| blob | be7d9b4fbd13c852f53661b51956ef1ce090c4ca |
2 @ ====================================================================
3 @ Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
4 @ project. The module is, however, dual licensed under OpenSSL and
5 @ CRYPTOGAMS licenses depending on where you obtain it. For further
6 @ details see http://www.openssl.org/~appro/cryptogams/.
7 @ ====================================================================
9 @ AES for ARMv4
11 @ January 2007.
12 @
13 @ Code uses single 1K S-box and is >2 times faster than code generated
14 @ by gcc-3.4.1. This is thanks to unique feature of ARMv4 ISA, which
15 @ allows to merge logical or arithmetic operation with shift or rotate
16 @ in one instruction and emit combined result every cycle. The module
17 @ is endian-neutral. The performance is ~42 cycles/byte for 128-bit
18 @ key [on single-issue Xscale PXA250 core].
20 @ May 2007.
21 @
22 @ AES_set_[en|de]crypt_key is added.
24 @ July 2010.
25 @
26 @ Rescheduling for dual-issue pipeline resulted in 12% improvement on
27 @ Cortex A8 core and ~25 cycles per byte processed with 128-bit key.
29 @ February 2011.
30 @
31 @ Profiler-assisted and platform-specific optimization resulted in 16%
32 @ improvement on Cortex A8 core and ~21.5 cycles per byte.
34 #ifndef __KERNEL__
35 # include "arm_arch.h"
36 # include "arm_asm.h"
37 #else
38 # define __ARM_ARCH__ __LINUX_ARM_ARCH__
39 #endif
41 .text
42 #if __ARM_ARCH__<7
43 .code 32
44 #else
45 .syntax unified
46 # ifdef __thumb2__
47 .thumb
48 # else
49 .code 32
50 # endif
51 #endif
53 .type AES_Te,%object
54 .align 5
55 AES_Te:
56 .word 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
57 .word 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
58 .word 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
59 .word 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
60 .word 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
61 .word 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
62 .word 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
63 .word 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
64 .word 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
65 .word 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
66 .word 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
67 .word 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
68 .word 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
69 .word 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
70 .word 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
71 .word 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
72 .word 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
73 .word 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
74 .word 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
75 .word 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
76 .word 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
77 .word 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
78 .word 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
79 .word 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
80 .word 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
81 .word 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
82 .word 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
83 .word 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
84 .word 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
85 .word 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
86 .word 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
87 .word 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
88 .word 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
89 .word 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
90 .word 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
91 .word 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
92 .word 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
93 .word 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
94 .word 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
95 .word 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
96 .word 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
97 .word 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
98 .word 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
99 .word 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
100 .word 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
101 .word 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
102 .word 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
103 .word 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
104 .word 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
105 .word 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
106 .word 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
107 .word 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
108 .word 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
109 .word 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
110 .word 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
111 .word 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
112 .word 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
113 .word 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
114 .word 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
115 .word 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
116 .word 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
117 .word 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
118 .word 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
119 .word 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
120 @ Te4[256]
121 .byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
122 .byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
123 .byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
124 .byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
125 .byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
126 .byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
127 .byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
128 .byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
129 .byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
130 .byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
131 .byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
132 .byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
133 .byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
134 .byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
135 .byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
136 .byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
137 .byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
138 .byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
139 .byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
140 .byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
141 .byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
142 .byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
143 .byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
144 .byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
145 .byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
146 .byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
147 .byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
148 .byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
149 .byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
150 .byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
151 .byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
152 .byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
153 @ rcon[]
154 .word 0x01000000, 0x02000000, 0x04000000, 0x08000000
155 .word 0x10000000, 0x20000000, 0x40000000, 0x80000000
156 .word 0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
157 .size AES_Te,.-AES_Te
159 @ void AES_encrypt(const unsigned char *in, unsigned char *out,
160 @ const AES_KEY *key) {
161 .global AES_encrypt
162 .type AES_encrypt,%function
163 .align 5
164 AES_encrypt:
165 #if __ARM_ARCH__<7
166 sub r3,pc,#8 @ AES_encrypt
167 #else
168 adr r3,AES_encrypt
169 #endif
170 stmdb sp!,{r1,r4-r12,lr}
171 mov r12,r0 @ inp
172 mov r11,r2
173 #ifdef __clang__
174 @ workaround for clang integrated assembler bug
175 ldr r0,=(AES_encrypt-AES_Te)
176 sub r10,r3,r0
177 #else
178 sub r10,r3,#AES_encrypt-AES_Te @ Te
179 #endif
180 #if __ARM_ARCH__<7
181 ldrb r0,[r12,#3] @ load input data in endian-neutral
182 ldrb r4,[r12,#2] @ manner...
183 ldrb r5,[r12,#1]
184 ldrb r6,[r12,#0]
185 orr r0,r0,r4,lsl#8
186 ldrb r1,[r12,#7]
187 orr r0,r0,r5,lsl#16
188 ldrb r4,[r12,#6]
189 orr r0,r0,r6,lsl#24
190 ldrb r5,[r12,#5]
191 ldrb r6,[r12,#4]
192 orr r1,r1,r4,lsl#8
193 ldrb r2,[r12,#11]
194 orr r1,r1,r5,lsl#16
195 ldrb r4,[r12,#10]
196 orr r1,r1,r6,lsl#24
197 ldrb r5,[r12,#9]
198 ldrb r6,[r12,#8]
199 orr r2,r2,r4,lsl#8
200 ldrb r3,[r12,#15]
201 orr r2,r2,r5,lsl#16
202 ldrb r4,[r12,#14]
203 orr r2,r2,r6,lsl#24
204 ldrb r5,[r12,#13]
205 ldrb r6,[r12,#12]
206 orr r3,r3,r4,lsl#8
207 orr r3,r3,r5,lsl#16
208 orr r3,r3,r6,lsl#24
209 #else
210 ldr r0,[r12,#0]
211 ldr r1,[r12,#4]
212 ldr r2,[r12,#8]
213 ldr r3,[r12,#12]
214 #ifdef __ARMEL__
215 rev r0,r0
216 rev r1,r1
217 rev r2,r2
218 rev r3,r3
219 #endif
220 #endif
221 bl _armv4_AES_encrypt
223 ldr r12,[sp],#4 @ pop out
224 #if __ARM_ARCH__>=7
225 #ifdef __ARMEL__
226 rev r0,r0
227 rev r1,r1
228 rev r2,r2
229 rev r3,r3
230 #endif
231 str r0,[r12,#0]
232 str r1,[r12,#4]
233 str r2,[r12,#8]
234 str r3,[r12,#12]
235 #else
236 mov r4,r0,lsr#24 @ write output in endian-neutral
237 mov r5,r0,lsr#16 @ manner...
238 mov r6,r0,lsr#8
239 strb r4,[r12,#0]
240 strb r5,[r12,#1]
241 mov r4,r1,lsr#24
242 strb r6,[r12,#2]
243 mov r5,r1,lsr#16
244 strb r0,[r12,#3]
245 mov r6,r1,lsr#8
246 strb r4,[r12,#4]
247 strb r5,[r12,#5]
248 mov r4,r2,lsr#24
249 strb r6,[r12,#6]
250 mov r5,r2,lsr#16
251 strb r1,[r12,#7]
252 mov r6,r2,lsr#8
253 strb r4,[r12,#8]
254 strb r5,[r12,#9]
255 mov r4,r3,lsr#24
256 strb r6,[r12,#10]
257 mov r5,r3,lsr#16
258 strb r2,[r12,#11]
259 mov r6,r3,lsr#8
260 strb r4,[r12,#12]
261 strb r5,[r12,#13]
262 strb r6,[r12,#14]
263 strb r3,[r12,#15]
264 #endif
265 #if __ARM_ARCH__>=5
266 ldmia sp!,{r4-r12,pc}
267 #else
268 ldmia sp!,{r4-r12,lr}
269 tst lr,#1
270 moveq pc,lr @ be binary compatible with V4, yet
271 .word 0xe12fff1e @ interoperable with Thumb ISA:-)
272 #endif
273 .size AES_encrypt,.-AES_encrypt
275 .type _armv4_AES_encrypt,%function
276 .align 2
277 _armv4_AES_encrypt:
278 str lr,[sp,#-4]! @ push lr
279 ldmia r11!,{r4-r7}
280 eor r0,r0,r4
281 ldr r12,[r11,#240-16]
282 eor r1,r1,r5
283 eor r2,r2,r6
284 eor r3,r3,r7
285 sub r12,r12,#1
286 mov lr,#255
288 and r7,lr,r0
289 and r8,lr,r0,lsr#8
290 and r9,lr,r0,lsr#16
291 mov r0,r0,lsr#24
292 .Lenc_loop:
293 ldr r4,[r10,r7,lsl#2] @ Te3[s0>>0]
294 and r7,lr,r1,lsr#16 @ i0
295 ldr r5,[r10,r8,lsl#2] @ Te2[s0>>8]
296 and r8,lr,r1
297 ldr r6,[r10,r9,lsl#2] @ Te1[s0>>16]
298 and r9,lr,r1,lsr#8
299 ldr r0,[r10,r0,lsl#2] @ Te0[s0>>24]
300 mov r1,r1,lsr#24
302 ldr r7,[r10,r7,lsl#2] @ Te1[s1>>16]
303 ldr r8,[r10,r8,lsl#2] @ Te3[s1>>0]
304 ldr r9,[r10,r9,lsl#2] @ Te2[s1>>8]
305 eor r0,r0,r7,ror#8
306 ldr r1,[r10,r1,lsl#2] @ Te0[s1>>24]
307 and r7,lr,r2,lsr#8 @ i0
308 eor r5,r5,r8,ror#8
309 and r8,lr,r2,lsr#16 @ i1
310 eor r6,r6,r9,ror#8
311 and r9,lr,r2
312 ldr r7,[r10,r7,lsl#2] @ Te2[s2>>8]
313 eor r1,r1,r4,ror#24
314 ldr r8,[r10,r8,lsl#2] @ Te1[s2>>16]
315 mov r2,r2,lsr#24
317 ldr r9,[r10,r9,lsl#2] @ Te3[s2>>0]
318 eor r0,r0,r7,ror#16
319 ldr r2,[r10,r2,lsl#2] @ Te0[s2>>24]
320 and r7,lr,r3 @ i0
321 eor r1,r1,r8,ror#8
322 and r8,lr,r3,lsr#8 @ i1
323 eor r6,r6,r9,ror#16
324 and r9,lr,r3,lsr#16 @ i2
325 ldr r7,[r10,r7,lsl#2] @ Te3[s3>>0]
326 eor r2,r2,r5,ror#16
327 ldr r8,[r10,r8,lsl#2] @ Te2[s3>>8]
328 mov r3,r3,lsr#24
330 ldr r9,[r10,r9,lsl#2] @ Te1[s3>>16]
331 eor r0,r0,r7,ror#24
332 ldr r7,[r11],#16
333 eor r1,r1,r8,ror#16
334 ldr r3,[r10,r3,lsl#2] @ Te0[s3>>24]
335 eor r2,r2,r9,ror#8
336 ldr r4,[r11,#-12]
337 eor r3,r3,r6,ror#8
339 ldr r5,[r11,#-8]
340 eor r0,r0,r7
341 ldr r6,[r11,#-4]
342 and r7,lr,r0
343 eor r1,r1,r4
344 and r8,lr,r0,lsr#8
345 eor r2,r2,r5
346 and r9,lr,r0,lsr#16
347 eor r3,r3,r6
348 mov r0,r0,lsr#24
350 subs r12,r12,#1
351 bne .Lenc_loop
353 add r10,r10,#2
355 ldrb r4,[r10,r7,lsl#2] @ Te4[s0>>0]
356 and r7,lr,r1,lsr#16 @ i0
357 ldrb r5,[r10,r8,lsl#2] @ Te4[s0>>8]
358 and r8,lr,r1
359 ldrb r6,[r10,r9,lsl#2] @ Te4[s0>>16]
360 and r9,lr,r1,lsr#8
361 ldrb r0,[r10,r0,lsl#2] @ Te4[s0>>24]
362 mov r1,r1,lsr#24
364 ldrb r7,[r10,r7,lsl#2] @ Te4[s1>>16]
365 ldrb r8,[r10,r8,lsl#2] @ Te4[s1>>0]
366 ldrb r9,[r10,r9,lsl#2] @ Te4[s1>>8]
367 eor r0,r7,r0,lsl#8
368 ldrb r1,[r10,r1,lsl#2] @ Te4[s1>>24]
369 and r7,lr,r2,lsr#8 @ i0
370 eor r5,r8,r5,lsl#8
371 and r8,lr,r2,lsr#16 @ i1
372 eor r6,r9,r6,lsl#8
373 and r9,lr,r2
374 ldrb r7,[r10,r7,lsl#2] @ Te4[s2>>8]
375 eor r1,r4,r1,lsl#24
376 ldrb r8,[r10,r8,lsl#2] @ Te4[s2>>16]
377 mov r2,r2,lsr#24
379 ldrb r9,[r10,r9,lsl#2] @ Te4[s2>>0]
380 eor r0,r7,r0,lsl#8
381 ldrb r2,[r10,r2,lsl#2] @ Te4[s2>>24]
382 and r7,lr,r3 @ i0
383 eor r1,r1,r8,lsl#16
384 and r8,lr,r3,lsr#8 @ i1
385 eor r6,r9,r6,lsl#8
386 and r9,lr,r3,lsr#16 @ i2
387 ldrb r7,[r10,r7,lsl#2] @ Te4[s3>>0]
388 eor r2,r5,r2,lsl#24
389 ldrb r8,[r10,r8,lsl#2] @ Te4[s3>>8]
390 mov r3,r3,lsr#24
392 ldrb r9,[r10,r9,lsl#2] @ Te4[s3>>16]
393 eor r0,r7,r0,lsl#8
394 ldr r7,[r11,#0]
395 ldrb r3,[r10,r3,lsl#2] @ Te4[s3>>24]
396 eor r1,r1,r8,lsl#8
397 ldr r4,[r11,#4]
398 eor r2,r2,r9,lsl#16
399 ldr r5,[r11,#8]
400 eor r3,r6,r3,lsl#24
401 ldr r6,[r11,#12]
403 eor r0,r0,r7
404 eor r1,r1,r4
405 eor r2,r2,r5
406 eor r3,r3,r6
408 sub r10,r10,#2
409 ldr pc,[sp],#4 @ pop and return
410 .size _armv4_AES_encrypt,.-_armv4_AES_encrypt
412 .global private_AES_set_encrypt_key
413 .type private_AES_set_encrypt_key,%function
414 .align 5
415 private_AES_set_encrypt_key:
416 _armv4_AES_set_encrypt_key:
417 #if __ARM_ARCH__<7
418 sub r3,pc,#8 @ AES_set_encrypt_key
419 #else
420 adr r3,private_AES_set_encrypt_key
421 #endif
422 teq r0,#0
423 #if __ARM_ARCH__>=7
424 itt eq @ Thumb2 thing, sanity check in ARM
425 #endif
426 moveq r0,#-1
427 beq .Labrt
428 teq r2,#0
429 #if __ARM_ARCH__>=7
430 itt eq @ Thumb2 thing, sanity check in ARM
431 #endif
432 moveq r0,#-1
433 beq .Labrt
435 teq r1,#128
436 beq .Lok
437 teq r1,#192
438 beq .Lok
439 teq r1,#256
440 #if __ARM_ARCH__>=7
441 itt ne @ Thumb2 thing, sanity check in ARM
442 #endif
443 movne r0,#-1
444 bne .Labrt
446 .Lok: stmdb sp!,{r4-r12,lr}
447 sub r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024 @ Te4
449 mov r12,r0 @ inp
450 mov lr,r1 @ bits
451 mov r11,r2 @ key
453 #if __ARM_ARCH__<7
454 ldrb r0,[r12,#3] @ load input data in endian-neutral
455 ldrb r4,[r12,#2] @ manner...
456 ldrb r5,[r12,#1]
457 ldrb r6,[r12,#0]
458 orr r0,r0,r4,lsl#8
459 ldrb r1,[r12,#7]
460 orr r0,r0,r5,lsl#16
461 ldrb r4,[r12,#6]
462 orr r0,r0,r6,lsl#24
463 ldrb r5,[r12,#5]
464 ldrb r6,[r12,#4]
465 orr r1,r1,r4,lsl#8
466 ldrb r2,[r12,#11]
467 orr r1,r1,r5,lsl#16
468 ldrb r4,[r12,#10]
469 orr r1,r1,r6,lsl#24
470 ldrb r5,[r12,#9]
471 ldrb r6,[r12,#8]
472 orr r2,r2,r4,lsl#8
473 ldrb r3,[r12,#15]
474 orr r2,r2,r5,lsl#16
475 ldrb r4,[r12,#14]
476 orr r2,r2,r6,lsl#24
477 ldrb r5,[r12,#13]
478 ldrb r6,[r12,#12]
479 orr r3,r3,r4,lsl#8
480 str r0,[r11],#16
481 orr r3,r3,r5,lsl#16
482 str r1,[r11,#-12]
483 orr r3,r3,r6,lsl#24
484 str r2,[r11,#-8]
485 str r3,[r11,#-4]
486 #else
487 ldr r0,[r12,#0]
488 ldr r1,[r12,#4]
489 ldr r2,[r12,#8]
490 ldr r3,[r12,#12]
491 #ifdef __ARMEL__
492 rev r0,r0
493 rev r1,r1
494 rev r2,r2
495 rev r3,r3
496 #endif
497 str r0,[r11],#16
498 str r1,[r11,#-12]
499 str r2,[r11,#-8]
500 str r3,[r11,#-4]
501 #endif
503 teq lr,#128
504 bne .Lnot128
505 mov r12,#10
506 str r12,[r11,#240-16]
507 add r6,r10,#256 @ rcon
508 mov lr,#255
510 .L128_loop:
511 and r5,lr,r3,lsr#24
512 and r7,lr,r3,lsr#16
513 ldrb r5,[r10,r5]
514 and r8,lr,r3,lsr#8
515 ldrb r7,[r10,r7]
516 and r9,lr,r3
517 ldrb r8,[r10,r8]
518 orr r5,r5,r7,lsl#24
519 ldrb r9,[r10,r9]
520 orr r5,r5,r8,lsl#16
521 ldr r4,[r6],#4 @ rcon[i++]
522 orr r5,r5,r9,lsl#8
523 eor r5,r5,r4
524 eor r0,r0,r5 @ rk[4]=rk[0]^...
525 eor r1,r1,r0 @ rk[5]=rk[1]^rk[4]
526 str r0,[r11],#16
527 eor r2,r2,r1 @ rk[6]=rk[2]^rk[5]
528 str r1,[r11,#-12]
529 eor r3,r3,r2 @ rk[7]=rk[3]^rk[6]
530 str r2,[r11,#-8]
531 subs r12,r12,#1
532 str r3,[r11,#-4]
533 bne .L128_loop
534 sub r2,r11,#176
535 b .Ldone
537 .Lnot128:
538 #if __ARM_ARCH__<7
539 ldrb r8,[r12,#19]
540 ldrb r4,[r12,#18]
541 ldrb r5,[r12,#17]
542 ldrb r6,[r12,#16]
543 orr r8,r8,r4,lsl#8
544 ldrb r9,[r12,#23]
545 orr r8,r8,r5,lsl#16
546 ldrb r4,[r12,#22]
547 orr r8,r8,r6,lsl#24
548 ldrb r5,[r12,#21]
549 ldrb r6,[r12,#20]
550 orr r9,r9,r4,lsl#8
551 orr r9,r9,r5,lsl#16
552 str r8,[r11],#8
553 orr r9,r9,r6,lsl#24
554 str r9,[r11,#-4]
555 #else
556 ldr r8,[r12,#16]
557 ldr r9,[r12,#20]
558 #ifdef __ARMEL__
559 rev r8,r8
560 rev r9,r9
561 #endif
562 str r8,[r11],#8
563 str r9,[r11,#-4]
564 #endif
566 teq lr,#192
567 bne .Lnot192
568 mov r12,#12
569 str r12,[r11,#240-24]
570 add r6,r10,#256 @ rcon
571 mov lr,#255
572 mov r12,#8
574 .L192_loop:
575 and r5,lr,r9,lsr#24
576 and r7,lr,r9,lsr#16
577 ldrb r5,[r10,r5]
578 and r8,lr,r9,lsr#8
579 ldrb r7,[r10,r7]
580 and r9,lr,r9
581 ldrb r8,[r10,r8]
582 orr r5,r5,r7,lsl#24
583 ldrb r9,[r10,r9]
584 orr r5,r5,r8,lsl#16
585 ldr r4,[r6],#4 @ rcon[i++]
586 orr r5,r5,r9,lsl#8
587 eor r9,r5,r4
588 eor r0,r0,r9 @ rk[6]=rk[0]^...
589 eor r1,r1,r0 @ rk[7]=rk[1]^rk[6]
590 str r0,[r11],#24
591 eor r2,r2,r1 @ rk[8]=rk[2]^rk[7]
592 str r1,[r11,#-20]
593 eor r3,r3,r2 @ rk[9]=rk[3]^rk[8]
594 str r2,[r11,#-16]
595 subs r12,r12,#1
596 str r3,[r11,#-12]
597 #if __ARM_ARCH__>=7
598 itt eq @ Thumb2 thing, sanity check in ARM
599 #endif
600 subeq r2,r11,#216
601 beq .Ldone
603 ldr r7,[r11,#-32]
604 ldr r8,[r11,#-28]
605 eor r7,r7,r3 @ rk[10]=rk[4]^rk[9]
606 eor r9,r8,r7 @ rk[11]=rk[5]^rk[10]
607 str r7,[r11,#-8]
608 str r9,[r11,#-4]
609 b .L192_loop
611 .Lnot192:
612 #if __ARM_ARCH__<7
613 ldrb r8,[r12,#27]
614 ldrb r4,[r12,#26]
615 ldrb r5,[r12,#25]
616 ldrb r6,[r12,#24]
617 orr r8,r8,r4,lsl#8
618 ldrb r9,[r12,#31]
619 orr r8,r8,r5,lsl#16
620 ldrb r4,[r12,#30]
621 orr r8,r8,r6,lsl#24
622 ldrb r5,[r12,#29]
623 ldrb r6,[r12,#28]
624 orr r9,r9,r4,lsl#8
625 orr r9,r9,r5,lsl#16
626 str r8,[r11],#8
627 orr r9,r9,r6,lsl#24
628 str r9,[r11,#-4]
629 #else
630 ldr r8,[r12,#24]
631 ldr r9,[r12,#28]
632 #ifdef __ARMEL__
633 rev r8,r8
634 rev r9,r9
635 #endif
636 str r8,[r11],#8
637 str r9,[r11,#-4]
638 #endif
640 mov r12,#14
641 str r12,[r11,#240-32]
642 add r6,r10,#256 @ rcon
643 mov lr,#255
644 mov r12,#7
646 .L256_loop:
647 and r5,lr,r9,lsr#24
648 and r7,lr,r9,lsr#16
649 ldrb r5,[r10,r5]
650 and r8,lr,r9,lsr#8
651 ldrb r7,[r10,r7]
652 and r9,lr,r9
653 ldrb r8,[r10,r8]
654 orr r5,r5,r7,lsl#24
655 ldrb r9,[r10,r9]
656 orr r5,r5,r8,lsl#16
657 ldr r4,[r6],#4 @ rcon[i++]
658 orr r5,r5,r9,lsl#8
659 eor r9,r5,r4
660 eor r0,r0,r9 @ rk[8]=rk[0]^...
661 eor r1,r1,r0 @ rk[9]=rk[1]^rk[8]
662 str r0,[r11],#32
663 eor r2,r2,r1 @ rk[10]=rk[2]^rk[9]
664 str r1,[r11,#-28]
665 eor r3,r3,r2 @ rk[11]=rk[3]^rk[10]
666 str r2,[r11,#-24]
667 subs r12,r12,#1
668 str r3,[r11,#-20]
669 #if __ARM_ARCH__>=7
670 itt eq @ Thumb2 thing, sanity check in ARM
671 #endif
672 subeq r2,r11,#256
673 beq .Ldone
675 and r5,lr,r3
676 and r7,lr,r3,lsr#8
677 ldrb r5,[r10,r5]
678 and r8,lr,r3,lsr#16
679 ldrb r7,[r10,r7]
680 and r9,lr,r3,lsr#24
681 ldrb r8,[r10,r8]
682 orr r5,r5,r7,lsl#8
683 ldrb r9,[r10,r9]
684 orr r5,r5,r8,lsl#16
685 ldr r4,[r11,#-48]
686 orr r5,r5,r9,lsl#24
688 ldr r7,[r11,#-44]
689 ldr r8,[r11,#-40]
690 eor r4,r4,r5 @ rk[12]=rk[4]^...
691 ldr r9,[r11,#-36]
692 eor r7,r7,r4 @ rk[13]=rk[5]^rk[12]
693 str r4,[r11,#-16]
694 eor r8,r8,r7 @ rk[14]=rk[6]^rk[13]
695 str r7,[r11,#-12]
696 eor r9,r9,r8 @ rk[15]=rk[7]^rk[14]
697 str r8,[r11,#-8]
698 str r9,[r11,#-4]
699 b .L256_loop
701 .align 2
702 .Ldone: mov r0,#0
703 ldmia sp!,{r4-r12,lr}
704 .Labrt:
705 #if __ARM_ARCH__>=5
706 RET @ .word 0xe12fff1e
707 #else
708 tst lr,#1
709 moveq pc,lr @ be binary compatible with V4, yet
710 .word 0xe12fff1e @ interoperable with Thumb ISA:-)
711 #endif
712 .size private_AES_set_encrypt_key,.-private_AES_set_encrypt_key
714 .global private_AES_set_decrypt_key
715 .type private_AES_set_decrypt_key,%function
716 .align 5
717 private_AES_set_decrypt_key:
718 str lr,[sp,#-4]! @ push lr
719 bl _armv4_AES_set_encrypt_key
720 teq r0,#0
721 ldr lr,[sp],#4 @ pop lr
722 bne .Labrt
724 mov r0,r2 @ AES_set_encrypt_key preserves r2,
725 mov r1,r2 @ which is AES_KEY *key
726 b _armv4_AES_set_enc2dec_key
727 .size private_AES_set_decrypt_key,.-private_AES_set_decrypt_key
729 @ void AES_set_enc2dec_key(const AES_KEY *inp,AES_KEY *out)
730 .global AES_set_enc2dec_key
731 .type AES_set_enc2dec_key,%function
732 .align 5
733 AES_set_enc2dec_key:
734 _armv4_AES_set_enc2dec_key:
735 stmdb sp!,{r4-r12,lr}
737 ldr r12,[r0,#240]
738 mov r7,r0 @ input
739 add r8,r0,r12,lsl#4
740 mov r11,r1 @ ouput
741 add r10,r1,r12,lsl#4
742 str r12,[r1,#240]
744 .Linv: ldr r0,[r7],#16
745 ldr r1,[r7,#-12]
746 ldr r2,[r7,#-8]
747 ldr r3,[r7,#-4]
748 ldr r4,[r8],#-16
749 ldr r5,[r8,#16+4]
750 ldr r6,[r8,#16+8]
751 ldr r9,[r8,#16+12]
752 str r0,[r10],#-16
753 str r1,[r10,#16+4]
754 str r2,[r10,#16+8]
755 str r3,[r10,#16+12]
756 str r4,[r11],#16
757 str r5,[r11,#-12]
758 str r6,[r11,#-8]
759 str r9,[r11,#-4]
760 teq r7,r8
761 bne .Linv
763 ldr r0,[r7]
764 ldr r1,[r7,#4]
765 ldr r2,[r7,#8]
766 ldr r3,[r7,#12]
767 str r0,[r11]
768 str r1,[r11,#4]
769 str r2,[r11,#8]
770 str r3,[r11,#12]
771 sub r11,r11,r12,lsl#3
772 ldr r0,[r11,#16]! @ prefetch tp1
773 mov r7,#0x80
774 mov r8,#0x1b
775 orr r7,r7,#0x8000
776 orr r8,r8,#0x1b00
777 orr r7,r7,r7,lsl#16
778 orr r8,r8,r8,lsl#16
779 sub r12,r12,#1
780 mvn r9,r7
781 mov r12,r12,lsl#2 @ (rounds-1)*4
783 .Lmix: and r4,r0,r7
784 and r1,r0,r9
785 sub r4,r4,r4,lsr#7
786 and r4,r4,r8
787 eor r1,r4,r1,lsl#1 @ tp2
789 and r4,r1,r7
790 and r2,r1,r9
791 sub r4,r4,r4,lsr#7
792 and r4,r4,r8
793 eor r2,r4,r2,lsl#1 @ tp4
795 and r4,r2,r7
796 and r3,r2,r9
797 sub r4,r4,r4,lsr#7
798 and r4,r4,r8
799 eor r3,r4,r3,lsl#1 @ tp8
801 eor r4,r1,r2
802 eor r5,r0,r3 @ tp9
803 eor r4,r4,r3 @ tpe
804 eor r4,r4,r1,ror#24
805 eor r4,r4,r5,ror#24 @ ^= ROTATE(tpb=tp9^tp2,8)
806 eor r4,r4,r2,ror#16
807 eor r4,r4,r5,ror#16 @ ^= ROTATE(tpd=tp9^tp4,16)
808 eor r4,r4,r5,ror#8 @ ^= ROTATE(tp9,24)
810 ldr r0,[r11,#4] @ prefetch tp1
811 str r4,[r11],#4
812 subs r12,r12,#1
813 bne .Lmix
815 mov r0,#0
816 #if __ARM_ARCH__>=5
817 ldmia sp!,{r4-r12,pc}
818 #else
819 ldmia sp!,{r4-r12,lr}
820 tst lr,#1
821 moveq pc,lr @ be binary compatible with V4, yet
822 .word 0xe12fff1e @ interoperable with Thumb ISA:-)
823 #endif
824 .size AES_set_enc2dec_key,.-AES_set_enc2dec_key
826 .type AES_Td,%object
827 .align 5
828 AES_Td:
829 .word 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
830 .word 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
831 .word 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
832 .word 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
833 .word 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
834 .word 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
835 .word 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
836 .word 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
837 .word 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
838 .word 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
839 .word 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
840 .word 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
841 .word 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
842 .word 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
843 .word 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
844 .word 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
845 .word 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
846 .word 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
847 .word 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
848 .word 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
849 .word 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
850 .word 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
851 .word 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
852 .word 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
853 .word 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
854 .word 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
855 .word 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
856 .word 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
857 .word 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
858 .word 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
859 .word 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
860 .word 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
861 .word 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
862 .word 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
863 .word 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
864 .word 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
865 .word 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
866 .word 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
867 .word 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
868 .word 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
869 .word 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
870 .word 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
871 .word 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
872 .word 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
873 .word 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
874 .word 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
875 .word 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
876 .word 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
877 .word 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
878 .word 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
879 .word 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
880 .word 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
881 .word 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
882 .word 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
883 .word 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
884 .word 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
885 .word 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
886 .word 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
887 .word 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
888 .word 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
889 .word 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
890 .word 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
891 .word 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
892 .word 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
893 @ Td4[256]
894 .byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
895 .byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
896 .byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
897 .byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
898 .byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
899 .byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
900 .byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
901 .byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
902 .byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
903 .byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
904 .byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
905 .byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
906 .byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
907 .byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
908 .byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
909 .byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
910 .byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
911 .byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
912 .byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
913 .byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
914 .byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
915 .byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
916 .byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
917 .byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
918 .byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
919 .byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
920 .byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
921 .byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
922 .byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
923 .byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
924 .byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
925 .byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
926 .size AES_Td,.-AES_Td
928 @ void AES_decrypt(const unsigned char *in, unsigned char *out,
929 @ const AES_KEY *key) {
930 .global AES_decrypt
931 .type AES_decrypt,%function
932 .align 5
933 AES_decrypt:
934 #if __ARM_ARCH__<7
935 sub r3,pc,#8 @ AES_decrypt
936 #else
937 adr r3,AES_decrypt
938 #endif
939 stmdb sp!,{r1,r4-r12,lr}
940 mov r12,r0 @ inp
941 mov r11,r2
942 sub r10,r3,#AES_decrypt-AES_Td @ Td
943 #if __ARM_ARCH__<7
944 ldrb r0,[r12,#3] @ load input data in endian-neutral
945 ldrb r4,[r12,#2] @ manner...
946 ldrb r5,[r12,#1]
947 ldrb r6,[r12,#0]
948 orr r0,r0,r4,lsl#8
949 ldrb r1,[r12,#7]
950 orr r0,r0,r5,lsl#16
951 ldrb r4,[r12,#6]
952 orr r0,r0,r6,lsl#24
953 ldrb r5,[r12,#5]
954 ldrb r6,[r12,#4]
955 orr r1,r1,r4,lsl#8
956 ldrb r2,[r12,#11]
957 orr r1,r1,r5,lsl#16
958 ldrb r4,[r12,#10]
959 orr r1,r1,r6,lsl#24
960 ldrb r5,[r12,#9]
961 ldrb r6,[r12,#8]
962 orr r2,r2,r4,lsl#8
963 ldrb r3,[r12,#15]
964 orr r2,r2,r5,lsl#16
965 ldrb r4,[r12,#14]
966 orr r2,r2,r6,lsl#24
967 ldrb r5,[r12,#13]
968 ldrb r6,[r12,#12]
969 orr r3,r3,r4,lsl#8
970 orr r3,r3,r5,lsl#16
971 orr r3,r3,r6,lsl#24
972 #else
973 ldr r0,[r12,#0]
974 ldr r1,[r12,#4]
975 ldr r2,[r12,#8]
976 ldr r3,[r12,#12]
977 #ifdef __ARMEL__
978 rev r0,r0
979 rev r1,r1
980 rev r2,r2
981 rev r3,r3
982 #endif
983 #endif
984 bl _armv4_AES_decrypt
986 ldr r12,[sp],#4 @ pop out
987 #if __ARM_ARCH__>=7
988 #ifdef __ARMEL__
989 rev r0,r0
990 rev r1,r1
991 rev r2,r2
992 rev r3,r3
993 #endif
994 str r0,[r12,#0]
995 str r1,[r12,#4]
996 str r2,[r12,#8]
997 str r3,[r12,#12]
998 #else
999 mov r4,r0,lsr#24 @ write output in endian-neutral
1000 mov r5,r0,lsr#16 @ manner...
1001 mov r6,r0,lsr#8
1002 strb r4,[r12,#0]
1003 strb r5,[r12,#1]
1004 mov r4,r1,lsr#24
1005 strb r6,[r12,#2]
1006 mov r5,r1,lsr#16
1007 strb r0,[r12,#3]
1008 mov r6,r1,lsr#8
1009 strb r4,[r12,#4]
1010 strb r5,[r12,#5]
1011 mov r4,r2,lsr#24
1012 strb r6,[r12,#6]
1013 mov r5,r2,lsr#16
1014 strb r1,[r12,#7]
1015 mov r6,r2,lsr#8
1016 strb r4,[r12,#8]
1017 strb r5,[r12,#9]
1018 mov r4,r3,lsr#24
1019 strb r6,[r12,#10]
1020 mov r5,r3,lsr#16
1021 strb r2,[r12,#11]
1022 mov r6,r3,lsr#8
1023 strb r4,[r12,#12]
1024 strb r5,[r12,#13]
1025 strb r6,[r12,#14]
1026 strb r3,[r12,#15]
1027 #endif
1028 #if __ARM_ARCH__>=5
1029 ldmia sp!,{r4-r12,pc}
1030 #else
1031 ldmia sp!,{r4-r12,lr}
1032 tst lr,#1
1033 moveq pc,lr @ be binary compatible with V4, yet
1034 .word 0xe12fff1e @ interoperable with Thumb ISA:-)
1035 #endif
1036 .size AES_decrypt,.-AES_decrypt
1038 .type _armv4_AES_decrypt,%function
1039 .align 2
1040 _armv4_AES_decrypt:
1041 str lr,[sp,#-4]! @ push lr
1042 ldmia r11!,{r4-r7}
1043 eor r0,r0,r4
1044 ldr r12,[r11,#240-16]
1045 eor r1,r1,r5
1046 eor r2,r2,r6
1047 eor r3,r3,r7
1048 sub r12,r12,#1
1049 mov lr,#255
1051 and r7,lr,r0,lsr#16
1052 and r8,lr,r0,lsr#8
1053 and r9,lr,r0
1054 mov r0,r0,lsr#24
1055 .Ldec_loop:
1056 ldr r4,[r10,r7,lsl#2] @ Td1[s0>>16]
1057 and r7,lr,r1 @ i0
1058 ldr r5,[r10,r8,lsl#2] @ Td2[s0>>8]
1059 and r8,lr,r1,lsr#16
1060 ldr r6,[r10,r9,lsl#2] @ Td3[s0>>0]
1061 and r9,lr,r1,lsr#8
1062 ldr r0,[r10,r0,lsl#2] @ Td0[s0>>24]
1063 mov r1,r1,lsr#24
1065 ldr r7,[r10,r7,lsl#2] @ Td3[s1>>0]
1066 ldr r8,[r10,r8,lsl#2] @ Td1[s1>>16]
1067 ldr r9,[r10,r9,lsl#2] @ Td2[s1>>8]
1068 eor r0,r0,r7,ror#24
1069 ldr r1,[r10,r1,lsl#2] @ Td0[s1>>24]
1070 and r7,lr,r2,lsr#8 @ i0
1071 eor r5,r8,r5,ror#8
1072 and r8,lr,r2 @ i1
1073 eor r6,r9,r6,ror#8
1074 and r9,lr,r2,lsr#16
1075 ldr r7,[r10,r7,lsl#2] @ Td2[s2>>8]
1076 eor r1,r1,r4,ror#8
1077 ldr r8,[r10,r8,lsl#2] @ Td3[s2>>0]
1078 mov r2,r2,lsr#24
1080 ldr r9,[r10,r9,lsl#2] @ Td1[s2>>16]
1081 eor r0,r0,r7,ror#16
1082 ldr r2,[r10,r2,lsl#2] @ Td0[s2>>24]
1083 and r7,lr,r3,lsr#16 @ i0
1084 eor r1,r1,r8,ror#24
1085 and r8,lr,r3,lsr#8 @ i1
1086 eor r6,r9,r6,ror#8
1087 and r9,lr,r3 @ i2
1088 ldr r7,[r10,r7,lsl#2] @ Td1[s3>>16]
1089 eor r2,r2,r5,ror#8
1090 ldr r8,[r10,r8,lsl#2] @ Td2[s3>>8]
1091 mov r3,r3,lsr#24
1093 ldr r9,[r10,r9,lsl#2] @ Td3[s3>>0]
1094 eor r0,r0,r7,ror#8
1095 ldr r7,[r11],#16
1096 eor r1,r1,r8,ror#16
1097 ldr r3,[r10,r3,lsl#2] @ Td0[s3>>24]
1098 eor r2,r2,r9,ror#24
1100 ldr r4,[r11,#-12]
1101 eor r0,r0,r7
1102 ldr r5,[r11,#-8]
1103 eor r3,r3,r6,ror#8
1104 ldr r6,[r11,#-4]
1105 and r7,lr,r0,lsr#16
1106 eor r1,r1,r4
1107 and r8,lr,r0,lsr#8
1108 eor r2,r2,r5
1109 and r9,lr,r0
1110 eor r3,r3,r6
1111 mov r0,r0,lsr#24
1113 subs r12,r12,#1
1114 bne .Ldec_loop
1116 add r10,r10,#1024
1118 ldr r5,[r10,#0] @ prefetch Td4
1119 ldr r6,[r10,#32]
1120 ldr r4,[r10,#64]
1121 ldr r5,[r10,#96]
1122 ldr r6,[r10,#128]
1123 ldr r4,[r10,#160]
1124 ldr r5,[r10,#192]
1125 ldr r6,[r10,#224]
1127 ldrb r0,[r10,r0] @ Td4[s0>>24]
1128 ldrb r4,[r10,r7] @ Td4[s0>>16]
1129 and r7,lr,r1 @ i0
1130 ldrb r5,[r10,r8] @ Td4[s0>>8]
1131 and r8,lr,r1,lsr#16
1132 ldrb r6,[r10,r9] @ Td4[s0>>0]
1133 and r9,lr,r1,lsr#8
1135 add r1,r10,r1,lsr#24
1136 ldrb r7,[r10,r7] @ Td4[s1>>0]
1137 ldrb r1,[r1] @ Td4[s1>>24]
1138 ldrb r8,[r10,r8] @ Td4[s1>>16]
1139 eor r0,r7,r0,lsl#24
1140 ldrb r9,[r10,r9] @ Td4[s1>>8]
1141 eor r1,r4,r1,lsl#8
1142 and r7,lr,r2,lsr#8 @ i0
1143 eor r5,r5,r8,lsl#8
1144 and r8,lr,r2 @ i1
1145 ldrb r7,[r10,r7] @ Td4[s2>>8]
1146 eor r6,r6,r9,lsl#8
1147 ldrb r8,[r10,r8] @ Td4[s2>>0]
1148 and r9,lr,r2,lsr#16
1150 add r2,r10,r2,lsr#24
1151 ldrb r2,[r2] @ Td4[s2>>24]
1152 eor r0,r0,r7,lsl#8
1153 ldrb r9,[r10,r9] @ Td4[s2>>16]
1154 eor r1,r8,r1,lsl#16
1155 and r7,lr,r3,lsr#16 @ i0
1156 eor r2,r5,r2,lsl#16
1157 and r8,lr,r3,lsr#8 @ i1
1158 ldrb r7,[r10,r7] @ Td4[s3>>16]
1159 eor r6,r6,r9,lsl#16
1160 ldrb r8,[r10,r8] @ Td4[s3>>8]
1161 and r9,lr,r3 @ i2
1163 add r3,r10,r3,lsr#24
1164 ldrb r9,[r10,r9] @ Td4[s3>>0]
1165 ldrb r3,[r3] @ Td4[s3>>24]
1166 eor r0,r0,r7,lsl#16
1167 ldr r7,[r11,#0]
1168 eor r1,r1,r8,lsl#8
1169 ldr r4,[r11,#4]
1170 eor r2,r9,r2,lsl#8
1171 ldr r5,[r11,#8]
1172 eor r3,r6,r3,lsl#24
1173 ldr r6,[r11,#12]
1175 eor r0,r0,r7
1176 eor r1,r1,r4
1177 eor r2,r2,r5
1178 eor r3,r3,r6
1180 sub r10,r10,#1024
1181 ldr pc,[sp],#4 @ pop and return
1182 .size _armv4_AES_decrypt,.-_armv4_AES_decrypt
1183 .asciz "AES for ARMv4, CRYPTOGAMS by <appro@openssl.org>"
1184 .align 2