blob: 3a14ea8fe97e5cac183ad8c9148b816e473b4ae5 [file] [log] [blame]
David McCulloughf0be44f2012-09-07 04:17:02 +08001#define __ARM_ARCH__ __LINUX_ARM_ARCH__
2@ ====================================================================
3@ Written by Andy Polyakov <appro@fy.chalmers.se> 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@ ====================================================================
8
9@ AES for ARMv4
10
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].
19
20@ May 2007.
21@
22@ AES_set_[en|de]crypt_key is added.
23
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.
28
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.
33
34@ A little glue here to select the correct code below for the ARM CPU
35@ that is being targetted.
36
Dave Martin638591c2013-01-10 12:20:15 +010037#include <linux/linkage.h>
38
David McCulloughf0be44f2012-09-07 04:17:02 +080039.text
David McCulloughf0be44f2012-09-07 04:17:02 +080040
41.type AES_Te,%object
42.align 5
43AES_Te:
44.word 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
45.word 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
46.word 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
47.word 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
48.word 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
49.word 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
50.word 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
51.word 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
52.word 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
53.word 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
54.word 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
55.word 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
56.word 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
57.word 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
58.word 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
59.word 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
60.word 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
61.word 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
62.word 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
63.word 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
64.word 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
65.word 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
66.word 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
67.word 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
68.word 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
69.word 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
70.word 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
71.word 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
72.word 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
73.word 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
74.word 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
75.word 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
76.word 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
77.word 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
78.word 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
79.word 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
80.word 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
81.word 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
82.word 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
83.word 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
84.word 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
85.word 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
86.word 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
87.word 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
88.word 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
89.word 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
90.word 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
91.word 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
92.word 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
93.word 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
94.word 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
95.word 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
96.word 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
97.word 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
98.word 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
99.word 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
100.word 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
101.word 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
102.word 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
103.word 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
104.word 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
105.word 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
106.word 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
107.word 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
108@ Te4[256]
109.byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
110.byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
111.byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
112.byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
113.byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
114.byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
115.byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
116.byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
117.byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
118.byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
119.byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
120.byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
121.byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
122.byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
123.byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
124.byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
125.byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
126.byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
127.byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
128.byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
129.byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
130.byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
131.byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
132.byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
133.byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
134.byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
135.byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
136.byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
137.byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
138.byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
139.byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
140.byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
141@ rcon[]
142.word 0x01000000, 0x02000000, 0x04000000, 0x08000000
143.word 0x10000000, 0x20000000, 0x40000000, 0x80000000
144.word 0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
145.size AES_Te,.-AES_Te
146
147@ void AES_encrypt(const unsigned char *in, unsigned char *out,
148@ const AES_KEY *key) {
David McCulloughf0be44f2012-09-07 04:17:02 +0800149.align 5
Dave Martin638591c2013-01-10 12:20:15 +0100150ENTRY(AES_encrypt)
Ard Biesheuvel40190c82013-09-21 11:23:50 +0100151 adr r3,AES_encrypt
David McCulloughf0be44f2012-09-07 04:17:02 +0800152 stmdb sp!,{r1,r4-r12,lr}
153 mov r12,r0 @ inp
154 mov r11,r2
155 sub r10,r3,#AES_encrypt-AES_Te @ Te
156#if __ARM_ARCH__<7
157 ldrb r0,[r12,#3] @ load input data in endian-neutral
158 ldrb r4,[r12,#2] @ manner...
159 ldrb r5,[r12,#1]
160 ldrb r6,[r12,#0]
161 orr r0,r0,r4,lsl#8
162 ldrb r1,[r12,#7]
163 orr r0,r0,r5,lsl#16
164 ldrb r4,[r12,#6]
165 orr r0,r0,r6,lsl#24
166 ldrb r5,[r12,#5]
167 ldrb r6,[r12,#4]
168 orr r1,r1,r4,lsl#8
169 ldrb r2,[r12,#11]
170 orr r1,r1,r5,lsl#16
171 ldrb r4,[r12,#10]
172 orr r1,r1,r6,lsl#24
173 ldrb r5,[r12,#9]
174 ldrb r6,[r12,#8]
175 orr r2,r2,r4,lsl#8
176 ldrb r3,[r12,#15]
177 orr r2,r2,r5,lsl#16
178 ldrb r4,[r12,#14]
179 orr r2,r2,r6,lsl#24
180 ldrb r5,[r12,#13]
181 ldrb r6,[r12,#12]
182 orr r3,r3,r4,lsl#8
183 orr r3,r3,r5,lsl#16
184 orr r3,r3,r6,lsl#24
185#else
186 ldr r0,[r12,#0]
187 ldr r1,[r12,#4]
188 ldr r2,[r12,#8]
189 ldr r3,[r12,#12]
190#ifdef __ARMEL__
191 rev r0,r0
192 rev r1,r1
193 rev r2,r2
194 rev r3,r3
195#endif
196#endif
197 bl _armv4_AES_encrypt
198
199 ldr r12,[sp],#4 @ pop out
200#if __ARM_ARCH__>=7
201#ifdef __ARMEL__
202 rev r0,r0
203 rev r1,r1
204 rev r2,r2
205 rev r3,r3
206#endif
207 str r0,[r12,#0]
208 str r1,[r12,#4]
209 str r2,[r12,#8]
210 str r3,[r12,#12]
211#else
212 mov r4,r0,lsr#24 @ write output in endian-neutral
213 mov r5,r0,lsr#16 @ manner...
214 mov r6,r0,lsr#8
215 strb r4,[r12,#0]
216 strb r5,[r12,#1]
217 mov r4,r1,lsr#24
218 strb r6,[r12,#2]
219 mov r5,r1,lsr#16
220 strb r0,[r12,#3]
221 mov r6,r1,lsr#8
222 strb r4,[r12,#4]
223 strb r5,[r12,#5]
224 mov r4,r2,lsr#24
225 strb r6,[r12,#6]
226 mov r5,r2,lsr#16
227 strb r1,[r12,#7]
228 mov r6,r2,lsr#8
229 strb r4,[r12,#8]
230 strb r5,[r12,#9]
231 mov r4,r3,lsr#24
232 strb r6,[r12,#10]
233 mov r5,r3,lsr#16
234 strb r2,[r12,#11]
235 mov r6,r3,lsr#8
236 strb r4,[r12,#12]
237 strb r5,[r12,#13]
238 strb r6,[r12,#14]
239 strb r3,[r12,#15]
240#endif
David McCulloughf0be44f2012-09-07 04:17:02 +0800241 ldmia sp!,{r4-r12,pc}
Dave Martin638591c2013-01-10 12:20:15 +0100242ENDPROC(AES_encrypt)
David McCulloughf0be44f2012-09-07 04:17:02 +0800243
244.type _armv4_AES_encrypt,%function
245.align 2
246_armv4_AES_encrypt:
247 str lr,[sp,#-4]! @ push lr
248 ldmia r11!,{r4-r7}
249 eor r0,r0,r4
250 ldr r12,[r11,#240-16]
251 eor r1,r1,r5
252 eor r2,r2,r6
253 eor r3,r3,r7
254 sub r12,r12,#1
255 mov lr,#255
256
257 and r7,lr,r0
258 and r8,lr,r0,lsr#8
259 and r9,lr,r0,lsr#16
260 mov r0,r0,lsr#24
261.Lenc_loop:
262 ldr r4,[r10,r7,lsl#2] @ Te3[s0>>0]
263 and r7,lr,r1,lsr#16 @ i0
264 ldr r5,[r10,r8,lsl#2] @ Te2[s0>>8]
265 and r8,lr,r1
266 ldr r6,[r10,r9,lsl#2] @ Te1[s0>>16]
267 and r9,lr,r1,lsr#8
268 ldr r0,[r10,r0,lsl#2] @ Te0[s0>>24]
269 mov r1,r1,lsr#24
270
271 ldr r7,[r10,r7,lsl#2] @ Te1[s1>>16]
272 ldr r8,[r10,r8,lsl#2] @ Te3[s1>>0]
273 ldr r9,[r10,r9,lsl#2] @ Te2[s1>>8]
274 eor r0,r0,r7,ror#8
275 ldr r1,[r10,r1,lsl#2] @ Te0[s1>>24]
276 and r7,lr,r2,lsr#8 @ i0
277 eor r5,r5,r8,ror#8
278 and r8,lr,r2,lsr#16 @ i1
279 eor r6,r6,r9,ror#8
280 and r9,lr,r2
281 ldr r7,[r10,r7,lsl#2] @ Te2[s2>>8]
282 eor r1,r1,r4,ror#24
283 ldr r8,[r10,r8,lsl#2] @ Te1[s2>>16]
284 mov r2,r2,lsr#24
285
286 ldr r9,[r10,r9,lsl#2] @ Te3[s2>>0]
287 eor r0,r0,r7,ror#16
288 ldr r2,[r10,r2,lsl#2] @ Te0[s2>>24]
289 and r7,lr,r3 @ i0
290 eor r1,r1,r8,ror#8
291 and r8,lr,r3,lsr#8 @ i1
292 eor r6,r6,r9,ror#16
293 and r9,lr,r3,lsr#16 @ i2
294 ldr r7,[r10,r7,lsl#2] @ Te3[s3>>0]
295 eor r2,r2,r5,ror#16
296 ldr r8,[r10,r8,lsl#2] @ Te2[s3>>8]
297 mov r3,r3,lsr#24
298
299 ldr r9,[r10,r9,lsl#2] @ Te1[s3>>16]
300 eor r0,r0,r7,ror#24
301 ldr r7,[r11],#16
302 eor r1,r1,r8,ror#16
303 ldr r3,[r10,r3,lsl#2] @ Te0[s3>>24]
304 eor r2,r2,r9,ror#8
305 ldr r4,[r11,#-12]
306 eor r3,r3,r6,ror#8
307
308 ldr r5,[r11,#-8]
309 eor r0,r0,r7
310 ldr r6,[r11,#-4]
311 and r7,lr,r0
312 eor r1,r1,r4
313 and r8,lr,r0,lsr#8
314 eor r2,r2,r5
315 and r9,lr,r0,lsr#16
316 eor r3,r3,r6
317 mov r0,r0,lsr#24
318
319 subs r12,r12,#1
320 bne .Lenc_loop
321
322 add r10,r10,#2
323
324 ldrb r4,[r10,r7,lsl#2] @ Te4[s0>>0]
325 and r7,lr,r1,lsr#16 @ i0
326 ldrb r5,[r10,r8,lsl#2] @ Te4[s0>>8]
327 and r8,lr,r1
328 ldrb r6,[r10,r9,lsl#2] @ Te4[s0>>16]
329 and r9,lr,r1,lsr#8
330 ldrb r0,[r10,r0,lsl#2] @ Te4[s0>>24]
331 mov r1,r1,lsr#24
332
333 ldrb r7,[r10,r7,lsl#2] @ Te4[s1>>16]
334 ldrb r8,[r10,r8,lsl#2] @ Te4[s1>>0]
335 ldrb r9,[r10,r9,lsl#2] @ Te4[s1>>8]
336 eor r0,r7,r0,lsl#8
337 ldrb r1,[r10,r1,lsl#2] @ Te4[s1>>24]
338 and r7,lr,r2,lsr#8 @ i0
339 eor r5,r8,r5,lsl#8
340 and r8,lr,r2,lsr#16 @ i1
341 eor r6,r9,r6,lsl#8
342 and r9,lr,r2
343 ldrb r7,[r10,r7,lsl#2] @ Te4[s2>>8]
344 eor r1,r4,r1,lsl#24
345 ldrb r8,[r10,r8,lsl#2] @ Te4[s2>>16]
346 mov r2,r2,lsr#24
347
348 ldrb r9,[r10,r9,lsl#2] @ Te4[s2>>0]
349 eor r0,r7,r0,lsl#8
350 ldrb r2,[r10,r2,lsl#2] @ Te4[s2>>24]
351 and r7,lr,r3 @ i0
352 eor r1,r1,r8,lsl#16
353 and r8,lr,r3,lsr#8 @ i1
354 eor r6,r9,r6,lsl#8
355 and r9,lr,r3,lsr#16 @ i2
356 ldrb r7,[r10,r7,lsl#2] @ Te4[s3>>0]
357 eor r2,r5,r2,lsl#24
358 ldrb r8,[r10,r8,lsl#2] @ Te4[s3>>8]
359 mov r3,r3,lsr#24
360
361 ldrb r9,[r10,r9,lsl#2] @ Te4[s3>>16]
362 eor r0,r7,r0,lsl#8
363 ldr r7,[r11,#0]
364 ldrb r3,[r10,r3,lsl#2] @ Te4[s3>>24]
365 eor r1,r1,r8,lsl#8
366 ldr r4,[r11,#4]
367 eor r2,r2,r9,lsl#16
368 ldr r5,[r11,#8]
369 eor r3,r6,r3,lsl#24
370 ldr r6,[r11,#12]
371
372 eor r0,r0,r7
373 eor r1,r1,r4
374 eor r2,r2,r5
375 eor r3,r3,r6
376
377 sub r10,r10,#2
378 ldr pc,[sp],#4 @ pop and return
379.size _armv4_AES_encrypt,.-_armv4_AES_encrypt
380
David McCulloughf0be44f2012-09-07 04:17:02 +0800381.align 5
Dave Martin638591c2013-01-10 12:20:15 +0100382ENTRY(private_AES_set_encrypt_key)
David McCulloughf0be44f2012-09-07 04:17:02 +0800383_armv4_AES_set_encrypt_key:
Ard Biesheuvel40190c82013-09-21 11:23:50 +0100384 adr r3,_armv4_AES_set_encrypt_key
David McCulloughf0be44f2012-09-07 04:17:02 +0800385 teq r0,#0
386 moveq r0,#-1
387 beq .Labrt
388 teq r2,#0
389 moveq r0,#-1
390 beq .Labrt
391
392 teq r1,#128
393 beq .Lok
394 teq r1,#192
395 beq .Lok
396 teq r1,#256
397 movne r0,#-1
398 bne .Labrt
399
400.Lok: stmdb sp!,{r4-r12,lr}
401 sub r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024 @ Te4
402
403 mov r12,r0 @ inp
404 mov lr,r1 @ bits
405 mov r11,r2 @ key
406
407#if __ARM_ARCH__<7
408 ldrb r0,[r12,#3] @ load input data in endian-neutral
409 ldrb r4,[r12,#2] @ manner...
410 ldrb r5,[r12,#1]
411 ldrb r6,[r12,#0]
412 orr r0,r0,r4,lsl#8
413 ldrb r1,[r12,#7]
414 orr r0,r0,r5,lsl#16
415 ldrb r4,[r12,#6]
416 orr r0,r0,r6,lsl#24
417 ldrb r5,[r12,#5]
418 ldrb r6,[r12,#4]
419 orr r1,r1,r4,lsl#8
420 ldrb r2,[r12,#11]
421 orr r1,r1,r5,lsl#16
422 ldrb r4,[r12,#10]
423 orr r1,r1,r6,lsl#24
424 ldrb r5,[r12,#9]
425 ldrb r6,[r12,#8]
426 orr r2,r2,r4,lsl#8
427 ldrb r3,[r12,#15]
428 orr r2,r2,r5,lsl#16
429 ldrb r4,[r12,#14]
430 orr r2,r2,r6,lsl#24
431 ldrb r5,[r12,#13]
432 ldrb r6,[r12,#12]
433 orr r3,r3,r4,lsl#8
434 str r0,[r11],#16
435 orr r3,r3,r5,lsl#16
436 str r1,[r11,#-12]
437 orr r3,r3,r6,lsl#24
438 str r2,[r11,#-8]
439 str r3,[r11,#-4]
440#else
441 ldr r0,[r12,#0]
442 ldr r1,[r12,#4]
443 ldr r2,[r12,#8]
444 ldr r3,[r12,#12]
445#ifdef __ARMEL__
446 rev r0,r0
447 rev r1,r1
448 rev r2,r2
449 rev r3,r3
450#endif
451 str r0,[r11],#16
452 str r1,[r11,#-12]
453 str r2,[r11,#-8]
454 str r3,[r11,#-4]
455#endif
456
457 teq lr,#128
458 bne .Lnot128
459 mov r12,#10
460 str r12,[r11,#240-16]
461 add r6,r10,#256 @ rcon
462 mov lr,#255
463
464.L128_loop:
465 and r5,lr,r3,lsr#24
466 and r7,lr,r3,lsr#16
467 ldrb r5,[r10,r5]
468 and r8,lr,r3,lsr#8
469 ldrb r7,[r10,r7]
470 and r9,lr,r3
471 ldrb r8,[r10,r8]
472 orr r5,r5,r7,lsl#24
473 ldrb r9,[r10,r9]
474 orr r5,r5,r8,lsl#16
475 ldr r4,[r6],#4 @ rcon[i++]
476 orr r5,r5,r9,lsl#8
477 eor r5,r5,r4
478 eor r0,r0,r5 @ rk[4]=rk[0]^...
479 eor r1,r1,r0 @ rk[5]=rk[1]^rk[4]
480 str r0,[r11],#16
481 eor r2,r2,r1 @ rk[6]=rk[2]^rk[5]
482 str r1,[r11,#-12]
483 eor r3,r3,r2 @ rk[7]=rk[3]^rk[6]
484 str r2,[r11,#-8]
485 subs r12,r12,#1
486 str r3,[r11,#-4]
487 bne .L128_loop
488 sub r2,r11,#176
489 b .Ldone
490
491.Lnot128:
492#if __ARM_ARCH__<7
493 ldrb r8,[r12,#19]
494 ldrb r4,[r12,#18]
495 ldrb r5,[r12,#17]
496 ldrb r6,[r12,#16]
497 orr r8,r8,r4,lsl#8
498 ldrb r9,[r12,#23]
499 orr r8,r8,r5,lsl#16
500 ldrb r4,[r12,#22]
501 orr r8,r8,r6,lsl#24
502 ldrb r5,[r12,#21]
503 ldrb r6,[r12,#20]
504 orr r9,r9,r4,lsl#8
505 orr r9,r9,r5,lsl#16
506 str r8,[r11],#8
507 orr r9,r9,r6,lsl#24
508 str r9,[r11,#-4]
509#else
510 ldr r8,[r12,#16]
511 ldr r9,[r12,#20]
512#ifdef __ARMEL__
513 rev r8,r8
514 rev r9,r9
515#endif
516 str r8,[r11],#8
517 str r9,[r11,#-4]
518#endif
519
520 teq lr,#192
521 bne .Lnot192
522 mov r12,#12
523 str r12,[r11,#240-24]
524 add r6,r10,#256 @ rcon
525 mov lr,#255
526 mov r12,#8
527
528.L192_loop:
529 and r5,lr,r9,lsr#24
530 and r7,lr,r9,lsr#16
531 ldrb r5,[r10,r5]
532 and r8,lr,r9,lsr#8
533 ldrb r7,[r10,r7]
534 and r9,lr,r9
535 ldrb r8,[r10,r8]
536 orr r5,r5,r7,lsl#24
537 ldrb r9,[r10,r9]
538 orr r5,r5,r8,lsl#16
539 ldr r4,[r6],#4 @ rcon[i++]
540 orr r5,r5,r9,lsl#8
541 eor r9,r5,r4
542 eor r0,r0,r9 @ rk[6]=rk[0]^...
543 eor r1,r1,r0 @ rk[7]=rk[1]^rk[6]
544 str r0,[r11],#24
545 eor r2,r2,r1 @ rk[8]=rk[2]^rk[7]
546 str r1,[r11,#-20]
547 eor r3,r3,r2 @ rk[9]=rk[3]^rk[8]
548 str r2,[r11,#-16]
549 subs r12,r12,#1
550 str r3,[r11,#-12]
551 subeq r2,r11,#216
552 beq .Ldone
553
554 ldr r7,[r11,#-32]
555 ldr r8,[r11,#-28]
556 eor r7,r7,r3 @ rk[10]=rk[4]^rk[9]
557 eor r9,r8,r7 @ rk[11]=rk[5]^rk[10]
558 str r7,[r11,#-8]
559 str r9,[r11,#-4]
560 b .L192_loop
561
562.Lnot192:
563#if __ARM_ARCH__<7
564 ldrb r8,[r12,#27]
565 ldrb r4,[r12,#26]
566 ldrb r5,[r12,#25]
567 ldrb r6,[r12,#24]
568 orr r8,r8,r4,lsl#8
569 ldrb r9,[r12,#31]
570 orr r8,r8,r5,lsl#16
571 ldrb r4,[r12,#30]
572 orr r8,r8,r6,lsl#24
573 ldrb r5,[r12,#29]
574 ldrb r6,[r12,#28]
575 orr r9,r9,r4,lsl#8
576 orr r9,r9,r5,lsl#16
577 str r8,[r11],#8
578 orr r9,r9,r6,lsl#24
579 str r9,[r11,#-4]
580#else
581 ldr r8,[r12,#24]
582 ldr r9,[r12,#28]
583#ifdef __ARMEL__
584 rev r8,r8
585 rev r9,r9
586#endif
587 str r8,[r11],#8
588 str r9,[r11,#-4]
589#endif
590
591 mov r12,#14
592 str r12,[r11,#240-32]
593 add r6,r10,#256 @ rcon
594 mov lr,#255
595 mov r12,#7
596
597.L256_loop:
598 and r5,lr,r9,lsr#24
599 and r7,lr,r9,lsr#16
600 ldrb r5,[r10,r5]
601 and r8,lr,r9,lsr#8
602 ldrb r7,[r10,r7]
603 and r9,lr,r9
604 ldrb r8,[r10,r8]
605 orr r5,r5,r7,lsl#24
606 ldrb r9,[r10,r9]
607 orr r5,r5,r8,lsl#16
608 ldr r4,[r6],#4 @ rcon[i++]
609 orr r5,r5,r9,lsl#8
610 eor r9,r5,r4
611 eor r0,r0,r9 @ rk[8]=rk[0]^...
612 eor r1,r1,r0 @ rk[9]=rk[1]^rk[8]
613 str r0,[r11],#32
614 eor r2,r2,r1 @ rk[10]=rk[2]^rk[9]
615 str r1,[r11,#-28]
616 eor r3,r3,r2 @ rk[11]=rk[3]^rk[10]
617 str r2,[r11,#-24]
618 subs r12,r12,#1
619 str r3,[r11,#-20]
620 subeq r2,r11,#256
621 beq .Ldone
622
623 and r5,lr,r3
624 and r7,lr,r3,lsr#8
625 ldrb r5,[r10,r5]
626 and r8,lr,r3,lsr#16
627 ldrb r7,[r10,r7]
628 and r9,lr,r3,lsr#24
629 ldrb r8,[r10,r8]
630 orr r5,r5,r7,lsl#8
631 ldrb r9,[r10,r9]
632 orr r5,r5,r8,lsl#16
633 ldr r4,[r11,#-48]
634 orr r5,r5,r9,lsl#24
635
636 ldr r7,[r11,#-44]
637 ldr r8,[r11,#-40]
638 eor r4,r4,r5 @ rk[12]=rk[4]^...
639 ldr r9,[r11,#-36]
640 eor r7,r7,r4 @ rk[13]=rk[5]^rk[12]
641 str r4,[r11,#-16]
642 eor r8,r8,r7 @ rk[14]=rk[6]^rk[13]
643 str r7,[r11,#-12]
644 eor r9,r9,r8 @ rk[15]=rk[7]^rk[14]
645 str r8,[r11,#-8]
646 str r9,[r11,#-4]
647 b .L256_loop
648
649.Ldone: mov r0,#0
650 ldmia sp!,{r4-r12,lr}
Dave Martin638591c2013-01-10 12:20:15 +0100651.Labrt: mov pc,lr
652ENDPROC(private_AES_set_encrypt_key)
David McCulloughf0be44f2012-09-07 04:17:02 +0800653
David McCulloughf0be44f2012-09-07 04:17:02 +0800654.align 5
Dave Martin638591c2013-01-10 12:20:15 +0100655ENTRY(private_AES_set_decrypt_key)
David McCulloughf0be44f2012-09-07 04:17:02 +0800656 str lr,[sp,#-4]! @ push lr
657#if 0
658 @ kernel does both of these in setkey so optimise this bit out by
659 @ expecting the key to already have the enc_key work done (see aes_glue.c)
660 bl _armv4_AES_set_encrypt_key
661#else
662 mov r0,#0
663#endif
664 teq r0,#0
665 ldrne lr,[sp],#4 @ pop lr
666 bne .Labrt
667
668 stmdb sp!,{r4-r12}
669
670 ldr r12,[r2,#240] @ AES_set_encrypt_key preserves r2,
671 mov r11,r2 @ which is AES_KEY *key
672 mov r7,r2
673 add r8,r2,r12,lsl#4
674
675.Linv: ldr r0,[r7]
676 ldr r1,[r7,#4]
677 ldr r2,[r7,#8]
678 ldr r3,[r7,#12]
679 ldr r4,[r8]
680 ldr r5,[r8,#4]
681 ldr r6,[r8,#8]
682 ldr r9,[r8,#12]
683 str r0,[r8],#-16
684 str r1,[r8,#16+4]
685 str r2,[r8,#16+8]
686 str r3,[r8,#16+12]
687 str r4,[r7],#16
688 str r5,[r7,#-12]
689 str r6,[r7,#-8]
690 str r9,[r7,#-4]
691 teq r7,r8
692 bne .Linv
693 ldr r0,[r11,#16]! @ prefetch tp1
694 mov r7,#0x80
695 mov r8,#0x1b
696 orr r7,r7,#0x8000
697 orr r8,r8,#0x1b00
698 orr r7,r7,r7,lsl#16
699 orr r8,r8,r8,lsl#16
700 sub r12,r12,#1
701 mvn r9,r7
702 mov r12,r12,lsl#2 @ (rounds-1)*4
703
704.Lmix: and r4,r0,r7
705 and r1,r0,r9
706 sub r4,r4,r4,lsr#7
707 and r4,r4,r8
708 eor r1,r4,r1,lsl#1 @ tp2
709
710 and r4,r1,r7
711 and r2,r1,r9
712 sub r4,r4,r4,lsr#7
713 and r4,r4,r8
714 eor r2,r4,r2,lsl#1 @ tp4
715
716 and r4,r2,r7
717 and r3,r2,r9
718 sub r4,r4,r4,lsr#7
719 and r4,r4,r8
720 eor r3,r4,r3,lsl#1 @ tp8
721
722 eor r4,r1,r2
723 eor r5,r0,r3 @ tp9
724 eor r4,r4,r3 @ tpe
725 eor r4,r4,r1,ror#24
726 eor r4,r4,r5,ror#24 @ ^= ROTATE(tpb=tp9^tp2,8)
727 eor r4,r4,r2,ror#16
728 eor r4,r4,r5,ror#16 @ ^= ROTATE(tpd=tp9^tp4,16)
729 eor r4,r4,r5,ror#8 @ ^= ROTATE(tp9,24)
730
731 ldr r0,[r11,#4] @ prefetch tp1
732 str r4,[r11],#4
733 subs r12,r12,#1
734 bne .Lmix
735
736 mov r0,#0
David McCulloughf0be44f2012-09-07 04:17:02 +0800737 ldmia sp!,{r4-r12,pc}
Dave Martin638591c2013-01-10 12:20:15 +0100738ENDPROC(private_AES_set_decrypt_key)
David McCulloughf0be44f2012-09-07 04:17:02 +0800739
740.type AES_Td,%object
741.align 5
742AES_Td:
743.word 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
744.word 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
745.word 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
746.word 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
747.word 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
748.word 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
749.word 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
750.word 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
751.word 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
752.word 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
753.word 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
754.word 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
755.word 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
756.word 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
757.word 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
758.word 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
759.word 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
760.word 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
761.word 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
762.word 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
763.word 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
764.word 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
765.word 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
766.word 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
767.word 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
768.word 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
769.word 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
770.word 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
771.word 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
772.word 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
773.word 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
774.word 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
775.word 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
776.word 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
777.word 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
778.word 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
779.word 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
780.word 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
781.word 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
782.word 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
783.word 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
784.word 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
785.word 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
786.word 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
787.word 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
788.word 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
789.word 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
790.word 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
791.word 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
792.word 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
793.word 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
794.word 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
795.word 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
796.word 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
797.word 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
798.word 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
799.word 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
800.word 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
801.word 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
802.word 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
803.word 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
804.word 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
805.word 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
806.word 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
807@ Td4[256]
808.byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
809.byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
810.byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
811.byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
812.byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
813.byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
814.byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
815.byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
816.byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
817.byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
818.byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
819.byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
820.byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
821.byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
822.byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
823.byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
824.byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
825.byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
826.byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
827.byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
828.byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
829.byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
830.byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
831.byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
832.byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
833.byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
834.byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
835.byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
836.byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
837.byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
838.byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
839.byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
840.size AES_Td,.-AES_Td
841
842@ void AES_decrypt(const unsigned char *in, unsigned char *out,
843@ const AES_KEY *key) {
David McCulloughf0be44f2012-09-07 04:17:02 +0800844.align 5
Dave Martin638591c2013-01-10 12:20:15 +0100845ENTRY(AES_decrypt)
Ard Biesheuvel40190c82013-09-21 11:23:50 +0100846 adr r3,AES_decrypt
David McCulloughf0be44f2012-09-07 04:17:02 +0800847 stmdb sp!,{r1,r4-r12,lr}
848 mov r12,r0 @ inp
849 mov r11,r2
850 sub r10,r3,#AES_decrypt-AES_Td @ Td
851#if __ARM_ARCH__<7
852 ldrb r0,[r12,#3] @ load input data in endian-neutral
853 ldrb r4,[r12,#2] @ manner...
854 ldrb r5,[r12,#1]
855 ldrb r6,[r12,#0]
856 orr r0,r0,r4,lsl#8
857 ldrb r1,[r12,#7]
858 orr r0,r0,r5,lsl#16
859 ldrb r4,[r12,#6]
860 orr r0,r0,r6,lsl#24
861 ldrb r5,[r12,#5]
862 ldrb r6,[r12,#4]
863 orr r1,r1,r4,lsl#8
864 ldrb r2,[r12,#11]
865 orr r1,r1,r5,lsl#16
866 ldrb r4,[r12,#10]
867 orr r1,r1,r6,lsl#24
868 ldrb r5,[r12,#9]
869 ldrb r6,[r12,#8]
870 orr r2,r2,r4,lsl#8
871 ldrb r3,[r12,#15]
872 orr r2,r2,r5,lsl#16
873 ldrb r4,[r12,#14]
874 orr r2,r2,r6,lsl#24
875 ldrb r5,[r12,#13]
876 ldrb r6,[r12,#12]
877 orr r3,r3,r4,lsl#8
878 orr r3,r3,r5,lsl#16
879 orr r3,r3,r6,lsl#24
880#else
881 ldr r0,[r12,#0]
882 ldr r1,[r12,#4]
883 ldr r2,[r12,#8]
884 ldr r3,[r12,#12]
885#ifdef __ARMEL__
886 rev r0,r0
887 rev r1,r1
888 rev r2,r2
889 rev r3,r3
890#endif
891#endif
892 bl _armv4_AES_decrypt
893
894 ldr r12,[sp],#4 @ pop out
895#if __ARM_ARCH__>=7
896#ifdef __ARMEL__
897 rev r0,r0
898 rev r1,r1
899 rev r2,r2
900 rev r3,r3
901#endif
902 str r0,[r12,#0]
903 str r1,[r12,#4]
904 str r2,[r12,#8]
905 str r3,[r12,#12]
906#else
907 mov r4,r0,lsr#24 @ write output in endian-neutral
908 mov r5,r0,lsr#16 @ manner...
909 mov r6,r0,lsr#8
910 strb r4,[r12,#0]
911 strb r5,[r12,#1]
912 mov r4,r1,lsr#24
913 strb r6,[r12,#2]
914 mov r5,r1,lsr#16
915 strb r0,[r12,#3]
916 mov r6,r1,lsr#8
917 strb r4,[r12,#4]
918 strb r5,[r12,#5]
919 mov r4,r2,lsr#24
920 strb r6,[r12,#6]
921 mov r5,r2,lsr#16
922 strb r1,[r12,#7]
923 mov r6,r2,lsr#8
924 strb r4,[r12,#8]
925 strb r5,[r12,#9]
926 mov r4,r3,lsr#24
927 strb r6,[r12,#10]
928 mov r5,r3,lsr#16
929 strb r2,[r12,#11]
930 mov r6,r3,lsr#8
931 strb r4,[r12,#12]
932 strb r5,[r12,#13]
933 strb r6,[r12,#14]
934 strb r3,[r12,#15]
935#endif
David McCulloughf0be44f2012-09-07 04:17:02 +0800936 ldmia sp!,{r4-r12,pc}
Dave Martin638591c2013-01-10 12:20:15 +0100937ENDPROC(AES_decrypt)
David McCulloughf0be44f2012-09-07 04:17:02 +0800938
939.type _armv4_AES_decrypt,%function
940.align 2
941_armv4_AES_decrypt:
942 str lr,[sp,#-4]! @ push lr
943 ldmia r11!,{r4-r7}
944 eor r0,r0,r4
945 ldr r12,[r11,#240-16]
946 eor r1,r1,r5
947 eor r2,r2,r6
948 eor r3,r3,r7
949 sub r12,r12,#1
950 mov lr,#255
951
952 and r7,lr,r0,lsr#16
953 and r8,lr,r0,lsr#8
954 and r9,lr,r0
955 mov r0,r0,lsr#24
956.Ldec_loop:
957 ldr r4,[r10,r7,lsl#2] @ Td1[s0>>16]
958 and r7,lr,r1 @ i0
959 ldr r5,[r10,r8,lsl#2] @ Td2[s0>>8]
960 and r8,lr,r1,lsr#16
961 ldr r6,[r10,r9,lsl#2] @ Td3[s0>>0]
962 and r9,lr,r1,lsr#8
963 ldr r0,[r10,r0,lsl#2] @ Td0[s0>>24]
964 mov r1,r1,lsr#24
965
966 ldr r7,[r10,r7,lsl#2] @ Td3[s1>>0]
967 ldr r8,[r10,r8,lsl#2] @ Td1[s1>>16]
968 ldr r9,[r10,r9,lsl#2] @ Td2[s1>>8]
969 eor r0,r0,r7,ror#24
970 ldr r1,[r10,r1,lsl#2] @ Td0[s1>>24]
971 and r7,lr,r2,lsr#8 @ i0
972 eor r5,r8,r5,ror#8
973 and r8,lr,r2 @ i1
974 eor r6,r9,r6,ror#8
975 and r9,lr,r2,lsr#16
976 ldr r7,[r10,r7,lsl#2] @ Td2[s2>>8]
977 eor r1,r1,r4,ror#8
978 ldr r8,[r10,r8,lsl#2] @ Td3[s2>>0]
979 mov r2,r2,lsr#24
980
981 ldr r9,[r10,r9,lsl#2] @ Td1[s2>>16]
982 eor r0,r0,r7,ror#16
983 ldr r2,[r10,r2,lsl#2] @ Td0[s2>>24]
984 and r7,lr,r3,lsr#16 @ i0
985 eor r1,r1,r8,ror#24
986 and r8,lr,r3,lsr#8 @ i1
987 eor r6,r9,r6,ror#8
988 and r9,lr,r3 @ i2
989 ldr r7,[r10,r7,lsl#2] @ Td1[s3>>16]
990 eor r2,r2,r5,ror#8
991 ldr r8,[r10,r8,lsl#2] @ Td2[s3>>8]
992 mov r3,r3,lsr#24
993
994 ldr r9,[r10,r9,lsl#2] @ Td3[s3>>0]
995 eor r0,r0,r7,ror#8
996 ldr r7,[r11],#16
997 eor r1,r1,r8,ror#16
998 ldr r3,[r10,r3,lsl#2] @ Td0[s3>>24]
999 eor r2,r2,r9,ror#24
1000
1001 ldr r4,[r11,#-12]
1002 eor r0,r0,r7
1003 ldr r5,[r11,#-8]
1004 eor r3,r3,r6,ror#8
1005 ldr r6,[r11,#-4]
1006 and r7,lr,r0,lsr#16
1007 eor r1,r1,r4
1008 and r8,lr,r0,lsr#8
1009 eor r2,r2,r5
1010 and r9,lr,r0
1011 eor r3,r3,r6
1012 mov r0,r0,lsr#24
1013
1014 subs r12,r12,#1
1015 bne .Ldec_loop
1016
1017 add r10,r10,#1024
1018
1019 ldr r5,[r10,#0] @ prefetch Td4
1020 ldr r6,[r10,#32]
1021 ldr r4,[r10,#64]
1022 ldr r5,[r10,#96]
1023 ldr r6,[r10,#128]
1024 ldr r4,[r10,#160]
1025 ldr r5,[r10,#192]
1026 ldr r6,[r10,#224]
1027
1028 ldrb r0,[r10,r0] @ Td4[s0>>24]
1029 ldrb r4,[r10,r7] @ Td4[s0>>16]
1030 and r7,lr,r1 @ i0
1031 ldrb r5,[r10,r8] @ Td4[s0>>8]
1032 and r8,lr,r1,lsr#16
1033 ldrb r6,[r10,r9] @ Td4[s0>>0]
1034 and r9,lr,r1,lsr#8
1035
1036 ldrb r7,[r10,r7] @ Td4[s1>>0]
Dave Martin638591c2013-01-10 12:20:15 +01001037 ARM( ldrb r1,[r10,r1,lsr#24] ) @ Td4[s1>>24]
1038 THUMB( add r1,r10,r1,lsr#24 ) @ Td4[s1>>24]
1039 THUMB( ldrb r1,[r1] )
David McCulloughf0be44f2012-09-07 04:17:02 +08001040 ldrb r8,[r10,r8] @ Td4[s1>>16]
1041 eor r0,r7,r0,lsl#24
1042 ldrb r9,[r10,r9] @ Td4[s1>>8]
1043 eor r1,r4,r1,lsl#8
1044 and r7,lr,r2,lsr#8 @ i0
1045 eor r5,r5,r8,lsl#8
1046 and r8,lr,r2 @ i1
1047 ldrb r7,[r10,r7] @ Td4[s2>>8]
1048 eor r6,r6,r9,lsl#8
1049 ldrb r8,[r10,r8] @ Td4[s2>>0]
1050 and r9,lr,r2,lsr#16
1051
Dave Martin638591c2013-01-10 12:20:15 +01001052 ARM( ldrb r2,[r10,r2,lsr#24] ) @ Td4[s2>>24]
1053 THUMB( add r2,r10,r2,lsr#24 ) @ Td4[s2>>24]
1054 THUMB( ldrb r2,[r2] )
David McCulloughf0be44f2012-09-07 04:17:02 +08001055 eor r0,r0,r7,lsl#8
1056 ldrb r9,[r10,r9] @ Td4[s2>>16]
1057 eor r1,r8,r1,lsl#16
1058 and r7,lr,r3,lsr#16 @ i0
1059 eor r2,r5,r2,lsl#16
1060 and r8,lr,r3,lsr#8 @ i1
1061 ldrb r7,[r10,r7] @ Td4[s3>>16]
1062 eor r6,r6,r9,lsl#16
1063 ldrb r8,[r10,r8] @ Td4[s3>>8]
1064 and r9,lr,r3 @ i2
1065
1066 ldrb r9,[r10,r9] @ Td4[s3>>0]
Dave Martin638591c2013-01-10 12:20:15 +01001067 ARM( ldrb r3,[r10,r3,lsr#24] ) @ Td4[s3>>24]
1068 THUMB( add r3,r10,r3,lsr#24 ) @ Td4[s3>>24]
1069 THUMB( ldrb r3,[r3] )
David McCulloughf0be44f2012-09-07 04:17:02 +08001070 eor r0,r0,r7,lsl#16
1071 ldr r7,[r11,#0]
1072 eor r1,r1,r8,lsl#8
1073 ldr r4,[r11,#4]
1074 eor r2,r9,r2,lsl#8
1075 ldr r5,[r11,#8]
1076 eor r3,r6,r3,lsl#24
1077 ldr r6,[r11,#12]
1078
1079 eor r0,r0,r7
1080 eor r1,r1,r4
1081 eor r2,r2,r5
1082 eor r3,r3,r6
1083
1084 sub r10,r10,#1024
1085 ldr pc,[sp],#4 @ pop and return
1086.size _armv4_AES_decrypt,.-_armv4_AES_decrypt
1087.asciz "AES for ARMv4, CRYPTOGAMS by <appro@openssl.org>"
1088.align 2