1*0957b409SSimon J. Gerraty /*
2*0957b409SSimon J. Gerraty * Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
3*0957b409SSimon J. Gerraty *
4*0957b409SSimon J. Gerraty * Permission is hereby granted, free of charge, to any person obtaining
5*0957b409SSimon J. Gerraty * a copy of this software and associated documentation files (the
6*0957b409SSimon J. Gerraty * "Software"), to deal in the Software without restriction, including
7*0957b409SSimon J. Gerraty * without limitation the rights to use, copy, modify, merge, publish,
8*0957b409SSimon J. Gerraty * distribute, sublicense, and/or sell copies of the Software, and to
9*0957b409SSimon J. Gerraty * permit persons to whom the Software is furnished to do so, subject to
10*0957b409SSimon J. Gerraty * the following conditions:
11*0957b409SSimon J. Gerraty *
12*0957b409SSimon J. Gerraty * The above copyright notice and this permission notice shall be
13*0957b409SSimon J. Gerraty * included in all copies or substantial portions of the Software.
14*0957b409SSimon J. Gerraty *
15*0957b409SSimon J. Gerraty * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16*0957b409SSimon J. Gerraty * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17*0957b409SSimon J. Gerraty * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18*0957b409SSimon J. Gerraty * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19*0957b409SSimon J. Gerraty * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20*0957b409SSimon J. Gerraty * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21*0957b409SSimon J. Gerraty * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22*0957b409SSimon J. Gerraty * SOFTWARE.
23*0957b409SSimon J. Gerraty */
24*0957b409SSimon J. Gerraty
25*0957b409SSimon J. Gerraty #include "inner.h"
26*0957b409SSimon J. Gerraty
27*0957b409SSimon J. Gerraty /* see inner.h */
28*0957b409SSimon J. Gerraty void
br_aes_ct64_bitslice_Sbox(uint64_t * q)29*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_Sbox(uint64_t *q)
30*0957b409SSimon J. Gerraty {
31*0957b409SSimon J. Gerraty /*
32*0957b409SSimon J. Gerraty * This S-box implementation is a straightforward translation of
33*0957b409SSimon J. Gerraty * the circuit described by Boyar and Peralta in "A new
34*0957b409SSimon J. Gerraty * combinational logic minimization technique with applications
35*0957b409SSimon J. Gerraty * to cryptology" (https://eprint.iacr.org/2009/191.pdf).
36*0957b409SSimon J. Gerraty *
37*0957b409SSimon J. Gerraty * Note that variables x* (input) and s* (output) are numbered
38*0957b409SSimon J. Gerraty * in "reverse" order (x0 is the high bit, x7 is the low bit).
39*0957b409SSimon J. Gerraty */
40*0957b409SSimon J. Gerraty
41*0957b409SSimon J. Gerraty uint64_t x0, x1, x2, x3, x4, x5, x6, x7;
42*0957b409SSimon J. Gerraty uint64_t y1, y2, y3, y4, y5, y6, y7, y8, y9;
43*0957b409SSimon J. Gerraty uint64_t y10, y11, y12, y13, y14, y15, y16, y17, y18, y19;
44*0957b409SSimon J. Gerraty uint64_t y20, y21;
45*0957b409SSimon J. Gerraty uint64_t z0, z1, z2, z3, z4, z5, z6, z7, z8, z9;
46*0957b409SSimon J. Gerraty uint64_t z10, z11, z12, z13, z14, z15, z16, z17;
47*0957b409SSimon J. Gerraty uint64_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9;
48*0957b409SSimon J. Gerraty uint64_t t10, t11, t12, t13, t14, t15, t16, t17, t18, t19;
49*0957b409SSimon J. Gerraty uint64_t t20, t21, t22, t23, t24, t25, t26, t27, t28, t29;
50*0957b409SSimon J. Gerraty uint64_t t30, t31, t32, t33, t34, t35, t36, t37, t38, t39;
51*0957b409SSimon J. Gerraty uint64_t t40, t41, t42, t43, t44, t45, t46, t47, t48, t49;
52*0957b409SSimon J. Gerraty uint64_t t50, t51, t52, t53, t54, t55, t56, t57, t58, t59;
53*0957b409SSimon J. Gerraty uint64_t t60, t61, t62, t63, t64, t65, t66, t67;
54*0957b409SSimon J. Gerraty uint64_t s0, s1, s2, s3, s4, s5, s6, s7;
55*0957b409SSimon J. Gerraty
56*0957b409SSimon J. Gerraty x0 = q[7];
57*0957b409SSimon J. Gerraty x1 = q[6];
58*0957b409SSimon J. Gerraty x2 = q[5];
59*0957b409SSimon J. Gerraty x3 = q[4];
60*0957b409SSimon J. Gerraty x4 = q[3];
61*0957b409SSimon J. Gerraty x5 = q[2];
62*0957b409SSimon J. Gerraty x6 = q[1];
63*0957b409SSimon J. Gerraty x7 = q[0];
64*0957b409SSimon J. Gerraty
65*0957b409SSimon J. Gerraty /*
66*0957b409SSimon J. Gerraty * Top linear transformation.
67*0957b409SSimon J. Gerraty */
68*0957b409SSimon J. Gerraty y14 = x3 ^ x5;
69*0957b409SSimon J. Gerraty y13 = x0 ^ x6;
70*0957b409SSimon J. Gerraty y9 = x0 ^ x3;
71*0957b409SSimon J. Gerraty y8 = x0 ^ x5;
72*0957b409SSimon J. Gerraty t0 = x1 ^ x2;
73*0957b409SSimon J. Gerraty y1 = t0 ^ x7;
74*0957b409SSimon J. Gerraty y4 = y1 ^ x3;
75*0957b409SSimon J. Gerraty y12 = y13 ^ y14;
76*0957b409SSimon J. Gerraty y2 = y1 ^ x0;
77*0957b409SSimon J. Gerraty y5 = y1 ^ x6;
78*0957b409SSimon J. Gerraty y3 = y5 ^ y8;
79*0957b409SSimon J. Gerraty t1 = x4 ^ y12;
80*0957b409SSimon J. Gerraty y15 = t1 ^ x5;
81*0957b409SSimon J. Gerraty y20 = t1 ^ x1;
82*0957b409SSimon J. Gerraty y6 = y15 ^ x7;
83*0957b409SSimon J. Gerraty y10 = y15 ^ t0;
84*0957b409SSimon J. Gerraty y11 = y20 ^ y9;
85*0957b409SSimon J. Gerraty y7 = x7 ^ y11;
86*0957b409SSimon J. Gerraty y17 = y10 ^ y11;
87*0957b409SSimon J. Gerraty y19 = y10 ^ y8;
88*0957b409SSimon J. Gerraty y16 = t0 ^ y11;
89*0957b409SSimon J. Gerraty y21 = y13 ^ y16;
90*0957b409SSimon J. Gerraty y18 = x0 ^ y16;
91*0957b409SSimon J. Gerraty
92*0957b409SSimon J. Gerraty /*
93*0957b409SSimon J. Gerraty * Non-linear section.
94*0957b409SSimon J. Gerraty */
95*0957b409SSimon J. Gerraty t2 = y12 & y15;
96*0957b409SSimon J. Gerraty t3 = y3 & y6;
97*0957b409SSimon J. Gerraty t4 = t3 ^ t2;
98*0957b409SSimon J. Gerraty t5 = y4 & x7;
99*0957b409SSimon J. Gerraty t6 = t5 ^ t2;
100*0957b409SSimon J. Gerraty t7 = y13 & y16;
101*0957b409SSimon J. Gerraty t8 = y5 & y1;
102*0957b409SSimon J. Gerraty t9 = t8 ^ t7;
103*0957b409SSimon J. Gerraty t10 = y2 & y7;
104*0957b409SSimon J. Gerraty t11 = t10 ^ t7;
105*0957b409SSimon J. Gerraty t12 = y9 & y11;
106*0957b409SSimon J. Gerraty t13 = y14 & y17;
107*0957b409SSimon J. Gerraty t14 = t13 ^ t12;
108*0957b409SSimon J. Gerraty t15 = y8 & y10;
109*0957b409SSimon J. Gerraty t16 = t15 ^ t12;
110*0957b409SSimon J. Gerraty t17 = t4 ^ t14;
111*0957b409SSimon J. Gerraty t18 = t6 ^ t16;
112*0957b409SSimon J. Gerraty t19 = t9 ^ t14;
113*0957b409SSimon J. Gerraty t20 = t11 ^ t16;
114*0957b409SSimon J. Gerraty t21 = t17 ^ y20;
115*0957b409SSimon J. Gerraty t22 = t18 ^ y19;
116*0957b409SSimon J. Gerraty t23 = t19 ^ y21;
117*0957b409SSimon J. Gerraty t24 = t20 ^ y18;
118*0957b409SSimon J. Gerraty
119*0957b409SSimon J. Gerraty t25 = t21 ^ t22;
120*0957b409SSimon J. Gerraty t26 = t21 & t23;
121*0957b409SSimon J. Gerraty t27 = t24 ^ t26;
122*0957b409SSimon J. Gerraty t28 = t25 & t27;
123*0957b409SSimon J. Gerraty t29 = t28 ^ t22;
124*0957b409SSimon J. Gerraty t30 = t23 ^ t24;
125*0957b409SSimon J. Gerraty t31 = t22 ^ t26;
126*0957b409SSimon J. Gerraty t32 = t31 & t30;
127*0957b409SSimon J. Gerraty t33 = t32 ^ t24;
128*0957b409SSimon J. Gerraty t34 = t23 ^ t33;
129*0957b409SSimon J. Gerraty t35 = t27 ^ t33;
130*0957b409SSimon J. Gerraty t36 = t24 & t35;
131*0957b409SSimon J. Gerraty t37 = t36 ^ t34;
132*0957b409SSimon J. Gerraty t38 = t27 ^ t36;
133*0957b409SSimon J. Gerraty t39 = t29 & t38;
134*0957b409SSimon J. Gerraty t40 = t25 ^ t39;
135*0957b409SSimon J. Gerraty
136*0957b409SSimon J. Gerraty t41 = t40 ^ t37;
137*0957b409SSimon J. Gerraty t42 = t29 ^ t33;
138*0957b409SSimon J. Gerraty t43 = t29 ^ t40;
139*0957b409SSimon J. Gerraty t44 = t33 ^ t37;
140*0957b409SSimon J. Gerraty t45 = t42 ^ t41;
141*0957b409SSimon J. Gerraty z0 = t44 & y15;
142*0957b409SSimon J. Gerraty z1 = t37 & y6;
143*0957b409SSimon J. Gerraty z2 = t33 & x7;
144*0957b409SSimon J. Gerraty z3 = t43 & y16;
145*0957b409SSimon J. Gerraty z4 = t40 & y1;
146*0957b409SSimon J. Gerraty z5 = t29 & y7;
147*0957b409SSimon J. Gerraty z6 = t42 & y11;
148*0957b409SSimon J. Gerraty z7 = t45 & y17;
149*0957b409SSimon J. Gerraty z8 = t41 & y10;
150*0957b409SSimon J. Gerraty z9 = t44 & y12;
151*0957b409SSimon J. Gerraty z10 = t37 & y3;
152*0957b409SSimon J. Gerraty z11 = t33 & y4;
153*0957b409SSimon J. Gerraty z12 = t43 & y13;
154*0957b409SSimon J. Gerraty z13 = t40 & y5;
155*0957b409SSimon J. Gerraty z14 = t29 & y2;
156*0957b409SSimon J. Gerraty z15 = t42 & y9;
157*0957b409SSimon J. Gerraty z16 = t45 & y14;
158*0957b409SSimon J. Gerraty z17 = t41 & y8;
159*0957b409SSimon J. Gerraty
160*0957b409SSimon J. Gerraty /*
161*0957b409SSimon J. Gerraty * Bottom linear transformation.
162*0957b409SSimon J. Gerraty */
163*0957b409SSimon J. Gerraty t46 = z15 ^ z16;
164*0957b409SSimon J. Gerraty t47 = z10 ^ z11;
165*0957b409SSimon J. Gerraty t48 = z5 ^ z13;
166*0957b409SSimon J. Gerraty t49 = z9 ^ z10;
167*0957b409SSimon J. Gerraty t50 = z2 ^ z12;
168*0957b409SSimon J. Gerraty t51 = z2 ^ z5;
169*0957b409SSimon J. Gerraty t52 = z7 ^ z8;
170*0957b409SSimon J. Gerraty t53 = z0 ^ z3;
171*0957b409SSimon J. Gerraty t54 = z6 ^ z7;
172*0957b409SSimon J. Gerraty t55 = z16 ^ z17;
173*0957b409SSimon J. Gerraty t56 = z12 ^ t48;
174*0957b409SSimon J. Gerraty t57 = t50 ^ t53;
175*0957b409SSimon J. Gerraty t58 = z4 ^ t46;
176*0957b409SSimon J. Gerraty t59 = z3 ^ t54;
177*0957b409SSimon J. Gerraty t60 = t46 ^ t57;
178*0957b409SSimon J. Gerraty t61 = z14 ^ t57;
179*0957b409SSimon J. Gerraty t62 = t52 ^ t58;
180*0957b409SSimon J. Gerraty t63 = t49 ^ t58;
181*0957b409SSimon J. Gerraty t64 = z4 ^ t59;
182*0957b409SSimon J. Gerraty t65 = t61 ^ t62;
183*0957b409SSimon J. Gerraty t66 = z1 ^ t63;
184*0957b409SSimon J. Gerraty s0 = t59 ^ t63;
185*0957b409SSimon J. Gerraty s6 = t56 ^ ~t62;
186*0957b409SSimon J. Gerraty s7 = t48 ^ ~t60;
187*0957b409SSimon J. Gerraty t67 = t64 ^ t65;
188*0957b409SSimon J. Gerraty s3 = t53 ^ t66;
189*0957b409SSimon J. Gerraty s4 = t51 ^ t66;
190*0957b409SSimon J. Gerraty s5 = t47 ^ t65;
191*0957b409SSimon J. Gerraty s1 = t64 ^ ~s3;
192*0957b409SSimon J. Gerraty s2 = t55 ^ ~t67;
193*0957b409SSimon J. Gerraty
194*0957b409SSimon J. Gerraty q[7] = s0;
195*0957b409SSimon J. Gerraty q[6] = s1;
196*0957b409SSimon J. Gerraty q[5] = s2;
197*0957b409SSimon J. Gerraty q[4] = s3;
198*0957b409SSimon J. Gerraty q[3] = s4;
199*0957b409SSimon J. Gerraty q[2] = s5;
200*0957b409SSimon J. Gerraty q[1] = s6;
201*0957b409SSimon J. Gerraty q[0] = s7;
202*0957b409SSimon J. Gerraty }
203*0957b409SSimon J. Gerraty
204*0957b409SSimon J. Gerraty /* see inner.h */
205*0957b409SSimon J. Gerraty void
br_aes_ct64_ortho(uint64_t * q)206*0957b409SSimon J. Gerraty br_aes_ct64_ortho(uint64_t *q)
207*0957b409SSimon J. Gerraty {
208*0957b409SSimon J. Gerraty #define SWAPN(cl, ch, s, x, y) do { \
209*0957b409SSimon J. Gerraty uint64_t a, b; \
210*0957b409SSimon J. Gerraty a = (x); \
211*0957b409SSimon J. Gerraty b = (y); \
212*0957b409SSimon J. Gerraty (x) = (a & (uint64_t)cl) | ((b & (uint64_t)cl) << (s)); \
213*0957b409SSimon J. Gerraty (y) = ((a & (uint64_t)ch) >> (s)) | (b & (uint64_t)ch); \
214*0957b409SSimon J. Gerraty } while (0)
215*0957b409SSimon J. Gerraty
216*0957b409SSimon J. Gerraty #define SWAP2(x, y) SWAPN(0x5555555555555555, 0xAAAAAAAAAAAAAAAA, 1, x, y)
217*0957b409SSimon J. Gerraty #define SWAP4(x, y) SWAPN(0x3333333333333333, 0xCCCCCCCCCCCCCCCC, 2, x, y)
218*0957b409SSimon J. Gerraty #define SWAP8(x, y) SWAPN(0x0F0F0F0F0F0F0F0F, 0xF0F0F0F0F0F0F0F0, 4, x, y)
219*0957b409SSimon J. Gerraty
220*0957b409SSimon J. Gerraty SWAP2(q[0], q[1]);
221*0957b409SSimon J. Gerraty SWAP2(q[2], q[3]);
222*0957b409SSimon J. Gerraty SWAP2(q[4], q[5]);
223*0957b409SSimon J. Gerraty SWAP2(q[6], q[7]);
224*0957b409SSimon J. Gerraty
225*0957b409SSimon J. Gerraty SWAP4(q[0], q[2]);
226*0957b409SSimon J. Gerraty SWAP4(q[1], q[3]);
227*0957b409SSimon J. Gerraty SWAP4(q[4], q[6]);
228*0957b409SSimon J. Gerraty SWAP4(q[5], q[7]);
229*0957b409SSimon J. Gerraty
230*0957b409SSimon J. Gerraty SWAP8(q[0], q[4]);
231*0957b409SSimon J. Gerraty SWAP8(q[1], q[5]);
232*0957b409SSimon J. Gerraty SWAP8(q[2], q[6]);
233*0957b409SSimon J. Gerraty SWAP8(q[3], q[7]);
234*0957b409SSimon J. Gerraty }
235*0957b409SSimon J. Gerraty
236*0957b409SSimon J. Gerraty /* see inner.h */
237*0957b409SSimon J. Gerraty void
br_aes_ct64_interleave_in(uint64_t * q0,uint64_t * q1,const uint32_t * w)238*0957b409SSimon J. Gerraty br_aes_ct64_interleave_in(uint64_t *q0, uint64_t *q1, const uint32_t *w)
239*0957b409SSimon J. Gerraty {
240*0957b409SSimon J. Gerraty uint64_t x0, x1, x2, x3;
241*0957b409SSimon J. Gerraty
242*0957b409SSimon J. Gerraty x0 = w[0];
243*0957b409SSimon J. Gerraty x1 = w[1];
244*0957b409SSimon J. Gerraty x2 = w[2];
245*0957b409SSimon J. Gerraty x3 = w[3];
246*0957b409SSimon J. Gerraty x0 |= (x0 << 16);
247*0957b409SSimon J. Gerraty x1 |= (x1 << 16);
248*0957b409SSimon J. Gerraty x2 |= (x2 << 16);
249*0957b409SSimon J. Gerraty x3 |= (x3 << 16);
250*0957b409SSimon J. Gerraty x0 &= (uint64_t)0x0000FFFF0000FFFF;
251*0957b409SSimon J. Gerraty x1 &= (uint64_t)0x0000FFFF0000FFFF;
252*0957b409SSimon J. Gerraty x2 &= (uint64_t)0x0000FFFF0000FFFF;
253*0957b409SSimon J. Gerraty x3 &= (uint64_t)0x0000FFFF0000FFFF;
254*0957b409SSimon J. Gerraty x0 |= (x0 << 8);
255*0957b409SSimon J. Gerraty x1 |= (x1 << 8);
256*0957b409SSimon J. Gerraty x2 |= (x2 << 8);
257*0957b409SSimon J. Gerraty x3 |= (x3 << 8);
258*0957b409SSimon J. Gerraty x0 &= (uint64_t)0x00FF00FF00FF00FF;
259*0957b409SSimon J. Gerraty x1 &= (uint64_t)0x00FF00FF00FF00FF;
260*0957b409SSimon J. Gerraty x2 &= (uint64_t)0x00FF00FF00FF00FF;
261*0957b409SSimon J. Gerraty x3 &= (uint64_t)0x00FF00FF00FF00FF;
262*0957b409SSimon J. Gerraty *q0 = x0 | (x2 << 8);
263*0957b409SSimon J. Gerraty *q1 = x1 | (x3 << 8);
264*0957b409SSimon J. Gerraty }
265*0957b409SSimon J. Gerraty
266*0957b409SSimon J. Gerraty /* see inner.h */
267*0957b409SSimon J. Gerraty void
br_aes_ct64_interleave_out(uint32_t * w,uint64_t q0,uint64_t q1)268*0957b409SSimon J. Gerraty br_aes_ct64_interleave_out(uint32_t *w, uint64_t q0, uint64_t q1)
269*0957b409SSimon J. Gerraty {
270*0957b409SSimon J. Gerraty uint64_t x0, x1, x2, x3;
271*0957b409SSimon J. Gerraty
272*0957b409SSimon J. Gerraty x0 = q0 & (uint64_t)0x00FF00FF00FF00FF;
273*0957b409SSimon J. Gerraty x1 = q1 & (uint64_t)0x00FF00FF00FF00FF;
274*0957b409SSimon J. Gerraty x2 = (q0 >> 8) & (uint64_t)0x00FF00FF00FF00FF;
275*0957b409SSimon J. Gerraty x3 = (q1 >> 8) & (uint64_t)0x00FF00FF00FF00FF;
276*0957b409SSimon J. Gerraty x0 |= (x0 >> 8);
277*0957b409SSimon J. Gerraty x1 |= (x1 >> 8);
278*0957b409SSimon J. Gerraty x2 |= (x2 >> 8);
279*0957b409SSimon J. Gerraty x3 |= (x3 >> 8);
280*0957b409SSimon J. Gerraty x0 &= (uint64_t)0x0000FFFF0000FFFF;
281*0957b409SSimon J. Gerraty x1 &= (uint64_t)0x0000FFFF0000FFFF;
282*0957b409SSimon J. Gerraty x2 &= (uint64_t)0x0000FFFF0000FFFF;
283*0957b409SSimon J. Gerraty x3 &= (uint64_t)0x0000FFFF0000FFFF;
284*0957b409SSimon J. Gerraty w[0] = (uint32_t)x0 | (uint32_t)(x0 >> 16);
285*0957b409SSimon J. Gerraty w[1] = (uint32_t)x1 | (uint32_t)(x1 >> 16);
286*0957b409SSimon J. Gerraty w[2] = (uint32_t)x2 | (uint32_t)(x2 >> 16);
287*0957b409SSimon J. Gerraty w[3] = (uint32_t)x3 | (uint32_t)(x3 >> 16);
288*0957b409SSimon J. Gerraty }
289*0957b409SSimon J. Gerraty
290*0957b409SSimon J. Gerraty static const unsigned char Rcon[] = {
291*0957b409SSimon J. Gerraty 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x1B, 0x36
292*0957b409SSimon J. Gerraty };
293*0957b409SSimon J. Gerraty
294*0957b409SSimon J. Gerraty static uint32_t
sub_word(uint32_t x)295*0957b409SSimon J. Gerraty sub_word(uint32_t x)
296*0957b409SSimon J. Gerraty {
297*0957b409SSimon J. Gerraty uint64_t q[8];
298*0957b409SSimon J. Gerraty
299*0957b409SSimon J. Gerraty memset(q, 0, sizeof q);
300*0957b409SSimon J. Gerraty q[0] = x;
301*0957b409SSimon J. Gerraty br_aes_ct64_ortho(q);
302*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_Sbox(q);
303*0957b409SSimon J. Gerraty br_aes_ct64_ortho(q);
304*0957b409SSimon J. Gerraty return (uint32_t)q[0];
305*0957b409SSimon J. Gerraty }
306*0957b409SSimon J. Gerraty
307*0957b409SSimon J. Gerraty /* see inner.h */
308*0957b409SSimon J. Gerraty unsigned
br_aes_ct64_keysched(uint64_t * comp_skey,const void * key,size_t key_len)309*0957b409SSimon J. Gerraty br_aes_ct64_keysched(uint64_t *comp_skey, const void *key, size_t key_len)
310*0957b409SSimon J. Gerraty {
311*0957b409SSimon J. Gerraty unsigned num_rounds;
312*0957b409SSimon J. Gerraty int i, j, k, nk, nkf;
313*0957b409SSimon J. Gerraty uint32_t tmp;
314*0957b409SSimon J. Gerraty uint32_t skey[60];
315*0957b409SSimon J. Gerraty
316*0957b409SSimon J. Gerraty switch (key_len) {
317*0957b409SSimon J. Gerraty case 16:
318*0957b409SSimon J. Gerraty num_rounds = 10;
319*0957b409SSimon J. Gerraty break;
320*0957b409SSimon J. Gerraty case 24:
321*0957b409SSimon J. Gerraty num_rounds = 12;
322*0957b409SSimon J. Gerraty break;
323*0957b409SSimon J. Gerraty case 32:
324*0957b409SSimon J. Gerraty num_rounds = 14;
325*0957b409SSimon J. Gerraty break;
326*0957b409SSimon J. Gerraty default:
327*0957b409SSimon J. Gerraty /* abort(); */
328*0957b409SSimon J. Gerraty return 0;
329*0957b409SSimon J. Gerraty }
330*0957b409SSimon J. Gerraty nk = (int)(key_len >> 2);
331*0957b409SSimon J. Gerraty nkf = (int)((num_rounds + 1) << 2);
332*0957b409SSimon J. Gerraty br_range_dec32le(skey, (key_len >> 2), key);
333*0957b409SSimon J. Gerraty tmp = skey[(key_len >> 2) - 1];
334*0957b409SSimon J. Gerraty for (i = nk, j = 0, k = 0; i < nkf; i ++) {
335*0957b409SSimon J. Gerraty if (j == 0) {
336*0957b409SSimon J. Gerraty tmp = (tmp << 24) | (tmp >> 8);
337*0957b409SSimon J. Gerraty tmp = sub_word(tmp) ^ Rcon[k];
338*0957b409SSimon J. Gerraty } else if (nk > 6 && j == 4) {
339*0957b409SSimon J. Gerraty tmp = sub_word(tmp);
340*0957b409SSimon J. Gerraty }
341*0957b409SSimon J. Gerraty tmp ^= skey[i - nk];
342*0957b409SSimon J. Gerraty skey[i] = tmp;
343*0957b409SSimon J. Gerraty if (++ j == nk) {
344*0957b409SSimon J. Gerraty j = 0;
345*0957b409SSimon J. Gerraty k ++;
346*0957b409SSimon J. Gerraty }
347*0957b409SSimon J. Gerraty }
348*0957b409SSimon J. Gerraty
349*0957b409SSimon J. Gerraty for (i = 0, j = 0; i < nkf; i += 4, j += 2) {
350*0957b409SSimon J. Gerraty uint64_t q[8];
351*0957b409SSimon J. Gerraty
352*0957b409SSimon J. Gerraty br_aes_ct64_interleave_in(&q[0], &q[4], skey + i);
353*0957b409SSimon J. Gerraty q[1] = q[0];
354*0957b409SSimon J. Gerraty q[2] = q[0];
355*0957b409SSimon J. Gerraty q[3] = q[0];
356*0957b409SSimon J. Gerraty q[5] = q[4];
357*0957b409SSimon J. Gerraty q[6] = q[4];
358*0957b409SSimon J. Gerraty q[7] = q[4];
359*0957b409SSimon J. Gerraty br_aes_ct64_ortho(q);
360*0957b409SSimon J. Gerraty comp_skey[j + 0] =
361*0957b409SSimon J. Gerraty (q[0] & (uint64_t)0x1111111111111111)
362*0957b409SSimon J. Gerraty | (q[1] & (uint64_t)0x2222222222222222)
363*0957b409SSimon J. Gerraty | (q[2] & (uint64_t)0x4444444444444444)
364*0957b409SSimon J. Gerraty | (q[3] & (uint64_t)0x8888888888888888);
365*0957b409SSimon J. Gerraty comp_skey[j + 1] =
366*0957b409SSimon J. Gerraty (q[4] & (uint64_t)0x1111111111111111)
367*0957b409SSimon J. Gerraty | (q[5] & (uint64_t)0x2222222222222222)
368*0957b409SSimon J. Gerraty | (q[6] & (uint64_t)0x4444444444444444)
369*0957b409SSimon J. Gerraty | (q[7] & (uint64_t)0x8888888888888888);
370*0957b409SSimon J. Gerraty }
371*0957b409SSimon J. Gerraty return num_rounds;
372*0957b409SSimon J. Gerraty }
373*0957b409SSimon J. Gerraty
374*0957b409SSimon J. Gerraty /* see inner.h */
375*0957b409SSimon J. Gerraty void
br_aes_ct64_skey_expand(uint64_t * skey,unsigned num_rounds,const uint64_t * comp_skey)376*0957b409SSimon J. Gerraty br_aes_ct64_skey_expand(uint64_t *skey,
377*0957b409SSimon J. Gerraty unsigned num_rounds, const uint64_t *comp_skey)
378*0957b409SSimon J. Gerraty {
379*0957b409SSimon J. Gerraty unsigned u, v, n;
380*0957b409SSimon J. Gerraty
381*0957b409SSimon J. Gerraty n = (num_rounds + 1) << 1;
382*0957b409SSimon J. Gerraty for (u = 0, v = 0; u < n; u ++, v += 4) {
383*0957b409SSimon J. Gerraty uint64_t x0, x1, x2, x3;
384*0957b409SSimon J. Gerraty
385*0957b409SSimon J. Gerraty x0 = x1 = x2 = x3 = comp_skey[u];
386*0957b409SSimon J. Gerraty x0 &= (uint64_t)0x1111111111111111;
387*0957b409SSimon J. Gerraty x1 &= (uint64_t)0x2222222222222222;
388*0957b409SSimon J. Gerraty x2 &= (uint64_t)0x4444444444444444;
389*0957b409SSimon J. Gerraty x3 &= (uint64_t)0x8888888888888888;
390*0957b409SSimon J. Gerraty x1 >>= 1;
391*0957b409SSimon J. Gerraty x2 >>= 2;
392*0957b409SSimon J. Gerraty x3 >>= 3;
393*0957b409SSimon J. Gerraty skey[v + 0] = (x0 << 4) - x0;
394*0957b409SSimon J. Gerraty skey[v + 1] = (x1 << 4) - x1;
395*0957b409SSimon J. Gerraty skey[v + 2] = (x2 << 4) - x2;
396*0957b409SSimon J. Gerraty skey[v + 3] = (x3 << 4) - x3;
397*0957b409SSimon J. Gerraty }
398*0957b409SSimon J. Gerraty }
399