xref: /freebsd/sys/crypto/openssl/arm/aes-armv4.S (revision 2a63c3be158216222d89a073dcbd6a72ee4aab5a) 
1/* Do not modify. This file is auto-generated from aes-armv4.pl. */
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#ifndef __KERNEL__
43# include "arm_arch.h"
44#else
45# define __ARM_ARCH__ __LINUX_ARM_ARCH__
46#endif
47
48.text
49#if defined(__thumb2__) && !defined(__APPLE__)
50.syntax	unified
51.thumb
52#else
53.code	32
54#undef __thumb2__
55#endif
56
57.type	AES_Te,%object
58.align	5
59AES_Te:
60.word	0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
61.word	0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
62.word	0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
63.word	0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
64.word	0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
65.word	0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
66.word	0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
67.word	0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
68.word	0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
69.word	0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
70.word	0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
71.word	0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
72.word	0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
73.word	0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
74.word	0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
75.word	0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
76.word	0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
77.word	0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
78.word	0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
79.word	0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
80.word	0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
81.word	0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
82.word	0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
83.word	0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
84.word	0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
85.word	0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
86.word	0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
87.word	0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
88.word	0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
89.word	0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
90.word	0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
91.word	0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
92.word	0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
93.word	0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
94.word	0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
95.word	0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
96.word	0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
97.word	0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
98.word	0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
99.word	0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
100.word	0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
101.word	0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
102.word	0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
103.word	0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
104.word	0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
105.word	0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
106.word	0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
107.word	0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
108.word	0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
109.word	0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
110.word	0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
111.word	0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
112.word	0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
113.word	0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
114.word	0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
115.word	0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
116.word	0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
117.word	0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
118.word	0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
119.word	0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
120.word	0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
121.word	0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
122.word	0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
123.word	0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
124@ Te4[256]
125.byte	0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
126.byte	0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
127.byte	0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
128.byte	0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
129.byte	0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
130.byte	0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
131.byte	0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
132.byte	0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
133.byte	0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
134.byte	0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
135.byte	0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
136.byte	0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
137.byte	0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
138.byte	0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
139.byte	0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
140.byte	0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
141.byte	0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
142.byte	0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
143.byte	0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
144.byte	0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
145.byte	0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
146.byte	0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
147.byte	0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
148.byte	0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
149.byte	0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
150.byte	0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
151.byte	0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
152.byte	0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
153.byte	0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
154.byte	0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
155.byte	0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
156.byte	0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
157@ rcon[]
158.word	0x01000000, 0x02000000, 0x04000000, 0x08000000
159.word	0x10000000, 0x20000000, 0x40000000, 0x80000000
160.word	0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
161.size	AES_Te,.-AES_Te
162
163@ void AES_encrypt(const unsigned char *in, unsigned char *out,
164@ 		 const AES_KEY *key) {
165.globl	AES_encrypt
166.type	AES_encrypt,%function
167.align	5
168AES_encrypt:
169#ifndef	__thumb2__
170	sub	r3,pc,#8		@ AES_encrypt
171#else
172	adr	r3,.
173#endif
174	stmdb	sp!,{r1,r4-r12,lr}
175#if defined(__thumb2__) || defined(__APPLE__)
176	adr	r10,AES_Te
177#else
178	sub	r10,r3,#AES_encrypt-AES_Te	@ Te
179#endif
180	mov	r12,r0		@ inp
181	mov	r11,r2
182#if __ARM_ARCH__<7
183	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
184	ldrb	r4,[r12,#2]	@ manner...
185	ldrb	r5,[r12,#1]
186	ldrb	r6,[r12,#0]
187	orr	r0,r0,r4,lsl#8
188	ldrb	r1,[r12,#7]
189	orr	r0,r0,r5,lsl#16
190	ldrb	r4,[r12,#6]
191	orr	r0,r0,r6,lsl#24
192	ldrb	r5,[r12,#5]
193	ldrb	r6,[r12,#4]
194	orr	r1,r1,r4,lsl#8
195	ldrb	r2,[r12,#11]
196	orr	r1,r1,r5,lsl#16
197	ldrb	r4,[r12,#10]
198	orr	r1,r1,r6,lsl#24
199	ldrb	r5,[r12,#9]
200	ldrb	r6,[r12,#8]
201	orr	r2,r2,r4,lsl#8
202	ldrb	r3,[r12,#15]
203	orr	r2,r2,r5,lsl#16
204	ldrb	r4,[r12,#14]
205	orr	r2,r2,r6,lsl#24
206	ldrb	r5,[r12,#13]
207	ldrb	r6,[r12,#12]
208	orr	r3,r3,r4,lsl#8
209	orr	r3,r3,r5,lsl#16
210	orr	r3,r3,r6,lsl#24
211#else
212	ldr	r0,[r12,#0]
213	ldr	r1,[r12,#4]
214	ldr	r2,[r12,#8]
215	ldr	r3,[r12,#12]
216#ifdef __ARMEL__
217	rev	r0,r0
218	rev	r1,r1
219	rev	r2,r2
220	rev	r3,r3
221#endif
222#endif
223	bl	_armv4_AES_encrypt
224
225	ldr	r12,[sp],#4		@ pop out
226#if __ARM_ARCH__>=7
227#ifdef __ARMEL__
228	rev	r0,r0
229	rev	r1,r1
230	rev	r2,r2
231	rev	r3,r3
232#endif
233	str	r0,[r12,#0]
234	str	r1,[r12,#4]
235	str	r2,[r12,#8]
236	str	r3,[r12,#12]
237#else
238	mov	r4,r0,lsr#24		@ write output in endian-neutral
239	mov	r5,r0,lsr#16		@ manner...
240	mov	r6,r0,lsr#8
241	strb	r4,[r12,#0]
242	strb	r5,[r12,#1]
243	mov	r4,r1,lsr#24
244	strb	r6,[r12,#2]
245	mov	r5,r1,lsr#16
246	strb	r0,[r12,#3]
247	mov	r6,r1,lsr#8
248	strb	r4,[r12,#4]
249	strb	r5,[r12,#5]
250	mov	r4,r2,lsr#24
251	strb	r6,[r12,#6]
252	mov	r5,r2,lsr#16
253	strb	r1,[r12,#7]
254	mov	r6,r2,lsr#8
255	strb	r4,[r12,#8]
256	strb	r5,[r12,#9]
257	mov	r4,r3,lsr#24
258	strb	r6,[r12,#10]
259	mov	r5,r3,lsr#16
260	strb	r2,[r12,#11]
261	mov	r6,r3,lsr#8
262	strb	r4,[r12,#12]
263	strb	r5,[r12,#13]
264	strb	r6,[r12,#14]
265	strb	r3,[r12,#15]
266#endif
267#if __ARM_ARCH__>=5
268	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
269#else
270	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
271	tst	lr,#1
272	moveq	pc,lr			@ be binary compatible with V4, yet
273.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
274#endif
275.size	AES_encrypt,.-AES_encrypt
276
277.type	_armv4_AES_encrypt,%function
278.align	2
279_armv4_AES_encrypt:
280	str	lr,[sp,#-4]!		@ push lr
281	ldmia	r11!,{r4,r5,r6,r7}
282	eor	r0,r0,r4
283	ldr	r12,[r11,#240-16]
284	eor	r1,r1,r5
285	eor	r2,r2,r6
286	eor	r3,r3,r7
287	sub	r12,r12,#1
288	mov	lr,#255
289
290	and	r7,lr,r0
291	and	r8,lr,r0,lsr#8
292	and	r9,lr,r0,lsr#16
293	mov	r0,r0,lsr#24
294.Lenc_loop:
295	ldr	r4,[r10,r7,lsl#2]	@ Te3[s0>>0]
296	and	r7,lr,r1,lsr#16	@ i0
297	ldr	r5,[r10,r8,lsl#2]	@ Te2[s0>>8]
298	and	r8,lr,r1
299	ldr	r6,[r10,r9,lsl#2]	@ Te1[s0>>16]
300	and	r9,lr,r1,lsr#8
301	ldr	r0,[r10,r0,lsl#2]	@ Te0[s0>>24]
302	mov	r1,r1,lsr#24
303
304	ldr	r7,[r10,r7,lsl#2]	@ Te1[s1>>16]
305	ldr	r8,[r10,r8,lsl#2]	@ Te3[s1>>0]
306	ldr	r9,[r10,r9,lsl#2]	@ Te2[s1>>8]
307	eor	r0,r0,r7,ror#8
308	ldr	r1,[r10,r1,lsl#2]	@ Te0[s1>>24]
309	and	r7,lr,r2,lsr#8	@ i0
310	eor	r5,r5,r8,ror#8
311	and	r8,lr,r2,lsr#16	@ i1
312	eor	r6,r6,r9,ror#8
313	and	r9,lr,r2
314	ldr	r7,[r10,r7,lsl#2]	@ Te2[s2>>8]
315	eor	r1,r1,r4,ror#24
316	ldr	r8,[r10,r8,lsl#2]	@ Te1[s2>>16]
317	mov	r2,r2,lsr#24
318
319	ldr	r9,[r10,r9,lsl#2]	@ Te3[s2>>0]
320	eor	r0,r0,r7,ror#16
321	ldr	r2,[r10,r2,lsl#2]	@ Te0[s2>>24]
322	and	r7,lr,r3		@ i0
323	eor	r1,r1,r8,ror#8
324	and	r8,lr,r3,lsr#8	@ i1
325	eor	r6,r6,r9,ror#16
326	and	r9,lr,r3,lsr#16	@ i2
327	ldr	r7,[r10,r7,lsl#2]	@ Te3[s3>>0]
328	eor	r2,r2,r5,ror#16
329	ldr	r8,[r10,r8,lsl#2]	@ Te2[s3>>8]
330	mov	r3,r3,lsr#24
331
332	ldr	r9,[r10,r9,lsl#2]	@ Te1[s3>>16]
333	eor	r0,r0,r7,ror#24
334	ldr	r7,[r11],#16
335	eor	r1,r1,r8,ror#16
336	ldr	r3,[r10,r3,lsl#2]	@ Te0[s3>>24]
337	eor	r2,r2,r9,ror#8
338	ldr	r4,[r11,#-12]
339	eor	r3,r3,r6,ror#8
340
341	ldr	r5,[r11,#-8]
342	eor	r0,r0,r7
343	ldr	r6,[r11,#-4]
344	and	r7,lr,r0
345	eor	r1,r1,r4
346	and	r8,lr,r0,lsr#8
347	eor	r2,r2,r5
348	and	r9,lr,r0,lsr#16
349	eor	r3,r3,r6
350	mov	r0,r0,lsr#24
351
352	subs	r12,r12,#1
353	bne	.Lenc_loop
354
355	add	r10,r10,#2
356
357	ldrb	r4,[r10,r7,lsl#2]	@ Te4[s0>>0]
358	and	r7,lr,r1,lsr#16	@ i0
359	ldrb	r5,[r10,r8,lsl#2]	@ Te4[s0>>8]
360	and	r8,lr,r1
361	ldrb	r6,[r10,r9,lsl#2]	@ Te4[s0>>16]
362	and	r9,lr,r1,lsr#8
363	ldrb	r0,[r10,r0,lsl#2]	@ Te4[s0>>24]
364	mov	r1,r1,lsr#24
365
366	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s1>>16]
367	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s1>>0]
368	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s1>>8]
369	eor	r0,r7,r0,lsl#8
370	ldrb	r1,[r10,r1,lsl#2]	@ Te4[s1>>24]
371	and	r7,lr,r2,lsr#8	@ i0
372	eor	r5,r8,r5,lsl#8
373	and	r8,lr,r2,lsr#16	@ i1
374	eor	r6,r9,r6,lsl#8
375	and	r9,lr,r2
376	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s2>>8]
377	eor	r1,r4,r1,lsl#24
378	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s2>>16]
379	mov	r2,r2,lsr#24
380
381	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s2>>0]
382	eor	r0,r7,r0,lsl#8
383	ldrb	r2,[r10,r2,lsl#2]	@ Te4[s2>>24]
384	and	r7,lr,r3		@ i0
385	eor	r1,r1,r8,lsl#16
386	and	r8,lr,r3,lsr#8	@ i1
387	eor	r6,r9,r6,lsl#8
388	and	r9,lr,r3,lsr#16	@ i2
389	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s3>>0]
390	eor	r2,r5,r2,lsl#24
391	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s3>>8]
392	mov	r3,r3,lsr#24
393
394	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s3>>16]
395	eor	r0,r7,r0,lsl#8
396	ldr	r7,[r11,#0]
397	ldrb	r3,[r10,r3,lsl#2]	@ Te4[s3>>24]
398	eor	r1,r1,r8,lsl#8
399	ldr	r4,[r11,#4]
400	eor	r2,r2,r9,lsl#16
401	ldr	r5,[r11,#8]
402	eor	r3,r6,r3,lsl#24
403	ldr	r6,[r11,#12]
404
405	eor	r0,r0,r7
406	eor	r1,r1,r4
407	eor	r2,r2,r5
408	eor	r3,r3,r6
409
410	sub	r10,r10,#2
411	ldr	pc,[sp],#4		@ pop and return
412.size	_armv4_AES_encrypt,.-_armv4_AES_encrypt
413
414.globl	AES_set_encrypt_key
415.type	AES_set_encrypt_key,%function
416.align	5
417AES_set_encrypt_key:
418_armv4_AES_set_encrypt_key:
419#ifndef	__thumb2__
420	sub	r3,pc,#8		@ AES_set_encrypt_key
421#else
422	adr	r3,.
423#endif
424	teq	r0,#0
425#ifdef	__thumb2__
426	itt	eq			@ Thumb2 thing, sanity check in ARM
427#endif
428	moveq	r0,#-1
429	beq	.Labrt
430	teq	r2,#0
431#ifdef	__thumb2__
432	itt	eq			@ Thumb2 thing, sanity check in ARM
433#endif
434	moveq	r0,#-1
435	beq	.Labrt
436
437	teq	r1,#128
438	beq	.Lok
439	teq	r1,#192
440	beq	.Lok
441	teq	r1,#256
442#ifdef	__thumb2__
443	itt	ne			@ Thumb2 thing, sanity check in ARM
444#endif
445	movne	r0,#-1
446	bne	.Labrt
447
448.Lok:	stmdb	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
449	mov	r12,r0		@ inp
450	mov	lr,r1			@ bits
451	mov	r11,r2			@ key
452
453#if defined(__thumb2__) || defined(__APPLE__)
454	adr	r10,AES_Te+1024				@ Te4
455#else
456	sub	r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024	@ Te4
457#endif
458
459#if __ARM_ARCH__<7
460	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
461	ldrb	r4,[r12,#2]	@ manner...
462	ldrb	r5,[r12,#1]
463	ldrb	r6,[r12,#0]
464	orr	r0,r0,r4,lsl#8
465	ldrb	r1,[r12,#7]
466	orr	r0,r0,r5,lsl#16
467	ldrb	r4,[r12,#6]
468	orr	r0,r0,r6,lsl#24
469	ldrb	r5,[r12,#5]
470	ldrb	r6,[r12,#4]
471	orr	r1,r1,r4,lsl#8
472	ldrb	r2,[r12,#11]
473	orr	r1,r1,r5,lsl#16
474	ldrb	r4,[r12,#10]
475	orr	r1,r1,r6,lsl#24
476	ldrb	r5,[r12,#9]
477	ldrb	r6,[r12,#8]
478	orr	r2,r2,r4,lsl#8
479	ldrb	r3,[r12,#15]
480	orr	r2,r2,r5,lsl#16
481	ldrb	r4,[r12,#14]
482	orr	r2,r2,r6,lsl#24
483	ldrb	r5,[r12,#13]
484	ldrb	r6,[r12,#12]
485	orr	r3,r3,r4,lsl#8
486	str	r0,[r11],#16
487	orr	r3,r3,r5,lsl#16
488	str	r1,[r11,#-12]
489	orr	r3,r3,r6,lsl#24
490	str	r2,[r11,#-8]
491	str	r3,[r11,#-4]
492#else
493	ldr	r0,[r12,#0]
494	ldr	r1,[r12,#4]
495	ldr	r2,[r12,#8]
496	ldr	r3,[r12,#12]
497#ifdef __ARMEL__
498	rev	r0,r0
499	rev	r1,r1
500	rev	r2,r2
501	rev	r3,r3
502#endif
503	str	r0,[r11],#16
504	str	r1,[r11,#-12]
505	str	r2,[r11,#-8]
506	str	r3,[r11,#-4]
507#endif
508
509	teq	lr,#128
510	bne	.Lnot128
511	mov	r12,#10
512	str	r12,[r11,#240-16]
513	add	r6,r10,#256			@ rcon
514	mov	lr,#255
515
516.L128_loop:
517	and	r5,lr,r3,lsr#24
518	and	r7,lr,r3,lsr#16
519	ldrb	r5,[r10,r5]
520	and	r8,lr,r3,lsr#8
521	ldrb	r7,[r10,r7]
522	and	r9,lr,r3
523	ldrb	r8,[r10,r8]
524	orr	r5,r5,r7,lsl#24
525	ldrb	r9,[r10,r9]
526	orr	r5,r5,r8,lsl#16
527	ldr	r4,[r6],#4			@ rcon[i++]
528	orr	r5,r5,r9,lsl#8
529	eor	r5,r5,r4
530	eor	r0,r0,r5			@ rk[4]=rk[0]^...
531	eor	r1,r1,r0			@ rk[5]=rk[1]^rk[4]
532	str	r0,[r11],#16
533	eor	r2,r2,r1			@ rk[6]=rk[2]^rk[5]
534	str	r1,[r11,#-12]
535	eor	r3,r3,r2			@ rk[7]=rk[3]^rk[6]
536	str	r2,[r11,#-8]
537	subs	r12,r12,#1
538	str	r3,[r11,#-4]
539	bne	.L128_loop
540	sub	r2,r11,#176
541	b	.Ldone
542
543.Lnot128:
544#if __ARM_ARCH__<7
545	ldrb	r8,[r12,#19]
546	ldrb	r4,[r12,#18]
547	ldrb	r5,[r12,#17]
548	ldrb	r6,[r12,#16]
549	orr	r8,r8,r4,lsl#8
550	ldrb	r9,[r12,#23]
551	orr	r8,r8,r5,lsl#16
552	ldrb	r4,[r12,#22]
553	orr	r8,r8,r6,lsl#24
554	ldrb	r5,[r12,#21]
555	ldrb	r6,[r12,#20]
556	orr	r9,r9,r4,lsl#8
557	orr	r9,r9,r5,lsl#16
558	str	r8,[r11],#8
559	orr	r9,r9,r6,lsl#24
560	str	r9,[r11,#-4]
561#else
562	ldr	r8,[r12,#16]
563	ldr	r9,[r12,#20]
564#ifdef __ARMEL__
565	rev	r8,r8
566	rev	r9,r9
567#endif
568	str	r8,[r11],#8
569	str	r9,[r11,#-4]
570#endif
571
572	teq	lr,#192
573	bne	.Lnot192
574	mov	r12,#12
575	str	r12,[r11,#240-24]
576	add	r6,r10,#256			@ rcon
577	mov	lr,#255
578	mov	r12,#8
579
580.L192_loop:
581	and	r5,lr,r9,lsr#24
582	and	r7,lr,r9,lsr#16
583	ldrb	r5,[r10,r5]
584	and	r8,lr,r9,lsr#8
585	ldrb	r7,[r10,r7]
586	and	r9,lr,r9
587	ldrb	r8,[r10,r8]
588	orr	r5,r5,r7,lsl#24
589	ldrb	r9,[r10,r9]
590	orr	r5,r5,r8,lsl#16
591	ldr	r4,[r6],#4			@ rcon[i++]
592	orr	r5,r5,r9,lsl#8
593	eor	r9,r5,r4
594	eor	r0,r0,r9			@ rk[6]=rk[0]^...
595	eor	r1,r1,r0			@ rk[7]=rk[1]^rk[6]
596	str	r0,[r11],#24
597	eor	r2,r2,r1			@ rk[8]=rk[2]^rk[7]
598	str	r1,[r11,#-20]
599	eor	r3,r3,r2			@ rk[9]=rk[3]^rk[8]
600	str	r2,[r11,#-16]
601	subs	r12,r12,#1
602	str	r3,[r11,#-12]
603#ifdef	__thumb2__
604	itt	eq				@ Thumb2 thing, sanity check in ARM
605#endif
606	subeq	r2,r11,#216
607	beq	.Ldone
608
609	ldr	r7,[r11,#-32]
610	ldr	r8,[r11,#-28]
611	eor	r7,r7,r3			@ rk[10]=rk[4]^rk[9]
612	eor	r9,r8,r7			@ rk[11]=rk[5]^rk[10]
613	str	r7,[r11,#-8]
614	str	r9,[r11,#-4]
615	b	.L192_loop
616
617.Lnot192:
618#if __ARM_ARCH__<7
619	ldrb	r8,[r12,#27]
620	ldrb	r4,[r12,#26]
621	ldrb	r5,[r12,#25]
622	ldrb	r6,[r12,#24]
623	orr	r8,r8,r4,lsl#8
624	ldrb	r9,[r12,#31]
625	orr	r8,r8,r5,lsl#16
626	ldrb	r4,[r12,#30]
627	orr	r8,r8,r6,lsl#24
628	ldrb	r5,[r12,#29]
629	ldrb	r6,[r12,#28]
630	orr	r9,r9,r4,lsl#8
631	orr	r9,r9,r5,lsl#16
632	str	r8,[r11],#8
633	orr	r9,r9,r6,lsl#24
634	str	r9,[r11,#-4]
635#else
636	ldr	r8,[r12,#24]
637	ldr	r9,[r12,#28]
638#ifdef __ARMEL__
639	rev	r8,r8
640	rev	r9,r9
641#endif
642	str	r8,[r11],#8
643	str	r9,[r11,#-4]
644#endif
645
646	mov	r12,#14
647	str	r12,[r11,#240-32]
648	add	r6,r10,#256			@ rcon
649	mov	lr,#255
650	mov	r12,#7
651
652.L256_loop:
653	and	r5,lr,r9,lsr#24
654	and	r7,lr,r9,lsr#16
655	ldrb	r5,[r10,r5]
656	and	r8,lr,r9,lsr#8
657	ldrb	r7,[r10,r7]
658	and	r9,lr,r9
659	ldrb	r8,[r10,r8]
660	orr	r5,r5,r7,lsl#24
661	ldrb	r9,[r10,r9]
662	orr	r5,r5,r8,lsl#16
663	ldr	r4,[r6],#4			@ rcon[i++]
664	orr	r5,r5,r9,lsl#8
665	eor	r9,r5,r4
666	eor	r0,r0,r9			@ rk[8]=rk[0]^...
667	eor	r1,r1,r0			@ rk[9]=rk[1]^rk[8]
668	str	r0,[r11],#32
669	eor	r2,r2,r1			@ rk[10]=rk[2]^rk[9]
670	str	r1,[r11,#-28]
671	eor	r3,r3,r2			@ rk[11]=rk[3]^rk[10]
672	str	r2,[r11,#-24]
673	subs	r12,r12,#1
674	str	r3,[r11,#-20]
675#ifdef	__thumb2__
676	itt	eq				@ Thumb2 thing, sanity check in ARM
677#endif
678	subeq	r2,r11,#256
679	beq	.Ldone
680
681	and	r5,lr,r3
682	and	r7,lr,r3,lsr#8
683	ldrb	r5,[r10,r5]
684	and	r8,lr,r3,lsr#16
685	ldrb	r7,[r10,r7]
686	and	r9,lr,r3,lsr#24
687	ldrb	r8,[r10,r8]
688	orr	r5,r5,r7,lsl#8
689	ldrb	r9,[r10,r9]
690	orr	r5,r5,r8,lsl#16
691	ldr	r4,[r11,#-48]
692	orr	r5,r5,r9,lsl#24
693
694	ldr	r7,[r11,#-44]
695	ldr	r8,[r11,#-40]
696	eor	r4,r4,r5			@ rk[12]=rk[4]^...
697	ldr	r9,[r11,#-36]
698	eor	r7,r7,r4			@ rk[13]=rk[5]^rk[12]
699	str	r4,[r11,#-16]
700	eor	r8,r8,r7			@ rk[14]=rk[6]^rk[13]
701	str	r7,[r11,#-12]
702	eor	r9,r9,r8			@ rk[15]=rk[7]^rk[14]
703	str	r8,[r11,#-8]
704	str	r9,[r11,#-4]
705	b	.L256_loop
706
707.align	2
708.Ldone:	mov	r0,#0
709	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
710.Labrt:
711#if __ARM_ARCH__>=5
712	bx	lr				@ .word	0xe12fff1e
713#else
714	tst	lr,#1
715	moveq	pc,lr			@ be binary compatible with V4, yet
716.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
717#endif
718.size	AES_set_encrypt_key,.-AES_set_encrypt_key
719
720.globl	AES_set_decrypt_key
721.type	AES_set_decrypt_key,%function
722.align	5
723AES_set_decrypt_key:
724	str	lr,[sp,#-4]!            @ push lr
725	bl	_armv4_AES_set_encrypt_key
726	teq	r0,#0
727	ldr	lr,[sp],#4              @ pop lr
728	bne	.Labrt
729
730	mov	r0,r2			@ AES_set_encrypt_key preserves r2,
731	mov	r1,r2			@ which is AES_KEY *key
732	b	_armv4_AES_set_enc2dec_key
733.size	AES_set_decrypt_key,.-AES_set_decrypt_key
734
735@ void AES_set_enc2dec_key(const AES_KEY *inp,AES_KEY *out)
736.globl	AES_set_enc2dec_key
737.type	AES_set_enc2dec_key,%function
738.align	5
739AES_set_enc2dec_key:
740_armv4_AES_set_enc2dec_key:
741	stmdb	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
742
743	ldr	r12,[r0,#240]
744	mov	r7,r0			@ input
745	add	r8,r0,r12,lsl#4
746	mov	r11,r1			@ output
747	add	r10,r1,r12,lsl#4
748	str	r12,[r1,#240]
749
750.Linv:	ldr	r0,[r7],#16
751	ldr	r1,[r7,#-12]
752	ldr	r2,[r7,#-8]
753	ldr	r3,[r7,#-4]
754	ldr	r4,[r8],#-16
755	ldr	r5,[r8,#16+4]
756	ldr	r6,[r8,#16+8]
757	ldr	r9,[r8,#16+12]
758	str	r0,[r10],#-16
759	str	r1,[r10,#16+4]
760	str	r2,[r10,#16+8]
761	str	r3,[r10,#16+12]
762	str	r4,[r11],#16
763	str	r5,[r11,#-12]
764	str	r6,[r11,#-8]
765	str	r9,[r11,#-4]
766	teq	r7,r8
767	bne	.Linv
768
769	ldr	r0,[r7]
770	ldr	r1,[r7,#4]
771	ldr	r2,[r7,#8]
772	ldr	r3,[r7,#12]
773	str	r0,[r11]
774	str	r1,[r11,#4]
775	str	r2,[r11,#8]
776	str	r3,[r11,#12]
777	sub	r11,r11,r12,lsl#3
778	ldr	r0,[r11,#16]!		@ prefetch tp1
779	mov	r7,#0x80
780	mov	r8,#0x1b
781	orr	r7,r7,#0x8000
782	orr	r8,r8,#0x1b00
783	orr	r7,r7,r7,lsl#16
784	orr	r8,r8,r8,lsl#16
785	sub	r12,r12,#1
786	mvn	r9,r7
787	mov	r12,r12,lsl#2	@ (rounds-1)*4
788
789.Lmix:	and	r4,r0,r7
790	and	r1,r0,r9
791	sub	r4,r4,r4,lsr#7
792	and	r4,r4,r8
793	eor	r1,r4,r1,lsl#1	@ tp2
794
795	and	r4,r1,r7
796	and	r2,r1,r9
797	sub	r4,r4,r4,lsr#7
798	and	r4,r4,r8
799	eor	r2,r4,r2,lsl#1	@ tp4
800
801	and	r4,r2,r7
802	and	r3,r2,r9
803	sub	r4,r4,r4,lsr#7
804	and	r4,r4,r8
805	eor	r3,r4,r3,lsl#1	@ tp8
806
807	eor	r4,r1,r2
808	eor	r5,r0,r3		@ tp9
809	eor	r4,r4,r3		@ tpe
810	eor	r4,r4,r1,ror#24
811	eor	r4,r4,r5,ror#24	@ ^= ROTATE(tpb=tp9^tp2,8)
812	eor	r4,r4,r2,ror#16
813	eor	r4,r4,r5,ror#16	@ ^= ROTATE(tpd=tp9^tp4,16)
814	eor	r4,r4,r5,ror#8	@ ^= ROTATE(tp9,24)
815
816	ldr	r0,[r11,#4]		@ prefetch tp1
817	str	r4,[r11],#4
818	subs	r12,r12,#1
819	bne	.Lmix
820
821	mov	r0,#0
822#if __ARM_ARCH__>=5
823	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
824#else
825	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
826	tst	lr,#1
827	moveq	pc,lr			@ be binary compatible with V4, yet
828.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
829#endif
830.size	AES_set_enc2dec_key,.-AES_set_enc2dec_key
831
832.type	AES_Td,%object
833.align	5
834AES_Td:
835.word	0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
836.word	0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
837.word	0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
838.word	0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
839.word	0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
840.word	0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
841.word	0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
842.word	0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
843.word	0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
844.word	0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
845.word	0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
846.word	0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
847.word	0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
848.word	0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
849.word	0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
850.word	0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
851.word	0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
852.word	0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
853.word	0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
854.word	0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
855.word	0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
856.word	0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
857.word	0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
858.word	0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
859.word	0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
860.word	0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
861.word	0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
862.word	0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
863.word	0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
864.word	0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
865.word	0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
866.word	0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
867.word	0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
868.word	0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
869.word	0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
870.word	0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
871.word	0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
872.word	0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
873.word	0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
874.word	0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
875.word	0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
876.word	0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
877.word	0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
878.word	0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
879.word	0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
880.word	0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
881.word	0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
882.word	0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
883.word	0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
884.word	0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
885.word	0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
886.word	0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
887.word	0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
888.word	0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
889.word	0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
890.word	0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
891.word	0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
892.word	0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
893.word	0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
894.word	0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
895.word	0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
896.word	0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
897.word	0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
898.word	0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
899@ Td4[256]
900.byte	0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
901.byte	0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
902.byte	0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
903.byte	0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
904.byte	0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
905.byte	0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
906.byte	0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
907.byte	0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
908.byte	0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
909.byte	0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
910.byte	0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
911.byte	0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
912.byte	0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
913.byte	0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
914.byte	0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
915.byte	0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
916.byte	0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
917.byte	0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
918.byte	0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
919.byte	0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
920.byte	0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
921.byte	0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
922.byte	0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
923.byte	0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
924.byte	0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
925.byte	0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
926.byte	0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
927.byte	0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
928.byte	0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
929.byte	0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
930.byte	0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
931.byte	0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
932.size	AES_Td,.-AES_Td
933
934@ void AES_decrypt(const unsigned char *in, unsigned char *out,
935@ 		 const AES_KEY *key) {
936.globl	AES_decrypt
937.type	AES_decrypt,%function
938.align	5
939AES_decrypt:
940#ifndef	__thumb2__
941	sub	r3,pc,#8		@ AES_decrypt
942#else
943	adr	r3,.
944#endif
945	stmdb	sp!,{r1,r4-r12,lr}
946#if defined(__thumb2__) || defined(__APPLE__)
947	adr	r10,AES_Td
948#else
949	sub	r10,r3,#AES_decrypt-AES_Td	@ Td
950#endif
951	mov	r12,r0		@ inp
952	mov	r11,r2
953#if __ARM_ARCH__<7
954	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
955	ldrb	r4,[r12,#2]	@ manner...
956	ldrb	r5,[r12,#1]
957	ldrb	r6,[r12,#0]
958	orr	r0,r0,r4,lsl#8
959	ldrb	r1,[r12,#7]
960	orr	r0,r0,r5,lsl#16
961	ldrb	r4,[r12,#6]
962	orr	r0,r0,r6,lsl#24
963	ldrb	r5,[r12,#5]
964	ldrb	r6,[r12,#4]
965	orr	r1,r1,r4,lsl#8
966	ldrb	r2,[r12,#11]
967	orr	r1,r1,r5,lsl#16
968	ldrb	r4,[r12,#10]
969	orr	r1,r1,r6,lsl#24
970	ldrb	r5,[r12,#9]
971	ldrb	r6,[r12,#8]
972	orr	r2,r2,r4,lsl#8
973	ldrb	r3,[r12,#15]
974	orr	r2,r2,r5,lsl#16
975	ldrb	r4,[r12,#14]
976	orr	r2,r2,r6,lsl#24
977	ldrb	r5,[r12,#13]
978	ldrb	r6,[r12,#12]
979	orr	r3,r3,r4,lsl#8
980	orr	r3,r3,r5,lsl#16
981	orr	r3,r3,r6,lsl#24
982#else
983	ldr	r0,[r12,#0]
984	ldr	r1,[r12,#4]
985	ldr	r2,[r12,#8]
986	ldr	r3,[r12,#12]
987#ifdef __ARMEL__
988	rev	r0,r0
989	rev	r1,r1
990	rev	r2,r2
991	rev	r3,r3
992#endif
993#endif
994	bl	_armv4_AES_decrypt
995
996	ldr	r12,[sp],#4		@ pop out
997#if __ARM_ARCH__>=7
998#ifdef __ARMEL__
999	rev	r0,r0
1000	rev	r1,r1
1001	rev	r2,r2
1002	rev	r3,r3
1003#endif
1004	str	r0,[r12,#0]
1005	str	r1,[r12,#4]
1006	str	r2,[r12,#8]
1007	str	r3,[r12,#12]
1008#else
1009	mov	r4,r0,lsr#24		@ write output in endian-neutral
1010	mov	r5,r0,lsr#16		@ manner...
1011	mov	r6,r0,lsr#8
1012	strb	r4,[r12,#0]
1013	strb	r5,[r12,#1]
1014	mov	r4,r1,lsr#24
1015	strb	r6,[r12,#2]
1016	mov	r5,r1,lsr#16
1017	strb	r0,[r12,#3]
1018	mov	r6,r1,lsr#8
1019	strb	r4,[r12,#4]
1020	strb	r5,[r12,#5]
1021	mov	r4,r2,lsr#24
1022	strb	r6,[r12,#6]
1023	mov	r5,r2,lsr#16
1024	strb	r1,[r12,#7]
1025	mov	r6,r2,lsr#8
1026	strb	r4,[r12,#8]
1027	strb	r5,[r12,#9]
1028	mov	r4,r3,lsr#24
1029	strb	r6,[r12,#10]
1030	mov	r5,r3,lsr#16
1031	strb	r2,[r12,#11]
1032	mov	r6,r3,lsr#8
1033	strb	r4,[r12,#12]
1034	strb	r5,[r12,#13]
1035	strb	r6,[r12,#14]
1036	strb	r3,[r12,#15]
1037#endif
1038#if __ARM_ARCH__>=5
1039	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,pc}
1040#else
1041	ldmia	sp!,{r4,r5,r6,r7,r8,r9,r10,r11,r12,lr}
1042	tst	lr,#1
1043	moveq	pc,lr			@ be binary compatible with V4, yet
1044.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
1045#endif
1046.size	AES_decrypt,.-AES_decrypt
1047
1048.type	_armv4_AES_decrypt,%function
1049.align	2
1050_armv4_AES_decrypt:
1051	str	lr,[sp,#-4]!		@ push lr
1052	ldmia	r11!,{r4,r5,r6,r7}
1053	eor	r0,r0,r4
1054	ldr	r12,[r11,#240-16]
1055	eor	r1,r1,r5
1056	eor	r2,r2,r6
1057	eor	r3,r3,r7
1058	sub	r12,r12,#1
1059	mov	lr,#255
1060
1061	and	r7,lr,r0,lsr#16
1062	and	r8,lr,r0,lsr#8
1063	and	r9,lr,r0
1064	mov	r0,r0,lsr#24
1065.Ldec_loop:
1066	ldr	r4,[r10,r7,lsl#2]	@ Td1[s0>>16]
1067	and	r7,lr,r1		@ i0
1068	ldr	r5,[r10,r8,lsl#2]	@ Td2[s0>>8]
1069	and	r8,lr,r1,lsr#16
1070	ldr	r6,[r10,r9,lsl#2]	@ Td3[s0>>0]
1071	and	r9,lr,r1,lsr#8
1072	ldr	r0,[r10,r0,lsl#2]	@ Td0[s0>>24]
1073	mov	r1,r1,lsr#24
1074
1075	ldr	r7,[r10,r7,lsl#2]	@ Td3[s1>>0]
1076	ldr	r8,[r10,r8,lsl#2]	@ Td1[s1>>16]
1077	ldr	r9,[r10,r9,lsl#2]	@ Td2[s1>>8]
1078	eor	r0,r0,r7,ror#24
1079	ldr	r1,[r10,r1,lsl#2]	@ Td0[s1>>24]
1080	and	r7,lr,r2,lsr#8	@ i0
1081	eor	r5,r8,r5,ror#8
1082	and	r8,lr,r2		@ i1
1083	eor	r6,r9,r6,ror#8
1084	and	r9,lr,r2,lsr#16
1085	ldr	r7,[r10,r7,lsl#2]	@ Td2[s2>>8]
1086	eor	r1,r1,r4,ror#8
1087	ldr	r8,[r10,r8,lsl#2]	@ Td3[s2>>0]
1088	mov	r2,r2,lsr#24
1089
1090	ldr	r9,[r10,r9,lsl#2]	@ Td1[s2>>16]
1091	eor	r0,r0,r7,ror#16
1092	ldr	r2,[r10,r2,lsl#2]	@ Td0[s2>>24]
1093	and	r7,lr,r3,lsr#16	@ i0
1094	eor	r1,r1,r8,ror#24
1095	and	r8,lr,r3,lsr#8	@ i1
1096	eor	r6,r9,r6,ror#8
1097	and	r9,lr,r3		@ i2
1098	ldr	r7,[r10,r7,lsl#2]	@ Td1[s3>>16]
1099	eor	r2,r2,r5,ror#8
1100	ldr	r8,[r10,r8,lsl#2]	@ Td2[s3>>8]
1101	mov	r3,r3,lsr#24
1102
1103	ldr	r9,[r10,r9,lsl#2]	@ Td3[s3>>0]
1104	eor	r0,r0,r7,ror#8
1105	ldr	r7,[r11],#16
1106	eor	r1,r1,r8,ror#16
1107	ldr	r3,[r10,r3,lsl#2]	@ Td0[s3>>24]
1108	eor	r2,r2,r9,ror#24
1109
1110	ldr	r4,[r11,#-12]
1111	eor	r0,r0,r7
1112	ldr	r5,[r11,#-8]
1113	eor	r3,r3,r6,ror#8
1114	ldr	r6,[r11,#-4]
1115	and	r7,lr,r0,lsr#16
1116	eor	r1,r1,r4
1117	and	r8,lr,r0,lsr#8
1118	eor	r2,r2,r5
1119	and	r9,lr,r0
1120	eor	r3,r3,r6
1121	mov	r0,r0,lsr#24
1122
1123	subs	r12,r12,#1
1124	bne	.Ldec_loop
1125
1126	add	r10,r10,#1024
1127
1128	ldr	r5,[r10,#0]		@ prefetch Td4
1129	ldr	r6,[r10,#32]
1130	ldr	r4,[r10,#64]
1131	ldr	r5,[r10,#96]
1132	ldr	r6,[r10,#128]
1133	ldr	r4,[r10,#160]
1134	ldr	r5,[r10,#192]
1135	ldr	r6,[r10,#224]
1136
1137	ldrb	r0,[r10,r0]		@ Td4[s0>>24]
1138	ldrb	r4,[r10,r7]		@ Td4[s0>>16]
1139	and	r7,lr,r1		@ i0
1140	ldrb	r5,[r10,r8]		@ Td4[s0>>8]
1141	and	r8,lr,r1,lsr#16
1142	ldrb	r6,[r10,r9]		@ Td4[s0>>0]
1143	and	r9,lr,r1,lsr#8
1144
1145	add	r1,r10,r1,lsr#24
1146	ldrb	r7,[r10,r7]		@ Td4[s1>>0]
1147	ldrb	r1,[r1]		@ Td4[s1>>24]
1148	ldrb	r8,[r10,r8]		@ Td4[s1>>16]
1149	eor	r0,r7,r0,lsl#24
1150	ldrb	r9,[r10,r9]		@ Td4[s1>>8]
1151	eor	r1,r4,r1,lsl#8
1152	and	r7,lr,r2,lsr#8	@ i0
1153	eor	r5,r5,r8,lsl#8
1154	and	r8,lr,r2		@ i1
1155	ldrb	r7,[r10,r7]		@ Td4[s2>>8]
1156	eor	r6,r6,r9,lsl#8
1157	ldrb	r8,[r10,r8]		@ Td4[s2>>0]
1158	and	r9,lr,r2,lsr#16
1159
1160	add	r2,r10,r2,lsr#24
1161	ldrb	r2,[r2]		@ Td4[s2>>24]
1162	eor	r0,r0,r7,lsl#8
1163	ldrb	r9,[r10,r9]		@ Td4[s2>>16]
1164	eor	r1,r8,r1,lsl#16
1165	and	r7,lr,r3,lsr#16	@ i0
1166	eor	r2,r5,r2,lsl#16
1167	and	r8,lr,r3,lsr#8	@ i1
1168	ldrb	r7,[r10,r7]		@ Td4[s3>>16]
1169	eor	r6,r6,r9,lsl#16
1170	ldrb	r8,[r10,r8]		@ Td4[s3>>8]
1171	and	r9,lr,r3		@ i2
1172
1173	add	r3,r10,r3,lsr#24
1174	ldrb	r9,[r10,r9]		@ Td4[s3>>0]
1175	ldrb	r3,[r3]		@ Td4[s3>>24]
1176	eor	r0,r0,r7,lsl#16
1177	ldr	r7,[r11,#0]
1178	eor	r1,r1,r8,lsl#8
1179	ldr	r4,[r11,#4]
1180	eor	r2,r9,r2,lsl#8
1181	ldr	r5,[r11,#8]
1182	eor	r3,r6,r3,lsl#24
1183	ldr	r6,[r11,#12]
1184
1185	eor	r0,r0,r7
1186	eor	r1,r1,r4
1187	eor	r2,r2,r5
1188	eor	r3,r3,r6
1189
1190	sub	r10,r10,#1024
1191	ldr	pc,[sp],#4		@ pop and return
1192.size	_armv4_AES_decrypt,.-_armv4_AES_decrypt
1193.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
1194.align	2
1195.align	2
1196