aes-armv4.S 28 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089
  1. #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. @ AES for ARMv4
  9. @ January 2007.
  10. @
  11. @ Code uses single 1K S-box and is >2 times faster than code generated
  12. @ by gcc-3.4.1. This is thanks to unique feature of ARMv4 ISA, which
  13. @ allows to merge logical or arithmetic operation with shift or rotate
  14. @ in one instruction and emit combined result every cycle. The module
  15. @ is endian-neutral. The performance is ~42 cycles/byte for 128-bit
  16. @ key [on single-issue Xscale PXA250 core].
  17. @ May 2007.
  18. @
  19. @ AES_set_[en|de]crypt_key is added.
  20. @ July 2010.
  21. @
  22. @ Rescheduling for dual-issue pipeline resulted in 12% improvement on
  23. @ Cortex A8 core and ~25 cycles per byte processed with 128-bit key.
  24. @ February 2011.
  25. @
  26. @ Profiler-assisted and platform-specific optimization resulted in 16%
  27. @ improvement on Cortex A8 core and ~21.5 cycles per byte.
  28. @ A little glue here to select the correct code below for the ARM CPU
  29. @ that is being targetted.
  30. #include <linux/linkage.h>
  31. .text
  32. .type AES_Te,%object
  33. .align 5
  34. AES_Te:
  35. .word 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
  36. .word 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
  37. .word 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
  38. .word 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
  39. .word 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
  40. .word 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
  41. .word 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
  42. .word 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
  43. .word 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
  44. .word 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
  45. .word 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
  46. .word 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
  47. .word 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
  48. .word 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
  49. .word 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
  50. .word 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
  51. .word 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
  52. .word 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
  53. .word 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
  54. .word 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
  55. .word 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
  56. .word 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
  57. .word 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
  58. .word 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
  59. .word 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
  60. .word 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
  61. .word 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
  62. .word 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
  63. .word 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
  64. .word 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
  65. .word 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
  66. .word 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
  67. .word 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
  68. .word 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
  69. .word 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
  70. .word 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
  71. .word 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
  72. .word 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
  73. .word 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
  74. .word 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
  75. .word 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
  76. .word 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
  77. .word 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
  78. .word 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
  79. .word 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
  80. .word 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
  81. .word 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
  82. .word 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
  83. .word 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
  84. .word 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
  85. .word 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
  86. .word 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
  87. .word 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
  88. .word 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
  89. .word 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
  90. .word 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
  91. .word 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
  92. .word 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
  93. .word 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
  94. .word 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
  95. .word 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
  96. .word 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
  97. .word 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
  98. .word 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
  99. @ Te4[256]
  100. .byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
  101. .byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
  102. .byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
  103. .byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
  104. .byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
  105. .byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
  106. .byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
  107. .byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
  108. .byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
  109. .byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
  110. .byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
  111. .byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
  112. .byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
  113. .byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
  114. .byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
  115. .byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
  116. .byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
  117. .byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
  118. .byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
  119. .byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
  120. .byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
  121. .byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
  122. .byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
  123. .byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
  124. .byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
  125. .byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
  126. .byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
  127. .byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
  128. .byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
  129. .byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
  130. .byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
  131. .byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
  132. @ rcon[]
  133. .word 0x01000000, 0x02000000, 0x04000000, 0x08000000
  134. .word 0x10000000, 0x20000000, 0x40000000, 0x80000000
  135. .word 0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0
  136. .size AES_Te,.-AES_Te
  137. @ void AES_encrypt(const unsigned char *in, unsigned char *out,
  138. @ const AES_KEY *key) {
  139. .align 5
  140. ENTRY(AES_encrypt)
  141. adr r3,AES_encrypt
  142. stmdb sp!,{r1,r4-r12,lr}
  143. mov r12,r0 @ inp
  144. mov r11,r2
  145. sub r10,r3,#AES_encrypt-AES_Te @ Te
  146. #if __ARM_ARCH__<7
  147. ldrb r0,[r12,#3] @ load input data in endian-neutral
  148. ldrb r4,[r12,#2] @ manner...
  149. ldrb r5,[r12,#1]
  150. ldrb r6,[r12,#0]
  151. orr r0,r0,r4,lsl#8
  152. ldrb r1,[r12,#7]
  153. orr r0,r0,r5,lsl#16
  154. ldrb r4,[r12,#6]
  155. orr r0,r0,r6,lsl#24
  156. ldrb r5,[r12,#5]
  157. ldrb r6,[r12,#4]
  158. orr r1,r1,r4,lsl#8
  159. ldrb r2,[r12,#11]
  160. orr r1,r1,r5,lsl#16
  161. ldrb r4,[r12,#10]
  162. orr r1,r1,r6,lsl#24
  163. ldrb r5,[r12,#9]
  164. ldrb r6,[r12,#8]
  165. orr r2,r2,r4,lsl#8
  166. ldrb r3,[r12,#15]
  167. orr r2,r2,r5,lsl#16
  168. ldrb r4,[r12,#14]
  169. orr r2,r2,r6,lsl#24
  170. ldrb r5,[r12,#13]
  171. ldrb r6,[r12,#12]
  172. orr r3,r3,r4,lsl#8
  173. orr r3,r3,r5,lsl#16
  174. orr r3,r3,r6,lsl#24
  175. #else
  176. ldr r0,[r12,#0]
  177. ldr r1,[r12,#4]
  178. ldr r2,[r12,#8]
  179. ldr r3,[r12,#12]
  180. #ifdef __ARMEL__
  181. rev r0,r0
  182. rev r1,r1
  183. rev r2,r2
  184. rev r3,r3
  185. #endif
  186. #endif
  187. bl _armv4_AES_encrypt
  188. ldr r12,[sp],#4 @ pop out
  189. #if __ARM_ARCH__>=7
  190. #ifdef __ARMEL__
  191. rev r0,r0
  192. rev r1,r1
  193. rev r2,r2
  194. rev r3,r3
  195. #endif
  196. str r0,[r12,#0]
  197. str r1,[r12,#4]
  198. str r2,[r12,#8]
  199. str r3,[r12,#12]
  200. #else
  201. mov r4,r0,lsr#24 @ write output in endian-neutral
  202. mov r5,r0,lsr#16 @ manner...
  203. mov r6,r0,lsr#8
  204. strb r4,[r12,#0]
  205. strb r5,[r12,#1]
  206. mov r4,r1,lsr#24
  207. strb r6,[r12,#2]
  208. mov r5,r1,lsr#16
  209. strb r0,[r12,#3]
  210. mov r6,r1,lsr#8
  211. strb r4,[r12,#4]
  212. strb r5,[r12,#5]
  213. mov r4,r2,lsr#24
  214. strb r6,[r12,#6]
  215. mov r5,r2,lsr#16
  216. strb r1,[r12,#7]
  217. mov r6,r2,lsr#8
  218. strb r4,[r12,#8]
  219. strb r5,[r12,#9]
  220. mov r4,r3,lsr#24
  221. strb r6,[r12,#10]
  222. mov r5,r3,lsr#16
  223. strb r2,[r12,#11]
  224. mov r6,r3,lsr#8
  225. strb r4,[r12,#12]
  226. strb r5,[r12,#13]
  227. strb r6,[r12,#14]
  228. strb r3,[r12,#15]
  229. #endif
  230. ldmia sp!,{r4-r12,pc}
  231. ENDPROC(AES_encrypt)
  232. .type _armv4_AES_encrypt,%function
  233. .align 2
  234. _armv4_AES_encrypt:
  235. str lr,[sp,#-4]! @ push lr
  236. ldmia r11!,{r4-r7}
  237. eor r0,r0,r4
  238. ldr r12,[r11,#240-16]
  239. eor r1,r1,r5
  240. eor r2,r2,r6
  241. eor r3,r3,r7
  242. sub r12,r12,#1
  243. mov lr,#255
  244. and r7,lr,r0
  245. and r8,lr,r0,lsr#8
  246. and r9,lr,r0,lsr#16
  247. mov r0,r0,lsr#24
  248. .Lenc_loop:
  249. ldr r4,[r10,r7,lsl#2] @ Te3[s0>>0]
  250. and r7,lr,r1,lsr#16 @ i0
  251. ldr r5,[r10,r8,lsl#2] @ Te2[s0>>8]
  252. and r8,lr,r1
  253. ldr r6,[r10,r9,lsl#2] @ Te1[s0>>16]
  254. and r9,lr,r1,lsr#8
  255. ldr r0,[r10,r0,lsl#2] @ Te0[s0>>24]
  256. mov r1,r1,lsr#24
  257. ldr r7,[r10,r7,lsl#2] @ Te1[s1>>16]
  258. ldr r8,[r10,r8,lsl#2] @ Te3[s1>>0]
  259. ldr r9,[r10,r9,lsl#2] @ Te2[s1>>8]
  260. eor r0,r0,r7,ror#8
  261. ldr r1,[r10,r1,lsl#2] @ Te0[s1>>24]
  262. and r7,lr,r2,lsr#8 @ i0
  263. eor r5,r5,r8,ror#8
  264. and r8,lr,r2,lsr#16 @ i1
  265. eor r6,r6,r9,ror#8
  266. and r9,lr,r2
  267. ldr r7,[r10,r7,lsl#2] @ Te2[s2>>8]
  268. eor r1,r1,r4,ror#24
  269. ldr r8,[r10,r8,lsl#2] @ Te1[s2>>16]
  270. mov r2,r2,lsr#24
  271. ldr r9,[r10,r9,lsl#2] @ Te3[s2>>0]
  272. eor r0,r0,r7,ror#16
  273. ldr r2,[r10,r2,lsl#2] @ Te0[s2>>24]
  274. and r7,lr,r3 @ i0
  275. eor r1,r1,r8,ror#8
  276. and r8,lr,r3,lsr#8 @ i1
  277. eor r6,r6,r9,ror#16
  278. and r9,lr,r3,lsr#16 @ i2
  279. ldr r7,[r10,r7,lsl#2] @ Te3[s3>>0]
  280. eor r2,r2,r5,ror#16
  281. ldr r8,[r10,r8,lsl#2] @ Te2[s3>>8]
  282. mov r3,r3,lsr#24
  283. ldr r9,[r10,r9,lsl#2] @ Te1[s3>>16]
  284. eor r0,r0,r7,ror#24
  285. ldr r7,[r11],#16
  286. eor r1,r1,r8,ror#16
  287. ldr r3,[r10,r3,lsl#2] @ Te0[s3>>24]
  288. eor r2,r2,r9,ror#8
  289. ldr r4,[r11,#-12]
  290. eor r3,r3,r6,ror#8
  291. ldr r5,[r11,#-8]
  292. eor r0,r0,r7
  293. ldr r6,[r11,#-4]
  294. and r7,lr,r0
  295. eor r1,r1,r4
  296. and r8,lr,r0,lsr#8
  297. eor r2,r2,r5
  298. and r9,lr,r0,lsr#16
  299. eor r3,r3,r6
  300. mov r0,r0,lsr#24
  301. subs r12,r12,#1
  302. bne .Lenc_loop
  303. add r10,r10,#2
  304. ldrb r4,[r10,r7,lsl#2] @ Te4[s0>>0]
  305. and r7,lr,r1,lsr#16 @ i0
  306. ldrb r5,[r10,r8,lsl#2] @ Te4[s0>>8]
  307. and r8,lr,r1
  308. ldrb r6,[r10,r9,lsl#2] @ Te4[s0>>16]
  309. and r9,lr,r1,lsr#8
  310. ldrb r0,[r10,r0,lsl#2] @ Te4[s0>>24]
  311. mov r1,r1,lsr#24
  312. ldrb r7,[r10,r7,lsl#2] @ Te4[s1>>16]
  313. ldrb r8,[r10,r8,lsl#2] @ Te4[s1>>0]
  314. ldrb r9,[r10,r9,lsl#2] @ Te4[s1>>8]
  315. eor r0,r7,r0,lsl#8
  316. ldrb r1,[r10,r1,lsl#2] @ Te4[s1>>24]
  317. and r7,lr,r2,lsr#8 @ i0
  318. eor r5,r8,r5,lsl#8
  319. and r8,lr,r2,lsr#16 @ i1
  320. eor r6,r9,r6,lsl#8
  321. and r9,lr,r2
  322. ldrb r7,[r10,r7,lsl#2] @ Te4[s2>>8]
  323. eor r1,r4,r1,lsl#24
  324. ldrb r8,[r10,r8,lsl#2] @ Te4[s2>>16]
  325. mov r2,r2,lsr#24
  326. ldrb r9,[r10,r9,lsl#2] @ Te4[s2>>0]
  327. eor r0,r7,r0,lsl#8
  328. ldrb r2,[r10,r2,lsl#2] @ Te4[s2>>24]
  329. and r7,lr,r3 @ i0
  330. eor r1,r1,r8,lsl#16
  331. and r8,lr,r3,lsr#8 @ i1
  332. eor r6,r9,r6,lsl#8
  333. and r9,lr,r3,lsr#16 @ i2
  334. ldrb r7,[r10,r7,lsl#2] @ Te4[s3>>0]
  335. eor r2,r5,r2,lsl#24
  336. ldrb r8,[r10,r8,lsl#2] @ Te4[s3>>8]
  337. mov r3,r3,lsr#24
  338. ldrb r9,[r10,r9,lsl#2] @ Te4[s3>>16]
  339. eor r0,r7,r0,lsl#8
  340. ldr r7,[r11,#0]
  341. ldrb r3,[r10,r3,lsl#2] @ Te4[s3>>24]
  342. eor r1,r1,r8,lsl#8
  343. ldr r4,[r11,#4]
  344. eor r2,r2,r9,lsl#16
  345. ldr r5,[r11,#8]
  346. eor r3,r6,r3,lsl#24
  347. ldr r6,[r11,#12]
  348. eor r0,r0,r7
  349. eor r1,r1,r4
  350. eor r2,r2,r5
  351. eor r3,r3,r6
  352. sub r10,r10,#2
  353. ldr pc,[sp],#4 @ pop and return
  354. .size _armv4_AES_encrypt,.-_armv4_AES_encrypt
  355. .align 5
  356. ENTRY(private_AES_set_encrypt_key)
  357. _armv4_AES_set_encrypt_key:
  358. adr r3,_armv4_AES_set_encrypt_key
  359. teq r0,#0
  360. moveq r0,#-1
  361. beq .Labrt
  362. teq r2,#0
  363. moveq r0,#-1
  364. beq .Labrt
  365. teq r1,#128
  366. beq .Lok
  367. teq r1,#192
  368. beq .Lok
  369. teq r1,#256
  370. movne r0,#-1
  371. bne .Labrt
  372. .Lok: stmdb sp!,{r4-r12,lr}
  373. sub r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024 @ Te4
  374. mov r12,r0 @ inp
  375. mov lr,r1 @ bits
  376. mov r11,r2 @ key
  377. #if __ARM_ARCH__<7
  378. ldrb r0,[r12,#3] @ load input data in endian-neutral
  379. ldrb r4,[r12,#2] @ manner...
  380. ldrb r5,[r12,#1]
  381. ldrb r6,[r12,#0]
  382. orr r0,r0,r4,lsl#8
  383. ldrb r1,[r12,#7]
  384. orr r0,r0,r5,lsl#16
  385. ldrb r4,[r12,#6]
  386. orr r0,r0,r6,lsl#24
  387. ldrb r5,[r12,#5]
  388. ldrb r6,[r12,#4]
  389. orr r1,r1,r4,lsl#8
  390. ldrb r2,[r12,#11]
  391. orr r1,r1,r5,lsl#16
  392. ldrb r4,[r12,#10]
  393. orr r1,r1,r6,lsl#24
  394. ldrb r5,[r12,#9]
  395. ldrb r6,[r12,#8]
  396. orr r2,r2,r4,lsl#8
  397. ldrb r3,[r12,#15]
  398. orr r2,r2,r5,lsl#16
  399. ldrb r4,[r12,#14]
  400. orr r2,r2,r6,lsl#24
  401. ldrb r5,[r12,#13]
  402. ldrb r6,[r12,#12]
  403. orr r3,r3,r4,lsl#8
  404. str r0,[r11],#16
  405. orr r3,r3,r5,lsl#16
  406. str r1,[r11,#-12]
  407. orr r3,r3,r6,lsl#24
  408. str r2,[r11,#-8]
  409. str r3,[r11,#-4]
  410. #else
  411. ldr r0,[r12,#0]
  412. ldr r1,[r12,#4]
  413. ldr r2,[r12,#8]
  414. ldr r3,[r12,#12]
  415. #ifdef __ARMEL__
  416. rev r0,r0
  417. rev r1,r1
  418. rev r2,r2
  419. rev r3,r3
  420. #endif
  421. str r0,[r11],#16
  422. str r1,[r11,#-12]
  423. str r2,[r11,#-8]
  424. str r3,[r11,#-4]
  425. #endif
  426. teq lr,#128
  427. bne .Lnot128
  428. mov r12,#10
  429. str r12,[r11,#240-16]
  430. add r6,r10,#256 @ rcon
  431. mov lr,#255
  432. .L128_loop:
  433. and r5,lr,r3,lsr#24
  434. and r7,lr,r3,lsr#16
  435. ldrb r5,[r10,r5]
  436. and r8,lr,r3,lsr#8
  437. ldrb r7,[r10,r7]
  438. and r9,lr,r3
  439. ldrb r8,[r10,r8]
  440. orr r5,r5,r7,lsl#24
  441. ldrb r9,[r10,r9]
  442. orr r5,r5,r8,lsl#16
  443. ldr r4,[r6],#4 @ rcon[i++]
  444. orr r5,r5,r9,lsl#8
  445. eor r5,r5,r4
  446. eor r0,r0,r5 @ rk[4]=rk[0]^...
  447. eor r1,r1,r0 @ rk[5]=rk[1]^rk[4]
  448. str r0,[r11],#16
  449. eor r2,r2,r1 @ rk[6]=rk[2]^rk[5]
  450. str r1,[r11,#-12]
  451. eor r3,r3,r2 @ rk[7]=rk[3]^rk[6]
  452. str r2,[r11,#-8]
  453. subs r12,r12,#1
  454. str r3,[r11,#-4]
  455. bne .L128_loop
  456. sub r2,r11,#176
  457. b .Ldone
  458. .Lnot128:
  459. #if __ARM_ARCH__<7
  460. ldrb r8,[r12,#19]
  461. ldrb r4,[r12,#18]
  462. ldrb r5,[r12,#17]
  463. ldrb r6,[r12,#16]
  464. orr r8,r8,r4,lsl#8
  465. ldrb r9,[r12,#23]
  466. orr r8,r8,r5,lsl#16
  467. ldrb r4,[r12,#22]
  468. orr r8,r8,r6,lsl#24
  469. ldrb r5,[r12,#21]
  470. ldrb r6,[r12,#20]
  471. orr r9,r9,r4,lsl#8
  472. orr r9,r9,r5,lsl#16
  473. str r8,[r11],#8
  474. orr r9,r9,r6,lsl#24
  475. str r9,[r11,#-4]
  476. #else
  477. ldr r8,[r12,#16]
  478. ldr r9,[r12,#20]
  479. #ifdef __ARMEL__
  480. rev r8,r8
  481. rev r9,r9
  482. #endif
  483. str r8,[r11],#8
  484. str r9,[r11,#-4]
  485. #endif
  486. teq lr,#192
  487. bne .Lnot192
  488. mov r12,#12
  489. str r12,[r11,#240-24]
  490. add r6,r10,#256 @ rcon
  491. mov lr,#255
  492. mov r12,#8
  493. .L192_loop:
  494. and r5,lr,r9,lsr#24
  495. and r7,lr,r9,lsr#16
  496. ldrb r5,[r10,r5]
  497. and r8,lr,r9,lsr#8
  498. ldrb r7,[r10,r7]
  499. and r9,lr,r9
  500. ldrb r8,[r10,r8]
  501. orr r5,r5,r7,lsl#24
  502. ldrb r9,[r10,r9]
  503. orr r5,r5,r8,lsl#16
  504. ldr r4,[r6],#4 @ rcon[i++]
  505. orr r5,r5,r9,lsl#8
  506. eor r9,r5,r4
  507. eor r0,r0,r9 @ rk[6]=rk[0]^...
  508. eor r1,r1,r0 @ rk[7]=rk[1]^rk[6]
  509. str r0,[r11],#24
  510. eor r2,r2,r1 @ rk[8]=rk[2]^rk[7]
  511. str r1,[r11,#-20]
  512. eor r3,r3,r2 @ rk[9]=rk[3]^rk[8]
  513. str r2,[r11,#-16]
  514. subs r12,r12,#1
  515. str r3,[r11,#-12]
  516. subeq r2,r11,#216
  517. beq .Ldone
  518. ldr r7,[r11,#-32]
  519. ldr r8,[r11,#-28]
  520. eor r7,r7,r3 @ rk[10]=rk[4]^rk[9]
  521. eor r9,r8,r7 @ rk[11]=rk[5]^rk[10]
  522. str r7,[r11,#-8]
  523. str r9,[r11,#-4]
  524. b .L192_loop
  525. .Lnot192:
  526. #if __ARM_ARCH__<7
  527. ldrb r8,[r12,#27]
  528. ldrb r4,[r12,#26]
  529. ldrb r5,[r12,#25]
  530. ldrb r6,[r12,#24]
  531. orr r8,r8,r4,lsl#8
  532. ldrb r9,[r12,#31]
  533. orr r8,r8,r5,lsl#16
  534. ldrb r4,[r12,#30]
  535. orr r8,r8,r6,lsl#24
  536. ldrb r5,[r12,#29]
  537. ldrb r6,[r12,#28]
  538. orr r9,r9,r4,lsl#8
  539. orr r9,r9,r5,lsl#16
  540. str r8,[r11],#8
  541. orr r9,r9,r6,lsl#24
  542. str r9,[r11,#-4]
  543. #else
  544. ldr r8,[r12,#24]
  545. ldr r9,[r12,#28]
  546. #ifdef __ARMEL__
  547. rev r8,r8
  548. rev r9,r9
  549. #endif
  550. str r8,[r11],#8
  551. str r9,[r11,#-4]
  552. #endif
  553. mov r12,#14
  554. str r12,[r11,#240-32]
  555. add r6,r10,#256 @ rcon
  556. mov lr,#255
  557. mov r12,#7
  558. .L256_loop:
  559. and r5,lr,r9,lsr#24
  560. and r7,lr,r9,lsr#16
  561. ldrb r5,[r10,r5]
  562. and r8,lr,r9,lsr#8
  563. ldrb r7,[r10,r7]
  564. and r9,lr,r9
  565. ldrb r8,[r10,r8]
  566. orr r5,r5,r7,lsl#24
  567. ldrb r9,[r10,r9]
  568. orr r5,r5,r8,lsl#16
  569. ldr r4,[r6],#4 @ rcon[i++]
  570. orr r5,r5,r9,lsl#8
  571. eor r9,r5,r4
  572. eor r0,r0,r9 @ rk[8]=rk[0]^...
  573. eor r1,r1,r0 @ rk[9]=rk[1]^rk[8]
  574. str r0,[r11],#32
  575. eor r2,r2,r1 @ rk[10]=rk[2]^rk[9]
  576. str r1,[r11,#-28]
  577. eor r3,r3,r2 @ rk[11]=rk[3]^rk[10]
  578. str r2,[r11,#-24]
  579. subs r12,r12,#1
  580. str r3,[r11,#-20]
  581. subeq r2,r11,#256
  582. beq .Ldone
  583. and r5,lr,r3
  584. and r7,lr,r3,lsr#8
  585. ldrb r5,[r10,r5]
  586. and r8,lr,r3,lsr#16
  587. ldrb r7,[r10,r7]
  588. and r9,lr,r3,lsr#24
  589. ldrb r8,[r10,r8]
  590. orr r5,r5,r7,lsl#8
  591. ldrb r9,[r10,r9]
  592. orr r5,r5,r8,lsl#16
  593. ldr r4,[r11,#-48]
  594. orr r5,r5,r9,lsl#24
  595. ldr r7,[r11,#-44]
  596. ldr r8,[r11,#-40]
  597. eor r4,r4,r5 @ rk[12]=rk[4]^...
  598. ldr r9,[r11,#-36]
  599. eor r7,r7,r4 @ rk[13]=rk[5]^rk[12]
  600. str r4,[r11,#-16]
  601. eor r8,r8,r7 @ rk[14]=rk[6]^rk[13]
  602. str r7,[r11,#-12]
  603. eor r9,r9,r8 @ rk[15]=rk[7]^rk[14]
  604. str r8,[r11,#-8]
  605. str r9,[r11,#-4]
  606. b .L256_loop
  607. .Ldone: mov r0,#0
  608. ldmia sp!,{r4-r12,lr}
  609. .Labrt: mov pc,lr
  610. ENDPROC(private_AES_set_encrypt_key)
  611. .align 5
  612. ENTRY(private_AES_set_decrypt_key)
  613. str lr,[sp,#-4]! @ push lr
  614. #if 0
  615. @ kernel does both of these in setkey so optimise this bit out by
  616. @ expecting the key to already have the enc_key work done (see aes_glue.c)
  617. bl _armv4_AES_set_encrypt_key
  618. #else
  619. mov r0,#0
  620. #endif
  621. teq r0,#0
  622. ldrne lr,[sp],#4 @ pop lr
  623. bne .Labrt
  624. stmdb sp!,{r4-r12}
  625. ldr r12,[r2,#240] @ AES_set_encrypt_key preserves r2,
  626. mov r11,r2 @ which is AES_KEY *key
  627. mov r7,r2
  628. add r8,r2,r12,lsl#4
  629. .Linv: ldr r0,[r7]
  630. ldr r1,[r7,#4]
  631. ldr r2,[r7,#8]
  632. ldr r3,[r7,#12]
  633. ldr r4,[r8]
  634. ldr r5,[r8,#4]
  635. ldr r6,[r8,#8]
  636. ldr r9,[r8,#12]
  637. str r0,[r8],#-16
  638. str r1,[r8,#16+4]
  639. str r2,[r8,#16+8]
  640. str r3,[r8,#16+12]
  641. str r4,[r7],#16
  642. str r5,[r7,#-12]
  643. str r6,[r7,#-8]
  644. str r9,[r7,#-4]
  645. teq r7,r8
  646. bne .Linv
  647. ldr r0,[r11,#16]! @ prefetch tp1
  648. mov r7,#0x80
  649. mov r8,#0x1b
  650. orr r7,r7,#0x8000
  651. orr r8,r8,#0x1b00
  652. orr r7,r7,r7,lsl#16
  653. orr r8,r8,r8,lsl#16
  654. sub r12,r12,#1
  655. mvn r9,r7
  656. mov r12,r12,lsl#2 @ (rounds-1)*4
  657. .Lmix: and r4,r0,r7
  658. and r1,r0,r9
  659. sub r4,r4,r4,lsr#7
  660. and r4,r4,r8
  661. eor r1,r4,r1,lsl#1 @ tp2
  662. and r4,r1,r7
  663. and r2,r1,r9
  664. sub r4,r4,r4,lsr#7
  665. and r4,r4,r8
  666. eor r2,r4,r2,lsl#1 @ tp4
  667. and r4,r2,r7
  668. and r3,r2,r9
  669. sub r4,r4,r4,lsr#7
  670. and r4,r4,r8
  671. eor r3,r4,r3,lsl#1 @ tp8
  672. eor r4,r1,r2
  673. eor r5,r0,r3 @ tp9
  674. eor r4,r4,r3 @ tpe
  675. eor r4,r4,r1,ror#24
  676. eor r4,r4,r5,ror#24 @ ^= ROTATE(tpb=tp9^tp2,8)
  677. eor r4,r4,r2,ror#16
  678. eor r4,r4,r5,ror#16 @ ^= ROTATE(tpd=tp9^tp4,16)
  679. eor r4,r4,r5,ror#8 @ ^= ROTATE(tp9,24)
  680. ldr r0,[r11,#4] @ prefetch tp1
  681. str r4,[r11],#4
  682. subs r12,r12,#1
  683. bne .Lmix
  684. mov r0,#0
  685. ldmia sp!,{r4-r12,pc}
  686. ENDPROC(private_AES_set_decrypt_key)
  687. .type AES_Td,%object
  688. .align 5
  689. AES_Td:
  690. .word 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
  691. .word 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
  692. .word 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
  693. .word 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
  694. .word 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
  695. .word 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
  696. .word 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
  697. .word 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
  698. .word 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
  699. .word 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
  700. .word 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
  701. .word 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
  702. .word 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
  703. .word 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
  704. .word 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
  705. .word 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
  706. .word 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
  707. .word 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
  708. .word 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
  709. .word 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
  710. .word 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
  711. .word 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
  712. .word 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
  713. .word 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
  714. .word 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
  715. .word 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
  716. .word 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
  717. .word 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
  718. .word 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
  719. .word 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
  720. .word 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
  721. .word 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
  722. .word 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
  723. .word 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
  724. .word 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
  725. .word 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
  726. .word 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
  727. .word 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
  728. .word 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
  729. .word 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
  730. .word 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
  731. .word 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
  732. .word 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
  733. .word 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
  734. .word 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
  735. .word 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
  736. .word 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
  737. .word 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
  738. .word 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
  739. .word 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
  740. .word 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
  741. .word 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
  742. .word 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
  743. .word 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
  744. .word 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
  745. .word 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
  746. .word 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
  747. .word 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
  748. .word 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
  749. .word 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
  750. .word 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
  751. .word 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
  752. .word 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
  753. .word 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
  754. @ Td4[256]
  755. .byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
  756. .byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
  757. .byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
  758. .byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
  759. .byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
  760. .byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
  761. .byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
  762. .byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
  763. .byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
  764. .byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
  765. .byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
  766. .byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
  767. .byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
  768. .byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
  769. .byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
  770. .byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
  771. .byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
  772. .byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
  773. .byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
  774. .byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
  775. .byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
  776. .byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
  777. .byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
  778. .byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
  779. .byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
  780. .byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
  781. .byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
  782. .byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
  783. .byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
  784. .byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
  785. .byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
  786. .byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
  787. .size AES_Td,.-AES_Td
  788. @ void AES_decrypt(const unsigned char *in, unsigned char *out,
  789. @ const AES_KEY *key) {
  790. .align 5
  791. ENTRY(AES_decrypt)
  792. adr r3,AES_decrypt
  793. stmdb sp!,{r1,r4-r12,lr}
  794. mov r12,r0 @ inp
  795. mov r11,r2
  796. sub r10,r3,#AES_decrypt-AES_Td @ Td
  797. #if __ARM_ARCH__<7
  798. ldrb r0,[r12,#3] @ load input data in endian-neutral
  799. ldrb r4,[r12,#2] @ manner...
  800. ldrb r5,[r12,#1]
  801. ldrb r6,[r12,#0]
  802. orr r0,r0,r4,lsl#8
  803. ldrb r1,[r12,#7]
  804. orr r0,r0,r5,lsl#16
  805. ldrb r4,[r12,#6]
  806. orr r0,r0,r6,lsl#24
  807. ldrb r5,[r12,#5]
  808. ldrb r6,[r12,#4]
  809. orr r1,r1,r4,lsl#8
  810. ldrb r2,[r12,#11]
  811. orr r1,r1,r5,lsl#16
  812. ldrb r4,[r12,#10]
  813. orr r1,r1,r6,lsl#24
  814. ldrb r5,[r12,#9]
  815. ldrb r6,[r12,#8]
  816. orr r2,r2,r4,lsl#8
  817. ldrb r3,[r12,#15]
  818. orr r2,r2,r5,lsl#16
  819. ldrb r4,[r12,#14]
  820. orr r2,r2,r6,lsl#24
  821. ldrb r5,[r12,#13]
  822. ldrb r6,[r12,#12]
  823. orr r3,r3,r4,lsl#8
  824. orr r3,r3,r5,lsl#16
  825. orr r3,r3,r6,lsl#24
  826. #else
  827. ldr r0,[r12,#0]
  828. ldr r1,[r12,#4]
  829. ldr r2,[r12,#8]
  830. ldr r3,[r12,#12]
  831. #ifdef __ARMEL__
  832. rev r0,r0
  833. rev r1,r1
  834. rev r2,r2
  835. rev r3,r3
  836. #endif
  837. #endif
  838. bl _armv4_AES_decrypt
  839. ldr r12,[sp],#4 @ pop out
  840. #if __ARM_ARCH__>=7
  841. #ifdef __ARMEL__
  842. rev r0,r0
  843. rev r1,r1
  844. rev r2,r2
  845. rev r3,r3
  846. #endif
  847. str r0,[r12,#0]
  848. str r1,[r12,#4]
  849. str r2,[r12,#8]
  850. str r3,[r12,#12]
  851. #else
  852. mov r4,r0,lsr#24 @ write output in endian-neutral
  853. mov r5,r0,lsr#16 @ manner...
  854. mov r6,r0,lsr#8
  855. strb r4,[r12,#0]
  856. strb r5,[r12,#1]
  857. mov r4,r1,lsr#24
  858. strb r6,[r12,#2]
  859. mov r5,r1,lsr#16
  860. strb r0,[r12,#3]
  861. mov r6,r1,lsr#8
  862. strb r4,[r12,#4]
  863. strb r5,[r12,#5]
  864. mov r4,r2,lsr#24
  865. strb r6,[r12,#6]
  866. mov r5,r2,lsr#16
  867. strb r1,[r12,#7]
  868. mov r6,r2,lsr#8
  869. strb r4,[r12,#8]
  870. strb r5,[r12,#9]
  871. mov r4,r3,lsr#24
  872. strb r6,[r12,#10]
  873. mov r5,r3,lsr#16
  874. strb r2,[r12,#11]
  875. mov r6,r3,lsr#8
  876. strb r4,[r12,#12]
  877. strb r5,[r12,#13]
  878. strb r6,[r12,#14]
  879. strb r3,[r12,#15]
  880. #endif
  881. ldmia sp!,{r4-r12,pc}
  882. ENDPROC(AES_decrypt)
  883. .type _armv4_AES_decrypt,%function
  884. .align 2
  885. _armv4_AES_decrypt:
  886. str lr,[sp,#-4]! @ push lr
  887. ldmia r11!,{r4-r7}
  888. eor r0,r0,r4
  889. ldr r12,[r11,#240-16]
  890. eor r1,r1,r5
  891. eor r2,r2,r6
  892. eor r3,r3,r7
  893. sub r12,r12,#1
  894. mov lr,#255
  895. and r7,lr,r0,lsr#16
  896. and r8,lr,r0,lsr#8
  897. and r9,lr,r0
  898. mov r0,r0,lsr#24
  899. .Ldec_loop:
  900. ldr r4,[r10,r7,lsl#2] @ Td1[s0>>16]
  901. and r7,lr,r1 @ i0
  902. ldr r5,[r10,r8,lsl#2] @ Td2[s0>>8]
  903. and r8,lr,r1,lsr#16
  904. ldr r6,[r10,r9,lsl#2] @ Td3[s0>>0]
  905. and r9,lr,r1,lsr#8
  906. ldr r0,[r10,r0,lsl#2] @ Td0[s0>>24]
  907. mov r1,r1,lsr#24
  908. ldr r7,[r10,r7,lsl#2] @ Td3[s1>>0]
  909. ldr r8,[r10,r8,lsl#2] @ Td1[s1>>16]
  910. ldr r9,[r10,r9,lsl#2] @ Td2[s1>>8]
  911. eor r0,r0,r7,ror#24
  912. ldr r1,[r10,r1,lsl#2] @ Td0[s1>>24]
  913. and r7,lr,r2,lsr#8 @ i0
  914. eor r5,r8,r5,ror#8
  915. and r8,lr,r2 @ i1
  916. eor r6,r9,r6,ror#8
  917. and r9,lr,r2,lsr#16
  918. ldr r7,[r10,r7,lsl#2] @ Td2[s2>>8]
  919. eor r1,r1,r4,ror#8
  920. ldr r8,[r10,r8,lsl#2] @ Td3[s2>>0]
  921. mov r2,r2,lsr#24
  922. ldr r9,[r10,r9,lsl#2] @ Td1[s2>>16]
  923. eor r0,r0,r7,ror#16
  924. ldr r2,[r10,r2,lsl#2] @ Td0[s2>>24]
  925. and r7,lr,r3,lsr#16 @ i0
  926. eor r1,r1,r8,ror#24
  927. and r8,lr,r3,lsr#8 @ i1
  928. eor r6,r9,r6,ror#8
  929. and r9,lr,r3 @ i2
  930. ldr r7,[r10,r7,lsl#2] @ Td1[s3>>16]
  931. eor r2,r2,r5,ror#8
  932. ldr r8,[r10,r8,lsl#2] @ Td2[s3>>8]
  933. mov r3,r3,lsr#24
  934. ldr r9,[r10,r9,lsl#2] @ Td3[s3>>0]
  935. eor r0,r0,r7,ror#8
  936. ldr r7,[r11],#16
  937. eor r1,r1,r8,ror#16
  938. ldr r3,[r10,r3,lsl#2] @ Td0[s3>>24]
  939. eor r2,r2,r9,ror#24
  940. ldr r4,[r11,#-12]
  941. eor r0,r0,r7
  942. ldr r5,[r11,#-8]
  943. eor r3,r3,r6,ror#8
  944. ldr r6,[r11,#-4]
  945. and r7,lr,r0,lsr#16
  946. eor r1,r1,r4
  947. and r8,lr,r0,lsr#8
  948. eor r2,r2,r5
  949. and r9,lr,r0
  950. eor r3,r3,r6
  951. mov r0,r0,lsr#24
  952. subs r12,r12,#1
  953. bne .Ldec_loop
  954. add r10,r10,#1024
  955. ldr r5,[r10,#0] @ prefetch Td4
  956. ldr r6,[r10,#32]
  957. ldr r4,[r10,#64]
  958. ldr r5,[r10,#96]
  959. ldr r6,[r10,#128]
  960. ldr r4,[r10,#160]
  961. ldr r5,[r10,#192]
  962. ldr r6,[r10,#224]
  963. ldrb r0,[r10,r0] @ Td4[s0>>24]
  964. ldrb r4,[r10,r7] @ Td4[s0>>16]
  965. and r7,lr,r1 @ i0
  966. ldrb r5,[r10,r8] @ Td4[s0>>8]
  967. and r8,lr,r1,lsr#16
  968. ldrb r6,[r10,r9] @ Td4[s0>>0]
  969. and r9,lr,r1,lsr#8
  970. ldrb r7,[r10,r7] @ Td4[s1>>0]
  971. ARM( ldrb r1,[r10,r1,lsr#24] ) @ Td4[s1>>24]
  972. THUMB( add r1,r10,r1,lsr#24 ) @ Td4[s1>>24]
  973. THUMB( ldrb r1,[r1] )
  974. ldrb r8,[r10,r8] @ Td4[s1>>16]
  975. eor r0,r7,r0,lsl#24
  976. ldrb r9,[r10,r9] @ Td4[s1>>8]
  977. eor r1,r4,r1,lsl#8
  978. and r7,lr,r2,lsr#8 @ i0
  979. eor r5,r5,r8,lsl#8
  980. and r8,lr,r2 @ i1
  981. ldrb r7,[r10,r7] @ Td4[s2>>8]
  982. eor r6,r6,r9,lsl#8
  983. ldrb r8,[r10,r8] @ Td4[s2>>0]
  984. and r9,lr,r2,lsr#16
  985. ARM( ldrb r2,[r10,r2,lsr#24] ) @ Td4[s2>>24]
  986. THUMB( add r2,r10,r2,lsr#24 ) @ Td4[s2>>24]
  987. THUMB( ldrb r2,[r2] )
  988. eor r0,r0,r7,lsl#8
  989. ldrb r9,[r10,r9] @ Td4[s2>>16]
  990. eor r1,r8,r1,lsl#16
  991. and r7,lr,r3,lsr#16 @ i0
  992. eor r2,r5,r2,lsl#16
  993. and r8,lr,r3,lsr#8 @ i1
  994. ldrb r7,[r10,r7] @ Td4[s3>>16]
  995. eor r6,r6,r9,lsl#16
  996. ldrb r8,[r10,r8] @ Td4[s3>>8]
  997. and r9,lr,r3 @ i2
  998. ldrb r9,[r10,r9] @ Td4[s3>>0]
  999. ARM( ldrb r3,[r10,r3,lsr#24] ) @ Td4[s3>>24]
  1000. THUMB( add r3,r10,r3,lsr#24 ) @ Td4[s3>>24]
  1001. THUMB( ldrb r3,[r3] )
  1002. eor r0,r0,r7,lsl#16
  1003. ldr r7,[r11,#0]
  1004. eor r1,r1,r8,lsl#8
  1005. ldr r4,[r11,#4]
  1006. eor r2,r9,r2,lsl#8
  1007. ldr r5,[r11,#8]
  1008. eor r3,r6,r3,lsl#24
  1009. ldr r6,[r11,#12]
  1010. eor r0,r0,r7
  1011. eor r1,r1,r4
  1012. eor r2,r2,r5
  1013. eor r3,r3,r6
  1014. sub r10,r10,#1024
  1015. ldr pc,[sp],#4 @ pop and return
  1016. .size _armv4_AES_decrypt,.-_armv4_AES_decrypt
  1017. .asciz "AES for ARMv4, CRYPTOGAMS by <appro@openssl.org>"
  1018. .align 2