• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1@ Copyright 2007-2020 The OpenSSL Project Authors. All Rights Reserved.
2@
3@ Licensed under the Apache License 2.0 (the "License").  You may not use
4@ this file except in compliance with the License.  You can obtain a copy
5@ in the file LICENSE in the source distribution or at
6@ https://www.openssl.org/source/license.html
7
8
9@ ====================================================================
10@ Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
11@ project. The module is, however, dual licensed under OpenSSL and
12@ CRYPTOGAMS licenses depending on where you obtain it. For further
13@ details see http://www.openssl.org/~appro/cryptogams/.
14@ ====================================================================
15
16@ AES for ARMv4
17
18@ January 2007.
19@
20@ Code uses single 1K S-box and is >2 times faster than code generated
21@ by gcc-3.4.1. This is thanks to unique feature of ARMv4 ISA, which
22@ allows to merge logical or arithmetic operation with shift or rotate
23@ in one instruction and emit combined result every cycle. The module
24@ is endian-neutral. The performance is ~42 cycles/byte for 128-bit
25@ key [on single-issue Xscale PXA250 core].
26
27@ May 2007.
28@
29@ AES_set_[en|de]crypt_key is added.
30
31@ July 2010.
32@
33@ Rescheduling for dual-issue pipeline resulted in 12% improvement on
34@ Cortex A8 core and ~25 cycles per byte processed with 128-bit key.
35
36@ February 2011.
37@
38@ Profiler-assisted and platform-specific optimization resulted in 16%
39@ improvement on Cortex A8 core and ~21.5 cycles per byte.
40
41@ $output is the last argument if it looks like a file (it has an extension)
42@ $flavour is the first argument if it doesn't look like a file
43#ifndef __KERNEL__
44# include "arm_arch.h"
45#else
46# define __ARM_ARCH__ __LINUX_ARM_ARCH__
47#endif
48
49#if defined(__thumb2__) && !defined(__APPLE__)
50.syntax	unified
51.thumb
52#else
53.code	32
54#undef __thumb2__
55#endif
56
57.text
58
59.type	AES_Te,%object
60.align	5
61AES_Te:
62.word	0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
63.word	0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
64.word	0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
65.word	0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
66.word	0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
67.word	0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
68.word	0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
69.word	0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
70.word	0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
71.word	0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
72.word	0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
73.word	0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
74.word	0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
75.word	0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
76.word	0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
77.word	0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
78.word	0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
79.word	0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
80.word	0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
81.word	0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
82.word	0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
83.word	0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
84.word	0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
85.word	0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
86.word	0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
87.word	0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
88.word	0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
89.word	0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
90.word	0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
91.word	0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
92.word	0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
93.word	0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
94.word	0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
95.word	0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
96.word	0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
97.word	0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
98.word	0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
99.word	0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
100.word	0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
101.word	0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
102.word	0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
103.word	0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
104.word	0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
105.word	0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
106.word	0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
107.word	0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
108.word	0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
109.word	0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
110.word	0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
111.word	0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
112.word	0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
113.word	0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
114.word	0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
115.word	0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
116.word	0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
117.word	0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
118.word	0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
119.word	0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
120.word	0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
121.word	0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
122.word	0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
123.word	0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
124.word	0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
125.word	0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
126@ Te4[256]
127.byte	0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
128.byte	0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
129.byte	0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
130.byte	0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
131.byte	0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
132.byte	0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
133.byte	0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
134.byte	0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
135.byte	0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
136.byte	0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
137.byte	0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
138.byte	0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
139.byte	0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
140.byte	0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
141.byte	0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
142.byte	0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
143.byte	0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
144.byte	0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
145.byte	0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
146.byte	0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
147.byte	0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
148.byte	0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
149.byte	0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
150.byte	0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
151.byte	0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
152.byte	0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
153.byte	0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
154.byte	0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
155.byte	0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
156.byte	0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
157.byte	0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
158.byte	0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
159@ rcon[]
160.word	0x01000000, 0x02000000, 0x04000000, 0x08000000
161.word	0x10000000, 0x20000000, 0x40000000, 0x80000000
162.word	0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
163.size	AES_Te,.-AES_Te
164
165@ void AES_encrypt(const unsigned char *in, unsigned char *out,
166@ 		 const AES_KEY *key) {
167.globl	AES_encrypt
168.type	AES_encrypt,%function
169.align	5
170AES_encrypt:
171#ifndef	__thumb2__
172	sub	r3,pc,#8		@ AES_encrypt
173#else
174	adr	r3,.
175#endif
176	stmdb	sp!,{r1,r4-r12,lr}
177#if defined(__thumb2__) || defined(__APPLE__)
178	adr	r10,AES_Te
179#else
180	sub	r10,r3,#AES_encrypt-AES_Te	@ Te
181#endif
182	mov	r12,r0		@ inp
183	mov	r11,r2
184#if __ARM_ARCH__<7
185	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
186	ldrb	r4,[r12,#2]	@ manner...
187	ldrb	r5,[r12,#1]
188	ldrb	r6,[r12,#0]
189	orr	r0,r0,r4,lsl#8
190	ldrb	r1,[r12,#7]
191	orr	r0,r0,r5,lsl#16
192	ldrb	r4,[r12,#6]
193	orr	r0,r0,r6,lsl#24
194	ldrb	r5,[r12,#5]
195	ldrb	r6,[r12,#4]
196	orr	r1,r1,r4,lsl#8
197	ldrb	r2,[r12,#11]
198	orr	r1,r1,r5,lsl#16
199	ldrb	r4,[r12,#10]
200	orr	r1,r1,r6,lsl#24
201	ldrb	r5,[r12,#9]
202	ldrb	r6,[r12,#8]
203	orr	r2,r2,r4,lsl#8
204	ldrb	r3,[r12,#15]
205	orr	r2,r2,r5,lsl#16
206	ldrb	r4,[r12,#14]
207	orr	r2,r2,r6,lsl#24
208	ldrb	r5,[r12,#13]
209	ldrb	r6,[r12,#12]
210	orr	r3,r3,r4,lsl#8
211	orr	r3,r3,r5,lsl#16
212	orr	r3,r3,r6,lsl#24
213#else
214	ldr	r0,[r12,#0]
215	ldr	r1,[r12,#4]
216	ldr	r2,[r12,#8]
217	ldr	r3,[r12,#12]
218#ifdef __ARMEL__
219	rev	r0,r0
220	rev	r1,r1
221	rev	r2,r2
222	rev	r3,r3
223#endif
224#endif
225	bl	_armv4_AES_encrypt
226
227	ldr	r12,[sp],#4		@ pop out
228#if __ARM_ARCH__>=7
229#ifdef __ARMEL__
230	rev	r0,r0
231	rev	r1,r1
232	rev	r2,r2
233	rev	r3,r3
234#endif
235	str	r0,[r12,#0]
236	str	r1,[r12,#4]
237	str	r2,[r12,#8]
238	str	r3,[r12,#12]
239#else
240	mov	r4,r0,lsr#24		@ write output in endian-neutral
241	mov	r5,r0,lsr#16		@ manner...
242	mov	r6,r0,lsr#8
243	strb	r4,[r12,#0]
244	strb	r5,[r12,#1]
245	mov	r4,r1,lsr#24
246	strb	r6,[r12,#2]
247	mov	r5,r1,lsr#16
248	strb	r0,[r12,#3]
249	mov	r6,r1,lsr#8
250	strb	r4,[r12,#4]
251	strb	r5,[r12,#5]
252	mov	r4,r2,lsr#24
253	strb	r6,[r12,#6]
254	mov	r5,r2,lsr#16
255	strb	r1,[r12,#7]
256	mov	r6,r2,lsr#8
257	strb	r4,[r12,#8]
258	strb	r5,[r12,#9]
259	mov	r4,r3,lsr#24
260	strb	r6,[r12,#10]
261	mov	r5,r3,lsr#16
262	strb	r2,[r12,#11]
263	mov	r6,r3,lsr#8
264	strb	r4,[r12,#12]
265	strb	r5,[r12,#13]
266	strb	r6,[r12,#14]
267	strb	r3,[r12,#15]
268#endif
269#if __ARM_ARCH__>=5
270	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
271#else
272	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
273	tst	lr,#1
274	moveq	pc,lr			@ be binary compatible with V4, yet
275.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
276#endif
277.size	AES_encrypt,.-AES_encrypt
278
279.type	_armv4_AES_encrypt,%function
280.align	2
281_armv4_AES_encrypt:
282	str	lr,[sp,#-4]!		@ push lr
283	ldmia	r11!,{r4,r5,r6,r7}
284	eor	r0,r0,r4
285	ldr	r12,[r11,#240-16]
286	eor	r1,r1,r5
287	eor	r2,r2,r6
288	eor	r3,r3,r7
289	sub	r12,r12,#1
290	mov	lr,#255
291
292	and	r7,lr,r0
293	and	r8,lr,r0,lsr#8
294	and	r9,lr,r0,lsr#16
295	mov	r0,r0,lsr#24
296.Lenc_loop:
297	ldr	r4,[r10,r7,lsl#2]	@ Te3[s0>>0]
298	and	r7,lr,r1,lsr#16	@ i0
299	ldr	r5,[r10,r8,lsl#2]	@ Te2[s0>>8]
300	and	r8,lr,r1
301	ldr	r6,[r10,r9,lsl#2]	@ Te1[s0>>16]
302	and	r9,lr,r1,lsr#8
303	ldr	r0,[r10,r0,lsl#2]	@ Te0[s0>>24]
304	mov	r1,r1,lsr#24
305
306	ldr	r7,[r10,r7,lsl#2]	@ Te1[s1>>16]
307	ldr	r8,[r10,r8,lsl#2]	@ Te3[s1>>0]
308	ldr	r9,[r10,r9,lsl#2]	@ Te2[s1>>8]
309	eor	r0,r0,r7,ror#8
310	ldr	r1,[r10,r1,lsl#2]	@ Te0[s1>>24]
311	and	r7,lr,r2,lsr#8	@ i0
312	eor	r5,r5,r8,ror#8
313	and	r8,lr,r2,lsr#16	@ i1
314	eor	r6,r6,r9,ror#8
315	and	r9,lr,r2
316	ldr	r7,[r10,r7,lsl#2]	@ Te2[s2>>8]
317	eor	r1,r1,r4,ror#24
318	ldr	r8,[r10,r8,lsl#2]	@ Te1[s2>>16]
319	mov	r2,r2,lsr#24
320
321	ldr	r9,[r10,r9,lsl#2]	@ Te3[s2>>0]
322	eor	r0,r0,r7,ror#16
323	ldr	r2,[r10,r2,lsl#2]	@ Te0[s2>>24]
324	and	r7,lr,r3		@ i0
325	eor	r1,r1,r8,ror#8
326	and	r8,lr,r3,lsr#8	@ i1
327	eor	r6,r6,r9,ror#16
328	and	r9,lr,r3,lsr#16	@ i2
329	ldr	r7,[r10,r7,lsl#2]	@ Te3[s3>>0]
330	eor	r2,r2,r5,ror#16
331	ldr	r8,[r10,r8,lsl#2]	@ Te2[s3>>8]
332	mov	r3,r3,lsr#24
333
334	ldr	r9,[r10,r9,lsl#2]	@ Te1[s3>>16]
335	eor	r0,r0,r7,ror#24
336	ldr	r7,[r11],#16
337	eor	r1,r1,r8,ror#16
338	ldr	r3,[r10,r3,lsl#2]	@ Te0[s3>>24]
339	eor	r2,r2,r9,ror#8
340	ldr	r4,[r11,#-12]
341	eor	r3,r3,r6,ror#8
342
343	ldr	r5,[r11,#-8]
344	eor	r0,r0,r7
345	ldr	r6,[r11,#-4]
346	and	r7,lr,r0
347	eor	r1,r1,r4
348	and	r8,lr,r0,lsr#8
349	eor	r2,r2,r5
350	and	r9,lr,r0,lsr#16
351	eor	r3,r3,r6
352	mov	r0,r0,lsr#24
353
354	subs	r12,r12,#1
355	bne	.Lenc_loop
356
357	add	r10,r10,#2
358
359	ldrb	r4,[r10,r7,lsl#2]	@ Te4[s0>>0]
360	and	r7,lr,r1,lsr#16	@ i0
361	ldrb	r5,[r10,r8,lsl#2]	@ Te4[s0>>8]
362	and	r8,lr,r1
363	ldrb	r6,[r10,r9,lsl#2]	@ Te4[s0>>16]
364	and	r9,lr,r1,lsr#8
365	ldrb	r0,[r10,r0,lsl#2]	@ Te4[s0>>24]
366	mov	r1,r1,lsr#24
367
368	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s1>>16]
369	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s1>>0]
370	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s1>>8]
371	eor	r0,r7,r0,lsl#8
372	ldrb	r1,[r10,r1,lsl#2]	@ Te4[s1>>24]
373	and	r7,lr,r2,lsr#8	@ i0
374	eor	r5,r8,r5,lsl#8
375	and	r8,lr,r2,lsr#16	@ i1
376	eor	r6,r9,r6,lsl#8
377	and	r9,lr,r2
378	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s2>>8]
379	eor	r1,r4,r1,lsl#24
380	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s2>>16]
381	mov	r2,r2,lsr#24
382
383	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s2>>0]
384	eor	r0,r7,r0,lsl#8
385	ldrb	r2,[r10,r2,lsl#2]	@ Te4[s2>>24]
386	and	r7,lr,r3		@ i0
387	eor	r1,r1,r8,lsl#16
388	and	r8,lr,r3,lsr#8	@ i1
389	eor	r6,r9,r6,lsl#8
390	and	r9,lr,r3,lsr#16	@ i2
391	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s3>>0]
392	eor	r2,r5,r2,lsl#24
393	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s3>>8]
394	mov	r3,r3,lsr#24
395
396	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s3>>16]
397	eor	r0,r7,r0,lsl#8
398	ldr	r7,[r11,#0]
399	ldrb	r3,[r10,r3,lsl#2]	@ Te4[s3>>24]
400	eor	r1,r1,r8,lsl#8
401	ldr	r4,[r11,#4]
402	eor	r2,r2,r9,lsl#16
403	ldr	r5,[r11,#8]
404	eor	r3,r6,r3,lsl#24
405	ldr	r6,[r11,#12]
406
407	eor	r0,r0,r7
408	eor	r1,r1,r4
409	eor	r2,r2,r5
410	eor	r3,r3,r6
411
412	sub	r10,r10,#2
413	ldr	pc,[sp],#4		@ pop and return
414.size	_armv4_AES_encrypt,.-_armv4_AES_encrypt
415
416.globl	AES_set_encrypt_key
417.type	AES_set_encrypt_key,%function
418.align	5
419AES_set_encrypt_key:
420_armv4_AES_set_encrypt_key:
421#ifndef	__thumb2__
422	sub	r3,pc,#8		@ AES_set_encrypt_key
423#else
424	adr	r3,.
425#endif
426	teq	r0,#0
427#ifdef	__thumb2__
428	itt	eq			@ Thumb2 thing, sanity check in ARM
429#endif
430	moveq	r0,#-1
431	beq	.Labrt
432	teq	r2,#0
433#ifdef	__thumb2__
434	itt	eq			@ Thumb2 thing, sanity check in ARM
435#endif
436	moveq	r0,#-1
437	beq	.Labrt
438
439	teq	r1,#128
440	beq	.Lok
441	teq	r1,#192
442	beq	.Lok
443	teq	r1,#256
444#ifdef	__thumb2__
445	itt	ne			@ Thumb2 thing, sanity check in ARM
446#endif
447	movne	r0,#-1
448	bne	.Labrt
449
450.Lok:	stmdb	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
451	mov	r12,r0		@ inp
452	mov	lr,r1			@ bits
453	mov	r11,r2			@ key
454
455#if defined(__thumb2__) || defined(__APPLE__)
456	adr	r10,AES_Te+1024				@ Te4
457#else
458	sub	r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024	@ Te4
459#endif
460
461#if __ARM_ARCH__<7
462	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
463	ldrb	r4,[r12,#2]	@ manner...
464	ldrb	r5,[r12,#1]
465	ldrb	r6,[r12,#0]
466	orr	r0,r0,r4,lsl#8
467	ldrb	r1,[r12,#7]
468	orr	r0,r0,r5,lsl#16
469	ldrb	r4,[r12,#6]
470	orr	r0,r0,r6,lsl#24
471	ldrb	r5,[r12,#5]
472	ldrb	r6,[r12,#4]
473	orr	r1,r1,r4,lsl#8
474	ldrb	r2,[r12,#11]
475	orr	r1,r1,r5,lsl#16
476	ldrb	r4,[r12,#10]
477	orr	r1,r1,r6,lsl#24
478	ldrb	r5,[r12,#9]
479	ldrb	r6,[r12,#8]
480	orr	r2,r2,r4,lsl#8
481	ldrb	r3,[r12,#15]
482	orr	r2,r2,r5,lsl#16
483	ldrb	r4,[r12,#14]
484	orr	r2,r2,r6,lsl#24
485	ldrb	r5,[r12,#13]
486	ldrb	r6,[r12,#12]
487	orr	r3,r3,r4,lsl#8
488	str	r0,[r11],#16
489	orr	r3,r3,r5,lsl#16
490	str	r1,[r11,#-12]
491	orr	r3,r3,r6,lsl#24
492	str	r2,[r11,#-8]
493	str	r3,[r11,#-4]
494#else
495	ldr	r0,[r12,#0]
496	ldr	r1,[r12,#4]
497	ldr	r2,[r12,#8]
498	ldr	r3,[r12,#12]
499#ifdef __ARMEL__
500	rev	r0,r0
501	rev	r1,r1
502	rev	r2,r2
503	rev	r3,r3
504#endif
505	str	r0,[r11],#16
506	str	r1,[r11,#-12]
507	str	r2,[r11,#-8]
508	str	r3,[r11,#-4]
509#endif
510
511	teq	lr,#128
512	bne	.Lnot128
513	mov	r12,#10
514	str	r12,[r11,#240-16]
515	add	r6,r10,#256			@ rcon
516	mov	lr,#255
517
518.L128_loop:
519	and	r5,lr,r3,lsr#24
520	and	r7,lr,r3,lsr#16
521	ldrb	r5,[r10,r5]
522	and	r8,lr,r3,lsr#8
523	ldrb	r7,[r10,r7]
524	and	r9,lr,r3
525	ldrb	r8,[r10,r8]
526	orr	r5,r5,r7,lsl#24
527	ldrb	r9,[r10,r9]
528	orr	r5,r5,r8,lsl#16
529	ldr	r4,[r6],#4			@ rcon[i++]
530	orr	r5,r5,r9,lsl#8
531	eor	r5,r5,r4
532	eor	r0,r0,r5			@ rk[4]=rk[0]^...
533	eor	r1,r1,r0			@ rk[5]=rk[1]^rk[4]
534	str	r0,[r11],#16
535	eor	r2,r2,r1			@ rk[6]=rk[2]^rk[5]
536	str	r1,[r11,#-12]
537	eor	r3,r3,r2			@ rk[7]=rk[3]^rk[6]
538	str	r2,[r11,#-8]
539	subs	r12,r12,#1
540	str	r3,[r11,#-4]
541	bne	.L128_loop
542	sub	r2,r11,#176
543	b	.Ldone
544
545.Lnot128:
546#if __ARM_ARCH__<7
547	ldrb	r8,[r12,#19]
548	ldrb	r4,[r12,#18]
549	ldrb	r5,[r12,#17]
550	ldrb	r6,[r12,#16]
551	orr	r8,r8,r4,lsl#8
552	ldrb	r9,[r12,#23]
553	orr	r8,r8,r5,lsl#16
554	ldrb	r4,[r12,#22]
555	orr	r8,r8,r6,lsl#24
556	ldrb	r5,[r12,#21]
557	ldrb	r6,[r12,#20]
558	orr	r9,r9,r4,lsl#8
559	orr	r9,r9,r5,lsl#16
560	str	r8,[r11],#8
561	orr	r9,r9,r6,lsl#24
562	str	r9,[r11,#-4]
563#else
564	ldr	r8,[r12,#16]
565	ldr	r9,[r12,#20]
566#ifdef __ARMEL__
567	rev	r8,r8
568	rev	r9,r9
569#endif
570	str	r8,[r11],#8
571	str	r9,[r11,#-4]
572#endif
573
574	teq	lr,#192
575	bne	.Lnot192
576	mov	r12,#12
577	str	r12,[r11,#240-24]
578	add	r6,r10,#256			@ rcon
579	mov	lr,#255
580	mov	r12,#8
581
582.L192_loop:
583	and	r5,lr,r9,lsr#24
584	and	r7,lr,r9,lsr#16
585	ldrb	r5,[r10,r5]
586	and	r8,lr,r9,lsr#8
587	ldrb	r7,[r10,r7]
588	and	r9,lr,r9
589	ldrb	r8,[r10,r8]
590	orr	r5,r5,r7,lsl#24
591	ldrb	r9,[r10,r9]
592	orr	r5,r5,r8,lsl#16
593	ldr	r4,[r6],#4			@ rcon[i++]
594	orr	r5,r5,r9,lsl#8
595	eor	r9,r5,r4
596	eor	r0,r0,r9			@ rk[6]=rk[0]^...
597	eor	r1,r1,r0			@ rk[7]=rk[1]^rk[6]
598	str	r0,[r11],#24
599	eor	r2,r2,r1			@ rk[8]=rk[2]^rk[7]
600	str	r1,[r11,#-20]
601	eor	r3,r3,r2			@ rk[9]=rk[3]^rk[8]
602	str	r2,[r11,#-16]
603	subs	r12,r12,#1
604	str	r3,[r11,#-12]
605#ifdef	__thumb2__
606	itt	eq				@ Thumb2 thing, sanity check in ARM
607#endif
608	subeq	r2,r11,#216
609	beq	.Ldone
610
611	ldr	r7,[r11,#-32]
612	ldr	r8,[r11,#-28]
613	eor	r7,r7,r3			@ rk[10]=rk[4]^rk[9]
614	eor	r9,r8,r7			@ rk[11]=rk[5]^rk[10]
615	str	r7,[r11,#-8]
616	str	r9,[r11,#-4]
617	b	.L192_loop
618
619.Lnot192:
620#if __ARM_ARCH__<7
621	ldrb	r8,[r12,#27]
622	ldrb	r4,[r12,#26]
623	ldrb	r5,[r12,#25]
624	ldrb	r6,[r12,#24]
625	orr	r8,r8,r4,lsl#8
626	ldrb	r9,[r12,#31]
627	orr	r8,r8,r5,lsl#16
628	ldrb	r4,[r12,#30]
629	orr	r8,r8,r6,lsl#24
630	ldrb	r5,[r12,#29]
631	ldrb	r6,[r12,#28]
632	orr	r9,r9,r4,lsl#8
633	orr	r9,r9,r5,lsl#16
634	str	r8,[r11],#8
635	orr	r9,r9,r6,lsl#24
636	str	r9,[r11,#-4]
637#else
638	ldr	r8,[r12,#24]
639	ldr	r9,[r12,#28]
640#ifdef __ARMEL__
641	rev	r8,r8
642	rev	r9,r9
643#endif
644	str	r8,[r11],#8
645	str	r9,[r11,#-4]
646#endif
647
648	mov	r12,#14
649	str	r12,[r11,#240-32]
650	add	r6,r10,#256			@ rcon
651	mov	lr,#255
652	mov	r12,#7
653
654.L256_loop:
655	and	r5,lr,r9,lsr#24
656	and	r7,lr,r9,lsr#16
657	ldrb	r5,[r10,r5]
658	and	r8,lr,r9,lsr#8
659	ldrb	r7,[r10,r7]
660	and	r9,lr,r9
661	ldrb	r8,[r10,r8]
662	orr	r5,r5,r7,lsl#24
663	ldrb	r9,[r10,r9]
664	orr	r5,r5,r8,lsl#16
665	ldr	r4,[r6],#4			@ rcon[i++]
666	orr	r5,r5,r9,lsl#8
667	eor	r9,r5,r4
668	eor	r0,r0,r9			@ rk[8]=rk[0]^...
669	eor	r1,r1,r0			@ rk[9]=rk[1]^rk[8]
670	str	r0,[r11],#32
671	eor	r2,r2,r1			@ rk[10]=rk[2]^rk[9]
672	str	r1,[r11,#-28]
673	eor	r3,r3,r2			@ rk[11]=rk[3]^rk[10]
674	str	r2,[r11,#-24]
675	subs	r12,r12,#1
676	str	r3,[r11,#-20]
677#ifdef	__thumb2__
678	itt	eq				@ Thumb2 thing, sanity check in ARM
679#endif
680	subeq	r2,r11,#256
681	beq	.Ldone
682
683	and	r5,lr,r3
684	and	r7,lr,r3,lsr#8
685	ldrb	r5,[r10,r5]
686	and	r8,lr,r3,lsr#16
687	ldrb	r7,[r10,r7]
688	and	r9,lr,r3,lsr#24
689	ldrb	r8,[r10,r8]
690	orr	r5,r5,r7,lsl#8
691	ldrb	r9,[r10,r9]
692	orr	r5,r5,r8,lsl#16
693	ldr	r4,[r11,#-48]
694	orr	r5,r5,r9,lsl#24
695
696	ldr	r7,[r11,#-44]
697	ldr	r8,[r11,#-40]
698	eor	r4,r4,r5			@ rk[12]=rk[4]^...
699	ldr	r9,[r11,#-36]
700	eor	r7,r7,r4			@ rk[13]=rk[5]^rk[12]
701	str	r4,[r11,#-16]
702	eor	r8,r8,r7			@ rk[14]=rk[6]^rk[13]
703	str	r7,[r11,#-12]
704	eor	r9,r9,r8			@ rk[15]=rk[7]^rk[14]
705	str	r8,[r11,#-8]
706	str	r9,[r11,#-4]
707	b	.L256_loop
708
709.align	2
710.Ldone:	mov	r0,#0
711	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
712.Labrt:
713#if __ARM_ARCH__>=5
714	bx	lr				@ .word	0xe12fff1e
715#else
716	tst	lr,#1
717	moveq	pc,lr			@ be binary compatible with V4, yet
718.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
719#endif
720.size	AES_set_encrypt_key,.-AES_set_encrypt_key
721
722.globl	AES_set_decrypt_key
723.type	AES_set_decrypt_key,%function
724.align	5
725AES_set_decrypt_key:
726	str	lr,[sp,#-4]!            @ push lr
727	bl	_armv4_AES_set_encrypt_key
728	teq	r0,#0
729	ldr	lr,[sp],#4              @ pop lr
730	bne	.Labrt
731
732	mov	r0,r2			@ AES_set_encrypt_key preserves r2,
733	mov	r1,r2			@ which is AES_KEY *key
734	b	_armv4_AES_set_enc2dec_key
735.size	AES_set_decrypt_key,.-AES_set_decrypt_key
736
737@ void AES_set_enc2dec_key(const AES_KEY *inp,AES_KEY *out)
738.globl	AES_set_enc2dec_key
739.type	AES_set_enc2dec_key,%function
740.align	5
741AES_set_enc2dec_key:
742_armv4_AES_set_enc2dec_key:
743	stmdb	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
744
745	ldr	r12,[r0,#240]
746	mov	r7,r0			@ input
747	add	r8,r0,r12,lsl#4
748	mov	r11,r1			@ output
749	add	r10,r1,r12,lsl#4
750	str	r12,[r1,#240]
751
752.Linv:	ldr	r0,[r7],#16
753	ldr	r1,[r7,#-12]
754	ldr	r2,[r7,#-8]
755	ldr	r3,[r7,#-4]
756	ldr	r4,[r8],#-16
757	ldr	r5,[r8,#16+4]
758	ldr	r6,[r8,#16+8]
759	ldr	r9,[r8,#16+12]
760	str	r0,[r10],#-16
761	str	r1,[r10,#16+4]
762	str	r2,[r10,#16+8]
763	str	r3,[r10,#16+12]
764	str	r4,[r11],#16
765	str	r5,[r11,#-12]
766	str	r6,[r11,#-8]
767	str	r9,[r11,#-4]
768	teq	r7,r8
769	bne	.Linv
770
771	ldr	r0,[r7]
772	ldr	r1,[r7,#4]
773	ldr	r2,[r7,#8]
774	ldr	r3,[r7,#12]
775	str	r0,[r11]
776	str	r1,[r11,#4]
777	str	r2,[r11,#8]
778	str	r3,[r11,#12]
779	sub	r11,r11,r12,lsl#3
780	ldr	r0,[r11,#16]!		@ prefetch tp1
781	mov	r7,#0x80
782	mov	r8,#0x1b
783	orr	r7,r7,#0x8000
784	orr	r8,r8,#0x1b00
785	orr	r7,r7,r7,lsl#16
786	orr	r8,r8,r8,lsl#16
787	sub	r12,r12,#1
788	mvn	r9,r7
789	mov	r12,r12,lsl#2	@ (rounds-1)*4
790
791.Lmix:	and	r4,r0,r7
792	and	r1,r0,r9
793	sub	r4,r4,r4,lsr#7
794	and	r4,r4,r8
795	eor	r1,r4,r1,lsl#1	@ tp2
796
797	and	r4,r1,r7
798	and	r2,r1,r9
799	sub	r4,r4,r4,lsr#7
800	and	r4,r4,r8
801	eor	r2,r4,r2,lsl#1	@ tp4
802
803	and	r4,r2,r7
804	and	r3,r2,r9
805	sub	r4,r4,r4,lsr#7
806	and	r4,r4,r8
807	eor	r3,r4,r3,lsl#1	@ tp8
808
809	eor	r4,r1,r2
810	eor	r5,r0,r3		@ tp9
811	eor	r4,r4,r3		@ tpe
812	eor	r4,r4,r1,ror#24
813	eor	r4,r4,r5,ror#24	@ ^= ROTATE(tpb=tp9^tp2,8)
814	eor	r4,r4,r2,ror#16
815	eor	r4,r4,r5,ror#16	@ ^= ROTATE(tpd=tp9^tp4,16)
816	eor	r4,r4,r5,ror#8	@ ^= ROTATE(tp9,24)
817
818	ldr	r0,[r11,#4]		@ prefetch tp1
819	str	r4,[r11],#4
820	subs	r12,r12,#1
821	bne	.Lmix
822
823	mov	r0,#0
824#if __ARM_ARCH__>=5
825	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
826#else
827	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
828	tst	lr,#1
829	moveq	pc,lr			@ be binary compatible with V4, yet
830.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
831#endif
832.size	AES_set_enc2dec_key,.-AES_set_enc2dec_key
833
834.type	AES_Td,%object
835.align	5
836AES_Td:
837.word	0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
838.word	0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
839.word	0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
840.word	0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
841.word	0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
842.word	0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
843.word	0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
844.word	0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
845.word	0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
846.word	0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
847.word	0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
848.word	0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
849.word	0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
850.word	0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
851.word	0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
852.word	0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
853.word	0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
854.word	0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
855.word	0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
856.word	0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
857.word	0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
858.word	0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
859.word	0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
860.word	0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
861.word	0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
862.word	0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
863.word	0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
864.word	0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
865.word	0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
866.word	0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
867.word	0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
868.word	0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
869.word	0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
870.word	0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
871.word	0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
872.word	0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
873.word	0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
874.word	0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
875.word	0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
876.word	0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
877.word	0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
878.word	0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
879.word	0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
880.word	0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
881.word	0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
882.word	0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
883.word	0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
884.word	0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
885.word	0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
886.word	0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
887.word	0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
888.word	0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
889.word	0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
890.word	0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
891.word	0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
892.word	0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
893.word	0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
894.word	0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
895.word	0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
896.word	0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
897.word	0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
898.word	0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
899.word	0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
900.word	0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
901@ Td4[256]
902.byte	0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
903.byte	0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
904.byte	0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
905.byte	0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
906.byte	0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
907.byte	0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
908.byte	0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
909.byte	0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
910.byte	0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
911.byte	0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
912.byte	0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
913.byte	0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
914.byte	0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
915.byte	0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
916.byte	0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
917.byte	0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
918.byte	0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
919.byte	0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
920.byte	0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
921.byte	0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
922.byte	0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
923.byte	0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
924.byte	0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
925.byte	0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
926.byte	0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
927.byte	0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
928.byte	0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
929.byte	0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
930.byte	0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
931.byte	0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
932.byte	0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
933.byte	0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
934.size	AES_Td,.-AES_Td
935
936@ void AES_decrypt(const unsigned char *in, unsigned char *out,
937@ 		 const AES_KEY *key) {
938.globl	AES_decrypt
939.type	AES_decrypt,%function
940.align	5
941AES_decrypt:
942#ifndef	__thumb2__
943	sub	r3,pc,#8		@ AES_decrypt
944#else
945	adr	r3,.
946#endif
947	stmdb	sp!,{r1,r4-r12,lr}
948#if defined(__thumb2__) || defined(__APPLE__)
949	adr	r10,AES_Td
950#else
951	sub	r10,r3,#AES_decrypt-AES_Td	@ Td
952#endif
953	mov	r12,r0		@ inp
954	mov	r11,r2
955#if __ARM_ARCH__<7
956	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
957	ldrb	r4,[r12,#2]	@ manner...
958	ldrb	r5,[r12,#1]
959	ldrb	r6,[r12,#0]
960	orr	r0,r0,r4,lsl#8
961	ldrb	r1,[r12,#7]
962	orr	r0,r0,r5,lsl#16
963	ldrb	r4,[r12,#6]
964	orr	r0,r0,r6,lsl#24
965	ldrb	r5,[r12,#5]
966	ldrb	r6,[r12,#4]
967	orr	r1,r1,r4,lsl#8
968	ldrb	r2,[r12,#11]
969	orr	r1,r1,r5,lsl#16
970	ldrb	r4,[r12,#10]
971	orr	r1,r1,r6,lsl#24
972	ldrb	r5,[r12,#9]
973	ldrb	r6,[r12,#8]
974	orr	r2,r2,r4,lsl#8
975	ldrb	r3,[r12,#15]
976	orr	r2,r2,r5,lsl#16
977	ldrb	r4,[r12,#14]
978	orr	r2,r2,r6,lsl#24
979	ldrb	r5,[r12,#13]
980	ldrb	r6,[r12,#12]
981	orr	r3,r3,r4,lsl#8
982	orr	r3,r3,r5,lsl#16
983	orr	r3,r3,r6,lsl#24
984#else
985	ldr	r0,[r12,#0]
986	ldr	r1,[r12,#4]
987	ldr	r2,[r12,#8]
988	ldr	r3,[r12,#12]
989#ifdef __ARMEL__
990	rev	r0,r0
991	rev	r1,r1
992	rev	r2,r2
993	rev	r3,r3
994#endif
995#endif
996	bl	_armv4_AES_decrypt
997
998	ldr	r12,[sp],#4		@ pop out
999#if __ARM_ARCH__>=7
1000#ifdef __ARMEL__
1001	rev	r0,r0
1002	rev	r1,r1
1003	rev	r2,r2
1004	rev	r3,r3
1005#endif
1006	str	r0,[r12,#0]
1007	str	r1,[r12,#4]
1008	str	r2,[r12,#8]
1009	str	r3,[r12,#12]
1010#else
1011	mov	r4,r0,lsr#24		@ write output in endian-neutral
1012	mov	r5,r0,lsr#16		@ manner...
1013	mov	r6,r0,lsr#8
1014	strb	r4,[r12,#0]
1015	strb	r5,[r12,#1]
1016	mov	r4,r1,lsr#24
1017	strb	r6,[r12,#2]
1018	mov	r5,r1,lsr#16
1019	strb	r0,[r12,#3]
1020	mov	r6,r1,lsr#8
1021	strb	r4,[r12,#4]
1022	strb	r5,[r12,#5]
1023	mov	r4,r2,lsr#24
1024	strb	r6,[r12,#6]
1025	mov	r5,r2,lsr#16
1026	strb	r1,[r12,#7]
1027	mov	r6,r2,lsr#8
1028	strb	r4,[r12,#8]
1029	strb	r5,[r12,#9]
1030	mov	r4,r3,lsr#24
1031	strb	r6,[r12,#10]
1032	mov	r5,r3,lsr#16
1033	strb	r2,[r12,#11]
1034	mov	r6,r3,lsr#8
1035	strb	r4,[r12,#12]
1036	strb	r5,[r12,#13]
1037	strb	r6,[r12,#14]
1038	strb	r3,[r12,#15]
1039#endif
1040#if __ARM_ARCH__>=5
1041	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
1042#else
1043	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
1044	tst	lr,#1
1045	moveq	pc,lr			@ be binary compatible with V4, yet
1046.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
1047#endif
1048.size	AES_decrypt,.-AES_decrypt
1049
1050.type	_armv4_AES_decrypt,%function
1051.align	2
1052_armv4_AES_decrypt:
1053	str	lr,[sp,#-4]!		@ push lr
1054	ldmia	r11!,{r4,r5,r6,r7}
1055	eor	r0,r0,r4
1056	ldr	r12,[r11,#240-16]
1057	eor	r1,r1,r5
1058	eor	r2,r2,r6
1059	eor	r3,r3,r7
1060	sub	r12,r12,#1
1061	mov	lr,#255
1062
1063	and	r7,lr,r0,lsr#16
1064	and	r8,lr,r0,lsr#8
1065	and	r9,lr,r0
1066	mov	r0,r0,lsr#24
1067.Ldec_loop:
1068	ldr	r4,[r10,r7,lsl#2]	@ Td1[s0>>16]
1069	and	r7,lr,r1		@ i0
1070	ldr	r5,[r10,r8,lsl#2]	@ Td2[s0>>8]
1071	and	r8,lr,r1,lsr#16
1072	ldr	r6,[r10,r9,lsl#2]	@ Td3[s0>>0]
1073	and	r9,lr,r1,lsr#8
1074	ldr	r0,[r10,r0,lsl#2]	@ Td0[s0>>24]
1075	mov	r1,r1,lsr#24
1076
1077	ldr	r7,[r10,r7,lsl#2]	@ Td3[s1>>0]
1078	ldr	r8,[r10,r8,lsl#2]	@ Td1[s1>>16]
1079	ldr	r9,[r10,r9,lsl#2]	@ Td2[s1>>8]
1080	eor	r0,r0,r7,ror#24
1081	ldr	r1,[r10,r1,lsl#2]	@ Td0[s1>>24]
1082	and	r7,lr,r2,lsr#8	@ i0
1083	eor	r5,r8,r5,ror#8
1084	and	r8,lr,r2		@ i1
1085	eor	r6,r9,r6,ror#8
1086	and	r9,lr,r2,lsr#16
1087	ldr	r7,[r10,r7,lsl#2]	@ Td2[s2>>8]
1088	eor	r1,r1,r4,ror#8
1089	ldr	r8,[r10,r8,lsl#2]	@ Td3[s2>>0]
1090	mov	r2,r2,lsr#24
1091
1092	ldr	r9,[r10,r9,lsl#2]	@ Td1[s2>>16]
1093	eor	r0,r0,r7,ror#16
1094	ldr	r2,[r10,r2,lsl#2]	@ Td0[s2>>24]
1095	and	r7,lr,r3,lsr#16	@ i0
1096	eor	r1,r1,r8,ror#24
1097	and	r8,lr,r3,lsr#8	@ i1
1098	eor	r6,r9,r6,ror#8
1099	and	r9,lr,r3		@ i2
1100	ldr	r7,[r10,r7,lsl#2]	@ Td1[s3>>16]
1101	eor	r2,r2,r5,ror#8
1102	ldr	r8,[r10,r8,lsl#2]	@ Td2[s3>>8]
1103	mov	r3,r3,lsr#24
1104
1105	ldr	r9,[r10,r9,lsl#2]	@ Td3[s3>>0]
1106	eor	r0,r0,r7,ror#8
1107	ldr	r7,[r11],#16
1108	eor	r1,r1,r8,ror#16
1109	ldr	r3,[r10,r3,lsl#2]	@ Td0[s3>>24]
1110	eor	r2,r2,r9,ror#24
1111
1112	ldr	r4,[r11,#-12]
1113	eor	r0,r0,r7
1114	ldr	r5,[r11,#-8]
1115	eor	r3,r3,r6,ror#8
1116	ldr	r6,[r11,#-4]
1117	and	r7,lr,r0,lsr#16
1118	eor	r1,r1,r4
1119	and	r8,lr,r0,lsr#8
1120	eor	r2,r2,r5
1121	and	r9,lr,r0
1122	eor	r3,r3,r6
1123	mov	r0,r0,lsr#24
1124
1125	subs	r12,r12,#1
1126	bne	.Ldec_loop
1127
1128	add	r10,r10,#1024
1129
1130	ldr	r5,[r10,#0]		@ prefetch Td4
1131	ldr	r6,[r10,#32]
1132	ldr	r4,[r10,#64]
1133	ldr	r5,[r10,#96]
1134	ldr	r6,[r10,#128]
1135	ldr	r4,[r10,#160]
1136	ldr	r5,[r10,#192]
1137	ldr	r6,[r10,#224]
1138
1139	ldrb	r0,[r10,r0]		@ Td4[s0>>24]
1140	ldrb	r4,[r10,r7]		@ Td4[s0>>16]
1141	and	r7,lr,r1		@ i0
1142	ldrb	r5,[r10,r8]		@ Td4[s0>>8]
1143	and	r8,lr,r1,lsr#16
1144	ldrb	r6,[r10,r9]		@ Td4[s0>>0]
1145	and	r9,lr,r1,lsr#8
1146
1147	add	r1,r10,r1,lsr#24
1148	ldrb	r7,[r10,r7]		@ Td4[s1>>0]
1149	ldrb	r1,[r1]		@ Td4[s1>>24]
1150	ldrb	r8,[r10,r8]		@ Td4[s1>>16]
1151	eor	r0,r7,r0,lsl#24
1152	ldrb	r9,[r10,r9]		@ Td4[s1>>8]
1153	eor	r1,r4,r1,lsl#8
1154	and	r7,lr,r2,lsr#8	@ i0
1155	eor	r5,r5,r8,lsl#8
1156	and	r8,lr,r2		@ i1
1157	ldrb	r7,[r10,r7]		@ Td4[s2>>8]
1158	eor	r6,r6,r9,lsl#8
1159	ldrb	r8,[r10,r8]		@ Td4[s2>>0]
1160	and	r9,lr,r2,lsr#16
1161
1162	add	r2,r10,r2,lsr#24
1163	ldrb	r2,[r2]		@ Td4[s2>>24]
1164	eor	r0,r0,r7,lsl#8
1165	ldrb	r9,[r10,r9]		@ Td4[s2>>16]
1166	eor	r1,r8,r1,lsl#16
1167	and	r7,lr,r3,lsr#16	@ i0
1168	eor	r2,r5,r2,lsl#16
1169	and	r8,lr,r3,lsr#8	@ i1
1170	ldrb	r7,[r10,r7]		@ Td4[s3>>16]
1171	eor	r6,r6,r9,lsl#16
1172	ldrb	r8,[r10,r8]		@ Td4[s3>>8]
1173	and	r9,lr,r3		@ i2
1174
1175	add	r3,r10,r3,lsr#24
1176	ldrb	r9,[r10,r9]		@ Td4[s3>>0]
1177	ldrb	r3,[r3]		@ Td4[s3>>24]
1178	eor	r0,r0,r7,lsl#16
1179	ldr	r7,[r11,#0]
1180	eor	r1,r1,r8,lsl#8
1181	ldr	r4,[r11,#4]
1182	eor	r2,r9,r2,lsl#8
1183	ldr	r5,[r11,#8]
1184	eor	r3,r6,r3,lsl#24
1185	ldr	r6,[r11,#12]
1186
1187	eor	r0,r0,r7
1188	eor	r1,r1,r4
1189	eor	r2,r2,r5
1190	eor	r3,r3,r6
1191
1192	sub	r10,r10,#1024
1193	ldr	pc,[sp],#4		@ pop and return
1194.size	_armv4_AES_decrypt,.-_armv4_AES_decrypt
1195.byte	65,69,83,32,102,111,114,32,65,82,77,118,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
1196.align	2
1197.align	2
1198