]>
Commit | Line | Data |
---|---|---|
6aa36e8e | 1 | #! /usr/bin/env perl |
b0edda11 | 2 | # Copyright 2012-2018 The OpenSSL Project Authors. All Rights Reserved. |
6aa36e8e RS |
3 | # |
4 | # Licensed under the OpenSSL license (the "License"). You may not use | |
5 | # this file except in compliance with the License. You can obtain a copy | |
6 | # in the file LICENSE in the source distribution or at | |
7 | # https://www.openssl.org/source/license.html | |
8 | ||
c4a52a6d AP |
9 | |
10 | # ==================================================================== | |
358c372d | 11 | # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL |
c4a52a6d AP |
12 | # project. The module is, however, dual licensed under OpenSSL and |
13 | # CRYPTOGAMS licenses depending on where you obtain it. For further | |
14 | # details see http://www.openssl.org/~appro/cryptogams/. | |
e0202d94 AP |
15 | # |
16 | # Specific modes and adaptation for Linux kernel by Ard Biesheuvel | |
e3713c36 | 17 | # of Linaro. Permission to use under GPL terms is granted. |
c4a52a6d AP |
18 | # ==================================================================== |
19 | ||
20 | # Bit-sliced AES for ARM NEON | |
21 | # | |
22 | # February 2012. | |
23 | # | |
24 | # This implementation is direct adaptation of bsaes-x86_64 module for | |
25 | # ARM NEON. Except that this module is endian-neutral [in sense that | |
26 | # it can be compiled for either endianness] by courtesy of vld1.8's | |
27 | # neutrality. Initial version doesn't implement interface to OpenSSL, | |
28 | # only low-level primitives and unsupported entry points, just enough | |
29 | # to collect performance results, which for Cortex-A8 core are: | |
30 | # | |
e7db9896 | 31 | # encrypt 19.5 cycles per byte processed with 128-bit key |
6f6a6130 | 32 | # decrypt 22.1 cycles per byte processed with 128-bit key |
e7db9896 | 33 | # key conv. 440 cycles per 128-bit key/0.18 of 8x block |
c4a52a6d | 34 | # |
6f6a6130 | 35 | # Snapdragon S4 encrypts byte in 17.6 cycles and decrypts in 19.7, |
e7db9896 AP |
36 | # which is [much] worse than anticipated (for further details see |
37 | # http://www.openssl.org/~appro/Snapdragon-S4.html). | |
a903e691 | 38 | # |
6f6a6130 | 39 | # Cortex-A15 manages in 14.2/16.1 cycles [when integer-only code |
75fe4223 AP |
40 | # manages in 20.0 cycles]. |
41 | # | |
c4a52a6d | 42 | # When comparing to x86_64 results keep in mind that NEON unit is |
a903e691 AP |
43 | # [mostly] single-issue and thus can't [fully] benefit from |
44 | # instruction-level parallelism. And when comparing to aes-armv4 | |
45 | # results keep in mind key schedule conversion overhead (see | |
46 | # bsaes-x86_64.pl for further details)... | |
c4a52a6d AP |
47 | # |
48 | # <appro@openssl.org> | |
49 | ||
e0202d94 | 50 | # April-August 2013 |
e3713c36 | 51 | # Add CBC, CTR and XTS subroutines and adapt for kernel use; courtesy of Ard. |
e0202d94 | 52 | |
313e6ec1 | 53 | $flavour = shift; |
a5aa63a4 RL |
54 | if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; } |
55 | else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} } | |
313e6ec1 AP |
56 | |
57 | if ($flavour && $flavour ne "void") { | |
58 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; | |
59 | ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or | |
60 | ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or | |
61 | die "can't locate arm-xlate.pl"; | |
62 | ||
63 | open STDOUT,"| \"$^X\" $xlate $flavour $output"; | |
64 | } else { | |
65 | open STDOUT,">$output"; | |
66 | } | |
c4a52a6d AP |
67 | |
68 | my ($inp,$out,$len,$key)=("r0","r1","r2","r3"); | |
69 | my @XMM=map("q$_",(0..15)); | |
70 | ||
71 | { | |
72 | my ($key,$rounds,$const)=("r4","r5","r6"); | |
73 | ||
74 | sub Dlo() { shift=~m|q([1]?[0-9])|?"d".($1*2):""; } | |
75 | sub Dhi() { shift=~m|q([1]?[0-9])|?"d".($1*2+1):""; } | |
76 | ||
77 | sub Sbox { | |
78 | # input in lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb | |
79 | # output in lsb > [b0, b1, b4, b6, b3, b7, b2, b5] < msb | |
80 | my @b=@_[0..7]; | |
81 | my @t=@_[8..11]; | |
82 | my @s=@_[12..15]; | |
83 | &InBasisChange (@b); | |
84 | &Inv_GF256 (@b[6,5,0,3,7,1,4,2],@t,@s); | |
85 | &OutBasisChange (@b[7,1,4,2,6,5,0,3]); | |
86 | } | |
87 | ||
88 | sub InBasisChange { | |
89 | # input in lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb | |
609b0852 | 90 | # output in lsb > [b6, b5, b0, b3, b7, b1, b4, b2] < msb |
c4a52a6d AP |
91 | my @b=@_[0..7]; |
92 | $code.=<<___; | |
c4a52a6d | 93 | veor @b[2], @b[2], @b[1] |
a903e691 | 94 | veor @b[5], @b[5], @b[6] |
c4a52a6d AP |
95 | veor @b[3], @b[3], @b[0] |
96 | veor @b[6], @b[6], @b[2] | |
97 | veor @b[5], @b[5], @b[0] | |
98 | ||
99 | veor @b[6], @b[6], @b[3] | |
100 | veor @b[3], @b[3], @b[7] | |
101 | veor @b[7], @b[7], @b[5] | |
102 | veor @b[3], @b[3], @b[4] | |
103 | veor @b[4], @b[4], @b[5] | |
c4a52a6d AP |
104 | |
105 | veor @b[2], @b[2], @b[7] | |
a903e691 | 106 | veor @b[3], @b[3], @b[1] |
c4a52a6d AP |
107 | veor @b[1], @b[1], @b[5] |
108 | ___ | |
109 | } | |
110 | ||
111 | sub OutBasisChange { | |
112 | # input in lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb | |
113 | # output in lsb > [b6, b1, b2, b4, b7, b0, b3, b5] < msb | |
114 | my @b=@_[0..7]; | |
115 | $code.=<<___; | |
116 | veor @b[0], @b[0], @b[6] | |
117 | veor @b[1], @b[1], @b[4] | |
c4a52a6d | 118 | veor @b[4], @b[4], @b[6] |
a903e691 | 119 | veor @b[2], @b[2], @b[0] |
c4a52a6d AP |
120 | veor @b[6], @b[6], @b[1] |
121 | ||
122 | veor @b[1], @b[1], @b[5] | |
123 | veor @b[5], @b[5], @b[3] | |
124 | veor @b[3], @b[3], @b[7] | |
125 | veor @b[7], @b[7], @b[5] | |
126 | veor @b[2], @b[2], @b[5] | |
127 | ||
128 | veor @b[4], @b[4], @b[7] | |
129 | ___ | |
130 | } | |
131 | ||
132 | sub InvSbox { | |
133 | # input in lsb > [b0, b1, b2, b3, b4, b5, b6, b7] < msb | |
134 | # output in lsb > [b0, b1, b6, b4, b2, b7, b3, b5] < msb | |
135 | my @b=@_[0..7]; | |
136 | my @t=@_[8..11]; | |
137 | my @s=@_[12..15]; | |
138 | &InvInBasisChange (@b); | |
139 | &Inv_GF256 (@b[5,1,2,6,3,7,0,4],@t,@s); | |
140 | &InvOutBasisChange (@b[3,7,0,4,5,1,2,6]); | |
141 | } | |
142 | ||
a903e691 | 143 | sub InvInBasisChange { # OutBasisChange in reverse (with twist) |
c4a52a6d AP |
144 | my @b=@_[5,1,2,6,3,7,0,4]; |
145 | $code.=<<___ | |
a903e691 | 146 | veor @b[1], @b[1], @b[7] |
c4a52a6d AP |
147 | veor @b[4], @b[4], @b[7] |
148 | ||
149 | veor @b[7], @b[7], @b[5] | |
a903e691 | 150 | veor @b[1], @b[1], @b[3] |
c4a52a6d AP |
151 | veor @b[2], @b[2], @b[5] |
152 | veor @b[3], @b[3], @b[7] | |
c4a52a6d AP |
153 | |
154 | veor @b[6], @b[6], @b[1] | |
155 | veor @b[2], @b[2], @b[0] | |
a903e691 | 156 | veor @b[5], @b[5], @b[3] |
c4a52a6d AP |
157 | veor @b[4], @b[4], @b[6] |
158 | veor @b[0], @b[0], @b[6] | |
159 | veor @b[1], @b[1], @b[4] | |
160 | ___ | |
161 | } | |
162 | ||
163 | sub InvOutBasisChange { # InBasisChange in reverse | |
164 | my @b=@_[2,5,7,3,6,1,0,4]; | |
165 | $code.=<<___; | |
166 | veor @b[1], @b[1], @b[5] | |
167 | veor @b[2], @b[2], @b[7] | |
168 | ||
169 | veor @b[3], @b[3], @b[1] | |
170 | veor @b[4], @b[4], @b[5] | |
171 | veor @b[7], @b[7], @b[5] | |
172 | veor @b[3], @b[3], @b[4] | |
173 | veor @b[5], @b[5], @b[0] | |
174 | veor @b[3], @b[3], @b[7] | |
175 | veor @b[6], @b[6], @b[2] | |
176 | veor @b[2], @b[2], @b[1] | |
177 | veor @b[6], @b[6], @b[3] | |
178 | ||
179 | veor @b[3], @b[3], @b[0] | |
180 | veor @b[5], @b[5], @b[6] | |
181 | ___ | |
182 | } | |
183 | ||
184 | sub Mul_GF4 { | |
185 | #;************************************************************* | |
186 | #;* Mul_GF4: Input x0-x1,y0-y1 Output x0-x1 Temp t0 (8) * | |
187 | #;************************************************************* | |
a903e691 | 188 | my ($x0,$x1,$y0,$y1,$t0,$t1)=@_; |
c4a52a6d AP |
189 | $code.=<<___; |
190 | veor $t0, $y0, $y1 | |
191 | vand $t0, $t0, $x0 | |
192 | veor $x0, $x0, $x1 | |
a903e691 | 193 | vand $t1, $x1, $y0 |
c4a52a6d | 194 | vand $x0, $x0, $y1 |
a903e691 AP |
195 | veor $x1, $t1, $t0 |
196 | veor $x0, $x0, $t1 | |
c4a52a6d AP |
197 | ___ |
198 | } | |
199 | ||
200 | sub Mul_GF4_N { # not used, see next subroutine | |
201 | # multiply and scale by N | |
202 | my ($x0,$x1,$y0,$y1,$t0)=@_; | |
203 | $code.=<<___; | |
204 | veor $t0, $y0, $y1 | |
205 | vand $t0, $t0, $x0 | |
206 | veor $x0, $x0, $x1 | |
207 | vand $x1, $x1, $y0 | |
208 | vand $x0, $x0, $y1 | |
209 | veor $x1, $x1, $x0 | |
210 | veor $x0, $x0, $t0 | |
211 | ___ | |
212 | } | |
213 | ||
214 | sub Mul_GF4_N_GF4 { | |
215 | # interleaved Mul_GF4_N and Mul_GF4 | |
216 | my ($x0,$x1,$y0,$y1,$t0, | |
217 | $x2,$x3,$y2,$y3,$t1)=@_; | |
218 | $code.=<<___; | |
219 | veor $t0, $y0, $y1 | |
220 | veor $t1, $y2, $y3 | |
221 | vand $t0, $t0, $x0 | |
222 | vand $t1, $t1, $x2 | |
223 | veor $x0, $x0, $x1 | |
224 | veor $x2, $x2, $x3 | |
225 | vand $x1, $x1, $y0 | |
226 | vand $x3, $x3, $y2 | |
227 | vand $x0, $x0, $y1 | |
228 | vand $x2, $x2, $y3 | |
229 | veor $x1, $x1, $x0 | |
230 | veor $x2, $x2, $x3 | |
231 | veor $x0, $x0, $t0 | |
232 | veor $x3, $x3, $t1 | |
233 | ___ | |
234 | } | |
235 | sub Mul_GF16_2 { | |
236 | my @x=@_[0..7]; | |
237 | my @y=@_[8..11]; | |
238 | my @t=@_[12..15]; | |
239 | $code.=<<___; | |
a903e691 AP |
240 | veor @t[0], @x[0], @x[2] |
241 | veor @t[1], @x[1], @x[3] | |
c4a52a6d | 242 | ___ |
a903e691 | 243 | &Mul_GF4 (@x[0], @x[1], @y[0], @y[1], @t[2..3]); |
c4a52a6d | 244 | $code.=<<___; |
c4a52a6d AP |
245 | veor @y[0], @y[0], @y[2] |
246 | veor @y[1], @y[1], @y[3] | |
247 | ___ | |
248 | Mul_GF4_N_GF4 (@t[0], @t[1], @y[0], @y[1], @t[3], | |
249 | @x[2], @x[3], @y[2], @y[3], @t[2]); | |
250 | $code.=<<___; | |
251 | veor @x[0], @x[0], @t[0] | |
252 | veor @x[2], @x[2], @t[0] | |
253 | veor @x[1], @x[1], @t[1] | |
254 | veor @x[3], @x[3], @t[1] | |
255 | ||
256 | veor @t[0], @x[4], @x[6] | |
257 | veor @t[1], @x[5], @x[7] | |
258 | ___ | |
259 | &Mul_GF4_N_GF4 (@t[0], @t[1], @y[0], @y[1], @t[3], | |
260 | @x[6], @x[7], @y[2], @y[3], @t[2]); | |
261 | $code.=<<___; | |
262 | veor @y[0], @y[0], @y[2] | |
263 | veor @y[1], @y[1], @y[3] | |
264 | ___ | |
a903e691 | 265 | &Mul_GF4 (@x[4], @x[5], @y[0], @y[1], @t[2..3]); |
c4a52a6d AP |
266 | $code.=<<___; |
267 | veor @x[4], @x[4], @t[0] | |
268 | veor @x[6], @x[6], @t[0] | |
269 | veor @x[5], @x[5], @t[1] | |
270 | veor @x[7], @x[7], @t[1] | |
271 | ___ | |
272 | } | |
273 | sub Inv_GF256 { | |
274 | #;******************************************************************** | |
275 | #;* Inv_GF256: Input x0-x7 Output x0-x7 Temp t0-t3,s0-s3 (144) * | |
276 | #;******************************************************************** | |
277 | my @x=@_[0..7]; | |
278 | my @t=@_[8..11]; | |
279 | my @s=@_[12..15]; | |
280 | # direct optimizations from hardware | |
281 | $code.=<<___; | |
282 | veor @t[3], @x[4], @x[6] | |
283 | veor @t[2], @x[5], @x[7] | |
284 | veor @t[1], @x[1], @x[3] | |
285 | veor @s[1], @x[7], @x[6] | |
286 | vmov @t[0], @t[2] | |
287 | veor @s[0], @x[0], @x[2] | |
288 | ||
289 | vorr @t[2], @t[2], @t[1] | |
290 | veor @s[3], @t[3], @t[0] | |
291 | vand @s[2], @t[3], @s[0] | |
292 | vorr @t[3], @t[3], @s[0] | |
293 | veor @s[0], @s[0], @t[1] | |
294 | vand @t[0], @t[0], @t[1] | |
4f16215b | 295 | veor @t[1], @x[3], @x[2] |
c4a52a6d | 296 | vand @s[3], @s[3], @s[0] |
4f16215b AP |
297 | vand @s[1], @s[1], @t[1] |
298 | veor @t[1], @x[4], @x[5] | |
299 | veor @s[0], @x[1], @x[0] | |
c4a52a6d AP |
300 | veor @t[3], @t[3], @s[1] |
301 | veor @t[2], @t[2], @s[1] | |
4f16215b AP |
302 | vand @s[1], @t[1], @s[0] |
303 | vorr @t[1], @t[1], @s[0] | |
c4a52a6d | 304 | veor @t[3], @t[3], @s[3] |
4f16215b | 305 | veor @t[0], @t[0], @s[1] |
c4a52a6d AP |
306 | veor @t[2], @t[2], @s[2] |
307 | veor @t[1], @t[1], @s[3] | |
308 | veor @t[0], @t[0], @s[2] | |
c4a52a6d | 309 | vand @s[0], @x[7], @x[3] |
4f16215b | 310 | veor @t[1], @t[1], @s[2] |
c4a52a6d AP |
311 | vand @s[1], @x[6], @x[2] |
312 | vand @s[2], @x[5], @x[1] | |
313 | vorr @s[3], @x[4], @x[0] | |
314 | veor @t[3], @t[3], @s[0] | |
c4a52a6d AP |
315 | veor @t[1], @t[1], @s[2] |
316 | veor @t[0], @t[0], @s[3] | |
e7db9896 | 317 | veor @t[2], @t[2], @s[1] |
c4a52a6d AP |
318 | |
319 | @ Inv_GF16 \t0, \t1, \t2, \t3, \s0, \s1, \s2, \s3 | |
320 | ||
321 | @ new smaller inversion | |
322 | ||
e7db9896 AP |
323 | vand @s[2], @t[3], @t[1] |
324 | vmov @s[0], @t[0] | |
c4a52a6d | 325 | |
e7db9896 AP |
326 | veor @s[1], @t[2], @s[2] |
327 | veor @s[3], @t[0], @s[2] | |
328 | veor @s[2], @t[0], @s[2] @ @s[2]=@s[3] | |
442c9f13 | 329 | |
442c9f13 | 330 | vbsl @s[1], @t[1], @t[0] |
e7db9896 AP |
331 | vbsl @s[3], @t[3], @t[2] |
332 | veor @t[3], @t[3], @t[2] | |
c4a52a6d | 333 | |
e7db9896 | 334 | vbsl @s[0], @s[1], @s[2] |
442c9f13 | 335 | vbsl @t[0], @s[2], @s[1] |
c4a52a6d | 336 | |
e7db9896 | 337 | vand @s[2], @s[0], @s[3] |
442c9f13 | 338 | veor @t[1], @t[1], @t[0] |
c4a52a6d | 339 | |
e7db9896 | 340 | veor @s[2], @s[2], @t[3] |
c4a52a6d AP |
341 | ___ |
342 | # output in s3, s2, s1, t1 | |
343 | ||
344 | # Mul_GF16_2 \x0, \x1, \x2, \x3, \x4, \x5, \x6, \x7, \t2, \t3, \t0, \t1, \s0, \s1, \s2, \s3 | |
345 | ||
346 | # Mul_GF16_2 \x0, \x1, \x2, \x3, \x4, \x5, \x6, \x7, \s3, \s2, \s1, \t1, \s0, \t0, \t2, \t3 | |
347 | &Mul_GF16_2(@x,@s[3,2,1],@t[1],@s[0],@t[0,2,3]); | |
348 | ||
349 | ### output msb > [x3,x2,x1,x0,x7,x6,x5,x4] < lsb | |
350 | } | |
351 | ||
352 | # AES linear components | |
353 | ||
354 | sub ShiftRows { | |
355 | my @x=@_[0..7]; | |
356 | my @t=@_[8..11]; | |
357 | my $mask=pop; | |
358 | $code.=<<___; | |
359 | vldmia $key!, {@t[0]-@t[3]} | |
360 | veor @t[0], @t[0], @x[0] | |
361 | veor @t[1], @t[1], @x[1] | |
362 | vtbl.8 `&Dlo(@x[0])`, {@t[0]}, `&Dlo($mask)` | |
363 | vtbl.8 `&Dhi(@x[0])`, {@t[0]}, `&Dhi($mask)` | |
364 | vldmia $key!, {@t[0]} | |
365 | veor @t[2], @t[2], @x[2] | |
366 | vtbl.8 `&Dlo(@x[1])`, {@t[1]}, `&Dlo($mask)` | |
367 | vtbl.8 `&Dhi(@x[1])`, {@t[1]}, `&Dhi($mask)` | |
368 | vldmia $key!, {@t[1]} | |
369 | veor @t[3], @t[3], @x[3] | |
370 | vtbl.8 `&Dlo(@x[2])`, {@t[2]}, `&Dlo($mask)` | |
371 | vtbl.8 `&Dhi(@x[2])`, {@t[2]}, `&Dhi($mask)` | |
372 | vldmia $key!, {@t[2]} | |
373 | vtbl.8 `&Dlo(@x[3])`, {@t[3]}, `&Dlo($mask)` | |
374 | vtbl.8 `&Dhi(@x[3])`, {@t[3]}, `&Dhi($mask)` | |
375 | vldmia $key!, {@t[3]} | |
376 | veor @t[0], @t[0], @x[4] | |
377 | veor @t[1], @t[1], @x[5] | |
378 | vtbl.8 `&Dlo(@x[4])`, {@t[0]}, `&Dlo($mask)` | |
379 | vtbl.8 `&Dhi(@x[4])`, {@t[0]}, `&Dhi($mask)` | |
380 | veor @t[2], @t[2], @x[6] | |
381 | vtbl.8 `&Dlo(@x[5])`, {@t[1]}, `&Dlo($mask)` | |
382 | vtbl.8 `&Dhi(@x[5])`, {@t[1]}, `&Dhi($mask)` | |
383 | veor @t[3], @t[3], @x[7] | |
384 | vtbl.8 `&Dlo(@x[6])`, {@t[2]}, `&Dlo($mask)` | |
385 | vtbl.8 `&Dhi(@x[6])`, {@t[2]}, `&Dhi($mask)` | |
386 | vtbl.8 `&Dlo(@x[7])`, {@t[3]}, `&Dlo($mask)` | |
387 | vtbl.8 `&Dhi(@x[7])`, {@t[3]}, `&Dhi($mask)` | |
388 | ___ | |
389 | } | |
390 | ||
391 | sub MixColumns { | |
392 | # modified to emit output in order suitable for feeding back to aesenc[last] | |
393 | my @x=@_[0..7]; | |
394 | my @t=@_[8..15]; | |
6f6a6130 | 395 | my $inv=@_[16]; # optional |
c4a52a6d AP |
396 | $code.=<<___; |
397 | vext.8 @t[0], @x[0], @x[0], #12 @ x0 <<< 32 | |
398 | vext.8 @t[1], @x[1], @x[1], #12 | |
399 | veor @x[0], @x[0], @t[0] @ x0 ^ (x0 <<< 32) | |
400 | vext.8 @t[2], @x[2], @x[2], #12 | |
401 | veor @x[1], @x[1], @t[1] | |
402 | vext.8 @t[3], @x[3], @x[3], #12 | |
403 | veor @x[2], @x[2], @t[2] | |
404 | vext.8 @t[4], @x[4], @x[4], #12 | |
405 | veor @x[3], @x[3], @t[3] | |
406 | vext.8 @t[5], @x[5], @x[5], #12 | |
407 | veor @x[4], @x[4], @t[4] | |
408 | vext.8 @t[6], @x[6], @x[6], #12 | |
409 | veor @x[5], @x[5], @t[5] | |
410 | vext.8 @t[7], @x[7], @x[7], #12 | |
411 | veor @x[6], @x[6], @t[6] | |
c4a52a6d AP |
412 | |
413 | veor @t[1], @t[1], @x[0] | |
4f16215b | 414 | veor @x[7], @x[7], @t[7] |
c4a52a6d | 415 | vext.8 @x[0], @x[0], @x[0], #8 @ (x0 ^ (x0 <<< 32)) <<< 64) |
4f16215b | 416 | veor @t[2], @t[2], @x[1] |
c4a52a6d AP |
417 | veor @t[0], @t[0], @x[7] |
418 | veor @t[1], @t[1], @x[7] | |
c4a52a6d AP |
419 | vext.8 @x[1], @x[1], @x[1], #8 |
420 | veor @t[5], @t[5], @x[4] | |
421 | veor @x[0], @x[0], @t[0] | |
422 | veor @t[6], @t[6], @x[5] | |
423 | veor @x[1], @x[1], @t[1] | |
424 | vext.8 @t[0], @x[4], @x[4], #8 | |
425 | veor @t[4], @t[4], @x[3] | |
426 | vext.8 @t[1], @x[5], @x[5], #8 | |
427 | veor @t[7], @t[7], @x[6] | |
428 | vext.8 @x[4], @x[3], @x[3], #8 | |
429 | veor @t[3], @t[3], @x[2] | |
430 | vext.8 @x[5], @x[7], @x[7], #8 | |
c4a52a6d | 431 | veor @t[4], @t[4], @x[7] |
4f16215b AP |
432 | vext.8 @x[3], @x[6], @x[6], #8 |
433 | veor @t[3], @t[3], @x[7] | |
c4a52a6d | 434 | vext.8 @x[6], @x[2], @x[2], #8 |
c4a52a6d | 435 | veor @x[7], @t[1], @t[5] |
6f6a6130 AP |
436 | ___ |
437 | $code.=<<___ if (!$inv); | |
a903e691 | 438 | veor @x[2], @t[0], @t[4] |
c4a52a6d AP |
439 | veor @x[4], @x[4], @t[3] |
440 | veor @x[5], @x[5], @t[7] | |
441 | veor @x[3], @x[3], @t[6] | |
442 | @ vmov @x[2], @t[0] | |
443 | veor @x[6], @x[6], @t[2] | |
444 | @ vmov @x[7], @t[1] | |
445 | ___ | |
6f6a6130 AP |
446 | $code.=<<___ if ($inv); |
447 | veor @t[3], @t[3], @x[4] | |
448 | veor @x[5], @x[5], @t[7] | |
449 | veor @x[2], @x[3], @t[6] | |
450 | veor @x[3], @t[0], @t[4] | |
451 | veor @x[4], @x[6], @t[2] | |
452 | vmov @x[6], @t[3] | |
453 | @ vmov @x[7], @t[1] | |
454 | ___ | |
c4a52a6d AP |
455 | } |
456 | ||
6f6a6130 | 457 | sub InvMixColumns_orig { |
c4a52a6d AP |
458 | my @x=@_[0..7]; |
459 | my @t=@_[8..15]; | |
460 | ||
461 | $code.=<<___; | |
462 | @ multiplication by 0x0e | |
463 | vext.8 @t[7], @x[7], @x[7], #12 | |
464 | vmov @t[2], @x[2] | |
c4a52a6d | 465 | veor @x[2], @x[2], @x[5] @ 2 5 |
a903e691 | 466 | veor @x[7], @x[7], @x[5] @ 7 5 |
c4a52a6d AP |
467 | vext.8 @t[0], @x[0], @x[0], #12 |
468 | vmov @t[5], @x[5] | |
469 | veor @x[5], @x[5], @x[0] @ 5 0 [1] | |
470 | veor @x[0], @x[0], @x[1] @ 0 1 | |
471 | vext.8 @t[1], @x[1], @x[1], #12 | |
472 | veor @x[1], @x[1], @x[2] @ 1 25 | |
473 | veor @x[0], @x[0], @x[6] @ 01 6 [2] | |
c4a52a6d | 474 | vext.8 @t[3], @x[3], @x[3], #12 |
a903e691 | 475 | veor @x[1], @x[1], @x[3] @ 125 3 [4] |
c4a52a6d AP |
476 | veor @x[2], @x[2], @x[0] @ 25 016 [3] |
477 | veor @x[3], @x[3], @x[7] @ 3 75 | |
478 | veor @x[7], @x[7], @x[6] @ 75 6 [0] | |
479 | vext.8 @t[6], @x[6], @x[6], #12 | |
480 | vmov @t[4], @x[4] | |
481 | veor @x[6], @x[6], @x[4] @ 6 4 | |
482 | veor @x[4], @x[4], @x[3] @ 4 375 [6] | |
483 | veor @x[3], @x[3], @x[7] @ 375 756=36 | |
484 | veor @x[6], @x[6], @t[5] @ 64 5 [7] | |
485 | veor @x[3], @x[3], @t[2] @ 36 2 | |
486 | vext.8 @t[5], @t[5], @t[5], #12 | |
487 | veor @x[3], @x[3], @t[4] @ 362 4 [5] | |
488 | ___ | |
489 | my @y = @x[7,5,0,2,1,3,4,6]; | |
490 | $code.=<<___; | |
491 | @ multiplication by 0x0b | |
492 | veor @y[1], @y[1], @y[0] | |
493 | veor @y[0], @y[0], @t[0] | |
c4a52a6d | 494 | vext.8 @t[2], @t[2], @t[2], #12 |
a903e691 | 495 | veor @y[1], @y[1], @t[1] |
c4a52a6d | 496 | veor @y[0], @y[0], @t[5] |
a903e691 | 497 | vext.8 @t[4], @t[4], @t[4], #12 |
c4a52a6d AP |
498 | veor @y[1], @y[1], @t[6] |
499 | veor @y[0], @y[0], @t[7] | |
c4a52a6d | 500 | veor @t[7], @t[7], @t[6] @ clobber t[7] |
c4a52a6d AP |
501 | |
502 | veor @y[3], @y[3], @t[0] | |
a903e691 | 503 | veor @y[1], @y[1], @y[0] |
c4a52a6d AP |
504 | vext.8 @t[0], @t[0], @t[0], #12 |
505 | veor @y[2], @y[2], @t[1] | |
506 | veor @y[4], @y[4], @t[1] | |
c4a52a6d | 507 | vext.8 @t[1], @t[1], @t[1], #12 |
a903e691 | 508 | veor @y[2], @y[2], @t[2] |
c4a52a6d AP |
509 | veor @y[3], @y[3], @t[2] |
510 | veor @y[5], @y[5], @t[2] | |
511 | veor @y[2], @y[2], @t[7] | |
512 | vext.8 @t[2], @t[2], @t[2], #12 | |
513 | veor @y[3], @y[3], @t[3] | |
514 | veor @y[6], @y[6], @t[3] | |
515 | veor @y[4], @y[4], @t[3] | |
c4a52a6d | 516 | veor @y[7], @y[7], @t[4] |
a903e691 | 517 | vext.8 @t[3], @t[3], @t[3], #12 |
c4a52a6d AP |
518 | veor @y[5], @y[5], @t[4] |
519 | veor @y[7], @y[7], @t[7] | |
4f16215b | 520 | veor @t[7], @t[7], @t[5] @ clobber t[7] even more |
c4a52a6d AP |
521 | veor @y[3], @y[3], @t[5] |
522 | veor @y[4], @y[4], @t[4] | |
c4a52a6d AP |
523 | |
524 | veor @y[5], @y[5], @t[7] | |
525 | vext.8 @t[4], @t[4], @t[4], #12 | |
526 | veor @y[6], @y[6], @t[7] | |
527 | veor @y[4], @y[4], @t[7] | |
528 | ||
529 | veor @t[7], @t[7], @t[5] | |
530 | vext.8 @t[5], @t[5], @t[5], #12 | |
c4a52a6d AP |
531 | |
532 | @ multiplication by 0x0d | |
533 | veor @y[4], @y[4], @y[7] | |
a903e691 | 534 | veor @t[7], @t[7], @t[6] @ restore t[7] |
c4a52a6d AP |
535 | veor @y[7], @y[7], @t[4] |
536 | vext.8 @t[6], @t[6], @t[6], #12 | |
537 | veor @y[2], @y[2], @t[0] | |
538 | veor @y[7], @y[7], @t[5] | |
c4a52a6d | 539 | vext.8 @t[7], @t[7], @t[7], #12 |
a903e691 | 540 | veor @y[2], @y[2], @t[2] |
c4a52a6d AP |
541 | |
542 | veor @y[3], @y[3], @y[1] | |
543 | veor @y[1], @y[1], @t[1] | |
544 | veor @y[0], @y[0], @t[0] | |
545 | veor @y[3], @y[3], @t[0] | |
546 | veor @y[1], @y[1], @t[5] | |
547 | veor @y[0], @y[0], @t[5] | |
c4a52a6d | 548 | vext.8 @t[0], @t[0], @t[0], #12 |
a903e691 | 549 | veor @y[1], @y[1], @t[7] |
c4a52a6d AP |
550 | veor @y[0], @y[0], @t[6] |
551 | veor @y[3], @y[3], @y[1] | |
552 | veor @y[4], @y[4], @t[1] | |
553 | vext.8 @t[1], @t[1], @t[1], #12 | |
554 | ||
555 | veor @y[7], @y[7], @t[7] | |
556 | veor @y[4], @y[4], @t[2] | |
557 | veor @y[5], @y[5], @t[2] | |
c4a52a6d AP |
558 | veor @y[2], @y[2], @t[6] |
559 | veor @t[6], @t[6], @t[3] @ clobber t[6] | |
a903e691 | 560 | vext.8 @t[2], @t[2], @t[2], #12 |
c4a52a6d AP |
561 | veor @y[4], @y[4], @y[7] |
562 | veor @y[3], @y[3], @t[6] | |
563 | ||
564 | veor @y[6], @y[6], @t[6] | |
565 | veor @y[5], @y[5], @t[5] | |
566 | vext.8 @t[5], @t[5], @t[5], #12 | |
567 | veor @y[6], @y[6], @t[4] | |
568 | vext.8 @t[4], @t[4], @t[4], #12 | |
569 | veor @y[5], @y[5], @t[6] | |
570 | veor @y[6], @y[6], @t[7] | |
571 | vext.8 @t[7], @t[7], @t[7], #12 | |
572 | veor @t[6], @t[6], @t[3] @ restore t[6] | |
573 | vext.8 @t[3], @t[3], @t[3], #12 | |
574 | ||
575 | @ multiplication by 0x09 | |
576 | veor @y[4], @y[4], @y[1] | |
577 | veor @t[1], @t[1], @y[1] @ t[1]=y[1] | |
578 | veor @t[0], @t[0], @t[5] @ clobber t[0] | |
579 | vext.8 @t[6], @t[6], @t[6], #12 | |
580 | veor @t[1], @t[1], @t[5] | |
581 | veor @y[3], @y[3], @t[0] | |
582 | veor @t[0], @t[0], @y[0] @ t[0]=y[0] | |
583 | veor @t[1], @t[1], @t[6] | |
584 | veor @t[6], @t[6], @t[7] @ clobber t[6] | |
585 | veor @y[4], @y[4], @t[1] | |
586 | veor @y[7], @y[7], @t[4] | |
c4a52a6d | 587 | veor @y[6], @y[6], @t[3] |
c4a52a6d | 588 | veor @y[5], @y[5], @t[2] |
a903e691 AP |
589 | veor @t[4], @t[4], @y[4] @ t[4]=y[4] |
590 | veor @t[3], @t[3], @y[3] @ t[3]=y[3] | |
591 | veor @t[5], @t[5], @y[5] @ t[5]=y[5] | |
c4a52a6d AP |
592 | veor @t[2], @t[2], @y[2] @ t[2]=y[2] |
593 | veor @t[3], @t[3], @t[7] | |
c4a52a6d AP |
594 | veor @XMM[5], @t[5], @t[6] |
595 | veor @XMM[6], @t[6], @y[6] @ t[6]=y[6] | |
596 | veor @XMM[2], @t[2], @t[6] | |
597 | veor @XMM[7], @t[7], @y[7] @ t[7]=y[7] | |
598 | ||
599 | vmov @XMM[0], @t[0] | |
600 | vmov @XMM[1], @t[1] | |
601 | @ vmov @XMM[2], @t[2] | |
602 | vmov @XMM[3], @t[3] | |
603 | vmov @XMM[4], @t[4] | |
604 | @ vmov @XMM[5], @t[5] | |
605 | @ vmov @XMM[6], @t[6] | |
606 | @ vmov @XMM[7], @t[7] | |
607 | ___ | |
608 | } | |
609 | ||
6f6a6130 AP |
610 | sub InvMixColumns { |
611 | my @x=@_[0..7]; | |
612 | my @t=@_[8..15]; | |
613 | ||
614 | # Thanks to Jussi Kivilinna for providing pointer to | |
615 | # | |
616 | # | 0e 0b 0d 09 | | 02 03 01 01 | | 05 00 04 00 | | |
617 | # | 09 0e 0b 0d | = | 01 02 03 01 | x | 00 05 00 04 | | |
618 | # | 0d 09 0e 0b | | 01 01 02 03 | | 04 00 05 00 | | |
619 | # | 0b 0d 09 0e | | 03 01 01 02 | | 00 04 00 05 | | |
620 | ||
621 | $code.=<<___; | |
622 | @ multiplication by 0x05-0x00-0x04-0x00 | |
623 | vext.8 @t[0], @x[0], @x[0], #8 | |
624 | vext.8 @t[6], @x[6], @x[6], #8 | |
625 | vext.8 @t[7], @x[7], @x[7], #8 | |
626 | veor @t[0], @t[0], @x[0] | |
627 | vext.8 @t[1], @x[1], @x[1], #8 | |
628 | veor @t[6], @t[6], @x[6] | |
629 | vext.8 @t[2], @x[2], @x[2], #8 | |
630 | veor @t[7], @t[7], @x[7] | |
631 | vext.8 @t[3], @x[3], @x[3], #8 | |
632 | veor @t[1], @t[1], @x[1] | |
633 | vext.8 @t[4], @x[4], @x[4], #8 | |
634 | veor @t[2], @t[2], @x[2] | |
635 | vext.8 @t[5], @x[5], @x[5], #8 | |
636 | veor @t[3], @t[3], @x[3] | |
637 | veor @t[4], @t[4], @x[4] | |
638 | veor @t[5], @t[5], @x[5] | |
639 | ||
640 | veor @x[0], @x[0], @t[6] | |
641 | veor @x[1], @x[1], @t[6] | |
642 | veor @x[2], @x[2], @t[0] | |
643 | veor @x[4], @x[4], @t[2] | |
644 | veor @x[3], @x[3], @t[1] | |
645 | veor @x[1], @x[1], @t[7] | |
646 | veor @x[2], @x[2], @t[7] | |
647 | veor @x[4], @x[4], @t[6] | |
648 | veor @x[5], @x[5], @t[3] | |
649 | veor @x[3], @x[3], @t[6] | |
650 | veor @x[6], @x[6], @t[4] | |
651 | veor @x[4], @x[4], @t[7] | |
652 | veor @x[5], @x[5], @t[7] | |
653 | veor @x[7], @x[7], @t[5] | |
654 | ___ | |
655 | &MixColumns (@x,@t,1); # flipped 2<->3 and 4<->6 | |
656 | } | |
657 | ||
c4a52a6d AP |
658 | sub swapmove { |
659 | my ($a,$b,$n,$mask,$t)=@_; | |
660 | $code.=<<___; | |
661 | vshr.u64 $t, $b, #$n | |
662 | veor $t, $t, $a | |
663 | vand $t, $t, $mask | |
664 | veor $a, $a, $t | |
665 | vshl.u64 $t, $t, #$n | |
666 | veor $b, $b, $t | |
667 | ___ | |
668 | } | |
669 | sub swapmove2x { | |
670 | my ($a0,$b0,$a1,$b1,$n,$mask,$t0,$t1)=@_; | |
671 | $code.=<<___; | |
672 | vshr.u64 $t0, $b0, #$n | |
673 | vshr.u64 $t1, $b1, #$n | |
674 | veor $t0, $t0, $a0 | |
675 | veor $t1, $t1, $a1 | |
676 | vand $t0, $t0, $mask | |
677 | vand $t1, $t1, $mask | |
678 | veor $a0, $a0, $t0 | |
679 | vshl.u64 $t0, $t0, #$n | |
680 | veor $a1, $a1, $t1 | |
681 | vshl.u64 $t1, $t1, #$n | |
682 | veor $b0, $b0, $t0 | |
683 | veor $b1, $b1, $t1 | |
684 | ___ | |
685 | } | |
686 | ||
687 | sub bitslice { | |
688 | my @x=reverse(@_[0..7]); | |
689 | my ($t0,$t1,$t2,$t3)=@_[8..11]; | |
690 | $code.=<<___; | |
691 | vmov.i8 $t0,#0x55 @ compose .LBS0 | |
692 | vmov.i8 $t1,#0x33 @ compose .LBS1 | |
693 | ___ | |
694 | &swapmove2x(@x[0,1,2,3],1,$t0,$t2,$t3); | |
695 | &swapmove2x(@x[4,5,6,7],1,$t0,$t2,$t3); | |
696 | $code.=<<___; | |
697 | vmov.i8 $t0,#0x0f @ compose .LBS2 | |
698 | ___ | |
699 | &swapmove2x(@x[0,2,1,3],2,$t1,$t2,$t3); | |
700 | &swapmove2x(@x[4,6,5,7],2,$t1,$t2,$t3); | |
701 | ||
702 | &swapmove2x(@x[0,4,1,5],4,$t0,$t2,$t3); | |
703 | &swapmove2x(@x[2,6,3,7],4,$t0,$t2,$t3); | |
704 | } | |
705 | ||
706 | $code.=<<___; | |
e0202d94 AP |
707 | #ifndef __KERNEL__ |
708 | # include "arm_arch.h" | |
709 | ||
710 | # define VFP_ABI_PUSH vstmdb sp!,{d8-d15} | |
711 | # define VFP_ABI_POP vldmia sp!,{d8-d15} | |
712 | # define VFP_ABI_FRAME 0x40 | |
713 | #else | |
714 | # define VFP_ABI_PUSH | |
715 | # define VFP_ABI_POP | |
716 | # define VFP_ABI_FRAME 0 | |
717 | # define BSAES_ASM_EXTENDED_KEY | |
718 | # define XTS_CHAIN_TWEAK | |
719 | # define __ARM_ARCH__ __LINUX_ARM_ARCH__ | |
e620e5ae | 720 | # define __ARM_MAX_ARCH__ 7 |
e0202d94 AP |
721 | #endif |
722 | ||
723 | #ifdef __thumb__ | |
724 | # define adrl adr | |
725 | #endif | |
75fe4223 | 726 | |
c1669e1c AP |
727 | #if __ARM_MAX_ARCH__>=7 |
728 | .arch armv7-a | |
729 | .fpu neon | |
730 | ||
c4a52a6d | 731 | .text |
e0202d94 | 732 | .syntax unified @ ARMv7-capable assembler is expected to handle this |
313e6ec1 | 733 | #if defined(__thumb2__) && !defined(__APPLE__) |
066caf05 AP |
734 | .thumb |
735 | #else | |
736 | .code 32 | |
a2859927 | 737 | # undef __thumb2__ |
066caf05 AP |
738 | #endif |
739 | ||
c4a52a6d AP |
740 | .type _bsaes_decrypt8,%function |
741 | .align 4 | |
742 | _bsaes_decrypt8: | |
b82acc3c | 743 | adr $const,. |
c4a52a6d | 744 | vldmia $key!, {@XMM[9]} @ round 0 key |
8a5d8bc4 | 745 | #if defined(__thumb2__) || defined(__APPLE__) |
313e6ec1 AP |
746 | adr $const,.LM0ISR |
747 | #else | |
c4a52a6d | 748 | add $const,$const,#.LM0ISR-_bsaes_decrypt8 |
313e6ec1 | 749 | #endif |
c4a52a6d AP |
750 | |
751 | vldmia $const!, {@XMM[8]} @ .LM0ISR | |
752 | veor @XMM[10], @XMM[0], @XMM[9] @ xor with round0 key | |
753 | veor @XMM[11], @XMM[1], @XMM[9] | |
754 | vtbl.8 `&Dlo(@XMM[0])`, {@XMM[10]}, `&Dlo(@XMM[8])` | |
755 | vtbl.8 `&Dhi(@XMM[0])`, {@XMM[10]}, `&Dhi(@XMM[8])` | |
756 | veor @XMM[12], @XMM[2], @XMM[9] | |
757 | vtbl.8 `&Dlo(@XMM[1])`, {@XMM[11]}, `&Dlo(@XMM[8])` | |
758 | vtbl.8 `&Dhi(@XMM[1])`, {@XMM[11]}, `&Dhi(@XMM[8])` | |
759 | veor @XMM[13], @XMM[3], @XMM[9] | |
760 | vtbl.8 `&Dlo(@XMM[2])`, {@XMM[12]}, `&Dlo(@XMM[8])` | |
761 | vtbl.8 `&Dhi(@XMM[2])`, {@XMM[12]}, `&Dhi(@XMM[8])` | |
762 | veor @XMM[14], @XMM[4], @XMM[9] | |
763 | vtbl.8 `&Dlo(@XMM[3])`, {@XMM[13]}, `&Dlo(@XMM[8])` | |
764 | vtbl.8 `&Dhi(@XMM[3])`, {@XMM[13]}, `&Dhi(@XMM[8])` | |
765 | veor @XMM[15], @XMM[5], @XMM[9] | |
766 | vtbl.8 `&Dlo(@XMM[4])`, {@XMM[14]}, `&Dlo(@XMM[8])` | |
767 | vtbl.8 `&Dhi(@XMM[4])`, {@XMM[14]}, `&Dhi(@XMM[8])` | |
768 | veor @XMM[10], @XMM[6], @XMM[9] | |
769 | vtbl.8 `&Dlo(@XMM[5])`, {@XMM[15]}, `&Dlo(@XMM[8])` | |
770 | vtbl.8 `&Dhi(@XMM[5])`, {@XMM[15]}, `&Dhi(@XMM[8])` | |
771 | veor @XMM[11], @XMM[7], @XMM[9] | |
772 | vtbl.8 `&Dlo(@XMM[6])`, {@XMM[10]}, `&Dlo(@XMM[8])` | |
773 | vtbl.8 `&Dhi(@XMM[6])`, {@XMM[10]}, `&Dhi(@XMM[8])` | |
774 | vtbl.8 `&Dlo(@XMM[7])`, {@XMM[11]}, `&Dlo(@XMM[8])` | |
775 | vtbl.8 `&Dhi(@XMM[7])`, {@XMM[11]}, `&Dhi(@XMM[8])` | |
776 | ___ | |
777 | &bitslice (@XMM[0..7, 8..11]); | |
778 | $code.=<<___; | |
779 | sub $rounds,$rounds,#1 | |
780 | b .Ldec_sbox | |
781 | .align 4 | |
782 | .Ldec_loop: | |
783 | ___ | |
784 | &ShiftRows (@XMM[0..7, 8..12]); | |
785 | $code.=".Ldec_sbox:\n"; | |
786 | &InvSbox (@XMM[0..7, 8..15]); | |
787 | $code.=<<___; | |
788 | subs $rounds,$rounds,#1 | |
789 | bcc .Ldec_done | |
790 | ___ | |
791 | &InvMixColumns (@XMM[0,1,6,4,2,7,3,5, 8..15]); | |
792 | $code.=<<___; | |
793 | vldmia $const, {@XMM[12]} @ .LISR | |
e0202d94 | 794 | ite eq @ Thumb2 thing, sanity check in ARM |
c4a52a6d AP |
795 | addeq $const,$const,#0x10 |
796 | bne .Ldec_loop | |
797 | vldmia $const, {@XMM[12]} @ .LISRM0 | |
798 | b .Ldec_loop | |
799 | .align 4 | |
800 | .Ldec_done: | |
801 | ___ | |
802 | &bitslice (@XMM[0,1,6,4,2,7,3,5, 8..11]); | |
803 | $code.=<<___; | |
804 | vldmia $key, {@XMM[8]} @ last round key | |
805 | veor @XMM[6], @XMM[6], @XMM[8] | |
806 | veor @XMM[4], @XMM[4], @XMM[8] | |
807 | veor @XMM[2], @XMM[2], @XMM[8] | |
808 | veor @XMM[7], @XMM[7], @XMM[8] | |
809 | veor @XMM[3], @XMM[3], @XMM[8] | |
810 | veor @XMM[5], @XMM[5], @XMM[8] | |
811 | veor @XMM[0], @XMM[0], @XMM[8] | |
812 | veor @XMM[1], @XMM[1], @XMM[8] | |
813 | bx lr | |
814 | .size _bsaes_decrypt8,.-_bsaes_decrypt8 | |
815 | ||
816 | .type _bsaes_const,%object | |
817 | .align 6 | |
818 | _bsaes_const: | |
819 | .LM0ISR: @ InvShiftRows constants | |
820 | .quad 0x0a0e0206070b0f03, 0x0004080c0d010509 | |
821 | .LISR: | |
822 | .quad 0x0504070602010003, 0x0f0e0d0c080b0a09 | |
823 | .LISRM0: | |
824 | .quad 0x01040b0e0205080f, 0x0306090c00070a0d | |
825 | .LM0SR: @ ShiftRows constants | |
826 | .quad 0x0a0e02060f03070b, 0x0004080c05090d01 | |
827 | .LSR: | |
828 | .quad 0x0504070600030201, 0x0f0e0d0c0a09080b | |
829 | .LSRM0: | |
830 | .quad 0x0304090e00050a0f, 0x01060b0c0207080d | |
831 | .LM0: | |
832 | .quad 0x02060a0e03070b0f, 0x0004080c0105090d | |
9575d1a9 | 833 | .LREVM0SR: |
612f4e23 | 834 | .quad 0x090d01050c000408, 0x03070b0f060a0e02 |
c4a52a6d AP |
835 | .asciz "Bit-sliced AES for NEON, CRYPTOGAMS by <appro\@openssl.org>" |
836 | .align 6 | |
837 | .size _bsaes_const,.-_bsaes_const | |
838 | ||
839 | .type _bsaes_encrypt8,%function | |
840 | .align 4 | |
841 | _bsaes_encrypt8: | |
b82acc3c | 842 | adr $const,. |
c4a52a6d | 843 | vldmia $key!, {@XMM[9]} @ round 0 key |
8a5d8bc4 | 844 | #if defined(__thumb2__) || defined(__APPLE__) |
313e6ec1 AP |
845 | adr $const,.LM0SR |
846 | #else | |
c4a52a6d | 847 | sub $const,$const,#_bsaes_encrypt8-.LM0SR |
313e6ec1 | 848 | #endif |
c4a52a6d AP |
849 | |
850 | vldmia $const!, {@XMM[8]} @ .LM0SR | |
9575d1a9 | 851 | _bsaes_encrypt8_alt: |
c4a52a6d AP |
852 | veor @XMM[10], @XMM[0], @XMM[9] @ xor with round0 key |
853 | veor @XMM[11], @XMM[1], @XMM[9] | |
854 | vtbl.8 `&Dlo(@XMM[0])`, {@XMM[10]}, `&Dlo(@XMM[8])` | |
855 | vtbl.8 `&Dhi(@XMM[0])`, {@XMM[10]}, `&Dhi(@XMM[8])` | |
856 | veor @XMM[12], @XMM[2], @XMM[9] | |
857 | vtbl.8 `&Dlo(@XMM[1])`, {@XMM[11]}, `&Dlo(@XMM[8])` | |
858 | vtbl.8 `&Dhi(@XMM[1])`, {@XMM[11]}, `&Dhi(@XMM[8])` | |
859 | veor @XMM[13], @XMM[3], @XMM[9] | |
860 | vtbl.8 `&Dlo(@XMM[2])`, {@XMM[12]}, `&Dlo(@XMM[8])` | |
861 | vtbl.8 `&Dhi(@XMM[2])`, {@XMM[12]}, `&Dhi(@XMM[8])` | |
862 | veor @XMM[14], @XMM[4], @XMM[9] | |
863 | vtbl.8 `&Dlo(@XMM[3])`, {@XMM[13]}, `&Dlo(@XMM[8])` | |
864 | vtbl.8 `&Dhi(@XMM[3])`, {@XMM[13]}, `&Dhi(@XMM[8])` | |
865 | veor @XMM[15], @XMM[5], @XMM[9] | |
866 | vtbl.8 `&Dlo(@XMM[4])`, {@XMM[14]}, `&Dlo(@XMM[8])` | |
867 | vtbl.8 `&Dhi(@XMM[4])`, {@XMM[14]}, `&Dhi(@XMM[8])` | |
868 | veor @XMM[10], @XMM[6], @XMM[9] | |
869 | vtbl.8 `&Dlo(@XMM[5])`, {@XMM[15]}, `&Dlo(@XMM[8])` | |
870 | vtbl.8 `&Dhi(@XMM[5])`, {@XMM[15]}, `&Dhi(@XMM[8])` | |
871 | veor @XMM[11], @XMM[7], @XMM[9] | |
872 | vtbl.8 `&Dlo(@XMM[6])`, {@XMM[10]}, `&Dlo(@XMM[8])` | |
873 | vtbl.8 `&Dhi(@XMM[6])`, {@XMM[10]}, `&Dhi(@XMM[8])` | |
874 | vtbl.8 `&Dlo(@XMM[7])`, {@XMM[11]}, `&Dlo(@XMM[8])` | |
875 | vtbl.8 `&Dhi(@XMM[7])`, {@XMM[11]}, `&Dhi(@XMM[8])` | |
876 | _bsaes_encrypt8_bitslice: | |
877 | ___ | |
878 | &bitslice (@XMM[0..7, 8..11]); | |
879 | $code.=<<___; | |
880 | sub $rounds,$rounds,#1 | |
881 | b .Lenc_sbox | |
882 | .align 4 | |
883 | .Lenc_loop: | |
884 | ___ | |
885 | &ShiftRows (@XMM[0..7, 8..12]); | |
886 | $code.=".Lenc_sbox:\n"; | |
887 | &Sbox (@XMM[0..7, 8..15]); | |
888 | $code.=<<___; | |
889 | subs $rounds,$rounds,#1 | |
890 | bcc .Lenc_done | |
891 | ___ | |
892 | &MixColumns (@XMM[0,1,4,6,3,7,2,5, 8..15]); | |
893 | $code.=<<___; | |
894 | vldmia $const, {@XMM[12]} @ .LSR | |
e0202d94 | 895 | ite eq @ Thumb2 thing, samity check in ARM |
c4a52a6d AP |
896 | addeq $const,$const,#0x10 |
897 | bne .Lenc_loop | |
898 | vldmia $const, {@XMM[12]} @ .LSRM0 | |
899 | b .Lenc_loop | |
900 | .align 4 | |
901 | .Lenc_done: | |
902 | ___ | |
903 | # output in lsb > [t0, t1, t4, t6, t3, t7, t2, t5] < msb | |
904 | &bitslice (@XMM[0,1,4,6,3,7,2,5, 8..11]); | |
905 | $code.=<<___; | |
906 | vldmia $key, {@XMM[8]} @ last round key | |
907 | veor @XMM[4], @XMM[4], @XMM[8] | |
908 | veor @XMM[6], @XMM[6], @XMM[8] | |
909 | veor @XMM[3], @XMM[3], @XMM[8] | |
910 | veor @XMM[7], @XMM[7], @XMM[8] | |
911 | veor @XMM[2], @XMM[2], @XMM[8] | |
912 | veor @XMM[5], @XMM[5], @XMM[8] | |
913 | veor @XMM[0], @XMM[0], @XMM[8] | |
914 | veor @XMM[1], @XMM[1], @XMM[8] | |
915 | bx lr | |
916 | .size _bsaes_encrypt8,.-_bsaes_encrypt8 | |
917 | ___ | |
918 | } | |
919 | { | |
920 | my ($out,$inp,$rounds,$const)=("r12","r4","r5","r6"); | |
921 | ||
922 | sub bitslice_key { | |
923 | my @x=reverse(@_[0..7]); | |
924 | my ($bs0,$bs1,$bs2,$t2,$t3)=@_[8..12]; | |
925 | ||
926 | &swapmove (@x[0,1],1,$bs0,$t2,$t3); | |
927 | $code.=<<___; | |
928 | @ &swapmove(@x[2,3],1,$t0,$t2,$t3); | |
929 | vmov @x[2], @x[0] | |
930 | vmov @x[3], @x[1] | |
931 | ___ | |
932 | #&swapmove2x(@x[4,5,6,7],1,$t0,$t2,$t3); | |
933 | ||
934 | &swapmove2x (@x[0,2,1,3],2,$bs1,$t2,$t3); | |
935 | $code.=<<___; | |
936 | @ &swapmove2x(@x[4,6,5,7],2,$t1,$t2,$t3); | |
937 | vmov @x[4], @x[0] | |
938 | vmov @x[6], @x[2] | |
939 | vmov @x[5], @x[1] | |
940 | vmov @x[7], @x[3] | |
941 | ___ | |
942 | &swapmove2x (@x[0,4,1,5],4,$bs2,$t2,$t3); | |
943 | &swapmove2x (@x[2,6,3,7],4,$bs2,$t2,$t3); | |
944 | } | |
945 | ||
946 | $code.=<<___; | |
947 | .type _bsaes_key_convert,%function | |
948 | .align 4 | |
949 | _bsaes_key_convert: | |
b82acc3c | 950 | adr $const,. |
c4a52a6d | 951 | vld1.8 {@XMM[7]}, [$inp]! @ load round 0 key |
8a5d8bc4 | 952 | #if defined(__thumb2__) || defined(__APPLE__) |
313e6ec1 AP |
953 | adr $const,.LM0 |
954 | #else | |
c4a52a6d | 955 | sub $const,$const,#_bsaes_key_convert-.LM0 |
313e6ec1 | 956 | #endif |
c4a52a6d AP |
957 | vld1.8 {@XMM[15]}, [$inp]! @ load round 1 key |
958 | ||
442c9f13 AP |
959 | vmov.i8 @XMM[8], #0x01 @ bit masks |
960 | vmov.i8 @XMM[9], #0x02 | |
961 | vmov.i8 @XMM[10], #0x04 | |
962 | vmov.i8 @XMM[11], #0x08 | |
963 | vmov.i8 @XMM[12], #0x10 | |
964 | vmov.i8 @XMM[13], #0x20 | |
965 | vldmia $const, {@XMM[14]} @ .LM0 | |
c4a52a6d AP |
966 | |
967 | #ifdef __ARMEL__ | |
968 | vrev32.8 @XMM[7], @XMM[7] | |
969 | vrev32.8 @XMM[15], @XMM[15] | |
970 | #endif | |
971 | sub $rounds,$rounds,#1 | |
972 | vstmia $out!, {@XMM[7]} @ save round 0 key | |
973 | b .Lkey_loop | |
974 | ||
975 | .align 4 | |
976 | .Lkey_loop: | |
442c9f13 AP |
977 | vtbl.8 `&Dlo(@XMM[7])`,{@XMM[15]},`&Dlo(@XMM[14])` |
978 | vtbl.8 `&Dhi(@XMM[7])`,{@XMM[15]},`&Dhi(@XMM[14])` | |
979 | vmov.i8 @XMM[6], #0x40 | |
980 | vmov.i8 @XMM[15], #0x80 | |
981 | ||
982 | vtst.8 @XMM[0], @XMM[7], @XMM[8] | |
983 | vtst.8 @XMM[1], @XMM[7], @XMM[9] | |
984 | vtst.8 @XMM[2], @XMM[7], @XMM[10] | |
985 | vtst.8 @XMM[3], @XMM[7], @XMM[11] | |
986 | vtst.8 @XMM[4], @XMM[7], @XMM[12] | |
987 | vtst.8 @XMM[5], @XMM[7], @XMM[13] | |
988 | vtst.8 @XMM[6], @XMM[7], @XMM[6] | |
989 | vtst.8 @XMM[7], @XMM[7], @XMM[15] | |
c4a52a6d | 990 | vld1.8 {@XMM[15]}, [$inp]! @ load next round key |
442c9f13 | 991 | vmvn @XMM[0], @XMM[0] @ "pnot" |
c4a52a6d | 992 | vmvn @XMM[1], @XMM[1] |
442c9f13 AP |
993 | vmvn @XMM[5], @XMM[5] |
994 | vmvn @XMM[6], @XMM[6] | |
c4a52a6d AP |
995 | #ifdef __ARMEL__ |
996 | vrev32.8 @XMM[15], @XMM[15] | |
997 | #endif | |
998 | subs $rounds,$rounds,#1 | |
999 | vstmia $out!,{@XMM[0]-@XMM[7]} @ write bit-sliced round key | |
1000 | bne .Lkey_loop | |
1001 | ||
1002 | vmov.i8 @XMM[7],#0x63 @ compose .L63 | |
1003 | @ don't save last round key | |
1004 | bx lr | |
1005 | .size _bsaes_key_convert,.-_bsaes_key_convert | |
1006 | ___ | |
1007 | } | |
1008 | ||
9575d1a9 | 1009 | if (0) { # following four functions are unsupported interface |
c4a52a6d AP |
1010 | # used for benchmarking... |
1011 | $code.=<<___; | |
1012 | .globl bsaes_enc_key_convert | |
1013 | .type bsaes_enc_key_convert,%function | |
1014 | .align 4 | |
1015 | bsaes_enc_key_convert: | |
1016 | stmdb sp!,{r4-r6,lr} | |
1017 | vstmdb sp!,{d8-d15} @ ABI specification says so | |
1018 | ||
1019 | ldr r5,[$inp,#240] @ pass rounds | |
1020 | mov r4,$inp @ pass key | |
1021 | mov r12,$out @ pass key schedule | |
1022 | bl _bsaes_key_convert | |
1023 | veor @XMM[7],@XMM[7],@XMM[15] @ fix up last round key | |
1024 | vstmia r12, {@XMM[7]} @ save last round key | |
1025 | ||
1026 | vldmia sp!,{d8-d15} | |
1027 | ldmia sp!,{r4-r6,pc} | |
1028 | .size bsaes_enc_key_convert,.-bsaes_enc_key_convert | |
1029 | ||
1030 | .globl bsaes_encrypt_128 | |
1031 | .type bsaes_encrypt_128,%function | |
1032 | .align 4 | |
1033 | bsaes_encrypt_128: | |
1034 | stmdb sp!,{r4-r6,lr} | |
1035 | vstmdb sp!,{d8-d15} @ ABI specification says so | |
1036 | .Lenc128_loop: | |
442c9f13 AP |
1037 | vld1.8 {@XMM[0]-@XMM[1]}, [$inp]! @ load input |
1038 | vld1.8 {@XMM[2]-@XMM[3]}, [$inp]! | |
c4a52a6d | 1039 | mov r4,$key @ pass the key |
442c9f13 | 1040 | vld1.8 {@XMM[4]-@XMM[5]}, [$inp]! |
c4a52a6d | 1041 | mov r5,#10 @ pass rounds |
442c9f13 | 1042 | vld1.8 {@XMM[6]-@XMM[7]}, [$inp]! |
c4a52a6d AP |
1043 | |
1044 | bl _bsaes_encrypt8 | |
1045 | ||
442c9f13 | 1046 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output |
c4a52a6d AP |
1047 | vst1.8 {@XMM[4]}, [$out]! |
1048 | vst1.8 {@XMM[6]}, [$out]! | |
1049 | vst1.8 {@XMM[3]}, [$out]! | |
1050 | vst1.8 {@XMM[7]}, [$out]! | |
1051 | vst1.8 {@XMM[2]}, [$out]! | |
1052 | subs $len,$len,#0x80 | |
1053 | vst1.8 {@XMM[5]}, [$out]! | |
1054 | bhi .Lenc128_loop | |
1055 | ||
1056 | vldmia sp!,{d8-d15} | |
1057 | ldmia sp!,{r4-r6,pc} | |
1058 | .size bsaes_encrypt_128,.-bsaes_encrypt_128 | |
1059 | ||
1060 | .globl bsaes_dec_key_convert | |
1061 | .type bsaes_dec_key_convert,%function | |
1062 | .align 4 | |
1063 | bsaes_dec_key_convert: | |
1064 | stmdb sp!,{r4-r6,lr} | |
1065 | vstmdb sp!,{d8-d15} @ ABI specification says so | |
1066 | ||
1067 | ldr r5,[$inp,#240] @ pass rounds | |
1068 | mov r4,$inp @ pass key | |
1069 | mov r12,$out @ pass key schedule | |
1070 | bl _bsaes_key_convert | |
1071 | vldmia $out, {@XMM[6]} | |
1072 | vstmia r12, {@XMM[15]} @ save last round key | |
1073 | veor @XMM[7], @XMM[7], @XMM[6] @ fix up round 0 key | |
1074 | vstmia $out, {@XMM[7]} | |
1075 | ||
1076 | vldmia sp!,{d8-d15} | |
1077 | ldmia sp!,{r4-r6,pc} | |
1078 | .size bsaes_dec_key_convert,.-bsaes_dec_key_convert | |
1079 | ||
1080 | .globl bsaes_decrypt_128 | |
1081 | .type bsaes_decrypt_128,%function | |
1082 | .align 4 | |
1083 | bsaes_decrypt_128: | |
1084 | stmdb sp!,{r4-r6,lr} | |
1085 | vstmdb sp!,{d8-d15} @ ABI specification says so | |
1086 | .Ldec128_loop: | |
442c9f13 AP |
1087 | vld1.8 {@XMM[0]-@XMM[1]}, [$inp]! @ load input |
1088 | vld1.8 {@XMM[2]-@XMM[3]}, [$inp]! | |
c4a52a6d | 1089 | mov r4,$key @ pass the key |
442c9f13 | 1090 | vld1.8 {@XMM[4]-@XMM[5]}, [$inp]! |
c4a52a6d | 1091 | mov r5,#10 @ pass rounds |
442c9f13 | 1092 | vld1.8 {@XMM[6]-@XMM[7]}, [$inp]! |
c4a52a6d AP |
1093 | |
1094 | bl _bsaes_decrypt8 | |
1095 | ||
442c9f13 | 1096 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output |
c4a52a6d AP |
1097 | vst1.8 {@XMM[6]}, [$out]! |
1098 | vst1.8 {@XMM[4]}, [$out]! | |
1099 | vst1.8 {@XMM[2]}, [$out]! | |
1100 | vst1.8 {@XMM[7]}, [$out]! | |
1101 | vst1.8 {@XMM[3]}, [$out]! | |
1102 | subs $len,$len,#0x80 | |
1103 | vst1.8 {@XMM[5]}, [$out]! | |
1104 | bhi .Ldec128_loop | |
1105 | ||
1106 | vldmia sp!,{d8-d15} | |
1107 | ldmia sp!,{r4-r6,pc} | |
1108 | .size bsaes_decrypt_128,.-bsaes_decrypt_128 | |
1109 | ___ | |
1110 | } | |
9575d1a9 AP |
1111 | { |
1112 | my ($inp,$out,$len,$key, $ivp,$fp,$rounds)=map("r$_",(0..3,8..10)); | |
1113 | my ($keysched)=("sp"); | |
1114 | ||
1115 | $code.=<<___; | |
1116 | .extern AES_cbc_encrypt | |
1117 | .extern AES_decrypt | |
1118 | ||
1119 | .global bsaes_cbc_encrypt | |
1120 | .type bsaes_cbc_encrypt,%function | |
1121 | .align 5 | |
1122 | bsaes_cbc_encrypt: | |
e0202d94 | 1123 | #ifndef __KERNEL__ |
9575d1a9 | 1124 | cmp $len, #128 |
e0202d94 | 1125 | #ifndef __thumb__ |
9575d1a9 | 1126 | blo AES_cbc_encrypt |
e0202d94 AP |
1127 | #else |
1128 | bhs 1f | |
1129 | b AES_cbc_encrypt | |
1130 | 1: | |
1131 | #endif | |
1132 | #endif | |
9575d1a9 AP |
1133 | |
1134 | @ it is up to the caller to make sure we are called with enc == 0 | |
1135 | ||
e0202d94 | 1136 | mov ip, sp |
9575d1a9 | 1137 | stmdb sp!, {r4-r10, lr} |
e0202d94 AP |
1138 | VFP_ABI_PUSH |
1139 | ldr $ivp, [ip] @ IV is 1st arg on the stack | |
9575d1a9 AP |
1140 | mov $len, $len, lsr#4 @ len in 16 byte blocks |
1141 | sub sp, #0x10 @ scratch space to carry over the IV | |
1142 | mov $fp, sp @ save sp | |
1143 | ||
9575d1a9 | 1144 | ldr $rounds, [$key, #240] @ get # of rounds |
e0202d94 AP |
1145 | #ifndef BSAES_ASM_EXTENDED_KEY |
1146 | @ allocate the key schedule on the stack | |
1147 | sub r12, sp, $rounds, lsl#7 @ 128 bytes per inner round key | |
1148 | add r12, #`128-32` @ sifze of bit-slices key schedule | |
9575d1a9 AP |
1149 | |
1150 | @ populate the key schedule | |
1151 | mov r4, $key @ pass key | |
1152 | mov r5, $rounds @ pass # of rounds | |
e0202d94 | 1153 | mov sp, r12 @ sp is $keysched |
9575d1a9 AP |
1154 | bl _bsaes_key_convert |
1155 | vldmia $keysched, {@XMM[6]} | |
1156 | vstmia r12, {@XMM[15]} @ save last round key | |
1157 | veor @XMM[7], @XMM[7], @XMM[6] @ fix up round 0 key | |
1158 | vstmia $keysched, {@XMM[7]} | |
e0202d94 AP |
1159 | #else |
1160 | ldr r12, [$key, #244] | |
1161 | eors r12, #1 | |
1162 | beq 0f | |
1163 | ||
1164 | @ populate the key schedule | |
1165 | str r12, [$key, #244] | |
1166 | mov r4, $key @ pass key | |
1167 | mov r5, $rounds @ pass # of rounds | |
1168 | add r12, $key, #248 @ pass key schedule | |
1169 | bl _bsaes_key_convert | |
1170 | add r4, $key, #248 | |
1171 | vldmia r4, {@XMM[6]} | |
1172 | vstmia r12, {@XMM[15]} @ save last round key | |
1173 | veor @XMM[7], @XMM[7], @XMM[6] @ fix up round 0 key | |
1174 | vstmia r4, {@XMM[7]} | |
1175 | ||
1176 | .align 2 | |
1177 | 0: | |
1178 | #endif | |
9575d1a9 AP |
1179 | |
1180 | vld1.8 {@XMM[15]}, [$ivp] @ load IV | |
1181 | b .Lcbc_dec_loop | |
1182 | ||
1183 | .align 4 | |
1184 | .Lcbc_dec_loop: | |
1185 | subs $len, $len, #0x8 | |
1186 | bmi .Lcbc_dec_loop_finish | |
1187 | ||
1188 | vld1.8 {@XMM[0]-@XMM[1]}, [$inp]! @ load input | |
1189 | vld1.8 {@XMM[2]-@XMM[3]}, [$inp]! | |
e0202d94 | 1190 | #ifndef BSAES_ASM_EXTENDED_KEY |
9575d1a9 | 1191 | mov r4, $keysched @ pass the key |
e0202d94 AP |
1192 | #else |
1193 | add r4, $key, #248 | |
1194 | #endif | |
9575d1a9 AP |
1195 | vld1.8 {@XMM[4]-@XMM[5]}, [$inp]! |
1196 | mov r5, $rounds | |
1197 | vld1.8 {@XMM[6]-@XMM[7]}, [$inp] | |
1198 | sub $inp, $inp, #0x60 | |
1199 | vstmia $fp, {@XMM[15]} @ put aside IV | |
1200 | ||
1201 | bl _bsaes_decrypt8 | |
1202 | ||
1203 | vldmia $fp, {@XMM[14]} @ reload IV | |
1204 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ reload input | |
1205 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1206 | vld1.8 {@XMM[10]-@XMM[11]}, [$inp]! | |
1207 | veor @XMM[1], @XMM[1], @XMM[8] | |
1208 | veor @XMM[6], @XMM[6], @XMM[9] | |
1209 | vld1.8 {@XMM[12]-@XMM[13]}, [$inp]! | |
1210 | veor @XMM[4], @XMM[4], @XMM[10] | |
1211 | veor @XMM[2], @XMM[2], @XMM[11] | |
1212 | vld1.8 {@XMM[14]-@XMM[15]}, [$inp]! | |
1213 | veor @XMM[7], @XMM[7], @XMM[12] | |
1214 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1215 | veor @XMM[3], @XMM[3], @XMM[13] | |
1216 | vst1.8 {@XMM[6]}, [$out]! | |
1217 | veor @XMM[5], @XMM[5], @XMM[14] | |
1218 | vst1.8 {@XMM[4]}, [$out]! | |
1219 | vst1.8 {@XMM[2]}, [$out]! | |
1220 | vst1.8 {@XMM[7]}, [$out]! | |
1221 | vst1.8 {@XMM[3]}, [$out]! | |
1222 | vst1.8 {@XMM[5]}, [$out]! | |
1223 | ||
1224 | b .Lcbc_dec_loop | |
1225 | ||
1226 | .Lcbc_dec_loop_finish: | |
1227 | adds $len, $len, #8 | |
1228 | beq .Lcbc_dec_done | |
1229 | ||
1230 | vld1.8 {@XMM[0]}, [$inp]! @ load input | |
1231 | cmp $len, #2 | |
1232 | blo .Lcbc_dec_one | |
1233 | vld1.8 {@XMM[1]}, [$inp]! | |
e0202d94 | 1234 | #ifndef BSAES_ASM_EXTENDED_KEY |
9575d1a9 | 1235 | mov r4, $keysched @ pass the key |
e0202d94 AP |
1236 | #else |
1237 | add r4, $key, #248 | |
1238 | #endif | |
9575d1a9 AP |
1239 | mov r5, $rounds |
1240 | vstmia $fp, {@XMM[15]} @ put aside IV | |
1241 | beq .Lcbc_dec_two | |
1242 | vld1.8 {@XMM[2]}, [$inp]! | |
1243 | cmp $len, #4 | |
1244 | blo .Lcbc_dec_three | |
1245 | vld1.8 {@XMM[3]}, [$inp]! | |
1246 | beq .Lcbc_dec_four | |
1247 | vld1.8 {@XMM[4]}, [$inp]! | |
1248 | cmp $len, #6 | |
1249 | blo .Lcbc_dec_five | |
1250 | vld1.8 {@XMM[5]}, [$inp]! | |
1251 | beq .Lcbc_dec_six | |
1252 | vld1.8 {@XMM[6]}, [$inp]! | |
1253 | sub $inp, $inp, #0x70 | |
1254 | ||
1255 | bl _bsaes_decrypt8 | |
1256 | ||
1257 | vldmia $fp, {@XMM[14]} @ reload IV | |
1258 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ reload input | |
1259 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1260 | vld1.8 {@XMM[10]-@XMM[11]}, [$inp]! | |
1261 | veor @XMM[1], @XMM[1], @XMM[8] | |
1262 | veor @XMM[6], @XMM[6], @XMM[9] | |
1263 | vld1.8 {@XMM[12]-@XMM[13]}, [$inp]! | |
1264 | veor @XMM[4], @XMM[4], @XMM[10] | |
1265 | veor @XMM[2], @XMM[2], @XMM[11] | |
1266 | vld1.8 {@XMM[15]}, [$inp]! | |
1267 | veor @XMM[7], @XMM[7], @XMM[12] | |
1268 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1269 | veor @XMM[3], @XMM[3], @XMM[13] | |
1270 | vst1.8 {@XMM[6]}, [$out]! | |
1271 | vst1.8 {@XMM[4]}, [$out]! | |
1272 | vst1.8 {@XMM[2]}, [$out]! | |
1273 | vst1.8 {@XMM[7]}, [$out]! | |
1274 | vst1.8 {@XMM[3]}, [$out]! | |
1275 | b .Lcbc_dec_done | |
1276 | .align 4 | |
1277 | .Lcbc_dec_six: | |
1278 | sub $inp, $inp, #0x60 | |
1279 | bl _bsaes_decrypt8 | |
1280 | vldmia $fp,{@XMM[14]} @ reload IV | |
1281 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ reload input | |
1282 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1283 | vld1.8 {@XMM[10]-@XMM[11]}, [$inp]! | |
1284 | veor @XMM[1], @XMM[1], @XMM[8] | |
1285 | veor @XMM[6], @XMM[6], @XMM[9] | |
1286 | vld1.8 {@XMM[12]}, [$inp]! | |
1287 | veor @XMM[4], @XMM[4], @XMM[10] | |
1288 | veor @XMM[2], @XMM[2], @XMM[11] | |
1289 | vld1.8 {@XMM[15]}, [$inp]! | |
1290 | veor @XMM[7], @XMM[7], @XMM[12] | |
1291 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1292 | vst1.8 {@XMM[6]}, [$out]! | |
1293 | vst1.8 {@XMM[4]}, [$out]! | |
1294 | vst1.8 {@XMM[2]}, [$out]! | |
1295 | vst1.8 {@XMM[7]}, [$out]! | |
1296 | b .Lcbc_dec_done | |
1297 | .align 4 | |
1298 | .Lcbc_dec_five: | |
1299 | sub $inp, $inp, #0x50 | |
1300 | bl _bsaes_decrypt8 | |
1301 | vldmia $fp, {@XMM[14]} @ reload IV | |
1302 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ reload input | |
1303 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1304 | vld1.8 {@XMM[10]-@XMM[11]}, [$inp]! | |
1305 | veor @XMM[1], @XMM[1], @XMM[8] | |
1306 | veor @XMM[6], @XMM[6], @XMM[9] | |
1307 | vld1.8 {@XMM[15]}, [$inp]! | |
1308 | veor @XMM[4], @XMM[4], @XMM[10] | |
1309 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1310 | veor @XMM[2], @XMM[2], @XMM[11] | |
1311 | vst1.8 {@XMM[6]}, [$out]! | |
1312 | vst1.8 {@XMM[4]}, [$out]! | |
1313 | vst1.8 {@XMM[2]}, [$out]! | |
1314 | b .Lcbc_dec_done | |
1315 | .align 4 | |
1316 | .Lcbc_dec_four: | |
1317 | sub $inp, $inp, #0x40 | |
1318 | bl _bsaes_decrypt8 | |
1319 | vldmia $fp, {@XMM[14]} @ reload IV | |
1320 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ reload input | |
1321 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1322 | vld1.8 {@XMM[10]}, [$inp]! | |
1323 | veor @XMM[1], @XMM[1], @XMM[8] | |
1324 | veor @XMM[6], @XMM[6], @XMM[9] | |
1325 | vld1.8 {@XMM[15]}, [$inp]! | |
1326 | veor @XMM[4], @XMM[4], @XMM[10] | |
1327 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1328 | vst1.8 {@XMM[6]}, [$out]! | |
1329 | vst1.8 {@XMM[4]}, [$out]! | |
1330 | b .Lcbc_dec_done | |
1331 | .align 4 | |
1332 | .Lcbc_dec_three: | |
1333 | sub $inp, $inp, #0x30 | |
1334 | bl _bsaes_decrypt8 | |
1335 | vldmia $fp, {@XMM[14]} @ reload IV | |
1336 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ reload input | |
1337 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1338 | vld1.8 {@XMM[15]}, [$inp]! | |
1339 | veor @XMM[1], @XMM[1], @XMM[8] | |
1340 | veor @XMM[6], @XMM[6], @XMM[9] | |
1341 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1342 | vst1.8 {@XMM[6]}, [$out]! | |
1343 | b .Lcbc_dec_done | |
1344 | .align 4 | |
1345 | .Lcbc_dec_two: | |
1346 | sub $inp, $inp, #0x20 | |
1347 | bl _bsaes_decrypt8 | |
1348 | vldmia $fp, {@XMM[14]} @ reload IV | |
1349 | vld1.8 {@XMM[8]}, [$inp]! @ reload input | |
1350 | veor @XMM[0], @XMM[0], @XMM[14] @ ^= IV | |
1351 | vld1.8 {@XMM[15]}, [$inp]! @ reload input | |
1352 | veor @XMM[1], @XMM[1], @XMM[8] | |
1353 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1354 | b .Lcbc_dec_done | |
1355 | .align 4 | |
1356 | .Lcbc_dec_one: | |
1357 | sub $inp, $inp, #0x10 | |
1358 | mov $rounds, $out @ save original out pointer | |
1359 | mov $out, $fp @ use the iv scratch space as out buffer | |
1360 | mov r2, $key | |
1361 | vmov @XMM[4],@XMM[15] @ just in case ensure that IV | |
1362 | vmov @XMM[5],@XMM[0] @ and input are preserved | |
1363 | bl AES_decrypt | |
0822d41b | 1364 | vld1.8 {@XMM[0]}, [$fp] @ load result |
9575d1a9 AP |
1365 | veor @XMM[0], @XMM[0], @XMM[4] @ ^= IV |
1366 | vmov @XMM[15], @XMM[5] @ @XMM[5] holds input | |
1367 | vst1.8 {@XMM[0]}, [$rounds] @ write output | |
1368 | ||
1369 | .Lcbc_dec_done: | |
e0202d94 | 1370 | #ifndef BSAES_ASM_EXTENDED_KEY |
9575d1a9 AP |
1371 | vmov.i32 q0, #0 |
1372 | vmov.i32 q1, #0 | |
1373 | .Lcbc_dec_bzero: @ wipe key schedule [if any] | |
1374 | vstmia $keysched!, {q0-q1} | |
e0202d94 | 1375 | cmp $keysched, $fp |
9575d1a9 | 1376 | bne .Lcbc_dec_bzero |
e0202d94 | 1377 | #endif |
9575d1a9 | 1378 | |
e0202d94 AP |
1379 | mov sp, $fp |
1380 | add sp, #0x10 @ add sp,$fp,#0x10 is no good for thumb | |
9575d1a9 | 1381 | vst1.8 {@XMM[15]}, [$ivp] @ return IV |
e0202d94 | 1382 | VFP_ABI_POP |
9575d1a9 AP |
1383 | ldmia sp!, {r4-r10, pc} |
1384 | .size bsaes_cbc_encrypt,.-bsaes_cbc_encrypt | |
1385 | ___ | |
1386 | } | |
1387 | { | |
1388 | my ($inp,$out,$len,$key, $ctr,$fp,$rounds)=(map("r$_",(0..3,8..10))); | |
1389 | my $const = "r6"; # shared with _bsaes_encrypt8_alt | |
1390 | my $keysched = "sp"; | |
1391 | ||
1392 | $code.=<<___; | |
1393 | .extern AES_encrypt | |
1394 | .global bsaes_ctr32_encrypt_blocks | |
1395 | .type bsaes_ctr32_encrypt_blocks,%function | |
1396 | .align 5 | |
1397 | bsaes_ctr32_encrypt_blocks: | |
1398 | cmp $len, #8 @ use plain AES for | |
1399 | blo .Lctr_enc_short @ small sizes | |
1400 | ||
e0202d94 | 1401 | mov ip, sp |
9575d1a9 | 1402 | stmdb sp!, {r4-r10, lr} |
e0202d94 AP |
1403 | VFP_ABI_PUSH |
1404 | ldr $ctr, [ip] @ ctr is 1st arg on the stack | |
9575d1a9 AP |
1405 | sub sp, sp, #0x10 @ scratch space to carry over the ctr |
1406 | mov $fp, sp @ save sp | |
1407 | ||
9575d1a9 | 1408 | ldr $rounds, [$key, #240] @ get # of rounds |
e0202d94 AP |
1409 | #ifndef BSAES_ASM_EXTENDED_KEY |
1410 | @ allocate the key schedule on the stack | |
1411 | sub r12, sp, $rounds, lsl#7 @ 128 bytes per inner round key | |
1412 | add r12, #`128-32` @ size of bit-sliced key schedule | |
9575d1a9 AP |
1413 | |
1414 | @ populate the key schedule | |
1415 | mov r4, $key @ pass key | |
1416 | mov r5, $rounds @ pass # of rounds | |
e0202d94 | 1417 | mov sp, r12 @ sp is $keysched |
9575d1a9 AP |
1418 | bl _bsaes_key_convert |
1419 | veor @XMM[7],@XMM[7],@XMM[15] @ fix up last round key | |
1420 | vstmia r12, {@XMM[7]} @ save last round key | |
1421 | ||
1422 | vld1.8 {@XMM[0]}, [$ctr] @ load counter | |
313e6ec1 | 1423 | #ifdef __APPLE__ |
c4cd8316 | 1424 | mov $ctr, #:lower16:(.LREVM0SR-.LM0) |
313e6ec1 AP |
1425 | add $ctr, $const, $ctr |
1426 | #else | |
9575d1a9 | 1427 | add $ctr, $const, #.LREVM0SR-.LM0 @ borrow $ctr |
313e6ec1 | 1428 | #endif |
9575d1a9 | 1429 | vldmia $keysched, {@XMM[4]} @ load round0 key |
e0202d94 AP |
1430 | #else |
1431 | ldr r12, [$key, #244] | |
1432 | eors r12, #1 | |
1433 | beq 0f | |
1434 | ||
1435 | @ populate the key schedule | |
1436 | str r12, [$key, #244] | |
1437 | mov r4, $key @ pass key | |
1438 | mov r5, $rounds @ pass # of rounds | |
1439 | add r12, $key, #248 @ pass key schedule | |
1440 | bl _bsaes_key_convert | |
1441 | veor @XMM[7],@XMM[7],@XMM[15] @ fix up last round key | |
1442 | vstmia r12, {@XMM[7]} @ save last round key | |
1443 | ||
1444 | .align 2 | |
1445 | 0: add r12, $key, #248 | |
1446 | vld1.8 {@XMM[0]}, [$ctr] @ load counter | |
1447 | adrl $ctr, .LREVM0SR @ borrow $ctr | |
1448 | vldmia r12, {@XMM[4]} @ load round0 key | |
1449 | sub sp, #0x10 @ place for adjusted round0 key | |
1450 | #endif | |
9575d1a9 | 1451 | |
612f4e23 AP |
1452 | vmov.i32 @XMM[8],#1 @ compose 1<<96 |
1453 | veor @XMM[9],@XMM[9],@XMM[9] | |
1454 | vrev32.8 @XMM[0],@XMM[0] | |
1455 | vext.8 @XMM[8],@XMM[9],@XMM[8],#4 | |
1456 | vrev32.8 @XMM[4],@XMM[4] | |
9575d1a9 AP |
1457 | vadd.u32 @XMM[9],@XMM[8],@XMM[8] @ compose 2<<96 |
1458 | vstmia $keysched, {@XMM[4]} @ save adjusted round0 key | |
1459 | b .Lctr_enc_loop | |
1460 | ||
1461 | .align 4 | |
1462 | .Lctr_enc_loop: | |
1463 | vadd.u32 @XMM[10], @XMM[8], @XMM[9] @ compose 3<<96 | |
1464 | vadd.u32 @XMM[1], @XMM[0], @XMM[8] @ +1 | |
1465 | vadd.u32 @XMM[2], @XMM[0], @XMM[9] @ +2 | |
1466 | vadd.u32 @XMM[3], @XMM[0], @XMM[10] @ +3 | |
1467 | vadd.u32 @XMM[4], @XMM[1], @XMM[10] | |
1468 | vadd.u32 @XMM[5], @XMM[2], @XMM[10] | |
1469 | vadd.u32 @XMM[6], @XMM[3], @XMM[10] | |
1470 | vadd.u32 @XMM[7], @XMM[4], @XMM[10] | |
1471 | vadd.u32 @XMM[10], @XMM[5], @XMM[10] @ next counter | |
1472 | ||
1473 | @ Borrow prologue from _bsaes_encrypt8 to use the opportunity | |
1474 | @ to flip byte order in 32-bit counter | |
1475 | ||
1476 | vldmia $keysched, {@XMM[9]} @ load round0 key | |
e0202d94 | 1477 | #ifndef BSAES_ASM_EXTENDED_KEY |
9575d1a9 | 1478 | add r4, $keysched, #0x10 @ pass next round key |
e0202d94 AP |
1479 | #else |
1480 | add r4, $key, #`248+16` | |
1481 | #endif | |
9575d1a9 AP |
1482 | vldmia $ctr, {@XMM[8]} @ .LREVM0SR |
1483 | mov r5, $rounds @ pass rounds | |
1484 | vstmia $fp, {@XMM[10]} @ save next counter | |
313e6ec1 | 1485 | #ifdef __APPLE__ |
c4cd8316 | 1486 | mov $const, #:lower16:(.LREVM0SR-.LSR) |
313e6ec1 AP |
1487 | sub $const, $ctr, $const |
1488 | #else | |
9575d1a9 | 1489 | sub $const, $ctr, #.LREVM0SR-.LSR @ pass constants |
313e6ec1 | 1490 | #endif |
9575d1a9 AP |
1491 | |
1492 | bl _bsaes_encrypt8_alt | |
1493 | ||
1494 | subs $len, $len, #8 | |
1495 | blo .Lctr_enc_loop_done | |
1496 | ||
1497 | vld1.8 {@XMM[8]-@XMM[9]}, [$inp]! @ load input | |
1498 | vld1.8 {@XMM[10]-@XMM[11]}, [$inp]! | |
1499 | veor @XMM[0], @XMM[8] | |
1500 | veor @XMM[1], @XMM[9] | |
1501 | vld1.8 {@XMM[12]-@XMM[13]}, [$inp]! | |
1502 | veor @XMM[4], @XMM[10] | |
1503 | veor @XMM[6], @XMM[11] | |
1504 | vld1.8 {@XMM[14]-@XMM[15]}, [$inp]! | |
1505 | veor @XMM[3], @XMM[12] | |
1506 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! @ write output | |
1507 | veor @XMM[7], @XMM[13] | |
1508 | veor @XMM[2], @XMM[14] | |
1509 | vst1.8 {@XMM[4]}, [$out]! | |
1510 | veor @XMM[5], @XMM[15] | |
1511 | vst1.8 {@XMM[6]}, [$out]! | |
612f4e23 | 1512 | vmov.i32 @XMM[8], #1 @ compose 1<<96 |
9575d1a9 | 1513 | vst1.8 {@XMM[3]}, [$out]! |
612f4e23 | 1514 | veor @XMM[9], @XMM[9], @XMM[9] |
9575d1a9 | 1515 | vst1.8 {@XMM[7]}, [$out]! |
612f4e23 | 1516 | vext.8 @XMM[8], @XMM[9], @XMM[8], #4 |
9575d1a9 AP |
1517 | vst1.8 {@XMM[2]}, [$out]! |
1518 | vadd.u32 @XMM[9],@XMM[8],@XMM[8] @ compose 2<<96 | |
1519 | vst1.8 {@XMM[5]}, [$out]! | |
1520 | vldmia $fp, {@XMM[0]} @ load counter | |
1521 | ||
1522 | bne .Lctr_enc_loop | |
1523 | b .Lctr_enc_done | |
1524 | ||
1525 | .align 4 | |
1526 | .Lctr_enc_loop_done: | |
1527 | add $len, $len, #8 | |
1528 | vld1.8 {@XMM[8]}, [$inp]! @ load input | |
1529 | veor @XMM[0], @XMM[8] | |
1530 | vst1.8 {@XMM[0]}, [$out]! @ write output | |
1531 | cmp $len, #2 | |
1532 | blo .Lctr_enc_done | |
1533 | vld1.8 {@XMM[9]}, [$inp]! | |
1534 | veor @XMM[1], @XMM[9] | |
1535 | vst1.8 {@XMM[1]}, [$out]! | |
1536 | beq .Lctr_enc_done | |
1537 | vld1.8 {@XMM[10]}, [$inp]! | |
1538 | veor @XMM[4], @XMM[10] | |
1539 | vst1.8 {@XMM[4]}, [$out]! | |
1540 | cmp $len, #4 | |
1541 | blo .Lctr_enc_done | |
1542 | vld1.8 {@XMM[11]}, [$inp]! | |
1543 | veor @XMM[6], @XMM[11] | |
1544 | vst1.8 {@XMM[6]}, [$out]! | |
1545 | beq .Lctr_enc_done | |
1546 | vld1.8 {@XMM[12]}, [$inp]! | |
1547 | veor @XMM[3], @XMM[12] | |
1548 | vst1.8 {@XMM[3]}, [$out]! | |
1549 | cmp $len, #6 | |
1550 | blo .Lctr_enc_done | |
1551 | vld1.8 {@XMM[13]}, [$inp]! | |
1552 | veor @XMM[7], @XMM[13] | |
1553 | vst1.8 {@XMM[7]}, [$out]! | |
1554 | beq .Lctr_enc_done | |
1555 | vld1.8 {@XMM[14]}, [$inp] | |
1556 | veor @XMM[2], @XMM[14] | |
1557 | vst1.8 {@XMM[2]}, [$out]! | |
1558 | ||
1559 | .Lctr_enc_done: | |
1560 | vmov.i32 q0, #0 | |
1561 | vmov.i32 q1, #0 | |
e0202d94 | 1562 | #ifndef BSAES_ASM_EXTENDED_KEY |
9575d1a9 AP |
1563 | .Lctr_enc_bzero: @ wipe key schedule [if any] |
1564 | vstmia $keysched!, {q0-q1} | |
e0202d94 | 1565 | cmp $keysched, $fp |
9575d1a9 | 1566 | bne .Lctr_enc_bzero |
e0202d94 AP |
1567 | #else |
1568 | vstmia $keysched, {q0-q1} | |
1569 | #endif | |
9575d1a9 | 1570 | |
e0202d94 AP |
1571 | mov sp, $fp |
1572 | add sp, #0x10 @ add sp,$fp,#0x10 is no good for thumb | |
1573 | VFP_ABI_POP | |
9575d1a9 AP |
1574 | ldmia sp!, {r4-r10, pc} @ return |
1575 | ||
1576 | .align 4 | |
1577 | .Lctr_enc_short: | |
1578 | ldr ip, [sp] @ ctr pointer is passed on stack | |
1579 | stmdb sp!, {r4-r8, lr} | |
1580 | ||
1581 | mov r4, $inp @ copy arguments | |
1582 | mov r5, $out | |
1583 | mov r6, $len | |
1584 | mov r7, $key | |
1585 | ldr r8, [ip, #12] @ load counter LSW | |
1586 | vld1.8 {@XMM[1]}, [ip] @ load whole counter value | |
1587 | #ifdef __ARMEL__ | |
1588 | rev r8, r8 | |
1589 | #endif | |
1590 | sub sp, sp, #0x10 | |
313e6ec1 | 1591 | vst1.8 {@XMM[1]}, [sp] @ copy counter value |
9575d1a9 AP |
1592 | sub sp, sp, #0x10 |
1593 | ||
1594 | .Lctr_enc_short_loop: | |
1595 | add r0, sp, #0x10 @ input counter value | |
1596 | mov r1, sp @ output on the stack | |
1597 | mov r2, r7 @ key | |
1598 | ||
1599 | bl AES_encrypt | |
1600 | ||
1601 | vld1.8 {@XMM[0]}, [r4]! @ load input | |
313e6ec1 | 1602 | vld1.8 {@XMM[1]}, [sp] @ load encrypted counter |
9575d1a9 AP |
1603 | add r8, r8, #1 |
1604 | #ifdef __ARMEL__ | |
1605 | rev r0, r8 | |
1606 | str r0, [sp, #0x1c] @ next counter value | |
1607 | #else | |
1608 | str r8, [sp, #0x1c] @ next counter value | |
1609 | #endif | |
1610 | veor @XMM[0],@XMM[0],@XMM[1] | |
1611 | vst1.8 {@XMM[0]}, [r5]! @ store output | |
1612 | subs r6, r6, #1 | |
1613 | bne .Lctr_enc_short_loop | |
1614 | ||
e0202d94 AP |
1615 | vmov.i32 q0, #0 |
1616 | vmov.i32 q1, #0 | |
1617 | vstmia sp!, {q0-q1} | |
1618 | ||
9575d1a9 AP |
1619 | ldmia sp!, {r4-r8, pc} |
1620 | .size bsaes_ctr32_encrypt_blocks,.-bsaes_ctr32_encrypt_blocks | |
1621 | ___ | |
1622 | } | |
a2ea9f3e AB |
1623 | { |
1624 | ###################################################################### | |
1625 | # void bsaes_xts_[en|de]crypt(const char *inp,char *out,size_t len, | |
1626 | # const AES_KEY *key1, const AES_KEY *key2, | |
1627 | # const unsigned char iv[16]); | |
1628 | # | |
29f41e8a AP |
1629 | my ($inp,$out,$len,$key,$rounds,$magic,$fp)=(map("r$_",(7..10,1..3))); |
1630 | my $const="r6"; # returned by _bsaes_key_convert | |
a2ea9f3e AB |
1631 | my $twmask=@XMM[5]; |
1632 | my @T=@XMM[6..7]; | |
1633 | ||
1634 | $code.=<<___; | |
1635 | .globl bsaes_xts_encrypt | |
1636 | .type bsaes_xts_encrypt,%function | |
1637 | .align 4 | |
1638 | bsaes_xts_encrypt: | |
e0202d94 | 1639 | mov ip, sp |
29f41e8a | 1640 | stmdb sp!, {r4-r10, lr} @ 0x20 |
e0202d94 | 1641 | VFP_ABI_PUSH |
29f41e8a | 1642 | mov r6, sp @ future $fp |
a2ea9f3e AB |
1643 | |
1644 | mov $inp, r0 | |
1645 | mov $out, r1 | |
1646 | mov $len, r2 | |
1647 | mov $key, r3 | |
1648 | ||
e0202d94 AP |
1649 | sub r0, sp, #0x10 @ 0x10 |
1650 | bic r0, #0xf @ align at 16 bytes | |
1651 | mov sp, r0 | |
1652 | ||
1653 | #ifdef XTS_CHAIN_TWEAK | |
1654 | ldr r0, [ip] @ pointer to input tweak | |
1655 | #else | |
a2ea9f3e | 1656 | @ generate initial tweak |
e0202d94 | 1657 | ldr r0, [ip, #4] @ iv[] |
a2ea9f3e | 1658 | mov r1, sp |
e0202d94 | 1659 | ldr r2, [ip, #0] @ key2 |
a2ea9f3e | 1660 | bl AES_encrypt |
e0202d94 AP |
1661 | mov r0,sp @ pointer to initial tweak |
1662 | #endif | |
a2ea9f3e | 1663 | |
a2ea9f3e | 1664 | ldr $rounds, [$key, #240] @ get # of rounds |
29f41e8a | 1665 | mov $fp, r6 |
e0202d94 AP |
1666 | #ifndef BSAES_ASM_EXTENDED_KEY |
1667 | @ allocate the key schedule on the stack | |
1668 | sub r12, sp, $rounds, lsl#7 @ 128 bytes per inner round key | |
1669 | @ add r12, #`128-32` @ size of bit-sliced key schedule | |
1670 | sub r12, #`32+16` @ place for tweak[9] | |
a2ea9f3e AB |
1671 | |
1672 | @ populate the key schedule | |
1673 | mov r4, $key @ pass key | |
1674 | mov r5, $rounds @ pass # of rounds | |
e0202d94 AP |
1675 | mov sp, r12 |
1676 | add r12, #0x90 @ pass key schedule | |
a2ea9f3e AB |
1677 | bl _bsaes_key_convert |
1678 | veor @XMM[7], @XMM[7], @XMM[15] @ fix up last round key | |
1679 | vstmia r12, {@XMM[7]} @ save last round key | |
e0202d94 AP |
1680 | #else |
1681 | ldr r12, [$key, #244] | |
1682 | eors r12, #1 | |
1683 | beq 0f | |
1684 | ||
1685 | str r12, [$key, #244] | |
1686 | mov r4, $key @ pass key | |
1687 | mov r5, $rounds @ pass # of rounds | |
1688 | add r12, $key, #248 @ pass key schedule | |
1689 | bl _bsaes_key_convert | |
1690 | veor @XMM[7], @XMM[7], @XMM[15] @ fix up last round key | |
1691 | vstmia r12, {@XMM[7]} | |
1692 | ||
1693 | .align 2 | |
1694 | 0: sub sp, #0x90 @ place for tweak[9] | |
1695 | #endif | |
a2ea9f3e | 1696 | |
29f41e8a | 1697 | vld1.8 {@XMM[8]}, [r0] @ initial tweak |
e0202d94 | 1698 | adr $magic, .Lxts_magic |
a2ea9f3e AB |
1699 | |
1700 | subs $len, #0x80 | |
1701 | blo .Lxts_enc_short | |
1702 | b .Lxts_enc_loop | |
1703 | ||
1704 | .align 4 | |
1705 | .Lxts_enc_loop: | |
29f41e8a | 1706 | vldmia $magic, {$twmask} @ load XTS magic |
a2ea9f3e AB |
1707 | vshr.s64 @T[0], @XMM[8], #63 |
1708 | mov r0, sp | |
1709 | vand @T[0], @T[0], $twmask | |
1710 | ___ | |
1711 | for($i=9;$i<16;$i++) { | |
1712 | $code.=<<___; | |
1713 | vadd.u64 @XMM[$i], @XMM[$i-1], @XMM[$i-1] | |
29f41e8a | 1714 | vst1.64 {@XMM[$i-1]}, [r0,:128]! |
a2ea9f3e AB |
1715 | vswp `&Dhi("@T[0]")`,`&Dlo("@T[0]")` |
1716 | vshr.s64 @T[1], @XMM[$i], #63 | |
1717 | veor @XMM[$i], @XMM[$i], @T[0] | |
1718 | vand @T[1], @T[1], $twmask | |
1719 | ___ | |
1720 | @T=reverse(@T); | |
1721 | ||
1722 | $code.=<<___ if ($i>=10); | |
1723 | vld1.8 {@XMM[$i-10]}, [$inp]! | |
1724 | ___ | |
1725 | $code.=<<___ if ($i>=11); | |
1726 | veor @XMM[$i-11], @XMM[$i-11], @XMM[$i-3] | |
1727 | ___ | |
1728 | } | |
1729 | $code.=<<___; | |
1730 | vadd.u64 @XMM[8], @XMM[15], @XMM[15] | |
29f41e8a | 1731 | vst1.64 {@XMM[15]}, [r0,:128]! |
a2ea9f3e AB |
1732 | vswp `&Dhi("@T[0]")`,`&Dlo("@T[0]")` |
1733 | veor @XMM[8], @XMM[8], @T[0] | |
29f41e8a | 1734 | vst1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
1735 | |
1736 | vld1.8 {@XMM[6]-@XMM[7]}, [$inp]! | |
1737 | veor @XMM[5], @XMM[5], @XMM[13] | |
e0202d94 | 1738 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1739 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1740 | #else |
1741 | add r4, $key, #248 @ pass key schedule | |
1742 | #endif | |
a2ea9f3e AB |
1743 | veor @XMM[6], @XMM[6], @XMM[14] |
1744 | mov r5, $rounds @ pass rounds | |
1745 | veor @XMM[7], @XMM[7], @XMM[15] | |
1746 | mov r0, sp | |
1747 | ||
1748 | bl _bsaes_encrypt8 | |
1749 | ||
29f41e8a AP |
1750 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
1751 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 1752 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 1753 | vld1.64 {@XMM[12]-@XMM[13]}, [r0,:128]! |
a2ea9f3e AB |
1754 | veor @XMM[1], @XMM[1], @XMM[ 9] |
1755 | veor @XMM[8], @XMM[4], @XMM[10] | |
1756 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1757 | veor @XMM[9], @XMM[6], @XMM[11] | |
29f41e8a | 1758 | vld1.64 {@XMM[14]-@XMM[15]}, [r0,:128]! |
a2ea9f3e AB |
1759 | veor @XMM[10], @XMM[3], @XMM[12] |
1760 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
1761 | veor @XMM[11], @XMM[7], @XMM[13] | |
1762 | veor @XMM[12], @XMM[2], @XMM[14] | |
1763 | vst1.8 {@XMM[10]-@XMM[11]}, [$out]! | |
1764 | veor @XMM[13], @XMM[5], @XMM[15] | |
1765 | vst1.8 {@XMM[12]-@XMM[13]}, [$out]! | |
1766 | ||
29f41e8a | 1767 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
1768 | |
1769 | subs $len, #0x80 | |
1770 | bpl .Lxts_enc_loop | |
1771 | ||
1772 | .Lxts_enc_short: | |
1773 | adds $len, #0x70 | |
1774 | bmi .Lxts_enc_done | |
1775 | ||
29f41e8a | 1776 | vldmia $magic, {$twmask} @ load XTS magic |
a2ea9f3e AB |
1777 | vshr.s64 @T[0], @XMM[8], #63 |
1778 | mov r0, sp | |
1779 | vand @T[0], @T[0], $twmask | |
1780 | ___ | |
1781 | for($i=9;$i<16;$i++) { | |
1782 | $code.=<<___; | |
1783 | vadd.u64 @XMM[$i], @XMM[$i-1], @XMM[$i-1] | |
29f41e8a | 1784 | vst1.64 {@XMM[$i-1]}, [r0,:128]! |
a2ea9f3e AB |
1785 | vswp `&Dhi("@T[0]")`,`&Dlo("@T[0]")` |
1786 | vshr.s64 @T[1], @XMM[$i], #63 | |
1787 | veor @XMM[$i], @XMM[$i], @T[0] | |
1788 | vand @T[1], @T[1], $twmask | |
1789 | ___ | |
1790 | @T=reverse(@T); | |
1791 | ||
1792 | $code.=<<___ if ($i>=10); | |
1793 | vld1.8 {@XMM[$i-10]}, [$inp]! | |
1794 | subs $len, #0x10 | |
1795 | bmi .Lxts_enc_`$i-9` | |
1796 | ___ | |
1797 | $code.=<<___ if ($i>=11); | |
1798 | veor @XMM[$i-11], @XMM[$i-11], @XMM[$i-3] | |
1799 | ___ | |
1800 | } | |
1801 | $code.=<<___; | |
1802 | sub $len, #0x10 | |
29f41e8a | 1803 | vst1.64 {@XMM[15]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
1804 | |
1805 | vld1.8 {@XMM[6]}, [$inp]! | |
1806 | veor @XMM[5], @XMM[5], @XMM[13] | |
e0202d94 | 1807 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1808 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1809 | #else |
1810 | add r4, $key, #248 @ pass key schedule | |
1811 | #endif | |
a2ea9f3e AB |
1812 | veor @XMM[6], @XMM[6], @XMM[14] |
1813 | mov r5, $rounds @ pass rounds | |
1814 | mov r0, sp | |
1815 | ||
1816 | bl _bsaes_encrypt8 | |
1817 | ||
29f41e8a AP |
1818 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
1819 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 1820 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 1821 | vld1.64 {@XMM[12]-@XMM[13]}, [r0,:128]! |
a2ea9f3e AB |
1822 | veor @XMM[1], @XMM[1], @XMM[ 9] |
1823 | veor @XMM[8], @XMM[4], @XMM[10] | |
1824 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1825 | veor @XMM[9], @XMM[6], @XMM[11] | |
29f41e8a | 1826 | vld1.64 {@XMM[14]}, [r0,:128]! |
a2ea9f3e AB |
1827 | veor @XMM[10], @XMM[3], @XMM[12] |
1828 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
1829 | veor @XMM[11], @XMM[7], @XMM[13] | |
1830 | veor @XMM[12], @XMM[2], @XMM[14] | |
1831 | vst1.8 {@XMM[10]-@XMM[11]}, [$out]! | |
1832 | vst1.8 {@XMM[12]}, [$out]! | |
1833 | ||
29f41e8a | 1834 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 1835 | b .Lxts_enc_done |
29f41e8a | 1836 | .align 4 |
a2ea9f3e | 1837 | .Lxts_enc_6: |
a2ea9f3e | 1838 | veor @XMM[4], @XMM[4], @XMM[12] |
e0202d94 | 1839 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1840 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1841 | #else |
1842 | add r4, $key, #248 @ pass key schedule | |
1843 | #endif | |
a2ea9f3e AB |
1844 | veor @XMM[5], @XMM[5], @XMM[13] |
1845 | mov r5, $rounds @ pass rounds | |
1846 | mov r0, sp | |
1847 | ||
1848 | bl _bsaes_encrypt8 | |
1849 | ||
29f41e8a AP |
1850 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
1851 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 1852 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 1853 | vld1.64 {@XMM[12]-@XMM[13]}, [r0,:128]! |
a2ea9f3e AB |
1854 | veor @XMM[1], @XMM[1], @XMM[ 9] |
1855 | veor @XMM[8], @XMM[4], @XMM[10] | |
1856 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1857 | veor @XMM[9], @XMM[6], @XMM[11] | |
1858 | veor @XMM[10], @XMM[3], @XMM[12] | |
1859 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
1860 | veor @XMM[11], @XMM[7], @XMM[13] | |
1861 | vst1.8 {@XMM[10]-@XMM[11]}, [$out]! | |
1862 | ||
29f41e8a | 1863 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 1864 | b .Lxts_enc_done |
e0202d94 AP |
1865 | |
1866 | @ put this in range for both ARM and Thumb mode adr instructions | |
1867 | .align 5 | |
1868 | .Lxts_magic: | |
1869 | .quad 1, 0x87 | |
1870 | ||
1871 | .align 5 | |
a2ea9f3e | 1872 | .Lxts_enc_5: |
a2ea9f3e | 1873 | veor @XMM[3], @XMM[3], @XMM[11] |
e0202d94 | 1874 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1875 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1876 | #else |
1877 | add r4, $key, #248 @ pass key schedule | |
1878 | #endif | |
a2ea9f3e AB |
1879 | veor @XMM[4], @XMM[4], @XMM[12] |
1880 | mov r5, $rounds @ pass rounds | |
1881 | mov r0, sp | |
1882 | ||
1883 | bl _bsaes_encrypt8 | |
1884 | ||
29f41e8a AP |
1885 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
1886 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 1887 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 1888 | vld1.64 {@XMM[12]}, [r0,:128]! |
a2ea9f3e AB |
1889 | veor @XMM[1], @XMM[1], @XMM[ 9] |
1890 | veor @XMM[8], @XMM[4], @XMM[10] | |
1891 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1892 | veor @XMM[9], @XMM[6], @XMM[11] | |
1893 | veor @XMM[10], @XMM[3], @XMM[12] | |
1894 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
1895 | vst1.8 {@XMM[10]}, [$out]! | |
1896 | ||
29f41e8a | 1897 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 1898 | b .Lxts_enc_done |
29f41e8a | 1899 | .align 4 |
a2ea9f3e | 1900 | .Lxts_enc_4: |
a2ea9f3e | 1901 | veor @XMM[2], @XMM[2], @XMM[10] |
e0202d94 | 1902 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1903 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1904 | #else |
1905 | add r4, $key, #248 @ pass key schedule | |
1906 | #endif | |
a2ea9f3e AB |
1907 | veor @XMM[3], @XMM[3], @XMM[11] |
1908 | mov r5, $rounds @ pass rounds | |
1909 | mov r0, sp | |
1910 | ||
1911 | bl _bsaes_encrypt8 | |
1912 | ||
29f41e8a AP |
1913 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
1914 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e AB |
1915 | veor @XMM[0], @XMM[0], @XMM[ 8] |
1916 | veor @XMM[1], @XMM[1], @XMM[ 9] | |
1917 | veor @XMM[8], @XMM[4], @XMM[10] | |
1918 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1919 | veor @XMM[9], @XMM[6], @XMM[11] | |
1920 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
1921 | ||
29f41e8a | 1922 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 1923 | b .Lxts_enc_done |
29f41e8a | 1924 | .align 4 |
a2ea9f3e | 1925 | .Lxts_enc_3: |
a2ea9f3e | 1926 | veor @XMM[1], @XMM[1], @XMM[9] |
e0202d94 | 1927 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1928 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1929 | #else |
1930 | add r4, $key, #248 @ pass key schedule | |
1931 | #endif | |
a2ea9f3e AB |
1932 | veor @XMM[2], @XMM[2], @XMM[10] |
1933 | mov r5, $rounds @ pass rounds | |
1934 | mov r0, sp | |
1935 | ||
1936 | bl _bsaes_encrypt8 | |
1937 | ||
29f41e8a AP |
1938 | vld1.64 {@XMM[8]-@XMM[9]}, [r0,:128]! |
1939 | vld1.64 {@XMM[10]}, [r0,:128]! | |
a2ea9f3e AB |
1940 | veor @XMM[0], @XMM[0], @XMM[ 8] |
1941 | veor @XMM[1], @XMM[1], @XMM[ 9] | |
1942 | veor @XMM[8], @XMM[4], @XMM[10] | |
1943 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1944 | vst1.8 {@XMM[8]}, [$out]! | |
1945 | ||
29f41e8a | 1946 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 1947 | b .Lxts_enc_done |
29f41e8a | 1948 | .align 4 |
a2ea9f3e | 1949 | .Lxts_enc_2: |
a2ea9f3e | 1950 | veor @XMM[0], @XMM[0], @XMM[8] |
e0202d94 | 1951 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 1952 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
1953 | #else |
1954 | add r4, $key, #248 @ pass key schedule | |
1955 | #endif | |
a2ea9f3e AB |
1956 | veor @XMM[1], @XMM[1], @XMM[9] |
1957 | mov r5, $rounds @ pass rounds | |
1958 | mov r0, sp | |
1959 | ||
1960 | bl _bsaes_encrypt8 | |
1961 | ||
29f41e8a | 1962 | vld1.64 {@XMM[8]-@XMM[9]}, [r0,:128]! |
a2ea9f3e AB |
1963 | veor @XMM[0], @XMM[0], @XMM[ 8] |
1964 | veor @XMM[1], @XMM[1], @XMM[ 9] | |
1965 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
1966 | ||
29f41e8a | 1967 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 1968 | b .Lxts_enc_done |
29f41e8a | 1969 | .align 4 |
a2ea9f3e AB |
1970 | .Lxts_enc_1: |
1971 | mov r0, sp | |
4973a60c | 1972 | veor @XMM[0], @XMM[0], @XMM[8] |
a2ea9f3e AB |
1973 | mov r1, sp |
1974 | vst1.8 {@XMM[0]}, [sp,:128] | |
1975 | mov r2, $key | |
1976 | mov r4, $fp @ preserve fp | |
1977 | ||
1978 | bl AES_encrypt | |
1979 | ||
1980 | vld1.8 {@XMM[0]}, [sp,:128] | |
1981 | veor @XMM[0], @XMM[0], @XMM[8] | |
1982 | vst1.8 {@XMM[0]}, [$out]! | |
1983 | mov $fp, r4 | |
1984 | ||
1985 | vmov @XMM[8], @XMM[9] @ next round tweak | |
1986 | ||
1987 | .Lxts_enc_done: | |
e0202d94 | 1988 | #ifndef XTS_CHAIN_TWEAK |
a2ea9f3e AB |
1989 | adds $len, #0x10 |
1990 | beq .Lxts_enc_ret | |
1991 | sub r6, $out, #0x10 | |
1992 | ||
1993 | .Lxts_enc_steal: | |
1994 | ldrb r0, [$inp], #1 | |
1995 | ldrb r1, [$out, #-0x10] | |
1996 | strb r0, [$out, #-0x10] | |
1997 | strb r1, [$out], #1 | |
1998 | ||
1999 | subs $len, #1 | |
2000 | bhi .Lxts_enc_steal | |
2001 | ||
2002 | vld1.8 {@XMM[0]}, [r6] | |
2003 | mov r0, sp | |
2004 | veor @XMM[0], @XMM[0], @XMM[8] | |
2005 | mov r1, sp | |
2006 | vst1.8 {@XMM[0]}, [sp,:128] | |
2007 | mov r2, $key | |
2008 | mov r4, $fp @ preserve fp | |
2009 | ||
2010 | bl AES_encrypt | |
2011 | ||
2012 | vld1.8 {@XMM[0]}, [sp,:128] | |
2013 | veor @XMM[0], @XMM[0], @XMM[8] | |
2014 | vst1.8 {@XMM[0]}, [r6] | |
2015 | mov $fp, r4 | |
e0202d94 | 2016 | #endif |
a2ea9f3e AB |
2017 | |
2018 | .Lxts_enc_ret: | |
29f41e8a AP |
2019 | bic r0, $fp, #0xf |
2020 | vmov.i32 q0, #0 | |
2021 | vmov.i32 q1, #0 | |
e0202d94 AP |
2022 | #ifdef XTS_CHAIN_TWEAK |
2023 | ldr r1, [$fp, #0x20+VFP_ABI_FRAME] @ chain tweak | |
2024 | #endif | |
a2ea9f3e | 2025 | .Lxts_enc_bzero: @ wipe key schedule [if any] |
29f41e8a | 2026 | vstmia sp!, {q0-q1} |
e0202d94 | 2027 | cmp sp, r0 |
a2ea9f3e AB |
2028 | bne .Lxts_enc_bzero |
2029 | ||
29f41e8a | 2030 | mov sp, $fp |
e0202d94 AP |
2031 | #ifdef XTS_CHAIN_TWEAK |
2032 | vst1.8 {@XMM[8]}, [r1] | |
2033 | #endif | |
2034 | VFP_ABI_POP | |
29f41e8a | 2035 | ldmia sp!, {r4-r10, pc} @ return |
a2ea9f3e AB |
2036 | |
2037 | .size bsaes_xts_encrypt,.-bsaes_xts_encrypt | |
2038 | ||
2039 | .globl bsaes_xts_decrypt | |
2040 | .type bsaes_xts_decrypt,%function | |
2041 | .align 4 | |
2042 | bsaes_xts_decrypt: | |
e0202d94 | 2043 | mov ip, sp |
29f41e8a | 2044 | stmdb sp!, {r4-r10, lr} @ 0x20 |
e0202d94 | 2045 | VFP_ABI_PUSH |
29f41e8a | 2046 | mov r6, sp @ future $fp |
a2ea9f3e AB |
2047 | |
2048 | mov $inp, r0 | |
2049 | mov $out, r1 | |
2050 | mov $len, r2 | |
2051 | mov $key, r3 | |
2052 | ||
e0202d94 AP |
2053 | sub r0, sp, #0x10 @ 0x10 |
2054 | bic r0, #0xf @ align at 16 bytes | |
2055 | mov sp, r0 | |
2056 | ||
2057 | #ifdef XTS_CHAIN_TWEAK | |
2058 | ldr r0, [ip] @ pointer to input tweak | |
2059 | #else | |
a2ea9f3e | 2060 | @ generate initial tweak |
e0202d94 | 2061 | ldr r0, [ip, #4] @ iv[] |
a2ea9f3e | 2062 | mov r1, sp |
e0202d94 | 2063 | ldr r2, [ip, #0] @ key2 |
a2ea9f3e | 2064 | bl AES_encrypt |
e0202d94 AP |
2065 | mov r0, sp @ pointer to initial tweak |
2066 | #endif | |
a2ea9f3e | 2067 | |
a2ea9f3e | 2068 | ldr $rounds, [$key, #240] @ get # of rounds |
29f41e8a | 2069 | mov $fp, r6 |
e0202d94 AP |
2070 | #ifndef BSAES_ASM_EXTENDED_KEY |
2071 | @ allocate the key schedule on the stack | |
2072 | sub r12, sp, $rounds, lsl#7 @ 128 bytes per inner round key | |
2073 | @ add r12, #`128-32` @ size of bit-sliced key schedule | |
2074 | sub r12, #`32+16` @ place for tweak[9] | |
a2ea9f3e AB |
2075 | |
2076 | @ populate the key schedule | |
2077 | mov r4, $key @ pass key | |
2078 | mov r5, $rounds @ pass # of rounds | |
e0202d94 AP |
2079 | mov sp, r12 |
2080 | add r12, #0x90 @ pass key schedule | |
a2ea9f3e | 2081 | bl _bsaes_key_convert |
29f41e8a AP |
2082 | add r4, sp, #0x90 |
2083 | vldmia r4, {@XMM[6]} | |
a2ea9f3e AB |
2084 | vstmia r12, {@XMM[15]} @ save last round key |
2085 | veor @XMM[7], @XMM[7], @XMM[6] @ fix up round 0 key | |
29f41e8a | 2086 | vstmia r4, {@XMM[7]} |
e0202d94 AP |
2087 | #else |
2088 | ldr r12, [$key, #244] | |
2089 | eors r12, #1 | |
2090 | beq 0f | |
2091 | ||
2092 | str r12, [$key, #244] | |
2093 | mov r4, $key @ pass key | |
2094 | mov r5, $rounds @ pass # of rounds | |
2095 | add r12, $key, #248 @ pass key schedule | |
2096 | bl _bsaes_key_convert | |
2097 | add r4, $key, #248 | |
2098 | vldmia r4, {@XMM[6]} | |
2099 | vstmia r12, {@XMM[15]} @ save last round key | |
2100 | veor @XMM[7], @XMM[7], @XMM[6] @ fix up round 0 key | |
2101 | vstmia r4, {@XMM[7]} | |
a2ea9f3e | 2102 | |
e0202d94 AP |
2103 | .align 2 |
2104 | 0: sub sp, #0x90 @ place for tweak[9] | |
2105 | #endif | |
29f41e8a | 2106 | vld1.8 {@XMM[8]}, [r0] @ initial tweak |
e0202d94 | 2107 | adr $magic, .Lxts_magic |
a2ea9f3e | 2108 | |
e620e5ae | 2109 | #ifndef XTS_CHAIN_TWEAK |
a2ea9f3e | 2110 | tst $len, #0xf @ if not multiple of 16 |
e0202d94 | 2111 | it ne @ Thumb2 thing, sanity check in ARM |
29f41e8a | 2112 | subne $len, #0x10 @ subtract another 16 bytes |
e620e5ae | 2113 | #endif |
a2ea9f3e AB |
2114 | subs $len, #0x80 |
2115 | ||
2116 | blo .Lxts_dec_short | |
2117 | b .Lxts_dec_loop | |
2118 | ||
2119 | .align 4 | |
2120 | .Lxts_dec_loop: | |
29f41e8a | 2121 | vldmia $magic, {$twmask} @ load XTS magic |
a2ea9f3e AB |
2122 | vshr.s64 @T[0], @XMM[8], #63 |
2123 | mov r0, sp | |
2124 | vand @T[0], @T[0], $twmask | |
2125 | ___ | |
2126 | for($i=9;$i<16;$i++) { | |
2127 | $code.=<<___; | |
2128 | vadd.u64 @XMM[$i], @XMM[$i-1], @XMM[$i-1] | |
29f41e8a | 2129 | vst1.64 {@XMM[$i-1]}, [r0,:128]! |
a2ea9f3e AB |
2130 | vswp `&Dhi("@T[0]")`,`&Dlo("@T[0]")` |
2131 | vshr.s64 @T[1], @XMM[$i], #63 | |
2132 | veor @XMM[$i], @XMM[$i], @T[0] | |
2133 | vand @T[1], @T[1], $twmask | |
2134 | ___ | |
2135 | @T=reverse(@T); | |
2136 | ||
2137 | $code.=<<___ if ($i>=10); | |
2138 | vld1.8 {@XMM[$i-10]}, [$inp]! | |
2139 | ___ | |
2140 | $code.=<<___ if ($i>=11); | |
2141 | veor @XMM[$i-11], @XMM[$i-11], @XMM[$i-3] | |
2142 | ___ | |
2143 | } | |
2144 | $code.=<<___; | |
2145 | vadd.u64 @XMM[8], @XMM[15], @XMM[15] | |
29f41e8a | 2146 | vst1.64 {@XMM[15]}, [r0,:128]! |
a2ea9f3e AB |
2147 | vswp `&Dhi("@T[0]")`,`&Dlo("@T[0]")` |
2148 | veor @XMM[8], @XMM[8], @T[0] | |
29f41e8a | 2149 | vst1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
2150 | |
2151 | vld1.8 {@XMM[6]-@XMM[7]}, [$inp]! | |
2152 | veor @XMM[5], @XMM[5], @XMM[13] | |
e0202d94 | 2153 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2154 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2155 | #else |
2156 | add r4, $key, #248 @ pass key schedule | |
2157 | #endif | |
a2ea9f3e AB |
2158 | veor @XMM[6], @XMM[6], @XMM[14] |
2159 | mov r5, $rounds @ pass rounds | |
2160 | veor @XMM[7], @XMM[7], @XMM[15] | |
2161 | mov r0, sp | |
2162 | ||
2163 | bl _bsaes_decrypt8 | |
2164 | ||
29f41e8a AP |
2165 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
2166 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 2167 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 2168 | vld1.64 {@XMM[12]-@XMM[13]}, [r0,:128]! |
a2ea9f3e AB |
2169 | veor @XMM[1], @XMM[1], @XMM[ 9] |
2170 | veor @XMM[8], @XMM[6], @XMM[10] | |
2171 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2172 | veor @XMM[9], @XMM[4], @XMM[11] | |
29f41e8a | 2173 | vld1.64 {@XMM[14]-@XMM[15]}, [r0,:128]! |
a2ea9f3e AB |
2174 | veor @XMM[10], @XMM[2], @XMM[12] |
2175 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
2176 | veor @XMM[11], @XMM[7], @XMM[13] | |
2177 | veor @XMM[12], @XMM[3], @XMM[14] | |
2178 | vst1.8 {@XMM[10]-@XMM[11]}, [$out]! | |
2179 | veor @XMM[13], @XMM[5], @XMM[15] | |
2180 | vst1.8 {@XMM[12]-@XMM[13]}, [$out]! | |
2181 | ||
29f41e8a | 2182 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
2183 | |
2184 | subs $len, #0x80 | |
2185 | bpl .Lxts_dec_loop | |
2186 | ||
2187 | .Lxts_dec_short: | |
2188 | adds $len, #0x70 | |
2189 | bmi .Lxts_dec_done | |
2190 | ||
29f41e8a | 2191 | vldmia $magic, {$twmask} @ load XTS magic |
a2ea9f3e AB |
2192 | vshr.s64 @T[0], @XMM[8], #63 |
2193 | mov r0, sp | |
2194 | vand @T[0], @T[0], $twmask | |
2195 | ___ | |
2196 | for($i=9;$i<16;$i++) { | |
2197 | $code.=<<___; | |
2198 | vadd.u64 @XMM[$i], @XMM[$i-1], @XMM[$i-1] | |
29f41e8a | 2199 | vst1.64 {@XMM[$i-1]}, [r0,:128]! |
a2ea9f3e AB |
2200 | vswp `&Dhi("@T[0]")`,`&Dlo("@T[0]")` |
2201 | vshr.s64 @T[1], @XMM[$i], #63 | |
2202 | veor @XMM[$i], @XMM[$i], @T[0] | |
2203 | vand @T[1], @T[1], $twmask | |
2204 | ___ | |
2205 | @T=reverse(@T); | |
2206 | ||
2207 | $code.=<<___ if ($i>=10); | |
2208 | vld1.8 {@XMM[$i-10]}, [$inp]! | |
2209 | subs $len, #0x10 | |
2210 | bmi .Lxts_dec_`$i-9` | |
2211 | ___ | |
2212 | $code.=<<___ if ($i>=11); | |
2213 | veor @XMM[$i-11], @XMM[$i-11], @XMM[$i-3] | |
2214 | ___ | |
2215 | } | |
2216 | $code.=<<___; | |
2217 | sub $len, #0x10 | |
29f41e8a | 2218 | vst1.64 {@XMM[15]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
2219 | |
2220 | vld1.8 {@XMM[6]}, [$inp]! | |
2221 | veor @XMM[5], @XMM[5], @XMM[13] | |
e0202d94 | 2222 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2223 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2224 | #else |
2225 | add r4, $key, #248 @ pass key schedule | |
2226 | #endif | |
a2ea9f3e AB |
2227 | veor @XMM[6], @XMM[6], @XMM[14] |
2228 | mov r5, $rounds @ pass rounds | |
2229 | mov r0, sp | |
2230 | ||
2231 | bl _bsaes_decrypt8 | |
2232 | ||
29f41e8a AP |
2233 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
2234 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 2235 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 2236 | vld1.64 {@XMM[12]-@XMM[13]}, [r0,:128]! |
a2ea9f3e AB |
2237 | veor @XMM[1], @XMM[1], @XMM[ 9] |
2238 | veor @XMM[8], @XMM[6], @XMM[10] | |
2239 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2240 | veor @XMM[9], @XMM[4], @XMM[11] | |
29f41e8a | 2241 | vld1.64 {@XMM[14]}, [r0,:128]! |
a2ea9f3e AB |
2242 | veor @XMM[10], @XMM[2], @XMM[12] |
2243 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
2244 | veor @XMM[11], @XMM[7], @XMM[13] | |
2245 | veor @XMM[12], @XMM[3], @XMM[14] | |
2246 | vst1.8 {@XMM[10]-@XMM[11]}, [$out]! | |
2247 | vst1.8 {@XMM[12]}, [$out]! | |
2248 | ||
29f41e8a | 2249 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 2250 | b .Lxts_dec_done |
29f41e8a | 2251 | .align 4 |
a2ea9f3e | 2252 | .Lxts_dec_6: |
29f41e8a | 2253 | vst1.64 {@XMM[14]}, [r0,:128] @ next round tweak |
a2ea9f3e AB |
2254 | |
2255 | veor @XMM[4], @XMM[4], @XMM[12] | |
e0202d94 | 2256 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2257 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2258 | #else |
2259 | add r4, $key, #248 @ pass key schedule | |
2260 | #endif | |
a2ea9f3e AB |
2261 | veor @XMM[5], @XMM[5], @XMM[13] |
2262 | mov r5, $rounds @ pass rounds | |
2263 | mov r0, sp | |
2264 | ||
2265 | bl _bsaes_decrypt8 | |
2266 | ||
29f41e8a AP |
2267 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
2268 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 2269 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 2270 | vld1.64 {@XMM[12]-@XMM[13]}, [r0,:128]! |
a2ea9f3e AB |
2271 | veor @XMM[1], @XMM[1], @XMM[ 9] |
2272 | veor @XMM[8], @XMM[6], @XMM[10] | |
2273 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2274 | veor @XMM[9], @XMM[4], @XMM[11] | |
2275 | veor @XMM[10], @XMM[2], @XMM[12] | |
2276 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
2277 | veor @XMM[11], @XMM[7], @XMM[13] | |
2278 | vst1.8 {@XMM[10]-@XMM[11]}, [$out]! | |
2279 | ||
29f41e8a | 2280 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 2281 | b .Lxts_dec_done |
29f41e8a | 2282 | .align 4 |
a2ea9f3e | 2283 | .Lxts_dec_5: |
a2ea9f3e | 2284 | veor @XMM[3], @XMM[3], @XMM[11] |
e0202d94 | 2285 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2286 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2287 | #else |
2288 | add r4, $key, #248 @ pass key schedule | |
2289 | #endif | |
a2ea9f3e AB |
2290 | veor @XMM[4], @XMM[4], @XMM[12] |
2291 | mov r5, $rounds @ pass rounds | |
2292 | mov r0, sp | |
2293 | ||
2294 | bl _bsaes_decrypt8 | |
2295 | ||
29f41e8a AP |
2296 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
2297 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e | 2298 | veor @XMM[0], @XMM[0], @XMM[ 8] |
29f41e8a | 2299 | vld1.64 {@XMM[12]}, [r0,:128]! |
a2ea9f3e AB |
2300 | veor @XMM[1], @XMM[1], @XMM[ 9] |
2301 | veor @XMM[8], @XMM[6], @XMM[10] | |
2302 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2303 | veor @XMM[9], @XMM[4], @XMM[11] | |
2304 | veor @XMM[10], @XMM[2], @XMM[12] | |
2305 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
2306 | vst1.8 {@XMM[10]}, [$out]! | |
2307 | ||
29f41e8a | 2308 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 2309 | b .Lxts_dec_done |
29f41e8a | 2310 | .align 4 |
a2ea9f3e | 2311 | .Lxts_dec_4: |
a2ea9f3e | 2312 | veor @XMM[2], @XMM[2], @XMM[10] |
e0202d94 | 2313 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2314 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2315 | #else |
2316 | add r4, $key, #248 @ pass key schedule | |
2317 | #endif | |
a2ea9f3e AB |
2318 | veor @XMM[3], @XMM[3], @XMM[11] |
2319 | mov r5, $rounds @ pass rounds | |
2320 | mov r0, sp | |
2321 | ||
2322 | bl _bsaes_decrypt8 | |
2323 | ||
29f41e8a AP |
2324 | vld1.64 {@XMM[ 8]-@XMM[ 9]}, [r0,:128]! |
2325 | vld1.64 {@XMM[10]-@XMM[11]}, [r0,:128]! | |
a2ea9f3e AB |
2326 | veor @XMM[0], @XMM[0], @XMM[ 8] |
2327 | veor @XMM[1], @XMM[1], @XMM[ 9] | |
2328 | veor @XMM[8], @XMM[6], @XMM[10] | |
2329 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2330 | veor @XMM[9], @XMM[4], @XMM[11] | |
2331 | vst1.8 {@XMM[8]-@XMM[9]}, [$out]! | |
2332 | ||
29f41e8a | 2333 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 2334 | b .Lxts_dec_done |
29f41e8a | 2335 | .align 4 |
a2ea9f3e | 2336 | .Lxts_dec_3: |
a2ea9f3e | 2337 | veor @XMM[1], @XMM[1], @XMM[9] |
e0202d94 | 2338 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2339 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2340 | #else |
2341 | add r4, $key, #248 @ pass key schedule | |
2342 | #endif | |
a2ea9f3e AB |
2343 | veor @XMM[2], @XMM[2], @XMM[10] |
2344 | mov r5, $rounds @ pass rounds | |
2345 | mov r0, sp | |
2346 | ||
2347 | bl _bsaes_decrypt8 | |
2348 | ||
29f41e8a AP |
2349 | vld1.64 {@XMM[8]-@XMM[9]}, [r0,:128]! |
2350 | vld1.64 {@XMM[10]}, [r0,:128]! | |
a2ea9f3e AB |
2351 | veor @XMM[0], @XMM[0], @XMM[ 8] |
2352 | veor @XMM[1], @XMM[1], @XMM[ 9] | |
2353 | veor @XMM[8], @XMM[6], @XMM[10] | |
2354 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2355 | vst1.8 {@XMM[8]}, [$out]! | |
2356 | ||
29f41e8a | 2357 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 2358 | b .Lxts_dec_done |
29f41e8a | 2359 | .align 4 |
a2ea9f3e | 2360 | .Lxts_dec_2: |
a2ea9f3e | 2361 | veor @XMM[0], @XMM[0], @XMM[8] |
e0202d94 | 2362 | #ifndef BSAES_ASM_EXTENDED_KEY |
29f41e8a | 2363 | add r4, sp, #0x90 @ pass key schedule |
e0202d94 AP |
2364 | #else |
2365 | add r4, $key, #248 @ pass key schedule | |
2366 | #endif | |
a2ea9f3e AB |
2367 | veor @XMM[1], @XMM[1], @XMM[9] |
2368 | mov r5, $rounds @ pass rounds | |
2369 | mov r0, sp | |
2370 | ||
2371 | bl _bsaes_decrypt8 | |
2372 | ||
29f41e8a | 2373 | vld1.64 {@XMM[8]-@XMM[9]}, [r0,:128]! |
a2ea9f3e AB |
2374 | veor @XMM[0], @XMM[0], @XMM[ 8] |
2375 | veor @XMM[1], @XMM[1], @XMM[ 9] | |
2376 | vst1.8 {@XMM[0]-@XMM[1]}, [$out]! | |
2377 | ||
29f41e8a | 2378 | vld1.64 {@XMM[8]}, [r0,:128] @ next round tweak |
a2ea9f3e | 2379 | b .Lxts_dec_done |
29f41e8a | 2380 | .align 4 |
a2ea9f3e AB |
2381 | .Lxts_dec_1: |
2382 | mov r0, sp | |
4973a60c | 2383 | veor @XMM[0], @XMM[0], @XMM[8] |
a2ea9f3e AB |
2384 | mov r1, sp |
2385 | vst1.8 {@XMM[0]}, [sp,:128] | |
3d32bab8 | 2386 | mov r5, $magic @ preserve magic |
a2ea9f3e AB |
2387 | mov r2, $key |
2388 | mov r4, $fp @ preserve fp | |
2389 | ||
2390 | bl AES_decrypt | |
2391 | ||
2392 | vld1.8 {@XMM[0]}, [sp,:128] | |
2393 | veor @XMM[0], @XMM[0], @XMM[8] | |
2394 | vst1.8 {@XMM[0]}, [$out]! | |
2395 | mov $fp, r4 | |
29f41e8a | 2396 | mov $magic, r5 |
a2ea9f3e AB |
2397 | |
2398 | vmov @XMM[8], @XMM[9] @ next round tweak | |
2399 | ||
2400 | .Lxts_dec_done: | |
e0202d94 | 2401 | #ifndef XTS_CHAIN_TWEAK |
a2ea9f3e AB |
2402 | adds $len, #0x10 |
2403 | beq .Lxts_dec_ret | |
2404 | ||
2405 | @ calculate one round of extra tweak for the stolen ciphertext | |
29f41e8a | 2406 | vldmia $magic, {$twmask} |
a2ea9f3e AB |
2407 | vshr.s64 @XMM[6], @XMM[8], #63 |
2408 | vand @XMM[6], @XMM[6], $twmask | |
2409 | vadd.u64 @XMM[9], @XMM[8], @XMM[8] | |
2410 | vswp `&Dhi("@XMM[6]")`,`&Dlo("@XMM[6]")` | |
2411 | veor @XMM[9], @XMM[9], @XMM[6] | |
2412 | ||
2413 | @ perform the final decryption with the last tweak value | |
2414 | vld1.8 {@XMM[0]}, [$inp]! | |
2415 | mov r0, sp | |
2416 | veor @XMM[0], @XMM[0], @XMM[9] | |
2417 | mov r1, sp | |
2418 | vst1.8 {@XMM[0]}, [sp,:128] | |
2419 | mov r2, $key | |
2420 | mov r4, $fp @ preserve fp | |
2421 | ||
2422 | bl AES_decrypt | |
2423 | ||
2424 | vld1.8 {@XMM[0]}, [sp,:128] | |
2425 | veor @XMM[0], @XMM[0], @XMM[9] | |
2426 | vst1.8 {@XMM[0]}, [$out] | |
2427 | ||
2428 | mov r6, $out | |
2429 | .Lxts_dec_steal: | |
2430 | ldrb r1, [$out] | |
2431 | ldrb r0, [$inp], #1 | |
2432 | strb r1, [$out, #0x10] | |
2433 | strb r0, [$out], #1 | |
2434 | ||
2435 | subs $len, #1 | |
2436 | bhi .Lxts_dec_steal | |
2437 | ||
2438 | vld1.8 {@XMM[0]}, [r6] | |
2439 | mov r0, sp | |
2440 | veor @XMM[0], @XMM[8] | |
2441 | mov r1, sp | |
2442 | vst1.8 {@XMM[0]}, [sp,:128] | |
2443 | mov r2, $key | |
2444 | ||
2445 | bl AES_decrypt | |
2446 | ||
2447 | vld1.8 {@XMM[0]}, [sp,:128] | |
2448 | veor @XMM[0], @XMM[0], @XMM[8] | |
2449 | vst1.8 {@XMM[0]}, [r6] | |
2450 | mov $fp, r4 | |
e0202d94 | 2451 | #endif |
a2ea9f3e AB |
2452 | |
2453 | .Lxts_dec_ret: | |
29f41e8a AP |
2454 | bic r0, $fp, #0xf |
2455 | vmov.i32 q0, #0 | |
2456 | vmov.i32 q1, #0 | |
e0202d94 AP |
2457 | #ifdef XTS_CHAIN_TWEAK |
2458 | ldr r1, [$fp, #0x20+VFP_ABI_FRAME] @ chain tweak | |
2459 | #endif | |
a2ea9f3e | 2460 | .Lxts_dec_bzero: @ wipe key schedule [if any] |
29f41e8a | 2461 | vstmia sp!, {q0-q1} |
e0202d94 | 2462 | cmp sp, r0 |
a2ea9f3e AB |
2463 | bne .Lxts_dec_bzero |
2464 | ||
29f41e8a | 2465 | mov sp, $fp |
e0202d94 AP |
2466 | #ifdef XTS_CHAIN_TWEAK |
2467 | vst1.8 {@XMM[8]}, [r1] | |
2468 | #endif | |
2469 | VFP_ABI_POP | |
29f41e8a | 2470 | ldmia sp!, {r4-r10, pc} @ return |
a2ea9f3e AB |
2471 | |
2472 | .size bsaes_xts_decrypt,.-bsaes_xts_decrypt | |
2473 | ___ | |
2474 | } | |
75fe4223 AP |
2475 | $code.=<<___; |
2476 | #endif | |
2477 | ___ | |
c4a52a6d AP |
2478 | |
2479 | $code =~ s/\`([^\`]*)\`/eval($1)/gem; | |
2480 | ||
e0202d94 AP |
2481 | open SELF,$0; |
2482 | while(<SELF>) { | |
2483 | next if (/^#!/); | |
2484 | last if (!s/^#/@/ and !/^$/); | |
2485 | print; | |
2486 | } | |
2487 | close SELF; | |
2488 | ||
c4a52a6d AP |
2489 | print $code; |
2490 | ||
2491 | close STDOUT; |