diff options
author | djm <> | 2009-04-06 06:30:10 +0000 |
---|---|---|
committer | djm <> | 2009-04-06 06:30:10 +0000 |
commit | 2b6e09b39ef1d803b50ee024a06d1c250fde442d (patch) | |
tree | f116109c359f26a2b149bbc752be39c16099bae1 /src/lib/libcrypto/aes/asm | |
parent | a0fdc9ec41594852f67ec77dfad9cb06bacc4186 (diff) | |
download | openbsd-2b6e09b39ef1d803b50ee024a06d1c250fde442d.tar.gz openbsd-2b6e09b39ef1d803b50ee024a06d1c250fde442d.tar.bz2 openbsd-2b6e09b39ef1d803b50ee024a06d1c250fde442d.zip |
import of OpenSSL 0.9.8k
Diffstat (limited to 'src/lib/libcrypto/aes/asm')
-rw-r--r-- | src/lib/libcrypto/aes/asm/aes-armv4.pl | 1030 | ||||
-rw-r--r-- | src/lib/libcrypto/aes/asm/aes-ppc.pl | 1176 | ||||
-rw-r--r-- | src/lib/libcrypto/aes/asm/aes-s390x.pl | 1333 | ||||
-rwxr-xr-x | src/lib/libcrypto/aes/asm/aes-sparcv9.pl | 1181 |
4 files changed, 4720 insertions, 0 deletions
diff --git a/src/lib/libcrypto/aes/asm/aes-armv4.pl b/src/lib/libcrypto/aes/asm/aes-armv4.pl new file mode 100644 index 0000000000..15742c1ec5 --- /dev/null +++ b/src/lib/libcrypto/aes/asm/aes-armv4.pl | |||
@@ -0,0 +1,1030 @@ | |||
1 | #!/usr/bin/env perl | ||
2 | |||
3 | # ==================================================================== | ||
4 | # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL | ||
5 | # project. The module is, however, dual licensed under OpenSSL and | ||
6 | # CRYPTOGAMS licenses depending on where you obtain it. For further | ||
7 | # details see http://www.openssl.org/~appro/cryptogams/. | ||
8 | # ==================================================================== | ||
9 | |||
10 | # AES for ARMv4 | ||
11 | |||
12 | # January 2007. | ||
13 | # | ||
14 | # Code uses single 1K S-box and is >2 times faster than code generated | ||
15 | # by gcc-3.4.1. This is thanks to unique feature of ARMv4 ISA, which | ||
16 | # allows to merge logical or arithmetic operation with shift or rotate | ||
17 | # in one instruction and emit combined result every cycle. The module | ||
18 | # is endian-neutral. The performance is ~42 cycles/byte for 128-bit | ||
19 | # key. | ||
20 | |||
21 | # May 2007. | ||
22 | # | ||
23 | # AES_set_[en|de]crypt_key is added. | ||
24 | |||
25 | $s0="r0"; | ||
26 | $s1="r1"; | ||
27 | $s2="r2"; | ||
28 | $s3="r3"; | ||
29 | $t1="r4"; | ||
30 | $t2="r5"; | ||
31 | $t3="r6"; | ||
32 | $i1="r7"; | ||
33 | $i2="r8"; | ||
34 | $i3="r9"; | ||
35 | |||
36 | $tbl="r10"; | ||
37 | $key="r11"; | ||
38 | $rounds="r12"; | ||
39 | |||
40 | $code=<<___; | ||
41 | .text | ||
42 | .code 32 | ||
43 | |||
44 | .type AES_Te,%object | ||
45 | .align 5 | ||
46 | AES_Te: | ||
47 | .word 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d | ||
48 | .word 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554 | ||
49 | .word 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d | ||
50 | .word 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a | ||
51 | .word 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87 | ||
52 | .word 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b | ||
53 | .word 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea | ||
54 | .word 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b | ||
55 | .word 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a | ||
56 | .word 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f | ||
57 | .word 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108 | ||
58 | .word 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f | ||
59 | .word 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e | ||
60 | .word 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5 | ||
61 | .word 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d | ||
62 | .word 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f | ||
63 | .word 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e | ||
64 | .word 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb | ||
65 | .word 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce | ||
66 | .word 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497 | ||
67 | .word 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c | ||
68 | .word 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed | ||
69 | .word 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b | ||
70 | .word 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a | ||
71 | .word 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16 | ||
72 | .word 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594 | ||
73 | .word 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81 | ||
74 | .word 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3 | ||
75 | .word 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a | ||
76 | .word 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504 | ||
77 | .word 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163 | ||
78 | .word 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d | ||
79 | .word 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f | ||
80 | .word 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739 | ||
81 | .word 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47 | ||
82 | .word 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395 | ||
83 | .word 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f | ||
84 | .word 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883 | ||
85 | .word 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c | ||
86 | .word 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76 | ||
87 | .word 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e | ||
88 | .word 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4 | ||
89 | .word 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6 | ||
90 | .word 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b | ||
91 | .word 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7 | ||
92 | .word 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0 | ||
93 | .word 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25 | ||
94 | .word 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818 | ||
95 | .word 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72 | ||
96 | .word 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651 | ||
97 | .word 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21 | ||
98 | .word 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85 | ||
99 | .word 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa | ||
100 | .word 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12 | ||
101 | .word 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0 | ||
102 | .word 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9 | ||
103 | .word 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133 | ||
104 | .word 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7 | ||
105 | .word 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920 | ||
106 | .word 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a | ||
107 | .word 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17 | ||
108 | .word 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8 | ||
109 | .word 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11 | ||
110 | .word 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a | ||
111 | @ Te4[256] | ||
112 | .byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 | ||
113 | .byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76 | ||
114 | .byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0 | ||
115 | .byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0 | ||
116 | .byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc | ||
117 | .byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15 | ||
118 | .byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a | ||
119 | .byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75 | ||
120 | .byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0 | ||
121 | .byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84 | ||
122 | .byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b | ||
123 | .byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf | ||
124 | .byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85 | ||
125 | .byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8 | ||
126 | .byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5 | ||
127 | .byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2 | ||
128 | .byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17 | ||
129 | .byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73 | ||
130 | .byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88 | ||
131 | .byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb | ||
132 | .byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c | ||
133 | .byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79 | ||
134 | .byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9 | ||
135 | .byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08 | ||
136 | .byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6 | ||
137 | .byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a | ||
138 | .byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e | ||
139 | .byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e | ||
140 | .byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94 | ||
141 | .byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf | ||
142 | .byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68 | ||
143 | .byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 | ||
144 | @ rcon[] | ||
145 | .word 0x01000000, 0x02000000, 0x04000000, 0x08000000 | ||
146 | .word 0x10000000, 0x20000000, 0x40000000, 0x80000000 | ||
147 | .word 0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0 | ||
148 | .size AES_Te,.-AES_Te | ||
149 | |||
150 | @ void AES_encrypt(const unsigned char *in, unsigned char *out, | ||
151 | @ const AES_KEY *key) { | ||
152 | .global AES_encrypt | ||
153 | .type AES_encrypt,%function | ||
154 | .align 5 | ||
155 | AES_encrypt: | ||
156 | sub r3,pc,#8 @ AES_encrypt | ||
157 | stmdb sp!,{r1,r4-r12,lr} | ||
158 | mov $rounds,r0 @ inp | ||
159 | mov $key,r2 | ||
160 | sub $tbl,r3,#AES_encrypt-AES_Te @ Te | ||
161 | |||
162 | ldrb $s0,[$rounds,#3] @ load input data in endian-neutral | ||
163 | ldrb $t1,[$rounds,#2] @ manner... | ||
164 | ldrb $t2,[$rounds,#1] | ||
165 | ldrb $t3,[$rounds,#0] | ||
166 | orr $s0,$s0,$t1,lsl#8 | ||
167 | orr $s0,$s0,$t2,lsl#16 | ||
168 | orr $s0,$s0,$t3,lsl#24 | ||
169 | ldrb $s1,[$rounds,#7] | ||
170 | ldrb $t1,[$rounds,#6] | ||
171 | ldrb $t2,[$rounds,#5] | ||
172 | ldrb $t3,[$rounds,#4] | ||
173 | orr $s1,$s1,$t1,lsl#8 | ||
174 | orr $s1,$s1,$t2,lsl#16 | ||
175 | orr $s1,$s1,$t3,lsl#24 | ||
176 | ldrb $s2,[$rounds,#11] | ||
177 | ldrb $t1,[$rounds,#10] | ||
178 | ldrb $t2,[$rounds,#9] | ||
179 | ldrb $t3,[$rounds,#8] | ||
180 | orr $s2,$s2,$t1,lsl#8 | ||
181 | orr $s2,$s2,$t2,lsl#16 | ||
182 | orr $s2,$s2,$t3,lsl#24 | ||
183 | ldrb $s3,[$rounds,#15] | ||
184 | ldrb $t1,[$rounds,#14] | ||
185 | ldrb $t2,[$rounds,#13] | ||
186 | ldrb $t3,[$rounds,#12] | ||
187 | orr $s3,$s3,$t1,lsl#8 | ||
188 | orr $s3,$s3,$t2,lsl#16 | ||
189 | orr $s3,$s3,$t3,lsl#24 | ||
190 | |||
191 | bl _armv4_AES_encrypt | ||
192 | |||
193 | ldr $rounds,[sp],#4 @ pop out | ||
194 | mov $t1,$s0,lsr#24 @ write output in endian-neutral | ||
195 | mov $t2,$s0,lsr#16 @ manner... | ||
196 | mov $t3,$s0,lsr#8 | ||
197 | strb $t1,[$rounds,#0] | ||
198 | strb $t2,[$rounds,#1] | ||
199 | strb $t3,[$rounds,#2] | ||
200 | strb $s0,[$rounds,#3] | ||
201 | mov $t1,$s1,lsr#24 | ||
202 | mov $t2,$s1,lsr#16 | ||
203 | mov $t3,$s1,lsr#8 | ||
204 | strb $t1,[$rounds,#4] | ||
205 | strb $t2,[$rounds,#5] | ||
206 | strb $t3,[$rounds,#6] | ||
207 | strb $s1,[$rounds,#7] | ||
208 | mov $t1,$s2,lsr#24 | ||
209 | mov $t2,$s2,lsr#16 | ||
210 | mov $t3,$s2,lsr#8 | ||
211 | strb $t1,[$rounds,#8] | ||
212 | strb $t2,[$rounds,#9] | ||
213 | strb $t3,[$rounds,#10] | ||
214 | strb $s2,[$rounds,#11] | ||
215 | mov $t1,$s3,lsr#24 | ||
216 | mov $t2,$s3,lsr#16 | ||
217 | mov $t3,$s3,lsr#8 | ||
218 | strb $t1,[$rounds,#12] | ||
219 | strb $t2,[$rounds,#13] | ||
220 | strb $t3,[$rounds,#14] | ||
221 | strb $s3,[$rounds,#15] | ||
222 | |||
223 | ldmia sp!,{r4-r12,lr} | ||
224 | tst lr,#1 | ||
225 | moveq pc,lr @ be binary compatible with V4, yet | ||
226 | bx lr @ interoperable with Thumb ISA:-) | ||
227 | .size AES_encrypt,.-AES_encrypt | ||
228 | |||
229 | .type _armv4_AES_encrypt,%function | ||
230 | .align 2 | ||
231 | _armv4_AES_encrypt: | ||
232 | str lr,[sp,#-4]! @ push lr | ||
233 | ldr $t1,[$key],#16 | ||
234 | ldr $t2,[$key,#-12] | ||
235 | ldr $t3,[$key,#-8] | ||
236 | ldr $i1,[$key,#-4] | ||
237 | ldr $rounds,[$key,#240-16] | ||
238 | eor $s0,$s0,$t1 | ||
239 | eor $s1,$s1,$t2 | ||
240 | eor $s2,$s2,$t3 | ||
241 | eor $s3,$s3,$i1 | ||
242 | sub $rounds,$rounds,#1 | ||
243 | mov lr,#255 | ||
244 | |||
245 | .Lenc_loop: | ||
246 | and $i2,lr,$s0,lsr#8 | ||
247 | and $i3,lr,$s0,lsr#16 | ||
248 | and $i1,lr,$s0 | ||
249 | mov $s0,$s0,lsr#24 | ||
250 | ldr $t1,[$tbl,$i1,lsl#2] @ Te3[s0>>0] | ||
251 | ldr $s0,[$tbl,$s0,lsl#2] @ Te0[s0>>24] | ||
252 | ldr $t2,[$tbl,$i2,lsl#2] @ Te2[s0>>8] | ||
253 | ldr $t3,[$tbl,$i3,lsl#2] @ Te1[s0>>16] | ||
254 | |||
255 | and $i1,lr,$s1,lsr#16 @ i0 | ||
256 | and $i2,lr,$s1 | ||
257 | and $i3,lr,$s1,lsr#8 | ||
258 | mov $s1,$s1,lsr#24 | ||
259 | ldr $i1,[$tbl,$i1,lsl#2] @ Te1[s1>>16] | ||
260 | ldr $s1,[$tbl,$s1,lsl#2] @ Te0[s1>>24] | ||
261 | ldr $i2,[$tbl,$i2,lsl#2] @ Te3[s1>>0] | ||
262 | ldr $i3,[$tbl,$i3,lsl#2] @ Te2[s1>>8] | ||
263 | eor $s0,$s0,$i1,ror#8 | ||
264 | eor $s1,$s1,$t1,ror#24 | ||
265 | eor $t2,$t2,$i2,ror#8 | ||
266 | eor $t3,$t3,$i3,ror#8 | ||
267 | |||
268 | and $i1,lr,$s2,lsr#8 @ i0 | ||
269 | and $i2,lr,$s2,lsr#16 @ i1 | ||
270 | and $i3,lr,$s2 | ||
271 | mov $s2,$s2,lsr#24 | ||
272 | ldr $i1,[$tbl,$i1,lsl#2] @ Te2[s2>>8] | ||
273 | ldr $i2,[$tbl,$i2,lsl#2] @ Te1[s2>>16] | ||
274 | ldr $s2,[$tbl,$s2,lsl#2] @ Te0[s2>>24] | ||
275 | ldr $i3,[$tbl,$i3,lsl#2] @ Te3[s2>>0] | ||
276 | eor $s0,$s0,$i1,ror#16 | ||
277 | eor $s1,$s1,$i2,ror#8 | ||
278 | eor $s2,$s2,$t2,ror#16 | ||
279 | eor $t3,$t3,$i3,ror#16 | ||
280 | |||
281 | and $i1,lr,$s3 @ i0 | ||
282 | and $i2,lr,$s3,lsr#8 @ i1 | ||
283 | and $i3,lr,$s3,lsr#16 @ i2 | ||
284 | mov $s3,$s3,lsr#24 | ||
285 | ldr $i1,[$tbl,$i1,lsl#2] @ Te3[s3>>0] | ||
286 | ldr $i2,[$tbl,$i2,lsl#2] @ Te2[s3>>8] | ||
287 | ldr $i3,[$tbl,$i3,lsl#2] @ Te1[s3>>16] | ||
288 | ldr $s3,[$tbl,$s3,lsl#2] @ Te0[s3>>24] | ||
289 | eor $s0,$s0,$i1,ror#24 | ||
290 | eor $s1,$s1,$i2,ror#16 | ||
291 | eor $s2,$s2,$i3,ror#8 | ||
292 | eor $s3,$s3,$t3,ror#8 | ||
293 | |||
294 | ldr $t1,[$key],#16 | ||
295 | ldr $t2,[$key,#-12] | ||
296 | ldr $t3,[$key,#-8] | ||
297 | ldr $i1,[$key,#-4] | ||
298 | eor $s0,$s0,$t1 | ||
299 | eor $s1,$s1,$t2 | ||
300 | eor $s2,$s2,$t3 | ||
301 | eor $s3,$s3,$i1 | ||
302 | |||
303 | subs $rounds,$rounds,#1 | ||
304 | bne .Lenc_loop | ||
305 | |||
306 | add $tbl,$tbl,#2 | ||
307 | |||
308 | and $i1,lr,$s0 | ||
309 | and $i2,lr,$s0,lsr#8 | ||
310 | and $i3,lr,$s0,lsr#16 | ||
311 | mov $s0,$s0,lsr#24 | ||
312 | ldrb $t1,[$tbl,$i1,lsl#2] @ Te4[s0>>0] | ||
313 | ldrb $s0,[$tbl,$s0,lsl#2] @ Te4[s0>>24] | ||
314 | ldrb $t2,[$tbl,$i2,lsl#2] @ Te4[s0>>8] | ||
315 | ldrb $t3,[$tbl,$i3,lsl#2] @ Te4[s0>>16] | ||
316 | |||
317 | and $i1,lr,$s1,lsr#16 @ i0 | ||
318 | and $i2,lr,$s1 | ||
319 | and $i3,lr,$s1,lsr#8 | ||
320 | mov $s1,$s1,lsr#24 | ||
321 | ldrb $i1,[$tbl,$i1,lsl#2] @ Te4[s1>>16] | ||
322 | ldrb $s1,[$tbl,$s1,lsl#2] @ Te4[s1>>24] | ||
323 | ldrb $i2,[$tbl,$i2,lsl#2] @ Te4[s1>>0] | ||
324 | ldrb $i3,[$tbl,$i3,lsl#2] @ Te4[s1>>8] | ||
325 | eor $s0,$i1,$s0,lsl#8 | ||
326 | eor $s1,$t1,$s1,lsl#24 | ||
327 | eor $t2,$i2,$t2,lsl#8 | ||
328 | eor $t3,$i3,$t3,lsl#8 | ||
329 | |||
330 | and $i1,lr,$s2,lsr#8 @ i0 | ||
331 | and $i2,lr,$s2,lsr#16 @ i1 | ||
332 | and $i3,lr,$s2 | ||
333 | mov $s2,$s2,lsr#24 | ||
334 | ldrb $i1,[$tbl,$i1,lsl#2] @ Te4[s2>>8] | ||
335 | ldrb $i2,[$tbl,$i2,lsl#2] @ Te4[s2>>16] | ||
336 | ldrb $s2,[$tbl,$s2,lsl#2] @ Te4[s2>>24] | ||
337 | ldrb $i3,[$tbl,$i3,lsl#2] @ Te4[s2>>0] | ||
338 | eor $s0,$i1,$s0,lsl#8 | ||
339 | eor $s1,$s1,$i2,lsl#16 | ||
340 | eor $s2,$t2,$s2,lsl#24 | ||
341 | eor $t3,$i3,$t3,lsl#8 | ||
342 | |||
343 | and $i1,lr,$s3 @ i0 | ||
344 | and $i2,lr,$s3,lsr#8 @ i1 | ||
345 | and $i3,lr,$s3,lsr#16 @ i2 | ||
346 | mov $s3,$s3,lsr#24 | ||
347 | ldrb $i1,[$tbl,$i1,lsl#2] @ Te4[s3>>0] | ||
348 | ldrb $i2,[$tbl,$i2,lsl#2] @ Te4[s3>>8] | ||
349 | ldrb $i3,[$tbl,$i3,lsl#2] @ Te4[s3>>16] | ||
350 | ldrb $s3,[$tbl,$s3,lsl#2] @ Te4[s3>>24] | ||
351 | eor $s0,$i1,$s0,lsl#8 | ||
352 | eor $s1,$s1,$i2,lsl#8 | ||
353 | eor $s2,$s2,$i3,lsl#16 | ||
354 | eor $s3,$t3,$s3,lsl#24 | ||
355 | |||
356 | ldr lr,[sp],#4 @ pop lr | ||
357 | ldr $t1,[$key,#0] | ||
358 | ldr $t2,[$key,#4] | ||
359 | ldr $t3,[$key,#8] | ||
360 | ldr $i1,[$key,#12] | ||
361 | eor $s0,$s0,$t1 | ||
362 | eor $s1,$s1,$t2 | ||
363 | eor $s2,$s2,$t3 | ||
364 | eor $s3,$s3,$i1 | ||
365 | |||
366 | sub $tbl,$tbl,#2 | ||
367 | mov pc,lr @ return | ||
368 | .size _armv4_AES_encrypt,.-_armv4_AES_encrypt | ||
369 | |||
370 | .global AES_set_encrypt_key | ||
371 | .type AES_set_encrypt_key,%function | ||
372 | .align 5 | ||
373 | AES_set_encrypt_key: | ||
374 | sub r3,pc,#8 @ AES_set_encrypt_key | ||
375 | teq r0,#0 | ||
376 | moveq r0,#-1 | ||
377 | beq .Labrt | ||
378 | teq r2,#0 | ||
379 | moveq r0,#-1 | ||
380 | beq .Labrt | ||
381 | |||
382 | teq r1,#128 | ||
383 | beq .Lok | ||
384 | teq r1,#192 | ||
385 | beq .Lok | ||
386 | teq r1,#256 | ||
387 | movne r0,#-1 | ||
388 | bne .Labrt | ||
389 | |||
390 | .Lok: stmdb sp!,{r4-r12,lr} | ||
391 | sub $tbl,r3,#AES_set_encrypt_key-AES_Te-1024 @ Te4 | ||
392 | |||
393 | mov $rounds,r0 @ inp | ||
394 | mov lr,r1 @ bits | ||
395 | mov $key,r2 @ key | ||
396 | |||
397 | ldrb $s0,[$rounds,#3] @ load input data in endian-neutral | ||
398 | ldrb $t1,[$rounds,#2] @ manner... | ||
399 | ldrb $t2,[$rounds,#1] | ||
400 | ldrb $t3,[$rounds,#0] | ||
401 | orr $s0,$s0,$t1,lsl#8 | ||
402 | orr $s0,$s0,$t2,lsl#16 | ||
403 | orr $s0,$s0,$t3,lsl#24 | ||
404 | ldrb $s1,[$rounds,#7] | ||
405 | ldrb $t1,[$rounds,#6] | ||
406 | ldrb $t2,[$rounds,#5] | ||
407 | ldrb $t3,[$rounds,#4] | ||
408 | orr $s1,$s1,$t1,lsl#8 | ||
409 | orr $s1,$s1,$t2,lsl#16 | ||
410 | orr $s1,$s1,$t3,lsl#24 | ||
411 | ldrb $s2,[$rounds,#11] | ||
412 | ldrb $t1,[$rounds,#10] | ||
413 | ldrb $t2,[$rounds,#9] | ||
414 | ldrb $t3,[$rounds,#8] | ||
415 | orr $s2,$s2,$t1,lsl#8 | ||
416 | orr $s2,$s2,$t2,lsl#16 | ||
417 | orr $s2,$s2,$t3,lsl#24 | ||
418 | ldrb $s3,[$rounds,#15] | ||
419 | ldrb $t1,[$rounds,#14] | ||
420 | ldrb $t2,[$rounds,#13] | ||
421 | ldrb $t3,[$rounds,#12] | ||
422 | orr $s3,$s3,$t1,lsl#8 | ||
423 | orr $s3,$s3,$t2,lsl#16 | ||
424 | orr $s3,$s3,$t3,lsl#24 | ||
425 | str $s0,[$key],#16 | ||
426 | str $s1,[$key,#-12] | ||
427 | str $s2,[$key,#-8] | ||
428 | str $s3,[$key,#-4] | ||
429 | |||
430 | teq lr,#128 | ||
431 | bne .Lnot128 | ||
432 | mov $rounds,#10 | ||
433 | str $rounds,[$key,#240-16] | ||
434 | add $t3,$tbl,#256 @ rcon | ||
435 | mov lr,#255 | ||
436 | |||
437 | .L128_loop: | ||
438 | and $t2,lr,$s3,lsr#24 | ||
439 | and $i1,lr,$s3,lsr#16 | ||
440 | and $i2,lr,$s3,lsr#8 | ||
441 | and $i3,lr,$s3 | ||
442 | ldrb $t2,[$tbl,$t2] | ||
443 | ldrb $i1,[$tbl,$i1] | ||
444 | ldrb $i2,[$tbl,$i2] | ||
445 | ldrb $i3,[$tbl,$i3] | ||
446 | ldr $t1,[$t3],#4 @ rcon[i++] | ||
447 | orr $t2,$t2,$i1,lsl#24 | ||
448 | orr $t2,$t2,$i2,lsl#16 | ||
449 | orr $t2,$t2,$i3,lsl#8 | ||
450 | eor $t2,$t2,$t1 | ||
451 | eor $s0,$s0,$t2 @ rk[4]=rk[0]^... | ||
452 | eor $s1,$s1,$s0 @ rk[5]=rk[1]^rk[4] | ||
453 | eor $s2,$s2,$s1 @ rk[6]=rk[2]^rk[5] | ||
454 | eor $s3,$s3,$s2 @ rk[7]=rk[3]^rk[6] | ||
455 | str $s0,[$key],#16 | ||
456 | str $s1,[$key,#-12] | ||
457 | str $s2,[$key,#-8] | ||
458 | str $s3,[$key,#-4] | ||
459 | |||
460 | subs $rounds,$rounds,#1 | ||
461 | bne .L128_loop | ||
462 | sub r2,$key,#176 | ||
463 | b .Ldone | ||
464 | |||
465 | .Lnot128: | ||
466 | ldrb $i2,[$rounds,#19] | ||
467 | ldrb $t1,[$rounds,#18] | ||
468 | ldrb $t2,[$rounds,#17] | ||
469 | ldrb $t3,[$rounds,#16] | ||
470 | orr $i2,$i2,$t1,lsl#8 | ||
471 | orr $i2,$i2,$t2,lsl#16 | ||
472 | orr $i2,$i2,$t3,lsl#24 | ||
473 | ldrb $i3,[$rounds,#23] | ||
474 | ldrb $t1,[$rounds,#22] | ||
475 | ldrb $t2,[$rounds,#21] | ||
476 | ldrb $t3,[$rounds,#20] | ||
477 | orr $i3,$i3,$t1,lsl#8 | ||
478 | orr $i3,$i3,$t2,lsl#16 | ||
479 | orr $i3,$i3,$t3,lsl#24 | ||
480 | str $i2,[$key],#8 | ||
481 | str $i3,[$key,#-4] | ||
482 | |||
483 | teq lr,#192 | ||
484 | bne .Lnot192 | ||
485 | mov $rounds,#12 | ||
486 | str $rounds,[$key,#240-24] | ||
487 | add $t3,$tbl,#256 @ rcon | ||
488 | mov lr,#255 | ||
489 | mov $rounds,#8 | ||
490 | |||
491 | .L192_loop: | ||
492 | and $t2,lr,$i3,lsr#24 | ||
493 | and $i1,lr,$i3,lsr#16 | ||
494 | and $i2,lr,$i3,lsr#8 | ||
495 | and $i3,lr,$i3 | ||
496 | ldrb $t2,[$tbl,$t2] | ||
497 | ldrb $i1,[$tbl,$i1] | ||
498 | ldrb $i2,[$tbl,$i2] | ||
499 | ldrb $i3,[$tbl,$i3] | ||
500 | ldr $t1,[$t3],#4 @ rcon[i++] | ||
501 | orr $t2,$t2,$i1,lsl#24 | ||
502 | orr $t2,$t2,$i2,lsl#16 | ||
503 | orr $t2,$t2,$i3,lsl#8 | ||
504 | eor $i3,$t2,$t1 | ||
505 | eor $s0,$s0,$i3 @ rk[6]=rk[0]^... | ||
506 | eor $s1,$s1,$s0 @ rk[7]=rk[1]^rk[6] | ||
507 | eor $s2,$s2,$s1 @ rk[8]=rk[2]^rk[7] | ||
508 | eor $s3,$s3,$s2 @ rk[9]=rk[3]^rk[8] | ||
509 | str $s0,[$key],#24 | ||
510 | str $s1,[$key,#-20] | ||
511 | str $s2,[$key,#-16] | ||
512 | str $s3,[$key,#-12] | ||
513 | |||
514 | subs $rounds,$rounds,#1 | ||
515 | subeq r2,$key,#216 | ||
516 | beq .Ldone | ||
517 | |||
518 | ldr $i1,[$key,#-32] | ||
519 | ldr $i2,[$key,#-28] | ||
520 | eor $i1,$i1,$s3 @ rk[10]=rk[4]^rk[9] | ||
521 | eor $i3,$i2,$i1 @ rk[11]=rk[5]^rk[10] | ||
522 | str $i1,[$key,#-8] | ||
523 | str $i3,[$key,#-4] | ||
524 | b .L192_loop | ||
525 | |||
526 | .Lnot192: | ||
527 | ldrb $i2,[$rounds,#27] | ||
528 | ldrb $t1,[$rounds,#26] | ||
529 | ldrb $t2,[$rounds,#25] | ||
530 | ldrb $t3,[$rounds,#24] | ||
531 | orr $i2,$i2,$t1,lsl#8 | ||
532 | orr $i2,$i2,$t2,lsl#16 | ||
533 | orr $i2,$i2,$t3,lsl#24 | ||
534 | ldrb $i3,[$rounds,#31] | ||
535 | ldrb $t1,[$rounds,#30] | ||
536 | ldrb $t2,[$rounds,#29] | ||
537 | ldrb $t3,[$rounds,#28] | ||
538 | orr $i3,$i3,$t1,lsl#8 | ||
539 | orr $i3,$i3,$t2,lsl#16 | ||
540 | orr $i3,$i3,$t3,lsl#24 | ||
541 | str $i2,[$key],#8 | ||
542 | str $i3,[$key,#-4] | ||
543 | |||
544 | mov $rounds,#14 | ||
545 | str $rounds,[$key,#240-32] | ||
546 | add $t3,$tbl,#256 @ rcon | ||
547 | mov lr,#255 | ||
548 | mov $rounds,#7 | ||
549 | |||
550 | .L256_loop: | ||
551 | and $t2,lr,$i3,lsr#24 | ||
552 | and $i1,lr,$i3,lsr#16 | ||
553 | and $i2,lr,$i3,lsr#8 | ||
554 | and $i3,lr,$i3 | ||
555 | ldrb $t2,[$tbl,$t2] | ||
556 | ldrb $i1,[$tbl,$i1] | ||
557 | ldrb $i2,[$tbl,$i2] | ||
558 | ldrb $i3,[$tbl,$i3] | ||
559 | ldr $t1,[$t3],#4 @ rcon[i++] | ||
560 | orr $t2,$t2,$i1,lsl#24 | ||
561 | orr $t2,$t2,$i2,lsl#16 | ||
562 | orr $t2,$t2,$i3,lsl#8 | ||
563 | eor $i3,$t2,$t1 | ||
564 | eor $s0,$s0,$i3 @ rk[8]=rk[0]^... | ||
565 | eor $s1,$s1,$s0 @ rk[9]=rk[1]^rk[8] | ||
566 | eor $s2,$s2,$s1 @ rk[10]=rk[2]^rk[9] | ||
567 | eor $s3,$s3,$s2 @ rk[11]=rk[3]^rk[10] | ||
568 | str $s0,[$key],#32 | ||
569 | str $s1,[$key,#-28] | ||
570 | str $s2,[$key,#-24] | ||
571 | str $s3,[$key,#-20] | ||
572 | |||
573 | subs $rounds,$rounds,#1 | ||
574 | subeq r2,$key,#256 | ||
575 | beq .Ldone | ||
576 | |||
577 | and $t2,lr,$s3 | ||
578 | and $i1,lr,$s3,lsr#8 | ||
579 | and $i2,lr,$s3,lsr#16 | ||
580 | and $i3,lr,$s3,lsr#24 | ||
581 | ldrb $t2,[$tbl,$t2] | ||
582 | ldrb $i1,[$tbl,$i1] | ||
583 | ldrb $i2,[$tbl,$i2] | ||
584 | ldrb $i3,[$tbl,$i3] | ||
585 | orr $t2,$t2,$i1,lsl#8 | ||
586 | orr $t2,$t2,$i2,lsl#16 | ||
587 | orr $t2,$t2,$i3,lsl#24 | ||
588 | |||
589 | ldr $t1,[$key,#-48] | ||
590 | ldr $i1,[$key,#-44] | ||
591 | ldr $i2,[$key,#-40] | ||
592 | ldr $i3,[$key,#-36] | ||
593 | eor $t1,$t1,$t2 @ rk[12]=rk[4]^... | ||
594 | eor $i1,$i1,$t1 @ rk[13]=rk[5]^rk[12] | ||
595 | eor $i2,$i2,$i1 @ rk[14]=rk[6]^rk[13] | ||
596 | eor $i3,$i3,$i2 @ rk[15]=rk[7]^rk[14] | ||
597 | str $t1,[$key,#-16] | ||
598 | str $i1,[$key,#-12] | ||
599 | str $i2,[$key,#-8] | ||
600 | str $i3,[$key,#-4] | ||
601 | b .L256_loop | ||
602 | |||
603 | .Ldone: mov r0,#0 | ||
604 | ldmia sp!,{r4-r12,lr} | ||
605 | .Labrt: tst lr,#1 | ||
606 | moveq pc,lr @ be binary compatible with V4, yet | ||
607 | bx lr @ interoperable with Thumb ISA:-) | ||
608 | .size AES_set_encrypt_key,.-AES_set_encrypt_key | ||
609 | |||
610 | .global AES_set_decrypt_key | ||
611 | .type AES_set_decrypt_key,%function | ||
612 | .align 5 | ||
613 | AES_set_decrypt_key: | ||
614 | str lr,[sp,#-4]! @ push lr | ||
615 | bl AES_set_encrypt_key | ||
616 | teq r0,#0 | ||
617 | ldrne lr,[sp],#4 @ pop lr | ||
618 | bne .Labrt | ||
619 | |||
620 | stmdb sp!,{r4-r12} | ||
621 | |||
622 | ldr $rounds,[r2,#240] @ AES_set_encrypt_key preserves r2, | ||
623 | mov $key,r2 @ which is AES_KEY *key | ||
624 | mov $i1,r2 | ||
625 | add $i2,r2,$rounds,lsl#4 | ||
626 | |||
627 | .Linv: ldr $s0,[$i1] | ||
628 | ldr $s1,[$i1,#4] | ||
629 | ldr $s2,[$i1,#8] | ||
630 | ldr $s3,[$i1,#12] | ||
631 | ldr $t1,[$i2] | ||
632 | ldr $t2,[$i2,#4] | ||
633 | ldr $t3,[$i2,#8] | ||
634 | ldr $i3,[$i2,#12] | ||
635 | str $s0,[$i2],#-16 | ||
636 | str $s1,[$i2,#16+4] | ||
637 | str $s2,[$i2,#16+8] | ||
638 | str $s3,[$i2,#16+12] | ||
639 | str $t1,[$i1],#16 | ||
640 | str $t2,[$i1,#-12] | ||
641 | str $t3,[$i1,#-8] | ||
642 | str $i3,[$i1,#-4] | ||
643 | teq $i1,$i2 | ||
644 | bne .Linv | ||
645 | ___ | ||
646 | $mask80=$i1; | ||
647 | $mask1b=$i2; | ||
648 | $mask7f=$i3; | ||
649 | $code.=<<___; | ||
650 | ldr $s0,[$key,#16]! @ prefetch tp1 | ||
651 | mov $mask80,#0x80 | ||
652 | mov $mask1b,#0x1b | ||
653 | orr $mask80,$mask80,#0x8000 | ||
654 | orr $mask1b,$mask1b,#0x1b00 | ||
655 | orr $mask80,$mask80,$mask80,lsl#16 | ||
656 | orr $mask1b,$mask1b,$mask1b,lsl#16 | ||
657 | sub $rounds,$rounds,#1 | ||
658 | mvn $mask7f,$mask80 | ||
659 | mov $rounds,$rounds,lsl#2 @ (rounds-1)*4 | ||
660 | |||
661 | .Lmix: and $t1,$s0,$mask80 | ||
662 | and $s1,$s0,$mask7f | ||
663 | sub $t1,$t1,$t1,lsr#7 | ||
664 | and $t1,$t1,$mask1b | ||
665 | eor $s1,$t1,$s1,lsl#1 @ tp2 | ||
666 | |||
667 | and $t1,$s1,$mask80 | ||
668 | and $s2,$s1,$mask7f | ||
669 | sub $t1,$t1,$t1,lsr#7 | ||
670 | and $t1,$t1,$mask1b | ||
671 | eor $s2,$t1,$s2,lsl#1 @ tp4 | ||
672 | |||
673 | and $t1,$s2,$mask80 | ||
674 | and $s3,$s2,$mask7f | ||
675 | sub $t1,$t1,$t1,lsr#7 | ||
676 | and $t1,$t1,$mask1b | ||
677 | eor $s3,$t1,$s3,lsl#1 @ tp8 | ||
678 | |||
679 | eor $t1,$s1,$s2 | ||
680 | eor $t2,$s0,$s3 @ tp9 | ||
681 | eor $t1,$t1,$s3 @ tpe | ||
682 | eor $t1,$t1,$s1,ror#24 | ||
683 | eor $t1,$t1,$t2,ror#24 @ ^= ROTATE(tpb=tp9^tp2,8) | ||
684 | eor $t1,$t1,$s2,ror#16 | ||
685 | eor $t1,$t1,$t2,ror#16 @ ^= ROTATE(tpd=tp9^tp4,16) | ||
686 | eor $t1,$t1,$t2,ror#8 @ ^= ROTATE(tp9,24) | ||
687 | |||
688 | ldr $s0,[$key,#4] @ prefetch tp1 | ||
689 | str $t1,[$key],#4 | ||
690 | subs $rounds,$rounds,#1 | ||
691 | bne .Lmix | ||
692 | |||
693 | mov r0,#0 | ||
694 | ldmia sp!,{r4-r12,lr} | ||
695 | tst lr,#1 | ||
696 | moveq pc,lr @ be binary compatible with V4, yet | ||
697 | bx lr @ interoperable with Thumb ISA:-) | ||
698 | .size AES_set_decrypt_key,.-AES_set_decrypt_key | ||
699 | |||
700 | .type AES_Td,%object | ||
701 | .align 5 | ||
702 | AES_Td: | ||
703 | .word 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96 | ||
704 | .word 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393 | ||
705 | .word 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25 | ||
706 | .word 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f | ||
707 | .word 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1 | ||
708 | .word 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6 | ||
709 | .word 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da | ||
710 | .word 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844 | ||
711 | .word 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd | ||
712 | .word 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4 | ||
713 | .word 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45 | ||
714 | .word 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94 | ||
715 | .word 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7 | ||
716 | .word 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a | ||
717 | .word 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5 | ||
718 | .word 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c | ||
719 | .word 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1 | ||
720 | .word 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a | ||
721 | .word 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75 | ||
722 | .word 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051 | ||
723 | .word 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46 | ||
724 | .word 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff | ||
725 | .word 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77 | ||
726 | .word 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb | ||
727 | .word 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000 | ||
728 | .word 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e | ||
729 | .word 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927 | ||
730 | .word 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a | ||
731 | .word 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e | ||
732 | .word 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16 | ||
733 | .word 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d | ||
734 | .word 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8 | ||
735 | .word 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd | ||
736 | .word 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34 | ||
737 | .word 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163 | ||
738 | .word 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120 | ||
739 | .word 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d | ||
740 | .word 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0 | ||
741 | .word 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422 | ||
742 | .word 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef | ||
743 | .word 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36 | ||
744 | .word 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4 | ||
745 | .word 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662 | ||
746 | .word 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5 | ||
747 | .word 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3 | ||
748 | .word 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b | ||
749 | .word 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8 | ||
750 | .word 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6 | ||
751 | .word 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6 | ||
752 | .word 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0 | ||
753 | .word 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815 | ||
754 | .word 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f | ||
755 | .word 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df | ||
756 | .word 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f | ||
757 | .word 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e | ||
758 | .word 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713 | ||
759 | .word 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89 | ||
760 | .word 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c | ||
761 | .word 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf | ||
762 | .word 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86 | ||
763 | .word 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f | ||
764 | .word 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541 | ||
765 | .word 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190 | ||
766 | .word 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742 | ||
767 | @ Td4[256] | ||
768 | .byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 | ||
769 | .byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb | ||
770 | .byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87 | ||
771 | .byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb | ||
772 | .byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d | ||
773 | .byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e | ||
774 | .byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2 | ||
775 | .byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25 | ||
776 | .byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16 | ||
777 | .byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92 | ||
778 | .byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda | ||
779 | .byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84 | ||
780 | .byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a | ||
781 | .byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06 | ||
782 | .byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02 | ||
783 | .byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b | ||
784 | .byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea | ||
785 | .byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73 | ||
786 | .byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85 | ||
787 | .byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e | ||
788 | .byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89 | ||
789 | .byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b | ||
790 | .byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20 | ||
791 | .byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4 | ||
792 | .byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31 | ||
793 | .byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f | ||
794 | .byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d | ||
795 | .byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef | ||
796 | .byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0 | ||
797 | .byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61 | ||
798 | .byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26 | ||
799 | .byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d | ||
800 | .size AES_Td,.-AES_Td | ||
801 | |||
802 | @ void AES_decrypt(const unsigned char *in, unsigned char *out, | ||
803 | @ const AES_KEY *key) { | ||
804 | .global AES_decrypt | ||
805 | .type AES_decrypt,%function | ||
806 | .align 5 | ||
807 | AES_decrypt: | ||
808 | sub r3,pc,#8 @ AES_decrypt | ||
809 | stmdb sp!,{r1,r4-r12,lr} | ||
810 | mov $rounds,r0 @ inp | ||
811 | mov $key,r2 | ||
812 | sub $tbl,r3,#AES_decrypt-AES_Td @ Td | ||
813 | |||
814 | ldrb $s0,[$rounds,#3] @ load input data in endian-neutral | ||
815 | ldrb $t1,[$rounds,#2] @ manner... | ||
816 | ldrb $t2,[$rounds,#1] | ||
817 | ldrb $t3,[$rounds,#0] | ||
818 | orr $s0,$s0,$t1,lsl#8 | ||
819 | orr $s0,$s0,$t2,lsl#16 | ||
820 | orr $s0,$s0,$t3,lsl#24 | ||
821 | ldrb $s1,[$rounds,#7] | ||
822 | ldrb $t1,[$rounds,#6] | ||
823 | ldrb $t2,[$rounds,#5] | ||
824 | ldrb $t3,[$rounds,#4] | ||
825 | orr $s1,$s1,$t1,lsl#8 | ||
826 | orr $s1,$s1,$t2,lsl#16 | ||
827 | orr $s1,$s1,$t3,lsl#24 | ||
828 | ldrb $s2,[$rounds,#11] | ||
829 | ldrb $t1,[$rounds,#10] | ||
830 | ldrb $t2,[$rounds,#9] | ||
831 | ldrb $t3,[$rounds,#8] | ||
832 | orr $s2,$s2,$t1,lsl#8 | ||
833 | orr $s2,$s2,$t2,lsl#16 | ||
834 | orr $s2,$s2,$t3,lsl#24 | ||
835 | ldrb $s3,[$rounds,#15] | ||
836 | ldrb $t1,[$rounds,#14] | ||
837 | ldrb $t2,[$rounds,#13] | ||
838 | ldrb $t3,[$rounds,#12] | ||
839 | orr $s3,$s3,$t1,lsl#8 | ||
840 | orr $s3,$s3,$t2,lsl#16 | ||
841 | orr $s3,$s3,$t3,lsl#24 | ||
842 | |||
843 | bl _armv4_AES_decrypt | ||
844 | |||
845 | ldr $rounds,[sp],#4 @ pop out | ||
846 | mov $t1,$s0,lsr#24 @ write output in endian-neutral | ||
847 | mov $t2,$s0,lsr#16 @ manner... | ||
848 | mov $t3,$s0,lsr#8 | ||
849 | strb $t1,[$rounds,#0] | ||
850 | strb $t2,[$rounds,#1] | ||
851 | strb $t3,[$rounds,#2] | ||
852 | strb $s0,[$rounds,#3] | ||
853 | mov $t1,$s1,lsr#24 | ||
854 | mov $t2,$s1,lsr#16 | ||
855 | mov $t3,$s1,lsr#8 | ||
856 | strb $t1,[$rounds,#4] | ||
857 | strb $t2,[$rounds,#5] | ||
858 | strb $t3,[$rounds,#6] | ||
859 | strb $s1,[$rounds,#7] | ||
860 | mov $t1,$s2,lsr#24 | ||
861 | mov $t2,$s2,lsr#16 | ||
862 | mov $t3,$s2,lsr#8 | ||
863 | strb $t1,[$rounds,#8] | ||
864 | strb $t2,[$rounds,#9] | ||
865 | strb $t3,[$rounds,#10] | ||
866 | strb $s2,[$rounds,#11] | ||
867 | mov $t1,$s3,lsr#24 | ||
868 | mov $t2,$s3,lsr#16 | ||
869 | mov $t3,$s3,lsr#8 | ||
870 | strb $t1,[$rounds,#12] | ||
871 | strb $t2,[$rounds,#13] | ||
872 | strb $t3,[$rounds,#14] | ||
873 | strb $s3,[$rounds,#15] | ||
874 | |||
875 | ldmia sp!,{r4-r12,lr} | ||
876 | tst lr,#1 | ||
877 | moveq pc,lr @ be binary compatible with V4, yet | ||
878 | bx lr @ interoperable with Thumb ISA:-) | ||
879 | .size AES_decrypt,.-AES_decrypt | ||
880 | |||
881 | .type _armv4_AES_decrypt,%function | ||
882 | .align 2 | ||
883 | _armv4_AES_decrypt: | ||
884 | str lr,[sp,#-4]! @ push lr | ||
885 | ldr $t1,[$key],#16 | ||
886 | ldr $t2,[$key,#-12] | ||
887 | ldr $t3,[$key,#-8] | ||
888 | ldr $i1,[$key,#-4] | ||
889 | ldr $rounds,[$key,#240-16] | ||
890 | eor $s0,$s0,$t1 | ||
891 | eor $s1,$s1,$t2 | ||
892 | eor $s2,$s2,$t3 | ||
893 | eor $s3,$s3,$i1 | ||
894 | sub $rounds,$rounds,#1 | ||
895 | mov lr,#255 | ||
896 | |||
897 | .Ldec_loop: | ||
898 | and $i1,lr,$s0,lsr#16 | ||
899 | and $i2,lr,$s0,lsr#8 | ||
900 | and $i3,lr,$s0 | ||
901 | mov $s0,$s0,lsr#24 | ||
902 | ldr $t1,[$tbl,$i1,lsl#2] @ Td1[s0>>16] | ||
903 | ldr $s0,[$tbl,$s0,lsl#2] @ Td0[s0>>24] | ||
904 | ldr $t2,[$tbl,$i2,lsl#2] @ Td2[s0>>8] | ||
905 | ldr $t3,[$tbl,$i3,lsl#2] @ Td3[s0>>0] | ||
906 | |||
907 | and $i1,lr,$s1 @ i0 | ||
908 | and $i2,lr,$s1,lsr#16 | ||
909 | and $i3,lr,$s1,lsr#8 | ||
910 | mov $s1,$s1,lsr#24 | ||
911 | ldr $i1,[$tbl,$i1,lsl#2] @ Td3[s1>>0] | ||
912 | ldr $s1,[$tbl,$s1,lsl#2] @ Td0[s1>>24] | ||
913 | ldr $i2,[$tbl,$i2,lsl#2] @ Td1[s1>>16] | ||
914 | ldr $i3,[$tbl,$i3,lsl#2] @ Td2[s1>>8] | ||
915 | eor $s0,$s0,$i1,ror#24 | ||
916 | eor $s1,$s1,$t1,ror#8 | ||
917 | eor $t2,$i2,$t2,ror#8 | ||
918 | eor $t3,$i3,$t3,ror#8 | ||
919 | |||
920 | and $i1,lr,$s2,lsr#8 @ i0 | ||
921 | and $i2,lr,$s2 @ i1 | ||
922 | and $i3,lr,$s2,lsr#16 | ||
923 | mov $s2,$s2,lsr#24 | ||
924 | ldr $i1,[$tbl,$i1,lsl#2] @ Td2[s2>>8] | ||
925 | ldr $i2,[$tbl,$i2,lsl#2] @ Td3[s2>>0] | ||
926 | ldr $s2,[$tbl,$s2,lsl#2] @ Td0[s2>>24] | ||
927 | ldr $i3,[$tbl,$i3,lsl#2] @ Td1[s2>>16] | ||
928 | eor $s0,$s0,$i1,ror#16 | ||
929 | eor $s1,$s1,$i2,ror#24 | ||
930 | eor $s2,$s2,$t2,ror#8 | ||
931 | eor $t3,$i3,$t3,ror#8 | ||
932 | |||
933 | and $i1,lr,$s3,lsr#16 @ i0 | ||
934 | and $i2,lr,$s3,lsr#8 @ i1 | ||
935 | and $i3,lr,$s3 @ i2 | ||
936 | mov $s3,$s3,lsr#24 | ||
937 | ldr $i1,[$tbl,$i1,lsl#2] @ Td1[s3>>16] | ||
938 | ldr $i2,[$tbl,$i2,lsl#2] @ Td2[s3>>8] | ||
939 | ldr $i3,[$tbl,$i3,lsl#2] @ Td3[s3>>0] | ||
940 | ldr $s3,[$tbl,$s3,lsl#2] @ Td0[s3>>24] | ||
941 | eor $s0,$s0,$i1,ror#8 | ||
942 | eor $s1,$s1,$i2,ror#16 | ||
943 | eor $s2,$s2,$i3,ror#24 | ||
944 | eor $s3,$s3,$t3,ror#8 | ||
945 | |||
946 | ldr $t1,[$key],#16 | ||
947 | ldr $t2,[$key,#-12] | ||
948 | ldr $t3,[$key,#-8] | ||
949 | ldr $i1,[$key,#-4] | ||
950 | eor $s0,$s0,$t1 | ||
951 | eor $s1,$s1,$t2 | ||
952 | eor $s2,$s2,$t3 | ||
953 | eor $s3,$s3,$i1 | ||
954 | |||
955 | subs $rounds,$rounds,#1 | ||
956 | bne .Ldec_loop | ||
957 | |||
958 | add $tbl,$tbl,#1024 | ||
959 | |||
960 | ldr $t1,[$tbl,#0] @ prefetch Td4 | ||
961 | ldr $t2,[$tbl,#32] | ||
962 | ldr $t3,[$tbl,#64] | ||
963 | ldr $i1,[$tbl,#96] | ||
964 | ldr $i2,[$tbl,#128] | ||
965 | ldr $i3,[$tbl,#160] | ||
966 | ldr $t1,[$tbl,#192] | ||
967 | ldr $t2,[$tbl,#224] | ||
968 | |||
969 | and $i1,lr,$s0,lsr#16 | ||
970 | and $i2,lr,$s0,lsr#8 | ||
971 | and $i3,lr,$s0 | ||
972 | ldrb $s0,[$tbl,$s0,lsr#24] @ Td4[s0>>24] | ||
973 | ldrb $t1,[$tbl,$i1] @ Td4[s0>>16] | ||
974 | ldrb $t2,[$tbl,$i2] @ Td4[s0>>8] | ||
975 | ldrb $t3,[$tbl,$i3] @ Td4[s0>>0] | ||
976 | |||
977 | and $i1,lr,$s1 @ i0 | ||
978 | and $i2,lr,$s1,lsr#16 | ||
979 | and $i3,lr,$s1,lsr#8 | ||
980 | ldrb $i1,[$tbl,$i1] @ Td4[s1>>0] | ||
981 | ldrb $s1,[$tbl,$s1,lsr#24] @ Td4[s1>>24] | ||
982 | ldrb $i2,[$tbl,$i2] @ Td4[s1>>16] | ||
983 | ldrb $i3,[$tbl,$i3] @ Td4[s1>>8] | ||
984 | eor $s0,$i1,$s0,lsl#24 | ||
985 | eor $s1,$t1,$s1,lsl#8 | ||
986 | eor $t2,$t2,$i2,lsl#8 | ||
987 | eor $t3,$t3,$i3,lsl#8 | ||
988 | |||
989 | and $i1,lr,$s2,lsr#8 @ i0 | ||
990 | and $i2,lr,$s2 @ i1 | ||
991 | and $i3,lr,$s2,lsr#16 | ||
992 | ldrb $i1,[$tbl,$i1] @ Td4[s2>>8] | ||
993 | ldrb $i2,[$tbl,$i2] @ Td4[s2>>0] | ||
994 | ldrb $s2,[$tbl,$s2,lsr#24] @ Td4[s2>>24] | ||
995 | ldrb $i3,[$tbl,$i3] @ Td4[s2>>16] | ||
996 | eor $s0,$s0,$i1,lsl#8 | ||
997 | eor $s1,$i2,$s1,lsl#16 | ||
998 | eor $s2,$t2,$s2,lsl#16 | ||
999 | eor $t3,$t3,$i3,lsl#16 | ||
1000 | |||
1001 | and $i1,lr,$s3,lsr#16 @ i0 | ||
1002 | and $i2,lr,$s3,lsr#8 @ i1 | ||
1003 | and $i3,lr,$s3 @ i2 | ||
1004 | ldrb $i1,[$tbl,$i1] @ Td4[s3>>16] | ||
1005 | ldrb $i2,[$tbl,$i2] @ Td4[s3>>8] | ||
1006 | ldrb $i3,[$tbl,$i3] @ Td4[s3>>0] | ||
1007 | ldrb $s3,[$tbl,$s3,lsr#24] @ Td4[s3>>24] | ||
1008 | eor $s0,$s0,$i1,lsl#16 | ||
1009 | eor $s1,$s1,$i2,lsl#8 | ||
1010 | eor $s2,$i3,$s2,lsl#8 | ||
1011 | eor $s3,$t3,$s3,lsl#24 | ||
1012 | |||
1013 | ldr lr,[sp],#4 @ pop lr | ||
1014 | ldr $t1,[$key,#0] | ||
1015 | ldr $t2,[$key,#4] | ||
1016 | ldr $t3,[$key,#8] | ||
1017 | ldr $i1,[$key,#12] | ||
1018 | eor $s0,$s0,$t1 | ||
1019 | eor $s1,$s1,$t2 | ||
1020 | eor $s2,$s2,$t3 | ||
1021 | eor $s3,$s3,$i1 | ||
1022 | |||
1023 | sub $tbl,$tbl,#1024 | ||
1024 | mov pc,lr @ return | ||
1025 | .size _armv4_AES_decrypt,.-_armv4_AES_decrypt | ||
1026 | .asciz "AES for ARMv4, CRYPTOGAMS by <appro\@openssl.org>" | ||
1027 | ___ | ||
1028 | |||
1029 | $code =~ s/\bbx\s+lr\b/.word\t0xe12fff1e/gm; # make it possible to compile with -march=armv4 | ||
1030 | print $code; | ||
diff --git a/src/lib/libcrypto/aes/asm/aes-ppc.pl b/src/lib/libcrypto/aes/asm/aes-ppc.pl new file mode 100644 index 0000000000..ce427655ef --- /dev/null +++ b/src/lib/libcrypto/aes/asm/aes-ppc.pl | |||
@@ -0,0 +1,1176 @@ | |||
1 | #!/usr/bin/env perl | ||
2 | |||
3 | # ==================================================================== | ||
4 | # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL | ||
5 | # project. The module is, however, dual licensed under OpenSSL and | ||
6 | # CRYPTOGAMS licenses depending on where you obtain it. For further | ||
7 | # details see http://www.openssl.org/~appro/cryptogams/. | ||
8 | # ==================================================================== | ||
9 | |||
10 | # Needs more work: key setup, page boundaries, CBC routine... | ||
11 | # | ||
12 | # ppc_AES_[en|de]crypt perform at 18 cycles per byte processed with | ||
13 | # 128-bit key, which is ~40% better than 64-bit code generated by gcc | ||
14 | # 4.0. But these are not the ones currently used! Their "compact" | ||
15 | # counterparts are, for security reason. ppc_AES_encrypt_compact runs | ||
16 | # at 1/2 of ppc_AES_encrypt speed, while ppc_AES_decrypt_compact - | ||
17 | # at 1/3 of ppc_AES_decrypt. | ||
18 | |||
19 | $flavour = shift; | ||
20 | |||
21 | if ($flavour =~ /64/) { | ||
22 | $SIZE_T =8; | ||
23 | $STU ="stdu"; | ||
24 | $POP ="ld"; | ||
25 | $PUSH ="std"; | ||
26 | } elsif ($flavour =~ /32/) { | ||
27 | $SIZE_T =4; | ||
28 | $STU ="stwu"; | ||
29 | $POP ="lwz"; | ||
30 | $PUSH ="stw"; | ||
31 | } else { die "nonsense $flavour"; } | ||
32 | |||
33 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; | ||
34 | ( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or | ||
35 | ( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or | ||
36 | die "can't locate ppc-xlate.pl"; | ||
37 | |||
38 | open STDOUT,"| $^X $xlate $flavour ".shift || die "can't call $xlate: $!"; | ||
39 | |||
40 | $FRAME=32*$SIZE_T; | ||
41 | |||
42 | sub _data_word() | ||
43 | { my $i; | ||
44 | while(defined($i=shift)) { $code.=sprintf"\t.long\t0x%08x,0x%08x\n",$i,$i; } | ||
45 | } | ||
46 | |||
47 | $sp="r1"; | ||
48 | $toc="r2"; | ||
49 | $inp="r3"; | ||
50 | $out="r4"; | ||
51 | $key="r5"; | ||
52 | |||
53 | $Tbl0="r3"; | ||
54 | $Tbl1="r6"; | ||
55 | $Tbl2="r7"; | ||
56 | $Tbl3="r2"; | ||
57 | |||
58 | $s0="r8"; | ||
59 | $s1="r9"; | ||
60 | $s2="r10"; | ||
61 | $s3="r11"; | ||
62 | |||
63 | $t0="r12"; | ||
64 | $t1="r13"; | ||
65 | $t2="r14"; | ||
66 | $t3="r15"; | ||
67 | |||
68 | $acc00="r16"; | ||
69 | $acc01="r17"; | ||
70 | $acc02="r18"; | ||
71 | $acc03="r19"; | ||
72 | |||
73 | $acc04="r20"; | ||
74 | $acc05="r21"; | ||
75 | $acc06="r22"; | ||
76 | $acc07="r23"; | ||
77 | |||
78 | $acc08="r24"; | ||
79 | $acc09="r25"; | ||
80 | $acc10="r26"; | ||
81 | $acc11="r27"; | ||
82 | |||
83 | $acc12="r28"; | ||
84 | $acc13="r29"; | ||
85 | $acc14="r30"; | ||
86 | $acc15="r31"; | ||
87 | |||
88 | # stay away from TLS pointer | ||
89 | if ($SIZE_T==8) { die if ($t1 ne "r13"); $t1="r0"; } | ||
90 | else { die if ($Tbl3 ne "r2"); $Tbl3=$t0; $t0="r0"; } | ||
91 | $mask80=$Tbl2; | ||
92 | $mask1b=$Tbl3; | ||
93 | |||
94 | $code.=<<___; | ||
95 | .machine "any" | ||
96 | .text | ||
97 | |||
98 | .align 7 | ||
99 | LAES_Te: | ||
100 | mflr r0 | ||
101 | bcl 20,31,\$+4 | ||
102 | mflr $Tbl0 ; vvvvv "distance" between . and 1st data entry | ||
103 | addi $Tbl0,$Tbl0,`128-8` | ||
104 | mtlr r0 | ||
105 | blr | ||
106 | .space `32-24` | ||
107 | LAES_Td: | ||
108 | mflr r0 | ||
109 | bcl 20,31,\$+4 | ||
110 | mflr $Tbl0 ; vvvvvvvv "distance" between . and 1st data entry | ||
111 | addi $Tbl0,$Tbl0,`128-8-32+2048+256` | ||
112 | mtlr r0 | ||
113 | blr | ||
114 | .space `128-32-24` | ||
115 | ___ | ||
116 | &_data_word( | ||
117 | 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, | ||
118 | 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554, | ||
119 | 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, | ||
120 | 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a, | ||
121 | 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, | ||
122 | 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b, | ||
123 | 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea, | ||
124 | 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b, | ||
125 | 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a, | ||
126 | 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f, | ||
127 | 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108, | ||
128 | 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f, | ||
129 | 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e, | ||
130 | 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5, | ||
131 | 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d, | ||
132 | 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f, | ||
133 | 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e, | ||
134 | 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb, | ||
135 | 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce, | ||
136 | 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497, | ||
137 | 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c, | ||
138 | 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed, | ||
139 | 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b, | ||
140 | 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a, | ||
141 | 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16, | ||
142 | 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594, | ||
143 | 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81, | ||
144 | 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3, | ||
145 | 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a, | ||
146 | 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504, | ||
147 | 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163, | ||
148 | 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d, | ||
149 | 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f, | ||
150 | 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739, | ||
151 | 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47, | ||
152 | 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395, | ||
153 | 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f, | ||
154 | 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883, | ||
155 | 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c, | ||
156 | 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76, | ||
157 | 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e, | ||
158 | 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4, | ||
159 | 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6, | ||
160 | 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b, | ||
161 | 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7, | ||
162 | 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0, | ||
163 | 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25, | ||
164 | 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818, | ||
165 | 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72, | ||
166 | 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651, | ||
167 | 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21, | ||
168 | 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85, | ||
169 | 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa, | ||
170 | 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12, | ||
171 | 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0, | ||
172 | 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9, | ||
173 | 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133, | ||
174 | 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7, | ||
175 | 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, | ||
176 | 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a, | ||
177 | 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, | ||
178 | 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8, | ||
179 | 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, | ||
180 | 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a); | ||
181 | $code.=<<___; | ||
182 | .byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 | ||
183 | .byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76 | ||
184 | .byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0 | ||
185 | .byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0 | ||
186 | .byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc | ||
187 | .byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15 | ||
188 | .byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a | ||
189 | .byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75 | ||
190 | .byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0 | ||
191 | .byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84 | ||
192 | .byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b | ||
193 | .byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf | ||
194 | .byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85 | ||
195 | .byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8 | ||
196 | .byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5 | ||
197 | .byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2 | ||
198 | .byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17 | ||
199 | .byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73 | ||
200 | .byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88 | ||
201 | .byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb | ||
202 | .byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c | ||
203 | .byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79 | ||
204 | .byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9 | ||
205 | .byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08 | ||
206 | .byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6 | ||
207 | .byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a | ||
208 | .byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e | ||
209 | .byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e | ||
210 | .byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94 | ||
211 | .byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf | ||
212 | .byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68 | ||
213 | .byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 | ||
214 | ___ | ||
215 | &_data_word( | ||
216 | 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, | ||
217 | 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393, | ||
218 | 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, | ||
219 | 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f, | ||
220 | 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, | ||
221 | 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6, | ||
222 | 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da, | ||
223 | 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844, | ||
224 | 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd, | ||
225 | 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4, | ||
226 | 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45, | ||
227 | 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94, | ||
228 | 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7, | ||
229 | 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a, | ||
230 | 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5, | ||
231 | 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c, | ||
232 | 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1, | ||
233 | 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a, | ||
234 | 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75, | ||
235 | 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051, | ||
236 | 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46, | ||
237 | 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff, | ||
238 | 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77, | ||
239 | 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb, | ||
240 | 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000, | ||
241 | 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e, | ||
242 | 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927, | ||
243 | 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a, | ||
244 | 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e, | ||
245 | 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16, | ||
246 | 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d, | ||
247 | 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8, | ||
248 | 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd, | ||
249 | 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34, | ||
250 | 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163, | ||
251 | 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120, | ||
252 | 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d, | ||
253 | 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0, | ||
254 | 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422, | ||
255 | 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef, | ||
256 | 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36, | ||
257 | 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4, | ||
258 | 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662, | ||
259 | 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5, | ||
260 | 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3, | ||
261 | 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b, | ||
262 | 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8, | ||
263 | 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6, | ||
264 | 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6, | ||
265 | 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0, | ||
266 | 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815, | ||
267 | 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f, | ||
268 | 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df, | ||
269 | 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f, | ||
270 | 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e, | ||
271 | 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713, | ||
272 | 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89, | ||
273 | 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c, | ||
274 | 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, | ||
275 | 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86, | ||
276 | 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, | ||
277 | 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541, | ||
278 | 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, | ||
279 | 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742); | ||
280 | $code.=<<___; | ||
281 | .byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 | ||
282 | .byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb | ||
283 | .byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87 | ||
284 | .byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb | ||
285 | .byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d | ||
286 | .byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e | ||
287 | .byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2 | ||
288 | .byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25 | ||
289 | .byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16 | ||
290 | .byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92 | ||
291 | .byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda | ||
292 | .byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84 | ||
293 | .byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a | ||
294 | .byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06 | ||
295 | .byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02 | ||
296 | .byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b | ||
297 | .byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea | ||
298 | .byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73 | ||
299 | .byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85 | ||
300 | .byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e | ||
301 | .byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89 | ||
302 | .byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b | ||
303 | .byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20 | ||
304 | .byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4 | ||
305 | .byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31 | ||
306 | .byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f | ||
307 | .byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d | ||
308 | .byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef | ||
309 | .byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0 | ||
310 | .byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61 | ||
311 | .byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26 | ||
312 | .byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d | ||
313 | |||
314 | |||
315 | .globl .AES_encrypt | ||
316 | .align 7 | ||
317 | .AES_encrypt: | ||
318 | mflr r0 | ||
319 | $STU $sp,-$FRAME($sp) | ||
320 | |||
321 | $PUSH r0,`$FRAME-$SIZE_T*21`($sp) | ||
322 | $PUSH $toc,`$FRAME-$SIZE_T*20`($sp) | ||
323 | $PUSH r13,`$FRAME-$SIZE_T*19`($sp) | ||
324 | $PUSH r14,`$FRAME-$SIZE_T*18`($sp) | ||
325 | $PUSH r15,`$FRAME-$SIZE_T*17`($sp) | ||
326 | $PUSH r16,`$FRAME-$SIZE_T*16`($sp) | ||
327 | $PUSH r17,`$FRAME-$SIZE_T*15`($sp) | ||
328 | $PUSH r18,`$FRAME-$SIZE_T*14`($sp) | ||
329 | $PUSH r19,`$FRAME-$SIZE_T*13`($sp) | ||
330 | $PUSH r20,`$FRAME-$SIZE_T*12`($sp) | ||
331 | $PUSH r21,`$FRAME-$SIZE_T*11`($sp) | ||
332 | $PUSH r22,`$FRAME-$SIZE_T*10`($sp) | ||
333 | $PUSH r23,`$FRAME-$SIZE_T*9`($sp) | ||
334 | $PUSH r24,`$FRAME-$SIZE_T*8`($sp) | ||
335 | $PUSH r25,`$FRAME-$SIZE_T*7`($sp) | ||
336 | $PUSH r26,`$FRAME-$SIZE_T*6`($sp) | ||
337 | $PUSH r27,`$FRAME-$SIZE_T*5`($sp) | ||
338 | $PUSH r28,`$FRAME-$SIZE_T*4`($sp) | ||
339 | $PUSH r29,`$FRAME-$SIZE_T*3`($sp) | ||
340 | $PUSH r30,`$FRAME-$SIZE_T*2`($sp) | ||
341 | $PUSH r31,`$FRAME-$SIZE_T*1`($sp) | ||
342 | |||
343 | lwz $s0,0($inp) | ||
344 | lwz $s1,4($inp) | ||
345 | lwz $s2,8($inp) | ||
346 | lwz $s3,12($inp) | ||
347 | bl LAES_Te | ||
348 | bl Lppc_AES_encrypt_compact | ||
349 | stw $s0,0($out) | ||
350 | stw $s1,4($out) | ||
351 | stw $s2,8($out) | ||
352 | stw $s3,12($out) | ||
353 | |||
354 | $POP r0,`$FRAME-$SIZE_T*21`($sp) | ||
355 | $POP $toc,`$FRAME-$SIZE_T*20`($sp) | ||
356 | $POP r13,`$FRAME-$SIZE_T*19`($sp) | ||
357 | $POP r14,`$FRAME-$SIZE_T*18`($sp) | ||
358 | $POP r15,`$FRAME-$SIZE_T*17`($sp) | ||
359 | $POP r16,`$FRAME-$SIZE_T*16`($sp) | ||
360 | $POP r17,`$FRAME-$SIZE_T*15`($sp) | ||
361 | $POP r18,`$FRAME-$SIZE_T*14`($sp) | ||
362 | $POP r19,`$FRAME-$SIZE_T*13`($sp) | ||
363 | $POP r20,`$FRAME-$SIZE_T*12`($sp) | ||
364 | $POP r21,`$FRAME-$SIZE_T*11`($sp) | ||
365 | $POP r22,`$FRAME-$SIZE_T*10`($sp) | ||
366 | $POP r23,`$FRAME-$SIZE_T*9`($sp) | ||
367 | $POP r24,`$FRAME-$SIZE_T*8`($sp) | ||
368 | $POP r25,`$FRAME-$SIZE_T*7`($sp) | ||
369 | $POP r26,`$FRAME-$SIZE_T*6`($sp) | ||
370 | $POP r27,`$FRAME-$SIZE_T*5`($sp) | ||
371 | $POP r28,`$FRAME-$SIZE_T*4`($sp) | ||
372 | $POP r29,`$FRAME-$SIZE_T*3`($sp) | ||
373 | $POP r30,`$FRAME-$SIZE_T*2`($sp) | ||
374 | $POP r31,`$FRAME-$SIZE_T*1`($sp) | ||
375 | mtlr r0 | ||
376 | addi $sp,$sp,$FRAME | ||
377 | blr | ||
378 | |||
379 | .align 4 | ||
380 | Lppc_AES_encrypt: | ||
381 | lwz $acc00,240($key) | ||
382 | lwz $t0,0($key) | ||
383 | lwz $t1,4($key) | ||
384 | lwz $t2,8($key) | ||
385 | lwz $t3,12($key) | ||
386 | addi $Tbl1,$Tbl0,3 | ||
387 | addi $Tbl2,$Tbl0,2 | ||
388 | addi $Tbl3,$Tbl0,1 | ||
389 | addi $acc00,$acc00,-1 | ||
390 | addi $key,$key,16 | ||
391 | xor $s0,$s0,$t0 | ||
392 | xor $s1,$s1,$t1 | ||
393 | xor $s2,$s2,$t2 | ||
394 | xor $s3,$s3,$t3 | ||
395 | mtctr $acc00 | ||
396 | .align 4 | ||
397 | Lenc_loop: | ||
398 | rlwinm $acc00,$s0,`32-24+3`,21,28 | ||
399 | rlwinm $acc01,$s1,`32-24+3`,21,28 | ||
400 | lwz $t0,0($key) | ||
401 | lwz $t1,4($key) | ||
402 | rlwinm $acc02,$s2,`32-24+3`,21,28 | ||
403 | rlwinm $acc03,$s3,`32-24+3`,21,28 | ||
404 | lwz $t2,8($key) | ||
405 | lwz $t3,12($key) | ||
406 | rlwinm $acc04,$s1,`32-16+3`,21,28 | ||
407 | rlwinm $acc05,$s2,`32-16+3`,21,28 | ||
408 | lwzx $acc00,$Tbl0,$acc00 | ||
409 | lwzx $acc01,$Tbl0,$acc01 | ||
410 | rlwinm $acc06,$s3,`32-16+3`,21,28 | ||
411 | rlwinm $acc07,$s0,`32-16+3`,21,28 | ||
412 | lwzx $acc02,$Tbl0,$acc02 | ||
413 | lwzx $acc03,$Tbl0,$acc03 | ||
414 | rlwinm $acc08,$s2,`32-8+3`,21,28 | ||
415 | rlwinm $acc09,$s3,`32-8+3`,21,28 | ||
416 | lwzx $acc04,$Tbl1,$acc04 | ||
417 | lwzx $acc05,$Tbl1,$acc05 | ||
418 | rlwinm $acc10,$s0,`32-8+3`,21,28 | ||
419 | rlwinm $acc11,$s1,`32-8+3`,21,28 | ||
420 | lwzx $acc06,$Tbl1,$acc06 | ||
421 | lwzx $acc07,$Tbl1,$acc07 | ||
422 | rlwinm $acc12,$s3,`0+3`,21,28 | ||
423 | rlwinm $acc13,$s0,`0+3`,21,28 | ||
424 | lwzx $acc08,$Tbl2,$acc08 | ||
425 | lwzx $acc09,$Tbl2,$acc09 | ||
426 | rlwinm $acc14,$s1,`0+3`,21,28 | ||
427 | rlwinm $acc15,$s2,`0+3`,21,28 | ||
428 | lwzx $acc10,$Tbl2,$acc10 | ||
429 | lwzx $acc11,$Tbl2,$acc11 | ||
430 | xor $t0,$t0,$acc00 | ||
431 | xor $t1,$t1,$acc01 | ||
432 | lwzx $acc12,$Tbl3,$acc12 | ||
433 | lwzx $acc13,$Tbl3,$acc13 | ||
434 | xor $t2,$t2,$acc02 | ||
435 | xor $t3,$t3,$acc03 | ||
436 | lwzx $acc14,$Tbl3,$acc14 | ||
437 | lwzx $acc15,$Tbl3,$acc15 | ||
438 | xor $t0,$t0,$acc04 | ||
439 | xor $t1,$t1,$acc05 | ||
440 | xor $t2,$t2,$acc06 | ||
441 | xor $t3,$t3,$acc07 | ||
442 | xor $t0,$t0,$acc08 | ||
443 | xor $t1,$t1,$acc09 | ||
444 | xor $t2,$t2,$acc10 | ||
445 | xor $t3,$t3,$acc11 | ||
446 | xor $s0,$t0,$acc12 | ||
447 | xor $s1,$t1,$acc13 | ||
448 | xor $s2,$t2,$acc14 | ||
449 | xor $s3,$t3,$acc15 | ||
450 | addi $key,$key,16 | ||
451 | bdnz- Lenc_loop | ||
452 | |||
453 | addi $Tbl2,$Tbl0,2048 | ||
454 | nop | ||
455 | lwz $acc08,`2048+0`($Tbl0) ! prefetch Te4 | ||
456 | lwz $acc09,`2048+32`($Tbl0) | ||
457 | lwz $acc10,`2048+64`($Tbl0) | ||
458 | lwz $acc11,`2048+96`($Tbl0) | ||
459 | lwz $acc08,`2048+128`($Tbl0) | ||
460 | lwz $acc09,`2048+160`($Tbl0) | ||
461 | lwz $acc10,`2048+192`($Tbl0) | ||
462 | lwz $acc11,`2048+224`($Tbl0) | ||
463 | rlwinm $acc00,$s0,`32-24`,24,31 | ||
464 | rlwinm $acc01,$s1,`32-24`,24,31 | ||
465 | lwz $t0,0($key) | ||
466 | lwz $t1,4($key) | ||
467 | rlwinm $acc02,$s2,`32-24`,24,31 | ||
468 | rlwinm $acc03,$s3,`32-24`,24,31 | ||
469 | lwz $t2,8($key) | ||
470 | lwz $t3,12($key) | ||
471 | rlwinm $acc04,$s1,`32-16`,24,31 | ||
472 | rlwinm $acc05,$s2,`32-16`,24,31 | ||
473 | lbzx $acc00,$Tbl2,$acc00 | ||
474 | lbzx $acc01,$Tbl2,$acc01 | ||
475 | rlwinm $acc06,$s3,`32-16`,24,31 | ||
476 | rlwinm $acc07,$s0,`32-16`,24,31 | ||
477 | lbzx $acc02,$Tbl2,$acc02 | ||
478 | lbzx $acc03,$Tbl2,$acc03 | ||
479 | rlwinm $acc08,$s2,`32-8`,24,31 | ||
480 | rlwinm $acc09,$s3,`32-8`,24,31 | ||
481 | lbzx $acc04,$Tbl2,$acc04 | ||
482 | lbzx $acc05,$Tbl2,$acc05 | ||
483 | rlwinm $acc10,$s0,`32-8`,24,31 | ||
484 | rlwinm $acc11,$s1,`32-8`,24,31 | ||
485 | lbzx $acc06,$Tbl2,$acc06 | ||
486 | lbzx $acc07,$Tbl2,$acc07 | ||
487 | rlwinm $acc12,$s3,`0`,24,31 | ||
488 | rlwinm $acc13,$s0,`0`,24,31 | ||
489 | lbzx $acc08,$Tbl2,$acc08 | ||
490 | lbzx $acc09,$Tbl2,$acc09 | ||
491 | rlwinm $acc14,$s1,`0`,24,31 | ||
492 | rlwinm $acc15,$s2,`0`,24,31 | ||
493 | lbzx $acc10,$Tbl2,$acc10 | ||
494 | lbzx $acc11,$Tbl2,$acc11 | ||
495 | rlwinm $s0,$acc00,24,0,7 | ||
496 | rlwinm $s1,$acc01,24,0,7 | ||
497 | lbzx $acc12,$Tbl2,$acc12 | ||
498 | lbzx $acc13,$Tbl2,$acc13 | ||
499 | rlwinm $s2,$acc02,24,0,7 | ||
500 | rlwinm $s3,$acc03,24,0,7 | ||
501 | lbzx $acc14,$Tbl2,$acc14 | ||
502 | lbzx $acc15,$Tbl2,$acc15 | ||
503 | rlwimi $s0,$acc04,16,8,15 | ||
504 | rlwimi $s1,$acc05,16,8,15 | ||
505 | rlwimi $s2,$acc06,16,8,15 | ||
506 | rlwimi $s3,$acc07,16,8,15 | ||
507 | rlwimi $s0,$acc08,8,16,23 | ||
508 | rlwimi $s1,$acc09,8,16,23 | ||
509 | rlwimi $s2,$acc10,8,16,23 | ||
510 | rlwimi $s3,$acc11,8,16,23 | ||
511 | or $s0,$s0,$acc12 | ||
512 | or $s1,$s1,$acc13 | ||
513 | or $s2,$s2,$acc14 | ||
514 | or $s3,$s3,$acc15 | ||
515 | xor $s0,$s0,$t0 | ||
516 | xor $s1,$s1,$t1 | ||
517 | xor $s2,$s2,$t2 | ||
518 | xor $s3,$s3,$t3 | ||
519 | blr | ||
520 | |||
521 | .align 4 | ||
522 | Lppc_AES_encrypt_compact: | ||
523 | lwz $acc00,240($key) | ||
524 | lwz $t0,0($key) | ||
525 | lwz $t1,4($key) | ||
526 | lwz $t2,8($key) | ||
527 | lwz $t3,12($key) | ||
528 | addi $Tbl1,$Tbl0,2048 | ||
529 | lis $mask80,0x8080 | ||
530 | lis $mask1b,0x1b1b | ||
531 | addi $key,$key,16 | ||
532 | ori $mask80,$mask80,0x8080 | ||
533 | ori $mask1b,$mask1b,0x1b1b | ||
534 | mtctr $acc00 | ||
535 | .align 4 | ||
536 | Lenc_compact_loop: | ||
537 | xor $s0,$s0,$t0 | ||
538 | xor $s1,$s1,$t1 | ||
539 | xor $s2,$s2,$t2 | ||
540 | xor $s3,$s3,$t3 | ||
541 | rlwinm $acc00,$s0,`32-24`,24,31 | ||
542 | rlwinm $acc01,$s1,`32-24`,24,31 | ||
543 | rlwinm $acc02,$s2,`32-24`,24,31 | ||
544 | rlwinm $acc03,$s3,`32-24`,24,31 | ||
545 | lbzx $acc00,$Tbl1,$acc00 | ||
546 | lbzx $acc01,$Tbl1,$acc01 | ||
547 | rlwinm $acc04,$s1,`32-16`,24,31 | ||
548 | rlwinm $acc05,$s2,`32-16`,24,31 | ||
549 | lbzx $acc02,$Tbl1,$acc02 | ||
550 | lbzx $acc03,$Tbl1,$acc03 | ||
551 | rlwinm $acc06,$s3,`32-16`,24,31 | ||
552 | rlwinm $acc07,$s0,`32-16`,24,31 | ||
553 | lbzx $acc04,$Tbl1,$acc04 | ||
554 | lbzx $acc05,$Tbl1,$acc05 | ||
555 | rlwinm $acc08,$s2,`32-8`,24,31 | ||
556 | rlwinm $acc09,$s3,`32-8`,24,31 | ||
557 | lbzx $acc06,$Tbl1,$acc06 | ||
558 | lbzx $acc07,$Tbl1,$acc07 | ||
559 | rlwinm $acc10,$s0,`32-8`,24,31 | ||
560 | rlwinm $acc11,$s1,`32-8`,24,31 | ||
561 | lbzx $acc08,$Tbl1,$acc08 | ||
562 | lbzx $acc09,$Tbl1,$acc09 | ||
563 | rlwinm $acc12,$s3,`0`,24,31 | ||
564 | rlwinm $acc13,$s0,`0`,24,31 | ||
565 | lbzx $acc10,$Tbl1,$acc10 | ||
566 | lbzx $acc11,$Tbl1,$acc11 | ||
567 | rlwinm $acc14,$s1,`0`,24,31 | ||
568 | rlwinm $acc15,$s2,`0`,24,31 | ||
569 | lbzx $acc12,$Tbl1,$acc12 | ||
570 | lbzx $acc13,$Tbl1,$acc13 | ||
571 | rlwinm $s0,$acc00,24,0,7 | ||
572 | rlwinm $s1,$acc01,24,0,7 | ||
573 | lbzx $acc14,$Tbl1,$acc14 | ||
574 | lbzx $acc15,$Tbl1,$acc15 | ||
575 | rlwinm $s2,$acc02,24,0,7 | ||
576 | rlwinm $s3,$acc03,24,0,7 | ||
577 | rlwimi $s0,$acc04,16,8,15 | ||
578 | rlwimi $s1,$acc05,16,8,15 | ||
579 | rlwimi $s2,$acc06,16,8,15 | ||
580 | rlwimi $s3,$acc07,16,8,15 | ||
581 | rlwimi $s0,$acc08,8,16,23 | ||
582 | rlwimi $s1,$acc09,8,16,23 | ||
583 | rlwimi $s2,$acc10,8,16,23 | ||
584 | rlwimi $s3,$acc11,8,16,23 | ||
585 | lwz $t0,0($key) | ||
586 | lwz $t1,4($key) | ||
587 | or $s0,$s0,$acc12 | ||
588 | or $s1,$s1,$acc13 | ||
589 | lwz $t2,8($key) | ||
590 | lwz $t3,12($key) | ||
591 | or $s2,$s2,$acc14 | ||
592 | or $s3,$s3,$acc15 | ||
593 | |||
594 | addi $key,$key,16 | ||
595 | bdz Lenc_compact_done | ||
596 | |||
597 | and $acc00,$s0,$mask80 # r1=r0&0x80808080 | ||
598 | and $acc01,$s1,$mask80 | ||
599 | and $acc02,$s2,$mask80 | ||
600 | and $acc03,$s3,$mask80 | ||
601 | srwi $acc04,$acc00,7 # r1>>7 | ||
602 | srwi $acc05,$acc01,7 | ||
603 | srwi $acc06,$acc02,7 | ||
604 | srwi $acc07,$acc03,7 | ||
605 | andc $acc08,$s0,$mask80 # r0&0x7f7f7f7f | ||
606 | andc $acc09,$s1,$mask80 | ||
607 | andc $acc10,$s2,$mask80 | ||
608 | andc $acc11,$s3,$mask80 | ||
609 | sub $acc00,$acc00,$acc04 # r1-(r1>>7) | ||
610 | sub $acc01,$acc01,$acc05 | ||
611 | sub $acc02,$acc02,$acc06 | ||
612 | sub $acc03,$acc03,$acc07 | ||
613 | add $acc08,$acc08,$acc08 # (r0&0x7f7f7f7f)<<1 | ||
614 | add $acc09,$acc09,$acc09 | ||
615 | add $acc10,$acc10,$acc10 | ||
616 | add $acc11,$acc11,$acc11 | ||
617 | and $acc00,$acc00,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
618 | and $acc01,$acc01,$mask1b | ||
619 | and $acc02,$acc02,$mask1b | ||
620 | and $acc03,$acc03,$mask1b | ||
621 | xor $acc00,$acc00,$acc08 # r2 | ||
622 | xor $acc01,$acc01,$acc09 | ||
623 | xor $acc02,$acc02,$acc10 | ||
624 | xor $acc03,$acc03,$acc11 | ||
625 | |||
626 | rotlwi $acc12,$s0,16 # ROTATE(r0,16) | ||
627 | rotlwi $acc13,$s1,16 | ||
628 | rotlwi $acc14,$s2,16 | ||
629 | rotlwi $acc15,$s3,16 | ||
630 | xor $s0,$s0,$acc00 # r0^r2 | ||
631 | xor $s1,$s1,$acc01 | ||
632 | xor $s2,$s2,$acc02 | ||
633 | xor $s3,$s3,$acc03 | ||
634 | rotrwi $s0,$s0,24 # ROTATE(r2^r0,24) | ||
635 | rotrwi $s1,$s1,24 | ||
636 | rotrwi $s2,$s2,24 | ||
637 | rotrwi $s3,$s3,24 | ||
638 | xor $s0,$s0,$acc00 # ROTATE(r2^r0,24)^r2 | ||
639 | xor $s1,$s1,$acc01 | ||
640 | xor $s2,$s2,$acc02 | ||
641 | xor $s3,$s3,$acc03 | ||
642 | rotlwi $acc08,$acc12,8 # ROTATE(r0,24) | ||
643 | rotlwi $acc09,$acc13,8 | ||
644 | rotlwi $acc10,$acc14,8 | ||
645 | rotlwi $acc11,$acc15,8 | ||
646 | xor $s0,$s0,$acc12 # | ||
647 | xor $s1,$s1,$acc13 | ||
648 | xor $s2,$s2,$acc14 | ||
649 | xor $s3,$s3,$acc15 | ||
650 | xor $s0,$s0,$acc08 # | ||
651 | xor $s1,$s1,$acc09 | ||
652 | xor $s2,$s2,$acc10 | ||
653 | xor $s3,$s3,$acc11 | ||
654 | |||
655 | b Lenc_compact_loop | ||
656 | .align 4 | ||
657 | Lenc_compact_done: | ||
658 | xor $s0,$s0,$t0 | ||
659 | xor $s1,$s1,$t1 | ||
660 | xor $s2,$s2,$t2 | ||
661 | xor $s3,$s3,$t3 | ||
662 | blr | ||
663 | |||
664 | .globl .AES_decrypt | ||
665 | .align 7 | ||
666 | .AES_decrypt: | ||
667 | mflr r0 | ||
668 | $STU $sp,-$FRAME($sp) | ||
669 | |||
670 | $PUSH r0,`$FRAME-$SIZE_T*21`($sp) | ||
671 | $PUSH $toc,`$FRAME-$SIZE_T*20`($sp) | ||
672 | $PUSH r13,`$FRAME-$SIZE_T*19`($sp) | ||
673 | $PUSH r14,`$FRAME-$SIZE_T*18`($sp) | ||
674 | $PUSH r15,`$FRAME-$SIZE_T*17`($sp) | ||
675 | $PUSH r16,`$FRAME-$SIZE_T*16`($sp) | ||
676 | $PUSH r17,`$FRAME-$SIZE_T*15`($sp) | ||
677 | $PUSH r18,`$FRAME-$SIZE_T*14`($sp) | ||
678 | $PUSH r19,`$FRAME-$SIZE_T*13`($sp) | ||
679 | $PUSH r20,`$FRAME-$SIZE_T*12`($sp) | ||
680 | $PUSH r21,`$FRAME-$SIZE_T*11`($sp) | ||
681 | $PUSH r22,`$FRAME-$SIZE_T*10`($sp) | ||
682 | $PUSH r23,`$FRAME-$SIZE_T*9`($sp) | ||
683 | $PUSH r24,`$FRAME-$SIZE_T*8`($sp) | ||
684 | $PUSH r25,`$FRAME-$SIZE_T*7`($sp) | ||
685 | $PUSH r26,`$FRAME-$SIZE_T*6`($sp) | ||
686 | $PUSH r27,`$FRAME-$SIZE_T*5`($sp) | ||
687 | $PUSH r28,`$FRAME-$SIZE_T*4`($sp) | ||
688 | $PUSH r29,`$FRAME-$SIZE_T*3`($sp) | ||
689 | $PUSH r30,`$FRAME-$SIZE_T*2`($sp) | ||
690 | $PUSH r31,`$FRAME-$SIZE_T*1`($sp) | ||
691 | |||
692 | lwz $s0,0($inp) | ||
693 | lwz $s1,4($inp) | ||
694 | lwz $s2,8($inp) | ||
695 | lwz $s3,12($inp) | ||
696 | bl LAES_Td | ||
697 | bl Lppc_AES_decrypt_compact | ||
698 | stw $s0,0($out) | ||
699 | stw $s1,4($out) | ||
700 | stw $s2,8($out) | ||
701 | stw $s3,12($out) | ||
702 | |||
703 | $POP r0,`$FRAME-$SIZE_T*21`($sp) | ||
704 | $POP $toc,`$FRAME-$SIZE_T*20`($sp) | ||
705 | $POP r13,`$FRAME-$SIZE_T*19`($sp) | ||
706 | $POP r14,`$FRAME-$SIZE_T*18`($sp) | ||
707 | $POP r15,`$FRAME-$SIZE_T*17`($sp) | ||
708 | $POP r16,`$FRAME-$SIZE_T*16`($sp) | ||
709 | $POP r17,`$FRAME-$SIZE_T*15`($sp) | ||
710 | $POP r18,`$FRAME-$SIZE_T*14`($sp) | ||
711 | $POP r19,`$FRAME-$SIZE_T*13`($sp) | ||
712 | $POP r20,`$FRAME-$SIZE_T*12`($sp) | ||
713 | $POP r21,`$FRAME-$SIZE_T*11`($sp) | ||
714 | $POP r22,`$FRAME-$SIZE_T*10`($sp) | ||
715 | $POP r23,`$FRAME-$SIZE_T*9`($sp) | ||
716 | $POP r24,`$FRAME-$SIZE_T*8`($sp) | ||
717 | $POP r25,`$FRAME-$SIZE_T*7`($sp) | ||
718 | $POP r26,`$FRAME-$SIZE_T*6`($sp) | ||
719 | $POP r27,`$FRAME-$SIZE_T*5`($sp) | ||
720 | $POP r28,`$FRAME-$SIZE_T*4`($sp) | ||
721 | $POP r29,`$FRAME-$SIZE_T*3`($sp) | ||
722 | $POP r30,`$FRAME-$SIZE_T*2`($sp) | ||
723 | $POP r31,`$FRAME-$SIZE_T*1`($sp) | ||
724 | mtlr r0 | ||
725 | addi $sp,$sp,$FRAME | ||
726 | blr | ||
727 | |||
728 | .align 4 | ||
729 | Lppc_AES_decrypt: | ||
730 | lwz $acc00,240($key) | ||
731 | lwz $t0,0($key) | ||
732 | lwz $t1,4($key) | ||
733 | lwz $t2,8($key) | ||
734 | lwz $t3,12($key) | ||
735 | addi $Tbl1,$Tbl0,3 | ||
736 | addi $Tbl2,$Tbl0,2 | ||
737 | addi $Tbl3,$Tbl0,1 | ||
738 | addi $acc00,$acc00,-1 | ||
739 | addi $key,$key,16 | ||
740 | xor $s0,$s0,$t0 | ||
741 | xor $s1,$s1,$t1 | ||
742 | xor $s2,$s2,$t2 | ||
743 | xor $s3,$s3,$t3 | ||
744 | mtctr $acc00 | ||
745 | .align 4 | ||
746 | Ldec_loop: | ||
747 | rlwinm $acc00,$s0,`32-24+3`,21,28 | ||
748 | rlwinm $acc01,$s1,`32-24+3`,21,28 | ||
749 | lwz $t0,0($key) | ||
750 | lwz $t1,4($key) | ||
751 | rlwinm $acc02,$s2,`32-24+3`,21,28 | ||
752 | rlwinm $acc03,$s3,`32-24+3`,21,28 | ||
753 | lwz $t2,8($key) | ||
754 | lwz $t3,12($key) | ||
755 | rlwinm $acc04,$s3,`32-16+3`,21,28 | ||
756 | rlwinm $acc05,$s0,`32-16+3`,21,28 | ||
757 | lwzx $acc00,$Tbl0,$acc00 | ||
758 | lwzx $acc01,$Tbl0,$acc01 | ||
759 | rlwinm $acc06,$s1,`32-16+3`,21,28 | ||
760 | rlwinm $acc07,$s2,`32-16+3`,21,28 | ||
761 | lwzx $acc02,$Tbl0,$acc02 | ||
762 | lwzx $acc03,$Tbl0,$acc03 | ||
763 | rlwinm $acc08,$s2,`32-8+3`,21,28 | ||
764 | rlwinm $acc09,$s3,`32-8+3`,21,28 | ||
765 | lwzx $acc04,$Tbl1,$acc04 | ||
766 | lwzx $acc05,$Tbl1,$acc05 | ||
767 | rlwinm $acc10,$s0,`32-8+3`,21,28 | ||
768 | rlwinm $acc11,$s1,`32-8+3`,21,28 | ||
769 | lwzx $acc06,$Tbl1,$acc06 | ||
770 | lwzx $acc07,$Tbl1,$acc07 | ||
771 | rlwinm $acc12,$s1,`0+3`,21,28 | ||
772 | rlwinm $acc13,$s2,`0+3`,21,28 | ||
773 | lwzx $acc08,$Tbl2,$acc08 | ||
774 | lwzx $acc09,$Tbl2,$acc09 | ||
775 | rlwinm $acc14,$s3,`0+3`,21,28 | ||
776 | rlwinm $acc15,$s0,`0+3`,21,28 | ||
777 | lwzx $acc10,$Tbl2,$acc10 | ||
778 | lwzx $acc11,$Tbl2,$acc11 | ||
779 | xor $t0,$t0,$acc00 | ||
780 | xor $t1,$t1,$acc01 | ||
781 | lwzx $acc12,$Tbl3,$acc12 | ||
782 | lwzx $acc13,$Tbl3,$acc13 | ||
783 | xor $t2,$t2,$acc02 | ||
784 | xor $t3,$t3,$acc03 | ||
785 | lwzx $acc14,$Tbl3,$acc14 | ||
786 | lwzx $acc15,$Tbl3,$acc15 | ||
787 | xor $t0,$t0,$acc04 | ||
788 | xor $t1,$t1,$acc05 | ||
789 | xor $t2,$t2,$acc06 | ||
790 | xor $t3,$t3,$acc07 | ||
791 | xor $t0,$t0,$acc08 | ||
792 | xor $t1,$t1,$acc09 | ||
793 | xor $t2,$t2,$acc10 | ||
794 | xor $t3,$t3,$acc11 | ||
795 | xor $s0,$t0,$acc12 | ||
796 | xor $s1,$t1,$acc13 | ||
797 | xor $s2,$t2,$acc14 | ||
798 | xor $s3,$t3,$acc15 | ||
799 | addi $key,$key,16 | ||
800 | bdnz- Ldec_loop | ||
801 | |||
802 | addi $Tbl2,$Tbl0,2048 | ||
803 | nop | ||
804 | lwz $acc08,`2048+0`($Tbl0) ! prefetch Td4 | ||
805 | lwz $acc09,`2048+32`($Tbl0) | ||
806 | lwz $acc10,`2048+64`($Tbl0) | ||
807 | lwz $acc11,`2048+96`($Tbl0) | ||
808 | lwz $acc08,`2048+128`($Tbl0) | ||
809 | lwz $acc09,`2048+160`($Tbl0) | ||
810 | lwz $acc10,`2048+192`($Tbl0) | ||
811 | lwz $acc11,`2048+224`($Tbl0) | ||
812 | rlwinm $acc00,$s0,`32-24`,24,31 | ||
813 | rlwinm $acc01,$s1,`32-24`,24,31 | ||
814 | lwz $t0,0($key) | ||
815 | lwz $t1,4($key) | ||
816 | rlwinm $acc02,$s2,`32-24`,24,31 | ||
817 | rlwinm $acc03,$s3,`32-24`,24,31 | ||
818 | lwz $t2,8($key) | ||
819 | lwz $t3,12($key) | ||
820 | rlwinm $acc04,$s3,`32-16`,24,31 | ||
821 | rlwinm $acc05,$s0,`32-16`,24,31 | ||
822 | lbzx $acc00,$Tbl2,$acc00 | ||
823 | lbzx $acc01,$Tbl2,$acc01 | ||
824 | rlwinm $acc06,$s1,`32-16`,24,31 | ||
825 | rlwinm $acc07,$s2,`32-16`,24,31 | ||
826 | lbzx $acc02,$Tbl2,$acc02 | ||
827 | lbzx $acc03,$Tbl2,$acc03 | ||
828 | rlwinm $acc08,$s2,`32-8`,24,31 | ||
829 | rlwinm $acc09,$s3,`32-8`,24,31 | ||
830 | lbzx $acc04,$Tbl2,$acc04 | ||
831 | lbzx $acc05,$Tbl2,$acc05 | ||
832 | rlwinm $acc10,$s0,`32-8`,24,31 | ||
833 | rlwinm $acc11,$s1,`32-8`,24,31 | ||
834 | lbzx $acc06,$Tbl2,$acc06 | ||
835 | lbzx $acc07,$Tbl2,$acc07 | ||
836 | rlwinm $acc12,$s1,`0`,24,31 | ||
837 | rlwinm $acc13,$s2,`0`,24,31 | ||
838 | lbzx $acc08,$Tbl2,$acc08 | ||
839 | lbzx $acc09,$Tbl2,$acc09 | ||
840 | rlwinm $acc14,$s3,`0`,24,31 | ||
841 | rlwinm $acc15,$s0,`0`,24,31 | ||
842 | lbzx $acc10,$Tbl2,$acc10 | ||
843 | lbzx $acc11,$Tbl2,$acc11 | ||
844 | rlwinm $s0,$acc00,24,0,7 | ||
845 | rlwinm $s1,$acc01,24,0,7 | ||
846 | lbzx $acc12,$Tbl2,$acc12 | ||
847 | lbzx $acc13,$Tbl2,$acc13 | ||
848 | rlwinm $s2,$acc02,24,0,7 | ||
849 | rlwinm $s3,$acc03,24,0,7 | ||
850 | lbzx $acc14,$Tbl2,$acc14 | ||
851 | lbzx $acc15,$Tbl2,$acc15 | ||
852 | rlwimi $s0,$acc04,16,8,15 | ||
853 | rlwimi $s1,$acc05,16,8,15 | ||
854 | rlwimi $s2,$acc06,16,8,15 | ||
855 | rlwimi $s3,$acc07,16,8,15 | ||
856 | rlwimi $s0,$acc08,8,16,23 | ||
857 | rlwimi $s1,$acc09,8,16,23 | ||
858 | rlwimi $s2,$acc10,8,16,23 | ||
859 | rlwimi $s3,$acc11,8,16,23 | ||
860 | or $s0,$s0,$acc12 | ||
861 | or $s1,$s1,$acc13 | ||
862 | or $s2,$s2,$acc14 | ||
863 | or $s3,$s3,$acc15 | ||
864 | xor $s0,$s0,$t0 | ||
865 | xor $s1,$s1,$t1 | ||
866 | xor $s2,$s2,$t2 | ||
867 | xor $s3,$s3,$t3 | ||
868 | blr | ||
869 | |||
870 | .align 4 | ||
871 | Lppc_AES_decrypt_compact: | ||
872 | lwz $acc00,240($key) | ||
873 | lwz $t0,0($key) | ||
874 | lwz $t1,4($key) | ||
875 | lwz $t2,8($key) | ||
876 | lwz $t3,12($key) | ||
877 | addi $Tbl1,$Tbl0,2048 | ||
878 | lis $mask80,0x8080 | ||
879 | lis $mask1b,0x1b1b | ||
880 | addi $key,$key,16 | ||
881 | ori $mask80,$mask80,0x8080 | ||
882 | ori $mask1b,$mask1b,0x1b1b | ||
883 | ___ | ||
884 | $code.=<<___ if ($SIZE_T==8); | ||
885 | insrdi $mask80,$mask80,32,0 | ||
886 | insrdi $mask1b,$mask1b,32,0 | ||
887 | ___ | ||
888 | $code.=<<___; | ||
889 | mtctr $acc00 | ||
890 | .align 4 | ||
891 | Ldec_compact_loop: | ||
892 | xor $s0,$s0,$t0 | ||
893 | xor $s1,$s1,$t1 | ||
894 | xor $s2,$s2,$t2 | ||
895 | xor $s3,$s3,$t3 | ||
896 | rlwinm $acc00,$s0,`32-24`,24,31 | ||
897 | rlwinm $acc01,$s1,`32-24`,24,31 | ||
898 | rlwinm $acc02,$s2,`32-24`,24,31 | ||
899 | rlwinm $acc03,$s3,`32-24`,24,31 | ||
900 | lbzx $acc00,$Tbl1,$acc00 | ||
901 | lbzx $acc01,$Tbl1,$acc01 | ||
902 | rlwinm $acc04,$s3,`32-16`,24,31 | ||
903 | rlwinm $acc05,$s0,`32-16`,24,31 | ||
904 | lbzx $acc02,$Tbl1,$acc02 | ||
905 | lbzx $acc03,$Tbl1,$acc03 | ||
906 | rlwinm $acc06,$s1,`32-16`,24,31 | ||
907 | rlwinm $acc07,$s2,`32-16`,24,31 | ||
908 | lbzx $acc04,$Tbl1,$acc04 | ||
909 | lbzx $acc05,$Tbl1,$acc05 | ||
910 | rlwinm $acc08,$s2,`32-8`,24,31 | ||
911 | rlwinm $acc09,$s3,`32-8`,24,31 | ||
912 | lbzx $acc06,$Tbl1,$acc06 | ||
913 | lbzx $acc07,$Tbl1,$acc07 | ||
914 | rlwinm $acc10,$s0,`32-8`,24,31 | ||
915 | rlwinm $acc11,$s1,`32-8`,24,31 | ||
916 | lbzx $acc08,$Tbl1,$acc08 | ||
917 | lbzx $acc09,$Tbl1,$acc09 | ||
918 | rlwinm $acc12,$s1,`0`,24,31 | ||
919 | rlwinm $acc13,$s2,`0`,24,31 | ||
920 | lbzx $acc10,$Tbl1,$acc10 | ||
921 | lbzx $acc11,$Tbl1,$acc11 | ||
922 | rlwinm $acc14,$s3,`0`,24,31 | ||
923 | rlwinm $acc15,$s0,`0`,24,31 | ||
924 | lbzx $acc12,$Tbl1,$acc12 | ||
925 | lbzx $acc13,$Tbl1,$acc13 | ||
926 | rlwinm $s0,$acc00,24,0,7 | ||
927 | rlwinm $s1,$acc01,24,0,7 | ||
928 | lbzx $acc14,$Tbl1,$acc14 | ||
929 | lbzx $acc15,$Tbl1,$acc15 | ||
930 | rlwinm $s2,$acc02,24,0,7 | ||
931 | rlwinm $s3,$acc03,24,0,7 | ||
932 | rlwimi $s0,$acc04,16,8,15 | ||
933 | rlwimi $s1,$acc05,16,8,15 | ||
934 | rlwimi $s2,$acc06,16,8,15 | ||
935 | rlwimi $s3,$acc07,16,8,15 | ||
936 | rlwimi $s0,$acc08,8,16,23 | ||
937 | rlwimi $s1,$acc09,8,16,23 | ||
938 | rlwimi $s2,$acc10,8,16,23 | ||
939 | rlwimi $s3,$acc11,8,16,23 | ||
940 | lwz $t0,0($key) | ||
941 | lwz $t1,4($key) | ||
942 | or $s0,$s0,$acc12 | ||
943 | or $s1,$s1,$acc13 | ||
944 | lwz $t2,8($key) | ||
945 | lwz $t3,12($key) | ||
946 | or $s2,$s2,$acc14 | ||
947 | or $s3,$s3,$acc15 | ||
948 | |||
949 | addi $key,$key,16 | ||
950 | bdz Ldec_compact_done | ||
951 | ___ | ||
952 | $code.=<<___ if ($SIZE_T==8); | ||
953 | # vectorized permutation improves decrypt performance by 10% | ||
954 | insrdi $s0,$s1,32,0 | ||
955 | insrdi $s2,$s3,32,0 | ||
956 | |||
957 | and $acc00,$s0,$mask80 # r1=r0&0x80808080 | ||
958 | and $acc02,$s2,$mask80 | ||
959 | srdi $acc04,$acc00,7 # r1>>7 | ||
960 | srdi $acc06,$acc02,7 | ||
961 | andc $acc08,$s0,$mask80 # r0&0x7f7f7f7f | ||
962 | andc $acc10,$s2,$mask80 | ||
963 | sub $acc00,$acc00,$acc04 # r1-(r1>>7) | ||
964 | sub $acc02,$acc02,$acc06 | ||
965 | add $acc08,$acc08,$acc08 # (r0&0x7f7f7f7f)<<1 | ||
966 | add $acc10,$acc10,$acc10 | ||
967 | and $acc00,$acc00,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
968 | and $acc02,$acc02,$mask1b | ||
969 | xor $acc00,$acc00,$acc08 # r2 | ||
970 | xor $acc02,$acc02,$acc10 | ||
971 | |||
972 | and $acc04,$acc00,$mask80 # r1=r2&0x80808080 | ||
973 | and $acc06,$acc02,$mask80 | ||
974 | srdi $acc08,$acc04,7 # r1>>7 | ||
975 | srdi $acc10,$acc06,7 | ||
976 | andc $acc12,$acc00,$mask80 # r2&0x7f7f7f7f | ||
977 | andc $acc14,$acc02,$mask80 | ||
978 | sub $acc04,$acc04,$acc08 # r1-(r1>>7) | ||
979 | sub $acc06,$acc06,$acc10 | ||
980 | add $acc12,$acc12,$acc12 # (r2&0x7f7f7f7f)<<1 | ||
981 | add $acc14,$acc14,$acc14 | ||
982 | and $acc04,$acc04,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
983 | and $acc06,$acc06,$mask1b | ||
984 | xor $acc04,$acc04,$acc12 # r4 | ||
985 | xor $acc06,$acc06,$acc14 | ||
986 | |||
987 | and $acc08,$acc04,$mask80 # r1=r4&0x80808080 | ||
988 | and $acc10,$acc06,$mask80 | ||
989 | srdi $acc12,$acc08,7 # r1>>7 | ||
990 | srdi $acc14,$acc10,7 | ||
991 | sub $acc08,$acc08,$acc12 # r1-(r1>>7) | ||
992 | sub $acc10,$acc10,$acc14 | ||
993 | andc $acc12,$acc04,$mask80 # r4&0x7f7f7f7f | ||
994 | andc $acc14,$acc06,$mask80 | ||
995 | add $acc12,$acc12,$acc12 # (r4&0x7f7f7f7f)<<1 | ||
996 | add $acc14,$acc14,$acc14 | ||
997 | and $acc08,$acc08,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
998 | and $acc10,$acc10,$mask1b | ||
999 | xor $acc08,$acc08,$acc12 # r8 | ||
1000 | xor $acc10,$acc10,$acc14 | ||
1001 | |||
1002 | xor $acc00,$acc00,$s0 # r2^r0 | ||
1003 | xor $acc02,$acc02,$s2 | ||
1004 | xor $acc04,$acc04,$s0 # r4^r0 | ||
1005 | xor $acc06,$acc06,$s2 | ||
1006 | |||
1007 | extrdi $acc01,$acc00,32,0 | ||
1008 | extrdi $acc03,$acc02,32,0 | ||
1009 | extrdi $acc05,$acc04,32,0 | ||
1010 | extrdi $acc07,$acc06,32,0 | ||
1011 | extrdi $acc09,$acc08,32,0 | ||
1012 | extrdi $acc11,$acc10,32,0 | ||
1013 | ___ | ||
1014 | $code.=<<___ if ($SIZE_T==4); | ||
1015 | and $acc00,$s0,$mask80 # r1=r0&0x80808080 | ||
1016 | and $acc01,$s1,$mask80 | ||
1017 | and $acc02,$s2,$mask80 | ||
1018 | and $acc03,$s3,$mask80 | ||
1019 | srwi $acc04,$acc00,7 # r1>>7 | ||
1020 | srwi $acc05,$acc01,7 | ||
1021 | srwi $acc06,$acc02,7 | ||
1022 | srwi $acc07,$acc03,7 | ||
1023 | andc $acc08,$s0,$mask80 # r0&0x7f7f7f7f | ||
1024 | andc $acc09,$s1,$mask80 | ||
1025 | andc $acc10,$s2,$mask80 | ||
1026 | andc $acc11,$s3,$mask80 | ||
1027 | sub $acc00,$acc00,$acc04 # r1-(r1>>7) | ||
1028 | sub $acc01,$acc01,$acc05 | ||
1029 | sub $acc02,$acc02,$acc06 | ||
1030 | sub $acc03,$acc03,$acc07 | ||
1031 | add $acc08,$acc08,$acc08 # (r0&0x7f7f7f7f)<<1 | ||
1032 | add $acc09,$acc09,$acc09 | ||
1033 | add $acc10,$acc10,$acc10 | ||
1034 | add $acc11,$acc11,$acc11 | ||
1035 | and $acc00,$acc00,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
1036 | and $acc01,$acc01,$mask1b | ||
1037 | and $acc02,$acc02,$mask1b | ||
1038 | and $acc03,$acc03,$mask1b | ||
1039 | xor $acc00,$acc00,$acc08 # r2 | ||
1040 | xor $acc01,$acc01,$acc09 | ||
1041 | xor $acc02,$acc02,$acc10 | ||
1042 | xor $acc03,$acc03,$acc11 | ||
1043 | |||
1044 | and $acc04,$acc00,$mask80 # r1=r2&0x80808080 | ||
1045 | and $acc05,$acc01,$mask80 | ||
1046 | and $acc06,$acc02,$mask80 | ||
1047 | and $acc07,$acc03,$mask80 | ||
1048 | srwi $acc08,$acc04,7 # r1>>7 | ||
1049 | srwi $acc09,$acc05,7 | ||
1050 | srwi $acc10,$acc06,7 | ||
1051 | srwi $acc11,$acc07,7 | ||
1052 | andc $acc12,$acc00,$mask80 # r2&0x7f7f7f7f | ||
1053 | andc $acc13,$acc01,$mask80 | ||
1054 | andc $acc14,$acc02,$mask80 | ||
1055 | andc $acc15,$acc03,$mask80 | ||
1056 | sub $acc04,$acc04,$acc08 # r1-(r1>>7) | ||
1057 | sub $acc05,$acc05,$acc09 | ||
1058 | sub $acc06,$acc06,$acc10 | ||
1059 | sub $acc07,$acc07,$acc11 | ||
1060 | add $acc12,$acc12,$acc12 # (r2&0x7f7f7f7f)<<1 | ||
1061 | add $acc13,$acc13,$acc13 | ||
1062 | add $acc14,$acc14,$acc14 | ||
1063 | add $acc15,$acc15,$acc15 | ||
1064 | and $acc04,$acc04,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
1065 | and $acc05,$acc05,$mask1b | ||
1066 | and $acc06,$acc06,$mask1b | ||
1067 | and $acc07,$acc07,$mask1b | ||
1068 | xor $acc04,$acc04,$acc12 # r4 | ||
1069 | xor $acc05,$acc05,$acc13 | ||
1070 | xor $acc06,$acc06,$acc14 | ||
1071 | xor $acc07,$acc07,$acc15 | ||
1072 | |||
1073 | and $acc08,$acc04,$mask80 # r1=r4&0x80808080 | ||
1074 | and $acc09,$acc05,$mask80 | ||
1075 | and $acc10,$acc06,$mask80 | ||
1076 | and $acc11,$acc07,$mask80 | ||
1077 | srwi $acc12,$acc08,7 # r1>>7 | ||
1078 | srwi $acc13,$acc09,7 | ||
1079 | srwi $acc14,$acc10,7 | ||
1080 | srwi $acc15,$acc11,7 | ||
1081 | sub $acc08,$acc08,$acc12 # r1-(r1>>7) | ||
1082 | sub $acc09,$acc09,$acc13 | ||
1083 | sub $acc10,$acc10,$acc14 | ||
1084 | sub $acc11,$acc11,$acc15 | ||
1085 | andc $acc12,$acc04,$mask80 # r4&0x7f7f7f7f | ||
1086 | andc $acc13,$acc05,$mask80 | ||
1087 | andc $acc14,$acc06,$mask80 | ||
1088 | andc $acc15,$acc07,$mask80 | ||
1089 | add $acc12,$acc12,$acc12 # (r4&0x7f7f7f7f)<<1 | ||
1090 | add $acc13,$acc13,$acc13 | ||
1091 | add $acc14,$acc14,$acc14 | ||
1092 | add $acc15,$acc15,$acc15 | ||
1093 | and $acc08,$acc08,$mask1b # (r1-(r1>>7))&0x1b1b1b1b | ||
1094 | and $acc09,$acc09,$mask1b | ||
1095 | and $acc10,$acc10,$mask1b | ||
1096 | and $acc11,$acc11,$mask1b | ||
1097 | xor $acc08,$acc08,$acc12 # r8 | ||
1098 | xor $acc09,$acc09,$acc13 | ||
1099 | xor $acc10,$acc10,$acc14 | ||
1100 | xor $acc11,$acc11,$acc15 | ||
1101 | |||
1102 | xor $acc00,$acc00,$s0 # r2^r0 | ||
1103 | xor $acc01,$acc01,$s1 | ||
1104 | xor $acc02,$acc02,$s2 | ||
1105 | xor $acc03,$acc03,$s3 | ||
1106 | xor $acc04,$acc04,$s0 # r4^r0 | ||
1107 | xor $acc05,$acc05,$s1 | ||
1108 | xor $acc06,$acc06,$s2 | ||
1109 | xor $acc07,$acc07,$s3 | ||
1110 | ___ | ||
1111 | $code.=<<___; | ||
1112 | rotrwi $s0,$s0,8 # = ROTATE(r0,8) | ||
1113 | rotrwi $s1,$s1,8 | ||
1114 | rotrwi $s2,$s2,8 | ||
1115 | rotrwi $s3,$s3,8 | ||
1116 | xor $s0,$s0,$acc00 # ^= r2^r0 | ||
1117 | xor $s1,$s1,$acc01 | ||
1118 | xor $s2,$s2,$acc02 | ||
1119 | xor $s3,$s3,$acc03 | ||
1120 | xor $acc00,$acc00,$acc08 | ||
1121 | xor $acc01,$acc01,$acc09 | ||
1122 | xor $acc02,$acc02,$acc10 | ||
1123 | xor $acc03,$acc03,$acc11 | ||
1124 | xor $s0,$s0,$acc04 # ^= r4^r0 | ||
1125 | xor $s1,$s1,$acc05 | ||
1126 | xor $s2,$s2,$acc06 | ||
1127 | xor $s3,$s3,$acc07 | ||
1128 | rotrwi $acc00,$acc00,24 | ||
1129 | rotrwi $acc01,$acc01,24 | ||
1130 | rotrwi $acc02,$acc02,24 | ||
1131 | rotrwi $acc03,$acc03,24 | ||
1132 | xor $acc04,$acc04,$acc08 | ||
1133 | xor $acc05,$acc05,$acc09 | ||
1134 | xor $acc06,$acc06,$acc10 | ||
1135 | xor $acc07,$acc07,$acc11 | ||
1136 | xor $s0,$s0,$acc08 # ^= r8 [^((r4^r0)^(r2^r0)=r4^r2)] | ||
1137 | xor $s1,$s1,$acc09 | ||
1138 | xor $s2,$s2,$acc10 | ||
1139 | xor $s3,$s3,$acc11 | ||
1140 | rotrwi $acc04,$acc04,16 | ||
1141 | rotrwi $acc05,$acc05,16 | ||
1142 | rotrwi $acc06,$acc06,16 | ||
1143 | rotrwi $acc07,$acc07,16 | ||
1144 | xor $s0,$s0,$acc00 # ^= ROTATE(r8^r2^r0,24) | ||
1145 | xor $s1,$s1,$acc01 | ||
1146 | xor $s2,$s2,$acc02 | ||
1147 | xor $s3,$s3,$acc03 | ||
1148 | rotrwi $acc08,$acc08,8 | ||
1149 | rotrwi $acc09,$acc09,8 | ||
1150 | rotrwi $acc10,$acc10,8 | ||
1151 | rotrwi $acc11,$acc11,8 | ||
1152 | xor $s0,$s0,$acc04 # ^= ROTATE(r8^r4^r0,16) | ||
1153 | xor $s1,$s1,$acc05 | ||
1154 | xor $s2,$s2,$acc06 | ||
1155 | xor $s3,$s3,$acc07 | ||
1156 | xor $s0,$s0,$acc08 # ^= ROTATE(r8,8) | ||
1157 | xor $s1,$s1,$acc09 | ||
1158 | xor $s2,$s2,$acc10 | ||
1159 | xor $s3,$s3,$acc11 | ||
1160 | |||
1161 | b Ldec_compact_loop | ||
1162 | .align 4 | ||
1163 | Ldec_compact_done: | ||
1164 | xor $s0,$s0,$t0 | ||
1165 | xor $s1,$s1,$t1 | ||
1166 | xor $s2,$s2,$t2 | ||
1167 | xor $s3,$s3,$t3 | ||
1168 | blr | ||
1169 | .long 0 | ||
1170 | .asciz "AES for PPC, CRYPTOGAMS by <appro\@openssl.org>" | ||
1171 | .align 7 | ||
1172 | ___ | ||
1173 | |||
1174 | $code =~ s/\`([^\`]*)\`/eval $1/gem; | ||
1175 | print $code; | ||
1176 | close STDOUT; | ||
diff --git a/src/lib/libcrypto/aes/asm/aes-s390x.pl b/src/lib/libcrypto/aes/asm/aes-s390x.pl new file mode 100644 index 0000000000..4b27afd92f --- /dev/null +++ b/src/lib/libcrypto/aes/asm/aes-s390x.pl | |||
@@ -0,0 +1,1333 @@ | |||
1 | #!/usr/bin/env perl | ||
2 | |||
3 | # ==================================================================== | ||
4 | # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL | ||
5 | # project. The module is, however, dual licensed under OpenSSL and | ||
6 | # CRYPTOGAMS licenses depending on where you obtain it. For further | ||
7 | # details see http://www.openssl.org/~appro/cryptogams/. | ||
8 | # ==================================================================== | ||
9 | |||
10 | # AES for s390x. | ||
11 | |||
12 | # April 2007. | ||
13 | # | ||
14 | # Software performance improvement over gcc-generated code is ~70% and | ||
15 | # in absolute terms is ~73 cycles per byte processed with 128-bit key. | ||
16 | # You're likely to exclaim "why so slow?" Keep in mind that z-CPUs are | ||
17 | # *strictly* in-order execution and issued instruction [in this case | ||
18 | # load value from memory is critical] has to complete before execution | ||
19 | # flow proceeds. S-boxes are compressed to 2KB[+256B]. | ||
20 | # | ||
21 | # As for hardware acceleration support. It's basically a "teaser," as | ||
22 | # it can and should be improved in several ways. Most notably support | ||
23 | # for CBC is not utilized, nor multiple blocks are ever processed. | ||
24 | # Then software key schedule can be postponed till hardware support | ||
25 | # detection... Performance improvement over assembler is reportedly | ||
26 | # ~2.5x, but can reach >8x [naturally on larger chunks] if proper | ||
27 | # support is implemented. | ||
28 | |||
29 | # May 2007. | ||
30 | # | ||
31 | # Implement AES_set_[en|de]crypt_key. Key schedule setup is avoided | ||
32 | # for 128-bit keys, if hardware support is detected. | ||
33 | |||
34 | # Januray 2009. | ||
35 | # | ||
36 | # Add support for hardware AES192/256 and reschedule instructions to | ||
37 | # minimize/avoid Address Generation Interlock hazard and to favour | ||
38 | # dual-issue z10 pipeline. This gave ~25% improvement on z10 and | ||
39 | # almost 50% on z9. The gain is smaller on z10, because being dual- | ||
40 | # issue z10 makes it improssible to eliminate the interlock condition: | ||
41 | # critial path is not long enough. Yet it spends ~24 cycles per byte | ||
42 | # processed with 128-bit key. | ||
43 | # | ||
44 | # Unlike previous version hardware support detection takes place only | ||
45 | # at the moment of key schedule setup, which is denoted in key->rounds. | ||
46 | # This is done, because deferred key setup can't be made MT-safe, not | ||
47 | # for key lengthes longer than 128 bits. | ||
48 | # | ||
49 | # Add AES_cbc_encrypt, which gives incredible performance improvement, | ||
50 | # it was measured to be ~6.6x. It's less than previously mentioned 8x, | ||
51 | # because software implementation was optimized. | ||
52 | |||
53 | $softonly=0; # allow hardware support | ||
54 | |||
55 | $t0="%r0"; $mask="%r0"; | ||
56 | $t1="%r1"; | ||
57 | $t2="%r2"; $inp="%r2"; | ||
58 | $t3="%r3"; $out="%r3"; $bits="%r3"; | ||
59 | $key="%r4"; | ||
60 | $i1="%r5"; | ||
61 | $i2="%r6"; | ||
62 | $i3="%r7"; | ||
63 | $s0="%r8"; | ||
64 | $s1="%r9"; | ||
65 | $s2="%r10"; | ||
66 | $s3="%r11"; | ||
67 | $tbl="%r12"; | ||
68 | $rounds="%r13"; | ||
69 | $ra="%r14"; | ||
70 | $sp="%r15"; | ||
71 | |||
72 | sub _data_word() | ||
73 | { my $i; | ||
74 | while(defined($i=shift)) { $code.=sprintf".long\t0x%08x,0x%08x\n",$i,$i; } | ||
75 | } | ||
76 | |||
77 | $code=<<___; | ||
78 | .text | ||
79 | |||
80 | .type AES_Te,\@object | ||
81 | .align 256 | ||
82 | AES_Te: | ||
83 | ___ | ||
84 | &_data_word( | ||
85 | 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, | ||
86 | 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554, | ||
87 | 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, | ||
88 | 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a, | ||
89 | 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, | ||
90 | 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b, | ||
91 | 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea, | ||
92 | 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b, | ||
93 | 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a, | ||
94 | 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f, | ||
95 | 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108, | ||
96 | 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f, | ||
97 | 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e, | ||
98 | 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5, | ||
99 | 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d, | ||
100 | 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f, | ||
101 | 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e, | ||
102 | 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb, | ||
103 | 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce, | ||
104 | 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497, | ||
105 | 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c, | ||
106 | 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed, | ||
107 | 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b, | ||
108 | 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a, | ||
109 | 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16, | ||
110 | 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594, | ||
111 | 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81, | ||
112 | 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3, | ||
113 | 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a, | ||
114 | 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504, | ||
115 | 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163, | ||
116 | 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d, | ||
117 | 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f, | ||
118 | 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739, | ||
119 | 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47, | ||
120 | 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395, | ||
121 | 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f, | ||
122 | 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883, | ||
123 | 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c, | ||
124 | 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76, | ||
125 | 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e, | ||
126 | 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4, | ||
127 | 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6, | ||
128 | 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b, | ||
129 | 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7, | ||
130 | 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0, | ||
131 | 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25, | ||
132 | 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818, | ||
133 | 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72, | ||
134 | 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651, | ||
135 | 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21, | ||
136 | 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85, | ||
137 | 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa, | ||
138 | 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12, | ||
139 | 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0, | ||
140 | 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9, | ||
141 | 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133, | ||
142 | 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7, | ||
143 | 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, | ||
144 | 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a, | ||
145 | 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, | ||
146 | 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8, | ||
147 | 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, | ||
148 | 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a); | ||
149 | $code.=<<___; | ||
150 | # Te4[256] | ||
151 | .byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 | ||
152 | .byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76 | ||
153 | .byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0 | ||
154 | .byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0 | ||
155 | .byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc | ||
156 | .byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15 | ||
157 | .byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a | ||
158 | .byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75 | ||
159 | .byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0 | ||
160 | .byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84 | ||
161 | .byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b | ||
162 | .byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf | ||
163 | .byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85 | ||
164 | .byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8 | ||
165 | .byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5 | ||
166 | .byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2 | ||
167 | .byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17 | ||
168 | .byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73 | ||
169 | .byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88 | ||
170 | .byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb | ||
171 | .byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c | ||
172 | .byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79 | ||
173 | .byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9 | ||
174 | .byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08 | ||
175 | .byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6 | ||
176 | .byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a | ||
177 | .byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e | ||
178 | .byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e | ||
179 | .byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94 | ||
180 | .byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf | ||
181 | .byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68 | ||
182 | .byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 | ||
183 | # rcon[] | ||
184 | .long 0x01000000, 0x02000000, 0x04000000, 0x08000000 | ||
185 | .long 0x10000000, 0x20000000, 0x40000000, 0x80000000 | ||
186 | .long 0x1B000000, 0x36000000, 0, 0, 0, 0, 0, 0 | ||
187 | .align 256 | ||
188 | .size AES_Te,.-AES_Te | ||
189 | |||
190 | # void AES_encrypt(const unsigned char *inp, unsigned char *out, | ||
191 | # const AES_KEY *key) { | ||
192 | .globl AES_encrypt | ||
193 | .type AES_encrypt,\@function | ||
194 | AES_encrypt: | ||
195 | ___ | ||
196 | $code.=<<___ if (!$softonly); | ||
197 | l %r0,240($key) | ||
198 | lhi %r1,16 | ||
199 | clr %r0,%r1 | ||
200 | jl .Lesoft | ||
201 | |||
202 | la %r1,0($key) | ||
203 | #la %r2,0($inp) | ||
204 | la %r4,0($out) | ||
205 | lghi %r3,16 # single block length | ||
206 | .long 0xb92e0042 # km %r4,%r2 | ||
207 | brc 1,.-4 # can this happen? | ||
208 | br %r14 | ||
209 | .align 64 | ||
210 | .Lesoft: | ||
211 | ___ | ||
212 | $code.=<<___; | ||
213 | stmg %r3,$ra,24($sp) | ||
214 | |||
215 | llgf $s0,0($inp) | ||
216 | llgf $s1,4($inp) | ||
217 | llgf $s2,8($inp) | ||
218 | llgf $s3,12($inp) | ||
219 | |||
220 | larl $tbl,AES_Te | ||
221 | bras $ra,_s390x_AES_encrypt | ||
222 | |||
223 | lg $out,24($sp) | ||
224 | st $s0,0($out) | ||
225 | st $s1,4($out) | ||
226 | st $s2,8($out) | ||
227 | st $s3,12($out) | ||
228 | |||
229 | lmg %r6,$ra,48($sp) | ||
230 | br $ra | ||
231 | .size AES_encrypt,.-AES_encrypt | ||
232 | |||
233 | .type _s390x_AES_encrypt,\@function | ||
234 | .align 16 | ||
235 | _s390x_AES_encrypt: | ||
236 | stg $ra,152($sp) | ||
237 | x $s0,0($key) | ||
238 | x $s1,4($key) | ||
239 | x $s2,8($key) | ||
240 | x $s3,12($key) | ||
241 | l $rounds,240($key) | ||
242 | llill $mask,`0xff<<3` | ||
243 | aghi $rounds,-1 | ||
244 | j .Lenc_loop | ||
245 | .align 16 | ||
246 | .Lenc_loop: | ||
247 | sllg $t1,$s0,`0+3` | ||
248 | srlg $t2,$s0,`8-3` | ||
249 | srlg $t3,$s0,`16-3` | ||
250 | srl $s0,`24-3` | ||
251 | nr $s0,$mask | ||
252 | ngr $t1,$mask | ||
253 | nr $t2,$mask | ||
254 | nr $t3,$mask | ||
255 | |||
256 | srlg $i1,$s1,`16-3` # i0 | ||
257 | sllg $i2,$s1,`0+3` | ||
258 | srlg $i3,$s1,`8-3` | ||
259 | srl $s1,`24-3` | ||
260 | nr $i1,$mask | ||
261 | nr $s1,$mask | ||
262 | ngr $i2,$mask | ||
263 | nr $i3,$mask | ||
264 | |||
265 | l $s0,0($s0,$tbl) # Te0[s0>>24] | ||
266 | l $t1,1($t1,$tbl) # Te3[s0>>0] | ||
267 | l $t2,2($t2,$tbl) # Te2[s0>>8] | ||
268 | l $t3,3($t3,$tbl) # Te1[s0>>16] | ||
269 | |||
270 | x $s0,3($i1,$tbl) # Te1[s1>>16] | ||
271 | l $s1,0($s1,$tbl) # Te0[s1>>24] | ||
272 | x $t2,1($i2,$tbl) # Te3[s1>>0] | ||
273 | x $t3,2($i3,$tbl) # Te2[s1>>8] | ||
274 | |||
275 | srlg $i1,$s2,`8-3` # i0 | ||
276 | srlg $i2,$s2,`16-3` # i1 | ||
277 | nr $i1,$mask | ||
278 | nr $i2,$mask | ||
279 | sllg $i3,$s2,`0+3` | ||
280 | srl $s2,`24-3` | ||
281 | nr $s2,$mask | ||
282 | ngr $i3,$mask | ||
283 | |||
284 | xr $s1,$t1 | ||
285 | srlg $ra,$s3,`8-3` # i1 | ||
286 | sllg $t1,$s3,`0+3` # i0 | ||
287 | nr $ra,$mask | ||
288 | la $key,16($key) | ||
289 | ngr $t1,$mask | ||
290 | |||
291 | x $s0,2($i1,$tbl) # Te2[s2>>8] | ||
292 | x $s1,3($i2,$tbl) # Te1[s2>>16] | ||
293 | l $s2,0($s2,$tbl) # Te0[s2>>24] | ||
294 | x $t3,1($i3,$tbl) # Te3[s2>>0] | ||
295 | |||
296 | srlg $i3,$s3,`16-3` # i2 | ||
297 | xr $s2,$t2 | ||
298 | srl $s3,`24-3` | ||
299 | nr $i3,$mask | ||
300 | nr $s3,$mask | ||
301 | |||
302 | x $s0,0($key) | ||
303 | x $s1,4($key) | ||
304 | x $s2,8($key) | ||
305 | x $t3,12($key) | ||
306 | |||
307 | x $s0,1($t1,$tbl) # Te3[s3>>0] | ||
308 | x $s1,2($ra,$tbl) # Te2[s3>>8] | ||
309 | x $s2,3($i3,$tbl) # Te1[s3>>16] | ||
310 | l $s3,0($s3,$tbl) # Te0[s3>>24] | ||
311 | xr $s3,$t3 | ||
312 | |||
313 | brct $rounds,.Lenc_loop | ||
314 | .align 16 | ||
315 | |||
316 | sllg $t1,$s0,`0+3` | ||
317 | srlg $t2,$s0,`8-3` | ||
318 | ngr $t1,$mask | ||
319 | srlg $t3,$s0,`16-3` | ||
320 | srl $s0,`24-3` | ||
321 | nr $s0,$mask | ||
322 | nr $t2,$mask | ||
323 | nr $t3,$mask | ||
324 | |||
325 | srlg $i1,$s1,`16-3` # i0 | ||
326 | sllg $i2,$s1,`0+3` | ||
327 | ngr $i2,$mask | ||
328 | srlg $i3,$s1,`8-3` | ||
329 | srl $s1,`24-3` | ||
330 | nr $i1,$mask | ||
331 | nr $s1,$mask | ||
332 | nr $i3,$mask | ||
333 | |||
334 | llgc $s0,2($s0,$tbl) # Te4[s0>>24] | ||
335 | llgc $t1,2($t1,$tbl) # Te4[s0>>0] | ||
336 | sll $s0,24 | ||
337 | llgc $t2,2($t2,$tbl) # Te4[s0>>8] | ||
338 | llgc $t3,2($t3,$tbl) # Te4[s0>>16] | ||
339 | sll $t2,8 | ||
340 | sll $t3,16 | ||
341 | |||
342 | llgc $i1,2($i1,$tbl) # Te4[s1>>16] | ||
343 | llgc $s1,2($s1,$tbl) # Te4[s1>>24] | ||
344 | llgc $i2,2($i2,$tbl) # Te4[s1>>0] | ||
345 | llgc $i3,2($i3,$tbl) # Te4[s1>>8] | ||
346 | sll $i1,16 | ||
347 | sll $s1,24 | ||
348 | sll $i3,8 | ||
349 | or $s0,$i1 | ||
350 | or $s1,$t1 | ||
351 | or $t2,$i2 | ||
352 | or $t3,$i3 | ||
353 | |||
354 | srlg $i1,$s2,`8-3` # i0 | ||
355 | srlg $i2,$s2,`16-3` # i1 | ||
356 | nr $i1,$mask | ||
357 | nr $i2,$mask | ||
358 | sllg $i3,$s2,`0+3` | ||
359 | srl $s2,`24-3` | ||
360 | ngr $i3,$mask | ||
361 | nr $s2,$mask | ||
362 | |||
363 | sllg $t1,$s3,`0+3` # i0 | ||
364 | srlg $ra,$s3,`8-3` # i1 | ||
365 | ngr $t1,$mask | ||
366 | |||
367 | llgc $i1,2($i1,$tbl) # Te4[s2>>8] | ||
368 | llgc $i2,2($i2,$tbl) # Te4[s2>>16] | ||
369 | sll $i1,8 | ||
370 | llgc $s2,2($s2,$tbl) # Te4[s2>>24] | ||
371 | llgc $i3,2($i3,$tbl) # Te4[s2>>0] | ||
372 | sll $i2,16 | ||
373 | nr $ra,$mask | ||
374 | sll $s2,24 | ||
375 | or $s0,$i1 | ||
376 | or $s1,$i2 | ||
377 | or $s2,$t2 | ||
378 | or $t3,$i3 | ||
379 | |||
380 | srlg $i3,$s3,`16-3` # i2 | ||
381 | srl $s3,`24-3` | ||
382 | nr $i3,$mask | ||
383 | nr $s3,$mask | ||
384 | |||
385 | l $t0,16($key) | ||
386 | l $t2,20($key) | ||
387 | |||
388 | llgc $i1,2($t1,$tbl) # Te4[s3>>0] | ||
389 | llgc $i2,2($ra,$tbl) # Te4[s3>>8] | ||
390 | llgc $i3,2($i3,$tbl) # Te4[s3>>16] | ||
391 | llgc $s3,2($s3,$tbl) # Te4[s3>>24] | ||
392 | sll $i2,8 | ||
393 | sll $i3,16 | ||
394 | sll $s3,24 | ||
395 | or $s0,$i1 | ||
396 | or $s1,$i2 | ||
397 | or $s2,$i3 | ||
398 | or $s3,$t3 | ||
399 | |||
400 | lg $ra,152($sp) | ||
401 | xr $s0,$t0 | ||
402 | xr $s1,$t2 | ||
403 | x $s2,24($key) | ||
404 | x $s3,28($key) | ||
405 | |||
406 | br $ra | ||
407 | .size _s390x_AES_encrypt,.-_s390x_AES_encrypt | ||
408 | ___ | ||
409 | |||
410 | $code.=<<___; | ||
411 | .type AES_Td,\@object | ||
412 | .align 256 | ||
413 | AES_Td: | ||
414 | ___ | ||
415 | &_data_word( | ||
416 | 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, | ||
417 | 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393, | ||
418 | 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, | ||
419 | 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f, | ||
420 | 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, | ||
421 | 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6, | ||
422 | 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da, | ||
423 | 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844, | ||
424 | 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd, | ||
425 | 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4, | ||
426 | 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45, | ||
427 | 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94, | ||
428 | 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7, | ||
429 | 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a, | ||
430 | 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5, | ||
431 | 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c, | ||
432 | 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1, | ||
433 | 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a, | ||
434 | 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75, | ||
435 | 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051, | ||
436 | 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46, | ||
437 | 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff, | ||
438 | 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77, | ||
439 | 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb, | ||
440 | 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000, | ||
441 | 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e, | ||
442 | 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927, | ||
443 | 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a, | ||
444 | 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e, | ||
445 | 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16, | ||
446 | 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d, | ||
447 | 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8, | ||
448 | 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd, | ||
449 | 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34, | ||
450 | 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163, | ||
451 | 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120, | ||
452 | 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d, | ||
453 | 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0, | ||
454 | 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422, | ||
455 | 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef, | ||
456 | 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36, | ||
457 | 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4, | ||
458 | 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662, | ||
459 | 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5, | ||
460 | 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3, | ||
461 | 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b, | ||
462 | 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8, | ||
463 | 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6, | ||
464 | 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6, | ||
465 | 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0, | ||
466 | 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815, | ||
467 | 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f, | ||
468 | 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df, | ||
469 | 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f, | ||
470 | 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e, | ||
471 | 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713, | ||
472 | 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89, | ||
473 | 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c, | ||
474 | 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, | ||
475 | 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86, | ||
476 | 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, | ||
477 | 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541, | ||
478 | 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, | ||
479 | 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742); | ||
480 | $code.=<<___; | ||
481 | # Td4[256] | ||
482 | .byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 | ||
483 | .byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb | ||
484 | .byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87 | ||
485 | .byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb | ||
486 | .byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d | ||
487 | .byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e | ||
488 | .byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2 | ||
489 | .byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25 | ||
490 | .byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16 | ||
491 | .byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92 | ||
492 | .byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda | ||
493 | .byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84 | ||
494 | .byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a | ||
495 | .byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06 | ||
496 | .byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02 | ||
497 | .byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b | ||
498 | .byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea | ||
499 | .byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73 | ||
500 | .byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85 | ||
501 | .byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e | ||
502 | .byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89 | ||
503 | .byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b | ||
504 | .byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20 | ||
505 | .byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4 | ||
506 | .byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31 | ||
507 | .byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f | ||
508 | .byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d | ||
509 | .byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef | ||
510 | .byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0 | ||
511 | .byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61 | ||
512 | .byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26 | ||
513 | .byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d | ||
514 | .size AES_Td,.-AES_Td | ||
515 | |||
516 | # void AES_decrypt(const unsigned char *inp, unsigned char *out, | ||
517 | # const AES_KEY *key) { | ||
518 | .globl AES_decrypt | ||
519 | .type AES_decrypt,\@function | ||
520 | AES_decrypt: | ||
521 | ___ | ||
522 | $code.=<<___ if (!$softonly); | ||
523 | l %r0,240($key) | ||
524 | lhi %r1,16 | ||
525 | clr %r0,%r1 | ||
526 | jl .Ldsoft | ||
527 | |||
528 | la %r1,0($key) | ||
529 | #la %r2,0($inp) | ||
530 | la %r4,0($out) | ||
531 | lghi %r3,16 # single block length | ||
532 | .long 0xb92e0042 # km %r4,%r2 | ||
533 | brc 1,.-4 # can this happen? | ||
534 | br %r14 | ||
535 | .align 64 | ||
536 | .Ldsoft: | ||
537 | ___ | ||
538 | $code.=<<___; | ||
539 | stmg %r3,$ra,24($sp) | ||
540 | |||
541 | llgf $s0,0($inp) | ||
542 | llgf $s1,4($inp) | ||
543 | llgf $s2,8($inp) | ||
544 | llgf $s3,12($inp) | ||
545 | |||
546 | larl $tbl,AES_Td | ||
547 | bras $ra,_s390x_AES_decrypt | ||
548 | |||
549 | lg $out,24($sp) | ||
550 | st $s0,0($out) | ||
551 | st $s1,4($out) | ||
552 | st $s2,8($out) | ||
553 | st $s3,12($out) | ||
554 | |||
555 | lmg %r6,$ra,48($sp) | ||
556 | br $ra | ||
557 | .size AES_decrypt,.-AES_decrypt | ||
558 | |||
559 | .type _s390x_AES_decrypt,\@function | ||
560 | .align 16 | ||
561 | _s390x_AES_decrypt: | ||
562 | stg $ra,152($sp) | ||
563 | x $s0,0($key) | ||
564 | x $s1,4($key) | ||
565 | x $s2,8($key) | ||
566 | x $s3,12($key) | ||
567 | l $rounds,240($key) | ||
568 | llill $mask,`0xff<<3` | ||
569 | aghi $rounds,-1 | ||
570 | j .Ldec_loop | ||
571 | .align 16 | ||
572 | .Ldec_loop: | ||
573 | srlg $t1,$s0,`16-3` | ||
574 | srlg $t2,$s0,`8-3` | ||
575 | sllg $t3,$s0,`0+3` | ||
576 | srl $s0,`24-3` | ||
577 | nr $s0,$mask | ||
578 | nr $t1,$mask | ||
579 | nr $t2,$mask | ||
580 | ngr $t3,$mask | ||
581 | |||
582 | sllg $i1,$s1,`0+3` # i0 | ||
583 | srlg $i2,$s1,`16-3` | ||
584 | srlg $i3,$s1,`8-3` | ||
585 | srl $s1,`24-3` | ||
586 | ngr $i1,$mask | ||
587 | nr $s1,$mask | ||
588 | nr $i2,$mask | ||
589 | nr $i3,$mask | ||
590 | |||
591 | l $s0,0($s0,$tbl) # Td0[s0>>24] | ||
592 | l $t1,3($t1,$tbl) # Td1[s0>>16] | ||
593 | l $t2,2($t2,$tbl) # Td2[s0>>8] | ||
594 | l $t3,1($t3,$tbl) # Td3[s0>>0] | ||
595 | |||
596 | x $s0,1($i1,$tbl) # Td3[s1>>0] | ||
597 | l $s1,0($s1,$tbl) # Td0[s1>>24] | ||
598 | x $t2,3($i2,$tbl) # Td1[s1>>16] | ||
599 | x $t3,2($i3,$tbl) # Td2[s1>>8] | ||
600 | |||
601 | srlg $i1,$s2,`8-3` # i0 | ||
602 | sllg $i2,$s2,`0+3` # i1 | ||
603 | srlg $i3,$s2,`16-3` | ||
604 | srl $s2,`24-3` | ||
605 | nr $i1,$mask | ||
606 | ngr $i2,$mask | ||
607 | nr $s2,$mask | ||
608 | nr $i3,$mask | ||
609 | |||
610 | xr $s1,$t1 | ||
611 | srlg $ra,$s3,`8-3` # i1 | ||
612 | srlg $t1,$s3,`16-3` # i0 | ||
613 | nr $ra,$mask | ||
614 | la $key,16($key) | ||
615 | nr $t1,$mask | ||
616 | |||
617 | x $s0,2($i1,$tbl) # Td2[s2>>8] | ||
618 | x $s1,1($i2,$tbl) # Td3[s2>>0] | ||
619 | l $s2,0($s2,$tbl) # Td0[s2>>24] | ||
620 | x $t3,3($i3,$tbl) # Td1[s2>>16] | ||
621 | |||
622 | sllg $i3,$s3,`0+3` # i2 | ||
623 | srl $s3,`24-3` | ||
624 | ngr $i3,$mask | ||
625 | nr $s3,$mask | ||
626 | |||
627 | xr $s2,$t2 | ||
628 | x $s0,0($key) | ||
629 | x $s1,4($key) | ||
630 | x $s2,8($key) | ||
631 | x $t3,12($key) | ||
632 | |||
633 | x $s0,3($t1,$tbl) # Td1[s3>>16] | ||
634 | x $s1,2($ra,$tbl) # Td2[s3>>8] | ||
635 | x $s2,1($i3,$tbl) # Td3[s3>>0] | ||
636 | l $s3,0($s3,$tbl) # Td0[s3>>24] | ||
637 | xr $s3,$t3 | ||
638 | |||
639 | brct $rounds,.Ldec_loop | ||
640 | .align 16 | ||
641 | |||
642 | l $t1,`2048+0`($tbl) # prefetch Td4 | ||
643 | l $t2,`2048+64`($tbl) | ||
644 | l $t3,`2048+128`($tbl) | ||
645 | l $i1,`2048+192`($tbl) | ||
646 | llill $mask,0xff | ||
647 | |||
648 | srlg $i3,$s0,24 # i0 | ||
649 | srlg $t1,$s0,16 | ||
650 | srlg $t2,$s0,8 | ||
651 | nr $s0,$mask # i3 | ||
652 | nr $t1,$mask | ||
653 | |||
654 | srlg $i1,$s1,24 | ||
655 | nr $t2,$mask | ||
656 | srlg $i2,$s1,16 | ||
657 | srlg $ra,$s1,8 | ||
658 | nr $s1,$mask # i0 | ||
659 | nr $i2,$mask | ||
660 | nr $ra,$mask | ||
661 | |||
662 | llgc $i3,2048($i3,$tbl) # Td4[s0>>24] | ||
663 | llgc $t1,2048($t1,$tbl) # Td4[s0>>16] | ||
664 | llgc $t2,2048($t2,$tbl) # Td4[s0>>8] | ||
665 | sll $t1,16 | ||
666 | llgc $t3,2048($s0,$tbl) # Td4[s0>>0] | ||
667 | sllg $s0,$i3,24 | ||
668 | sll $t2,8 | ||
669 | |||
670 | llgc $s1,2048($s1,$tbl) # Td4[s1>>0] | ||
671 | llgc $i1,2048($i1,$tbl) # Td4[s1>>24] | ||
672 | llgc $i2,2048($i2,$tbl) # Td4[s1>>16] | ||
673 | sll $i1,24 | ||
674 | llgc $i3,2048($ra,$tbl) # Td4[s1>>8] | ||
675 | sll $i2,16 | ||
676 | sll $i3,8 | ||
677 | or $s0,$s1 | ||
678 | or $t1,$i1 | ||
679 | or $t2,$i2 | ||
680 | or $t3,$i3 | ||
681 | |||
682 | srlg $i1,$s2,8 # i0 | ||
683 | srlg $i2,$s2,24 | ||
684 | srlg $i3,$s2,16 | ||
685 | nr $s2,$mask # i1 | ||
686 | nr $i1,$mask | ||
687 | nr $i3,$mask | ||
688 | llgc $i1,2048($i1,$tbl) # Td4[s2>>8] | ||
689 | llgc $s1,2048($s2,$tbl) # Td4[s2>>0] | ||
690 | llgc $i2,2048($i2,$tbl) # Td4[s2>>24] | ||
691 | llgc $i3,2048($i3,$tbl) # Td4[s2>>16] | ||
692 | sll $i1,8 | ||
693 | sll $i2,24 | ||
694 | or $s0,$i1 | ||
695 | sll $i3,16 | ||
696 | or $t2,$i2 | ||
697 | or $t3,$i3 | ||
698 | |||
699 | srlg $i1,$s3,16 # i0 | ||
700 | srlg $i2,$s3,8 # i1 | ||
701 | srlg $i3,$s3,24 | ||
702 | nr $s3,$mask # i2 | ||
703 | nr $i1,$mask | ||
704 | nr $i2,$mask | ||
705 | |||
706 | lg $ra,152($sp) | ||
707 | or $s1,$t1 | ||
708 | l $t0,16($key) | ||
709 | l $t1,20($key) | ||
710 | |||
711 | llgc $i1,2048($i1,$tbl) # Td4[s3>>16] | ||
712 | llgc $i2,2048($i2,$tbl) # Td4[s3>>8] | ||
713 | sll $i1,16 | ||
714 | llgc $s2,2048($s3,$tbl) # Td4[s3>>0] | ||
715 | llgc $s3,2048($i3,$tbl) # Td4[s3>>24] | ||
716 | sll $i2,8 | ||
717 | sll $s3,24 | ||
718 | or $s0,$i1 | ||
719 | or $s1,$i2 | ||
720 | or $s2,$t2 | ||
721 | or $s3,$t3 | ||
722 | |||
723 | xr $s0,$t0 | ||
724 | xr $s1,$t1 | ||
725 | x $s2,24($key) | ||
726 | x $s3,28($key) | ||
727 | |||
728 | br $ra | ||
729 | .size _s390x_AES_decrypt,.-_s390x_AES_decrypt | ||
730 | ___ | ||
731 | |||
732 | $code.=<<___; | ||
733 | # void AES_set_encrypt_key(const unsigned char *in, int bits, | ||
734 | # AES_KEY *key) { | ||
735 | .globl AES_set_encrypt_key | ||
736 | .type AES_set_encrypt_key,\@function | ||
737 | .align 16 | ||
738 | AES_set_encrypt_key: | ||
739 | lghi $t0,0 | ||
740 | clgr $inp,$t0 | ||
741 | je .Lminus1 | ||
742 | clgr $key,$t0 | ||
743 | je .Lminus1 | ||
744 | |||
745 | lghi $t0,128 | ||
746 | clr $bits,$t0 | ||
747 | je .Lproceed | ||
748 | lghi $t0,192 | ||
749 | clr $bits,$t0 | ||
750 | je .Lproceed | ||
751 | lghi $t0,256 | ||
752 | clr $bits,$t0 | ||
753 | je .Lproceed | ||
754 | lghi %r2,-2 | ||
755 | br %r14 | ||
756 | |||
757 | .align 16 | ||
758 | .Lproceed: | ||
759 | ___ | ||
760 | $code.=<<___ if (!$softonly); | ||
761 | # convert bits to km code, [128,192,256]->[18,19,20] | ||
762 | lhi %r5,-128 | ||
763 | lhi %r0,18 | ||
764 | ar %r5,$bits | ||
765 | srl %r5,6 | ||
766 | ar %r5,%r0 | ||
767 | |||
768 | lghi %r0,0 # query capability vector | ||
769 | la %r1,16($sp) | ||
770 | .long 0xb92f0042 # kmc %r4,%r2 | ||
771 | |||
772 | llihh %r1,0x8000 | ||
773 | srlg %r1,%r1,0(%r5) | ||
774 | ng %r1,16($sp) | ||
775 | jz .Lekey_internal | ||
776 | |||
777 | lmg %r0,%r1,0($inp) # just copy 128 bits... | ||
778 | stmg %r0,%r1,0($key) | ||
779 | lhi %r0,192 | ||
780 | cr $bits,%r0 | ||
781 | jl 1f | ||
782 | lg %r1,16($inp) | ||
783 | stg %r1,16($key) | ||
784 | je 1f | ||
785 | lg %r1,24($inp) | ||
786 | stg %r1,24($key) | ||
787 | 1: st $bits,236($key) # save bits | ||
788 | st %r5,240($key) # save km code | ||
789 | lghi %r2,0 | ||
790 | br %r14 | ||
791 | ___ | ||
792 | $code.=<<___; | ||
793 | .align 16 | ||
794 | .Lekey_internal: | ||
795 | stmg %r6,%r13,48($sp) # all non-volatile regs | ||
796 | |||
797 | larl $tbl,AES_Te+2048 | ||
798 | |||
799 | llgf $s0,0($inp) | ||
800 | llgf $s1,4($inp) | ||
801 | llgf $s2,8($inp) | ||
802 | llgf $s3,12($inp) | ||
803 | st $s0,0($key) | ||
804 | st $s1,4($key) | ||
805 | st $s2,8($key) | ||
806 | st $s3,12($key) | ||
807 | lghi $t0,128 | ||
808 | cr $bits,$t0 | ||
809 | jne .Lnot128 | ||
810 | |||
811 | llill $mask,0xff | ||
812 | lghi $t3,0 # i=0 | ||
813 | lghi $rounds,10 | ||
814 | st $rounds,240($key) | ||
815 | |||
816 | llgfr $t2,$s3 # temp=rk[3] | ||
817 | srlg $i1,$s3,8 | ||
818 | srlg $i2,$s3,16 | ||
819 | srlg $i3,$s3,24 | ||
820 | nr $t2,$mask | ||
821 | nr $i1,$mask | ||
822 | nr $i2,$mask | ||
823 | |||
824 | .align 16 | ||
825 | .L128_loop: | ||
826 | la $t2,0($t2,$tbl) | ||
827 | la $i1,0($i1,$tbl) | ||
828 | la $i2,0($i2,$tbl) | ||
829 | la $i3,0($i3,$tbl) | ||
830 | icm $t2,2,0($t2) # Te4[rk[3]>>0]<<8 | ||
831 | icm $t2,4,0($i1) # Te4[rk[3]>>8]<<16 | ||
832 | icm $t2,8,0($i2) # Te4[rk[3]>>16]<<24 | ||
833 | icm $t2,1,0($i3) # Te4[rk[3]>>24] | ||
834 | x $t2,256($t3,$tbl) # rcon[i] | ||
835 | xr $s0,$t2 # rk[4]=rk[0]^... | ||
836 | xr $s1,$s0 # rk[5]=rk[1]^rk[4] | ||
837 | xr $s2,$s1 # rk[6]=rk[2]^rk[5] | ||
838 | xr $s3,$s2 # rk[7]=rk[3]^rk[6] | ||
839 | |||
840 | llgfr $t2,$s3 # temp=rk[3] | ||
841 | srlg $i1,$s3,8 | ||
842 | srlg $i2,$s3,16 | ||
843 | nr $t2,$mask | ||
844 | nr $i1,$mask | ||
845 | srlg $i3,$s3,24 | ||
846 | nr $i2,$mask | ||
847 | |||
848 | st $s0,16($key) | ||
849 | st $s1,20($key) | ||
850 | st $s2,24($key) | ||
851 | st $s3,28($key) | ||
852 | la $key,16($key) # key+=4 | ||
853 | la $t3,4($t3) # i++ | ||
854 | brct $rounds,.L128_loop | ||
855 | lghi %r2,0 | ||
856 | lmg %r6,%r13,48($sp) | ||
857 | br $ra | ||
858 | |||
859 | .align 16 | ||
860 | .Lnot128: | ||
861 | llgf $t0,16($inp) | ||
862 | llgf $t1,20($inp) | ||
863 | st $t0,16($key) | ||
864 | st $t1,20($key) | ||
865 | lghi $t0,192 | ||
866 | cr $bits,$t0 | ||
867 | jne .Lnot192 | ||
868 | |||
869 | llill $mask,0xff | ||
870 | lghi $t3,0 # i=0 | ||
871 | lghi $rounds,12 | ||
872 | st $rounds,240($key) | ||
873 | lghi $rounds,8 | ||
874 | |||
875 | srlg $i1,$t1,8 | ||
876 | srlg $i2,$t1,16 | ||
877 | srlg $i3,$t1,24 | ||
878 | nr $t1,$mask | ||
879 | nr $i1,$mask | ||
880 | nr $i2,$mask | ||
881 | |||
882 | .align 16 | ||
883 | .L192_loop: | ||
884 | la $t1,0($t1,$tbl) | ||
885 | la $i1,0($i1,$tbl) | ||
886 | la $i2,0($i2,$tbl) | ||
887 | la $i3,0($i3,$tbl) | ||
888 | icm $t1,2,0($t1) # Te4[rk[5]>>0]<<8 | ||
889 | icm $t1,4,0($i1) # Te4[rk[5]>>8]<<16 | ||
890 | icm $t1,8,0($i2) # Te4[rk[5]>>16]<<24 | ||
891 | icm $t1,1,0($i3) # Te4[rk[5]>>24] | ||
892 | x $t1,256($t3,$tbl) # rcon[i] | ||
893 | xr $s0,$t1 # rk[6]=rk[0]^... | ||
894 | xr $s1,$s0 # rk[7]=rk[1]^rk[6] | ||
895 | xr $s2,$s1 # rk[8]=rk[2]^rk[7] | ||
896 | xr $s3,$s2 # rk[9]=rk[3]^rk[8] | ||
897 | |||
898 | st $s0,24($key) | ||
899 | st $s1,28($key) | ||
900 | st $s2,32($key) | ||
901 | st $s3,36($key) | ||
902 | brct $rounds,.L192_continue | ||
903 | lghi %r2,0 | ||
904 | lmg %r6,%r13,48($sp) | ||
905 | br $ra | ||
906 | |||
907 | .align 16 | ||
908 | .L192_continue: | ||
909 | lgr $t1,$s3 | ||
910 | x $t1,16($key) # rk[10]=rk[4]^rk[9] | ||
911 | st $t1,40($key) | ||
912 | x $t1,20($key) # rk[11]=rk[5]^rk[10] | ||
913 | st $t1,44($key) | ||
914 | |||
915 | srlg $i1,$t1,8 | ||
916 | srlg $i2,$t1,16 | ||
917 | srlg $i3,$t1,24 | ||
918 | nr $t1,$mask | ||
919 | nr $i1,$mask | ||
920 | nr $i2,$mask | ||
921 | |||
922 | la $key,24($key) # key+=6 | ||
923 | la $t3,4($t3) # i++ | ||
924 | j .L192_loop | ||
925 | |||
926 | .align 16 | ||
927 | .Lnot192: | ||
928 | llgf $t0,24($inp) | ||
929 | llgf $t1,28($inp) | ||
930 | st $t0,24($key) | ||
931 | st $t1,28($key) | ||
932 | llill $mask,0xff | ||
933 | lghi $t3,0 # i=0 | ||
934 | lghi $rounds,14 | ||
935 | st $rounds,240($key) | ||
936 | lghi $rounds,7 | ||
937 | |||
938 | srlg $i1,$t1,8 | ||
939 | srlg $i2,$t1,16 | ||
940 | srlg $i3,$t1,24 | ||
941 | nr $t1,$mask | ||
942 | nr $i1,$mask | ||
943 | nr $i2,$mask | ||
944 | |||
945 | .align 16 | ||
946 | .L256_loop: | ||
947 | la $t1,0($t1,$tbl) | ||
948 | la $i1,0($i1,$tbl) | ||
949 | la $i2,0($i2,$tbl) | ||
950 | la $i3,0($i3,$tbl) | ||
951 | icm $t1,2,0($t1) # Te4[rk[7]>>0]<<8 | ||
952 | icm $t1,4,0($i1) # Te4[rk[7]>>8]<<16 | ||
953 | icm $t1,8,0($i2) # Te4[rk[7]>>16]<<24 | ||
954 | icm $t1,1,0($i3) # Te4[rk[7]>>24] | ||
955 | x $t1,256($t3,$tbl) # rcon[i] | ||
956 | xr $s0,$t1 # rk[8]=rk[0]^... | ||
957 | xr $s1,$s0 # rk[9]=rk[1]^rk[8] | ||
958 | xr $s2,$s1 # rk[10]=rk[2]^rk[9] | ||
959 | xr $s3,$s2 # rk[11]=rk[3]^rk[10] | ||
960 | st $s0,32($key) | ||
961 | st $s1,36($key) | ||
962 | st $s2,40($key) | ||
963 | st $s3,44($key) | ||
964 | brct $rounds,.L256_continue | ||
965 | lghi %r2,0 | ||
966 | lmg %r6,%r13,48($sp) | ||
967 | br $ra | ||
968 | |||
969 | .align 16 | ||
970 | .L256_continue: | ||
971 | lgr $t1,$s3 # temp=rk[11] | ||
972 | srlg $i1,$s3,8 | ||
973 | srlg $i2,$s3,16 | ||
974 | srlg $i3,$s3,24 | ||
975 | nr $t1,$mask | ||
976 | nr $i1,$mask | ||
977 | nr $i2,$mask | ||
978 | la $t1,0($t1,$tbl) | ||
979 | la $i1,0($i1,$tbl) | ||
980 | la $i2,0($i2,$tbl) | ||
981 | la $i3,0($i3,$tbl) | ||
982 | llgc $t1,0($t1) # Te4[rk[11]>>0] | ||
983 | icm $t1,2,0($i1) # Te4[rk[11]>>8]<<8 | ||
984 | icm $t1,4,0($i2) # Te4[rk[11]>>16]<<16 | ||
985 | icm $t1,8,0($i3) # Te4[rk[11]>>24]<<24 | ||
986 | x $t1,16($key) # rk[12]=rk[4]^... | ||
987 | st $t1,48($key) | ||
988 | x $t1,20($key) # rk[13]=rk[5]^rk[12] | ||
989 | st $t1,52($key) | ||
990 | x $t1,24($key) # rk[14]=rk[6]^rk[13] | ||
991 | st $t1,56($key) | ||
992 | x $t1,28($key) # rk[15]=rk[7]^rk[14] | ||
993 | st $t1,60($key) | ||
994 | |||
995 | srlg $i1,$t1,8 | ||
996 | srlg $i2,$t1,16 | ||
997 | srlg $i3,$t1,24 | ||
998 | nr $t1,$mask | ||
999 | nr $i1,$mask | ||
1000 | nr $i2,$mask | ||
1001 | |||
1002 | la $key,32($key) # key+=8 | ||
1003 | la $t3,4($t3) # i++ | ||
1004 | j .L256_loop | ||
1005 | |||
1006 | .Lminus1: | ||
1007 | lghi %r2,-1 | ||
1008 | br $ra | ||
1009 | .size AES_set_encrypt_key,.-AES_set_encrypt_key | ||
1010 | |||
1011 | # void AES_set_decrypt_key(const unsigned char *in, int bits, | ||
1012 | # AES_KEY *key) { | ||
1013 | .globl AES_set_decrypt_key | ||
1014 | .type AES_set_decrypt_key,\@function | ||
1015 | .align 16 | ||
1016 | AES_set_decrypt_key: | ||
1017 | stg $key,32($sp) # I rely on AES_set_encrypt_key to | ||
1018 | stg $ra,112($sp) # save non-volatile registers! | ||
1019 | bras $ra,AES_set_encrypt_key | ||
1020 | lg $key,32($sp) | ||
1021 | lg $ra,112($sp) | ||
1022 | ltgr %r2,%r2 | ||
1023 | bnzr $ra | ||
1024 | ___ | ||
1025 | $code.=<<___ if (!$softonly); | ||
1026 | l $t0,240($key) | ||
1027 | lhi $t1,16 | ||
1028 | cr $t0,$t1 | ||
1029 | jl .Lgo | ||
1030 | oill $t0,0x80 # set "decrypt" bit | ||
1031 | st $t0,240($key) | ||
1032 | br $ra | ||
1033 | |||
1034 | .align 16 | ||
1035 | .Ldkey_internal: | ||
1036 | stg $key,32($sp) | ||
1037 | stg $ra,40($sp) | ||
1038 | bras $ra,.Lekey_internal | ||
1039 | lg $key,32($sp) | ||
1040 | lg $ra,40($sp) | ||
1041 | ___ | ||
1042 | $code.=<<___; | ||
1043 | |||
1044 | .Lgo: llgf $rounds,240($key) | ||
1045 | la $i1,0($key) | ||
1046 | sllg $i2,$rounds,4 | ||
1047 | la $i2,0($i2,$key) | ||
1048 | srl $rounds,1 | ||
1049 | lghi $t1,-16 | ||
1050 | |||
1051 | .align 16 | ||
1052 | .Linv: lmg $s0,$s1,0($i1) | ||
1053 | lmg $s2,$s3,0($i2) | ||
1054 | stmg $s0,$s1,0($i2) | ||
1055 | stmg $s2,$s3,0($i1) | ||
1056 | la $i1,16($i1) | ||
1057 | la $i2,0($t1,$i2) | ||
1058 | brct $rounds,.Linv | ||
1059 | ___ | ||
1060 | $mask80=$i1; | ||
1061 | $mask1b=$i2; | ||
1062 | $maskfe=$i3; | ||
1063 | $code.=<<___; | ||
1064 | llgf $rounds,240($key) | ||
1065 | aghi $rounds,-1 | ||
1066 | sll $rounds,2 # (rounds-1)*4 | ||
1067 | llilh $mask80,0x8080 | ||
1068 | llilh $mask1b,0x1b1b | ||
1069 | llilh $maskfe,0xfefe | ||
1070 | oill $mask80,0x8080 | ||
1071 | oill $mask1b,0x1b1b | ||
1072 | oill $maskfe,0xfefe | ||
1073 | |||
1074 | .align 16 | ||
1075 | .Lmix: l $s0,16($key) # tp1 | ||
1076 | lr $s1,$s0 | ||
1077 | ngr $s1,$mask80 | ||
1078 | srlg $t1,$s1,7 | ||
1079 | slr $s1,$t1 | ||
1080 | nr $s1,$mask1b | ||
1081 | sllg $t1,$s0,1 | ||
1082 | nr $t1,$maskfe | ||
1083 | xr $s1,$t1 # tp2 | ||
1084 | |||
1085 | lr $s2,$s1 | ||
1086 | ngr $s2,$mask80 | ||
1087 | srlg $t1,$s2,7 | ||
1088 | slr $s2,$t1 | ||
1089 | nr $s2,$mask1b | ||
1090 | sllg $t1,$s1,1 | ||
1091 | nr $t1,$maskfe | ||
1092 | xr $s2,$t1 # tp4 | ||
1093 | |||
1094 | lr $s3,$s2 | ||
1095 | ngr $s3,$mask80 | ||
1096 | srlg $t1,$s3,7 | ||
1097 | slr $s3,$t1 | ||
1098 | nr $s3,$mask1b | ||
1099 | sllg $t1,$s2,1 | ||
1100 | nr $t1,$maskfe | ||
1101 | xr $s3,$t1 # tp8 | ||
1102 | |||
1103 | xr $s1,$s0 # tp2^tp1 | ||
1104 | xr $s2,$s0 # tp4^tp1 | ||
1105 | rll $s0,$s0,24 # = ROTATE(tp1,8) | ||
1106 | xr $s2,$s3 # ^=tp8 | ||
1107 | xr $s0,$s1 # ^=tp2^tp1 | ||
1108 | xr $s1,$s3 # tp2^tp1^tp8 | ||
1109 | xr $s0,$s2 # ^=tp4^tp1^tp8 | ||
1110 | rll $s1,$s1,8 | ||
1111 | rll $s2,$s2,16 | ||
1112 | xr $s0,$s1 # ^= ROTATE(tp8^tp2^tp1,24) | ||
1113 | rll $s3,$s3,24 | ||
1114 | xr $s0,$s2 # ^= ROTATE(tp8^tp4^tp1,16) | ||
1115 | xr $s0,$s3 # ^= ROTATE(tp8,8) | ||
1116 | |||
1117 | st $s0,16($key) | ||
1118 | la $key,4($key) | ||
1119 | brct $rounds,.Lmix | ||
1120 | |||
1121 | lmg %r6,%r13,48($sp)# as was saved by AES_set_encrypt_key! | ||
1122 | lghi %r2,0 | ||
1123 | br $ra | ||
1124 | .size AES_set_decrypt_key,.-AES_set_decrypt_key | ||
1125 | ___ | ||
1126 | |||
1127 | #void AES_cbc_encrypt(const unsigned char *in, unsigned char *out, | ||
1128 | # size_t length, const AES_KEY *key, | ||
1129 | # unsigned char *ivec, const int enc) | ||
1130 | { | ||
1131 | my $inp="%r2"; | ||
1132 | my $out="%r4"; # length and out are swapped | ||
1133 | my $len="%r3"; | ||
1134 | my $key="%r5"; | ||
1135 | my $ivp="%r6"; | ||
1136 | |||
1137 | $code.=<<___; | ||
1138 | .globl AES_cbc_encrypt | ||
1139 | .type AES_cbc_encrypt,\@function | ||
1140 | .align 16 | ||
1141 | AES_cbc_encrypt: | ||
1142 | xgr %r3,%r4 # flip %r3 and %r4, out and len | ||
1143 | xgr %r4,%r3 | ||
1144 | xgr %r3,%r4 | ||
1145 | ___ | ||
1146 | $code.=<<___ if (!$softonly); | ||
1147 | lhi %r0,16 | ||
1148 | cl %r0,240($key) | ||
1149 | jh .Lcbc_software | ||
1150 | |||
1151 | lg %r0,0($ivp) # copy ivec | ||
1152 | lg %r1,8($ivp) | ||
1153 | stmg %r0,%r1,16($sp) | ||
1154 | lmg %r0,%r1,0($key) # copy key, cover 256 bit | ||
1155 | stmg %r0,%r1,32($sp) | ||
1156 | lmg %r0,%r1,16($key) | ||
1157 | stmg %r0,%r1,48($sp) | ||
1158 | l %r0,240($key) # load kmc code | ||
1159 | lghi $key,15 # res=len%16, len-=res; | ||
1160 | ngr $key,$len | ||
1161 | slgr $len,$key | ||
1162 | la %r1,16($sp) # parameter block - ivec || key | ||
1163 | jz .Lkmc_truncated | ||
1164 | .long 0xb92f0042 # kmc %r4,%r2 | ||
1165 | brc 1,.-4 # pay attention to "partial completion" | ||
1166 | ltr $key,$key | ||
1167 | jnz .Lkmc_truncated | ||
1168 | .Lkmc_done: | ||
1169 | lmg %r0,%r1,16($sp) # copy ivec to caller | ||
1170 | stg %r0,0($ivp) | ||
1171 | stg %r1,8($ivp) | ||
1172 | br $ra | ||
1173 | .align 16 | ||
1174 | .Lkmc_truncated: | ||
1175 | ahi $key,-1 # it's the way it's encoded in mvc | ||
1176 | tmll %r0,0x80 | ||
1177 | jnz .Lkmc_truncated_dec | ||
1178 | lghi %r1,0 | ||
1179 | stg %r1,128($sp) | ||
1180 | stg %r1,136($sp) | ||
1181 | bras %r1,1f | ||
1182 | mvc 128(1,$sp),0($inp) | ||
1183 | 1: ex $key,0(%r1) | ||
1184 | la %r1,16($sp) # restore parameter block | ||
1185 | la $inp,128($sp) | ||
1186 | lghi $len,16 | ||
1187 | .long 0xb92f0042 # kmc %r4,%r2 | ||
1188 | j .Lkmc_done | ||
1189 | .align 16 | ||
1190 | .Lkmc_truncated_dec: | ||
1191 | stg $out,64($sp) | ||
1192 | la $out,128($sp) | ||
1193 | lghi $len,16 | ||
1194 | .long 0xb92f0042 # kmc %r4,%r2 | ||
1195 | lg $out,64($sp) | ||
1196 | bras %r1,2f | ||
1197 | mvc 0(1,$out),128($sp) | ||
1198 | 2: ex $key,0(%r1) | ||
1199 | j .Lkmc_done | ||
1200 | .align 16 | ||
1201 | .Lcbc_software: | ||
1202 | ___ | ||
1203 | $code.=<<___; | ||
1204 | stmg $key,$ra,40($sp) | ||
1205 | lhi %r0,0 | ||
1206 | cl %r0,164($sp) | ||
1207 | je .Lcbc_decrypt | ||
1208 | |||
1209 | larl $tbl,AES_Te | ||
1210 | |||
1211 | llgf $s0,0($ivp) | ||
1212 | llgf $s1,4($ivp) | ||
1213 | llgf $s2,8($ivp) | ||
1214 | llgf $s3,12($ivp) | ||
1215 | |||
1216 | lghi $t0,16 | ||
1217 | slgr $len,$t0 | ||
1218 | brc 4,.Lcbc_enc_tail # if borrow | ||
1219 | .Lcbc_enc_loop: | ||
1220 | stmg $inp,$out,16($sp) | ||
1221 | x $s0,0($inp) | ||
1222 | x $s1,4($inp) | ||
1223 | x $s2,8($inp) | ||
1224 | x $s3,12($inp) | ||
1225 | lgr %r4,$key | ||
1226 | |||
1227 | bras $ra,_s390x_AES_encrypt | ||
1228 | |||
1229 | lmg $inp,$key,16($sp) | ||
1230 | st $s0,0($out) | ||
1231 | st $s1,4($out) | ||
1232 | st $s2,8($out) | ||
1233 | st $s3,12($out) | ||
1234 | |||
1235 | la $inp,16($inp) | ||
1236 | la $out,16($out) | ||
1237 | lghi $t0,16 | ||
1238 | ltgr $len,$len | ||
1239 | jz .Lcbc_enc_done | ||
1240 | slgr $len,$t0 | ||
1241 | brc 4,.Lcbc_enc_tail # if borrow | ||
1242 | j .Lcbc_enc_loop | ||
1243 | .align 16 | ||
1244 | .Lcbc_enc_done: | ||
1245 | lg $ivp,48($sp) | ||
1246 | st $s0,0($ivp) | ||
1247 | st $s1,4($ivp) | ||
1248 | st $s2,8($ivp) | ||
1249 | st $s3,12($ivp) | ||
1250 | |||
1251 | lmg %r7,$ra,56($sp) | ||
1252 | br $ra | ||
1253 | |||
1254 | .align 16 | ||
1255 | .Lcbc_enc_tail: | ||
1256 | aghi $len,15 | ||
1257 | lghi $t0,0 | ||
1258 | stg $t0,128($sp) | ||
1259 | stg $t0,136($sp) | ||
1260 | bras $t1,3f | ||
1261 | mvc 128(1,$sp),0($inp) | ||
1262 | 3: ex $len,0($t1) | ||
1263 | lghi $len,0 | ||
1264 | la $inp,128($sp) | ||
1265 | j .Lcbc_enc_loop | ||
1266 | |||
1267 | .align 16 | ||
1268 | .Lcbc_decrypt: | ||
1269 | larl $tbl,AES_Td | ||
1270 | |||
1271 | lg $t0,0($ivp) | ||
1272 | lg $t1,8($ivp) | ||
1273 | stmg $t0,$t1,128($sp) | ||
1274 | |||
1275 | .Lcbc_dec_loop: | ||
1276 | stmg $inp,$out,16($sp) | ||
1277 | llgf $s0,0($inp) | ||
1278 | llgf $s1,4($inp) | ||
1279 | llgf $s2,8($inp) | ||
1280 | llgf $s3,12($inp) | ||
1281 | lgr %r4,$key | ||
1282 | |||
1283 | bras $ra,_s390x_AES_decrypt | ||
1284 | |||
1285 | lmg $inp,$key,16($sp) | ||
1286 | sllg $s0,$s0,32 | ||
1287 | sllg $s2,$s2,32 | ||
1288 | lr $s0,$s1 | ||
1289 | lr $s2,$s3 | ||
1290 | |||
1291 | lg $t0,0($inp) | ||
1292 | lg $t1,8($inp) | ||
1293 | xg $s0,128($sp) | ||
1294 | xg $s2,136($sp) | ||
1295 | lghi $s1,16 | ||
1296 | slgr $len,$s1 | ||
1297 | brc 4,.Lcbc_dec_tail # if borrow | ||
1298 | brc 2,.Lcbc_dec_done # if zero | ||
1299 | stg $s0,0($out) | ||
1300 | stg $s2,8($out) | ||
1301 | stmg $t0,$t1,128($sp) | ||
1302 | |||
1303 | la $inp,16($inp) | ||
1304 | la $out,16($out) | ||
1305 | j .Lcbc_dec_loop | ||
1306 | |||
1307 | .Lcbc_dec_done: | ||
1308 | stg $s0,0($out) | ||
1309 | stg $s2,8($out) | ||
1310 | .Lcbc_dec_exit: | ||
1311 | lmg $ivp,$ra,48($sp) | ||
1312 | stmg $t0,$t1,0($ivp) | ||
1313 | |||
1314 | br $ra | ||
1315 | |||
1316 | .align 16 | ||
1317 | .Lcbc_dec_tail: | ||
1318 | aghi $len,15 | ||
1319 | stg $s0,128($sp) | ||
1320 | stg $s2,136($sp) | ||
1321 | bras $s1,4f | ||
1322 | mvc 0(1,$out),128($sp) | ||
1323 | 4: ex $len,0($s1) | ||
1324 | j .Lcbc_dec_exit | ||
1325 | .size AES_cbc_encrypt,.-AES_cbc_encrypt | ||
1326 | ___ | ||
1327 | } | ||
1328 | $code.=<<___; | ||
1329 | .string "AES for s390x, CRYPTOGAMS by <appro\@openssl.org>" | ||
1330 | ___ | ||
1331 | |||
1332 | $code =~ s/\`([^\`]*)\`/eval $1/gem; | ||
1333 | print $code; | ||
diff --git a/src/lib/libcrypto/aes/asm/aes-sparcv9.pl b/src/lib/libcrypto/aes/asm/aes-sparcv9.pl new file mode 100755 index 0000000000..c57b3a2d6d --- /dev/null +++ b/src/lib/libcrypto/aes/asm/aes-sparcv9.pl | |||
@@ -0,0 +1,1181 @@ | |||
1 | #!/usr/bin/env perl | ||
2 | # | ||
3 | # ==================================================================== | ||
4 | # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL | ||
5 | # project. Rights for redistribution and usage in source and binary | ||
6 | # forms are granted according to the OpenSSL license. | ||
7 | # ==================================================================== | ||
8 | # | ||
9 | # Version 1.1 | ||
10 | # | ||
11 | # The major reason for undertaken effort was to mitigate the hazard of | ||
12 | # cache-timing attack. This is [currently and initially!] addressed in | ||
13 | # two ways. 1. S-boxes are compressed from 5KB to 2KB+256B size each. | ||
14 | # 2. References to them are scheduled for L2 cache latency, meaning | ||
15 | # that the tables don't have to reside in L1 cache. Once again, this | ||
16 | # is an initial draft and one should expect more countermeasures to | ||
17 | # be implemented... | ||
18 | # | ||
19 | # Version 1.1 prefetches T[ed]4 in order to mitigate attack on last | ||
20 | # round. | ||
21 | # | ||
22 | # Even though performance was not the primary goal [on the contrary, | ||
23 | # extra shifts "induced" by compressed S-box and longer loop epilogue | ||
24 | # "induced" by scheduling for L2 have negative effect on performance], | ||
25 | # the code turned out to run in ~23 cycles per processed byte en-/ | ||
26 | # decrypted with 128-bit key. This is pretty good result for code | ||
27 | # with mentioned qualities and UltraSPARC core. Compared to Sun C | ||
28 | # generated code my encrypt procedure runs just few percents faster, | ||
29 | # while decrypt one - whole 50% faster [yes, Sun C failed to generate | ||
30 | # optimal decrypt procedure]. Compared to GNU C generated code both | ||
31 | # procedures are more than 60% faster:-) | ||
32 | |||
33 | $bits=32; | ||
34 | for (@ARGV) { $bits=64 if (/\-m64/ || /\-xarch\=v9/); } | ||
35 | if ($bits==64) { $bias=2047; $frame=192; } | ||
36 | else { $bias=0; $frame=112; } | ||
37 | $locals=16; | ||
38 | |||
39 | $acc0="%l0"; | ||
40 | $acc1="%o0"; | ||
41 | $acc2="%o1"; | ||
42 | $acc3="%o2"; | ||
43 | |||
44 | $acc4="%l1"; | ||
45 | $acc5="%o3"; | ||
46 | $acc6="%o4"; | ||
47 | $acc7="%o5"; | ||
48 | |||
49 | $acc8="%l2"; | ||
50 | $acc9="%o7"; | ||
51 | $acc10="%g1"; | ||
52 | $acc11="%g2"; | ||
53 | |||
54 | $acc12="%l3"; | ||
55 | $acc13="%g3"; | ||
56 | $acc14="%g4"; | ||
57 | $acc15="%g5"; | ||
58 | |||
59 | $t0="%l4"; | ||
60 | $t1="%l5"; | ||
61 | $t2="%l6"; | ||
62 | $t3="%l7"; | ||
63 | |||
64 | $s0="%i0"; | ||
65 | $s1="%i1"; | ||
66 | $s2="%i2"; | ||
67 | $s3="%i3"; | ||
68 | $tbl="%i4"; | ||
69 | $key="%i5"; | ||
70 | $rounds="%i7"; # aliases with return address, which is off-loaded to stack | ||
71 | |||
72 | sub _data_word() | ||
73 | { my $i; | ||
74 | while(defined($i=shift)) { $code.=sprintf"\t.long\t0x%08x,0x%08x\n",$i,$i; } | ||
75 | } | ||
76 | |||
77 | $code.=<<___ if ($bits==64); | ||
78 | .register %g2,#scratch | ||
79 | .register %g3,#scratch | ||
80 | ___ | ||
81 | $code.=<<___; | ||
82 | .section ".text",#alloc,#execinstr | ||
83 | |||
84 | .align 256 | ||
85 | AES_Te: | ||
86 | ___ | ||
87 | &_data_word( | ||
88 | 0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, | ||
89 | 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554, | ||
90 | 0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, | ||
91 | 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a, | ||
92 | 0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, | ||
93 | 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b, | ||
94 | 0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea, | ||
95 | 0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b, | ||
96 | 0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a, | ||
97 | 0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f, | ||
98 | 0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108, | ||
99 | 0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f, | ||
100 | 0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e, | ||
101 | 0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5, | ||
102 | 0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d, | ||
103 | 0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f, | ||
104 | 0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e, | ||
105 | 0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb, | ||
106 | 0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce, | ||
107 | 0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497, | ||
108 | 0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c, | ||
109 | 0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed, | ||
110 | 0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b, | ||
111 | 0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a, | ||
112 | 0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16, | ||
113 | 0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594, | ||
114 | 0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81, | ||
115 | 0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3, | ||
116 | 0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a, | ||
117 | 0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504, | ||
118 | 0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163, | ||
119 | 0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d, | ||
120 | 0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f, | ||
121 | 0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739, | ||
122 | 0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47, | ||
123 | 0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395, | ||
124 | 0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f, | ||
125 | 0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883, | ||
126 | 0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c, | ||
127 | 0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76, | ||
128 | 0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e, | ||
129 | 0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4, | ||
130 | 0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6, | ||
131 | 0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b, | ||
132 | 0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7, | ||
133 | 0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0, | ||
134 | 0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25, | ||
135 | 0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818, | ||
136 | 0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72, | ||
137 | 0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651, | ||
138 | 0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21, | ||
139 | 0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85, | ||
140 | 0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa, | ||
141 | 0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12, | ||
142 | 0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0, | ||
143 | 0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9, | ||
144 | 0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133, | ||
145 | 0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7, | ||
146 | 0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, | ||
147 | 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a, | ||
148 | 0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, | ||
149 | 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8, | ||
150 | 0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, | ||
151 | 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a); | ||
152 | $code.=<<___; | ||
153 | .byte 0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5 | ||
154 | .byte 0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76 | ||
155 | .byte 0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0 | ||
156 | .byte 0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0 | ||
157 | .byte 0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc | ||
158 | .byte 0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15 | ||
159 | .byte 0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a | ||
160 | .byte 0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75 | ||
161 | .byte 0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0 | ||
162 | .byte 0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84 | ||
163 | .byte 0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b | ||
164 | .byte 0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf | ||
165 | .byte 0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85 | ||
166 | .byte 0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8 | ||
167 | .byte 0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5 | ||
168 | .byte 0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2 | ||
169 | .byte 0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17 | ||
170 | .byte 0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73 | ||
171 | .byte 0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88 | ||
172 | .byte 0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb | ||
173 | .byte 0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c | ||
174 | .byte 0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79 | ||
175 | .byte 0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9 | ||
176 | .byte 0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08 | ||
177 | .byte 0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6 | ||
178 | .byte 0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a | ||
179 | .byte 0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e | ||
180 | .byte 0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e | ||
181 | .byte 0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94 | ||
182 | .byte 0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf | ||
183 | .byte 0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68 | ||
184 | .byte 0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16 | ||
185 | .type AES_Te,#object | ||
186 | .size AES_Te,(.-AES_Te) | ||
187 | |||
188 | .align 64 | ||
189 | .skip 16 | ||
190 | _sparcv9_AES_encrypt: | ||
191 | save %sp,-$frame-$locals,%sp | ||
192 | stx %i7,[%sp+$bias+$frame+0] ! off-load return address | ||
193 | ld [$key+240],$rounds | ||
194 | ld [$key+0],$t0 | ||
195 | ld [$key+4],$t1 ! | ||
196 | ld [$key+8],$t2 | ||
197 | srl $rounds,1,$rounds | ||
198 | xor $t0,$s0,$s0 | ||
199 | ld [$key+12],$t3 | ||
200 | srl $s0,21,$acc0 | ||
201 | xor $t1,$s1,$s1 | ||
202 | ld [$key+16],$t0 | ||
203 | srl $s1,13,$acc1 ! | ||
204 | xor $t2,$s2,$s2 | ||
205 | ld [$key+20],$t1 | ||
206 | xor $t3,$s3,$s3 | ||
207 | ld [$key+24],$t2 | ||
208 | and $acc0,2040,$acc0 | ||
209 | ld [$key+28],$t3 | ||
210 | nop | ||
211 | .Lenc_loop: | ||
212 | srl $s2,5,$acc2 ! | ||
213 | and $acc1,2040,$acc1 | ||
214 | ldx [$tbl+$acc0],$acc0 | ||
215 | sll $s3,3,$acc3 | ||
216 | and $acc2,2040,$acc2 | ||
217 | ldx [$tbl+$acc1],$acc1 | ||
218 | srl $s1,21,$acc4 | ||
219 | and $acc3,2040,$acc3 | ||
220 | ldx [$tbl+$acc2],$acc2 ! | ||
221 | srl $s2,13,$acc5 | ||
222 | and $acc4,2040,$acc4 | ||
223 | ldx [$tbl+$acc3],$acc3 | ||
224 | srl $s3,5,$acc6 | ||
225 | and $acc5,2040,$acc5 | ||
226 | ldx [$tbl+$acc4],$acc4 | ||
227 | fmovs %f0,%f0 | ||
228 | sll $s0,3,$acc7 ! | ||
229 | and $acc6,2040,$acc6 | ||
230 | ldx [$tbl+$acc5],$acc5 | ||
231 | srl $s2,21,$acc8 | ||
232 | and $acc7,2040,$acc7 | ||
233 | ldx [$tbl+$acc6],$acc6 | ||
234 | srl $s3,13,$acc9 | ||
235 | and $acc8,2040,$acc8 | ||
236 | ldx [$tbl+$acc7],$acc7 ! | ||
237 | srl $s0,5,$acc10 | ||
238 | and $acc9,2040,$acc9 | ||
239 | ldx [$tbl+$acc8],$acc8 | ||
240 | sll $s1,3,$acc11 | ||
241 | and $acc10,2040,$acc10 | ||
242 | ldx [$tbl+$acc9],$acc9 | ||
243 | fmovs %f0,%f0 | ||
244 | srl $s3,21,$acc12 ! | ||
245 | and $acc11,2040,$acc11 | ||
246 | ldx [$tbl+$acc10],$acc10 | ||
247 | srl $s0,13,$acc13 | ||
248 | and $acc12,2040,$acc12 | ||
249 | ldx [$tbl+$acc11],$acc11 | ||
250 | srl $s1,5,$acc14 | ||
251 | and $acc13,2040,$acc13 | ||
252 | ldx [$tbl+$acc12],$acc12 ! | ||
253 | sll $s2,3,$acc15 | ||
254 | and $acc14,2040,$acc14 | ||
255 | ldx [$tbl+$acc13],$acc13 | ||
256 | and $acc15,2040,$acc15 | ||
257 | add $key,32,$key | ||
258 | ldx [$tbl+$acc14],$acc14 | ||
259 | fmovs %f0,%f0 | ||
260 | subcc $rounds,1,$rounds ! | ||
261 | ldx [$tbl+$acc15],$acc15 | ||
262 | bz,a,pn %icc,.Lenc_last | ||
263 | add $tbl,2048,$rounds | ||
264 | |||
265 | srlx $acc1,8,$acc1 | ||
266 | xor $acc0,$t0,$t0 | ||
267 | ld [$key+0],$s0 | ||
268 | fmovs %f0,%f0 | ||
269 | srlx $acc2,16,$acc2 ! | ||
270 | xor $acc1,$t0,$t0 | ||
271 | ld [$key+4],$s1 | ||
272 | srlx $acc3,24,$acc3 | ||
273 | xor $acc2,$t0,$t0 | ||
274 | ld [$key+8],$s2 | ||
275 | srlx $acc5,8,$acc5 | ||
276 | xor $acc3,$t0,$t0 | ||
277 | ld [$key+12],$s3 ! | ||
278 | srlx $acc6,16,$acc6 | ||
279 | xor $acc4,$t1,$t1 | ||
280 | fmovs %f0,%f0 | ||
281 | srlx $acc7,24,$acc7 | ||
282 | xor $acc5,$t1,$t1 | ||
283 | srlx $acc9,8,$acc9 | ||
284 | xor $acc6,$t1,$t1 | ||
285 | srlx $acc10,16,$acc10 ! | ||
286 | xor $acc7,$t1,$t1 | ||
287 | srlx $acc11,24,$acc11 | ||
288 | xor $acc8,$t2,$t2 | ||
289 | srlx $acc13,8,$acc13 | ||
290 | xor $acc9,$t2,$t2 | ||
291 | srlx $acc14,16,$acc14 | ||
292 | xor $acc10,$t2,$t2 | ||
293 | srlx $acc15,24,$acc15 ! | ||
294 | xor $acc11,$t2,$t2 | ||
295 | xor $acc12,$acc14,$acc14 | ||
296 | xor $acc13,$t3,$t3 | ||
297 | srl $t0,21,$acc0 | ||
298 | xor $acc14,$t3,$t3 | ||
299 | srl $t1,13,$acc1 | ||
300 | xor $acc15,$t3,$t3 | ||
301 | |||
302 | and $acc0,2040,$acc0 ! | ||
303 | srl $t2,5,$acc2 | ||
304 | and $acc1,2040,$acc1 | ||
305 | ldx [$tbl+$acc0],$acc0 | ||
306 | sll $t3,3,$acc3 | ||
307 | and $acc2,2040,$acc2 | ||
308 | ldx [$tbl+$acc1],$acc1 | ||
309 | fmovs %f0,%f0 | ||
310 | srl $t1,21,$acc4 ! | ||
311 | and $acc3,2040,$acc3 | ||
312 | ldx [$tbl+$acc2],$acc2 | ||
313 | srl $t2,13,$acc5 | ||
314 | and $acc4,2040,$acc4 | ||
315 | ldx [$tbl+$acc3],$acc3 | ||
316 | srl $t3,5,$acc6 | ||
317 | and $acc5,2040,$acc5 | ||
318 | ldx [$tbl+$acc4],$acc4 ! | ||
319 | sll $t0,3,$acc7 | ||
320 | and $acc6,2040,$acc6 | ||
321 | ldx [$tbl+$acc5],$acc5 | ||
322 | srl $t2,21,$acc8 | ||
323 | and $acc7,2040,$acc7 | ||
324 | ldx [$tbl+$acc6],$acc6 | ||
325 | fmovs %f0,%f0 | ||
326 | srl $t3,13,$acc9 ! | ||
327 | and $acc8,2040,$acc8 | ||
328 | ldx [$tbl+$acc7],$acc7 | ||
329 | srl $t0,5,$acc10 | ||
330 | and $acc9,2040,$acc9 | ||
331 | ldx [$tbl+$acc8],$acc8 | ||
332 | sll $t1,3,$acc11 | ||
333 | and $acc10,2040,$acc10 | ||
334 | ldx [$tbl+$acc9],$acc9 ! | ||
335 | srl $t3,21,$acc12 | ||
336 | and $acc11,2040,$acc11 | ||
337 | ldx [$tbl+$acc10],$acc10 | ||
338 | srl $t0,13,$acc13 | ||
339 | and $acc12,2040,$acc12 | ||
340 | ldx [$tbl+$acc11],$acc11 | ||
341 | fmovs %f0,%f0 | ||
342 | srl $t1,5,$acc14 ! | ||
343 | and $acc13,2040,$acc13 | ||
344 | ldx [$tbl+$acc12],$acc12 | ||
345 | sll $t2,3,$acc15 | ||
346 | and $acc14,2040,$acc14 | ||
347 | ldx [$tbl+$acc13],$acc13 | ||
348 | srlx $acc1,8,$acc1 | ||
349 | and $acc15,2040,$acc15 | ||
350 | ldx [$tbl+$acc14],$acc14 ! | ||
351 | |||
352 | srlx $acc2,16,$acc2 | ||
353 | xor $acc0,$s0,$s0 | ||
354 | ldx [$tbl+$acc15],$acc15 | ||
355 | srlx $acc3,24,$acc3 | ||
356 | xor $acc1,$s0,$s0 | ||
357 | ld [$key+16],$t0 | ||
358 | fmovs %f0,%f0 | ||
359 | srlx $acc5,8,$acc5 ! | ||
360 | xor $acc2,$s0,$s0 | ||
361 | ld [$key+20],$t1 | ||
362 | srlx $acc6,16,$acc6 | ||
363 | xor $acc3,$s0,$s0 | ||
364 | ld [$key+24],$t2 | ||
365 | srlx $acc7,24,$acc7 | ||
366 | xor $acc4,$s1,$s1 | ||
367 | ld [$key+28],$t3 ! | ||
368 | srlx $acc9,8,$acc9 | ||
369 | xor $acc5,$s1,$s1 | ||
370 | ldx [$tbl+2048+0],%g0 ! prefetch te4 | ||
371 | srlx $acc10,16,$acc10 | ||
372 | xor $acc6,$s1,$s1 | ||
373 | ldx [$tbl+2048+32],%g0 ! prefetch te4 | ||
374 | srlx $acc11,24,$acc11 | ||
375 | xor $acc7,$s1,$s1 | ||
376 | ldx [$tbl+2048+64],%g0 ! prefetch te4 | ||
377 | srlx $acc13,8,$acc13 | ||
378 | xor $acc8,$s2,$s2 | ||
379 | ldx [$tbl+2048+96],%g0 ! prefetch te4 | ||
380 | srlx $acc14,16,$acc14 ! | ||
381 | xor $acc9,$s2,$s2 | ||
382 | ldx [$tbl+2048+128],%g0 ! prefetch te4 | ||
383 | srlx $acc15,24,$acc15 | ||
384 | xor $acc10,$s2,$s2 | ||
385 | ldx [$tbl+2048+160],%g0 ! prefetch te4 | ||
386 | srl $s0,21,$acc0 | ||
387 | xor $acc11,$s2,$s2 | ||
388 | ldx [$tbl+2048+192],%g0 ! prefetch te4 | ||
389 | xor $acc12,$acc14,$acc14 | ||
390 | xor $acc13,$s3,$s3 | ||
391 | ldx [$tbl+2048+224],%g0 ! prefetch te4 | ||
392 | srl $s1,13,$acc1 ! | ||
393 | xor $acc14,$s3,$s3 | ||
394 | xor $acc15,$s3,$s3 | ||
395 | ba .Lenc_loop | ||
396 | and $acc0,2040,$acc0 | ||
397 | |||
398 | .align 32 | ||
399 | .Lenc_last: | ||
400 | srlx $acc1,8,$acc1 ! | ||
401 | xor $acc0,$t0,$t0 | ||
402 | ld [$key+0],$s0 | ||
403 | srlx $acc2,16,$acc2 | ||
404 | xor $acc1,$t0,$t0 | ||
405 | ld [$key+4],$s1 | ||
406 | srlx $acc3,24,$acc3 | ||
407 | xor $acc2,$t0,$t0 | ||
408 | ld [$key+8],$s2 ! | ||
409 | srlx $acc5,8,$acc5 | ||
410 | xor $acc3,$t0,$t0 | ||
411 | ld [$key+12],$s3 | ||
412 | srlx $acc6,16,$acc6 | ||
413 | xor $acc4,$t1,$t1 | ||
414 | srlx $acc7,24,$acc7 | ||
415 | xor $acc5,$t1,$t1 | ||
416 | srlx $acc9,8,$acc9 ! | ||
417 | xor $acc6,$t1,$t1 | ||
418 | srlx $acc10,16,$acc10 | ||
419 | xor $acc7,$t1,$t1 | ||
420 | srlx $acc11,24,$acc11 | ||
421 | xor $acc8,$t2,$t2 | ||
422 | srlx $acc13,8,$acc13 | ||
423 | xor $acc9,$t2,$t2 | ||
424 | srlx $acc14,16,$acc14 ! | ||
425 | xor $acc10,$t2,$t2 | ||
426 | srlx $acc15,24,$acc15 | ||
427 | xor $acc11,$t2,$t2 | ||
428 | xor $acc12,$acc14,$acc14 | ||
429 | xor $acc13,$t3,$t3 | ||
430 | srl $t0,24,$acc0 | ||
431 | xor $acc14,$t3,$t3 | ||
432 | srl $t1,16,$acc1 ! | ||
433 | xor $acc15,$t3,$t3 | ||
434 | |||
435 | srl $t2,8,$acc2 | ||
436 | and $acc1,255,$acc1 | ||
437 | ldub [$rounds+$acc0],$acc0 | ||
438 | srl $t1,24,$acc4 | ||
439 | and $acc2,255,$acc2 | ||
440 | ldub [$rounds+$acc1],$acc1 | ||
441 | srl $t2,16,$acc5 ! | ||
442 | and $t3,255,$acc3 | ||
443 | ldub [$rounds+$acc2],$acc2 | ||
444 | ldub [$rounds+$acc3],$acc3 | ||
445 | srl $t3,8,$acc6 | ||
446 | and $acc5,255,$acc5 | ||
447 | ldub [$rounds+$acc4],$acc4 | ||
448 | fmovs %f0,%f0 | ||
449 | srl $t2,24,$acc8 ! | ||
450 | and $acc6,255,$acc6 | ||
451 | ldub [$rounds+$acc5],$acc5 | ||
452 | srl $t3,16,$acc9 | ||
453 | and $t0,255,$acc7 | ||
454 | ldub [$rounds+$acc6],$acc6 | ||
455 | ldub [$rounds+$acc7],$acc7 | ||
456 | fmovs %f0,%f0 | ||
457 | srl $t0,8,$acc10 ! | ||
458 | and $acc9,255,$acc9 | ||
459 | ldub [$rounds+$acc8],$acc8 | ||
460 | srl $t3,24,$acc12 | ||
461 | and $acc10,255,$acc10 | ||
462 | ldub [$rounds+$acc9],$acc9 | ||
463 | srl $t0,16,$acc13 | ||
464 | and $t1,255,$acc11 | ||
465 | ldub [$rounds+$acc10],$acc10 ! | ||
466 | srl $t1,8,$acc14 | ||
467 | and $acc13,255,$acc13 | ||
468 | ldub [$rounds+$acc11],$acc11 | ||
469 | ldub [$rounds+$acc12],$acc12 | ||
470 | and $acc14,255,$acc14 | ||
471 | ldub [$rounds+$acc13],$acc13 | ||
472 | and $t2,255,$acc15 | ||
473 | ldub [$rounds+$acc14],$acc14 ! | ||
474 | |||
475 | sll $acc0,24,$acc0 | ||
476 | xor $acc3,$s0,$s0 | ||
477 | ldub [$rounds+$acc15],$acc15 | ||
478 | sll $acc1,16,$acc1 | ||
479 | xor $acc0,$s0,$s0 | ||
480 | ldx [%sp+$bias+$frame+0],%i7 ! restore return address | ||
481 | fmovs %f0,%f0 | ||
482 | sll $acc2,8,$acc2 ! | ||
483 | xor $acc1,$s0,$s0 | ||
484 | sll $acc4,24,$acc4 | ||
485 | xor $acc2,$s0,$s0 | ||
486 | sll $acc5,16,$acc5 | ||
487 | xor $acc7,$s1,$s1 | ||
488 | sll $acc6,8,$acc6 | ||
489 | xor $acc4,$s1,$s1 | ||
490 | sll $acc8,24,$acc8 ! | ||
491 | xor $acc5,$s1,$s1 | ||
492 | sll $acc9,16,$acc9 | ||
493 | xor $acc11,$s2,$s2 | ||
494 | sll $acc10,8,$acc10 | ||
495 | xor $acc6,$s1,$s1 | ||
496 | sll $acc12,24,$acc12 | ||
497 | xor $acc8,$s2,$s2 | ||
498 | sll $acc13,16,$acc13 ! | ||
499 | xor $acc9,$s2,$s2 | ||
500 | sll $acc14,8,$acc14 | ||
501 | xor $acc10,$s2,$s2 | ||
502 | xor $acc12,$acc14,$acc14 | ||
503 | xor $acc13,$s3,$s3 | ||
504 | xor $acc14,$s3,$s3 | ||
505 | xor $acc15,$s3,$s3 | ||
506 | |||
507 | ret | ||
508 | restore | ||
509 | .type _sparcv9_AES_encrypt,#function | ||
510 | .size _sparcv9_AES_encrypt,(.-_sparcv9_AES_encrypt) | ||
511 | |||
512 | .align 32 | ||
513 | .globl AES_encrypt | ||
514 | AES_encrypt: | ||
515 | or %o0,%o1,%g1 | ||
516 | andcc %g1,3,%g0 | ||
517 | bnz,pn %xcc,.Lunaligned_enc | ||
518 | save %sp,-$frame,%sp | ||
519 | |||
520 | ld [%i0+0],%o0 | ||
521 | ld [%i0+4],%o1 | ||
522 | ld [%i0+8],%o2 | ||
523 | ld [%i0+12],%o3 | ||
524 | |||
525 | 1: call .+8 | ||
526 | add %o7,AES_Te-1b,%o4 | ||
527 | call _sparcv9_AES_encrypt | ||
528 | mov %i2,%o5 | ||
529 | |||
530 | st %o0,[%i1+0] | ||
531 | st %o1,[%i1+4] | ||
532 | st %o2,[%i1+8] | ||
533 | st %o3,[%i1+12] | ||
534 | |||
535 | ret | ||
536 | restore | ||
537 | |||
538 | .align 32 | ||
539 | .Lunaligned_enc: | ||
540 | ldub [%i0+0],%l0 | ||
541 | ldub [%i0+1],%l1 | ||
542 | ldub [%i0+2],%l2 | ||
543 | |||
544 | sll %l0,24,%l0 | ||
545 | ldub [%i0+3],%l3 | ||
546 | sll %l1,16,%l1 | ||
547 | ldub [%i0+4],%l4 | ||
548 | sll %l2,8,%l2 | ||
549 | or %l1,%l0,%l0 | ||
550 | ldub [%i0+5],%l5 | ||
551 | sll %l4,24,%l4 | ||
552 | or %l3,%l2,%l2 | ||
553 | ldub [%i0+6],%l6 | ||
554 | sll %l5,16,%l5 | ||
555 | or %l0,%l2,%o0 | ||
556 | ldub [%i0+7],%l7 | ||
557 | |||
558 | sll %l6,8,%l6 | ||
559 | or %l5,%l4,%l4 | ||
560 | ldub [%i0+8],%l0 | ||
561 | or %l7,%l6,%l6 | ||
562 | ldub [%i0+9],%l1 | ||
563 | or %l4,%l6,%o1 | ||
564 | ldub [%i0+10],%l2 | ||
565 | |||
566 | sll %l0,24,%l0 | ||
567 | ldub [%i0+11],%l3 | ||
568 | sll %l1,16,%l1 | ||
569 | ldub [%i0+12],%l4 | ||
570 | sll %l2,8,%l2 | ||
571 | or %l1,%l0,%l0 | ||
572 | ldub [%i0+13],%l5 | ||
573 | sll %l4,24,%l4 | ||
574 | or %l3,%l2,%l2 | ||
575 | ldub [%i0+14],%l6 | ||
576 | sll %l5,16,%l5 | ||
577 | or %l0,%l2,%o2 | ||
578 | ldub [%i0+15],%l7 | ||
579 | |||
580 | sll %l6,8,%l6 | ||
581 | or %l5,%l4,%l4 | ||
582 | or %l7,%l6,%l6 | ||
583 | or %l4,%l6,%o3 | ||
584 | |||
585 | 1: call .+8 | ||
586 | add %o7,AES_Te-1b,%o4 | ||
587 | call _sparcv9_AES_encrypt | ||
588 | mov %i2,%o5 | ||
589 | |||
590 | srl %o0,24,%l0 | ||
591 | srl %o0,16,%l1 | ||
592 | stb %l0,[%i1+0] | ||
593 | srl %o0,8,%l2 | ||
594 | stb %l1,[%i1+1] | ||
595 | stb %l2,[%i1+2] | ||
596 | srl %o1,24,%l4 | ||
597 | stb %o0,[%i1+3] | ||
598 | |||
599 | srl %o1,16,%l5 | ||
600 | stb %l4,[%i1+4] | ||
601 | srl %o1,8,%l6 | ||
602 | stb %l5,[%i1+5] | ||
603 | stb %l6,[%i1+6] | ||
604 | srl %o2,24,%l0 | ||
605 | stb %o1,[%i1+7] | ||
606 | |||
607 | srl %o2,16,%l1 | ||
608 | stb %l0,[%i1+8] | ||
609 | srl %o2,8,%l2 | ||
610 | stb %l1,[%i1+9] | ||
611 | stb %l2,[%i1+10] | ||
612 | srl %o3,24,%l4 | ||
613 | stb %o2,[%i1+11] | ||
614 | |||
615 | srl %o3,16,%l5 | ||
616 | stb %l4,[%i1+12] | ||
617 | srl %o3,8,%l6 | ||
618 | stb %l5,[%i1+13] | ||
619 | stb %l6,[%i1+14] | ||
620 | stb %o3,[%i1+15] | ||
621 | |||
622 | ret | ||
623 | restore | ||
624 | .type AES_encrypt,#function | ||
625 | .size AES_encrypt,(.-AES_encrypt) | ||
626 | |||
627 | ___ | ||
628 | |||
629 | $code.=<<___; | ||
630 | .align 256 | ||
631 | AES_Td: | ||
632 | ___ | ||
633 | &_data_word( | ||
634 | 0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, | ||
635 | 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393, | ||
636 | 0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, | ||
637 | 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f, | ||
638 | 0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, | ||
639 | 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6, | ||
640 | 0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da, | ||
641 | 0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844, | ||
642 | 0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd, | ||
643 | 0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4, | ||
644 | 0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45, | ||
645 | 0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94, | ||
646 | 0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7, | ||
647 | 0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a, | ||
648 | 0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5, | ||
649 | 0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c, | ||
650 | 0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1, | ||
651 | 0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a, | ||
652 | 0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75, | ||
653 | 0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051, | ||
654 | 0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46, | ||
655 | 0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff, | ||
656 | 0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77, | ||
657 | 0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb, | ||
658 | 0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000, | ||
659 | 0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e, | ||
660 | 0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927, | ||
661 | 0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a, | ||
662 | 0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e, | ||
663 | 0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16, | ||
664 | 0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d, | ||
665 | 0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8, | ||
666 | 0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd, | ||
667 | 0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34, | ||
668 | 0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163, | ||
669 | 0xd731dcca, 0x42638510, 0x13972240, 0x84c61120, | ||
670 | 0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d, | ||
671 | 0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0, | ||
672 | 0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422, | ||
673 | 0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef, | ||
674 | 0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36, | ||
675 | 0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4, | ||
676 | 0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662, | ||
677 | 0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5, | ||
678 | 0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3, | ||
679 | 0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b, | ||
680 | 0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8, | ||
681 | 0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6, | ||
682 | 0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6, | ||
683 | 0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0, | ||
684 | 0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815, | ||
685 | 0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f, | ||
686 | 0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df, | ||
687 | 0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f, | ||
688 | 0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e, | ||
689 | 0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713, | ||
690 | 0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89, | ||
691 | 0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c, | ||
692 | 0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, | ||
693 | 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86, | ||
694 | 0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, | ||
695 | 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541, | ||
696 | 0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, | ||
697 | 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742); | ||
698 | $code.=<<___; | ||
699 | .byte 0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38 | ||
700 | .byte 0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb | ||
701 | .byte 0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87 | ||
702 | .byte 0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb | ||
703 | .byte 0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d | ||
704 | .byte 0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e | ||
705 | .byte 0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2 | ||
706 | .byte 0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25 | ||
707 | .byte 0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16 | ||
708 | .byte 0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92 | ||
709 | .byte 0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda | ||
710 | .byte 0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84 | ||
711 | .byte 0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a | ||
712 | .byte 0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06 | ||
713 | .byte 0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02 | ||
714 | .byte 0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b | ||
715 | .byte 0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea | ||
716 | .byte 0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73 | ||
717 | .byte 0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85 | ||
718 | .byte 0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e | ||
719 | .byte 0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89 | ||
720 | .byte 0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b | ||
721 | .byte 0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20 | ||
722 | .byte 0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4 | ||
723 | .byte 0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31 | ||
724 | .byte 0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f | ||
725 | .byte 0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d | ||
726 | .byte 0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef | ||
727 | .byte 0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0 | ||
728 | .byte 0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61 | ||
729 | .byte 0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26 | ||
730 | .byte 0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d | ||
731 | .type AES_Td,#object | ||
732 | .size AES_Td,(.-AES_Td) | ||
733 | |||
734 | .align 64 | ||
735 | .skip 16 | ||
736 | _sparcv9_AES_decrypt: | ||
737 | save %sp,-$frame-$locals,%sp | ||
738 | stx %i7,[%sp+$bias+$frame+0] ! off-load return address | ||
739 | ld [$key+240],$rounds | ||
740 | ld [$key+0],$t0 | ||
741 | ld [$key+4],$t1 ! | ||
742 | ld [$key+8],$t2 | ||
743 | ld [$key+12],$t3 | ||
744 | srl $rounds,1,$rounds | ||
745 | xor $t0,$s0,$s0 | ||
746 | ld [$key+16],$t0 | ||
747 | xor $t1,$s1,$s1 | ||
748 | ld [$key+20],$t1 | ||
749 | srl $s0,21,$acc0 ! | ||
750 | xor $t2,$s2,$s2 | ||
751 | ld [$key+24],$t2 | ||
752 | xor $t3,$s3,$s3 | ||
753 | and $acc0,2040,$acc0 | ||
754 | ld [$key+28],$t3 | ||
755 | srl $s3,13,$acc1 | ||
756 | nop | ||
757 | .Ldec_loop: | ||
758 | srl $s2,5,$acc2 ! | ||
759 | and $acc1,2040,$acc1 | ||
760 | ldx [$tbl+$acc0],$acc0 | ||
761 | sll $s1,3,$acc3 | ||
762 | and $acc2,2040,$acc2 | ||
763 | ldx [$tbl+$acc1],$acc1 | ||
764 | srl $s1,21,$acc4 | ||
765 | and $acc3,2040,$acc3 | ||
766 | ldx [$tbl+$acc2],$acc2 ! | ||
767 | srl $s0,13,$acc5 | ||
768 | and $acc4,2040,$acc4 | ||
769 | ldx [$tbl+$acc3],$acc3 | ||
770 | srl $s3,5,$acc6 | ||
771 | and $acc5,2040,$acc5 | ||
772 | ldx [$tbl+$acc4],$acc4 | ||
773 | fmovs %f0,%f0 | ||
774 | sll $s2,3,$acc7 ! | ||
775 | and $acc6,2040,$acc6 | ||
776 | ldx [$tbl+$acc5],$acc5 | ||
777 | srl $s2,21,$acc8 | ||
778 | and $acc7,2040,$acc7 | ||
779 | ldx [$tbl+$acc6],$acc6 | ||
780 | srl $s1,13,$acc9 | ||
781 | and $acc8,2040,$acc8 | ||
782 | ldx [$tbl+$acc7],$acc7 ! | ||
783 | srl $s0,5,$acc10 | ||
784 | and $acc9,2040,$acc9 | ||
785 | ldx [$tbl+$acc8],$acc8 | ||
786 | sll $s3,3,$acc11 | ||
787 | and $acc10,2040,$acc10 | ||
788 | ldx [$tbl+$acc9],$acc9 | ||
789 | fmovs %f0,%f0 | ||
790 | srl $s3,21,$acc12 ! | ||
791 | and $acc11,2040,$acc11 | ||
792 | ldx [$tbl+$acc10],$acc10 | ||
793 | srl $s2,13,$acc13 | ||
794 | and $acc12,2040,$acc12 | ||
795 | ldx [$tbl+$acc11],$acc11 | ||
796 | srl $s1,5,$acc14 | ||
797 | and $acc13,2040,$acc13 | ||
798 | ldx [$tbl+$acc12],$acc12 ! | ||
799 | sll $s0,3,$acc15 | ||
800 | and $acc14,2040,$acc14 | ||
801 | ldx [$tbl+$acc13],$acc13 | ||
802 | and $acc15,2040,$acc15 | ||
803 | add $key,32,$key | ||
804 | ldx [$tbl+$acc14],$acc14 | ||
805 | fmovs %f0,%f0 | ||
806 | subcc $rounds,1,$rounds ! | ||
807 | ldx [$tbl+$acc15],$acc15 | ||
808 | bz,a,pn %icc,.Ldec_last | ||
809 | add $tbl,2048,$rounds | ||
810 | |||
811 | srlx $acc1,8,$acc1 | ||
812 | xor $acc0,$t0,$t0 | ||
813 | ld [$key+0],$s0 | ||
814 | fmovs %f0,%f0 | ||
815 | srlx $acc2,16,$acc2 ! | ||
816 | xor $acc1,$t0,$t0 | ||
817 | ld [$key+4],$s1 | ||
818 | srlx $acc3,24,$acc3 | ||
819 | xor $acc2,$t0,$t0 | ||
820 | ld [$key+8],$s2 | ||
821 | srlx $acc5,8,$acc5 | ||
822 | xor $acc3,$t0,$t0 | ||
823 | ld [$key+12],$s3 ! | ||
824 | srlx $acc6,16,$acc6 | ||
825 | xor $acc4,$t1,$t1 | ||
826 | fmovs %f0,%f0 | ||
827 | srlx $acc7,24,$acc7 | ||
828 | xor $acc5,$t1,$t1 | ||
829 | srlx $acc9,8,$acc9 | ||
830 | xor $acc6,$t1,$t1 | ||
831 | srlx $acc10,16,$acc10 ! | ||
832 | xor $acc7,$t1,$t1 | ||
833 | srlx $acc11,24,$acc11 | ||
834 | xor $acc8,$t2,$t2 | ||
835 | srlx $acc13,8,$acc13 | ||
836 | xor $acc9,$t2,$t2 | ||
837 | srlx $acc14,16,$acc14 | ||
838 | xor $acc10,$t2,$t2 | ||
839 | srlx $acc15,24,$acc15 ! | ||
840 | xor $acc11,$t2,$t2 | ||
841 | xor $acc12,$acc14,$acc14 | ||
842 | xor $acc13,$t3,$t3 | ||
843 | srl $t0,21,$acc0 | ||
844 | xor $acc14,$t3,$t3 | ||
845 | xor $acc15,$t3,$t3 | ||
846 | srl $t3,13,$acc1 | ||
847 | |||
848 | and $acc0,2040,$acc0 ! | ||
849 | srl $t2,5,$acc2 | ||
850 | and $acc1,2040,$acc1 | ||
851 | ldx [$tbl+$acc0],$acc0 | ||
852 | sll $t1,3,$acc3 | ||
853 | and $acc2,2040,$acc2 | ||
854 | ldx [$tbl+$acc1],$acc1 | ||
855 | fmovs %f0,%f0 | ||
856 | srl $t1,21,$acc4 ! | ||
857 | and $acc3,2040,$acc3 | ||
858 | ldx [$tbl+$acc2],$acc2 | ||
859 | srl $t0,13,$acc5 | ||
860 | and $acc4,2040,$acc4 | ||
861 | ldx [$tbl+$acc3],$acc3 | ||
862 | srl $t3,5,$acc6 | ||
863 | and $acc5,2040,$acc5 | ||
864 | ldx [$tbl+$acc4],$acc4 ! | ||
865 | sll $t2,3,$acc7 | ||
866 | and $acc6,2040,$acc6 | ||
867 | ldx [$tbl+$acc5],$acc5 | ||
868 | srl $t2,21,$acc8 | ||
869 | and $acc7,2040,$acc7 | ||
870 | ldx [$tbl+$acc6],$acc6 | ||
871 | fmovs %f0,%f0 | ||
872 | srl $t1,13,$acc9 ! | ||
873 | and $acc8,2040,$acc8 | ||
874 | ldx [$tbl+$acc7],$acc7 | ||
875 | srl $t0,5,$acc10 | ||
876 | and $acc9,2040,$acc9 | ||
877 | ldx [$tbl+$acc8],$acc8 | ||
878 | sll $t3,3,$acc11 | ||
879 | and $acc10,2040,$acc10 | ||
880 | ldx [$tbl+$acc9],$acc9 ! | ||
881 | srl $t3,21,$acc12 | ||
882 | and $acc11,2040,$acc11 | ||
883 | ldx [$tbl+$acc10],$acc10 | ||
884 | srl $t2,13,$acc13 | ||
885 | and $acc12,2040,$acc12 | ||
886 | ldx [$tbl+$acc11],$acc11 | ||
887 | fmovs %f0,%f0 | ||
888 | srl $t1,5,$acc14 ! | ||
889 | and $acc13,2040,$acc13 | ||
890 | ldx [$tbl+$acc12],$acc12 | ||
891 | sll $t0,3,$acc15 | ||
892 | and $acc14,2040,$acc14 | ||
893 | ldx [$tbl+$acc13],$acc13 | ||
894 | srlx $acc1,8,$acc1 | ||
895 | and $acc15,2040,$acc15 | ||
896 | ldx [$tbl+$acc14],$acc14 ! | ||
897 | |||
898 | srlx $acc2,16,$acc2 | ||
899 | xor $acc0,$s0,$s0 | ||
900 | ldx [$tbl+$acc15],$acc15 | ||
901 | srlx $acc3,24,$acc3 | ||
902 | xor $acc1,$s0,$s0 | ||
903 | ld [$key+16],$t0 | ||
904 | fmovs %f0,%f0 | ||
905 | srlx $acc5,8,$acc5 ! | ||
906 | xor $acc2,$s0,$s0 | ||
907 | ld [$key+20],$t1 | ||
908 | srlx $acc6,16,$acc6 | ||
909 | xor $acc3,$s0,$s0 | ||
910 | ld [$key+24],$t2 | ||
911 | srlx $acc7,24,$acc7 | ||
912 | xor $acc4,$s1,$s1 | ||
913 | ld [$key+28],$t3 ! | ||
914 | srlx $acc9,8,$acc9 | ||
915 | xor $acc5,$s1,$s1 | ||
916 | ldx [$tbl+2048+0],%g0 ! prefetch td4 | ||
917 | srlx $acc10,16,$acc10 | ||
918 | xor $acc6,$s1,$s1 | ||
919 | ldx [$tbl+2048+32],%g0 ! prefetch td4 | ||
920 | srlx $acc11,24,$acc11 | ||
921 | xor $acc7,$s1,$s1 | ||
922 | ldx [$tbl+2048+64],%g0 ! prefetch td4 | ||
923 | srlx $acc13,8,$acc13 | ||
924 | xor $acc8,$s2,$s2 | ||
925 | ldx [$tbl+2048+96],%g0 ! prefetch td4 | ||
926 | srlx $acc14,16,$acc14 ! | ||
927 | xor $acc9,$s2,$s2 | ||
928 | ldx [$tbl+2048+128],%g0 ! prefetch td4 | ||
929 | srlx $acc15,24,$acc15 | ||
930 | xor $acc10,$s2,$s2 | ||
931 | ldx [$tbl+2048+160],%g0 ! prefetch td4 | ||
932 | srl $s0,21,$acc0 | ||
933 | xor $acc11,$s2,$s2 | ||
934 | ldx [$tbl+2048+192],%g0 ! prefetch td4 | ||
935 | xor $acc12,$acc14,$acc14 | ||
936 | xor $acc13,$s3,$s3 | ||
937 | ldx [$tbl+2048+224],%g0 ! prefetch td4 | ||
938 | and $acc0,2040,$acc0 ! | ||
939 | xor $acc14,$s3,$s3 | ||
940 | xor $acc15,$s3,$s3 | ||
941 | ba .Ldec_loop | ||
942 | srl $s3,13,$acc1 | ||
943 | |||
944 | .align 32 | ||
945 | .Ldec_last: | ||
946 | srlx $acc1,8,$acc1 ! | ||
947 | xor $acc0,$t0,$t0 | ||
948 | ld [$key+0],$s0 | ||
949 | srlx $acc2,16,$acc2 | ||
950 | xor $acc1,$t0,$t0 | ||
951 | ld [$key+4],$s1 | ||
952 | srlx $acc3,24,$acc3 | ||
953 | xor $acc2,$t0,$t0 | ||
954 | ld [$key+8],$s2 ! | ||
955 | srlx $acc5,8,$acc5 | ||
956 | xor $acc3,$t0,$t0 | ||
957 | ld [$key+12],$s3 | ||
958 | srlx $acc6,16,$acc6 | ||
959 | xor $acc4,$t1,$t1 | ||
960 | srlx $acc7,24,$acc7 | ||
961 | xor $acc5,$t1,$t1 | ||
962 | srlx $acc9,8,$acc9 ! | ||
963 | xor $acc6,$t1,$t1 | ||
964 | srlx $acc10,16,$acc10 | ||
965 | xor $acc7,$t1,$t1 | ||
966 | srlx $acc11,24,$acc11 | ||
967 | xor $acc8,$t2,$t2 | ||
968 | srlx $acc13,8,$acc13 | ||
969 | xor $acc9,$t2,$t2 | ||
970 | srlx $acc14,16,$acc14 ! | ||
971 | xor $acc10,$t2,$t2 | ||
972 | srlx $acc15,24,$acc15 | ||
973 | xor $acc11,$t2,$t2 | ||
974 | xor $acc12,$acc14,$acc14 | ||
975 | xor $acc13,$t3,$t3 | ||
976 | srl $t0,24,$acc0 | ||
977 | xor $acc14,$t3,$t3 | ||
978 | xor $acc15,$t3,$t3 ! | ||
979 | srl $t3,16,$acc1 | ||
980 | |||
981 | srl $t2,8,$acc2 | ||
982 | and $acc1,255,$acc1 | ||
983 | ldub [$rounds+$acc0],$acc0 | ||
984 | srl $t1,24,$acc4 | ||
985 | and $acc2,255,$acc2 | ||
986 | ldub [$rounds+$acc1],$acc1 | ||
987 | srl $t0,16,$acc5 ! | ||
988 | and $t1,255,$acc3 | ||
989 | ldub [$rounds+$acc2],$acc2 | ||
990 | ldub [$rounds+$acc3],$acc3 | ||
991 | srl $t3,8,$acc6 | ||
992 | and $acc5,255,$acc5 | ||
993 | ldub [$rounds+$acc4],$acc4 | ||
994 | fmovs %f0,%f0 | ||
995 | srl $t2,24,$acc8 ! | ||
996 | and $acc6,255,$acc6 | ||
997 | ldub [$rounds+$acc5],$acc5 | ||
998 | srl $t1,16,$acc9 | ||
999 | and $t2,255,$acc7 | ||
1000 | ldub [$rounds+$acc6],$acc6 | ||
1001 | ldub [$rounds+$acc7],$acc7 | ||
1002 | fmovs %f0,%f0 | ||
1003 | srl $t0,8,$acc10 ! | ||
1004 | and $acc9,255,$acc9 | ||
1005 | ldub [$rounds+$acc8],$acc8 | ||
1006 | srl $t3,24,$acc12 | ||
1007 | and $acc10,255,$acc10 | ||
1008 | ldub [$rounds+$acc9],$acc9 | ||
1009 | srl $t2,16,$acc13 | ||
1010 | and $t3,255,$acc11 | ||
1011 | ldub [$rounds+$acc10],$acc10 ! | ||
1012 | srl $t1,8,$acc14 | ||
1013 | and $acc13,255,$acc13 | ||
1014 | ldub [$rounds+$acc11],$acc11 | ||
1015 | ldub [$rounds+$acc12],$acc12 | ||
1016 | and $acc14,255,$acc14 | ||
1017 | ldub [$rounds+$acc13],$acc13 | ||
1018 | and $t0,255,$acc15 | ||
1019 | ldub [$rounds+$acc14],$acc14 ! | ||
1020 | |||
1021 | sll $acc0,24,$acc0 | ||
1022 | xor $acc3,$s0,$s0 | ||
1023 | ldub [$rounds+$acc15],$acc15 | ||
1024 | sll $acc1,16,$acc1 | ||
1025 | xor $acc0,$s0,$s0 | ||
1026 | ldx [%sp+$bias+$frame+0],%i7 ! restore return address | ||
1027 | fmovs %f0,%f0 | ||
1028 | sll $acc2,8,$acc2 ! | ||
1029 | xor $acc1,$s0,$s0 | ||
1030 | sll $acc4,24,$acc4 | ||
1031 | xor $acc2,$s0,$s0 | ||
1032 | sll $acc5,16,$acc5 | ||
1033 | xor $acc7,$s1,$s1 | ||
1034 | sll $acc6,8,$acc6 | ||
1035 | xor $acc4,$s1,$s1 | ||
1036 | sll $acc8,24,$acc8 ! | ||
1037 | xor $acc5,$s1,$s1 | ||
1038 | sll $acc9,16,$acc9 | ||
1039 | xor $acc11,$s2,$s2 | ||
1040 | sll $acc10,8,$acc10 | ||
1041 | xor $acc6,$s1,$s1 | ||
1042 | sll $acc12,24,$acc12 | ||
1043 | xor $acc8,$s2,$s2 | ||
1044 | sll $acc13,16,$acc13 ! | ||
1045 | xor $acc9,$s2,$s2 | ||
1046 | sll $acc14,8,$acc14 | ||
1047 | xor $acc10,$s2,$s2 | ||
1048 | xor $acc12,$acc14,$acc14 | ||
1049 | xor $acc13,$s3,$s3 | ||
1050 | xor $acc14,$s3,$s3 | ||
1051 | xor $acc15,$s3,$s3 | ||
1052 | |||
1053 | ret | ||
1054 | restore | ||
1055 | .type _sparcv9_AES_decrypt,#function | ||
1056 | .size _sparcv9_AES_decrypt,(.-_sparcv9_AES_decrypt) | ||
1057 | |||
1058 | .align 32 | ||
1059 | .globl AES_decrypt | ||
1060 | AES_decrypt: | ||
1061 | or %o0,%o1,%g1 | ||
1062 | andcc %g1,3,%g0 | ||
1063 | bnz,pn %xcc,.Lunaligned_dec | ||
1064 | save %sp,-$frame,%sp | ||
1065 | |||
1066 | ld [%i0+0],%o0 | ||
1067 | ld [%i0+4],%o1 | ||
1068 | ld [%i0+8],%o2 | ||
1069 | ld [%i0+12],%o3 | ||
1070 | |||
1071 | 1: call .+8 | ||
1072 | add %o7,AES_Td-1b,%o4 | ||
1073 | call _sparcv9_AES_decrypt | ||
1074 | mov %i2,%o5 | ||
1075 | |||
1076 | st %o0,[%i1+0] | ||
1077 | st %o1,[%i1+4] | ||
1078 | st %o2,[%i1+8] | ||
1079 | st %o3,[%i1+12] | ||
1080 | |||
1081 | ret | ||
1082 | restore | ||
1083 | |||
1084 | .align 32 | ||
1085 | .Lunaligned_dec: | ||
1086 | ldub [%i0+0],%l0 | ||
1087 | ldub [%i0+1],%l1 | ||
1088 | ldub [%i0+2],%l2 | ||
1089 | |||
1090 | sll %l0,24,%l0 | ||
1091 | ldub [%i0+3],%l3 | ||
1092 | sll %l1,16,%l1 | ||
1093 | ldub [%i0+4],%l4 | ||
1094 | sll %l2,8,%l2 | ||
1095 | or %l1,%l0,%l0 | ||
1096 | ldub [%i0+5],%l5 | ||
1097 | sll %l4,24,%l4 | ||
1098 | or %l3,%l2,%l2 | ||
1099 | ldub [%i0+6],%l6 | ||
1100 | sll %l5,16,%l5 | ||
1101 | or %l0,%l2,%o0 | ||
1102 | ldub [%i0+7],%l7 | ||
1103 | |||
1104 | sll %l6,8,%l6 | ||
1105 | or %l5,%l4,%l4 | ||
1106 | ldub [%i0+8],%l0 | ||
1107 | or %l7,%l6,%l6 | ||
1108 | ldub [%i0+9],%l1 | ||
1109 | or %l4,%l6,%o1 | ||
1110 | ldub [%i0+10],%l2 | ||
1111 | |||
1112 | sll %l0,24,%l0 | ||
1113 | ldub [%i0+11],%l3 | ||
1114 | sll %l1,16,%l1 | ||
1115 | ldub [%i0+12],%l4 | ||
1116 | sll %l2,8,%l2 | ||
1117 | or %l1,%l0,%l0 | ||
1118 | ldub [%i0+13],%l5 | ||
1119 | sll %l4,24,%l4 | ||
1120 | or %l3,%l2,%l2 | ||
1121 | ldub [%i0+14],%l6 | ||
1122 | sll %l5,16,%l5 | ||
1123 | or %l0,%l2,%o2 | ||
1124 | ldub [%i0+15],%l7 | ||
1125 | |||
1126 | sll %l6,8,%l6 | ||
1127 | or %l5,%l4,%l4 | ||
1128 | or %l7,%l6,%l6 | ||
1129 | or %l4,%l6,%o3 | ||
1130 | |||
1131 | 1: call .+8 | ||
1132 | add %o7,AES_Td-1b,%o4 | ||
1133 | call _sparcv9_AES_decrypt | ||
1134 | mov %i2,%o5 | ||
1135 | |||
1136 | srl %o0,24,%l0 | ||
1137 | srl %o0,16,%l1 | ||
1138 | stb %l0,[%i1+0] | ||
1139 | srl %o0,8,%l2 | ||
1140 | stb %l1,[%i1+1] | ||
1141 | stb %l2,[%i1+2] | ||
1142 | srl %o1,24,%l4 | ||
1143 | stb %o0,[%i1+3] | ||
1144 | |||
1145 | srl %o1,16,%l5 | ||
1146 | stb %l4,[%i1+4] | ||
1147 | srl %o1,8,%l6 | ||
1148 | stb %l5,[%i1+5] | ||
1149 | stb %l6,[%i1+6] | ||
1150 | srl %o2,24,%l0 | ||
1151 | stb %o1,[%i1+7] | ||
1152 | |||
1153 | srl %o2,16,%l1 | ||
1154 | stb %l0,[%i1+8] | ||
1155 | srl %o2,8,%l2 | ||
1156 | stb %l1,[%i1+9] | ||
1157 | stb %l2,[%i1+10] | ||
1158 | srl %o3,24,%l4 | ||
1159 | stb %o2,[%i1+11] | ||
1160 | |||
1161 | srl %o3,16,%l5 | ||
1162 | stb %l4,[%i1+12] | ||
1163 | srl %o3,8,%l6 | ||
1164 | stb %l5,[%i1+13] | ||
1165 | stb %l6,[%i1+14] | ||
1166 | stb %o3,[%i1+15] | ||
1167 | |||
1168 | ret | ||
1169 | restore | ||
1170 | .type AES_decrypt,#function | ||
1171 | .size AES_decrypt,(.-AES_decrypt) | ||
1172 | ___ | ||
1173 | |||
1174 | # fmovs instructions substituting for FP nops were originally added | ||
1175 | # to meet specific instruction alignment requirements to maximize ILP. | ||
1176 | # As UltraSPARC T1, a.k.a. Niagara, has shared FPU, FP nops can have | ||
1177 | # undesired effect, so just omit them and sacrifice some portion of | ||
1178 | # percent in performance... | ||
1179 | $code =~ s/fmovs.*$//gem; | ||
1180 | |||
1181 | print $code; | ||