Home | History | Annotate | Download | only in aes
      1 
      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 @ ====================================================================
      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 #if defined(__arm__)
     35 #ifndef __KERNEL__
     36 # include "arm_arch.h"
     37 #else
     38 # define __ARM_ARCH__ __LINUX_ARM_ARCH__
     39 #endif
     40 
     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
     52 
     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
    158 
    159 @ void asm_AES_encrypt(const unsigned char *in, unsigned char *out,
    160 @ 		       const AES_KEY *key) {
    161 .global asm_AES_encrypt
    162 .hidden asm_AES_encrypt
    163 .type   asm_AES_encrypt,%function
    164 .align	5
    165 asm_AES_encrypt:
    166 #if __ARM_ARCH__<7
    167 	sub	r3,pc,#8		@ asm_AES_encrypt
    168 #else
    169 	adr	r3,asm_AES_encrypt
    170 #endif
    171 	stmdb   sp!,{r1,r4-r12,lr}
    172 	mov	r12,r0		@ inp
    173 	mov	r11,r2
    174 	sub	r10,r3,#asm_AES_encrypt-AES_Te	@ Te
    175 #if __ARM_ARCH__<7
    176 	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
    177 	ldrb	r4,[r12,#2]	@ manner...
    178 	ldrb	r5,[r12,#1]
    179 	ldrb	r6,[r12,#0]
    180 	orr	r0,r0,r4,lsl#8
    181 	ldrb	r1,[r12,#7]
    182 	orr	r0,r0,r5,lsl#16
    183 	ldrb	r4,[r12,#6]
    184 	orr	r0,r0,r6,lsl#24
    185 	ldrb	r5,[r12,#5]
    186 	ldrb	r6,[r12,#4]
    187 	orr	r1,r1,r4,lsl#8
    188 	ldrb	r2,[r12,#11]
    189 	orr	r1,r1,r5,lsl#16
    190 	ldrb	r4,[r12,#10]
    191 	orr	r1,r1,r6,lsl#24
    192 	ldrb	r5,[r12,#9]
    193 	ldrb	r6,[r12,#8]
    194 	orr	r2,r2,r4,lsl#8
    195 	ldrb	r3,[r12,#15]
    196 	orr	r2,r2,r5,lsl#16
    197 	ldrb	r4,[r12,#14]
    198 	orr	r2,r2,r6,lsl#24
    199 	ldrb	r5,[r12,#13]
    200 	ldrb	r6,[r12,#12]
    201 	orr	r3,r3,r4,lsl#8
    202 	orr	r3,r3,r5,lsl#16
    203 	orr	r3,r3,r6,lsl#24
    204 #else
    205 	ldr	r0,[r12,#0]
    206 	ldr	r1,[r12,#4]
    207 	ldr	r2,[r12,#8]
    208 	ldr	r3,[r12,#12]
    209 #ifdef __ARMEL__
    210 	rev	r0,r0
    211 	rev	r1,r1
    212 	rev	r2,r2
    213 	rev	r3,r3
    214 #endif
    215 #endif
    216 	bl	_armv4_AES_encrypt
    217 
    218 	ldr	r12,[sp],#4		@ pop out
    219 #if __ARM_ARCH__>=7
    220 #ifdef __ARMEL__
    221 	rev	r0,r0
    222 	rev	r1,r1
    223 	rev	r2,r2
    224 	rev	r3,r3
    225 #endif
    226 	str	r0,[r12,#0]
    227 	str	r1,[r12,#4]
    228 	str	r2,[r12,#8]
    229 	str	r3,[r12,#12]
    230 #else
    231 	mov	r4,r0,lsr#24		@ write output in endian-neutral
    232 	mov	r5,r0,lsr#16		@ manner...
    233 	mov	r6,r0,lsr#8
    234 	strb	r4,[r12,#0]
    235 	strb	r5,[r12,#1]
    236 	mov	r4,r1,lsr#24
    237 	strb	r6,[r12,#2]
    238 	mov	r5,r1,lsr#16
    239 	strb	r0,[r12,#3]
    240 	mov	r6,r1,lsr#8
    241 	strb	r4,[r12,#4]
    242 	strb	r5,[r12,#5]
    243 	mov	r4,r2,lsr#24
    244 	strb	r6,[r12,#6]
    245 	mov	r5,r2,lsr#16
    246 	strb	r1,[r12,#7]
    247 	mov	r6,r2,lsr#8
    248 	strb	r4,[r12,#8]
    249 	strb	r5,[r12,#9]
    250 	mov	r4,r3,lsr#24
    251 	strb	r6,[r12,#10]
    252 	mov	r5,r3,lsr#16
    253 	strb	r2,[r12,#11]
    254 	mov	r6,r3,lsr#8
    255 	strb	r4,[r12,#12]
    256 	strb	r5,[r12,#13]
    257 	strb	r6,[r12,#14]
    258 	strb	r3,[r12,#15]
    259 #endif
    260 #if __ARM_ARCH__>=5
    261 	ldmia	sp!,{r4-r12,pc}
    262 #else
    263 	ldmia   sp!,{r4-r12,lr}
    264 	tst	lr,#1
    265 	moveq	pc,lr			@ be binary compatible with V4, yet
    266 	.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
    267 #endif
    268 .size	asm_AES_encrypt,.-asm_AES_encrypt
    269 
    270 .type   _armv4_AES_encrypt,%function
    271 .align	2
    272 _armv4_AES_encrypt:
    273 	str	lr,[sp,#-4]!		@ push lr
    274 	ldmia	r11!,{r4-r7}
    275 	eor	r0,r0,r4
    276 	ldr	r12,[r11,#240-16]
    277 	eor	r1,r1,r5
    278 	eor	r2,r2,r6
    279 	eor	r3,r3,r7
    280 	sub	r12,r12,#1
    281 	mov	lr,#255
    282 
    283 	and	r7,lr,r0
    284 	and	r8,lr,r0,lsr#8
    285 	and	r9,lr,r0,lsr#16
    286 	mov	r0,r0,lsr#24
    287 .Lenc_loop:
    288 	ldr	r4,[r10,r7,lsl#2]	@ Te3[s0>>0]
    289 	and	r7,lr,r1,lsr#16	@ i0
    290 	ldr	r5,[r10,r8,lsl#2]	@ Te2[s0>>8]
    291 	and	r8,lr,r1
    292 	ldr	r6,[r10,r9,lsl#2]	@ Te1[s0>>16]
    293 	and	r9,lr,r1,lsr#8
    294 	ldr	r0,[r10,r0,lsl#2]	@ Te0[s0>>24]
    295 	mov	r1,r1,lsr#24
    296 
    297 	ldr	r7,[r10,r7,lsl#2]	@ Te1[s1>>16]
    298 	ldr	r8,[r10,r8,lsl#2]	@ Te3[s1>>0]
    299 	ldr	r9,[r10,r9,lsl#2]	@ Te2[s1>>8]
    300 	eor	r0,r0,r7,ror#8
    301 	ldr	r1,[r10,r1,lsl#2]	@ Te0[s1>>24]
    302 	and	r7,lr,r2,lsr#8	@ i0
    303 	eor	r5,r5,r8,ror#8
    304 	and	r8,lr,r2,lsr#16	@ i1
    305 	eor	r6,r6,r9,ror#8
    306 	and	r9,lr,r2
    307 	ldr	r7,[r10,r7,lsl#2]	@ Te2[s2>>8]
    308 	eor	r1,r1,r4,ror#24
    309 	ldr	r8,[r10,r8,lsl#2]	@ Te1[s2>>16]
    310 	mov	r2,r2,lsr#24
    311 
    312 	ldr	r9,[r10,r9,lsl#2]	@ Te3[s2>>0]
    313 	eor	r0,r0,r7,ror#16
    314 	ldr	r2,[r10,r2,lsl#2]	@ Te0[s2>>24]
    315 	and	r7,lr,r3		@ i0
    316 	eor	r1,r1,r8,ror#8
    317 	and	r8,lr,r3,lsr#8	@ i1
    318 	eor	r6,r6,r9,ror#16
    319 	and	r9,lr,r3,lsr#16	@ i2
    320 	ldr	r7,[r10,r7,lsl#2]	@ Te3[s3>>0]
    321 	eor	r2,r2,r5,ror#16
    322 	ldr	r8,[r10,r8,lsl#2]	@ Te2[s3>>8]
    323 	mov	r3,r3,lsr#24
    324 
    325 	ldr	r9,[r10,r9,lsl#2]	@ Te1[s3>>16]
    326 	eor	r0,r0,r7,ror#24
    327 	ldr	r7,[r11],#16
    328 	eor	r1,r1,r8,ror#16
    329 	ldr	r3,[r10,r3,lsl#2]	@ Te0[s3>>24]
    330 	eor	r2,r2,r9,ror#8
    331 	ldr	r4,[r11,#-12]
    332 	eor	r3,r3,r6,ror#8
    333 
    334 	ldr	r5,[r11,#-8]
    335 	eor	r0,r0,r7
    336 	ldr	r6,[r11,#-4]
    337 	and	r7,lr,r0
    338 	eor	r1,r1,r4
    339 	and	r8,lr,r0,lsr#8
    340 	eor	r2,r2,r5
    341 	and	r9,lr,r0,lsr#16
    342 	eor	r3,r3,r6
    343 	mov	r0,r0,lsr#24
    344 
    345 	subs	r12,r12,#1
    346 	bne	.Lenc_loop
    347 
    348 	add	r10,r10,#2
    349 
    350 	ldrb	r4,[r10,r7,lsl#2]	@ Te4[s0>>0]
    351 	and	r7,lr,r1,lsr#16	@ i0
    352 	ldrb	r5,[r10,r8,lsl#2]	@ Te4[s0>>8]
    353 	and	r8,lr,r1
    354 	ldrb	r6,[r10,r9,lsl#2]	@ Te4[s0>>16]
    355 	and	r9,lr,r1,lsr#8
    356 	ldrb	r0,[r10,r0,lsl#2]	@ Te4[s0>>24]
    357 	mov	r1,r1,lsr#24
    358 
    359 	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s1>>16]
    360 	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s1>>0]
    361 	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s1>>8]
    362 	eor	r0,r7,r0,lsl#8
    363 	ldrb	r1,[r10,r1,lsl#2]	@ Te4[s1>>24]
    364 	and	r7,lr,r2,lsr#8	@ i0
    365 	eor	r5,r8,r5,lsl#8
    366 	and	r8,lr,r2,lsr#16	@ i1
    367 	eor	r6,r9,r6,lsl#8
    368 	and	r9,lr,r2
    369 	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s2>>8]
    370 	eor	r1,r4,r1,lsl#24
    371 	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s2>>16]
    372 	mov	r2,r2,lsr#24
    373 
    374 	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s2>>0]
    375 	eor	r0,r7,r0,lsl#8
    376 	ldrb	r2,[r10,r2,lsl#2]	@ Te4[s2>>24]
    377 	and	r7,lr,r3		@ i0
    378 	eor	r1,r1,r8,lsl#16
    379 	and	r8,lr,r3,lsr#8	@ i1
    380 	eor	r6,r9,r6,lsl#8
    381 	and	r9,lr,r3,lsr#16	@ i2
    382 	ldrb	r7,[r10,r7,lsl#2]	@ Te4[s3>>0]
    383 	eor	r2,r5,r2,lsl#24
    384 	ldrb	r8,[r10,r8,lsl#2]	@ Te4[s3>>8]
    385 	mov	r3,r3,lsr#24
    386 
    387 	ldrb	r9,[r10,r9,lsl#2]	@ Te4[s3>>16]
    388 	eor	r0,r7,r0,lsl#8
    389 	ldr	r7,[r11,#0]
    390 	ldrb	r3,[r10,r3,lsl#2]	@ Te4[s3>>24]
    391 	eor	r1,r1,r8,lsl#8
    392 	ldr	r4,[r11,#4]
    393 	eor	r2,r2,r9,lsl#16
    394 	ldr	r5,[r11,#8]
    395 	eor	r3,r6,r3,lsl#24
    396 	ldr	r6,[r11,#12]
    397 
    398 	eor	r0,r0,r7
    399 	eor	r1,r1,r4
    400 	eor	r2,r2,r5
    401 	eor	r3,r3,r6
    402 
    403 	sub	r10,r10,#2
    404 	ldr	pc,[sp],#4		@ pop and return
    405 .size	_armv4_AES_encrypt,.-_armv4_AES_encrypt
    406 
    407 .global asm_AES_set_encrypt_key
    408 .hidden asm_AES_set_encrypt_key
    409 .type   asm_AES_set_encrypt_key,%function
    410 .align	5
    411 asm_AES_set_encrypt_key:
    412 _armv4_AES_set_encrypt_key:
    413 #if __ARM_ARCH__<7
    414 	sub	r3,pc,#8		@ asm_AES_set_encrypt_key
    415 #else
    416 	adr	r3,asm_AES_set_encrypt_key
    417 #endif
    418 	teq	r0,#0
    419 #if __ARM_ARCH__>=7
    420 	itt	eq			@ Thumb2 thing, sanity check in ARM
    421 #endif
    422 	moveq	r0,#-1
    423 	beq	.Labrt
    424 	teq	r2,#0
    425 #if __ARM_ARCH__>=7
    426 	itt	eq			@ Thumb2 thing, sanity check in ARM
    427 #endif
    428 	moveq	r0,#-1
    429 	beq	.Labrt
    430 
    431 	teq	r1,#128
    432 	beq	.Lok
    433 	teq	r1,#192
    434 	beq	.Lok
    435 	teq	r1,#256
    436 #if __ARM_ARCH__>=7
    437 	itt	ne			@ Thumb2 thing, sanity check in ARM
    438 #endif
    439 	movne	r0,#-1
    440 	bne	.Labrt
    441 
    442 .Lok:	stmdb   sp!,{r4-r12,lr}
    443 	sub	r10,r3,#_armv4_AES_set_encrypt_key-AES_Te-1024	@ Te4
    444 
    445 	mov	r12,r0		@ inp
    446 	mov	lr,r1			@ bits
    447 	mov	r11,r2			@ key
    448 
    449 #if __ARM_ARCH__<7
    450 	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
    451 	ldrb	r4,[r12,#2]	@ manner...
    452 	ldrb	r5,[r12,#1]
    453 	ldrb	r6,[r12,#0]
    454 	orr	r0,r0,r4,lsl#8
    455 	ldrb	r1,[r12,#7]
    456 	orr	r0,r0,r5,lsl#16
    457 	ldrb	r4,[r12,#6]
    458 	orr	r0,r0,r6,lsl#24
    459 	ldrb	r5,[r12,#5]
    460 	ldrb	r6,[r12,#4]
    461 	orr	r1,r1,r4,lsl#8
    462 	ldrb	r2,[r12,#11]
    463 	orr	r1,r1,r5,lsl#16
    464 	ldrb	r4,[r12,#10]
    465 	orr	r1,r1,r6,lsl#24
    466 	ldrb	r5,[r12,#9]
    467 	ldrb	r6,[r12,#8]
    468 	orr	r2,r2,r4,lsl#8
    469 	ldrb	r3,[r12,#15]
    470 	orr	r2,r2,r5,lsl#16
    471 	ldrb	r4,[r12,#14]
    472 	orr	r2,r2,r6,lsl#24
    473 	ldrb	r5,[r12,#13]
    474 	ldrb	r6,[r12,#12]
    475 	orr	r3,r3,r4,lsl#8
    476 	str	r0,[r11],#16
    477 	orr	r3,r3,r5,lsl#16
    478 	str	r1,[r11,#-12]
    479 	orr	r3,r3,r6,lsl#24
    480 	str	r2,[r11,#-8]
    481 	str	r3,[r11,#-4]
    482 #else
    483 	ldr	r0,[r12,#0]
    484 	ldr	r1,[r12,#4]
    485 	ldr	r2,[r12,#8]
    486 	ldr	r3,[r12,#12]
    487 #ifdef __ARMEL__
    488 	rev	r0,r0
    489 	rev	r1,r1
    490 	rev	r2,r2
    491 	rev	r3,r3
    492 #endif
    493 	str	r0,[r11],#16
    494 	str	r1,[r11,#-12]
    495 	str	r2,[r11,#-8]
    496 	str	r3,[r11,#-4]
    497 #endif
    498 
    499 	teq	lr,#128
    500 	bne	.Lnot128
    501 	mov	r12,#10
    502 	str	r12,[r11,#240-16]
    503 	add	r6,r10,#256			@ rcon
    504 	mov	lr,#255
    505 
    506 .L128_loop:
    507 	and	r5,lr,r3,lsr#24
    508 	and	r7,lr,r3,lsr#16
    509 	ldrb	r5,[r10,r5]
    510 	and	r8,lr,r3,lsr#8
    511 	ldrb	r7,[r10,r7]
    512 	and	r9,lr,r3
    513 	ldrb	r8,[r10,r8]
    514 	orr	r5,r5,r7,lsl#24
    515 	ldrb	r9,[r10,r9]
    516 	orr	r5,r5,r8,lsl#16
    517 	ldr	r4,[r6],#4			@ rcon[i++]
    518 	orr	r5,r5,r9,lsl#8
    519 	eor	r5,r5,r4
    520 	eor	r0,r0,r5			@ rk[4]=rk[0]^...
    521 	eor	r1,r1,r0			@ rk[5]=rk[1]^rk[4]
    522 	str	r0,[r11],#16
    523 	eor	r2,r2,r1			@ rk[6]=rk[2]^rk[5]
    524 	str	r1,[r11,#-12]
    525 	eor	r3,r3,r2			@ rk[7]=rk[3]^rk[6]
    526 	str	r2,[r11,#-8]
    527 	subs	r12,r12,#1
    528 	str	r3,[r11,#-4]
    529 	bne	.L128_loop
    530 	sub	r2,r11,#176
    531 	b	.Ldone
    532 
    533 .Lnot128:
    534 #if __ARM_ARCH__<7
    535 	ldrb	r8,[r12,#19]
    536 	ldrb	r4,[r12,#18]
    537 	ldrb	r5,[r12,#17]
    538 	ldrb	r6,[r12,#16]
    539 	orr	r8,r8,r4,lsl#8
    540 	ldrb	r9,[r12,#23]
    541 	orr	r8,r8,r5,lsl#16
    542 	ldrb	r4,[r12,#22]
    543 	orr	r8,r8,r6,lsl#24
    544 	ldrb	r5,[r12,#21]
    545 	ldrb	r6,[r12,#20]
    546 	orr	r9,r9,r4,lsl#8
    547 	orr	r9,r9,r5,lsl#16
    548 	str	r8,[r11],#8
    549 	orr	r9,r9,r6,lsl#24
    550 	str	r9,[r11,#-4]
    551 #else
    552 	ldr	r8,[r12,#16]
    553 	ldr	r9,[r12,#20]
    554 #ifdef __ARMEL__
    555 	rev	r8,r8
    556 	rev	r9,r9
    557 #endif
    558 	str	r8,[r11],#8
    559 	str	r9,[r11,#-4]
    560 #endif
    561 
    562 	teq	lr,#192
    563 	bne	.Lnot192
    564 	mov	r12,#12
    565 	str	r12,[r11,#240-24]
    566 	add	r6,r10,#256			@ rcon
    567 	mov	lr,#255
    568 	mov	r12,#8
    569 
    570 .L192_loop:
    571 	and	r5,lr,r9,lsr#24
    572 	and	r7,lr,r9,lsr#16
    573 	ldrb	r5,[r10,r5]
    574 	and	r8,lr,r9,lsr#8
    575 	ldrb	r7,[r10,r7]
    576 	and	r9,lr,r9
    577 	ldrb	r8,[r10,r8]
    578 	orr	r5,r5,r7,lsl#24
    579 	ldrb	r9,[r10,r9]
    580 	orr	r5,r5,r8,lsl#16
    581 	ldr	r4,[r6],#4			@ rcon[i++]
    582 	orr	r5,r5,r9,lsl#8
    583 	eor	r9,r5,r4
    584 	eor	r0,r0,r9			@ rk[6]=rk[0]^...
    585 	eor	r1,r1,r0			@ rk[7]=rk[1]^rk[6]
    586 	str	r0,[r11],#24
    587 	eor	r2,r2,r1			@ rk[8]=rk[2]^rk[7]
    588 	str	r1,[r11,#-20]
    589 	eor	r3,r3,r2			@ rk[9]=rk[3]^rk[8]
    590 	str	r2,[r11,#-16]
    591 	subs	r12,r12,#1
    592 	str	r3,[r11,#-12]
    593 #if __ARM_ARCH__>=7
    594 	itt	eq				@ Thumb2 thing, sanity check in ARM
    595 #endif
    596 	subeq	r2,r11,#216
    597 	beq	.Ldone
    598 
    599 	ldr	r7,[r11,#-32]
    600 	ldr	r8,[r11,#-28]
    601 	eor	r7,r7,r3			@ rk[10]=rk[4]^rk[9]
    602 	eor	r9,r8,r7			@ rk[11]=rk[5]^rk[10]
    603 	str	r7,[r11,#-8]
    604 	str	r9,[r11,#-4]
    605 	b	.L192_loop
    606 
    607 .Lnot192:
    608 #if __ARM_ARCH__<7
    609 	ldrb	r8,[r12,#27]
    610 	ldrb	r4,[r12,#26]
    611 	ldrb	r5,[r12,#25]
    612 	ldrb	r6,[r12,#24]
    613 	orr	r8,r8,r4,lsl#8
    614 	ldrb	r9,[r12,#31]
    615 	orr	r8,r8,r5,lsl#16
    616 	ldrb	r4,[r12,#30]
    617 	orr	r8,r8,r6,lsl#24
    618 	ldrb	r5,[r12,#29]
    619 	ldrb	r6,[r12,#28]
    620 	orr	r9,r9,r4,lsl#8
    621 	orr	r9,r9,r5,lsl#16
    622 	str	r8,[r11],#8
    623 	orr	r9,r9,r6,lsl#24
    624 	str	r9,[r11,#-4]
    625 #else
    626 	ldr	r8,[r12,#24]
    627 	ldr	r9,[r12,#28]
    628 #ifdef __ARMEL__
    629 	rev	r8,r8
    630 	rev	r9,r9
    631 #endif
    632 	str	r8,[r11],#8
    633 	str	r9,[r11,#-4]
    634 #endif
    635 
    636 	mov	r12,#14
    637 	str	r12,[r11,#240-32]
    638 	add	r6,r10,#256			@ rcon
    639 	mov	lr,#255
    640 	mov	r12,#7
    641 
    642 .L256_loop:
    643 	and	r5,lr,r9,lsr#24
    644 	and	r7,lr,r9,lsr#16
    645 	ldrb	r5,[r10,r5]
    646 	and	r8,lr,r9,lsr#8
    647 	ldrb	r7,[r10,r7]
    648 	and	r9,lr,r9
    649 	ldrb	r8,[r10,r8]
    650 	orr	r5,r5,r7,lsl#24
    651 	ldrb	r9,[r10,r9]
    652 	orr	r5,r5,r8,lsl#16
    653 	ldr	r4,[r6],#4			@ rcon[i++]
    654 	orr	r5,r5,r9,lsl#8
    655 	eor	r9,r5,r4
    656 	eor	r0,r0,r9			@ rk[8]=rk[0]^...
    657 	eor	r1,r1,r0			@ rk[9]=rk[1]^rk[8]
    658 	str	r0,[r11],#32
    659 	eor	r2,r2,r1			@ rk[10]=rk[2]^rk[9]
    660 	str	r1,[r11,#-28]
    661 	eor	r3,r3,r2			@ rk[11]=rk[3]^rk[10]
    662 	str	r2,[r11,#-24]
    663 	subs	r12,r12,#1
    664 	str	r3,[r11,#-20]
    665 #if __ARM_ARCH__>=7
    666 	itt	eq				@ Thumb2 thing, sanity check in ARM
    667 #endif
    668 	subeq	r2,r11,#256
    669 	beq	.Ldone
    670 
    671 	and	r5,lr,r3
    672 	and	r7,lr,r3,lsr#8
    673 	ldrb	r5,[r10,r5]
    674 	and	r8,lr,r3,lsr#16
    675 	ldrb	r7,[r10,r7]
    676 	and	r9,lr,r3,lsr#24
    677 	ldrb	r8,[r10,r8]
    678 	orr	r5,r5,r7,lsl#8
    679 	ldrb	r9,[r10,r9]
    680 	orr	r5,r5,r8,lsl#16
    681 	ldr	r4,[r11,#-48]
    682 	orr	r5,r5,r9,lsl#24
    683 
    684 	ldr	r7,[r11,#-44]
    685 	ldr	r8,[r11,#-40]
    686 	eor	r4,r4,r5			@ rk[12]=rk[4]^...
    687 	ldr	r9,[r11,#-36]
    688 	eor	r7,r7,r4			@ rk[13]=rk[5]^rk[12]
    689 	str	r4,[r11,#-16]
    690 	eor	r8,r8,r7			@ rk[14]=rk[6]^rk[13]
    691 	str	r7,[r11,#-12]
    692 	eor	r9,r9,r8			@ rk[15]=rk[7]^rk[14]
    693 	str	r8,[r11,#-8]
    694 	str	r9,[r11,#-4]
    695 	b	.L256_loop
    696 
    697 .align	2
    698 .Ldone:	mov	r0,#0
    699 	ldmia   sp!,{r4-r12,lr}
    700 .Labrt:
    701 #if defined(__thumb2__) && __ARM_ARCH__>=7
    702 	.short	0x4770			@ .word	0xe12fff1e in Thumb2 encoding
    703 #else
    704 	tst	lr,#1
    705 	moveq	pc,lr			@ be binary compatible with V4, yet
    706 	.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
    707 #endif
    708 .size	asm_AES_set_encrypt_key,.-asm_AES_set_encrypt_key
    709 
    710 .global asm_AES_set_decrypt_key
    711 .hidden asm_AES_set_decrypt_key
    712 .type   asm_AES_set_decrypt_key,%function
    713 .align	5
    714 asm_AES_set_decrypt_key:
    715 	str	lr,[sp,#-4]!            @ push lr
    716 	bl	_armv4_AES_set_encrypt_key
    717 	teq	r0,#0
    718 	ldr	lr,[sp],#4              @ pop lr
    719 	bne	.Labrt
    720 
    721 	mov	r0,r2			@ asm_AES_set_encrypt_key preserves r2,
    722 	mov	r1,r2			@ which is AES_KEY *key
    723 	b	_armv4_AES_set_enc2dec_key
    724 .size	asm_AES_set_decrypt_key,.-asm_AES_set_decrypt_key
    725 
    726 @ void AES_set_enc2dec_key(const AES_KEY *inp,AES_KEY *out)
    727 .global	AES_set_enc2dec_key
    728 .hidden	AES_set_enc2dec_key
    729 .type	AES_set_enc2dec_key,%function
    730 .align	5
    731 AES_set_enc2dec_key:
    732 _armv4_AES_set_enc2dec_key:
    733 	stmdb   sp!,{r4-r12,lr}
    734 
    735 	ldr	r12,[r0,#240]
    736 	mov	r7,r0			@ input
    737 	add	r8,r0,r12,lsl#4
    738 	mov	r11,r1			@ ouput
    739 	add	r10,r1,r12,lsl#4
    740 	str	r12,[r1,#240]
    741 
    742 .Linv:	ldr	r0,[r7],#16
    743 	ldr	r1,[r7,#-12]
    744 	ldr	r2,[r7,#-8]
    745 	ldr	r3,[r7,#-4]
    746 	ldr	r4,[r8],#-16
    747 	ldr	r5,[r8,#16+4]
    748 	ldr	r6,[r8,#16+8]
    749 	ldr	r9,[r8,#16+12]
    750 	str	r0,[r10],#-16
    751 	str	r1,[r10,#16+4]
    752 	str	r2,[r10,#16+8]
    753 	str	r3,[r10,#16+12]
    754 	str	r4,[r11],#16
    755 	str	r5,[r11,#-12]
    756 	str	r6,[r11,#-8]
    757 	str	r9,[r11,#-4]
    758 	teq	r7,r8
    759 	bne	.Linv
    760 
    761 	ldr	r0,[r7]
    762 	ldr	r1,[r7,#4]
    763 	ldr	r2,[r7,#8]
    764 	ldr	r3,[r7,#12]
    765 	str	r0,[r11]
    766 	str	r1,[r11,#4]
    767 	str	r2,[r11,#8]
    768 	str	r3,[r11,#12]
    769 	sub	r11,r11,r12,lsl#3
    770 	ldr	r0,[r11,#16]!		@ prefetch tp1
    771 	mov	r7,#0x80
    772 	mov	r8,#0x1b
    773 	orr	r7,r7,#0x8000
    774 	orr	r8,r8,#0x1b00
    775 	orr	r7,r7,r7,lsl#16
    776 	orr	r8,r8,r8,lsl#16
    777 	sub	r12,r12,#1
    778 	mvn	r9,r7
    779 	mov	r12,r12,lsl#2	@ (rounds-1)*4
    780 
    781 .Lmix:	and	r4,r0,r7
    782 	and	r1,r0,r9
    783 	sub	r4,r4,r4,lsr#7
    784 	and	r4,r4,r8
    785 	eor	r1,r4,r1,lsl#1	@ tp2
    786 
    787 	and	r4,r1,r7
    788 	and	r2,r1,r9
    789 	sub	r4,r4,r4,lsr#7
    790 	and	r4,r4,r8
    791 	eor	r2,r4,r2,lsl#1	@ tp4
    792 
    793 	and	r4,r2,r7
    794 	and	r3,r2,r9
    795 	sub	r4,r4,r4,lsr#7
    796 	and	r4,r4,r8
    797 	eor	r3,r4,r3,lsl#1	@ tp8
    798 
    799 	eor	r4,r1,r2
    800 	eor	r5,r0,r3		@ tp9
    801 	eor	r4,r4,r3		@ tpe
    802 	eor	r4,r4,r1,ror#24
    803 	eor	r4,r4,r5,ror#24	@ ^= ROTATE(tpb=tp9^tp2,8)
    804 	eor	r4,r4,r2,ror#16
    805 	eor	r4,r4,r5,ror#16	@ ^= ROTATE(tpd=tp9^tp4,16)
    806 	eor	r4,r4,r5,ror#8	@ ^= ROTATE(tp9,24)
    807 
    808 	ldr	r0,[r11,#4]		@ prefetch tp1
    809 	str	r4,[r11],#4
    810 	subs	r12,r12,#1
    811 	bne	.Lmix
    812 
    813 	mov	r0,#0
    814 #if __ARM_ARCH__>=5
    815 	ldmia	sp!,{r4-r12,pc}
    816 #else
    817 	ldmia   sp!,{r4-r12,lr}
    818 	tst	lr,#1
    819 	moveq	pc,lr			@ be binary compatible with V4, yet
    820 	.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
    821 #endif
    822 .size	AES_set_enc2dec_key,.-AES_set_enc2dec_key
    823 
    824 .type	AES_Td,%object
    825 .align	5
    826 AES_Td:
    827 .word	0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
    828 .word	0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
    829 .word	0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
    830 .word	0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
    831 .word	0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
    832 .word	0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
    833 .word	0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
    834 .word	0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
    835 .word	0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
    836 .word	0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
    837 .word	0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
    838 .word	0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
    839 .word	0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
    840 .word	0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
    841 .word	0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
    842 .word	0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
    843 .word	0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
    844 .word	0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
    845 .word	0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
    846 .word	0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
    847 .word	0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
    848 .word	0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
    849 .word	0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
    850 .word	0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
    851 .word	0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
    852 .word	0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
    853 .word	0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
    854 .word	0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
    855 .word	0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
    856 .word	0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
    857 .word	0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
    858 .word	0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
    859 .word	0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
    860 .word	0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
    861 .word	0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
    862 .word	0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
    863 .word	0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
    864 .word	0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
    865 .word	0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
    866 .word	0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
    867 .word	0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
    868 .word	0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
    869 .word	0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
    870 .word	0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
    871 .word	0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
    872 .word	0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
    873 .word	0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
    874 .word	0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
    875 .word	0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
    876 .word	0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
    877 .word	0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
    878 .word	0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
    879 .word	0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
    880 .word	0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
    881 .word	0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
    882 .word	0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
    883 .word	0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
    884 .word	0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
    885 .word	0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
    886 .word	0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
    887 .word	0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
    888 .word	0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
    889 .word	0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
    890 .word	0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
    891 @ Td4[256]
    892 .byte	0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
    893 .byte	0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
    894 .byte	0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
    895 .byte	0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
    896 .byte	0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
    897 .byte	0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
    898 .byte	0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
    899 .byte	0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
    900 .byte	0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
    901 .byte	0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
    902 .byte	0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
    903 .byte	0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
    904 .byte	0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
    905 .byte	0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
    906 .byte	0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
    907 .byte	0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
    908 .byte	0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
    909 .byte	0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
    910 .byte	0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
    911 .byte	0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
    912 .byte	0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
    913 .byte	0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
    914 .byte	0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
    915 .byte	0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
    916 .byte	0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
    917 .byte	0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
    918 .byte	0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
    919 .byte	0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
    920 .byte	0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
    921 .byte	0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
    922 .byte	0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
    923 .byte	0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
    924 .size	AES_Td,.-AES_Td
    925 
    926 @ void asm_AES_decrypt(const unsigned char *in, unsigned char *out,
    927 @ 		       const AES_KEY *key) {
    928 .global asm_AES_decrypt
    929 .hidden asm_AES_decrypt
    930 .type   asm_AES_decrypt,%function
    931 .align	5
    932 asm_AES_decrypt:
    933 #if __ARM_ARCH__<7
    934 	sub	r3,pc,#8		@ asm_AES_decrypt
    935 #else
    936 	adr	r3,asm_AES_decrypt
    937 #endif
    938 	stmdb   sp!,{r1,r4-r12,lr}
    939 	mov	r12,r0		@ inp
    940 	mov	r11,r2
    941 	sub	r10,r3,#asm_AES_decrypt-AES_Td		@ Td
    942 #if __ARM_ARCH__<7
    943 	ldrb	r0,[r12,#3]	@ load input data in endian-neutral
    944 	ldrb	r4,[r12,#2]	@ manner...
    945 	ldrb	r5,[r12,#1]
    946 	ldrb	r6,[r12,#0]
    947 	orr	r0,r0,r4,lsl#8
    948 	ldrb	r1,[r12,#7]
    949 	orr	r0,r0,r5,lsl#16
    950 	ldrb	r4,[r12,#6]
    951 	orr	r0,r0,r6,lsl#24
    952 	ldrb	r5,[r12,#5]
    953 	ldrb	r6,[r12,#4]
    954 	orr	r1,r1,r4,lsl#8
    955 	ldrb	r2,[r12,#11]
    956 	orr	r1,r1,r5,lsl#16
    957 	ldrb	r4,[r12,#10]
    958 	orr	r1,r1,r6,lsl#24
    959 	ldrb	r5,[r12,#9]
    960 	ldrb	r6,[r12,#8]
    961 	orr	r2,r2,r4,lsl#8
    962 	ldrb	r3,[r12,#15]
    963 	orr	r2,r2,r5,lsl#16
    964 	ldrb	r4,[r12,#14]
    965 	orr	r2,r2,r6,lsl#24
    966 	ldrb	r5,[r12,#13]
    967 	ldrb	r6,[r12,#12]
    968 	orr	r3,r3,r4,lsl#8
    969 	orr	r3,r3,r5,lsl#16
    970 	orr	r3,r3,r6,lsl#24
    971 #else
    972 	ldr	r0,[r12,#0]
    973 	ldr	r1,[r12,#4]
    974 	ldr	r2,[r12,#8]
    975 	ldr	r3,[r12,#12]
    976 #ifdef __ARMEL__
    977 	rev	r0,r0
    978 	rev	r1,r1
    979 	rev	r2,r2
    980 	rev	r3,r3
    981 #endif
    982 #endif
    983 	bl	_armv4_AES_decrypt
    984 
    985 	ldr	r12,[sp],#4		@ pop out
    986 #if __ARM_ARCH__>=7
    987 #ifdef __ARMEL__
    988 	rev	r0,r0
    989 	rev	r1,r1
    990 	rev	r2,r2
    991 	rev	r3,r3
    992 #endif
    993 	str	r0,[r12,#0]
    994 	str	r1,[r12,#4]
    995 	str	r2,[r12,#8]
    996 	str	r3,[r12,#12]
    997 #else
    998 	mov	r4,r0,lsr#24		@ write output in endian-neutral
    999 	mov	r5,r0,lsr#16		@ manner...
   1000 	mov	r6,r0,lsr#8
   1001 	strb	r4,[r12,#0]
   1002 	strb	r5,[r12,#1]
   1003 	mov	r4,r1,lsr#24
   1004 	strb	r6,[r12,#2]
   1005 	mov	r5,r1,lsr#16
   1006 	strb	r0,[r12,#3]
   1007 	mov	r6,r1,lsr#8
   1008 	strb	r4,[r12,#4]
   1009 	strb	r5,[r12,#5]
   1010 	mov	r4,r2,lsr#24
   1011 	strb	r6,[r12,#6]
   1012 	mov	r5,r2,lsr#16
   1013 	strb	r1,[r12,#7]
   1014 	mov	r6,r2,lsr#8
   1015 	strb	r4,[r12,#8]
   1016 	strb	r5,[r12,#9]
   1017 	mov	r4,r3,lsr#24
   1018 	strb	r6,[r12,#10]
   1019 	mov	r5,r3,lsr#16
   1020 	strb	r2,[r12,#11]
   1021 	mov	r6,r3,lsr#8
   1022 	strb	r4,[r12,#12]
   1023 	strb	r5,[r12,#13]
   1024 	strb	r6,[r12,#14]
   1025 	strb	r3,[r12,#15]
   1026 #endif
   1027 #if __ARM_ARCH__>=5
   1028 	ldmia	sp!,{r4-r12,pc}
   1029 #else
   1030 	ldmia   sp!,{r4-r12,lr}
   1031 	tst	lr,#1
   1032 	moveq	pc,lr			@ be binary compatible with V4, yet
   1033 	.word	0xe12fff1e			@ interoperable with Thumb ISA:-)
   1034 #endif
   1035 .size	asm_AES_decrypt,.-asm_AES_decrypt
   1036 
   1037 .type   _armv4_AES_decrypt,%function
   1038 .align	2
   1039 _armv4_AES_decrypt:
   1040 	str	lr,[sp,#-4]!		@ push lr
   1041 	ldmia	r11!,{r4-r7}
   1042 	eor	r0,r0,r4
   1043 	ldr	r12,[r11,#240-16]
   1044 	eor	r1,r1,r5
   1045 	eor	r2,r2,r6
   1046 	eor	r3,r3,r7
   1047 	sub	r12,r12,#1
   1048 	mov	lr,#255
   1049 
   1050 	and	r7,lr,r0,lsr#16
   1051 	and	r8,lr,r0,lsr#8
   1052 	and	r9,lr,r0
   1053 	mov	r0,r0,lsr#24
   1054 .Ldec_loop:
   1055 	ldr	r4,[r10,r7,lsl#2]	@ Td1[s0>>16]
   1056 	and	r7,lr,r1		@ i0
   1057 	ldr	r5,[r10,r8,lsl#2]	@ Td2[s0>>8]
   1058 	and	r8,lr,r1,lsr#16
   1059 	ldr	r6,[r10,r9,lsl#2]	@ Td3[s0>>0]
   1060 	and	r9,lr,r1,lsr#8
   1061 	ldr	r0,[r10,r0,lsl#2]	@ Td0[s0>>24]
   1062 	mov	r1,r1,lsr#24
   1063 
   1064 	ldr	r7,[r10,r7,lsl#2]	@ Td3[s1>>0]
   1065 	ldr	r8,[r10,r8,lsl#2]	@ Td1[s1>>16]
   1066 	ldr	r9,[r10,r9,lsl#2]	@ Td2[s1>>8]
   1067 	eor	r0,r0,r7,ror#24
   1068 	ldr	r1,[r10,r1,lsl#2]	@ Td0[s1>>24]
   1069 	and	r7,lr,r2,lsr#8	@ i0
   1070 	eor	r5,r8,r5,ror#8
   1071 	and	r8,lr,r2		@ i1
   1072 	eor	r6,r9,r6,ror#8
   1073 	and	r9,lr,r2,lsr#16
   1074 	ldr	r7,[r10,r7,lsl#2]	@ Td2[s2>>8]
   1075 	eor	r1,r1,r4,ror#8
   1076 	ldr	r8,[r10,r8,lsl#2]	@ Td3[s2>>0]
   1077 	mov	r2,r2,lsr#24
   1078 
   1079 	ldr	r9,[r10,r9,lsl#2]	@ Td1[s2>>16]
   1080 	eor	r0,r0,r7,ror#16
   1081 	ldr	r2,[r10,r2,lsl#2]	@ Td0[s2>>24]
   1082 	and	r7,lr,r3,lsr#16	@ i0
   1083 	eor	r1,r1,r8,ror#24
   1084 	and	r8,lr,r3,lsr#8	@ i1
   1085 	eor	r6,r9,r6,ror#8
   1086 	and	r9,lr,r3		@ i2
   1087 	ldr	r7,[r10,r7,lsl#2]	@ Td1[s3>>16]
   1088 	eor	r2,r2,r5,ror#8
   1089 	ldr	r8,[r10,r8,lsl#2]	@ Td2[s3>>8]
   1090 	mov	r3,r3,lsr#24
   1091 
   1092 	ldr	r9,[r10,r9,lsl#2]	@ Td3[s3>>0]
   1093 	eor	r0,r0,r7,ror#8
   1094 	ldr	r7,[r11],#16
   1095 	eor	r1,r1,r8,ror#16
   1096 	ldr	r3,[r10,r3,lsl#2]	@ Td0[s3>>24]
   1097 	eor	r2,r2,r9,ror#24
   1098 
   1099 	ldr	r4,[r11,#-12]
   1100 	eor	r0,r0,r7
   1101 	ldr	r5,[r11,#-8]
   1102 	eor	r3,r3,r6,ror#8
   1103 	ldr	r6,[r11,#-4]
   1104 	and	r7,lr,r0,lsr#16
   1105 	eor	r1,r1,r4
   1106 	and	r8,lr,r0,lsr#8
   1107 	eor	r2,r2,r5
   1108 	and	r9,lr,r0
   1109 	eor	r3,r3,r6
   1110 	mov	r0,r0,lsr#24
   1111 
   1112 	subs	r12,r12,#1
   1113 	bne	.Ldec_loop
   1114 
   1115 	add	r10,r10,#1024
   1116 
   1117 	ldr	r5,[r10,#0]		@ prefetch Td4
   1118 	ldr	r6,[r10,#32]
   1119 	ldr	r4,[r10,#64]
   1120 	ldr	r5,[r10,#96]
   1121 	ldr	r6,[r10,#128]
   1122 	ldr	r4,[r10,#160]
   1123 	ldr	r5,[r10,#192]
   1124 	ldr	r6,[r10,#224]
   1125 
   1126 	ldrb	r0,[r10,r0]		@ Td4[s0>>24]
   1127 	ldrb	r4,[r10,r7]		@ Td4[s0>>16]
   1128 	and	r7,lr,r1		@ i0
   1129 	ldrb	r5,[r10,r8]		@ Td4[s0>>8]
   1130 	and	r8,lr,r1,lsr#16
   1131 	ldrb	r6,[r10,r9]		@ Td4[s0>>0]
   1132 	and	r9,lr,r1,lsr#8
   1133 
   1134 	add	r1,r10,r1,lsr#24
   1135 	ldrb	r7,[r10,r7]		@ Td4[s1>>0]
   1136 	ldrb	r1,[r1]		@ Td4[s1>>24]
   1137 	ldrb	r8,[r10,r8]		@ Td4[s1>>16]
   1138 	eor	r0,r7,r0,lsl#24
   1139 	ldrb	r9,[r10,r9]		@ Td4[s1>>8]
   1140 	eor	r1,r4,r1,lsl#8
   1141 	and	r7,lr,r2,lsr#8	@ i0
   1142 	eor	r5,r5,r8,lsl#8
   1143 	and	r8,lr,r2		@ i1
   1144 	ldrb	r7,[r10,r7]		@ Td4[s2>>8]
   1145 	eor	r6,r6,r9,lsl#8
   1146 	ldrb	r8,[r10,r8]		@ Td4[s2>>0]
   1147 	and	r9,lr,r2,lsr#16
   1148 
   1149 	add	r2,r10,r2,lsr#24
   1150 	ldrb	r2,[r2]		@ Td4[s2>>24]
   1151 	eor	r0,r0,r7,lsl#8
   1152 	ldrb	r9,[r10,r9]		@ Td4[s2>>16]
   1153 	eor	r1,r8,r1,lsl#16
   1154 	and	r7,lr,r3,lsr#16	@ i0
   1155 	eor	r2,r5,r2,lsl#16
   1156 	and	r8,lr,r3,lsr#8	@ i1
   1157 	ldrb	r7,[r10,r7]		@ Td4[s3>>16]
   1158 	eor	r6,r6,r9,lsl#16
   1159 	ldrb	r8,[r10,r8]		@ Td4[s3>>8]
   1160 	and	r9,lr,r3		@ i2
   1161 
   1162 	add	r3,r10,r3,lsr#24
   1163 	ldrb	r9,[r10,r9]		@ Td4[s3>>0]
   1164 	ldrb	r3,[r3]		@ Td4[s3>>24]
   1165 	eor	r0,r0,r7,lsl#16
   1166 	ldr	r7,[r11,#0]
   1167 	eor	r1,r1,r8,lsl#8
   1168 	ldr	r4,[r11,#4]
   1169 	eor	r2,r9,r2,lsl#8
   1170 	ldr	r5,[r11,#8]
   1171 	eor	r3,r6,r3,lsl#24
   1172 	ldr	r6,[r11,#12]
   1173 
   1174 	eor	r0,r0,r7
   1175 	eor	r1,r1,r4
   1176 	eor	r2,r2,r5
   1177 	eor	r3,r3,r6
   1178 
   1179 	sub	r10,r10,#1024
   1180 	ldr	pc,[sp],#4		@ pop and return
   1181 .size	_armv4_AES_decrypt,.-_armv4_AES_decrypt
   1182 .asciz	"AES for ARMv4, CRYPTOGAMS by <appro (at) openssl.org>"
   1183 .align	2
   1184 
   1185 #endif
   1186