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_invSbox(uint64_t * q)29*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_invSbox(uint64_t *q)
30*0957b409SSimon J. Gerraty {
31*0957b409SSimon J. Gerraty /*
32*0957b409SSimon J. Gerraty * See br_aes_ct_bitslice_invSbox(). This is the natural extension
33*0957b409SSimon J. Gerraty * to 64-bit registers.
34*0957b409SSimon J. Gerraty */
35*0957b409SSimon J. Gerraty uint64_t q0, q1, q2, q3, q4, q5, q6, q7;
36*0957b409SSimon J. Gerraty
37*0957b409SSimon J. Gerraty q0 = ~q[0];
38*0957b409SSimon J. Gerraty q1 = ~q[1];
39*0957b409SSimon J. Gerraty q2 = q[2];
40*0957b409SSimon J. Gerraty q3 = q[3];
41*0957b409SSimon J. Gerraty q4 = q[4];
42*0957b409SSimon J. Gerraty q5 = ~q[5];
43*0957b409SSimon J. Gerraty q6 = ~q[6];
44*0957b409SSimon J. Gerraty q7 = q[7];
45*0957b409SSimon J. Gerraty q[7] = q1 ^ q4 ^ q6;
46*0957b409SSimon J. Gerraty q[6] = q0 ^ q3 ^ q5;
47*0957b409SSimon J. Gerraty q[5] = q7 ^ q2 ^ q4;
48*0957b409SSimon J. Gerraty q[4] = q6 ^ q1 ^ q3;
49*0957b409SSimon J. Gerraty q[3] = q5 ^ q0 ^ q2;
50*0957b409SSimon J. Gerraty q[2] = q4 ^ q7 ^ q1;
51*0957b409SSimon J. Gerraty q[1] = q3 ^ q6 ^ q0;
52*0957b409SSimon J. Gerraty q[0] = q2 ^ q5 ^ q7;
53*0957b409SSimon J. Gerraty
54*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_Sbox(q);
55*0957b409SSimon J. Gerraty
56*0957b409SSimon J. Gerraty q0 = ~q[0];
57*0957b409SSimon J. Gerraty q1 = ~q[1];
58*0957b409SSimon J. Gerraty q2 = q[2];
59*0957b409SSimon J. Gerraty q3 = q[3];
60*0957b409SSimon J. Gerraty q4 = q[4];
61*0957b409SSimon J. Gerraty q5 = ~q[5];
62*0957b409SSimon J. Gerraty q6 = ~q[6];
63*0957b409SSimon J. Gerraty q7 = q[7];
64*0957b409SSimon J. Gerraty q[7] = q1 ^ q4 ^ q6;
65*0957b409SSimon J. Gerraty q[6] = q0 ^ q3 ^ q5;
66*0957b409SSimon J. Gerraty q[5] = q7 ^ q2 ^ q4;
67*0957b409SSimon J. Gerraty q[4] = q6 ^ q1 ^ q3;
68*0957b409SSimon J. Gerraty q[3] = q5 ^ q0 ^ q2;
69*0957b409SSimon J. Gerraty q[2] = q4 ^ q7 ^ q1;
70*0957b409SSimon J. Gerraty q[1] = q3 ^ q6 ^ q0;
71*0957b409SSimon J. Gerraty q[0] = q2 ^ q5 ^ q7;
72*0957b409SSimon J. Gerraty }
73*0957b409SSimon J. Gerraty
74*0957b409SSimon J. Gerraty static void
add_round_key(uint64_t * q,const uint64_t * sk)75*0957b409SSimon J. Gerraty add_round_key(uint64_t *q, const uint64_t *sk)
76*0957b409SSimon J. Gerraty {
77*0957b409SSimon J. Gerraty int i;
78*0957b409SSimon J. Gerraty
79*0957b409SSimon J. Gerraty for (i = 0; i < 8; i ++) {
80*0957b409SSimon J. Gerraty q[i] ^= sk[i];
81*0957b409SSimon J. Gerraty }
82*0957b409SSimon J. Gerraty }
83*0957b409SSimon J. Gerraty
84*0957b409SSimon J. Gerraty static void
inv_shift_rows(uint64_t * q)85*0957b409SSimon J. Gerraty inv_shift_rows(uint64_t *q)
86*0957b409SSimon J. Gerraty {
87*0957b409SSimon J. Gerraty int i;
88*0957b409SSimon J. Gerraty
89*0957b409SSimon J. Gerraty for (i = 0; i < 8; i ++) {
90*0957b409SSimon J. Gerraty uint64_t x;
91*0957b409SSimon J. Gerraty
92*0957b409SSimon J. Gerraty x = q[i];
93*0957b409SSimon J. Gerraty q[i] = (x & (uint64_t)0x000000000000FFFF)
94*0957b409SSimon J. Gerraty | ((x & (uint64_t)0x000000000FFF0000) << 4)
95*0957b409SSimon J. Gerraty | ((x & (uint64_t)0x00000000F0000000) >> 12)
96*0957b409SSimon J. Gerraty | ((x & (uint64_t)0x000000FF00000000) << 8)
97*0957b409SSimon J. Gerraty | ((x & (uint64_t)0x0000FF0000000000) >> 8)
98*0957b409SSimon J. Gerraty | ((x & (uint64_t)0x000F000000000000) << 12)
99*0957b409SSimon J. Gerraty | ((x & (uint64_t)0xFFF0000000000000) >> 4);
100*0957b409SSimon J. Gerraty }
101*0957b409SSimon J. Gerraty }
102*0957b409SSimon J. Gerraty
103*0957b409SSimon J. Gerraty static inline uint64_t
rotr32(uint64_t x)104*0957b409SSimon J. Gerraty rotr32(uint64_t x)
105*0957b409SSimon J. Gerraty {
106*0957b409SSimon J. Gerraty return (x << 32) | (x >> 32);
107*0957b409SSimon J. Gerraty }
108*0957b409SSimon J. Gerraty
109*0957b409SSimon J. Gerraty static void
inv_mix_columns(uint64_t * q)110*0957b409SSimon J. Gerraty inv_mix_columns(uint64_t *q)
111*0957b409SSimon J. Gerraty {
112*0957b409SSimon J. Gerraty uint64_t q0, q1, q2, q3, q4, q5, q6, q7;
113*0957b409SSimon J. Gerraty uint64_t r0, r1, r2, r3, r4, r5, r6, r7;
114*0957b409SSimon J. Gerraty
115*0957b409SSimon J. Gerraty q0 = q[0];
116*0957b409SSimon J. Gerraty q1 = q[1];
117*0957b409SSimon J. Gerraty q2 = q[2];
118*0957b409SSimon J. Gerraty q3 = q[3];
119*0957b409SSimon J. Gerraty q4 = q[4];
120*0957b409SSimon J. Gerraty q5 = q[5];
121*0957b409SSimon J. Gerraty q6 = q[6];
122*0957b409SSimon J. Gerraty q7 = q[7];
123*0957b409SSimon J. Gerraty r0 = (q0 >> 16) | (q0 << 48);
124*0957b409SSimon J. Gerraty r1 = (q1 >> 16) | (q1 << 48);
125*0957b409SSimon J. Gerraty r2 = (q2 >> 16) | (q2 << 48);
126*0957b409SSimon J. Gerraty r3 = (q3 >> 16) | (q3 << 48);
127*0957b409SSimon J. Gerraty r4 = (q4 >> 16) | (q4 << 48);
128*0957b409SSimon J. Gerraty r5 = (q5 >> 16) | (q5 << 48);
129*0957b409SSimon J. Gerraty r6 = (q6 >> 16) | (q6 << 48);
130*0957b409SSimon J. Gerraty r7 = (q7 >> 16) | (q7 << 48);
131*0957b409SSimon J. Gerraty
132*0957b409SSimon J. Gerraty q[0] = q5 ^ q6 ^ q7 ^ r0 ^ r5 ^ r7 ^ rotr32(q0 ^ q5 ^ q6 ^ r0 ^ r5);
133*0957b409SSimon J. Gerraty q[1] = q0 ^ q5 ^ r0 ^ r1 ^ r5 ^ r6 ^ r7 ^ rotr32(q1 ^ q5 ^ q7 ^ r1 ^ r5 ^ r6);
134*0957b409SSimon J. Gerraty q[2] = q0 ^ q1 ^ q6 ^ r1 ^ r2 ^ r6 ^ r7 ^ rotr32(q0 ^ q2 ^ q6 ^ r2 ^ r6 ^ r7);
135*0957b409SSimon J. Gerraty q[3] = q0 ^ q1 ^ q2 ^ q5 ^ q6 ^ r0 ^ r2 ^ r3 ^ r5 ^ rotr32(q0 ^ q1 ^ q3 ^ q5 ^ q6 ^ q7 ^ r0 ^ r3 ^ r5 ^ r7);
136*0957b409SSimon J. Gerraty q[4] = q1 ^ q2 ^ q3 ^ q5 ^ r1 ^ r3 ^ r4 ^ r5 ^ r6 ^ r7 ^ rotr32(q1 ^ q2 ^ q4 ^ q5 ^ q7 ^ r1 ^ r4 ^ r5 ^ r6);
137*0957b409SSimon J. Gerraty q[5] = q2 ^ q3 ^ q4 ^ q6 ^ r2 ^ r4 ^ r5 ^ r6 ^ r7 ^ rotr32(q2 ^ q3 ^ q5 ^ q6 ^ r2 ^ r5 ^ r6 ^ r7);
138*0957b409SSimon J. Gerraty q[6] = q3 ^ q4 ^ q5 ^ q7 ^ r3 ^ r5 ^ r6 ^ r7 ^ rotr32(q3 ^ q4 ^ q6 ^ q7 ^ r3 ^ r6 ^ r7);
139*0957b409SSimon J. Gerraty q[7] = q4 ^ q5 ^ q6 ^ r4 ^ r6 ^ r7 ^ rotr32(q4 ^ q5 ^ q7 ^ r4 ^ r7);
140*0957b409SSimon J. Gerraty }
141*0957b409SSimon J. Gerraty
142*0957b409SSimon J. Gerraty /* see inner.h */
143*0957b409SSimon J. Gerraty void
br_aes_ct64_bitslice_decrypt(unsigned num_rounds,const uint64_t * skey,uint64_t * q)144*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_decrypt(unsigned num_rounds,
145*0957b409SSimon J. Gerraty const uint64_t *skey, uint64_t *q)
146*0957b409SSimon J. Gerraty {
147*0957b409SSimon J. Gerraty unsigned u;
148*0957b409SSimon J. Gerraty
149*0957b409SSimon J. Gerraty add_round_key(q, skey + (num_rounds << 3));
150*0957b409SSimon J. Gerraty for (u = num_rounds - 1; u > 0; u --) {
151*0957b409SSimon J. Gerraty inv_shift_rows(q);
152*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_invSbox(q);
153*0957b409SSimon J. Gerraty add_round_key(q, skey + (u << 3));
154*0957b409SSimon J. Gerraty inv_mix_columns(q);
155*0957b409SSimon J. Gerraty }
156*0957b409SSimon J. Gerraty inv_shift_rows(q);
157*0957b409SSimon J. Gerraty br_aes_ct64_bitslice_invSbox(q);
158*0957b409SSimon J. Gerraty add_round_key(q, skey);
159*0957b409SSimon J. Gerraty }
160