r1
+	rll	%r10,%r10,8
+	rll	%r11,%r11,8
+	alr	%r8,%r10
+	alr	%r9,%r11
+	xr	%r4,%r8
+	xr	%r5,%r9
+	rll	%r4,%r4,7
+	rll	%r5,%r5,7
+	stm	%r8,%r9,160+4*8+4*8(%r15)
+	lm	%r8,%r9,160+4*8+4*10(%r15)
+	alr	%r2,%r6
+	alr	%r3,%r7
+	xr	%r12,%r2
+	xr	%r13,%r3
+	rll	%r12,%r12,16
+	rll	%r13,%r13,16
+	alr	%r8,%r12
+	alr	%r9,%r13
+	xr	%r6,%r8
+	xr	%r7,%r9
+	rll	%r6,%r6,12
+	rll	%r7,%r7,12
+	alr	%r2,%r6
+	alr	%r3,%r7
+	xr	%r12,%r2
+	xr	%r13,%r3
+	rll	%r12,%r12,8
+	rll	%r13,%r13,8
+	alr	%r8,%r12
+	alr	%r9,%r13
+	xr	%r6,%r8
+	xr	%r7,%r9
+	rll	%r6,%r6,7
+	rll	%r7,%r7,7
+	alr	%r0,%r5
+	alr	%r1,%r6
+	xr	%r13,%r0
+	xr	%r10,%r1
+	rll	%r13,%r13,16
+	rll	%r10,%r10,16
+	alr	%r8,%r13
+	alr	%r9,%r10
+	xr	%r5,%r8
+	xr	%r6,%r9
+	rll	%r5,%r5,12
+	rll	%r6,%r6,12
+	alr	%r0,%r5
+	alr	%r1,%r6
+	xr	%r13,%r0
+	xr	%r10,%r1
+	rll	%r13,%r13,8
+	rll	%r10,%r10,8
+	alr	%r8,%r13
+	alr	%r9,%r10
+	xr	%r5,%r8
+	xr	%r6,%r9
+	rll	%r5,%r5,7
+	rll	%r6,%r6,7
+	stm	%r8,%r9,160+4*8+4*10(%r15)
+	lm	%r8,%r9,160+4*8+4*8(%r15)
+	alr	%r2,%r7
+	alr	%r3,%r4
+	xr	%r11,%r2
+	xr	%r12,%r3
+	rll	%r11,%r11,16
+	rll	%r12,%r12,16
+	alr	%r8,%r11
+	alr	%r9,%r12
+	xr	%r7,%r8
+	xr	%r4,%r9
+	rll	%r7,%r7,12
+	rll	%r4,%r4,12
+	alr	%r2,%r7
+	alr	%r3,%r4
+	xr	%r11,%r2
+	xr	%r12,%r3
+	rll	%r11,%r11,8
+	rll	%r12,%r12,8
+	alr	%r8,%r11
+	alr	%r9,%r12
+	xr	%r7,%r8
+	xr	%r4,%r9
+	rll	%r7,%r7,7
+	rll	%r4,%r4,7
+	brct	%r14,.Loop
+
+	lg	%r14,240+2*8(%r15)		# pull input pointer
+	stm	%r8,%r9,160+4*8+4*8(%r15)	# offload x[8]-x[9]
+	lmg	%r8,%r9,240+3*8(%r15)
+
+	al	%r0,160+4*0(%r15)	# accumulate key schedule
+	al	%r1,160+4*1(%r15)
+	al	%r2,160+4*2(%r15)
+	al	%r3,160+4*3(%r15)
+	al	%r4,160+4*4(%r15)
+	al	%r5,160+4*5(%r15)
+	al	%r6,160+4*6(%r15)
+	al	%r7,160+4*7(%r15)
+	lrvr	%r0,%r0
+	lrvr	%r1,%r1
+	lrvr	%r2,%r2
+	lrvr	%r3,%r3
+	lrvr	%r4,%r4
+	lrvr	%r5,%r5
+	lrvr	%r6,%r6
+	lrvr	%r7,%r7
+	al	%r10,160+4*12(%r15)
+	al	%r11,160+4*13(%r15)
+	al	%r12,160+4*14(%r15)
+	al	%r13,160+4*15(%r15)
+	lrvr	%r10,%r10
+	lrvr	%r11,%r11
+	lrvr	%r12,%r12
+	lrvr	%r13,%r13
+
+	la	%r8,0(%r8,%r14)		# reconstruct output pointer
+	clgr	%r14,%r9
+	jh	.Ltail
+
+	x	%r0,4*0(%r14)			# xor with input
+	x	%r1,4*1(%r14)
+	st	%r0,4*0(%r8)		# store output
+	x	%r2,4*2(%r14)
+	st	%r1,4*1(%r8)
+	x	%r3,4*3(%r14)
+	st	%r2,4*2(%r8)
+	x	%r4,4*4(%r14)
+	st	%r3,4*3(%r8)
+	 lm	%r0,%r3,160+4*8+4*8(%r15)	# load x[8]-x[11]
+	x	%r5,4*5(%r14)
+	st	%r4,4*4(%r8)
+	x	%r6,4*6(%r14)
+	 al	%r0,160+4*8(%r15)
+	st	%r5,4*5(%r8)
+	x	%r7,4*7(%r14)
+	 al	%r1,160+4*9(%r15)
+	st	%r6,4*6(%r8)
+	x	%r10,4*12(%r14)
+	 al	%r2,160+4*10(%r15)
+	st	%r7,4*7(%r8)
+	x	%r11,4*13(%r14)
+	 al	%r3,160+4*11(%r15)
+	st	%r10,4*12(%r8)
+	x	%r12,4*14(%r14)
+	st	%r11,4*13(%r8)
+	x	%r13,4*15(%r14)
+	st	%r12,4*14(%r8)
+	 lrvr	%r0,%r0
+	st	%r13,4*15(%r8)
+	 lrvr	%r1,%r1
+	 lrvr	%r2,%r2
+	 lrvr	%r3,%r3
+	lhi	%r10,1
+	 x	%r0,4*8(%r14)
+	al	%r10,160+4*12(%r15)	# increment counter
+	 x	%r1,4*9(%r14)
+	 st	%r0,4*8(%r8)
+	 x	%r2,4*10(%r14)
+	 st	%r1,4*9(%r8)
+	 x	%r3,4*11(%r14)
+	 st	%r2,4*10(%r8)
+	 st	%r3,4*11(%r8)
+
+	clgr	%r14,%r9			# done yet?
+	la	%r14,64(%r14)
+	jl	.Loop_outer
+
+.Ldone:
+	xgr	%r0,%r0
+	xgr	%r1,%r1
+	xgr	%r2,%r2
+	xgr	%r3,%r3
+	stmg	%r0,%r3,160+4*4(%r15)	# wipe key copy
+	stmg	%r0,%r3,160+4*12(%r15)
+
+	lmg	%r6,%r15,288(%r15)
+	br	%r14
+
+.align	16
+.Ltail:
+	la	%r9,64(%r9)
+	stm	%r0,%r7,160+4*0(%r15)
+	slgr	%r9,%r14
+	lm	%r0,%r3,160+4*8+4*8(%r15)
+	lghi	%r6,0
+	stm	%r10,%r13,160+4*12(%r15)
+	al	%r0,160+4*8(%r15)
+	al	%r1,160+4*9(%r15)
+	al	%r2,160+4*10(%r15)
+	al	%r3,160+4*11(%r15)
+	lrvr	%r0,%r0
+	lrvr	%r1,%r1
+	lrvr	%r2,%r2
+	lrvr	%r3,%r3
+	stm	%r0,%r3,160+4*8(%r15)
+
+.Loop_tail:
+	llgc	%r4,0(%r6,%r14)
+	llgc	%r5,160(%r6,%r15)
+	xr	%r5,%r4
+	stc	%r5,0(%r6,%r8)
+	la	%r6,1(%r6)
+	brct	%r9,.Loop_tail
+
+	j	.Ldone
+.size	ChaCha20_ctr32,.-ChaCha20_ctr32
+
+.align	32
+.Lsigma:
+.long	0x61707865,0x3320646e,0x79622d32,0x6b206574	# endian-neutral
+.asciz	"ChaCha20 for s390x, CRYPTOGAMS by <appro@openssl.org>"
+.align	4
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/include/internal/bn_conf.h b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/include/internal/bn_conf.h
new file mode 100644
index 0000000000..34bd8b78b4
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/include/internal/bn_conf.h
@@ -0,0 +1,28 @@
+/* WARNING: do not edit! */
+/* Generated by Makefile from crypto/include/internal/bn_conf.h.in */
+/*
+ * Copyright 2016 The OpenSSL Project Authors. All Rights Reserved.
+ *
+ * Licensed under the OpenSSL license (the "License").  You may not use
+ * this file except in compliance with the License.  You can obtain a copy
+ * in the file LICENSE in the source distribution or at
+ * https://www.openssl.org/source/license.html
+ */
+
+#ifndef HEADER_BN_CONF_H
+# define HEADER_BN_CONF_H
+
+/*
+ * The contents of this file are not used in the UEFI build, as
+ * both 32-bit and 64-bit builds are supported from a single run
+ * of the Configure script.
+ */
+
+/* Should we define BN_DIV2W here? */
+
+/* Only one for the following should be defined */
+#define SIXTY_FOUR_BIT_LONG
+#undef SIXTY_FOUR_BIT
+#undef THIRTY_TWO_BIT
+
+#endif
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/include/internal/dso_conf.h b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/include/internal/dso_conf.h
new file mode 100644
index 0000000000..18f6ac3bff
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/include/internal/dso_conf.h
@@ -0,0 +1,19 @@
+/* WARNING: do not edit! */
+/* Generated by Makefile from crypto/include/internal/dso_conf.h.in */
+/*
+ * Copyright 2016-2018 The OpenSSL Project Authors. All Rights Reserved.
+ *
+ * Licensed under the OpenSSL license (the "License").  You may not use
+ * this file except in compliance with the License.  You can obtain a copy
+ * in the file LICENSE in the source distribution or at
+ * https://www.openssl.org/source/license.html
+ */
+
+#ifndef HEADER_DSO_CONF_H
+# define HEADER_DSO_CONF_H
+
+# define DSO_DLFCN
+# define HAVE_DLFCN_H
+# define DSO_EXTENSION ".so"
+
+#endif
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/modes/ghash-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/modes/ghash-s390x.S
new file mode 100644
index 0000000000..f6bd2d0857
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/modes/ghash-s390x.S
@@ -0,0 +1,147 @@
+#include "s390x_arch.h"
+
+.text
+
+.globl	gcm_gmult_4bit
+.align	32
+gcm_gmult_4bit:
+	stmg	%r6,%r14,6*8(%r15)
+
+	aghi	%r2,-1
+	lghi	%r5,1
+	lghi	%r13,120
+	larl	%r14,rem_4bit
+
+	lg	%r1,8+1(%r2)		# Xi
+	j	.Lgmult_shortcut
+.type	gcm_gmult_4bit,@function
+.size	gcm_gmult_4bit,(.-gcm_gmult_4bit)
+
+.globl	gcm_ghash_4bit
+.align	32
+gcm_ghash_4bit:
+	larl	%r1,OPENSSL_s390xcap_P
+	lg	%r0,S390X_KIMD+8(%r1)	# load second word of kimd capabilities
+					#  vector
+	tmhh	%r0,0x4000	# check for function 65
+	jz	.Lsoft_ghash
+	lghi	%r0,S390X_GHASH	# function 65
+	la	%r1,0(%r2)	# H lies right after Xi in gcm128_context
+	.long	0xb93e0004	# kimd %r0,%r4
+	brc	1,.-4		# pay attention to "partial completion"
+	br	%r14
+.align	32
+.Lsoft_ghash:
+	stmg	%r6,%r14,6*8(%r15)
+
+	aghi	%r2,-1
+	srlg	%r5,%r5,4
+	lghi	%r13,120
+	larl	%r14,rem_4bit
+
+	lg	%r1,8+1(%r2)		# Xi
+	lg	%r0,0+1(%r2)
+	lghi	%r12,0
+.Louter:
+	xg	%r0,0(%r4)		# Xi ^= inp
+	xg	%r1,8(%r4)
+	xgr	%r0,%r12
+	stg	%r1,8+1(%r2)
+	stg	%r0,0+1(%r2)
+
+.Lgmult_shortcut:
+	lghi	%r12,0xf0
+	sllg	%r8,%r1,4
+	srlg	%r10,%r1,8		# extract second byte
+	ngr	%r8,%r12
+	lgr	%r9,%r1
+	lghi	%r11,14
+	ngr	%r9,%r12
+
+	lg	%r1,8(%r8,%r3)
+	lg	%r0,0(%r8,%r3)
+
+	sllg	%r8,%r10,4
+	sllg	%r6,%r1,3
+	ngr	%r8,%r12
+	ngr	%r6,%r13
+	ngr	%r10,%r12
+
+	sllg	%r12,%r0,60
+	srlg	%r1,%r1,4
+	srlg	%r0,%r0,4
+	xg	%r1,8(%r9,%r3)
+	xg	%r0,0(%r9,%r3)
+	lgr	%r9,%r10
+	sllg	%r7,%r1,3
+	xgr	%r1,%r12
+	ngr	%r7,%r13
+	sllg	%r12,%r0,60
+	j	.Lghash_inner
+.align	16
+.Lghash_inner:
+	srlg	%r1,%r1,4
+	srlg	%r0,%r0,4
+	xg	%r1,8(%r8,%r3)
+	llgc	%r10,0(%r11,%r2)
+	xg	%r0,0(%r8,%r3)
+	sllg	%r8,%r10,4
+	xg	%r0,0(%r6,%r14)
+	nill	%r8,0xf0
+	sllg	%r6,%r1,3
+	xgr	%r1,%r12
+	ngr	%r6,%r13
+	nill	%r10,0xf0
+
+	sllg	%r12,%r0,60
+	srlg	%r1,%r1,4
+	srlg	%r0,%r0,4
+	xg	%r1,8(%r9,%r3)
+	xg	%r0,0(%r9,%r3)
+	lgr	%r9,%r10
+	xg	%r0,0(%r7,%r14)
+	sllg	%r7,%r1,3
+	xgr	%r1,%r12
+	ngr	%r7,%r13
+	sllg	%r12,%r0,60
+	brct	%r11,.Lghash_inner
+
+	srlg	%r1,%r1,4
+	srlg	%r0,%r0,4
+	xg	%r1,8(%r8,%r3)
+	xg	%r0,0(%r8,%r3)
+	sllg	%r10,%r1,3
+	xg	%r0,0(%r6,%r14)
+	xgr	%r1,%r12
+	ngr	%r10,%r13
+
+	sllg	%r12,%r0,60
+	srlg	%r1,%r1,4
+	srlg	%r0,%r0,4
+	xg	%r1,8(%r9,%r3)
+	xg	%r0,0(%r9,%r3)
+	xgr	%r1,%r12
+	xg	%r0,0(%r7,%r14)
+
+	lg	%r12,0(%r10,%r14)
+	la	%r4,16(%r4)
+	sllg	%r12,%r12,4		# correct last rem_4bit[rem]
+	brctg	%r5,.Louter
+
+	xgr	%r0,%r12
+	stg	%r1,8+1(%r2)
+	stg	%r0,0+1(%r2)
+	lmg	%r6,%r14,6*8(%r15)
+	br	%r14
+.type	gcm_ghash_4bit,@function
+.size	gcm_ghash_4bit,(.-gcm_ghash_4bit)
+
+.align	64
+rem_4bit:
+	.long	0,0,29491200,0,58982400,0,38141952,0
+	.long	117964800,0,113901568,0,76283904,0,88997888,0
+	.long	235929600,0,265420800,0,227803136,0,206962688,0
+	.long	152567808,0,148504576,0,177995776,0,190709760,0
+.type	rem_4bit,@object
+.size	rem_4bit,(.-rem_4bit)
+.string	"GHASH for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/poly1305/poly1305-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/poly1305/poly1305-s390x.S
new file mode 100644
index 0000000000..262ee2dca7
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/poly1305/poly1305-s390x.S
@@ -0,0 +1,162 @@
+.text
+
+.globl	poly1305_init
+.type	poly1305_init,@function
+.align	16
+poly1305_init:
+	lghi	%r0,0
+	lghi	%r1,-1
+	stg	%r0,0(%r2)		# zero hash value
+	stg	%r0,8(%r2)
+	stg	%r0,16(%r2)
+
+	clgr	%r3,%r0
+	je	.Lno_key
+
+	lrvg	%r4,0(%r3)		# load little-endian key
+	lrvg	%r5,8(%r3)
+
+	nihl	%r1,0xffc0		# 0xffffffc0ffffffff
+	srlg	%r0,%r1,4		# 0x0ffffffc0fffffff
+	srlg	%r1,%r1,4
+	nill	%r1,0xfffc		# 0x0ffffffc0ffffffc
+
+	ngr	%r4,%r0
+	ngr	%r5,%r1
+
+	stg	%r4,32(%r2)
+	stg	%r5,40(%r2)
+
+.Lno_key:
+	lghi	%r2,0
+	br	%r14
+.size	poly1305_init,.-poly1305_init
+.globl	poly1305_blocks
+.type	poly1305_blocks,@function
+.align	16
+poly1305_blocks:
+	srlg	%r4,%r4,4			# fixed-up in 64-bit build
+	lghi	%r0,0
+	clgr	%r4,%r0
+	je	.Lno_data
+
+	stmg	%r6,%r14,48(%r15)
+
+	llgfr   %r5,%r5		# clear upper half, much needed with
+					# non-64-bit ABI
+	lg	%r0,32(%r2)		# load key
+	lg	%r1,40(%r2)
+
+	lg	%r11,0(%r2)		# load hash value
+	lg	%r13,8(%r2)
+	lg	%r14,16(%r2)
+
+	stg	%r2,16(%r15)	# off-load %r2
+	srlg	%r2,%r1,2
+	algr	%r2,%r1			# s1 = r1 + r1>>2
+	j	.Loop
+
+.align	16
+.Loop:
+	lrvg	%r7,0(%r3)		# load little-endian input
+	lrvg	%r9,8(%r3)
+	la	%r3,16(%r3)
+
+	algr	%r7,%r11		# accumulate input
+	alcgr	%r9,%r13
+
+	lgr	%r11,%r7
+	mlgr	%r6,%r0		# h0*r0	  -> %r6:%r7
+	lgr	%r13,%r9
+	mlgr	%r8,%r2		# h1*5*r1 -> %r8:%r9
+
+	mlgr	%r10,%r1			# h0*r1   -> %r10:%r11
+	mlgr	%r12,%r0			# h1*r0   -> %r12:%r13
+	alcgr	%r14,%r5
+
+	algr	%r7,%r9
+	lgr	%r9,%r14
+	alcgr	%r6,%r8
+	lghi	%r8,0
+
+	algr	%r13,%r11
+	alcgr	%r12,%r10
+
+	msgr	%r9,%r2		# h2*s1
+	msgr	%r14,%r0			# h2*r0
+
+	algr	%r13,%r9
+	alcgr	%r12,%r8		# %r8 is zero
+
+	algr	%r13,%r6
+	alcgr	%r14,%r12
+
+	lghi	%r11,-4			# final reduction step
+	ngr	%r11,%r14
+	srlg	%r10,%r14,2
+	algr	%r11,%r10
+	lghi	%r12,3
+	ngr	%r14,%r12
+
+	algr	%r11,%r7
+	alcgr	%r13,%r8		# %r8 is still zero
+	alcgr	%r14,%r8		# %r8 is still zero
+
+	brctg	%r4,.Loop
+
+	lg	%r2,16(%r15)	# restore %r2
+
+	stg	%r11,0(%r2)		# store hash value
+	stg	%r13,8(%r2)
+	stg	%r14,16(%r2)
+
+	lmg	%r6,%r14,48(%r15)
+.Lno_data:
+	br	%r14
+.size	poly1305_blocks,.-poly1305_blocks
+.globl	poly1305_emit
+.type	poly1305_emit,@function
+.align	16
+poly1305_emit:
+	stmg	%r6,%r9,48(%r15)
+
+	lg	%r5,0(%r2)
+	lg	%r6,8(%r2)
+	lg	%r7,16(%r2)
+
+	lghi	%r0,5
+	lghi	%r1,0
+	lgr	%r8,%r5
+	lgr	%r9,%r6
+
+	algr	%r5,%r0			# compare to modulus
+	alcgr	%r6,%r1
+	alcgr	%r7,%r1
+
+	srlg	%r7,%r7,2		# did it borrow/carry?
+	slgr	%r1,%r7			# 0-%r7>>2
+	lg	%r7,0(%r4)		# load nonce
+	lghi	%r0,-1
+	lg	%r2,8(%r4)
+	xgr	%r0,%r1			# ~%r1
+
+	ngr	%r5,%r1
+	ngr	%r8,%r0
+	ngr	%r6,%r1
+	ngr	%r9,%r0
+	ogr	%r5,%r8
+	rllg	%r8,%r7,32		# flip nonce words
+	ogr	%r6,%r9
+	rllg	%r9,%r2,32
+
+	algr	%r5,%r8			# accumulate nonce
+	alcgr	%r6,%r9
+
+	strvg	%r5,0(%r3)		# write little-endian result
+	strvg	%r6,8(%r3)
+
+	lmg	%r6,%r9,48(%r15)
+	br	%r14
+.size	poly1305_emit,.-poly1305_emit
+
+.string	"Poly1305 for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/rc4/rc4-s390x.s b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/rc4/rc4-s390x.s
new file mode 100644
index 0000000000..e5883fea42
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/rc4/rc4-s390x.s
@@ -0,0 +1,240 @@
+.text
+
+.globl	RC4
+.type	RC4,@function
+.align	64
+RC4:
+	stmg	%r6,%r11,6*8(%r15)
+	llgc	%r6,0(%r2)
+	llgc	%r10,1(%r2)
+	la	%r6,1(%r6)
+	nill	%r6,0xff
+	srlg	%r1,%r3,3
+	ltgr	%r1,%r1
+	llgc	%r8,2(%r6,%r2)
+	jz	.Lshort
+	j	.Loop8
+
+.align	64
+.Loop8:
+	la	%r10,0(%r10,%r8)	# 0
+	nill	%r10,255
+	la	%r7,1(%r6)
+	nill	%r7,255
+	llgc	%r11,2(%r10,%r2)
+	stc	%r8,2(%r10,%r2)
+	llgc	%r9,2(%r7,%r2)
+	stc	%r11,2(%r6,%r2)
+	cr	%r7,%r10
+	jne	.Lcmov0
+	la	%r9,0(%r8)
+.Lcmov0:
+	la	%r11,0(%r11,%r8)
+	nill	%r11,255
+	la	%r10,0(%r10,%r9)	# 1
+	nill	%r10,255
+	la	%r6,1(%r7)
+	nill	%r6,255
+	llgc	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r9,2(%r10,%r2)
+	llgc	%r8,2(%r6,%r2)
+	stc	%r11,2(%r7,%r2)
+	cr	%r6,%r10
+	jne	.Lcmov1
+	la	%r8,0(%r9)
+.Lcmov1:
+	la	%r11,0(%r11,%r9)
+	nill	%r11,255
+	la	%r10,0(%r10,%r8)	# 2
+	nill	%r10,255
+	la	%r7,1(%r6)
+	nill	%r7,255
+	sllg	%r0,%r0,8
+	ic	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r8,2(%r10,%r2)
+	llgc	%r9,2(%r7,%r2)
+	stc	%r11,2(%r6,%r2)
+	cr	%r7,%r10
+	jne	.Lcmov2
+	la	%r9,0(%r8)
+.Lcmov2:
+	la	%r11,0(%r11,%r8)
+	nill	%r11,255
+	la	%r10,0(%r10,%r9)	# 3
+	nill	%r10,255
+	la	%r6,1(%r7)
+	nill	%r6,255
+	sllg	%r0,%r0,8
+	ic	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r9,2(%r10,%r2)
+	llgc	%r8,2(%r6,%r2)
+	stc	%r11,2(%r7,%r2)
+	cr	%r6,%r10
+	jne	.Lcmov3
+	la	%r8,0(%r9)
+.Lcmov3:
+	la	%r11,0(%r11,%r9)
+	nill	%r11,255
+	la	%r10,0(%r10,%r8)	# 4
+	nill	%r10,255
+	la	%r7,1(%r6)
+	nill	%r7,255
+	sllg	%r0,%r0,8
+	ic	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r8,2(%r10,%r2)
+	llgc	%r9,2(%r7,%r2)
+	stc	%r11,2(%r6,%r2)
+	cr	%r7,%r10
+	jne	.Lcmov4
+	la	%r9,0(%r8)
+.Lcmov4:
+	la	%r11,0(%r11,%r8)
+	nill	%r11,255
+	la	%r10,0(%r10,%r9)	# 5
+	nill	%r10,255
+	la	%r6,1(%r7)
+	nill	%r6,255
+	sllg	%r0,%r0,8
+	ic	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r9,2(%r10,%r2)
+	llgc	%r8,2(%r6,%r2)
+	stc	%r11,2(%r7,%r2)
+	cr	%r6,%r10
+	jne	.Lcmov5
+	la	%r8,0(%r9)
+.Lcmov5:
+	la	%r11,0(%r11,%r9)
+	nill	%r11,255
+	la	%r10,0(%r10,%r8)	# 6
+	nill	%r10,255
+	la	%r7,1(%r6)
+	nill	%r7,255
+	sllg	%r0,%r0,8
+	ic	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r8,2(%r10,%r2)
+	llgc	%r9,2(%r7,%r2)
+	stc	%r11,2(%r6,%r2)
+	cr	%r7,%r10
+	jne	.Lcmov6
+	la	%r9,0(%r8)
+.Lcmov6:
+	la	%r11,0(%r11,%r8)
+	nill	%r11,255
+	la	%r10,0(%r10,%r9)	# 7
+	nill	%r10,255
+	la	%r6,1(%r7)
+	nill	%r6,255
+	sllg	%r0,%r0,8
+	ic	%r0,2(%r11,%r2)
+	llgc	%r11,2(%r10,%r2)
+	stc	%r9,2(%r10,%r2)
+	llgc	%r8,2(%r6,%r2)
+	stc	%r11,2(%r7,%r2)
+	cr	%r6,%r10
+	jne	.Lcmov7
+	la	%r8,0(%r9)
+.Lcmov7:
+	la	%r11,0(%r11,%r9)
+	nill	%r11,255
+	lg	%r9,0(%r4)
+	sllg	%r0,%r0,8
+	la	%r4,8(%r4)
+	ic	%r0,2(%r11,%r2)
+	xgr	%r0,%r9
+	stg	%r0,0(%r5)
+	la	%r5,8(%r5)
+	brctg	%r1,.Loop8
+
+.Lshort:
+	lghi	%r0,7
+	ngr	%r3,%r0
+	jz	.Lexit
+	j	.Loop1
+
+.align	16
+.Loop1:
+	la	%r10,0(%r10,%r8)
+	nill	%r10,255
+	llgc	%r11,2(%r10,%r2)
+	stc	%r8,2(%r10,%r2)
+	stc	%r11,2(%r6,%r2)
+	ar	%r11,%r8
+	ahi	%r6,1
+	nill	%r11,255
+	nill	%r6,255
+	llgc	%r0,0(%r4)
+	la	%r4,1(%r4)
+	llgc	%r11,2(%r11,%r2)
+	llgc	%r8,2(%r6,%r2)
+	xr	%r0,%r11
+	stc	%r0,0(%r5)
+	la	%r5,1(%r5)
+	brct	%r3,.Loop1
+
+.Lexit:
+	ahi	%r6,-1
+	stc	%r6,0(%r2)
+	stc	%r10,1(%r2)
+	lmg	%r6,%r11,6*8(%r15)
+	br	%r14
+.size	RC4,.-RC4
+.string	"RC4 for s390x, CRYPTOGAMS by <appro@openssl.org>"
+
+.globl	RC4_set_key
+.type	RC4_set_key,@function
+.align	64
+RC4_set_key:
+	stmg	%r6,%r8,6*8(%r15)
+	lhi	%r0,256
+	la	%r1,0(%r0)
+	sth	%r1,0(%r2)
+.align	4
+.L1stloop:
+	stc	%r1,2(%r1,%r2)
+	la	%r1,1(%r1)
+	brct	%r0,.L1stloop
+
+	lghi	%r7,-256
+	lr	%r0,%r3
+	la	%r8,0(%r0)
+	la	%r1,0(%r0)
+.align	16
+.L2ndloop:
+	llgc	%r5,2+256(%r7,%r2)
+	llgc	%r6,0(%r8,%r4)
+	la	%r1,0(%r1,%r5)
+	la	%r7,1(%r7)
+	la	%r1,0(%r1,%r6)
+	nill	%r1,255
+	la	%r8,1(%r8)
+	tml	%r7,255
+	llgc	%r6,2(%r1,%r2)
+	stc	%r6,2+256-1(%r7,%r2)
+	stc	%r5,2(%r1,%r2)
+	jz	.Ldone
+	brct	%r0,.L2ndloop
+	lr	%r0,%r3
+	la	%r8,0(%r0)
+	j	.L2ndloop
+.Ldone:
+	lmg	%r6,%r8,6*8(%r15)
+	br	%r14
+.size	RC4_set_key,.-RC4_set_key
+
+.globl	RC4_options
+.type	RC4_options,@function
+.align	16
+RC4_options:
+	larl	%r2,.Loptions
+	br	%r14
+.size	RC4_options,.-RC4_options
+.section	.rodata
+.Loptions:
+.align	8
+.string	"rc4(8x,char)"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/s390xcpuid.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/s390xcpuid.S
new file mode 100644
index 0000000000..7c0a5a0720
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/s390xcpuid.S
@@ -0,0 +1,323 @@
+#include "s390x_arch.h"
+
+.text
+
+.globl	OPENSSL_s390x_facilities
+.type	OPENSSL_s390x_facilities,@function
+.align	16
+OPENSSL_s390x_facilities:
+	lghi	%r0,0
+	larl	%r4,OPENSSL_s390xcap_P
+
+	stg	%r0,S390X_STFLE+8(%r4)	# wipe capability vectors
+	stg	%r0,S390X_STFLE+16(%r4)
+	stg	%r0,S390X_STFLE+24(%r4)
+	stg	%r0,S390X_KIMD(%r4)
+	stg	%r0,S390X_KIMD+8(%r4)
+	stg	%r0,S390X_KLMD(%r4)
+	stg	%r0,S390X_KLMD+8(%r4)
+	stg	%r0,S390X_KM(%r4)
+	stg	%r0,S390X_KM+8(%r4)
+	stg	%r0,S390X_KMC(%r4)
+	stg	%r0,S390X_KMC+8(%r4)
+	stg	%r0,S390X_KMAC(%r4)
+	stg	%r0,S390X_KMAC+8(%r4)
+	stg	%r0,S390X_KMCTR(%r4)
+	stg	%r0,S390X_KMCTR+8(%r4)
+	stg	%r0,S390X_KMO(%r4)
+	stg	%r0,S390X_KMO+8(%r4)
+	stg	%r0,S390X_KMF(%r4)
+	stg	%r0,S390X_KMF+8(%r4)
+	stg	%r0,S390X_PRNO(%r4)
+	stg	%r0,S390X_PRNO+8(%r4)
+	stg	%r0,S390X_KMA(%r4)
+	stg	%r0,S390X_KMA+8(%r4)
+
+	.long	0xb2b04000		# stfle	0(%r4)
+	brc	8,.Ldone
+	lghi	%r0,1
+	.long	0xb2b04000		# stfle 0(%r4)
+	brc	8,.Ldone
+	lghi	%r0,2
+	.long	0xb2b04000		# stfle 0(%r4)
+.Ldone:
+	lmg	%r2,%r3,S390X_STFLE(%r4)
+	tmhl	%r2,0x4000		# check for message-security-assist
+	jz	.Lret
+
+	lghi	%r0,S390X_QUERY		# query kimd capabilities
+	la	%r1,S390X_KIMD(%r4)
+	.long	0xb93e0002		# kimd %r0,%r2
+
+	lghi	%r0,S390X_QUERY		# query klmd capabilities
+	la	%r1,S390X_KLMD(%r4)
+	.long	0xb93f0002		# klmd %r0,%r2
+
+	lghi	%r0,S390X_QUERY		# query km capability vector
+	la	%r1,S390X_KM(%r4)
+	.long	0xb92e0042		# km %r4,%r2
+
+	lghi	%r0,S390X_QUERY		# query kmc capability vector
+	la	%r1,S390X_KMC(%r4)
+	.long	0xb92f0042		# kmc %r4,%r2
+
+	lghi	%r0,S390X_QUERY		# query kmac capability vector
+	la	%r1,S390X_KMAC(%r4)
+	.long	0xb91e0042		# kmac %r4,%r2
+
+	tmhh	%r3,0x0004		# check for message-security-assist-4
+	jz	.Lret
+
+	lghi	%r0,S390X_QUERY		# query kmctr capability vector
+	la	%r1,S390X_KMCTR(%r4)
+	.long	0xb92d2042		# kmctr %r4,%r2,%r2
+
+	lghi	%r0,S390X_QUERY		# query kmo capability vector
+	la	%r1,S390X_KMO(%r4)
+	.long	0xb92b0042		# kmo %r4,%r2
+
+	lghi	%r0,S390X_QUERY		# query kmf capability vector
+	la	%r1,S390X_KMF(%r4)
+	.long	0xb92a0042		# kmf %r4,%r2
+
+	tml	%r2,0x40		# check for message-security-assist-5
+	jz	.Lret
+
+	lghi	%r0,S390X_QUERY		# query prno capability vector
+	la	%r1,S390X_PRNO(%r4)
+	.long	0xb93c0042		# prno %r4,%r2
+
+	lg	%r2,S390X_STFLE+16(%r4)
+	tmhl	%r2,0x2000		# check for message-security-assist-8
+	jz	.Lret
+
+	lghi	%r0,S390X_QUERY		# query kma capability vector
+	la	%r1,S390X_KMA(%r4)
+	.long	0xb9294022		# kma %r2,%r4,%r2
+
+.Lret:
+	br	%r14
+.size	OPENSSL_s390x_facilities,.-OPENSSL_s390x_facilities
+
+.globl	OPENSSL_rdtsc
+.type	OPENSSL_rdtsc,@function
+.align	16
+OPENSSL_rdtsc:
+	larl	%r4,OPENSSL_s390xcap_P
+	tm	S390X_STFLE+3(%r4),0x40	# check for store-clock-fast facility
+	jz	.Lstck
+
+	.long	0xb27cf010	# stckf 16(%r15)
+	lg	%r2,16(%r15)
+	br	%r14
+.Lstck:
+	stck	16(%r15)
+	lg	%r2,16(%r15)
+	br	%r14
+.size	OPENSSL_rdtsc,.-OPENSSL_rdtsc
+
+.globl	OPENSSL_atomic_add
+.type	OPENSSL_atomic_add,@function
+.align	16
+OPENSSL_atomic_add:
+	l	%r1,0(%r2)
+.Lspin:	lr	%r0,%r1
+	ar	%r0,%r3
+	cs	%r1,%r0,0(%r2)
+	brc	4,.Lspin
+	lgfr	%r2,%r0		# OpenSSL expects the new value
+	br	%r14
+.size	OPENSSL_atomic_add,.-OPENSSL_atomic_add
+
+.globl	OPENSSL_wipe_cpu
+.type	OPENSSL_wipe_cpu,@function
+.align	16
+OPENSSL_wipe_cpu:
+	xgr	%r0,%r0
+	xgr	%r1,%r1
+	lgr	%r2,%r15
+	xgr	%r3,%r3
+	xgr	%r4,%r4
+	lzdr	%f0
+	lzdr	%f1
+	lzdr	%f2
+	lzdr	%f3
+	lzdr	%f4
+	lzdr	%f5
+	lzdr	%f6
+	lzdr	%f7
+	br	%r14
+.size	OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
+
+.globl	OPENSSL_cleanse
+.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
+	jh	.Lot
+	clgr	%r3,%r0
+	bcr	8,%r14
+.Little:
+	stc	%r0,0(%r2)
+	la	%r2,1(%r2)
+	brctg	%r3,.Little
+	br	%r14
+.align	4
+.Lot:	tmll	%r2,7
+	jz	.Laligned
+	stc	%r0,0(%r2)
+	la	%r2,1(%r2)
+	brctg	%r3,.Lot
+.Laligned:
+	srlg	%r4,%r3,3
+.Loop:	stg	%r0,0(%r2)
+	la	%r2,8(%r2)
+	brctg	%r4,.Loop
+	lghi	%r4,7
+	ngr	%r3,%r4
+	jnz	.Little
+	br	%r14
+.size	OPENSSL_cleanse,.-OPENSSL_cleanse
+
+.globl	CRYPTO_memcmp
+.type	CRYPTO_memcmp,@function
+.align	16
+CRYPTO_memcmp:
+#if !defined(__s390x__) && !defined(__s390x)
+	llgfr	%r4,%r4
+#endif
+	lghi	%r5,0
+	clgr	%r4,%r5
+	je	.Lno_data
+
+.Loop_cmp:
+	llgc	%r0,0(%r2)
+	la	%r2,1(%r2)
+	llgc	%r1,0(%r3)
+	la	%r3,1(%r3)
+	xr	%r1,%r0
+	or	%r5,%r1
+	brctg	%r4,.Loop_cmp
+
+	lnr	%r5,%r5
+	srl	%r5,31
+.Lno_data:
+	lgr	%r2,%r5
+	br	%r14
+.size	CRYPTO_memcmp,.-CRYPTO_memcmp
+
+.globl	OPENSSL_instrument_bus
+.type	OPENSSL_instrument_bus,@function
+.align	16
+OPENSSL_instrument_bus:
+	lghi	%r2,0
+	br	%r14
+.size	OPENSSL_instrument_bus,.-OPENSSL_instrument_bus
+
+.globl	OPENSSL_instrument_bus2
+.type	OPENSSL_instrument_bus2,@function
+.align	16
+OPENSSL_instrument_bus2:
+	lghi	%r2,0
+	br	%r14
+.size	OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2
+
+.globl	OPENSSL_vx_probe
+.type	OPENSSL_vx_probe,@function
+.align	16
+OPENSSL_vx_probe:
+	.word	0xe700,0x0000,0x0044	# vzero %v0
+	br	%r14
+.size	OPENSSL_vx_probe,.-OPENSSL_vx_probe
+.globl	s390x_kimd
+.type	s390x_kimd,@function
+.align	16
+s390x_kimd:
+	llgfr	%r0,%r4
+	lgr	%r1,%r5
+
+	.long	0xb93e0002	# kimd %r0,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	br	%r14
+.size	s390x_kimd,.-s390x_kimd
+.globl	s390x_klmd
+.type	s390x_klmd,@function
+.align	32
+s390x_klmd:
+	llgfr	%r0,%r6
+	lg	%r1,160(%r15)
+
+	.long	0xb93f0042	# klmd %r4,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	br	%r14
+.size	s390x_klmd,.-s390x_klmd
+.globl	s390x_km
+.type	s390x_km,@function
+.align	16
+s390x_km:
+	lr	%r0,%r5
+	lgr	%r1,%r6
+
+	.long	0xb92e0042	# km %r4,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	br	%r14
+.size	s390x_km,.-s390x_km
+.globl	s390x_kmac
+.type	s390x_kmac,@function
+.align	16
+s390x_kmac:
+	lr	%r0,%r4
+	lgr	%r1,%r5
+
+	.long	0xb91e0002	# kmac %r0,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	br	%r14
+.size	s390x_kmac,.-s390x_kmac
+.globl	s390x_kmo
+.type	s390x_kmo,@function
+.align	16
+s390x_kmo:
+	lr	%r0,%r5
+	lgr	%r1,%r6
+
+	.long	0xb92b0042	# kmo %r4,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	br	%r14
+.size	s390x_kmo,.-s390x_kmo
+.globl	s390x_kmf
+.type	s390x_kmf,@function
+.align	16
+s390x_kmf:
+	lr	%r0,%r5
+	lgr	%r1,%r6
+
+	.long	0xb92a0042	# kmf %r4,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+
+	br	%r14
+.size	s390x_kmf,.-s390x_kmf
+.globl	s390x_kma
+.type	s390x_kma,@function
+.align	16
+s390x_kma:
+	stg	%r6,6*8(%r15)
+	lmg	%r0,%r1,160(%r15)
+
+	.long	0xb9292064	# kma %r6,%r2,%r4
+	brc	1,.-4		# pay attention to "partial completion"
+
+	lg	%r6,6*8(%r15)
+	br	%r14
+.size	s390x_kma,.-s390x_kma
+.section	.init
+	brasl	%r14,OPENSSL_cpuid_setup
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/keccak1600-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/keccak1600-s390x.S
new file mode 100644
index 0000000000..1a74fdb20e
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/keccak1600-s390x.S
@@ -0,0 +1,464 @@
+.text
+
+.type	__KeccakF1600,@function
+.align	32
+__KeccakF1600:
+	stg	%r14,8*14(%r15)
+	lg	%r0,160(%r2)
+	lg	%r1,168(%r2)
+	lg	%r5,176(%r2)
+	lg	%r6,184(%r2)
+	lg	%r7,192(%r2)
+	larl	%r4,iotas
+	j	.Loop
+
+.align	16
+.Loop:
+	lg	%r8,0(%r2)
+	lg	%r9,48(%r2)
+	lg	%r10,96(%r2)
+	lg	%r11,144(%r2)
+
+	xgr	%r0,%r8
+	xg	%r1,8(%r2)
+	xg	%r5,16(%r2)
+	xg	%r6,24(%r2)
+	lgr	%r12,%r7
+	xg	%r7,32(%r2)
+
+	xg	%r0,40(%r2)
+	xgr	%r1,%r9
+	xg	%r5,56(%r2)
+	xg	%r6,64(%r2)
+	xg	%r7,72(%r2)
+
+	xg	%r0,80(%r2)
+	xg	%r1,88(%r2)
+	xgr	%r5,%r10
+	xg	%r6,104(%r2)
+	xg	%r7,112(%r2)
+
+	xg	%r0,120(%r2)
+	xg	%r1,128(%r2)
+	xg	%r5,136(%r2)
+	xgr	%r6,%r11
+	xg	%r7,152(%r2)
+
+	lgr	%r13,%r5
+	rllg	%r5,%r5,1
+	xgr	%r5,%r0		# D[1] = ROL64(C[2], 1) ^ C[0]
+
+	rllg	%r0,%r0,1
+	xgr	%r0,%r6		# D[4] = ROL64(C[0], 1) ^ C[3]
+
+	rllg	%r6,%r6,1
+	xgr	%r6,%r1		# D[2] = ROL64(C[3], 1) ^ C[1]
+
+	rllg	%r1,%r1,1
+	xgr	%r1,%r7		# D[0] = ROL64(C[1], 1) ^ C[4]
+
+	rllg	%r7,%r7,1
+	xgr	%r7,%r13		# D[3] = ROL64(C[4], 1) ^ C[2]
+	xgr	%r9,%r5
+	xgr	%r10,%r6
+	xgr	%r11,%r7
+	 rllg	%r9,%r9,44
+	xgr	%r12,%r0
+	 rllg	%r10,%r10,43
+	xgr	%r8,%r1
+
+	lgr	%r13,%r9
+	ogr	%r9,%r10
+	 rllg	%r11,%r11,21
+	xgr	%r9,%r8		#	    C[0] ^ ( C[1] | C[2])
+	 rllg	%r12,%r12,14
+	xg	%r9,0(%r4)
+	la	%r4,8(%r4)
+	stg	%r9,0(%r3)	# R[0][0] = C[0] ^ ( C[1] | C[2]) ^ iotas[i]
+
+	lgr	%r14,%r12
+	ngr	%r12,%r11
+	 lghi	%r9,-1		# no 'not' instruction :-(
+	xgr	%r12,%r10		#	    C[2] ^ ( C[4] & C[3])
+	 xgr	%r10,%r9		# not	%r10
+	stg	%r12,16(%r3)	# R[0][2] = C[2] ^ ( C[4] & C[3])
+	 ogr	%r10,%r11
+	 xgr	%r10,%r13		#	    C[1] ^ (~C[2] | C[3])
+
+	ngr	%r13,%r8
+	 stg	%r10,8(%r3)	# R[0][1] = C[1] ^ (~C[2] | C[3])
+	xgr	%r13,%r14		#	    C[4] ^ ( C[1] & C[0])
+	 ogr	%r14,%r8
+	stg	%r13,32(%r3)	# R[0][4] = C[4] ^ ( C[1] & C[0])
+	 xgr	%r14,%r11		#	    C[3] ^ ( C[4] | C[0])
+	 stg	%r14,24(%r3)	# R[0][3] = C[3] ^ ( C[4] | C[0])
+
+
+	lg	%r8,24(%r2)
+	lg	%r12,176(%r2)
+	lg	%r11,128(%r2)
+	lg	%r9,72(%r2)
+	lg	%r10,80(%r2)
+
+	xgr	%r8,%r7
+	xgr	%r12,%r6
+	 rllg	%r8,%r8,28
+	xgr	%r11,%r5
+	 rllg	%r12,%r12,61
+	xgr	%r9,%r0
+	 rllg	%r11,%r11,45
+	xgr	%r10,%r1
+
+	lgr	%r13,%r8
+	ogr	%r8,%r12
+	 rllg	%r9,%r9,20
+	xgr	%r8,%r11		#	    C[3] ^ (C[0] |  C[4])
+	 rllg	%r10,%r10,3
+	stg	%r8,64(%r3)	# R[1][3] = C[3] ^ (C[0] |  C[4])
+
+	lgr	%r14,%r9
+	ngr	%r9,%r13
+	 lghi	%r8,-1		# no 'not' instruction :-(
+	xgr	%r9,%r12		#	    C[4] ^ (C[1] &  C[0])
+	 xgr	%r12,%r8		# not	%r12
+	stg	%r9,72(%r3)	# R[1][4] = C[4] ^ (C[1] &  C[0])
+
+	 ogr	%r12,%r11
+	 xgr	%r12,%r10		#	    C[2] ^ (~C[4] | C[3])
+
+	ngr	%r11,%r10
+	 stg	%r12,56(%r3)	# R[1][2] = C[2] ^ (~C[4] | C[3])
+	xgr	%r11,%r14		#	    C[1] ^ (C[3] &  C[2])
+	 ogr	%r14,%r10
+	stg	%r11,48(%r3)	# R[1][1] = C[1] ^ (C[3] &  C[2])
+	 xgr	%r14,%r13		#	    C[0] ^ (C[1] |  C[2])
+	 stg	%r14,40(%r3)	# R[1][0] = C[0] ^ (C[1] |  C[2])
+
+
+	lg	%r10,104(%r2)
+	lg	%r11,152(%r2)
+	lg	%r9,56(%r2)
+	lg	%r12,160(%r2)
+	lg	%r8,8(%r2)
+
+	xgr	%r10,%r7
+	xgr	%r11,%r0
+	 rllg	%r10,%r10,25
+	xgr	%r9,%r6
+	 rllg	%r11,%r11,8
+	xgr	%r12,%r1
+	 rllg	%r9,%r9,6
+	xgr	%r8,%r5
+
+	lgr	%r13,%r10
+	ngr	%r10,%r11
+	 rllg	%r12,%r12,18
+	xgr	%r10,%r9		#	     C[1] ^ ( C[2] & C[3])
+	lghi	%r14,-1		# no 'not' instruction :-(
+	stg	%r10,88(%r3)	# R[2][1] =  C[1] ^ ( C[2] & C[3])
+
+	xgr	%r11,%r14		# not	%r11
+	lgr	%r14,%r12
+	ngr	%r12,%r11
+	 rllg	%r8,%r8,1
+	xgr	%r12,%r13		#	     C[2] ^ ( C[4] & ~C[3])
+	 ogr	%r13,%r9
+	stg	%r12,96(%r3)	# R[2][2] =  C[2] ^ ( C[4] & ~C[3])
+	 xgr	%r13,%r8		#	     C[0] ^ ( C[2] | C[1])
+
+	ngr	%r9,%r8
+	 stg	%r13,80(%r3)	# R[2][0] =  C[0] ^ ( C[2] | C[1])
+	xgr	%r9,%r14		#	     C[4] ^ ( C[1] & C[0])
+	 ogr	%r8,%r14
+	stg	%r9,112(%r3)	# R[2][4] =  C[4] ^ ( C[1] & C[0])
+	 xgr	%r8,%r11		#	    ~C[3] ^ ( C[0] | C[4])
+	 stg	%r8,104(%r3)	# R[2][3] = ~C[3] ^ ( C[0] | C[4])
+
+
+	lg	%r10,88(%r2)
+	lg	%r11,136(%r2)
+	lg	%r9,40(%r2)
+	lg	%r12,184(%r2)
+	lg	%r8,32(%r2)
+
+	xgr	%r10,%r5
+	xgr	%r11,%r6
+	 rllg	%r10,%r10,10
+	xgr	%r9,%r1
+	 rllg	%r11,%r11,15
+	xgr	%r12,%r7
+	 rllg	%r9,%r9,36
+	xgr	%r8,%r0
+	 rllg	%r12,%r12,56
+
+	lgr	%r13,%r10
+	ogr	%r10,%r11
+	lghi	%r14,-1		# no 'not' instruction :-(
+	xgr	%r10,%r9		#	     C[1] ^ ( C[2] | C[3])
+	xgr	%r11,%r14		# not	%r11
+	stg	%r10,128(%r3)	# R[3][1] =  C[1] ^ ( C[2] | C[3])
+
+	lgr	%r14,%r12
+	ogr	%r12,%r11
+	 rllg	%r8,%r8,27
+	xgr	%r12,%r13		#	     C[2] ^ ( C[4] | ~C[3])
+	 ngr	%r13,%r9
+	stg	%r12,136(%r3)	# R[3][2] =  C[2] ^ ( C[4] | ~C[3])
+	 xgr	%r13,%r8		#	     C[0] ^ ( C[2] & C[1])
+
+	ogr	%r9,%r8
+	 stg	%r13,120(%r3)	# R[3][0] =  C[0] ^ ( C[2] & C[1])
+	xgr	%r9,%r14		#	     C[4] ^ ( C[1] | C[0])
+	 ngr	%r8,%r14
+	stg	%r9,152(%r3)	# R[3][4] =  C[4] ^ ( C[1] | C[0])
+	 xgr	%r8,%r11		#	    ~C[3] ^ ( C[0] & C[4])
+	 stg	%r8,144(%r3)	# R[3][3] = ~C[3] ^ ( C[0] & C[4])
+
+
+	xg	%r6,16(%r2)
+	xg	%r7,64(%r2)
+	xg	%r5,168(%r2)
+	xg	%r0,112(%r2)
+	xgr	%r3,%r2		# xchg	%r3,%r2
+	 rllg	%r6,%r6,62
+	xg	%r1,120(%r2)
+	 rllg	%r7,%r7,55
+	xgr	%r2,%r3
+	 rllg	%r5,%r5,2
+	xgr	%r3,%r2
+	 rllg	%r0,%r0,39
+	lgr	%r13,%r6
+	ngr	%r6,%r7
+	lghi	%r14,-1		# no 'not' instruction :-(
+	xgr	%r6,%r5		#	     C[4] ^ ( C[0] & C[1])
+	xgr	%r7,%r14		# not	%r7
+	stg	%r6,192(%r2)	# R[4][4] =  C[4] ^ ( C[0] & C[1])
+
+	lgr	%r14,%r0
+	ngr	%r0,%r7
+	 rllg	%r1,%r1,41
+	xgr	%r0,%r13		#	     C[0] ^ ( C[2] & ~C[1])
+	 ogr	%r13,%r5
+	stg	%r0,160(%r2)	# R[4][0] =  C[0] ^ ( C[2] & ~C[1])
+	 xgr	%r13,%r1		#	     C[3] ^ ( C[0] | C[4])
+
+	ngr	%r5,%r1
+	 stg	%r13,184(%r2)	# R[4][3] =  C[3] ^ ( C[0] | C[4])
+	xgr	%r5,%r14		#	     C[2] ^ ( C[4] & C[3])
+	 ogr	%r1,%r14
+	stg	%r5,176(%r2)	# R[4][2] =  C[2] ^ ( C[4] & C[3])
+	 xgr	%r1,%r7		#	    ~C[1] ^ ( C[2] | C[3])
+
+	lgr	%r7,%r6		# harmonize with the loop top
+	lgr	%r6,%r13
+	 stg	%r1,168(%r2)	# R[4][1] = ~C[1] ^ ( C[2] | C[3])
+
+	tmll	%r4,255
+	jnz	.Loop
+
+	lg	%r14,8*14(%r15)
+	br	%r14
+.size	__KeccakF1600,.-__KeccakF1600
+.type	KeccakF1600,@function
+.align	32
+KeccakF1600:
+.LKeccakF1600:
+	lghi	%r1,-360
+	stmg	%r6,%r15,8*6(%r15)
+	lgr	%r0,%r15
+	la	%r15,0(%r1,%r15)
+	stg	%r0,0(%r15)
+
+	lghi	%r8,-1		# no 'not' instruction :-(
+	lghi	%r9,-1
+	lghi	%r10,-1
+	lghi	%r11,-1
+	lghi	%r12,-1
+	lghi	%r13,-1
+	xg	%r8,8(%r2)
+	xg	%r9,16(%r2)
+	xg	%r10,64(%r2)
+	xg	%r11,96(%r2)
+	xg	%r12,136(%r2)
+	xg	%r13,160(%r2)
+	stmg	%r8,%r9,8(%r2)
+	stg	%r10,64(%r2)
+	stg	%r11,96(%r2)
+	stg	%r12,136(%r2)
+	stg	%r13,160(%r2)
+
+	la	%r3,160(%r15)
+
+	bras	%r14,__KeccakF1600
+
+	lghi	%r8,-1		# no 'not' instruction :-(
+	lghi	%r9,-1
+	lghi	%r10,-1
+	lghi	%r11,-1
+	lghi	%r12,-1
+	lghi	%r13,-1
+	xg	%r8,8(%r2)
+	xg	%r9,16(%r2)
+	xg	%r10,64(%r2)
+	xg	%r11,96(%r2)
+	xg	%r12,136(%r2)
+	xg	%r13,160(%r2)
+	stmg	%r8,%r9,8(%r2)
+	stg	%r10,64(%r2)
+	stg	%r11,96(%r2)
+	stg	%r12,136(%r2)
+	stg	%r13,160(%r2)
+
+	lmg	%r6,%r15,360+6*8(%r15)
+	br	%r14
+.size	KeccakF1600,.-KeccakF1600
+.globl	SHA3_absorb
+.type	SHA3_absorb,@function
+.align	32
+SHA3_absorb:
+	lghi	%r1,-360
+	stmg	%r5,%r15,8*5(%r15)
+	lgr	%r0,%r15
+	la	%r15,0(%r1,%r15)
+	stg	%r0,0(%r15)
+
+	lghi	%r8,-1		# no 'not' instruction :-(
+	lghi	%r9,-1
+	lghi	%r10,-1
+	lghi	%r11,-1
+	lghi	%r12,-1
+	lghi	%r13,-1
+	xg	%r8,8(%r2)
+	xg	%r9,16(%r2)
+	xg	%r10,64(%r2)
+	xg	%r11,96(%r2)
+	xg	%r12,136(%r2)
+	xg	%r13,160(%r2)
+	stmg	%r8,%r9,8(%r2)
+	stg	%r10,64(%r2)
+	stg	%r11,96(%r2)
+	stg	%r12,136(%r2)
+	stg	%r13,160(%r2)
+
+.Loop_absorb:
+	clgr	%r4,%r5
+	jl	.Ldone_absorb
+
+	srlg	%r5,%r5,3
+	la	%r1,0(%r2)
+
+.Lblock_absorb:
+	lrvg	%r0,0(%r3)
+	la	%r3,8(%r3)
+	xg	%r0,0(%r1)
+	aghi	%r4,-8
+	stg	%r0,0(%r1)
+	la	%r1,8(%r1)
+	brct	%r5,.Lblock_absorb
+
+	stmg	%r3,%r4,360+3*8(%r15)
+	la	%r3,160(%r15)
+	bras	%r14,__KeccakF1600
+	lmg	%r3,%r5,360+3*8(%r15)
+	j	.Loop_absorb
+
+.align	16
+.Ldone_absorb:
+	lghi	%r8,-1		# no 'not' instruction :-(
+	lghi	%r9,-1
+	lghi	%r10,-1
+	lghi	%r11,-1
+	lghi	%r12,-1
+	lghi	%r13,-1
+	xg	%r8,8(%r2)
+	xg	%r9,16(%r2)
+	xg	%r10,64(%r2)
+	xg	%r11,96(%r2)
+	xg	%r12,136(%r2)
+	xg	%r13,160(%r2)
+	stmg	%r8,%r9,8(%r2)
+	stg	%r10,64(%r2)
+	stg	%r11,96(%r2)
+	stg	%r12,136(%r2)
+	stg	%r13,160(%r2)
+
+	lgr	%r2,%r4		# return value
+
+	lmg	%r6,%r15,360+6*8(%r15)
+	br	%r14
+.size	SHA3_absorb,.-SHA3_absorb
+.globl	SHA3_squeeze
+.type	SHA3_squeeze,@function
+.align	32
+SHA3_squeeze:
+	srlg	%r5,%r5,3
+	stg	%r14,2*8(%r15)
+	lghi	%r14,8
+	stg	%r5,5*8(%r15)
+	la	%r1,0(%r2)
+
+	j	.Loop_squeeze
+
+.align	16
+.Loop_squeeze:
+	clgr %r4,%r14
+	jl	.Ltail_squeeze
+
+	lrvg	%r0,0(%r1)
+	la	%r1,8(%r1)
+	stg	%r0,0(%r3)
+	la	%r3,8(%r3)
+	aghi	%r4,-8			# len -= 8
+	jz	.Ldone_squeeze
+
+	brct	%r5,.Loop_squeeze	# bsz--
+
+	stmg	%r3,%r4,3*8(%r15)
+	bras	%r14,.LKeccakF1600
+	lmg	%r3,%r5,3*8(%r15)
+	lghi	%r14,8
+	la	%r1,0(%r2)
+	j	.Loop_squeeze
+
+.Ltail_squeeze:
+	lg	%r0,0(%r1)
+.Loop_tail_squeeze:
+	stc	%r0,0(%r3)
+	la	%r3,1(%r3)
+	srlg	%r0,%r0,8
+	brct	%r4,.Loop_tail_squeeze
+
+.Ldone_squeeze:
+	lg	%r14,2*8(%r15)
+	br	%r14
+.size	SHA3_squeeze,.-SHA3_squeeze
+.align	256
+	.quad	0,0,0,0,0,0,0,0
+.type	iotas,@object
+iotas:
+	.quad	0x0000000000000001
+	.quad	0x0000000000008082
+	.quad	0x800000000000808a
+	.quad	0x8000000080008000
+	.quad	0x000000000000808b
+	.quad	0x0000000080000001
+	.quad	0x8000000080008081
+	.quad	0x8000000000008009
+	.quad	0x000000000000008a
+	.quad	0x0000000000000088
+	.quad	0x0000000080008009
+	.quad	0x000000008000000a
+	.quad	0x000000008000808b
+	.quad	0x800000000000008b
+	.quad	0x8000000000008089
+	.quad	0x8000000000008003
+	.quad	0x8000000000008002
+	.quad	0x8000000000000080
+	.quad	0x000000000000800a
+	.quad	0x800000008000000a
+	.quad	0x8000000080008081
+	.quad	0x8000000000008080
+	.quad	0x0000000080000001
+	.quad	0x8000000080008008
+.size	iotas,.-iotas
+.asciz	"Keccak-1600 absorb and squeeze for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha1-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha1-s390x.S
new file mode 100644
index 0000000000..37413bfb08
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha1-s390x.S
@@ -0,0 +1,1207 @@
+#include "s390x_arch.h"
+
+.text
+.align	64
+.type	Ktable,@object
+Ktable: .long	0x5a827999,0x6ed9eba1,0x8f1bbcdc,0xca62c1d6
+	.skip	48	#.long	0,0,0,0,0,0,0,0,0,0,0,0
+.size	Ktable,.-Ktable
+.globl	sha1_block_data_order
+.type	sha1_block_data_order,@function
+sha1_block_data_order:
+	larl	%r1,OPENSSL_s390xcap_P
+	lg	%r0,S390X_KIMD(%r1)	# check kimd capabilities
+	tmhh	%r0,16384
+	jz	.Lsoftware
+	lghi	%r0,1
+	lgr	%r1,%r2
+	lgr	%r2,%r3
+	sllg	%r3,%r4,6
+	.long	0xb93e0002	# kimd %r0,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+	br	%r14
+.align	16
+.Lsoftware:
+	lghi	%r1,-224
+	stg	%r2,16(%r15)
+	stmg	%r6,%r15,48(%r15)
+	lgr	%r0,%r15
+	la	%r15,0(%r1,%r15)
+	stg	%r0,0(%r15)
+
+	larl	%r10,Ktable
+	llgf	%r5,0(%r2)
+	llgf	%r6,4(%r2)
+	llgf	%r7,8(%r2)
+	llgf	%r8,12(%r2)
+	llgf	%r9,16(%r2)
+
+	lg	%r0,0(%r10)
+	lg	%r1,8(%r10)
+
+.Lloop:
+	rllg	%r0,%r0,32
+	lg	%r12,0(%r3)	### Xload(0)
+	rllg	%r13,%r12,32
+	stg	%r12,160(%r15)
+	alr	%r9,%r0		### 0
+	rll	%r11,%r5,5
+	lr	%r10,%r8
+	xr	%r10,%r7
+	alr	%r9,%r11
+	nr	%r10,%r6
+	alr	%r9,%r13
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r0		### 1
+	rll	%r11,%r9,5
+	lr	%r10,%r7
+	xr	%r10,%r6
+	alr	%r8,%r11
+	nr	%r10,%r5
+	alr	%r8,%r12
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	lg	%r14,8(%r3)	### Xload(2)
+	rllg	%r12,%r14,32
+	stg	%r14,168(%r15)
+	alr	%r7,%r0		### 2
+	rll	%r11,%r8,5
+	lr	%r10,%r6
+	xr	%r10,%r5
+	alr	%r7,%r11
+	nr	%r10,%r9
+	alr	%r7,%r12
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r0		### 3
+	rll	%r11,%r7,5
+	lr	%r10,%r5
+	xr	%r10,%r9
+	alr	%r6,%r11
+	nr	%r10,%r8
+	alr	%r6,%r14
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	lg	%r13,16(%r3)	### Xload(4)
+	rllg	%r14,%r13,32
+	stg	%r13,176(%r15)
+	alr	%r5,%r0		### 4
+	rll	%r11,%r6,5
+	lr	%r10,%r9
+	xr	%r10,%r8
+	alr	%r5,%r11
+	nr	%r10,%r7
+	alr	%r5,%r14
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r0		### 5
+	rll	%r11,%r5,5
+	lr	%r10,%r8
+	xr	%r10,%r7
+	alr	%r9,%r11
+	nr	%r10,%r6
+	alr	%r9,%r13
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	lg	%r12,24(%r3)	### Xload(6)
+	rllg	%r13,%r12,32
+	stg	%r12,184(%r15)
+	alr	%r8,%r0		### 6
+	rll	%r11,%r9,5
+	lr	%r10,%r7
+	xr	%r10,%r6
+	alr	%r8,%r11
+	nr	%r10,%r5
+	alr	%r8,%r13
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r0		### 7
+	rll	%r11,%r8,5
+	lr	%r10,%r6
+	xr	%r10,%r5
+	alr	%r7,%r11
+	nr	%r10,%r9
+	alr	%r7,%r12
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	lg	%r14,32(%r3)	### Xload(8)
+	rllg	%r12,%r14,32
+	stg	%r14,192(%r15)
+	alr	%r6,%r0		### 8
+	rll	%r11,%r7,5
+	lr	%r10,%r5
+	xr	%r10,%r9
+	alr	%r6,%r11
+	nr	%r10,%r8
+	alr	%r6,%r12
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r0		### 9
+	rll	%r11,%r6,5
+	lr	%r10,%r9
+	xr	%r10,%r8
+	alr	%r5,%r11
+	nr	%r10,%r7
+	alr	%r5,%r14
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	lg	%r13,40(%r3)	### Xload(10)
+	rllg	%r14,%r13,32
+	stg	%r13,200(%r15)
+	alr	%r9,%r0		### 10
+	rll	%r11,%r5,5
+	lr	%r10,%r8
+	xr	%r10,%r7
+	alr	%r9,%r11
+	nr	%r10,%r6
+	alr	%r9,%r14
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r0		### 11
+	rll	%r11,%r9,5
+	lr	%r10,%r7
+	xr	%r10,%r6
+	alr	%r8,%r11
+	nr	%r10,%r5
+	alr	%r8,%r13
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	lg	%r12,48(%r3)	### Xload(12)
+	rllg	%r13,%r12,32
+	stg	%r12,208(%r15)
+	alr	%r7,%r0		### 12
+	rll	%r11,%r8,5
+	lr	%r10,%r6
+	xr	%r10,%r5
+	alr	%r7,%r11
+	nr	%r10,%r9
+	alr	%r7,%r13
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r0		### 13
+	rll	%r11,%r7,5
+	lr	%r10,%r5
+	xr	%r10,%r9
+	alr	%r6,%r11
+	nr	%r10,%r8
+	alr	%r6,%r12
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	lg	%r14,56(%r3)	### Xload(14)
+	rllg	%r12,%r14,32
+	stg	%r14,216(%r15)
+	alr	%r5,%r0		### 14
+	rll	%r11,%r6,5
+	lr	%r10,%r9
+	xr	%r10,%r8
+	alr	%r5,%r11
+	nr	%r10,%r7
+	alr	%r5,%r12
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	lg	%r2,160(%r15)	### Xupdate(16) warm-up
+	lr	%r13,%r12
+	alr	%r9,%r0		### 15
+	rll	%r11,%r5,5
+	lr	%r10,%r8
+	xr	%r10,%r7
+	alr	%r9,%r11
+	nr	%r10,%r6
+	alr	%r9,%r14
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r13,%r2		### Xupdate(16)
+	lg	%r2,168(%r15)
+	xg	%r13,192(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,160(%r15)
+	alr	%r8,%r0		### 16
+	rll	%r11,%r9,5
+	lr	%r10,%r7
+	xr	%r10,%r6
+	alr	%r8,%r11
+	nr	%r10,%r5
+	alr	%r8,%r14
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r0		### 17
+	rll	%r11,%r8,5
+	lr	%r10,%r6
+	xr	%r10,%r5
+	alr	%r7,%r11
+	nr	%r10,%r9
+	alr	%r7,%r13
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r12,%r2		### Xupdate(18)
+	lg	%r2,176(%r15)
+	xg	%r12,200(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,168(%r15)
+	alr	%r6,%r0		### 18
+	rll	%r11,%r7,5
+	lr	%r10,%r5
+	xr	%r10,%r9
+	alr	%r6,%r11
+	nr	%r10,%r8
+	alr	%r6,%r13
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r0		### 19
+	rll	%r11,%r6,5
+	lr	%r10,%r9
+	xr	%r10,%r8
+	alr	%r5,%r11
+	nr	%r10,%r7
+	alr	%r5,%r12
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	rllg	%r0,%r0,32
+	xgr	%r14,%r2		### Xupdate(20)
+	lg	%r2,184(%r15)
+	xg	%r14,208(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,176(%r15)
+	alr	%r9,%r0		### 20
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r12
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r0		### 21
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r14
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	xgr	%r13,%r2		### Xupdate(22)
+	lg	%r2,192(%r15)
+	xg	%r13,216(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,184(%r15)
+	alr	%r7,%r0		### 22
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r14
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r0		### 23
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r13
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	xgr	%r12,%r2		### Xupdate(24)
+	lg	%r2,200(%r15)
+	xg	%r12,160(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,192(%r15)
+	alr	%r5,%r0		### 24
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r13
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r0		### 25
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r12
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r14,%r2		### Xupdate(26)
+	lg	%r2,208(%r15)
+	xg	%r14,168(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,200(%r15)
+	alr	%r8,%r0		### 26
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r12
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r0		### 27
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r14
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r13,%r2		### Xupdate(28)
+	lg	%r2,216(%r15)
+	xg	%r13,176(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,208(%r15)
+	alr	%r6,%r0		### 28
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r14
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r0		### 29
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r13
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	xgr	%r12,%r2		### Xupdate(30)
+	lg	%r2,160(%r15)
+	xg	%r12,184(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,216(%r15)
+	alr	%r9,%r0		### 30
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r13
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r0		### 31
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r12
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	xgr	%r14,%r2		### Xupdate(32)
+	lg	%r2,168(%r15)
+	xg	%r14,192(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,160(%r15)
+	alr	%r7,%r0		### 32
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r12
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r0		### 33
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r14
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	xgr	%r13,%r2		### Xupdate(34)
+	lg	%r2,176(%r15)
+	xg	%r13,200(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,168(%r15)
+	alr	%r5,%r0		### 34
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r14
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r0		### 35
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r13
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r12,%r2		### Xupdate(36)
+	lg	%r2,184(%r15)
+	xg	%r12,208(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,176(%r15)
+	alr	%r8,%r0		### 36
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r13
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r0		### 37
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r12
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r14,%r2		### Xupdate(38)
+	lg	%r2,192(%r15)
+	xg	%r14,216(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,184(%r15)
+	alr	%r6,%r0		### 38
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r12
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r0		### 39
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r14
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	rllg	%r1,%r1,32
+	xgr	%r13,%r2		### Xupdate(40)
+	lg	%r2,200(%r15)
+	xg	%r13,160(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,192(%r15)
+	alr	%r9,%r1		### 40
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	or	%r10,%r7
+	lr	%r11,%r6
+	nr	%r10,%r8
+	nr	%r11,%r7
+	alr	%r9,%r14
+	or	%r10,%r11
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r1		### 41
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	or	%r10,%r6
+	lr	%r11,%r5
+	nr	%r10,%r7
+	nr	%r11,%r6
+	alr	%r8,%r13
+	or	%r10,%r11
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	xgr	%r12,%r2		### Xupdate(42)
+	lg	%r2,208(%r15)
+	xg	%r12,168(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,200(%r15)
+	alr	%r7,%r1		### 42
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	or	%r10,%r5
+	lr	%r11,%r9
+	nr	%r10,%r6
+	nr	%r11,%r5
+	alr	%r7,%r13
+	or	%r10,%r11
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r1		### 43
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	or	%r10,%r9
+	lr	%r11,%r8
+	nr	%r10,%r5
+	nr	%r11,%r9
+	alr	%r6,%r12
+	or	%r10,%r11
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	xgr	%r14,%r2		### Xupdate(44)
+	lg	%r2,216(%r15)
+	xg	%r14,176(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,208(%r15)
+	alr	%r5,%r1		### 44
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	or	%r10,%r8
+	lr	%r11,%r7
+	nr	%r10,%r9
+	nr	%r11,%r8
+	alr	%r5,%r12
+	or	%r10,%r11
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r1		### 45
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	or	%r10,%r7
+	lr	%r11,%r6
+	nr	%r10,%r8
+	nr	%r11,%r7
+	alr	%r9,%r14
+	or	%r10,%r11
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r13,%r2		### Xupdate(46)
+	lg	%r2,160(%r15)
+	xg	%r13,184(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,216(%r15)
+	alr	%r8,%r1		### 46
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	or	%r10,%r6
+	lr	%r11,%r5
+	nr	%r10,%r7
+	nr	%r11,%r6
+	alr	%r8,%r14
+	or	%r10,%r11
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r1		### 47
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	or	%r10,%r5
+	lr	%r11,%r9
+	nr	%r10,%r6
+	nr	%r11,%r5
+	alr	%r7,%r13
+	or	%r10,%r11
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r12,%r2		### Xupdate(48)
+	lg	%r2,168(%r15)
+	xg	%r12,192(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,160(%r15)
+	alr	%r6,%r1		### 48
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	or	%r10,%r9
+	lr	%r11,%r8
+	nr	%r10,%r5
+	nr	%r11,%r9
+	alr	%r6,%r13
+	or	%r10,%r11
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r1		### 49
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	or	%r10,%r8
+	lr	%r11,%r7
+	nr	%r10,%r9
+	nr	%r11,%r8
+	alr	%r5,%r12
+	or	%r10,%r11
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	xgr	%r14,%r2		### Xupdate(50)
+	lg	%r2,176(%r15)
+	xg	%r14,200(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,168(%r15)
+	alr	%r9,%r1		### 50
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	or	%r10,%r7
+	lr	%r11,%r6
+	nr	%r10,%r8
+	nr	%r11,%r7
+	alr	%r9,%r12
+	or	%r10,%r11
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r1		### 51
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	or	%r10,%r6
+	lr	%r11,%r5
+	nr	%r10,%r7
+	nr	%r11,%r6
+	alr	%r8,%r14
+	or	%r10,%r11
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	xgr	%r13,%r2		### Xupdate(52)
+	lg	%r2,184(%r15)
+	xg	%r13,208(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,176(%r15)
+	alr	%r7,%r1		### 52
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	or	%r10,%r5
+	lr	%r11,%r9
+	nr	%r10,%r6
+	nr	%r11,%r5
+	alr	%r7,%r14
+	or	%r10,%r11
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r1		### 53
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	or	%r10,%r9
+	lr	%r11,%r8
+	nr	%r10,%r5
+	nr	%r11,%r9
+	alr	%r6,%r13
+	or	%r10,%r11
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	xgr	%r12,%r2		### Xupdate(54)
+	lg	%r2,192(%r15)
+	xg	%r12,216(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,184(%r15)
+	alr	%r5,%r1		### 54
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	or	%r10,%r8
+	lr	%r11,%r7
+	nr	%r10,%r9
+	nr	%r11,%r8
+	alr	%r5,%r13
+	or	%r10,%r11
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r1		### 55
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	or	%r10,%r7
+	lr	%r11,%r6
+	nr	%r10,%r8
+	nr	%r11,%r7
+	alr	%r9,%r12
+	or	%r10,%r11
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r14,%r2		### Xupdate(56)
+	lg	%r2,200(%r15)
+	xg	%r14,160(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,192(%r15)
+	alr	%r8,%r1		### 56
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	or	%r10,%r6
+	lr	%r11,%r5
+	nr	%r10,%r7
+	nr	%r11,%r6
+	alr	%r8,%r12
+	or	%r10,%r11
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r1		### 57
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	or	%r10,%r5
+	lr	%r11,%r9
+	nr	%r10,%r6
+	nr	%r11,%r5
+	alr	%r7,%r14
+	or	%r10,%r11
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r13,%r2		### Xupdate(58)
+	lg	%r2,208(%r15)
+	xg	%r13,168(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,200(%r15)
+	alr	%r6,%r1		### 58
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	or	%r10,%r9
+	lr	%r11,%r8
+	nr	%r10,%r5
+	nr	%r11,%r9
+	alr	%r6,%r14
+	or	%r10,%r11
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r1		### 59
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	or	%r10,%r8
+	lr	%r11,%r7
+	nr	%r10,%r9
+	nr	%r11,%r8
+	alr	%r5,%r13
+	or	%r10,%r11
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	rllg	%r1,%r1,32
+	xgr	%r12,%r2		### Xupdate(60)
+	lg	%r2,216(%r15)
+	xg	%r12,176(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,208(%r15)
+	alr	%r9,%r1		### 60
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r13
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r1		### 61
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r12
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	xgr	%r14,%r2		### Xupdate(62)
+	lg	%r2,160(%r15)
+	xg	%r14,184(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,216(%r15)
+	alr	%r7,%r1		### 62
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r12
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r1		### 63
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r14
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	xgr	%r13,%r2		### Xupdate(64)
+	lg	%r2,168(%r15)
+	xg	%r13,192(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,160(%r15)
+	alr	%r5,%r1		### 64
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r14
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r1		### 65
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r13
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r12,%r2		### Xupdate(66)
+	lg	%r2,176(%r15)
+	xg	%r12,200(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	stg	%r12,168(%r15)
+	alr	%r8,%r1		### 66
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r13
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r1		### 67
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r12
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r14,%r2		### Xupdate(68)
+	lg	%r2,184(%r15)
+	xg	%r14,208(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	stg	%r14,176(%r15)
+	alr	%r6,%r1		### 68
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r12
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r1		### 69
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r14
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	xgr	%r13,%r2		### Xupdate(70)
+	lg	%r2,192(%r15)
+	xg	%r13,216(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	stg	%r13,184(%r15)
+	alr	%r9,%r1		### 70
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r14
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	alr	%r8,%r1		### 71
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r13
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	xgr	%r12,%r2		### Xupdate(72)
+	lg	%r2,200(%r15)
+	xg	%r12,160(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	alr	%r7,%r1		### 72
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r13
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	alr	%r6,%r1		### 73
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r12
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	xgr	%r14,%r2		### Xupdate(74)
+	lg	%r2,208(%r15)
+	xg	%r14,168(%r15)
+	xgr	%r14,%r2
+	rll	%r14,%r14,1
+	rllg	%r12,%r14,32
+	rll	%r12,%r12,1
+	rllg	%r14,%r12,32
+	lr	%r13,%r12		# feedback
+	alr	%r5,%r1		### 74
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r12
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+	alr	%r9,%r1		### 75
+	rll	%r11,%r5,5
+	lr	%r10,%r6
+	alr	%r9,%r11
+	xr	%r10,%r7
+	alr	%r9,%r14
+	xr	%r10,%r8
+	rll	%r6,%r6,30
+	alr	%r9,%r10
+	xgr	%r13,%r2		### Xupdate(76)
+	lg	%r2,216(%r15)
+	xg	%r13,176(%r15)
+	xgr	%r13,%r2
+	rll	%r13,%r13,1
+	rllg	%r14,%r13,32
+	rll	%r14,%r14,1
+	rllg	%r13,%r14,32
+	lr	%r12,%r14		# feedback
+	alr	%r8,%r1		### 76
+	rll	%r11,%r9,5
+	lr	%r10,%r5
+	alr	%r8,%r11
+	xr	%r10,%r6
+	alr	%r8,%r14
+	xr	%r10,%r7
+	rll	%r5,%r5,30
+	alr	%r8,%r10
+	alr	%r7,%r1		### 77
+	rll	%r11,%r8,5
+	lr	%r10,%r9
+	alr	%r7,%r11
+	xr	%r10,%r5
+	alr	%r7,%r13
+	xr	%r10,%r6
+	rll	%r9,%r9,30
+	alr	%r7,%r10
+	xgr	%r12,%r2		### Xupdate(78)
+	lg	%r2,160(%r15)
+	xg	%r12,184(%r15)
+	xgr	%r12,%r2
+	rll	%r12,%r12,1
+	rllg	%r13,%r12,32
+	rll	%r13,%r13,1
+	rllg	%r12,%r13,32
+	lr	%r14,%r13		# feedback
+	alr	%r6,%r1		### 78
+	rll	%r11,%r7,5
+	lr	%r10,%r8
+	alr	%r6,%r11
+	xr	%r10,%r9
+	alr	%r6,%r13
+	xr	%r10,%r5
+	rll	%r8,%r8,30
+	alr	%r6,%r10
+	alr	%r5,%r1		### 79
+	rll	%r11,%r6,5
+	lr	%r10,%r7
+	alr	%r5,%r11
+	xr	%r10,%r8
+	alr	%r5,%r12
+	xr	%r10,%r9
+	rll	%r7,%r7,30
+	alr	%r5,%r10
+
+	lg	%r2,240(%r15)
+	la	%r3,64(%r3)
+	al	%r5,0(%r2)
+	al	%r6,4(%r2)
+	al	%r7,8(%r2)
+	al	%r8,12(%r2)
+	al	%r9,16(%r2)
+	st	%r5,0(%r2)
+	st	%r6,4(%r2)
+	st	%r7,8(%r2)
+	st	%r8,12(%r2)
+	st	%r9,16(%r2)
+	brctg %r4,.Lloop
+
+	lmg	%r6,%r15,272(%r15)
+	br	%r14
+.size	sha1_block_data_order,.-sha1_block_data_order
+.string	"SHA1 block transform for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha256-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha256-s390x.S
new file mode 100644
index 0000000000..45d3601ac2
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha256-s390x.S
@@ -0,0 +1,1242 @@
+#include "s390x_arch.h"
+
+.text
+.align	64
+.type	K256,@object
+K256:
+	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
+	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
+	.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
+	.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
+	.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
+	.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
+	.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
+	.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
+	.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
+	.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
+	.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
+	.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
+	.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
+	.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
+	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
+	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
+.size	K256,.-K256
+.globl	sha256_block_data_order
+.type	sha256_block_data_order,@function
+sha256_block_data_order:
+	sllg	%r4,%r4,6
+	larl	%r1,OPENSSL_s390xcap_P
+	lg	%r0,S390X_KIMD(%r1)	# check kimd capabilities
+	tmhh	%r0,8192
+	jz	.Lsoftware
+	lghi	%r0,2
+	lgr	%r1,%r2
+	lgr	%r2,%r3
+	lgr	%r3,%r4
+	.long	0xb93e0002	# kimd %r0,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+	br	%r14
+.align	16
+.Lsoftware:
+	lghi	%r1,-224
+	la	%r4,0(%r4,%r3)
+	stmg	%r2,%r15,16(%r15)
+	lgr	%r0,%r15
+	la	%r15,0(%r1,%r15)
+	stg	%r0,0(%r15)
+
+	larl	%r13,K256
+	llgf	%r5,0(%r2)
+	llgf	%r6,4(%r2)
+	llgf	%r7,8(%r2)
+	llgf	%r8,12(%r2)
+	llgf	%r9,16(%r2)
+	llgf	%r10,20(%r2)
+	llgf	%r11,24(%r2)
+	llgf	%r12,28(%r2)
+
+.Lloop:
+	lghi	%r4,0
+	llgf	%r14,0(%r3)	### 0
+	rll	%r0,%r9,7
+	rll	%r1,%r9,21
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r11
+	st	%r14,160(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r12,%r5,10
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	al	%r14,0(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r5,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rll	%r0,%r0,11
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,4(%r3)	### 1
+	rll	%r0,%r8,7
+	rll	%r1,%r8,21
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r10
+	st	%r14,164(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r11,%r12,10
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	al	%r14,4(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r12,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rll	%r0,%r0,11
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,8(%r3)	### 2
+	rll	%r0,%r7,7
+	rll	%r1,%r7,21
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r9
+	st	%r14,168(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r10,%r11,10
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	al	%r14,8(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r11,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rll	%r0,%r0,11
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,12(%r3)	### 3
+	rll	%r0,%r6,7
+	rll	%r1,%r6,21
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r8
+	st	%r14,172(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r9,%r10,10
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	al	%r14,12(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r10,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rll	%r0,%r0,11
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,16(%r3)	### 4
+	rll	%r0,%r5,7
+	rll	%r1,%r5,21
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r7
+	st	%r14,176(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r8,%r9,10
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	al	%r14,16(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r9,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rll	%r0,%r0,11
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,20(%r3)	### 5
+	rll	%r0,%r12,7
+	rll	%r1,%r12,21
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r6
+	st	%r14,180(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r7,%r8,10
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	al	%r14,20(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r8,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rll	%r0,%r0,11
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,24(%r3)	### 6
+	rll	%r0,%r11,7
+	rll	%r1,%r11,21
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r5
+	st	%r14,184(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r6,%r7,10
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	al	%r14,24(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r7,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rll	%r0,%r0,11
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,28(%r3)	### 7
+	rll	%r0,%r10,7
+	rll	%r1,%r10,21
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r12
+	st	%r14,188(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r5,%r6,10
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	al	%r14,28(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r6,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rll	%r0,%r0,11
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,32(%r3)	### 8
+	rll	%r0,%r9,7
+	rll	%r1,%r9,21
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r11
+	st	%r14,192(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r12,%r5,10
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	al	%r14,32(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r5,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rll	%r0,%r0,11
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,36(%r3)	### 9
+	rll	%r0,%r8,7
+	rll	%r1,%r8,21
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r10
+	st	%r14,196(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r11,%r12,10
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	al	%r14,36(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r12,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rll	%r0,%r0,11
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,40(%r3)	### 10
+	rll	%r0,%r7,7
+	rll	%r1,%r7,21
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r9
+	st	%r14,200(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r10,%r11,10
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	al	%r14,40(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r11,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rll	%r0,%r0,11
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,44(%r3)	### 11
+	rll	%r0,%r6,7
+	rll	%r1,%r6,21
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r8
+	st	%r14,204(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r9,%r10,10
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	al	%r14,44(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r10,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rll	%r0,%r0,11
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,48(%r3)	### 12
+	rll	%r0,%r5,7
+	rll	%r1,%r5,21
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r7
+	st	%r14,208(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r8,%r9,10
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	al	%r14,48(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r9,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rll	%r0,%r0,11
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,52(%r3)	### 13
+	rll	%r0,%r12,7
+	rll	%r1,%r12,21
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r6
+	st	%r14,212(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r7,%r8,10
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	al	%r14,52(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r8,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rll	%r0,%r0,11
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,56(%r3)	### 14
+	rll	%r0,%r11,7
+	rll	%r1,%r11,21
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r5
+	st	%r14,216(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r6,%r7,10
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	al	%r14,56(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r7,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rll	%r0,%r0,11
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,60(%r3)	### 15
+	rll	%r0,%r10,7
+	rll	%r1,%r10,21
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r12
+	st	%r14,220(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r5,%r6,10
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	al	%r14,60(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r6,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rll	%r0,%r0,11
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+.Lrounds_16_xx:
+	llgf	%r14,164(%r15)	### 16
+	llgf	%r1,216(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,160(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,196(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r9,7
+	rll	%r1,%r9,21
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r11
+	st	%r14,160(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r12,%r5,10
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	al	%r14,64(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r5,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rll	%r0,%r0,11
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,168(%r15)	### 17
+	llgf	%r1,220(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,164(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,200(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r8,7
+	rll	%r1,%r8,21
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r10
+	st	%r14,164(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r11,%r12,10
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	al	%r14,68(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r12,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rll	%r0,%r0,11
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,172(%r15)	### 18
+	llgf	%r1,160(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,168(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,204(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r7,7
+	rll	%r1,%r7,21
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r9
+	st	%r14,168(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r10,%r11,10
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	al	%r14,72(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r11,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rll	%r0,%r0,11
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,176(%r15)	### 19
+	llgf	%r1,164(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,172(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,208(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r6,7
+	rll	%r1,%r6,21
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r8
+	st	%r14,172(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r9,%r10,10
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	al	%r14,76(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r10,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rll	%r0,%r0,11
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,180(%r15)	### 20
+	llgf	%r1,168(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,176(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,212(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r5,7
+	rll	%r1,%r5,21
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r7
+	st	%r14,176(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r8,%r9,10
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	al	%r14,80(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r9,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rll	%r0,%r0,11
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,184(%r15)	### 21
+	llgf	%r1,172(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,180(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,216(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r12,7
+	rll	%r1,%r12,21
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r6
+	st	%r14,180(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r7,%r8,10
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	al	%r14,84(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r8,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rll	%r0,%r0,11
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,188(%r15)	### 22
+	llgf	%r1,176(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,184(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,220(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r11,7
+	rll	%r1,%r11,21
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r5
+	st	%r14,184(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r6,%r7,10
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	al	%r14,88(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r7,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rll	%r0,%r0,11
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,192(%r15)	### 23
+	llgf	%r1,180(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,188(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,160(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r10,7
+	rll	%r1,%r10,21
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r12
+	st	%r14,188(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r5,%r6,10
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	al	%r14,92(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r6,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rll	%r0,%r0,11
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,196(%r15)	### 24
+	llgf	%r1,184(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,192(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,164(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r9,7
+	rll	%r1,%r9,21
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r11
+	st	%r14,192(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r12,%r5,10
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	al	%r14,96(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r5,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rll	%r0,%r0,11
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,200(%r15)	### 25
+	llgf	%r1,188(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,196(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,168(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r8,7
+	rll	%r1,%r8,21
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r10
+	st	%r14,196(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r11,%r12,10
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	al	%r14,100(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r12,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rll	%r0,%r0,11
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,204(%r15)	### 26
+	llgf	%r1,192(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,200(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,172(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r7,7
+	rll	%r1,%r7,21
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r9
+	st	%r14,200(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r10,%r11,10
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	al	%r14,104(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r11,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rll	%r0,%r0,11
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,208(%r15)	### 27
+	llgf	%r1,196(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,204(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,176(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r6,7
+	rll	%r1,%r6,21
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r8
+	st	%r14,204(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r9,%r10,10
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	al	%r14,108(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r10,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rll	%r0,%r0,11
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,212(%r15)	### 28
+	llgf	%r1,200(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,208(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,180(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r5,7
+	rll	%r1,%r5,21
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r7
+	st	%r14,208(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r8,%r9,10
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	al	%r14,112(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r9,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rll	%r0,%r0,11
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,216(%r15)	### 29
+	llgf	%r1,204(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,212(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,184(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r12,7
+	rll	%r1,%r12,21
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r6
+	st	%r14,212(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r7,%r8,10
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	al	%r14,116(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r8,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rll	%r0,%r0,11
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,220(%r15)	### 30
+	llgf	%r1,208(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,216(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,188(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r11,7
+	rll	%r1,%r11,21
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r5
+	st	%r14,216(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r6,%r7,10
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	al	%r14,120(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r7,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rll	%r0,%r0,11
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	llgf	%r14,160(%r15)	### 31
+	llgf	%r1,212(%r15)
+	rll	%r0,%r14,14
+	srl	%r14,3
+	rll	%r2,%r0,11
+	xgr	%r14,%r0
+	rll	%r0,%r1,13
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srl	%r1,10
+	al	%r14,220(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rll	%r0,%r0,2
+	al	%r14,192(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rll	%r0,%r10,7
+	rll	%r1,%r10,21
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rll	%r1,%r1,5
+	 xgr	%r2,%r12
+	st	%r14,220(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rll	%r5,%r6,10
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	al	%r14,124(%r4,%r13)	# T1+=K[i]
+	rll	%r0,%r6,19
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rll	%r0,%r0,11
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+	aghi	%r4,64
+	lghi	%r0,192
+	clgr	%r4,%r0
+	jne	.Lrounds_16_xx
+
+	lg	%r2,240(%r15)
+	la	%r3,64(%r3)
+	al	%r5,0(%r2)
+	al	%r6,4(%r2)
+	al	%r7,8(%r2)
+	al	%r8,12(%r2)
+	al	%r9,16(%r2)
+	al	%r10,20(%r2)
+	al	%r11,24(%r2)
+	al	%r12,28(%r2)
+	st	%r5,0(%r2)
+	st	%r6,4(%r2)
+	st	%r7,8(%r2)
+	st	%r8,12(%r2)
+	st	%r9,16(%r2)
+	st	%r10,20(%r2)
+	st	%r11,24(%r2)
+	st	%r12,28(%r2)
+	clg	%r3,256(%r15)
+	jne	.Lloop
+
+	lmg	%r6,%r15,272(%r15)
+	br	%r14
+.size	sha256_block_data_order,.-sha256_block_data_order
+.string	"SHA256 block transform for s390x, CRYPTOGAMS by <appro@openssl.org>"
diff --git a/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha512-s390x.S b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha512-s390x.S
new file mode 100644
index 0000000000..79e4277351
--- /dev/null
+++ b/deps/openssl/config/archs/linux64-s390x/asm_avx2/crypto/sha/sha512-s390x.S
@@ -0,0 +1,1266 @@
+#include "s390x_arch.h"
+
+.text
+.align	64
+.type	K512,@object
+K512:
+	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
+	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
+	.quad	0x3956c25bf348b538,0x59f111f1b605d019
+	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
+	.quad	0xd807aa98a3030242,0x12835b0145706fbe
+	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
+	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
+	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
+	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
+	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
+	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
+	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
+	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
+	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
+	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
+	.quad	0x06ca6351e003826f,0x142929670a0e6e70
+	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
+	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
+	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
+	.quad	0x81c2c92e47edaee6,0x92722c851482353b
+	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
+	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
+	.quad	0xd192e819d6ef5218,0xd69906245565a910
+	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
+	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
+	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
+	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
+	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
+	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
+	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
+	.quad	0x90befffa23631e28,0xa4506cebde82bde9
+	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
+	.quad	0xca273eceea26619c,0xd186b8c721c0c207
+	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
+	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
+	.quad	0x113f9804bef90dae,0x1b710b35131c471b
+	.quad	0x28db77f523047d84,0x32caab7b40c72493
+	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
+	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
+	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
+.size	K512,.-K512
+.globl	sha512_block_data_order
+.type	sha512_block_data_order,@function
+sha512_block_data_order:
+	sllg	%r4,%r4,7
+	larl	%r1,OPENSSL_s390xcap_P
+	lg	%r0,S390X_KIMD(%r1)	# check kimd capabilities
+	tmhh	%r0,4096
+	jz	.Lsoftware
+	lghi	%r0,3
+	lgr	%r1,%r2
+	lgr	%r2,%r3
+	lgr	%r3,%r4
+	.long	0xb93e0002	# kimd %r0,%r2
+	brc	1,.-4		# pay attention to "partial completion"
+	br	%r14
+.align	16
+.Lsoftware:
+	lghi	%r1,-288
+	la	%r4,0(%r4,%r3)
+	stmg	%r2,%r15,16(%r15)
+	lgr	%r0,%r15
+	la	%r15,0(%r1,%r15)
+	stg	%r0,0(%r15)
+
+	larl	%r13,K512
+	lg	%r5,0(%r2)
+	lg	%r6,8(%r2)
+	lg	%r7,16(%r2)
+	lg	%r8,24(%r2)
+	lg	%r9,32(%r2)
+	lg	%r10,40(%r2)
+	lg	%r11,48(%r2)
+	lg	%r12,56(%r2)
+
+.Lloop:
+	lghi	%r4,0
+	lg	%r14,0(%r3)	### 0
+	rllg	%r0,%r9,23
+	rllg	%r1,%r9,46
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r11
+	stg	%r14,160(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r12,%r5,25
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	alg	%r14,0(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r5,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rllg	%r0,%r0,6
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	lg	%r14,8(%r3)	### 1
+	rllg	%r0,%r8,23
+	rllg	%r1,%r8,46
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r10
+	stg	%r14,168(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r11,%r12,25
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	alg	%r14,8(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r12,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rllg	%r0,%r0,6
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	lg	%r14,16(%r3)	### 2
+	rllg	%r0,%r7,23
+	rllg	%r1,%r7,46
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r9
+	stg	%r14,176(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r10,%r11,25
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	alg	%r14,16(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r11,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rllg	%r0,%r0,6
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	lg	%r14,24(%r3)	### 3
+	rllg	%r0,%r6,23
+	rllg	%r1,%r6,46
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r8
+	stg	%r14,184(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r9,%r10,25
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	alg	%r14,24(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r10,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rllg	%r0,%r0,6
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	lg	%r14,32(%r3)	### 4
+	rllg	%r0,%r5,23
+	rllg	%r1,%r5,46
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r7
+	stg	%r14,192(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r8,%r9,25
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	alg	%r14,32(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r9,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rllg	%r0,%r0,6
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	lg	%r14,40(%r3)	### 5
+	rllg	%r0,%r12,23
+	rllg	%r1,%r12,46
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r6
+	stg	%r14,200(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r7,%r8,25
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	alg	%r14,40(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r8,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rllg	%r0,%r0,6
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	lg	%r14,48(%r3)	### 6
+	rllg	%r0,%r11,23
+	rllg	%r1,%r11,46
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r5
+	stg	%r14,208(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r6,%r7,25
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	alg	%r14,48(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r7,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rllg	%r0,%r0,6
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	lg	%r14,56(%r3)	### 7
+	rllg	%r0,%r10,23
+	rllg	%r1,%r10,46
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r12
+	stg	%r14,216(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r5,%r6,25
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	alg	%r14,56(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r6,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rllg	%r0,%r0,6
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+	lg	%r14,64(%r3)	### 8
+	rllg	%r0,%r9,23
+	rllg	%r1,%r9,46
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r11
+	stg	%r14,224(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r12,%r5,25
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	alg	%r14,64(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r5,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rllg	%r0,%r0,6
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	lg	%r14,72(%r3)	### 9
+	rllg	%r0,%r8,23
+	rllg	%r1,%r8,46
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r10
+	stg	%r14,232(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r11,%r12,25
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	alg	%r14,72(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r12,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rllg	%r0,%r0,6
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	lg	%r14,80(%r3)	### 10
+	rllg	%r0,%r7,23
+	rllg	%r1,%r7,46
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r9
+	stg	%r14,240(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r10,%r11,25
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	alg	%r14,80(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r11,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rllg	%r0,%r0,6
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	lg	%r14,88(%r3)	### 11
+	rllg	%r0,%r6,23
+	rllg	%r1,%r6,46
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r8
+	stg	%r14,248(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r9,%r10,25
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	alg	%r14,88(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r10,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rllg	%r0,%r0,6
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	lg	%r14,96(%r3)	### 12
+	rllg	%r0,%r5,23
+	rllg	%r1,%r5,46
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r7
+	stg	%r14,256(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r8,%r9,25
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	alg	%r14,96(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r9,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rllg	%r0,%r0,6
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	lg	%r14,104(%r3)	### 13
+	rllg	%r0,%r12,23
+	rllg	%r1,%r12,46
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r6
+	stg	%r14,264(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r7,%r8,25
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	alg	%r14,104(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r8,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rllg	%r0,%r0,6
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	lg	%r14,112(%r3)	### 14
+	rllg	%r0,%r11,23
+	rllg	%r1,%r11,46
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r5
+	stg	%r14,272(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r6,%r7,25
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	alg	%r14,112(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r7,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rllg	%r0,%r0,6
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	lg	%r14,120(%r3)	### 15
+	rllg	%r0,%r10,23
+	rllg	%r1,%r10,46
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r12
+	stg	%r14,280(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r5,%r6,25
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	alg	%r14,120(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r6,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rllg	%r0,%r0,6
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+.Lrounds_16_xx:
+	lg	%r14,168(%r15)	### 16
+	lg	%r1,272(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,160(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,232(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r9,23
+	rllg	%r1,%r9,46
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r11
+	stg	%r14,160(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r12,%r5,25
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	alg	%r14,128(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r5,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rllg	%r0,%r0,6
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	lg	%r14,176(%r15)	### 17
+	lg	%r1,280(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,168(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,240(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r8,23
+	rllg	%r1,%r8,46
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r10
+	stg	%r14,168(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r11,%r12,25
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	alg	%r14,136(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r12,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rllg	%r0,%r0,6
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	lg	%r14,184(%r15)	### 18
+	lg	%r1,160(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,176(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,248(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r7,23
+	rllg	%r1,%r7,46
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r9
+	stg	%r14,176(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r10,%r11,25
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	alg	%r14,144(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r11,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rllg	%r0,%r0,6
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	lg	%r14,192(%r15)	### 19
+	lg	%r1,168(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,184(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,256(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r6,23
+	rllg	%r1,%r6,46
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r8
+	stg	%r14,184(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r9,%r10,25
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	alg	%r14,152(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r10,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rllg	%r0,%r0,6
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	lg	%r14,200(%r15)	### 20
+	lg	%r1,176(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,192(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,264(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r5,23
+	rllg	%r1,%r5,46
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r7
+	stg	%r14,192(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r8,%r9,25
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	alg	%r14,160(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r9,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rllg	%r0,%r0,6
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	lg	%r14,208(%r15)	### 21
+	lg	%r1,184(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,200(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,272(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r12,23
+	rllg	%r1,%r12,46
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r6
+	stg	%r14,200(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r7,%r8,25
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	alg	%r14,168(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r8,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rllg	%r0,%r0,6
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	lg	%r14,216(%r15)	### 22
+	lg	%r1,192(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,208(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,280(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r11,23
+	rllg	%r1,%r11,46
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r5
+	stg	%r14,208(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r6,%r7,25
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	alg	%r14,176(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r7,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rllg	%r0,%r0,6
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	lg	%r14,224(%r15)	### 23
+	lg	%r1,200(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,216(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,160(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r10,23
+	rllg	%r1,%r10,46
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r12
+	stg	%r14,216(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r5,%r6,25
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	alg	%r14,184(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r6,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rllg	%r0,%r0,6
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+	lg	%r14,232(%r15)	### 24
+	lg	%r1,208(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,224(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,168(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r9,23
+	rllg	%r1,%r9,46
+	 lgr	%r2,%r10
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r11
+	stg	%r14,224(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r12			# T1+=h
+	 ngr	%r2,%r9
+	 lgr	%r1,%r5
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r12,%r5,25
+	 xgr	%r2,%r11			# Ch(e,f,g)
+	alg	%r14,192(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r5,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r6
+	xgr	%r12,%r0
+	 lgr	%r2,%r5
+	 ngr	%r1,%r7
+	rllg	%r0,%r0,6
+	xgr	%r12,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r6
+	algr	%r12,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r8,%r14			# d+=T1
+	algr	%r12,%r2			# h+=Maj(a,b,c)
+	lg	%r14,240(%r15)	### 25
+	lg	%r1,216(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,232(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,176(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r8,23
+	rllg	%r1,%r8,46
+	 lgr	%r2,%r9
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r10
+	stg	%r14,232(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r11			# T1+=h
+	 ngr	%r2,%r8
+	 lgr	%r1,%r12
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r11,%r12,25
+	 xgr	%r2,%r10			# Ch(e,f,g)
+	alg	%r14,200(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r12,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r5
+	xgr	%r11,%r0
+	 lgr	%r2,%r12
+	 ngr	%r1,%r6
+	rllg	%r0,%r0,6
+	xgr	%r11,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r5
+	algr	%r11,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r7,%r14			# d+=T1
+	algr	%r11,%r2			# h+=Maj(a,b,c)
+	lg	%r14,248(%r15)	### 26
+	lg	%r1,224(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,240(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,184(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r7,23
+	rllg	%r1,%r7,46
+	 lgr	%r2,%r8
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r9
+	stg	%r14,240(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r10			# T1+=h
+	 ngr	%r2,%r7
+	 lgr	%r1,%r11
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r10,%r11,25
+	 xgr	%r2,%r9			# Ch(e,f,g)
+	alg	%r14,208(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r11,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r12
+	xgr	%r10,%r0
+	 lgr	%r2,%r11
+	 ngr	%r1,%r5
+	rllg	%r0,%r0,6
+	xgr	%r10,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r12
+	algr	%r10,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r6,%r14			# d+=T1
+	algr	%r10,%r2			# h+=Maj(a,b,c)
+	lg	%r14,256(%r15)	### 27
+	lg	%r1,232(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,248(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,192(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r6,23
+	rllg	%r1,%r6,46
+	 lgr	%r2,%r7
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r8
+	stg	%r14,248(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r9			# T1+=h
+	 ngr	%r2,%r6
+	 lgr	%r1,%r10
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r9,%r10,25
+	 xgr	%r2,%r8			# Ch(e,f,g)
+	alg	%r14,216(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r10,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r11
+	xgr	%r9,%r0
+	 lgr	%r2,%r10
+	 ngr	%r1,%r12
+	rllg	%r0,%r0,6
+	xgr	%r9,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r11
+	algr	%r9,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r5,%r14			# d+=T1
+	algr	%r9,%r2			# h+=Maj(a,b,c)
+	lg	%r14,264(%r15)	### 28
+	lg	%r1,240(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,256(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,200(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r5,23
+	rllg	%r1,%r5,46
+	 lgr	%r2,%r6
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r7
+	stg	%r14,256(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r8			# T1+=h
+	 ngr	%r2,%r5
+	 lgr	%r1,%r9
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r8,%r9,25
+	 xgr	%r2,%r7			# Ch(e,f,g)
+	alg	%r14,224(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r9,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r10
+	xgr	%r8,%r0
+	 lgr	%r2,%r9
+	 ngr	%r1,%r11
+	rllg	%r0,%r0,6
+	xgr	%r8,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r10
+	algr	%r8,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r12,%r14			# d+=T1
+	algr	%r8,%r2			# h+=Maj(a,b,c)
+	lg	%r14,272(%r15)	### 29
+	lg	%r1,248(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,264(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,208(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r12,23
+	rllg	%r1,%r12,46
+	 lgr	%r2,%r5
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r6
+	stg	%r14,264(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r7			# T1+=h
+	 ngr	%r2,%r12
+	 lgr	%r1,%r8
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r7,%r8,25
+	 xgr	%r2,%r6			# Ch(e,f,g)
+	alg	%r14,232(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r8,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r9
+	xgr	%r7,%r0
+	 lgr	%r2,%r8
+	 ngr	%r1,%r10
+	rllg	%r0,%r0,6
+	xgr	%r7,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r9
+	algr	%r7,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r11,%r14			# d+=T1
+	algr	%r7,%r2			# h+=Maj(a,b,c)
+	lg	%r14,280(%r15)	### 30
+	lg	%r1,256(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,272(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,216(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r11,23
+	rllg	%r1,%r11,46
+	 lgr	%r2,%r12
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r5
+	stg	%r14,272(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r6			# T1+=h
+	 ngr	%r2,%r11
+	 lgr	%r1,%r7
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r6,%r7,25
+	 xgr	%r2,%r5			# Ch(e,f,g)
+	alg	%r14,240(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r7,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r8
+	xgr	%r6,%r0
+	 lgr	%r2,%r7
+	 ngr	%r1,%r9
+	rllg	%r0,%r0,6
+	xgr	%r6,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r8
+	algr	%r6,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r10,%r14			# d+=T1
+	algr	%r6,%r2			# h+=Maj(a,b,c)
+	lg	%r14,160(%r15)	### 31
+	lg	%r1,264(%r15)
+	rllg	%r0,%r14,56
+	srlg	%r14,%r14,7
+	rllg	%r2,%r0,7
+	xgr	%r14,%r0
+	rllg	%r0,%r1,3
+	xgr	%r14,%r2					# sigma0(X[i+1])
+	srlg	%r1,%r1,6
+	alg	%r14,280(%r15)	# +=X[i]
+	xgr	%r1,%r0
+	rllg	%r0,%r0,42
+	alg	%r14,224(%r15)	# +=X[i+9]
+	xgr	%r1,%r0				# sigma1(X[i+14])
+	algr	%r14,%r1				# +=sigma1(X[i+14])
+	rllg	%r0,%r10,23
+	rllg	%r1,%r10,46
+	 lgr	%r2,%r11
+	xgr	%r0,%r1
+	rllg	%r1,%r1,4
+	 xgr	%r2,%r12
+	stg	%r14,280(%r15)
+	xgr	%r0,%r1			# Sigma1(e)
+	algr	%r14,%r5			# T1+=h
+	 ngr	%r2,%r10
+	 lgr	%r1,%r6
+	algr	%r14,%r0			# T1+=Sigma1(e)
+	rllg	%r5,%r6,25
+	 xgr	%r2,%r12			# Ch(e,f,g)
+	alg	%r14,248(%r4,%r13)	# T1+=K[i]
+	rllg	%r0,%r6,30
+	algr	%r14,%r2			# T1+=Ch(e,f,g)
+	 ogr	%r1,%r7
+	xgr	%r5,%r0
+	 lgr	%r2,%r6
+	 ngr	%r1,%r8
+	rllg	%r0,%r0,6
+	xgr	%r5,%r0			# h=Sigma0(a)
+	 ngr	%r2,%r7
+	algr	%r5,%r14			# h+=T1
+	 ogr	%r2,%r1			# Maj(a,b,c)
+	algr	%r9,%r14			# d+=T1
+	algr	%r5,%r2			# h+=Maj(a,b,c)
+	aghi	%r4,128
+	lghi	%r0,512
+	clgr	%r4,%r0
+	jne	.Lrounds_16_xx
+
+	lg	%r2,304(%r15)
+	la	%r3,128(%r3)
+	alg	%r5,0(%r2)
+	alg	%r6,8(%r2)
+	alg	%r7,16(%r2)
+	alg	%r8,24(%r2)
+	alg	%r9,32(%r2)
+	alg	%r10,40(%r2)
+	alg	%r11,48(%r2)
+	alg	%r12,56(%r2)
+	stg	%r5,0(%r2)
+	stg	%r6,8(%r2)
+	stg	%r7,16(%r2)
+	stg	%r8,24(%r2)
+	stg	%r9,32(%r2)
+	stg	%r10,40(%r2)
+	stg	%r11,48(%r2)
+	stg	%r12,56(%r2)
+	clg	%r3,320(%r15)
+	jne	.Lloop
+
+	lmg	%r6,%r15,336(%r15)
+	br	%r14
+.size	sha512_block_data_order,.-sha512_block_data_order
+.string	"SHA512 block transform for s390x, CRYPTOGAMS by <appro@openssl.org>"