• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#! /usr/bin/env perl
2# Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16
17# AES for ARMv4
18
19# January 2007.
20#
21# Code uses single 1K S-box and is >2 times faster than code generated
22# by gcc-3.4.1. This is thanks to unique feature of ARMv4 ISA, which
23# allows to merge logical or arithmetic operation with shift or rotate
24# in one instruction and emit combined result every cycle. The module
25# is endian-neutral. The performance is ~42 cycles/byte for 128-bit
26# key [on single-issue Xscale PXA250 core].
27
28# May 2007.
29#
30# AES_set_[en|de]crypt_key is added.
31
32# July 2010.
33#
34# Rescheduling for dual-issue pipeline resulted in 12% improvement on
35# Cortex A8 core and ~25 cycles per byte processed with 128-bit key.
36
37# February 2011.
38#
39# Profiler-assisted and platform-specific optimization resulted in 16%
40# improvement on Cortex A8 core and ~21.5 cycles per byte.
41
42$flavour = shift;
43if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
44else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} }
45
46if ($flavour && $flavour ne "void") {
47    $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
48    ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
49    ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
50    die "can't locate arm-xlate.pl";
51
52    open STDOUT,"| \"$^X\" $xlate $flavour $output";
53} else {
54    open STDOUT,">$output";
55}
56
57$s0="r0";
58$s1="r1";
59$s2="r2";
60$s3="r3";
61$t1="r4";
62$t2="r5";
63$t3="r6";
64$i1="r7";
65$i2="r8";
66$i3="r9";
67
68$tbl="r10";
69$key="r11";
70$rounds="r12";
71
72$code=<<___;
73#ifndef __KERNEL__
74# include "arm_arch.h"
75#else
76# define __ARM_ARCH__ __LINUX_ARM_ARCH__
77#endif
78
79.text
80#if defined(__thumb2__) && !defined(__APPLE__)
81.syntax	unified
82.thumb
83#else
84.code	32
85#undef __thumb2__
86#endif
87
88.type	AES_Te,%object
89.align	5
90AES_Te:
91.word	0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
92.word	0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
93.word	0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
94.word	0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
95.word	0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
96.word	0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
97.word	0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
98.word	0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
99.word	0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
100.word	0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
101.word	0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
102.word	0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
103.word	0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
104.word	0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
105.word	0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
106.word	0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
107.word	0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
108.word	0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
109.word	0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
110.word	0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
111.word	0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
112.word	0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
113.word	0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
114.word	0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
115.word	0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
116.word	0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
117.word	0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
118.word	0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
119.word	0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
120.word	0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
121.word	0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
122.word	0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
123.word	0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
124.word	0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
125.word	0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
126.word	0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
127.word	0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
128.word	0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
129.word	0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
130.word	0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
131.word	0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
132.word	0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
133.word	0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
134.word	0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
135.word	0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
136.word	0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
137.word	0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
138.word	0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
139.word	0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
140.word	0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
141.word	0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
142.word	0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
143.word	0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
144.word	0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
145.word	0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
146.word	0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
147.word	0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
148.word	0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
149.word	0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
150.word	0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
151.word	0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
152.word	0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
153.word	0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
154.word	0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
155@ Te4[256]
156.byte	0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
157.byte	0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
158.byte	0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
159.byte	0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
160.byte	0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
161.byte	0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
162.byte	0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
163.byte	0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
164.byte	0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
165.byte	0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
166.byte	0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
167.byte	0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
168.byte	0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
169.byte	0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
170.byte	0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
171.byte	0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
172.byte	0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
173.byte	0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
174.byte	0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
175.byte	0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
176.byte	0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
177.byte	0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
178.byte	0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
179.byte	0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
180.byte	0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
181.byte	0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
182.byte	0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
183.byte	0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
184.byte	0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
185.byte	0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
186.byte	0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
187.byte	0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
188@ rcon[]
189.word	0x01000000, 0x02000000, 0x04000000, 0x08000000
190.word	0x10000000, 0x20000000, 0x40000000, 0x80000000
191.word	0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
192.size	AES_Te,.-AES_Te
193
194@ void AES_encrypt(const unsigned char *in, unsigned char *out,
195@ 		 const AES_KEY *key) {
196.global AES_encrypt
197.type   AES_encrypt,%function
198.align	5
199AES_encrypt:
200#ifndef	__thumb2__
201	sub	r3,pc,#8		@ AES_encrypt
202#else
203	adr	r3,.
204#endif
205	stmdb   sp!,{r1,r4-r12,lr}
206#if defined(__thumb2__) || defined(__APPLE__)
207	adr	$tbl,AES_Te
208#else
209	sub	$tbl,r3,#AES_encrypt-AES_Te	@ Te
210#endif
211	mov	$rounds,r0		@ inp
212	mov	$key,r2
213#if __ARM_ARCH__<7
214	ldrb	$s0,[$rounds,#3]	@ load input data in endian-neutral
215	ldrb	$t1,[$rounds,#2]	@ manner...
216	ldrb	$t2,[$rounds,#1]
217	ldrb	$t3,[$rounds,#0]
218	orr	$s0,$s0,$t1,lsl#8
219	ldrb	$s1,[$rounds,#7]
220	orr	$s0,$s0,$t2,lsl#16
221	ldrb	$t1,[$rounds,#6]
222	orr	$s0,$s0,$t3,lsl#24
223	ldrb	$t2,[$rounds,#5]
224	ldrb	$t3,[$rounds,#4]
225	orr	$s1,$s1,$t1,lsl#8
226	ldrb	$s2,[$rounds,#11]
227	orr	$s1,$s1,$t2,lsl#16
228	ldrb	$t1,[$rounds,#10]
229	orr	$s1,$s1,$t3,lsl#24
230	ldrb	$t2,[$rounds,#9]
231	ldrb	$t3,[$rounds,#8]
232	orr	$s2,$s2,$t1,lsl#8
233	ldrb	$s3,[$rounds,#15]
234	orr	$s2,$s2,$t2,lsl#16
235	ldrb	$t1,[$rounds,#14]
236	orr	$s2,$s2,$t3,lsl#24
237	ldrb	$t2,[$rounds,#13]
238	ldrb	$t3,[$rounds,#12]
239	orr	$s3,$s3,$t1,lsl#8
240	orr	$s3,$s3,$t2,lsl#16
241	orr	$s3,$s3,$t3,lsl#24
242#else
243	ldr	$s0,[$rounds,#0]
244	ldr	$s1,[$rounds,#4]
245	ldr	$s2,[$rounds,#8]
246	ldr	$s3,[$rounds,#12]
247#ifdef __ARMEL__
248	rev	$s0,$s0
249	rev	$s1,$s1
250	rev	$s2,$s2
251	rev	$s3,$s3
252#endif
253#endif
254	bl	_armv4_AES_encrypt
255
256	ldr	$rounds,[sp],#4		@ pop out
257#if __ARM_ARCH__>=7
258#ifdef __ARMEL__
259	rev	$s0,$s0
260	rev	$s1,$s1
261	rev	$s2,$s2
262	rev	$s3,$s3
263#endif
264	str	$s0,[$rounds,#0]
265	str	$s1,[$rounds,#4]
266	str	$s2,[$rounds,#8]
267	str	$s3,[$rounds,#12]
268#else
269	mov	$t1,$s0,lsr#24		@ write output in endian-neutral
270	mov	$t2,$s0,lsr#16		@ manner...
271	mov	$t3,$s0,lsr#8
272	strb	$t1,[$rounds,#0]
273	strb	$t2,[$rounds,#1]
274	mov	$t1,$s1,lsr#24
275	strb	$t3,[$rounds,#2]
276	mov	$t2,$s1,lsr#16
277	strb	$s0,[$rounds,#3]
278	mov	$t3,$s1,lsr#8
279	strb	$t1,[$rounds,#4]
280	strb	$t2,[$rounds,#5]
281	mov	$t1,$s2,lsr#24
282	strb	$t3,[$rounds,#6]
283	mov	$t2,$s2,lsr#16
284	strb	$s1,[$rounds,#7]
285	mov	$t3,$s2,lsr#8
286	strb	$t1,[$rounds,#8]
287	strb	$t2,[$rounds,#9]
288	mov	$t1,$s3,lsr#24
289	strb	$t3,[$rounds,#10]
290	mov	$t2,$s3,lsr#16
291	strb	$s2,[$rounds,#11]
292	mov	$t3,$s3,lsr#8
293	strb	$t1,[$rounds,#12]
294	strb	$t2,[$rounds,#13]
295	strb	$t3,[$rounds,#14]
296	strb	$s3,[$rounds,#15]
297#endif
298#if __ARM_ARCH__>=5
299	ldmia	sp!,{r4-r12,pc}
300#else
301	ldmia   sp!,{r4-r12,lr}
302	tst	lr,#1
303	moveq	pc,lr			@ be binary compatible with V4, yet
304	bx	lr			@ interoperable with Thumb ISA:-)
305#endif
306.size	AES_encrypt,.-AES_encrypt
307
308.type   _armv4_AES_encrypt,%function
309.align	2
310_armv4_AES_encrypt:
311	str	lr,[sp,#-4]!		@ push lr
312	ldmia	$key!,{$t1-$i1}
313	eor	$s0,$s0,$t1
314	ldr	$rounds,[$key,#240-16]
315	eor	$s1,$s1,$t2
316	eor	$s2,$s2,$t3
317	eor	$s3,$s3,$i1
318	sub	$rounds,$rounds,#1
319	mov	lr,#255
320
321	and	$i1,lr,$s0
322	and	$i2,lr,$s0,lsr#8
323	and	$i3,lr,$s0,lsr#16
324	mov	$s0,$s0,lsr#24
325.Lenc_loop:
326	ldr	$t1,[$tbl,$i1,lsl#2]	@ Te3[s0>>0]
327	and	$i1,lr,$s1,lsr#16	@ i0
328	ldr	$t2,[$tbl,$i2,lsl#2]	@ Te2[s0>>8]
329	and	$i2,lr,$s1
330	ldr	$t3,[$tbl,$i3,lsl#2]	@ Te1[s0>>16]
331	and	$i3,lr,$s1,lsr#8
332	ldr	$s0,[$tbl,$s0,lsl#2]	@ Te0[s0>>24]
333	mov	$s1,$s1,lsr#24
334
335	ldr	$i1,[$tbl,$i1,lsl#2]	@ Te1[s1>>16]
336	ldr	$i2,[$tbl,$i2,lsl#2]	@ Te3[s1>>0]
337	ldr	$i3,[$tbl,$i3,lsl#2]	@ Te2[s1>>8]
338	eor	$s0,$s0,$i1,ror#8
339	ldr	$s1,[$tbl,$s1,lsl#2]	@ Te0[s1>>24]
340	and	$i1,lr,$s2,lsr#8	@ i0
341	eor	$t2,$t2,$i2,ror#8
342	and	$i2,lr,$s2,lsr#16	@ i1
343	eor	$t3,$t3,$i3,ror#8
344	and	$i3,lr,$s2
345	ldr	$i1,[$tbl,$i1,lsl#2]	@ Te2[s2>>8]
346	eor	$s1,$s1,$t1,ror#24
347	ldr	$i2,[$tbl,$i2,lsl#2]	@ Te1[s2>>16]
348	mov	$s2,$s2,lsr#24
349
350	ldr	$i3,[$tbl,$i3,lsl#2]	@ Te3[s2>>0]
351	eor	$s0,$s0,$i1,ror#16
352	ldr	$s2,[$tbl,$s2,lsl#2]	@ Te0[s2>>24]
353	and	$i1,lr,$s3		@ i0
354	eor	$s1,$s1,$i2,ror#8
355	and	$i2,lr,$s3,lsr#8	@ i1
356	eor	$t3,$t3,$i3,ror#16
357	and	$i3,lr,$s3,lsr#16	@ i2
358	ldr	$i1,[$tbl,$i1,lsl#2]	@ Te3[s3>>0]
359	eor	$s2,$s2,$t2,ror#16
360	ldr	$i2,[$tbl,$i2,lsl#2]	@ Te2[s3>>8]
361	mov	$s3,$s3,lsr#24
362
363	ldr	$i3,[$tbl,$i3,lsl#2]	@ Te1[s3>>16]
364	eor	$s0,$s0,$i1,ror#24
365	ldr	$i1,[$key],#16
366	eor	$s1,$s1,$i2,ror#16
367	ldr	$s3,[$tbl,$s3,lsl#2]	@ Te0[s3>>24]
368	eor	$s2,$s2,$i3,ror#8
369	ldr	$t1,[$key,#-12]
370	eor	$s3,$s3,$t3,ror#8
371
372	ldr	$t2,[$key,#-8]
373	eor	$s0,$s0,$i1
374	ldr	$t3,[$key,#-4]
375	and	$i1,lr,$s0
376	eor	$s1,$s1,$t1
377	and	$i2,lr,$s0,lsr#8
378	eor	$s2,$s2,$t2
379	and	$i3,lr,$s0,lsr#16
380	eor	$s3,$s3,$t3
381	mov	$s0,$s0,lsr#24
382
383	subs	$rounds,$rounds,#1
384	bne	.Lenc_loop
385
386	add	$tbl,$tbl,#2
387
388	ldrb	$t1,[$tbl,$i1,lsl#2]	@ Te4[s0>>0]
389	and	$i1,lr,$s1,lsr#16	@ i0
390	ldrb	$t2,[$tbl,$i2,lsl#2]	@ Te4[s0>>8]
391	and	$i2,lr,$s1
392	ldrb	$t3,[$tbl,$i3,lsl#2]	@ Te4[s0>>16]
393	and	$i3,lr,$s1,lsr#8
394	ldrb	$s0,[$tbl,$s0,lsl#2]	@ Te4[s0>>24]
395	mov	$s1,$s1,lsr#24
396
397	ldrb	$i1,[$tbl,$i1,lsl#2]	@ Te4[s1>>16]
398	ldrb	$i2,[$tbl,$i2,lsl#2]	@ Te4[s1>>0]
399	ldrb	$i3,[$tbl,$i3,lsl#2]	@ Te4[s1>>8]
400	eor	$s0,$i1,$s0,lsl#8
401	ldrb	$s1,[$tbl,$s1,lsl#2]	@ Te4[s1>>24]
402	and	$i1,lr,$s2,lsr#8	@ i0
403	eor	$t2,$i2,$t2,lsl#8
404	and	$i2,lr,$s2,lsr#16	@ i1
405	eor	$t3,$i3,$t3,lsl#8
406	and	$i3,lr,$s2
407	ldrb	$i1,[$tbl,$i1,lsl#2]	@ Te4[s2>>8]
408	eor	$s1,$t1,$s1,lsl#24
409	ldrb	$i2,[$tbl,$i2,lsl#2]	@ Te4[s2>>16]
410	mov	$s2,$s2,lsr#24
411
412	ldrb	$i3,[$tbl,$i3,lsl#2]	@ Te4[s2>>0]
413	eor	$s0,$i1,$s0,lsl#8
414	ldrb	$s2,[$tbl,$s2,lsl#2]	@ Te4[s2>>24]
415	and	$i1,lr,$s3		@ i0
416	eor	$s1,$s1,$i2,lsl#16
417	and	$i2,lr,$s3,lsr#8	@ i1
418	eor	$t3,$i3,$t3,lsl#8
419	and	$i3,lr,$s3,lsr#16	@ i2
420	ldrb	$i1,[$tbl,$i1,lsl#2]	@ Te4[s3>>0]
421	eor	$s2,$t2,$s2,lsl#24
422	ldrb	$i2,[$tbl,$i2,lsl#2]	@ Te4[s3>>8]
423	mov	$s3,$s3,lsr#24
424
425	ldrb	$i3,[$tbl,$i3,lsl#2]	@ Te4[s3>>16]
426	eor	$s0,$i1,$s0,lsl#8
427	ldr	$i1,[$key,#0]
428	ldrb	$s3,[$tbl,$s3,lsl#2]	@ Te4[s3>>24]
429	eor	$s1,$s1,$i2,lsl#8
430	ldr	$t1,[$key,#4]
431	eor	$s2,$s2,$i3,lsl#16
432	ldr	$t2,[$key,#8]
433	eor	$s3,$t3,$s3,lsl#24
434	ldr	$t3,[$key,#12]
435
436	eor	$s0,$s0,$i1
437	eor	$s1,$s1,$t1
438	eor	$s2,$s2,$t2
439	eor	$s3,$s3,$t3
440
441	sub	$tbl,$tbl,#2
442	ldr	pc,[sp],#4		@ pop and return
443.size	_armv4_AES_encrypt,.-_armv4_AES_encrypt
444
445.global AES_set_encrypt_key
446.type   AES_set_encrypt_key,%function
447.align	5
448AES_set_encrypt_key:
449_armv4_AES_set_encrypt_key:
450#ifndef	__thumb2__
451	sub	r3,pc,#8		@ AES_set_encrypt_key
452#else
453	adr	r3,.
454#endif
455	teq	r0,#0
456#ifdef	__thumb2__
457	itt	eq			@ Thumb2 thing, sanity check in ARM
458#endif
459	moveq	r0,#-1
460	beq	.Labrt
461	teq	r2,#0
462#ifdef	__thumb2__
463	itt	eq			@ Thumb2 thing, sanity check in ARM
464#endif
465	moveq	r0,#-1
466	beq	.Labrt
467
468	teq	r1,#128
469	beq	.Lok
470	teq	r1,#192
471	beq	.Lok
472	teq	r1,#256
473#ifdef	__thumb2__
474	itt	ne			@ Thumb2 thing, sanity check in ARM
475#endif
476	movne	r0,#-1
477	bne	.Labrt
478
479.Lok:	stmdb   sp!,{r4-r12,lr}
480	mov	$rounds,r0		@ inp
481	mov	lr,r1			@ bits
482	mov	$key,r2			@ key
483
484#if defined(__thumb2__) || defined(__APPLE__)
485	adr	$tbl,AES_Te+1024				@ Te4
486#else
487	sub	$tbl,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024	@ Te4
488#endif
489
490#if __ARM_ARCH__<7
491	ldrb	$s0,[$rounds,#3]	@ load input data in endian-neutral
492	ldrb	$t1,[$rounds,#2]	@ manner...
493	ldrb	$t2,[$rounds,#1]
494	ldrb	$t3,[$rounds,#0]
495	orr	$s0,$s0,$t1,lsl#8
496	ldrb	$s1,[$rounds,#7]
497	orr	$s0,$s0,$t2,lsl#16
498	ldrb	$t1,[$rounds,#6]
499	orr	$s0,$s0,$t3,lsl#24
500	ldrb	$t2,[$rounds,#5]
501	ldrb	$t3,[$rounds,#4]
502	orr	$s1,$s1,$t1,lsl#8
503	ldrb	$s2,[$rounds,#11]
504	orr	$s1,$s1,$t2,lsl#16
505	ldrb	$t1,[$rounds,#10]
506	orr	$s1,$s1,$t3,lsl#24
507	ldrb	$t2,[$rounds,#9]
508	ldrb	$t3,[$rounds,#8]
509	orr	$s2,$s2,$t1,lsl#8
510	ldrb	$s3,[$rounds,#15]
511	orr	$s2,$s2,$t2,lsl#16
512	ldrb	$t1,[$rounds,#14]
513	orr	$s2,$s2,$t3,lsl#24
514	ldrb	$t2,[$rounds,#13]
515	ldrb	$t3,[$rounds,#12]
516	orr	$s3,$s3,$t1,lsl#8
517	str	$s0,[$key],#16
518	orr	$s3,$s3,$t2,lsl#16
519	str	$s1,[$key,#-12]
520	orr	$s3,$s3,$t3,lsl#24
521	str	$s2,[$key,#-8]
522	str	$s3,[$key,#-4]
523#else
524	ldr	$s0,[$rounds,#0]
525	ldr	$s1,[$rounds,#4]
526	ldr	$s2,[$rounds,#8]
527	ldr	$s3,[$rounds,#12]
528#ifdef __ARMEL__
529	rev	$s0,$s0
530	rev	$s1,$s1
531	rev	$s2,$s2
532	rev	$s3,$s3
533#endif
534	str	$s0,[$key],#16
535	str	$s1,[$key,#-12]
536	str	$s2,[$key,#-8]
537	str	$s3,[$key,#-4]
538#endif
539
540	teq	lr,#128
541	bne	.Lnot128
542	mov	$rounds,#10
543	str	$rounds,[$key,#240-16]
544	add	$t3,$tbl,#256			@ rcon
545	mov	lr,#255
546
547.L128_loop:
548	and	$t2,lr,$s3,lsr#24
549	and	$i1,lr,$s3,lsr#16
550	ldrb	$t2,[$tbl,$t2]
551	and	$i2,lr,$s3,lsr#8
552	ldrb	$i1,[$tbl,$i1]
553	and	$i3,lr,$s3
554	ldrb	$i2,[$tbl,$i2]
555	orr	$t2,$t2,$i1,lsl#24
556	ldrb	$i3,[$tbl,$i3]
557	orr	$t2,$t2,$i2,lsl#16
558	ldr	$t1,[$t3],#4			@ rcon[i++]
559	orr	$t2,$t2,$i3,lsl#8
560	eor	$t2,$t2,$t1
561	eor	$s0,$s0,$t2			@ rk[4]=rk[0]^...
562	eor	$s1,$s1,$s0			@ rk[5]=rk[1]^rk[4]
563	str	$s0,[$key],#16
564	eor	$s2,$s2,$s1			@ rk[6]=rk[2]^rk[5]
565	str	$s1,[$key,#-12]
566	eor	$s3,$s3,$s2			@ rk[7]=rk[3]^rk[6]
567	str	$s2,[$key,#-8]
568	subs	$rounds,$rounds,#1
569	str	$s3,[$key,#-4]
570	bne	.L128_loop
571	sub	r2,$key,#176
572	b	.Ldone
573
574.Lnot128:
575#if __ARM_ARCH__<7
576	ldrb	$i2,[$rounds,#19]
577	ldrb	$t1,[$rounds,#18]
578	ldrb	$t2,[$rounds,#17]
579	ldrb	$t3,[$rounds,#16]
580	orr	$i2,$i2,$t1,lsl#8
581	ldrb	$i3,[$rounds,#23]
582	orr	$i2,$i2,$t2,lsl#16
583	ldrb	$t1,[$rounds,#22]
584	orr	$i2,$i2,$t3,lsl#24
585	ldrb	$t2,[$rounds,#21]
586	ldrb	$t3,[$rounds,#20]
587	orr	$i3,$i3,$t1,lsl#8
588	orr	$i3,$i3,$t2,lsl#16
589	str	$i2,[$key],#8
590	orr	$i3,$i3,$t3,lsl#24
591	str	$i3,[$key,#-4]
592#else
593	ldr	$i2,[$rounds,#16]
594	ldr	$i3,[$rounds,#20]
595#ifdef __ARMEL__
596	rev	$i2,$i2
597	rev	$i3,$i3
598#endif
599	str	$i2,[$key],#8
600	str	$i3,[$key,#-4]
601#endif
602
603	teq	lr,#192
604	bne	.Lnot192
605	mov	$rounds,#12
606	str	$rounds,[$key,#240-24]
607	add	$t3,$tbl,#256			@ rcon
608	mov	lr,#255
609	mov	$rounds,#8
610
611.L192_loop:
612	and	$t2,lr,$i3,lsr#24
613	and	$i1,lr,$i3,lsr#16
614	ldrb	$t2,[$tbl,$t2]
615	and	$i2,lr,$i3,lsr#8
616	ldrb	$i1,[$tbl,$i1]
617	and	$i3,lr,$i3
618	ldrb	$i2,[$tbl,$i2]
619	orr	$t2,$t2,$i1,lsl#24
620	ldrb	$i3,[$tbl,$i3]
621	orr	$t2,$t2,$i2,lsl#16
622	ldr	$t1,[$t3],#4			@ rcon[i++]
623	orr	$t2,$t2,$i3,lsl#8
624	eor	$i3,$t2,$t1
625	eor	$s0,$s0,$i3			@ rk[6]=rk[0]^...
626	eor	$s1,$s1,$s0			@ rk[7]=rk[1]^rk[6]
627	str	$s0,[$key],#24
628	eor	$s2,$s2,$s1			@ rk[8]=rk[2]^rk[7]
629	str	$s1,[$key,#-20]
630	eor	$s3,$s3,$s2			@ rk[9]=rk[3]^rk[8]
631	str	$s2,[$key,#-16]
632	subs	$rounds,$rounds,#1
633	str	$s3,[$key,#-12]
634#ifdef	__thumb2__
635	itt	eq				@ Thumb2 thing, sanity check in ARM
636#endif
637	subeq	r2,$key,#216
638	beq	.Ldone
639
640	ldr	$i1,[$key,#-32]
641	ldr	$i2,[$key,#-28]
642	eor	$i1,$i1,$s3			@ rk[10]=rk[4]^rk[9]
643	eor	$i3,$i2,$i1			@ rk[11]=rk[5]^rk[10]
644	str	$i1,[$key,#-8]
645	str	$i3,[$key,#-4]
646	b	.L192_loop
647
648.Lnot192:
649#if __ARM_ARCH__<7
650	ldrb	$i2,[$rounds,#27]
651	ldrb	$t1,[$rounds,#26]
652	ldrb	$t2,[$rounds,#25]
653	ldrb	$t3,[$rounds,#24]
654	orr	$i2,$i2,$t1,lsl#8
655	ldrb	$i3,[$rounds,#31]
656	orr	$i2,$i2,$t2,lsl#16
657	ldrb	$t1,[$rounds,#30]
658	orr	$i2,$i2,$t3,lsl#24
659	ldrb	$t2,[$rounds,#29]
660	ldrb	$t3,[$rounds,#28]
661	orr	$i3,$i3,$t1,lsl#8
662	orr	$i3,$i3,$t2,lsl#16
663	str	$i2,[$key],#8
664	orr	$i3,$i3,$t3,lsl#24
665	str	$i3,[$key,#-4]
666#else
667	ldr	$i2,[$rounds,#24]
668	ldr	$i3,[$rounds,#28]
669#ifdef __ARMEL__
670	rev	$i2,$i2
671	rev	$i3,$i3
672#endif
673	str	$i2,[$key],#8
674	str	$i3,[$key,#-4]
675#endif
676
677	mov	$rounds,#14
678	str	$rounds,[$key,#240-32]
679	add	$t3,$tbl,#256			@ rcon
680	mov	lr,#255
681	mov	$rounds,#7
682
683.L256_loop:
684	and	$t2,lr,$i3,lsr#24
685	and	$i1,lr,$i3,lsr#16
686	ldrb	$t2,[$tbl,$t2]
687	and	$i2,lr,$i3,lsr#8
688	ldrb	$i1,[$tbl,$i1]
689	and	$i3,lr,$i3
690	ldrb	$i2,[$tbl,$i2]
691	orr	$t2,$t2,$i1,lsl#24
692	ldrb	$i3,[$tbl,$i3]
693	orr	$t2,$t2,$i2,lsl#16
694	ldr	$t1,[$t3],#4			@ rcon[i++]
695	orr	$t2,$t2,$i3,lsl#8
696	eor	$i3,$t2,$t1
697	eor	$s0,$s0,$i3			@ rk[8]=rk[0]^...
698	eor	$s1,$s1,$s0			@ rk[9]=rk[1]^rk[8]
699	str	$s0,[$key],#32
700	eor	$s2,$s2,$s1			@ rk[10]=rk[2]^rk[9]
701	str	$s1,[$key,#-28]
702	eor	$s3,$s3,$s2			@ rk[11]=rk[3]^rk[10]
703	str	$s2,[$key,#-24]
704	subs	$rounds,$rounds,#1
705	str	$s3,[$key,#-20]
706#ifdef	__thumb2__
707	itt	eq				@ Thumb2 thing, sanity check in ARM
708#endif
709	subeq	r2,$key,#256
710	beq	.Ldone
711
712	and	$t2,lr,$s3
713	and	$i1,lr,$s3,lsr#8
714	ldrb	$t2,[$tbl,$t2]
715	and	$i2,lr,$s3,lsr#16
716	ldrb	$i1,[$tbl,$i1]
717	and	$i3,lr,$s3,lsr#24
718	ldrb	$i2,[$tbl,$i2]
719	orr	$t2,$t2,$i1,lsl#8
720	ldrb	$i3,[$tbl,$i3]
721	orr	$t2,$t2,$i2,lsl#16
722	ldr	$t1,[$key,#-48]
723	orr	$t2,$t2,$i3,lsl#24
724
725	ldr	$i1,[$key,#-44]
726	ldr	$i2,[$key,#-40]
727	eor	$t1,$t1,$t2			@ rk[12]=rk[4]^...
728	ldr	$i3,[$key,#-36]
729	eor	$i1,$i1,$t1			@ rk[13]=rk[5]^rk[12]
730	str	$t1,[$key,#-16]
731	eor	$i2,$i2,$i1			@ rk[14]=rk[6]^rk[13]
732	str	$i1,[$key,#-12]
733	eor	$i3,$i3,$i2			@ rk[15]=rk[7]^rk[14]
734	str	$i2,[$key,#-8]
735	str	$i3,[$key,#-4]
736	b	.L256_loop
737
738.align	2
739.Ldone:	mov	r0,#0
740	ldmia   sp!,{r4-r12,lr}
741.Labrt:
742#if __ARM_ARCH__>=5
743	ret				@ bx lr
744#else
745	tst	lr,#1
746	moveq	pc,lr			@ be binary compatible with V4, yet
747	bx	lr			@ interoperable with Thumb ISA:-)
748#endif
749.size	AES_set_encrypt_key,.-AES_set_encrypt_key
750
751.global AES_set_decrypt_key
752.type   AES_set_decrypt_key,%function
753.align	5
754AES_set_decrypt_key:
755	str	lr,[sp,#-4]!            @ push lr
756	bl	_armv4_AES_set_encrypt_key
757	teq	r0,#0
758	ldr	lr,[sp],#4              @ pop lr
759	bne	.Labrt
760
761	mov	r0,r2			@ AES_set_encrypt_key preserves r2,
762	mov	r1,r2			@ which is AES_KEY *key
763	b	_armv4_AES_set_enc2dec_key
764.size	AES_set_decrypt_key,.-AES_set_decrypt_key
765
766@ void AES_set_enc2dec_key(const AES_KEY *inp,AES_KEY *out)
767.global	AES_set_enc2dec_key
768.type	AES_set_enc2dec_key,%function
769.align	5
770AES_set_enc2dec_key:
771_armv4_AES_set_enc2dec_key:
772	stmdb   sp!,{r4-r12,lr}
773
774	ldr	$rounds,[r0,#240]
775	mov	$i1,r0			@ input
776	add	$i2,r0,$rounds,lsl#4
777	mov	$key,r1			@ output
778	add	$tbl,r1,$rounds,lsl#4
779	str	$rounds,[r1,#240]
780
781.Linv:	ldr	$s0,[$i1],#16
782	ldr	$s1,[$i1,#-12]
783	ldr	$s2,[$i1,#-8]
784	ldr	$s3,[$i1,#-4]
785	ldr	$t1,[$i2],#-16
786	ldr	$t2,[$i2,#16+4]
787	ldr	$t3,[$i2,#16+8]
788	ldr	$i3,[$i2,#16+12]
789	str	$s0,[$tbl],#-16
790	str	$s1,[$tbl,#16+4]
791	str	$s2,[$tbl,#16+8]
792	str	$s3,[$tbl,#16+12]
793	str	$t1,[$key],#16
794	str	$t2,[$key,#-12]
795	str	$t3,[$key,#-8]
796	str	$i3,[$key,#-4]
797	teq	$i1,$i2
798	bne	.Linv
799
800	ldr	$s0,[$i1]
801	ldr	$s1,[$i1,#4]
802	ldr	$s2,[$i1,#8]
803	ldr	$s3,[$i1,#12]
804	str	$s0,[$key]
805	str	$s1,[$key,#4]
806	str	$s2,[$key,#8]
807	str	$s3,[$key,#12]
808	sub	$key,$key,$rounds,lsl#3
809___
810$mask80=$i1;
811$mask1b=$i2;
812$mask7f=$i3;
813$code.=<<___;
814	ldr	$s0,[$key,#16]!		@ prefetch tp1
815	mov	$mask80,#0x80
816	mov	$mask1b,#0x1b
817	orr	$mask80,$mask80,#0x8000
818	orr	$mask1b,$mask1b,#0x1b00
819	orr	$mask80,$mask80,$mask80,lsl#16
820	orr	$mask1b,$mask1b,$mask1b,lsl#16
821	sub	$rounds,$rounds,#1
822	mvn	$mask7f,$mask80
823	mov	$rounds,$rounds,lsl#2	@ (rounds-1)*4
824
825.Lmix:	and	$t1,$s0,$mask80
826	and	$s1,$s0,$mask7f
827	sub	$t1,$t1,$t1,lsr#7
828	and	$t1,$t1,$mask1b
829	eor	$s1,$t1,$s1,lsl#1	@ tp2
830
831	and	$t1,$s1,$mask80
832	and	$s2,$s1,$mask7f
833	sub	$t1,$t1,$t1,lsr#7
834	and	$t1,$t1,$mask1b
835	eor	$s2,$t1,$s2,lsl#1	@ tp4
836
837	and	$t1,$s2,$mask80
838	and	$s3,$s2,$mask7f
839	sub	$t1,$t1,$t1,lsr#7
840	and	$t1,$t1,$mask1b
841	eor	$s3,$t1,$s3,lsl#1	@ tp8
842
843	eor	$t1,$s1,$s2
844	eor	$t2,$s0,$s3		@ tp9
845	eor	$t1,$t1,$s3		@ tpe
846	eor	$t1,$t1,$s1,ror#24
847	eor	$t1,$t1,$t2,ror#24	@ ^= ROTATE(tpb=tp9^tp2,8)
848	eor	$t1,$t1,$s2,ror#16
849	eor	$t1,$t1,$t2,ror#16	@ ^= ROTATE(tpd=tp9^tp4,16)
850	eor	$t1,$t1,$t2,ror#8	@ ^= ROTATE(tp9,24)
851
852	ldr	$s0,[$key,#4]		@ prefetch tp1
853	str	$t1,[$key],#4
854	subs	$rounds,$rounds,#1
855	bne	.Lmix
856
857	mov	r0,#0
858#if __ARM_ARCH__>=5
859	ldmia	sp!,{r4-r12,pc}
860#else
861	ldmia   sp!,{r4-r12,lr}
862	tst	lr,#1
863	moveq	pc,lr			@ be binary compatible with V4, yet
864	bx	lr			@ interoperable with Thumb ISA:-)
865#endif
866.size	AES_set_enc2dec_key,.-AES_set_enc2dec_key
867
868.type	AES_Td,%object
869.align	5
870AES_Td:
871.word	0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
872.word	0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
873.word	0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
874.word	0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
875.word	0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
876.word	0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
877.word	0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
878.word	0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
879.word	0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
880.word	0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
881.word	0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
882.word	0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
883.word	0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
884.word	0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
885.word	0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
886.word	0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
887.word	0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
888.word	0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
889.word	0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
890.word	0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
891.word	0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
892.word	0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
893.word	0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
894.word	0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
895.word	0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
896.word	0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
897.word	0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
898.word	0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
899.word	0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
900.word	0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
901.word	0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
902.word	0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
903.word	0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
904.word	0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
905.word	0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
906.word	0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
907.word	0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
908.word	0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
909.word	0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
910.word	0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
911.word	0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
912.word	0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
913.word	0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
914.word	0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
915.word	0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
916.word	0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
917.word	0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
918.word	0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
919.word	0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
920.word	0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
921.word	0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
922.word	0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
923.word	0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
924.word	0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
925.word	0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
926.word	0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
927.word	0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
928.word	0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
929.word	0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
930.word	0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
931.word	0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
932.word	0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
933.word	0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
934.word	0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
935@ Td4[256]
936.byte	0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
937.byte	0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
938.byte	0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
939.byte	0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
940.byte	0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
941.byte	0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
942.byte	0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
943.byte	0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
944.byte	0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
945.byte	0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
946.byte	0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
947.byte	0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
948.byte	0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
949.byte	0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
950.byte	0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
951.byte	0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
952.byte	0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
953.byte	0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
954.byte	0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
955.byte	0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
956.byte	0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
957.byte	0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
958.byte	0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
959.byte	0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
960.byte	0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
961.byte	0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
962.byte	0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
963.byte	0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
964.byte	0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
965.byte	0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
966.byte	0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
967.byte	0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
968.size	AES_Td,.-AES_Td
969
970@ void AES_decrypt(const unsigned char *in, unsigned char *out,
971@ 		 const AES_KEY *key) {
972.global AES_decrypt
973.type   AES_decrypt,%function
974.align	5
975AES_decrypt:
976#ifndef	__thumb2__
977	sub	r3,pc,#8		@ AES_decrypt
978#else
979	adr	r3,.
980#endif
981	stmdb   sp!,{r1,r4-r12,lr}
982#if defined(__thumb2__) || defined(__APPLE__)
983	adr	$tbl,AES_Td
984#else
985	sub	$tbl,r3,#AES_decrypt-AES_Td	@ Td
986#endif
987	mov	$rounds,r0		@ inp
988	mov	$key,r2
989#if __ARM_ARCH__<7
990	ldrb	$s0,[$rounds,#3]	@ load input data in endian-neutral
991	ldrb	$t1,[$rounds,#2]	@ manner...
992	ldrb	$t2,[$rounds,#1]
993	ldrb	$t3,[$rounds,#0]
994	orr	$s0,$s0,$t1,lsl#8
995	ldrb	$s1,[$rounds,#7]
996	orr	$s0,$s0,$t2,lsl#16
997	ldrb	$t1,[$rounds,#6]
998	orr	$s0,$s0,$t3,lsl#24
999	ldrb	$t2,[$rounds,#5]
1000	ldrb	$t3,[$rounds,#4]
1001	orr	$s1,$s1,$t1,lsl#8
1002	ldrb	$s2,[$rounds,#11]
1003	orr	$s1,$s1,$t2,lsl#16
1004	ldrb	$t1,[$rounds,#10]
1005	orr	$s1,$s1,$t3,lsl#24
1006	ldrb	$t2,[$rounds,#9]
1007	ldrb	$t3,[$rounds,#8]
1008	orr	$s2,$s2,$t1,lsl#8
1009	ldrb	$s3,[$rounds,#15]
1010	orr	$s2,$s2,$t2,lsl#16
1011	ldrb	$t1,[$rounds,#14]
1012	orr	$s2,$s2,$t3,lsl#24
1013	ldrb	$t2,[$rounds,#13]
1014	ldrb	$t3,[$rounds,#12]
1015	orr	$s3,$s3,$t1,lsl#8
1016	orr	$s3,$s3,$t2,lsl#16
1017	orr	$s3,$s3,$t3,lsl#24
1018#else
1019	ldr	$s0,[$rounds,#0]
1020	ldr	$s1,[$rounds,#4]
1021	ldr	$s2,[$rounds,#8]
1022	ldr	$s3,[$rounds,#12]
1023#ifdef __ARMEL__
1024	rev	$s0,$s0
1025	rev	$s1,$s1
1026	rev	$s2,$s2
1027	rev	$s3,$s3
1028#endif
1029#endif
1030	bl	_armv4_AES_decrypt
1031
1032	ldr	$rounds,[sp],#4		@ pop out
1033#if __ARM_ARCH__>=7
1034#ifdef __ARMEL__
1035	rev	$s0,$s0
1036	rev	$s1,$s1
1037	rev	$s2,$s2
1038	rev	$s3,$s3
1039#endif
1040	str	$s0,[$rounds,#0]
1041	str	$s1,[$rounds,#4]
1042	str	$s2,[$rounds,#8]
1043	str	$s3,[$rounds,#12]
1044#else
1045	mov	$t1,$s0,lsr#24		@ write output in endian-neutral
1046	mov	$t2,$s0,lsr#16		@ manner...
1047	mov	$t3,$s0,lsr#8
1048	strb	$t1,[$rounds,#0]
1049	strb	$t2,[$rounds,#1]
1050	mov	$t1,$s1,lsr#24
1051	strb	$t3,[$rounds,#2]
1052	mov	$t2,$s1,lsr#16
1053	strb	$s0,[$rounds,#3]
1054	mov	$t3,$s1,lsr#8
1055	strb	$t1,[$rounds,#4]
1056	strb	$t2,[$rounds,#5]
1057	mov	$t1,$s2,lsr#24
1058	strb	$t3,[$rounds,#6]
1059	mov	$t2,$s2,lsr#16
1060	strb	$s1,[$rounds,#7]
1061	mov	$t3,$s2,lsr#8
1062	strb	$t1,[$rounds,#8]
1063	strb	$t2,[$rounds,#9]
1064	mov	$t1,$s3,lsr#24
1065	strb	$t3,[$rounds,#10]
1066	mov	$t2,$s3,lsr#16
1067	strb	$s2,[$rounds,#11]
1068	mov	$t3,$s3,lsr#8
1069	strb	$t1,[$rounds,#12]
1070	strb	$t2,[$rounds,#13]
1071	strb	$t3,[$rounds,#14]
1072	strb	$s3,[$rounds,#15]
1073#endif
1074#if __ARM_ARCH__>=5
1075	ldmia	sp!,{r4-r12,pc}
1076#else
1077	ldmia   sp!,{r4-r12,lr}
1078	tst	lr,#1
1079	moveq	pc,lr			@ be binary compatible with V4, yet
1080	bx	lr			@ interoperable with Thumb ISA:-)
1081#endif
1082.size	AES_decrypt,.-AES_decrypt
1083
1084.type   _armv4_AES_decrypt,%function
1085.align	2
1086_armv4_AES_decrypt:
1087	str	lr,[sp,#-4]!		@ push lr
1088	ldmia	$key!,{$t1-$i1}
1089	eor	$s0,$s0,$t1
1090	ldr	$rounds,[$key,#240-16]
1091	eor	$s1,$s1,$t2
1092	eor	$s2,$s2,$t3
1093	eor	$s3,$s3,$i1
1094	sub	$rounds,$rounds,#1
1095	mov	lr,#255
1096
1097	and	$i1,lr,$s0,lsr#16
1098	and	$i2,lr,$s0,lsr#8
1099	and	$i3,lr,$s0
1100	mov	$s0,$s0,lsr#24
1101.Ldec_loop:
1102	ldr	$t1,[$tbl,$i1,lsl#2]	@ Td1[s0>>16]
1103	and	$i1,lr,$s1		@ i0
1104	ldr	$t2,[$tbl,$i2,lsl#2]	@ Td2[s0>>8]
1105	and	$i2,lr,$s1,lsr#16
1106	ldr	$t3,[$tbl,$i3,lsl#2]	@ Td3[s0>>0]
1107	and	$i3,lr,$s1,lsr#8
1108	ldr	$s0,[$tbl,$s0,lsl#2]	@ Td0[s0>>24]
1109	mov	$s1,$s1,lsr#24
1110
1111	ldr	$i1,[$tbl,$i1,lsl#2]	@ Td3[s1>>0]
1112	ldr	$i2,[$tbl,$i2,lsl#2]	@ Td1[s1>>16]
1113	ldr	$i3,[$tbl,$i3,lsl#2]	@ Td2[s1>>8]
1114	eor	$s0,$s0,$i1,ror#24
1115	ldr	$s1,[$tbl,$s1,lsl#2]	@ Td0[s1>>24]
1116	and	$i1,lr,$s2,lsr#8	@ i0
1117	eor	$t2,$i2,$t2,ror#8
1118	and	$i2,lr,$s2		@ i1
1119	eor	$t3,$i3,$t3,ror#8
1120	and	$i3,lr,$s2,lsr#16
1121	ldr	$i1,[$tbl,$i1,lsl#2]	@ Td2[s2>>8]
1122	eor	$s1,$s1,$t1,ror#8
1123	ldr	$i2,[$tbl,$i2,lsl#2]	@ Td3[s2>>0]
1124	mov	$s2,$s2,lsr#24
1125
1126	ldr	$i3,[$tbl,$i3,lsl#2]	@ Td1[s2>>16]
1127	eor	$s0,$s0,$i1,ror#16
1128	ldr	$s2,[$tbl,$s2,lsl#2]	@ Td0[s2>>24]
1129	and	$i1,lr,$s3,lsr#16	@ i0
1130	eor	$s1,$s1,$i2,ror#24
1131	and	$i2,lr,$s3,lsr#8	@ i1
1132	eor	$t3,$i3,$t3,ror#8
1133	and	$i3,lr,$s3		@ i2
1134	ldr	$i1,[$tbl,$i1,lsl#2]	@ Td1[s3>>16]
1135	eor	$s2,$s2,$t2,ror#8
1136	ldr	$i2,[$tbl,$i2,lsl#2]	@ Td2[s3>>8]
1137	mov	$s3,$s3,lsr#24
1138
1139	ldr	$i3,[$tbl,$i3,lsl#2]	@ Td3[s3>>0]
1140	eor	$s0,$s0,$i1,ror#8
1141	ldr	$i1,[$key],#16
1142	eor	$s1,$s1,$i2,ror#16
1143	ldr	$s3,[$tbl,$s3,lsl#2]	@ Td0[s3>>24]
1144	eor	$s2,$s2,$i3,ror#24
1145
1146	ldr	$t1,[$key,#-12]
1147	eor	$s0,$s0,$i1
1148	ldr	$t2,[$key,#-8]
1149	eor	$s3,$s3,$t3,ror#8
1150	ldr	$t3,[$key,#-4]
1151	and	$i1,lr,$s0,lsr#16
1152	eor	$s1,$s1,$t1
1153	and	$i2,lr,$s0,lsr#8
1154	eor	$s2,$s2,$t2
1155	and	$i3,lr,$s0
1156	eor	$s3,$s3,$t3
1157	mov	$s0,$s0,lsr#24
1158
1159	subs	$rounds,$rounds,#1
1160	bne	.Ldec_loop
1161
1162	add	$tbl,$tbl,#1024
1163
1164	ldr	$t2,[$tbl,#0]		@ prefetch Td4
1165	ldr	$t3,[$tbl,#32]
1166	ldr	$t1,[$tbl,#64]
1167	ldr	$t2,[$tbl,#96]
1168	ldr	$t3,[$tbl,#128]
1169	ldr	$t1,[$tbl,#160]
1170	ldr	$t2,[$tbl,#192]
1171	ldr	$t3,[$tbl,#224]
1172
1173	ldrb	$s0,[$tbl,$s0]		@ Td4[s0>>24]
1174	ldrb	$t1,[$tbl,$i1]		@ Td4[s0>>16]
1175	and	$i1,lr,$s1		@ i0
1176	ldrb	$t2,[$tbl,$i2]		@ Td4[s0>>8]
1177	and	$i2,lr,$s1,lsr#16
1178	ldrb	$t3,[$tbl,$i3]		@ Td4[s0>>0]
1179	and	$i3,lr,$s1,lsr#8
1180
1181	add	$s1,$tbl,$s1,lsr#24
1182	ldrb	$i1,[$tbl,$i1]		@ Td4[s1>>0]
1183	ldrb	$s1,[$s1]		@ Td4[s1>>24]
1184	ldrb	$i2,[$tbl,$i2]		@ Td4[s1>>16]
1185	eor	$s0,$i1,$s0,lsl#24
1186	ldrb	$i3,[$tbl,$i3]		@ Td4[s1>>8]
1187	eor	$s1,$t1,$s1,lsl#8
1188	and	$i1,lr,$s2,lsr#8	@ i0
1189	eor	$t2,$t2,$i2,lsl#8
1190	and	$i2,lr,$s2		@ i1
1191	ldrb	$i1,[$tbl,$i1]		@ Td4[s2>>8]
1192	eor	$t3,$t3,$i3,lsl#8
1193	ldrb	$i2,[$tbl,$i2]		@ Td4[s2>>0]
1194	and	$i3,lr,$s2,lsr#16
1195
1196	add	$s2,$tbl,$s2,lsr#24
1197	ldrb	$s2,[$s2]		@ Td4[s2>>24]
1198	eor	$s0,$s0,$i1,lsl#8
1199	ldrb	$i3,[$tbl,$i3]		@ Td4[s2>>16]
1200	eor	$s1,$i2,$s1,lsl#16
1201	and	$i1,lr,$s3,lsr#16	@ i0
1202	eor	$s2,$t2,$s2,lsl#16
1203	and	$i2,lr,$s3,lsr#8	@ i1
1204	ldrb	$i1,[$tbl,$i1]		@ Td4[s3>>16]
1205	eor	$t3,$t3,$i3,lsl#16
1206	ldrb	$i2,[$tbl,$i2]		@ Td4[s3>>8]
1207	and	$i3,lr,$s3		@ i2
1208
1209	add	$s3,$tbl,$s3,lsr#24
1210	ldrb	$i3,[$tbl,$i3]		@ Td4[s3>>0]
1211	ldrb	$s3,[$s3]		@ Td4[s3>>24]
1212	eor	$s0,$s0,$i1,lsl#16
1213	ldr	$i1,[$key,#0]
1214	eor	$s1,$s1,$i2,lsl#8
1215	ldr	$t1,[$key,#4]
1216	eor	$s2,$i3,$s2,lsl#8
1217	ldr	$t2,[$key,#8]
1218	eor	$s3,$t3,$s3,lsl#24
1219	ldr	$t3,[$key,#12]
1220
1221	eor	$s0,$s0,$i1
1222	eor	$s1,$s1,$t1
1223	eor	$s2,$s2,$t2
1224	eor	$s3,$s3,$t3
1225
1226	sub	$tbl,$tbl,#1024
1227	ldr	pc,[sp],#4		@ pop and return
1228.size	_armv4_AES_decrypt,.-_armv4_AES_decrypt
1229.asciz	"AES for ARMv4, CRYPTOGAMS by <appro\@openssl.org>"
1230.align	2
1231___
1232
1233$code =~ s/\bbx\s+lr\b/.word\t0xe12fff1e/gm;	# make it possible to compile with -march=armv4
1234$code =~ s/\bret\b/bx\tlr/gm;
1235
1236open SELF,$0;
1237while(<SELF>) {
1238	next if (/^#!/);
1239	last if (!s/^#/@/ and !/^$/);
1240	print;
1241}
1242close SELF;
1243
1244print $code;
1245close STDOUT or die "error closing STDOUT: $!";	# enforce flush
1246