]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - secure/lib/libcrypto/arm/aes-armv4.S
MFC: r304636
[FreeBSD/FreeBSD.git] / secure / lib / libcrypto / arm / aes-armv4.S
1 # $FreeBSD$
2 # Do not modify. This file is auto-generated from aes-armv4.pl.
3
4 @ ====================================================================
5 @ Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
6 @ project. The module is, however, dual licensed under OpenSSL and
7 @ CRYPTOGAMS licenses depending on where you obtain it. For further
8 @ details see http://www.openssl.org/~appro/cryptogams/.
9 @ ====================================================================
10
11 @ AES for ARMv4
12
13 @ January 2007.
14 @
15 @ Code uses single 1K S-box and is >2 times faster than code generated
16 @ by gcc-3.4.1. This is thanks to unique feature of ARMv4 ISA, which
17 @ allows to merge logical or arithmetic operation with shift or rotate
18 @ in one instruction and emit combined result every cycle. The module
19 @ is endian-neutral. The performance is ~42 cycles/byte for 128-bit
20 @ key [on single-issue Xscale PXA250 core].
21
22 @ May 2007.
23 @
24 @ AES_set_[en|de]crypt_key is added.
25
26 @ July 2010.
27 @
28 @ Rescheduling for dual-issue pipeline resulted in 12% improvement on
29 @ Cortex A8 core and ~25 cycles per byte processed with 128-bit key.
30
31 @ February 2011.
32 @
33 @ Profiler-assisted and platform-specific optimization resulted in 16%
34 @ improvement on Cortex A8 core and ~21.5 cycles per byte.
35
36 #ifndef __KERNEL__
37 # include "arm_arch.h"
38 #else
39 # define __ARM_ARCH__ __LINUX_ARM_ARCH__
40 #endif
41
42 .text
43 #if __ARM_ARCH__<7
44 .code   32
45 #else
46 .syntax unified
47 # ifdef __thumb2__
48 .thumb
49 # else
50 .code   32
51 # endif
52 #endif
53
54 .type   AES_Te,%object
55 .align  5
56 AES_Te:
57 .word   0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
58 .word   0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
59 .word   0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
60 .word   0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
61 .word   0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
62 .word   0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
63 .word   0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
64 .word   0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
65 .word   0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
66 .word   0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
67 .word   0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
68 .word   0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
69 .word   0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
70 .word   0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
71 .word   0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
72 .word   0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
73 .word   0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
74 .word   0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
75 .word   0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
76 .word   0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
77 .word   0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
78 .word   0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
79 .word   0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
80 .word   0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
81 .word   0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
82 .word   0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
83 .word   0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
84 .word   0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
85 .word   0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
86 .word   0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
87 .word   0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
88 .word   0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
89 .word   0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
90 .word   0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
91 .word   0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
92 .word   0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
93 .word   0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
94 .word   0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
95 .word   0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
96 .word   0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
97 .word   0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
98 .word   0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
99 .word   0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
100 .word   0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
101 .word   0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
102 .word   0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
103 .word   0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
104 .word   0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
105 .word   0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
106 .word   0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
107 .word   0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
108 .word   0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
109 .word   0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
110 .word   0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
111 .word   0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
112 .word   0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
113 .word   0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
114 .word   0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
115 .word   0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
116 .word   0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
117 .word   0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
118 .word   0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
119 .word   0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
120 .word   0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
121 @ Te4[256]
122 .byte   0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
123 .byte   0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
124 .byte   0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
125 .byte   0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
126 .byte   0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
127 .byte   0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
128 .byte   0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
129 .byte   0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
130 .byte   0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
131 .byte   0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
132 .byte   0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
133 .byte   0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
134 .byte   0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
135 .byte   0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
136 .byte   0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
137 .byte   0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
138 .byte   0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
139 .byte   0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
140 .byte   0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
141 .byte   0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
142 .byte   0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
143 .byte   0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
144 .byte   0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
145 .byte   0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
146 .byte   0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
147 .byte   0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
148 .byte   0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
149 .byte   0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
150 .byte   0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
151 .byte   0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
152 .byte   0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
153 .byte   0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
154 @ rcon[]
155 .word   0x01000000, 0x02000000, 0x04000000, 0x08000000
156 .word   0x10000000, 0x20000000, 0x40000000, 0x80000000
157 .word   0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
158 .size   AES_Te,.-AES_Te
159
160 @ void AES_encrypt(const unsigned char *in, unsigned char *out,
161 @                const AES_KEY *key) {
162 .global AES_encrypt
163 .type   AES_encrypt,%function
164 .align  5
165 AES_encrypt:
166 #if __ARM_ARCH__<7
167         sub     r3,pc,#8                @ AES_encrypt
168 #else
169         adr     r3,AES_encrypt
170 #endif
171         stmdb   sp!,{r1,r4-r12,lr}
172         mov     r12,r0          @ inp
173         mov     r11,r2
174         sub     r10,r3,#AES_encrypt-AES_Te      @ Te
175 #if __ARM_ARCH__<7
176         ldrb    r0,[r12,#3]     @ load input data in endian-neutral
177         ldrb    r4,[r12,#2]     @ manner...
178         ldrb    r5,[r12,#1]
179         ldrb    r6,[r12,#0]
180         orr     r0,r0,r4,lsl#8
181         ldrb    r1,[r12,#7]
182         orr     r0,r0,r5,lsl#16
183         ldrb    r4,[r12,#6]
184         orr     r0,r0,r6,lsl#24
185         ldrb    r5,[r12,#5]
186         ldrb    r6,[r12,#4]
187         orr     r1,r1,r4,lsl#8
188         ldrb    r2,[r12,#11]
189         orr     r1,r1,r5,lsl#16
190         ldrb    r4,[r12,#10]
191         orr     r1,r1,r6,lsl#24
192         ldrb    r5,[r12,#9]
193         ldrb    r6,[r12,#8]
194         orr     r2,r2,r4,lsl#8
195         ldrb    r3,[r12,#15]
196         orr     r2,r2,r5,lsl#16
197         ldrb    r4,[r12,#14]
198         orr     r2,r2,r6,lsl#24
199         ldrb    r5,[r12,#13]
200         ldrb    r6,[r12,#12]
201         orr     r3,r3,r4,lsl#8
202         orr     r3,r3,r5,lsl#16
203         orr     r3,r3,r6,lsl#24
204 #else
205         ldr     r0,[r12,#0]
206         ldr     r1,[r12,#4]
207         ldr     r2,[r12,#8]
208         ldr     r3,[r12,#12]
209 #ifdef __ARMEL__
210         rev     r0,r0
211         rev     r1,r1
212         rev     r2,r2
213         rev     r3,r3
214 #endif
215 #endif
216         bl      _armv4_AES_encrypt
217
218         ldr     r12,[sp],#4             @ pop out
219 #if __ARM_ARCH__>=7
220 #ifdef __ARMEL__
221         rev     r0,r0
222         rev     r1,r1
223         rev     r2,r2
224         rev     r3,r3
225 #endif
226         str     r0,[r12,#0]
227         str     r1,[r12,#4]
228         str     r2,[r12,#8]
229         str     r3,[r12,#12]
230 #else
231         mov     r4,r0,lsr#24            @ write output in endian-neutral
232         mov     r5,r0,lsr#16            @ manner...
233         mov     r6,r0,lsr#8
234         strb    r4,[r12,#0]
235         strb    r5,[r12,#1]
236         mov     r4,r1,lsr#24
237         strb    r6,[r12,#2]
238         mov     r5,r1,lsr#16
239         strb    r0,[r12,#3]
240         mov     r6,r1,lsr#8
241         strb    r4,[r12,#4]
242         strb    r5,[r12,#5]
243         mov     r4,r2,lsr#24
244         strb    r6,[r12,#6]
245         mov     r5,r2,lsr#16
246         strb    r1,[r12,#7]
247         mov     r6,r2,lsr#8
248         strb    r4,[r12,#8]
249         strb    r5,[r12,#9]
250         mov     r4,r3,lsr#24
251         strb    r6,[r12,#10]
252         mov     r5,r3,lsr#16
253         strb    r2,[r12,#11]
254         mov     r6,r3,lsr#8
255         strb    r4,[r12,#12]
256         strb    r5,[r12,#13]
257         strb    r6,[r12,#14]
258         strb    r3,[r12,#15]
259 #endif
260 #if __ARM_ARCH__>=5
261         ldmia   sp!,{r4-r12,pc}
262 #else
263         ldmia   sp!,{r4-r12,lr}
264         tst     lr,#1
265         moveq   pc,lr                   @ be binary compatible with V4, yet
266         .word   0xe12fff1e                      @ interoperable with Thumb ISA:-)
267 #endif
268 .size   AES_encrypt,.-AES_encrypt
269
270 .type   _armv4_AES_encrypt,%function
271 .align  2
272 _armv4_AES_encrypt:
273         str     lr,[sp,#-4]!            @ push lr
274         ldmia   r11!,{r4-r7}
275         eor     r0,r0,r4
276         ldr     r12,[r11,#240-16]
277         eor     r1,r1,r5
278         eor     r2,r2,r6
279         eor     r3,r3,r7
280         sub     r12,r12,#1
281         mov     lr,#255
282
283         and     r7,lr,r0
284         and     r8,lr,r0,lsr#8
285         and     r9,lr,r0,lsr#16
286         mov     r0,r0,lsr#24
287 .Lenc_loop:
288         ldr     r4,[r10,r7,lsl#2]       @ Te3[s0>>0]
289         and     r7,lr,r1,lsr#16 @ i0
290         ldr     r5,[r10,r8,lsl#2]       @ Te2[s0>>8]
291         and     r8,lr,r1
292         ldr     r6,[r10,r9,lsl#2]       @ Te1[s0>>16]
293         and     r9,lr,r1,lsr#8
294         ldr     r0,[r10,r0,lsl#2]       @ Te0[s0>>24]
295         mov     r1,r1,lsr#24
296
297         ldr     r7,[r10,r7,lsl#2]       @ Te1[s1>>16]
298         ldr     r8,[r10,r8,lsl#2]       @ Te3[s1>>0]
299         ldr     r9,[r10,r9,lsl#2]       @ Te2[s1>>8]
300         eor     r0,r0,r7,ror#8
301         ldr     r1,[r10,r1,lsl#2]       @ Te0[s1>>24]
302         and     r7,lr,r2,lsr#8  @ i0
303         eor     r5,r5,r8,ror#8
304         and     r8,lr,r2,lsr#16 @ i1
305         eor     r6,r6,r9,ror#8
306         and     r9,lr,r2
307         ldr     r7,[r10,r7,lsl#2]       @ Te2[s2>>8]
308         eor     r1,r1,r4,ror#24
309         ldr     r8,[r10,r8,lsl#2]       @ Te1[s2>>16]
310         mov     r2,r2,lsr#24
311
312         ldr     r9,[r10,r9,lsl#2]       @ Te3[s2>>0]
313         eor     r0,r0,r7,ror#16
314         ldr     r2,[r10,r2,lsl#2]       @ Te0[s2>>24]
315         and     r7,lr,r3                @ i0
316         eor     r1,r1,r8,ror#8
317         and     r8,lr,r3,lsr#8  @ i1
318         eor     r6,r6,r9,ror#16
319         and     r9,lr,r3,lsr#16 @ i2
320         ldr     r7,[r10,r7,lsl#2]       @ Te3[s3>>0]
321         eor     r2,r2,r5,ror#16
322         ldr     r8,[r10,r8,lsl#2]       @ Te2[s3>>8]
323         mov     r3,r3,lsr#24
324
325         ldr     r9,[r10,r9,lsl#2]       @ Te1[s3>>16]
326         eor     r0,r0,r7,ror#24
327         ldr     r7,[r11],#16
328         eor     r1,r1,r8,ror#16
329         ldr     r3,[r10,r3,lsl#2]       @ Te0[s3>>24]
330         eor     r2,r2,r9,ror#8
331         ldr     r4,[r11,#-12]
332         eor     r3,r3,r6,ror#8
333
334         ldr     r5,[r11,#-8]
335         eor     r0,r0,r7
336         ldr     r6,[r11,#-4]
337         and     r7,lr,r0
338         eor     r1,r1,r4
339         and     r8,lr,r0,lsr#8
340         eor     r2,r2,r5
341         and     r9,lr,r0,lsr#16
342         eor     r3,r3,r6
343         mov     r0,r0,lsr#24
344
345         subs    r12,r12,#1
346         bne     .Lenc_loop
347
348         add     r10,r10,#2
349
350         ldrb    r4,[r10,r7,lsl#2]       @ Te4[s0>>0]
351         and     r7,lr,r1,lsr#16 @ i0
352         ldrb    r5,[r10,r8,lsl#2]       @ Te4[s0>>8]
353         and     r8,lr,r1
354         ldrb    r6,[r10,r9,lsl#2]       @ Te4[s0>>16]
355         and     r9,lr,r1,lsr#8
356         ldrb    r0,[r10,r0,lsl#2]       @ Te4[s0>>24]
357         mov     r1,r1,lsr#24
358
359         ldrb    r7,[r10,r7,lsl#2]       @ Te4[s1>>16]
360         ldrb    r8,[r10,r8,lsl#2]       @ Te4[s1>>0]
361         ldrb    r9,[r10,r9,lsl#2]       @ Te4[s1>>8]
362         eor     r0,r7,r0,lsl#8
363         ldrb    r1,[r10,r1,lsl#2]       @ Te4[s1>>24]
364         and     r7,lr,r2,lsr#8  @ i0
365         eor     r5,r8,r5,lsl#8
366         and     r8,lr,r2,lsr#16 @ i1
367         eor     r6,r9,r6,lsl#8
368         and     r9,lr,r2
369         ldrb    r7,[r10,r7,lsl#2]       @ Te4[s2>>8]
370         eor     r1,r4,r1,lsl#24
371         ldrb    r8,[r10,r8,lsl#2]       @ Te4[s2>>16]
372         mov     r2,r2,lsr#24
373
374         ldrb    r9,[r10,r9,lsl#2]       @ Te4[s2>>0]
375         eor     r0,r7,r0,lsl#8
376         ldrb    r2,[r10,r2,lsl#2]       @ Te4[s2>>24]
377         and     r7,lr,r3                @ i0
378         eor     r1,r1,r8,lsl#16
379         and     r8,lr,r3,lsr#8  @ i1
380         eor     r6,r9,r6,lsl#8
381         and     r9,lr,r3,lsr#16 @ i2
382         ldrb    r7,[r10,r7,lsl#2]       @ Te4[s3>>0]
383         eor     r2,r5,r2,lsl#24
384         ldrb    r8,[r10,r8,lsl#2]       @ Te4[s3>>8]
385         mov     r3,r3,lsr#24
386
387         ldrb    r9,[r10,r9,lsl#2]       @ Te4[s3>>16]
388         eor     r0,r7,r0,lsl#8
389         ldr     r7,[r11,#0]
390         ldrb    r3,[r10,r3,lsl#2]       @ Te4[s3>>24]
391         eor     r1,r1,r8,lsl#8
392         ldr     r4,[r11,#4]
393         eor     r2,r2,r9,lsl#16
394         ldr     r5,[r11,#8]
395         eor     r3,r6,r3,lsl#24
396         ldr     r6,[r11,#12]
397
398         eor     r0,r0,r7
399         eor     r1,r1,r4
400         eor     r2,r2,r5
401         eor     r3,r3,r6
402
403         sub     r10,r10,#2
404         ldr     pc,[sp],#4              @ pop and return
405 .size   _armv4_AES_encrypt,.-_armv4_AES_encrypt
406
407 .global private_AES_set_encrypt_key
408 .type   private_AES_set_encrypt_key,%function
409 .align  5
410 private_AES_set_encrypt_key:
411 _armv4_AES_set_encrypt_key:
412 #if __ARM_ARCH__<7
413         sub     r3,pc,#8                @ AES_set_encrypt_key
414 #else
415         adr     r3,private_AES_set_encrypt_key
416 #endif
417         teq     r0,#0
418 #if __ARM_ARCH__>=7
419         itt     eq                      @ Thumb2 thing, sanity check in ARM
420 #endif
421         moveq   r0,#-1
422         beq     .Labrt
423         teq     r2,#0
424 #if __ARM_ARCH__>=7
425         itt     eq                      @ Thumb2 thing, sanity check in ARM
426 #endif
427         moveq   r0,#-1
428         beq     .Labrt
429
430         teq     r1,#128
431         beq     .Lok
432         teq     r1,#192
433         beq     .Lok
434         teq     r1,#256
435 #if __ARM_ARCH__>=7
436         itt     ne                      @ Thumb2 thing, sanity check in ARM
437 #endif
438         movne   r0,#-1
439         bne     .Labrt
440
441 .Lok:   stmdb   sp!,{r4-r12,lr}
442         sub     r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024  @ Te4
443
444         mov     r12,r0          @ inp
445         mov     lr,r1                   @ bits
446         mov     r11,r2                  @ key
447
448 #if __ARM_ARCH__<7
449         ldrb    r0,[r12,#3]     @ load input data in endian-neutral
450         ldrb    r4,[r12,#2]     @ manner...
451         ldrb    r5,[r12,#1]
452         ldrb    r6,[r12,#0]
453         orr     r0,r0,r4,lsl#8
454         ldrb    r1,[r12,#7]
455         orr     r0,r0,r5,lsl#16
456         ldrb    r4,[r12,#6]
457         orr     r0,r0,r6,lsl#24
458         ldrb    r5,[r12,#5]
459         ldrb    r6,[r12,#4]
460         orr     r1,r1,r4,lsl#8
461         ldrb    r2,[r12,#11]
462         orr     r1,r1,r5,lsl#16
463         ldrb    r4,[r12,#10]
464         orr     r1,r1,r6,lsl#24
465         ldrb    r5,[r12,#9]
466         ldrb    r6,[r12,#8]
467         orr     r2,r2,r4,lsl#8
468         ldrb    r3,[r12,#15]
469         orr     r2,r2,r5,lsl#16
470         ldrb    r4,[r12,#14]
471         orr     r2,r2,r6,lsl#24
472         ldrb    r5,[r12,#13]
473         ldrb    r6,[r12,#12]
474         orr     r3,r3,r4,lsl#8
475         str     r0,[r11],#16
476         orr     r3,r3,r5,lsl#16
477         str     r1,[r11,#-12]
478         orr     r3,r3,r6,lsl#24
479         str     r2,[r11,#-8]
480         str     r3,[r11,#-4]
481 #else
482         ldr     r0,[r12,#0]
483         ldr     r1,[r12,#4]
484         ldr     r2,[r12,#8]
485         ldr     r3,[r12,#12]
486 #ifdef __ARMEL__
487         rev     r0,r0
488         rev     r1,r1
489         rev     r2,r2
490         rev     r3,r3
491 #endif
492         str     r0,[r11],#16
493         str     r1,[r11,#-12]
494         str     r2,[r11,#-8]
495         str     r3,[r11,#-4]
496 #endif
497
498         teq     lr,#128
499         bne     .Lnot128
500         mov     r12,#10
501         str     r12,[r11,#240-16]
502         add     r6,r10,#256                     @ rcon
503         mov     lr,#255
504
505 .L128_loop:
506         and     r5,lr,r3,lsr#24
507         and     r7,lr,r3,lsr#16
508         ldrb    r5,[r10,r5]
509         and     r8,lr,r3,lsr#8
510         ldrb    r7,[r10,r7]
511         and     r9,lr,r3
512         ldrb    r8,[r10,r8]
513         orr     r5,r5,r7,lsl#24
514         ldrb    r9,[r10,r9]
515         orr     r5,r5,r8,lsl#16
516         ldr     r4,[r6],#4                      @ rcon[i++]
517         orr     r5,r5,r9,lsl#8
518         eor     r5,r5,r4
519         eor     r0,r0,r5                        @ rk[4]=rk[0]^...
520         eor     r1,r1,r0                        @ rk[5]=rk[1]^rk[4]
521         str     r0,[r11],#16
522         eor     r2,r2,r1                        @ rk[6]=rk[2]^rk[5]
523         str     r1,[r11,#-12]
524         eor     r3,r3,r2                        @ rk[7]=rk[3]^rk[6]
525         str     r2,[r11,#-8]
526         subs    r12,r12,#1
527         str     r3,[r11,#-4]
528         bne     .L128_loop
529         sub     r2,r11,#176
530         b       .Ldone
531
532 .Lnot128:
533 #if __ARM_ARCH__<7
534         ldrb    r8,[r12,#19]
535         ldrb    r4,[r12,#18]
536         ldrb    r5,[r12,#17]
537         ldrb    r6,[r12,#16]
538         orr     r8,r8,r4,lsl#8
539         ldrb    r9,[r12,#23]
540         orr     r8,r8,r5,lsl#16
541         ldrb    r4,[r12,#22]
542         orr     r8,r8,r6,lsl#24
543         ldrb    r5,[r12,#21]
544         ldrb    r6,[r12,#20]
545         orr     r9,r9,r4,lsl#8
546         orr     r9,r9,r5,lsl#16
547         str     r8,[r11],#8
548         orr     r9,r9,r6,lsl#24
549         str     r9,[r11,#-4]
550 #else
551         ldr     r8,[r12,#16]
552         ldr     r9,[r12,#20]
553 #ifdef __ARMEL__
554         rev     r8,r8
555         rev     r9,r9
556 #endif
557         str     r8,[r11],#8
558         str     r9,[r11,#-4]
559 #endif
560
561         teq     lr,#192
562         bne     .Lnot192
563         mov     r12,#12
564         str     r12,[r11,#240-24]
565         add     r6,r10,#256                     @ rcon
566         mov     lr,#255
567         mov     r12,#8
568
569 .L192_loop:
570         and     r5,lr,r9,lsr#24
571         and     r7,lr,r9,lsr#16
572         ldrb    r5,[r10,r5]
573         and     r8,lr,r9,lsr#8
574         ldrb    r7,[r10,r7]
575         and     r9,lr,r9
576         ldrb    r8,[r10,r8]
577         orr     r5,r5,r7,lsl#24
578         ldrb    r9,[r10,r9]
579         orr     r5,r5,r8,lsl#16
580         ldr     r4,[r6],#4                      @ rcon[i++]
581         orr     r5,r5,r9,lsl#8
582         eor     r9,r5,r4
583         eor     r0,r0,r9                        @ rk[6]=rk[0]^...
584         eor     r1,r1,r0                        @ rk[7]=rk[1]^rk[6]
585         str     r0,[r11],#24
586         eor     r2,r2,r1                        @ rk[8]=rk[2]^rk[7]
587         str     r1,[r11,#-20]
588         eor     r3,r3,r2                        @ rk[9]=rk[3]^rk[8]
589         str     r2,[r11,#-16]
590         subs    r12,r12,#1
591         str     r3,[r11,#-12]
592 #if __ARM_ARCH__>=7
593         itt     eq                              @ Thumb2 thing, sanity check in ARM
594 #endif
595         subeq   r2,r11,#216
596         beq     .Ldone
597
598         ldr     r7,[r11,#-32]
599         ldr     r8,[r11,#-28]
600         eor     r7,r7,r3                        @ rk[10]=rk[4]^rk[9]
601         eor     r9,r8,r7                        @ rk[11]=rk[5]^rk[10]
602         str     r7,[r11,#-8]
603         str     r9,[r11,#-4]
604         b       .L192_loop
605
606 .Lnot192:
607 #if __ARM_ARCH__<7
608         ldrb    r8,[r12,#27]
609         ldrb    r4,[r12,#26]
610         ldrb    r5,[r12,#25]
611         ldrb    r6,[r12,#24]
612         orr     r8,r8,r4,lsl#8
613         ldrb    r9,[r12,#31]
614         orr     r8,r8,r5,lsl#16
615         ldrb    r4,[r12,#30]
616         orr     r8,r8,r6,lsl#24
617         ldrb    r5,[r12,#29]
618         ldrb    r6,[r12,#28]
619         orr     r9,r9,r4,lsl#8
620         orr     r9,r9,r5,lsl#16
621         str     r8,[r11],#8
622         orr     r9,r9,r6,lsl#24
623         str     r9,[r11,#-4]
624 #else
625         ldr     r8,[r12,#24]
626         ldr     r9,[r12,#28]
627 #ifdef __ARMEL__
628         rev     r8,r8
629         rev     r9,r9
630 #endif
631         str     r8,[r11],#8
632         str     r9,[r11,#-4]
633 #endif
634
635         mov     r12,#14
636         str     r12,[r11,#240-32]
637         add     r6,r10,#256                     @ rcon
638         mov     lr,#255
639         mov     r12,#7
640
641 .L256_loop:
642         and     r5,lr,r9,lsr#24
643         and     r7,lr,r9,lsr#16
644         ldrb    r5,[r10,r5]
645         and     r8,lr,r9,lsr#8
646         ldrb    r7,[r10,r7]
647         and     r9,lr,r9
648         ldrb    r8,[r10,r8]
649         orr     r5,r5,r7,lsl#24
650         ldrb    r9,[r10,r9]
651         orr     r5,r5,r8,lsl#16
652         ldr     r4,[r6],#4                      @ rcon[i++]
653         orr     r5,r5,r9,lsl#8
654         eor     r9,r5,r4
655         eor     r0,r0,r9                        @ rk[8]=rk[0]^...
656         eor     r1,r1,r0                        @ rk[9]=rk[1]^rk[8]
657         str     r0,[r11],#32
658         eor     r2,r2,r1                        @ rk[10]=rk[2]^rk[9]
659         str     r1,[r11,#-28]
660         eor     r3,r3,r2                        @ rk[11]=rk[3]^rk[10]
661         str     r2,[r11,#-24]
662         subs    r12,r12,#1
663         str     r3,[r11,#-20]
664 #if __ARM_ARCH__>=7
665         itt     eq                              @ Thumb2 thing, sanity check in ARM
666 #endif
667         subeq   r2,r11,#256
668         beq     .Ldone
669
670         and     r5,lr,r3
671         and     r7,lr,r3,lsr#8
672         ldrb    r5,[r10,r5]
673         and     r8,lr,r3,lsr#16
674         ldrb    r7,[r10,r7]
675         and     r9,lr,r3,lsr#24
676         ldrb    r8,[r10,r8]
677         orr     r5,r5,r7,lsl#8
678         ldrb    r9,[r10,r9]
679         orr     r5,r5,r8,lsl#16
680         ldr     r4,[r11,#-48]
681         orr     r5,r5,r9,lsl#24
682
683         ldr     r7,[r11,#-44]
684         ldr     r8,[r11,#-40]
685         eor     r4,r4,r5                        @ rk[12]=rk[4]^...
686         ldr     r9,[r11,#-36]
687         eor     r7,r7,r4                        @ rk[13]=rk[5]^rk[12]
688         str     r4,[r11,#-16]
689         eor     r8,r8,r7                        @ rk[14]=rk[6]^rk[13]
690         str     r7,[r11,#-12]
691         eor     r9,r9,r8                        @ rk[15]=rk[7]^rk[14]
692         str     r8,[r11,#-8]
693         str     r9,[r11,#-4]
694         b       .L256_loop
695
696 .align  2
697 .Ldone: mov     r0,#0
698         ldmia   sp!,{r4-r12,lr}
699 .Labrt:
700 #if __ARM_ARCH__>=5
701         bx      lr                              @ .word 0xe12fff1e
702 #else
703         tst     lr,#1
704         moveq   pc,lr                   @ be binary compatible with V4, yet
705         .word   0xe12fff1e                      @ interoperable with Thumb ISA:-)
706 #endif
707 .size   private_AES_set_encrypt_key,.-private_AES_set_encrypt_key
708
709 .global private_AES_set_decrypt_key
710 .type   private_AES_set_decrypt_key,%function
711 .align  5
712 private_AES_set_decrypt_key:
713         str     lr,[sp,#-4]!            @ push lr
714         bl      _armv4_AES_set_encrypt_key
715         teq     r0,#0
716         ldr     lr,[sp],#4              @ pop lr
717         bne     .Labrt
718
719         mov     r0,r2                   @ AES_set_encrypt_key preserves r2,
720         mov     r1,r2                   @ which is AES_KEY *key
721         b       _armv4_AES_set_enc2dec_key
722 .size   private_AES_set_decrypt_key,.-private_AES_set_decrypt_key
723
724 @ void AES_set_enc2dec_key(const AES_KEY *inp,AES_KEY *out)
725 .global AES_set_enc2dec_key
726 .type   AES_set_enc2dec_key,%function
727 .align  5
728 AES_set_enc2dec_key:
729 _armv4_AES_set_enc2dec_key:
730         stmdb   sp!,{r4-r12,lr}
731
732         ldr     r12,[r0,#240]
733         mov     r7,r0                   @ input
734         add     r8,r0,r12,lsl#4
735         mov     r11,r1                  @ ouput
736         add     r10,r1,r12,lsl#4
737         str     r12,[r1,#240]
738
739 .Linv:  ldr     r0,[r7],#16
740         ldr     r1,[r7,#-12]
741         ldr     r2,[r7,#-8]
742         ldr     r3,[r7,#-4]
743         ldr     r4,[r8],#-16
744         ldr     r5,[r8,#16+4]
745         ldr     r6,[r8,#16+8]
746         ldr     r9,[r8,#16+12]
747         str     r0,[r10],#-16
748         str     r1,[r10,#16+4]
749         str     r2,[r10,#16+8]
750         str     r3,[r10,#16+12]
751         str     r4,[r11],#16
752         str     r5,[r11,#-12]
753         str     r6,[r11,#-8]
754         str     r9,[r11,#-4]
755         teq     r7,r8
756         bne     .Linv
757
758         ldr     r0,[r7]
759         ldr     r1,[r7,#4]
760         ldr     r2,[r7,#8]
761         ldr     r3,[r7,#12]
762         str     r0,[r11]
763         str     r1,[r11,#4]
764         str     r2,[r11,#8]
765         str     r3,[r11,#12]
766         sub     r11,r11,r12,lsl#3
767         ldr     r0,[r11,#16]!           @ prefetch tp1
768         mov     r7,#0x80
769         mov     r8,#0x1b
770         orr     r7,r7,#0x8000
771         orr     r8,r8,#0x1b00
772         orr     r7,r7,r7,lsl#16
773         orr     r8,r8,r8,lsl#16
774         sub     r12,r12,#1
775         mvn     r9,r7
776         mov     r12,r12,lsl#2   @ (rounds-1)*4
777
778 .Lmix:  and     r4,r0,r7
779         and     r1,r0,r9
780         sub     r4,r4,r4,lsr#7
781         and     r4,r4,r8
782         eor     r1,r4,r1,lsl#1  @ tp2
783
784         and     r4,r1,r7
785         and     r2,r1,r9
786         sub     r4,r4,r4,lsr#7
787         and     r4,r4,r8
788         eor     r2,r4,r2,lsl#1  @ tp4
789
790         and     r4,r2,r7
791         and     r3,r2,r9
792         sub     r4,r4,r4,lsr#7
793         and     r4,r4,r8
794         eor     r3,r4,r3,lsl#1  @ tp8
795
796         eor     r4,r1,r2
797         eor     r5,r0,r3                @ tp9
798         eor     r4,r4,r3                @ tpe
799         eor     r4,r4,r1,ror#24
800         eor     r4,r4,r5,ror#24 @ ^= ROTATE(tpb=tp9^tp2,8)
801         eor     r4,r4,r2,ror#16
802         eor     r4,r4,r5,ror#16 @ ^= ROTATE(tpd=tp9^tp4,16)
803         eor     r4,r4,r5,ror#8  @ ^= ROTATE(tp9,24)
804
805         ldr     r0,[r11,#4]             @ prefetch tp1
806         str     r4,[r11],#4
807         subs    r12,r12,#1
808         bne     .Lmix
809
810         mov     r0,#0
811 #if __ARM_ARCH__>=5
812         ldmia   sp!,{r4-r12,pc}
813 #else
814         ldmia   sp!,{r4-r12,lr}
815         tst     lr,#1
816         moveq   pc,lr                   @ be binary compatible with V4, yet
817         .word   0xe12fff1e                      @ interoperable with Thumb ISA:-)
818 #endif
819 .size   AES_set_enc2dec_key,.-AES_set_enc2dec_key
820
821 .type   AES_Td,%object
822 .align  5
823 AES_Td:
824 .word   0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
825 .word   0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
826 .word   0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
827 .word   0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
828 .word   0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
829 .word   0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
830 .word   0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
831 .word   0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
832 .word   0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
833 .word   0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
834 .word   0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
835 .word   0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
836 .word   0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
837 .word   0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
838 .word   0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
839 .word   0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
840 .word   0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
841 .word   0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
842 .word   0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
843 .word   0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
844 .word   0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
845 .word   0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
846 .word   0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
847 .word   0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
848 .word   0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
849 .word   0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
850 .word   0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
851 .word   0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
852 .word   0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
853 .word   0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
854 .word   0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
855 .word   0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
856 .word   0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
857 .word   0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
858 .word   0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
859 .word   0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
860 .word   0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
861 .word   0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
862 .word   0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
863 .word   0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
864 .word   0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
865 .word   0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
866 .word   0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
867 .word   0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
868 .word   0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
869 .word   0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
870 .word   0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
871 .word   0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
872 .word   0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
873 .word   0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
874 .word   0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
875 .word   0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
876 .word   0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
877 .word   0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
878 .word   0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
879 .word   0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
880 .word   0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
881 .word   0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
882 .word   0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
883 .word   0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
884 .word   0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
885 .word   0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
886 .word   0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
887 .word   0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
888 @ Td4[256]
889 .byte   0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
890 .byte   0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
891 .byte   0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
892 .byte   0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
893 .byte   0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
894 .byte   0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
895 .byte   0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
896 .byte   0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
897 .byte   0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
898 .byte   0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
899 .byte   0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
900 .byte   0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
901 .byte   0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
902 .byte   0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
903 .byte   0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
904 .byte   0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
905 .byte   0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
906 .byte   0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
907 .byte   0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
908 .byte   0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
909 .byte   0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
910 .byte   0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
911 .byte   0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
912 .byte   0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
913 .byte   0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
914 .byte   0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
915 .byte   0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
916 .byte   0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
917 .byte   0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
918 .byte   0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
919 .byte   0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
920 .byte   0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
921 .size   AES_Td,.-AES_Td
922
923 @ void AES_decrypt(const unsigned char *in, unsigned char *out,
924 @                const AES_KEY *key) {
925 .global AES_decrypt
926 .type   AES_decrypt,%function
927 .align  5
928 AES_decrypt:
929 #if __ARM_ARCH__<7
930         sub     r3,pc,#8                @ AES_decrypt
931 #else
932         adr     r3,AES_decrypt
933 #endif
934         stmdb   sp!,{r1,r4-r12,lr}
935         mov     r12,r0          @ inp
936         mov     r11,r2
937         sub     r10,r3,#AES_decrypt-AES_Td              @ Td
938 #if __ARM_ARCH__<7
939         ldrb    r0,[r12,#3]     @ load input data in endian-neutral
940         ldrb    r4,[r12,#2]     @ manner...
941         ldrb    r5,[r12,#1]
942         ldrb    r6,[r12,#0]
943         orr     r0,r0,r4,lsl#8
944         ldrb    r1,[r12,#7]
945         orr     r0,r0,r5,lsl#16
946         ldrb    r4,[r12,#6]
947         orr     r0,r0,r6,lsl#24
948         ldrb    r5,[r12,#5]
949         ldrb    r6,[r12,#4]
950         orr     r1,r1,r4,lsl#8
951         ldrb    r2,[r12,#11]
952         orr     r1,r1,r5,lsl#16
953         ldrb    r4,[r12,#10]
954         orr     r1,r1,r6,lsl#24
955         ldrb    r5,[r12,#9]
956         ldrb    r6,[r12,#8]
957         orr     r2,r2,r4,lsl#8
958         ldrb    r3,[r12,#15]
959         orr     r2,r2,r5,lsl#16
960         ldrb    r4,[r12,#14]
961         orr     r2,r2,r6,lsl#24
962         ldrb    r5,[r12,#13]
963         ldrb    r6,[r12,#12]
964         orr     r3,r3,r4,lsl#8
965         orr     r3,r3,r5,lsl#16
966         orr     r3,r3,r6,lsl#24
967 #else
968         ldr     r0,[r12,#0]
969         ldr     r1,[r12,#4]
970         ldr     r2,[r12,#8]
971         ldr     r3,[r12,#12]
972 #ifdef __ARMEL__
973         rev     r0,r0
974         rev     r1,r1
975         rev     r2,r2
976         rev     r3,r3
977 #endif
978 #endif
979         bl      _armv4_AES_decrypt
980
981         ldr     r12,[sp],#4             @ pop out
982 #if __ARM_ARCH__>=7
983 #ifdef __ARMEL__
984         rev     r0,r0
985         rev     r1,r1
986         rev     r2,r2
987         rev     r3,r3
988 #endif
989         str     r0,[r12,#0]
990         str     r1,[r12,#4]
991         str     r2,[r12,#8]
992         str     r3,[r12,#12]
993 #else
994         mov     r4,r0,lsr#24            @ write output in endian-neutral
995         mov     r5,r0,lsr#16            @ manner...
996         mov     r6,r0,lsr#8
997         strb    r4,[r12,#0]
998         strb    r5,[r12,#1]
999         mov     r4,r1,lsr#24
1000         strb    r6,[r12,#2]
1001         mov     r5,r1,lsr#16
1002         strb    r0,[r12,#3]
1003         mov     r6,r1,lsr#8
1004         strb    r4,[r12,#4]
1005         strb    r5,[r12,#5]
1006         mov     r4,r2,lsr#24
1007         strb    r6,[r12,#6]
1008         mov     r5,r2,lsr#16
1009         strb    r1,[r12,#7]
1010         mov     r6,r2,lsr#8
1011         strb    r4,[r12,#8]
1012         strb    r5,[r12,#9]
1013         mov     r4,r3,lsr#24
1014         strb    r6,[r12,#10]
1015         mov     r5,r3,lsr#16
1016         strb    r2,[r12,#11]
1017         mov     r6,r3,lsr#8
1018         strb    r4,[r12,#12]
1019         strb    r5,[r12,#13]
1020         strb    r6,[r12,#14]
1021         strb    r3,[r12,#15]
1022 #endif
1023 #if __ARM_ARCH__>=5
1024         ldmia   sp!,{r4-r12,pc}
1025 #else
1026         ldmia   sp!,{r4-r12,lr}
1027         tst     lr,#1
1028         moveq   pc,lr                   @ be binary compatible with V4, yet
1029         .word   0xe12fff1e                      @ interoperable with Thumb ISA:-)
1030 #endif
1031 .size   AES_decrypt,.-AES_decrypt
1032
1033 .type   _armv4_AES_decrypt,%function
1034 .align  2
1035 _armv4_AES_decrypt:
1036         str     lr,[sp,#-4]!            @ push lr
1037         ldmia   r11!,{r4-r7}
1038         eor     r0,r0,r4
1039         ldr     r12,[r11,#240-16]
1040         eor     r1,r1,r5
1041         eor     r2,r2,r6
1042         eor     r3,r3,r7
1043         sub     r12,r12,#1
1044         mov     lr,#255
1045
1046         and     r7,lr,r0,lsr#16
1047         and     r8,lr,r0,lsr#8
1048         and     r9,lr,r0
1049         mov     r0,r0,lsr#24
1050 .Ldec_loop:
1051         ldr     r4,[r10,r7,lsl#2]       @ Td1[s0>>16]
1052         and     r7,lr,r1                @ i0
1053         ldr     r5,[r10,r8,lsl#2]       @ Td2[s0>>8]
1054         and     r8,lr,r1,lsr#16
1055         ldr     r6,[r10,r9,lsl#2]       @ Td3[s0>>0]
1056         and     r9,lr,r1,lsr#8
1057         ldr     r0,[r10,r0,lsl#2]       @ Td0[s0>>24]
1058         mov     r1,r1,lsr#24
1059
1060         ldr     r7,[r10,r7,lsl#2]       @ Td3[s1>>0]
1061         ldr     r8,[r10,r8,lsl#2]       @ Td1[s1>>16]
1062         ldr     r9,[r10,r9,lsl#2]       @ Td2[s1>>8]
1063         eor     r0,r0,r7,ror#24
1064         ldr     r1,[r10,r1,lsl#2]       @ Td0[s1>>24]
1065         and     r7,lr,r2,lsr#8  @ i0
1066         eor     r5,r8,r5,ror#8
1067         and     r8,lr,r2                @ i1
1068         eor     r6,r9,r6,ror#8
1069         and     r9,lr,r2,lsr#16
1070         ldr     r7,[r10,r7,lsl#2]       @ Td2[s2>>8]
1071         eor     r1,r1,r4,ror#8
1072         ldr     r8,[r10,r8,lsl#2]       @ Td3[s2>>0]
1073         mov     r2,r2,lsr#24
1074
1075         ldr     r9,[r10,r9,lsl#2]       @ Td1[s2>>16]
1076         eor     r0,r0,r7,ror#16
1077         ldr     r2,[r10,r2,lsl#2]       @ Td0[s2>>24]
1078         and     r7,lr,r3,lsr#16 @ i0
1079         eor     r1,r1,r8,ror#24
1080         and     r8,lr,r3,lsr#8  @ i1
1081         eor     r6,r9,r6,ror#8
1082         and     r9,lr,r3                @ i2
1083         ldr     r7,[r10,r7,lsl#2]       @ Td1[s3>>16]
1084         eor     r2,r2,r5,ror#8
1085         ldr     r8,[r10,r8,lsl#2]       @ Td2[s3>>8]
1086         mov     r3,r3,lsr#24
1087
1088         ldr     r9,[r10,r9,lsl#2]       @ Td3[s3>>0]
1089         eor     r0,r0,r7,ror#8
1090         ldr     r7,[r11],#16
1091         eor     r1,r1,r8,ror#16
1092         ldr     r3,[r10,r3,lsl#2]       @ Td0[s3>>24]
1093         eor     r2,r2,r9,ror#24
1094
1095         ldr     r4,[r11,#-12]
1096         eor     r0,r0,r7
1097         ldr     r5,[r11,#-8]
1098         eor     r3,r3,r6,ror#8
1099         ldr     r6,[r11,#-4]
1100         and     r7,lr,r0,lsr#16
1101         eor     r1,r1,r4
1102         and     r8,lr,r0,lsr#8
1103         eor     r2,r2,r5
1104         and     r9,lr,r0
1105         eor     r3,r3,r6
1106         mov     r0,r0,lsr#24
1107
1108         subs    r12,r12,#1
1109         bne     .Ldec_loop
1110
1111         add     r10,r10,#1024
1112
1113         ldr     r5,[r10,#0]             @ prefetch Td4
1114         ldr     r6,[r10,#32]
1115         ldr     r4,[r10,#64]
1116         ldr     r5,[r10,#96]
1117         ldr     r6,[r10,#128]
1118         ldr     r4,[r10,#160]
1119         ldr     r5,[r10,#192]
1120         ldr     r6,[r10,#224]
1121
1122         ldrb    r0,[r10,r0]             @ Td4[s0>>24]
1123         ldrb    r4,[r10,r7]             @ Td4[s0>>16]
1124         and     r7,lr,r1                @ i0
1125         ldrb    r5,[r10,r8]             @ Td4[s0>>8]
1126         and     r8,lr,r1,lsr#16
1127         ldrb    r6,[r10,r9]             @ Td4[s0>>0]
1128         and     r9,lr,r1,lsr#8
1129
1130         add     r1,r10,r1,lsr#24
1131         ldrb    r7,[r10,r7]             @ Td4[s1>>0]
1132         ldrb    r1,[r1]         @ Td4[s1>>24]
1133         ldrb    r8,[r10,r8]             @ Td4[s1>>16]
1134         eor     r0,r7,r0,lsl#24
1135         ldrb    r9,[r10,r9]             @ Td4[s1>>8]
1136         eor     r1,r4,r1,lsl#8
1137         and     r7,lr,r2,lsr#8  @ i0
1138         eor     r5,r5,r8,lsl#8
1139         and     r8,lr,r2                @ i1
1140         ldrb    r7,[r10,r7]             @ Td4[s2>>8]
1141         eor     r6,r6,r9,lsl#8
1142         ldrb    r8,[r10,r8]             @ Td4[s2>>0]
1143         and     r9,lr,r2,lsr#16
1144
1145         add     r2,r10,r2,lsr#24
1146         ldrb    r2,[r2]         @ Td4[s2>>24]
1147         eor     r0,r0,r7,lsl#8
1148         ldrb    r9,[r10,r9]             @ Td4[s2>>16]
1149         eor     r1,r8,r1,lsl#16
1150         and     r7,lr,r3,lsr#16 @ i0
1151         eor     r2,r5,r2,lsl#16
1152         and     r8,lr,r3,lsr#8  @ i1
1153         ldrb    r7,[r10,r7]             @ Td4[s3>>16]
1154         eor     r6,r6,r9,lsl#16
1155         ldrb    r8,[r10,r8]             @ Td4[s3>>8]
1156         and     r9,lr,r3                @ i2
1157
1158         add     r3,r10,r3,lsr#24
1159         ldrb    r9,[r10,r9]             @ Td4[s3>>0]
1160         ldrb    r3,[r3]         @ Td4[s3>>24]
1161         eor     r0,r0,r7,lsl#16
1162         ldr     r7,[r11,#0]
1163         eor     r1,r1,r8,lsl#8
1164         ldr     r4,[r11,#4]
1165         eor     r2,r9,r2,lsl#8
1166         ldr     r5,[r11,#8]
1167         eor     r3,r6,r3,lsl#24
1168         ldr     r6,[r11,#12]
1169
1170         eor     r0,r0,r7
1171         eor     r1,r1,r4
1172         eor     r2,r2,r5
1173         eor     r3,r3,r6
1174
1175         sub     r10,r10,#1024
1176         ldr     pc,[sp],#4              @ pop and return
1177 .size   _armv4_AES_decrypt,.-_armv4_AES_decrypt
1178 .asciz  "AES for ARMv4, CRYPTOGAMS by <appro@openssl.org>"
1179 .align  2