cast5.c

   1 /* cast5.c */
   2 /*
   3     This file is part of the Crypto-avr-lib/microcrypt-lib.
   4     Copyright (C) 2008  Daniel Otte (daniel.otte@rub.de)
   5
   6     This program is free software: you can redistribute it and/or modify
   7     it under the terms of the GNU General Public License as published by
   8     the Free Software Foundation, either version 3 of the License, or
   9     (at your option) any later version.
  10
  11     This program is distributed in the hope that it will be useful,
  12     but WITHOUT ANY WARRANTY; without even the implied warranty of
  13     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14     GNU General Public License for more details.
  15
  16     You should have received a copy of the GNU General Public License
  17     along with this program.  If not, see <http://www.gnu.org/licenses/>.
  18 */
  19 /*
  20  * \file        cast5.c
  21  * \author      Daniel Otte
  22  * \email       daniel.otte@rub.de
  23  * \date        2006-07-26
  24  * \par License:
  25  *  GPLv3 or later
  26  * \brief Implementation of the CAST5 (aka CAST-128) cipher algorithm as described in RFC 2144
  27  *
  28  */
  29
  30  #include <stdint.h>
  31  #include <string.h>
  32  #include "cast5.h"
  33  #include "config.h"
  34  #include "debug.h"
  35
  36  #undef DEBUG
  37
  38  #ifdef DEBUG
  39   #include "uart.h"
  40  #endif
  41
  42 #include "cast5-sbox.h"
  43
  44
  45
  46 #define S5(x) pgm_read_dword(&s5[(x)])
  47 #define S6(x) pgm_read_dword(&s6[(x)])
  48 #define S7(x) pgm_read_dword(&s7[(x)])
  49 #define S8(x) pgm_read_dword(&s8[(x)])
  50
  51
  52 void cast5_init_A(uint8_t *dest, uint8_t *src, bool bmode){
  53         uint8_t mask = bmode?0x8:0;
  54         *((uint32_t*)(&dest[0x0])) = *((uint32_t*)(&src[0x0^mask])) ^ S5(src[0xD^mask]) ^ S6(src[0xF^mask]) ^ S7(src[0xC^mask]) ^ S8(src[0xE^mask]) ^ S7(src[0x8^mask]);
  55         *((uint32_t*)(&dest[0x4])) = *((uint32_t*)(&src[0x8^mask])) ^ S5(dest[0x0]) ^ S6(dest[0x2]) ^ S7(dest[0x1]) ^ S8(dest[0x3]) ^ S8(src[0xA^mask]);
  56         *((uint32_t*)(&dest[0x8])) = *((uint32_t*)(&src[0xC^mask])) ^ S5(dest[0x7]) ^ S6(dest[0x6]) ^ S7(dest[0x5]) ^ S8(dest[0x4]) ^ S5(src[0x9^mask]);
  57         *((uint32_t*)(&dest[0xC])) = *((uint32_t*)(&src[0x4^mask])) ^ S5(dest[0xA]) ^ S6(dest[0x9]) ^ S7(dest[0xB]) ^ S8(dest[0x8]) ^ S6(src[0xB^mask]);
  58 }
  59
  60 void cast5_init_M(uint8_t *dest, uint8_t *src, bool nmode, bool xmode){
  61         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  62         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  63         #define NMT(x) (src[nmode?nmt[(x)]:(x)])
  64         #define XMT(x) (src[xmt[(xmode<<1) + nmode][(x)]])
  65         *((uint32_t*)(&dest[0x0])) = S5(NMT(0x8)) ^ S6(NMT(0x9)) ^ S7(NMT(0x7)) ^ S8(NMT(0x6)) ^ S5(XMT(0));
  66         *((uint32_t*)(&dest[0x4])) = S5(NMT(0xA)) ^ S6(NMT(0xB)) ^ S7(NMT(0x5)) ^ S8(NMT(0x4)) ^ S6(XMT(1));
  67         *((uint32_t*)(&dest[0x8])) = S5(NMT(0xC)) ^ S6(NMT(0xD)) ^ S7(NMT(0x3)) ^ S8(NMT(0x2)) ^ S7(XMT(2));
  68         *((uint32_t*)(&dest[0xC])) = S5(NMT(0xE)) ^ S6(NMT(0xF)) ^ S7(NMT(0x1)) ^ S8(NMT(0x0)) ^ S8(XMT(3));
  69 }
  70
  71 #define S5B(x) pgm_read_byte(3+(uint8_t*)(&s5[(x)]))
  72 #define S6B(x) pgm_read_byte(3+(uint8_t*)(&s6[(x)]))
  73 #define S7B(x) pgm_read_byte(3+(uint8_t*)(&s7[(x)]))
  74 #define S8B(x) pgm_read_byte(3+(uint8_t*)(&s8[(x)]))
  75
  76 void cast5_init_rM(uint8_t *klo, uint8_t *khi, uint8_t offset, uint8_t *src, bool nmode, bool xmode){
  77         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  78         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  79         uint8_t t, h=0;
  80         t = S5B(NMT(0x8)) ^ S6B(NMT(0x9)) ^ S7B(NMT(0x7)) ^ S8B(NMT(0x6)) ^ S5B(XMT(0));
  81                 klo[offset*2] |= (t & 0x0f);
  82                 h |= (t&0x10); h>>=1;
  83         t = S5B(NMT(0xA)) ^ S6B(NMT(0xB)) ^ S7B(NMT(0x5)) ^ S8B(NMT(0x4)) ^ S6B(XMT(1));
  84                 klo[offset*2] |= (t<<4) & 0xf0;
  85                 h |= t&0x10; h>>=1;
  86         t = S5B(NMT(0xC)) ^ S6B(NMT(0xD)) ^ S7B(NMT(0x3)) ^ S8B(NMT(0x2)) ^ S7B(XMT(2));
  87                 klo[offset*2+1] |= t&0xf;
  88                 h |= t&0x10; h>>=1;
  89         t = S5B(NMT(0xE)) ^ S6B(NMT(0xF)) ^ S7B(NMT(0x1)) ^ S8B(NMT(0x0)) ^ S8B(XMT(3));
  90                 klo[offset*2+1] |= t<<4;
  91                 h |= t&0x10; h >>=1;
  92         #ifdef DEBUG
  93                 uart_putstr("\r\n\t h="); uart_hexdump(&h,1);
  94         #endif
  95         khi[offset>>1] |= h<<((offset&0x1)?4:0);
  96 }
  97
  98 #define S_5X(s) pgm_read_dword(&s5[BPX[(s)]])
  99 #define S_6X(s) pgm_read_dword(&s6[BPX[(s)]])
 100 #define S_7X(s) pgm_read_dword(&s7[BPX[(s)]])
 101 #define S_8X(s) pgm_read_dword(&s8[BPX[(s)]])
 102
 103 #define S_5Z(s) pgm_read_dword(&s5[BPZ[(s)]])
 104 #define S_6Z(s) pgm_read_dword(&s6[BPZ[(s)]])
 105 #define S_7Z(s) pgm_read_dword(&s7[BPZ[(s)]])
 106 #define S_8Z(s) pgm_read_dword(&s8[BPZ[(s)]])
 107
 108
 109
 110
 111 void cast5_init(void* key, uint8_t keylength_b, cast5_ctx_t* s){
 112          /* we migth return if the key is valid and if setup was sucessfull */
 113         uint32_t x[4], z[4];
 114         #define BPX ((uint8_t*)&(x[0]))
 115         #define BPZ ((uint8_t*)&(z[0]))
 116         s->shortkey = (keylength_b<=80);
 117         /* littel endian only! */
 118         memset(&(x[0]), 0 ,16); /* set x to zero */
 119         if(keylength_b > 128)
 120                 keylength_b=128;
 121         memcpy(&(x[0]), key, (keylength_b+7)/8);
 122
 123
 124         /* todo: merge a and b and compress the whole stuff */
 125         /***** A *****/
 126         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 127         /***** M *****/
 128         cast5_init_M((uint8_t*)(&(s->mask[0])), (uint8_t*)(&z[0]), false, false);
 129         /***** B *****/
 130         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 131         /***** N *****/
 132         cast5_init_M((uint8_t*)(&(s->mask[4])), (uint8_t*)(&x[0]), true, false);
 133         /***** A *****/
 134         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 135         /***** N' *****/
 136         cast5_init_M((uint8_t*)(&(s->mask[8])), (uint8_t*)(&z[0]), true, true);
 137         /***** B *****/
 138         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 139         /***** M' *****/
 140         cast5_init_M((uint8_t*)(&(s->mask[12])), (uint8_t*)(&x[0]), false, true);
 141
 142         /* that were the masking keys, now the rotation keys */
 143         /* set the keys to zero */
 144         memset(&(s->rotl[0]),0,8);
 145         s->roth[0]=s->roth[1]=0;
 146         /***** A *****/
 147         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 148         /***** M *****/
 149         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 0, (uint8_t*)(&z[0]), false, false);
 150         /***** B *****/
 151         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 152         /***** N *****/
 153         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 1, (uint8_t*)(&x[0]), true, false);
 154         /***** A *****/
 155         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 156         /***** N' *****/
 157         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 2, (uint8_t*)(&z[0]), true, true);
 158         /***** B *****/
 159         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 160         /***** M' *****/
 161         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 3, (uint8_t*)(&x[0]), false, true);
 162         /* done ;-) */
 163 }
 164
 165
 166
 167 /********************************************************************************************************/
 168
 169 #define ROTL32(a,n) ((a)<<(n) | (a)>>(32-(n)))
 170 #define CHANGE_ENDIAN32(x) ((x)<<24 | (x)>>24 | ((x)&0xff00)<<8 | ((x)&0xff0000)>>8 )
 171
 172 typedef uint32_t cast5_f_t(uint32_t,uint32_t,uint8_t);
 173
 174 #define IA 3
 175 #define IB 2
 176 #define IC 1
 177 #define ID 0
 178
 179
 180 uint32_t cast5_f1(uint32_t d, uint32_t m, uint8_t r){
 181         uint32_t t;
 182         t = ROTL32((d + m),r);
 183 #ifdef DEBUG
 184         uint32_t ia,ib,ic,id;
 185         uart_putstr("\r\n f1("); uart_hexdump(&d, 4); uart_putc(',');
 186                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 187                 uart_hexdump(&t, 4);
 188         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 189         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 190         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 191         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 192         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 193         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 194         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 195         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 196
 197         return (((ia ^ ib) - ic) + id);
 198
 199 #else
 200
 201         return (((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) ^ pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 202                 - pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) + pgm_read_dword(&s4[((uint8_t*)&t)[ID]]));
 203
 204 #endif
 205 }
 206
 207
 208 uint32_t cast5_f2(uint32_t d, uint32_t m, uint8_t r){
 209         uint32_t t;
 210         t = ROTL32((d ^ m),r);
 211 #ifdef DEBUG
 212         uint32_t ia,ib,ic,id;
 213         uart_putstr("\r\n f2("); uart_hexdump(&d, 4); uart_putc(',');
 214                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 215                 uart_hexdump(&t, 4);
 216
 217         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 218         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 219         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 220         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 221
 222         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 223         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 224         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 225         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 226
 227         return (((ia - ib) + ic) ^ id);
 228 #else
 229
 230         return (((pgm_read_dword(&s1[((uint8_t*)&t)[IA]])
 231                 - pgm_read_dword(&s2[((uint8_t*)&t)[IB]]) )
 232                     + pgm_read_dword(&s3[((uint8_t*)&t)[IC]]) )
 233                     ^ pgm_read_dword(&s4[((uint8_t*)&t)[ID]]) );
 234
 235 #endif
 236 }
 237
 238 uint32_t cast5_f3(uint32_t d, uint32_t m, uint8_t r){
 239         uint32_t t;
 240         t = ROTL32((m - d),r);
 241
 242 #ifdef DEBUG
 243         uint32_t ia,ib,ic,id;
 244
 245         uart_putstr("\r\n f3("); uart_hexdump(&d, 4); uart_putc(',');
 246                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 247                 uart_hexdump(&t, 4);
 248
 249         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 250         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 251         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 252         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 253
 254         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 255         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 256         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 257         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 258         return (((ia + ib) ^ ic) - id);
 259 #else
 260         return ((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) + pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 261                 ^ pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) - pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 262
 263 #endif
 264 }
 265
 266 /*************************************************************************/
 267
 268 void cast5_enc(void* block, const cast5_ctx_t *s){
 269         uint32_t l,r, x, y;
 270         uint8_t i;
 271         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 272         l=((uint32_t*)block)[0];
 273         r=((uint32_t*)block)[1];
 274 //      uart_putstr("\r\n round[-1] = ");
 275 //      uart_hexdump(&r, 4);
 276         for (i=0;i<(s->shortkey?12:16);++i){
 277                 x = r;
 278                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 279                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 280                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 281                 r = l ^ CHANGE_ENDIAN32(y);
 282 //              uart_putstr("\r\n round["); DEBUG_B(i); uart_putstr("] = ");
 283 //              uart_hexdump(&r, 4);
 284                 l = x;
 285         }
 286         ((uint32_t*)block)[0]=r;
 287         ((uint32_t*)block)[1]=l;
 288 }
 289
 290 /*************************************************************************/
 291
 292 void cast5_dec(void* block, const cast5_ctx_t *s){
 293         uint32_t l,r, x, y;
 294         int8_t i, rounds;
 295         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 296         l=((uint32_t*)block)[0];
 297         r=((uint32_t*)block)[1];
 298         rounds = (s->shortkey?12:16);
 299         for (i=rounds-1; i>=0 ;--i){
 300                 x = r;
 301                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 302                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 303                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 304                 r = l ^ CHANGE_ENDIAN32(y);
 305                 l = x;
 306         }
 307         ((uint32_t*)block)[0]=r;
 308         ((uint32_t*)block)[1]=l;
 309 }
 310
 311
 312 /*********************************************************************************************************/
 313 /*********************************************************************************************************/
 314 /*********************************************************************************************************/
 315
 316 #if 0
 317
 318 void cast5_old_init(cast5_ctx_t* s, uint8_t* key, uint8_t keylength){
 319          /* we migth return if the key is valid and if setup was sucessfull */
 320         uint32_t x[4], z[4], t;
 321         #define BPX ((uint8_t*)&(x[0]))
 322         #define BPZ ((uint8_t*)&(z[0]))
 323         s->shortkey = (keylength<=80);
 324         /* littel endian only! */
 325         memset(&(x[0]), 0 ,16); /* set x to zero */
 326         memcpy(&(x[0]), key, keylength/8);
 327
 328
 329         /* todo: merge a and b and compress the whole stuff */
 330         /***** A *****/
 331         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 332         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 333         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 334         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 335         /***** M *****/
 336         s->mask[0] = S_5Z(0x8) ^ S_6Z(0x9) ^ S_7Z(0x7) ^ S_8Z(0x6) ^ S_5Z(0x2);
 337         s->mask[1] = S_5Z(0xA) ^ S_6Z(0xB) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_6Z(0x6);
 338         s->mask[2] = S_5Z(0xC) ^ S_6Z(0xD) ^ S_7Z(0x3) ^ S_8Z(0x2) ^ S_7Z(0x9);
 339         s->mask[3] = S_5Z(0xE) ^ S_6Z(0xF) ^ S_7Z(0x1) ^ S_8Z(0x0) ^ S_8Z(0xC);
 340         /***** B *****/
 341         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 342         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 343         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 344         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 345         /***** N *****/
 346         s->mask[4] = S_5X(0x3) ^ S_6X(0x2) ^ S_7X(0xC) ^ S_8X(0xD) ^ S_5X(0x8);
 347         s->mask[5] = S_5X(0x1) ^ S_6X(0x0) ^ S_7X(0xE) ^ S_8X(0xF) ^ S_6X(0xD);
 348         s->mask[6] = S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x8) ^ S_8X(0x9) ^ S_7X(0x3);
 349         s->mask[7] = S_5X(0x5) ^ S_6X(0x4) ^ S_7X(0xA) ^ S_8X(0xB) ^ S_8X(0x7);
 350         /***** A *****/
 351         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 352         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 353         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 354         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 355         /***** N' *****/
 356         s->mask[8] = S_5Z(0x3) ^ S_6Z(0x2) ^ S_7Z(0xC) ^ S_8Z(0xD) ^ S_5Z(0x9);
 357         s->mask[9] = S_5Z(0x1) ^ S_6Z(0x0) ^ S_7Z(0xE) ^ S_8Z(0xF) ^ S_6Z(0xC);
 358         s->mask[10] = S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x8) ^ S_8Z(0x9) ^ S_7Z(0x2);
 359         s->mask[11] = S_5Z(0x5) ^ S_6Z(0x4) ^ S_7Z(0xA) ^ S_8Z(0xB) ^ S_8Z(0x6);
 360         /***** B *****/
 361         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 362         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 363         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 364         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 365         /***** M' *****/
 366         s->mask[12] = S_5X(0x8) ^ S_6X(0x9) ^ S_7X(0x7) ^ S_8X(0x6) ^ S_5X(0x3);
 367         s->mask[13] = S_5X(0xA) ^ S_6X(0xB) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_6X(0x7);
 368         s->mask[14] = S_5X(0xC) ^ S_6X(0xD) ^ S_7X(0x3) ^ S_8X(0x2) ^ S_7X(0x8);
 369         s->mask[15] = S_5X(0xE) ^ S_6X(0xF) ^ S_7X(0x1) ^ S_8X(0x0) ^ S_8X(0xD);
 370
 371         /* that were the masking keys, now the rotation keys */
 372         /* set the keys to zero */
 373         memset(&(s->rotl[0]),0,8);
 374         s->roth[0]=s->roth[1]=0;
 375         /***** A *****/
 376         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 377         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 378         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 379         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 380         /***** M *****/
 381         t = S_5Z(0x8) ^ S_6Z(0x9) ^ S_7Z(0x7) ^ S_8Z(0x6) ^ S_5Z(0x2);
 382         t >>= 24;
 383         s->rotl[0] |= t & 0x0f;
 384         s->roth[0] |= (t >> 4) & (1<<0);
 385         t = S_5Z(0xA) ^ S_6Z(0xB) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_6Z(0x6);
 386         t >>= 24;
 387         s->rotl[0] |= (t<<4) & 0xf0;
 388         s->roth[0] |= (t >> 3) & (1<<1);
 389         t = S_5Z(0xC) ^ S_6Z(0xD) ^ S_7Z(0x3) ^ S_8Z(0x2) ^ S_7Z(0x9);
 390         t >>= 24;
 391         s->rotl[1] |= t & 0x0f;
 392         s->roth[0] |= (t >> 2) & (1<<2);
 393         t = S_5Z(0xE) ^ S_6Z(0xF) ^ S_7Z(0x1) ^ S_8Z(0x0) ^ S_8Z(0xC);
 394         t >>= 24;
 395         s->rotl[1] |= (t<<4) & 0xf0;
 396         s->roth[0] |= (t >> 1) & (1<<3);
 397         /***** B *****/
 398         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 399         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 400         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 401         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 402         /***** N *****/
 403         t = S_5X(0x3) ^ S_6X(0x2) ^ S_7X(0xC) ^ S_8X(0xD) ^ S_5X(0x8);
 404         t >>= 24;
 405         s->rotl[2] |= t & 0x0f;
 406         s->roth[0] |= t & (1<<4);
 407         t = S_5X(0x1) ^ S_6X(0x0) ^ S_7X(0xE) ^ S_8X(0xF) ^ S_6X(0xD);
 408         t >>= 24;
 409         s->rotl[2] |= (t<<4) & 0xf0;
 410         s->roth[0] |= (t<<1) & (1<<5);
 411         t = S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x8) ^ S_8X(0x9) ^ S_7X(0x3);
 412         t >>= 24;
 413         s->rotl[3] |= t & 0x0f;
 414         s->roth[0] |= (t<<2) & (1<<6);
 415         t = S_5X(0x5) ^ S_6X(0x4) ^ S_7X(0xA) ^ S_8X(0xB) ^ S_8X(0x7);
 416         t >>= 24;
 417         s->rotl[3] |= (t<<4) & 0xf0;
 418         s->roth[0] |= (t<<3) & (1<<7);
 419         /***** A *****/
 420         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 421         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 422         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 423         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 424         /***** N' *****/
 425         t = S_5Z(0x3) ^ S_6Z(0x2) ^ S_7Z(0xC) ^ S_8Z(0xD) ^ S_5Z(0x9);
 426         t >>= 24;
 427         s->rotl[4] |= t & 0x0f;
 428         s->roth[1] |= (t>>4) & (1<<0);
 429         t = S_5Z(0x1) ^ S_6Z(0x0) ^ S_7Z(0xE) ^ S_8Z(0xF) ^ S_6Z(0xC);
 430         t >>= 24;
 431         s->rotl[4] |= (t<<4) & 0xf0;
 432         s->roth[1] |= (t>>3) & (1<<1);
 433         t = S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x8) ^ S_8Z(0x9) ^ S_7Z(0x2);
 434         t >>= 24;
 435         s->rotl[5] |= t & 0x0f;
 436         s->roth[1] |= (t>>2) & (1<<2);
 437         t = S_5Z(0x5) ^ S_6Z(0x4) ^ S_7Z(0xA) ^ S_8Z(0xB) ^ S_8Z(0x6);
 438         t >>= 24;
 439         s->rotl[5] |= (t<<4) & 0xf0;
 440         s->roth[1] |= (t>>1) & (1<<3);
 441         /***** B *****/
 442         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 443         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 444         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 445         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 446         /***** M' *****/
 447         t = S_5X(0x8) ^ S_6X(0x9) ^ S_7X(0x7) ^ S_8X(0x6) ^ S_5X(0x3);
 448         t >>= 24;
 449         s->rotl[6] |= t & 0x0f;
 450         s->roth[1] |= t & (1<<4);
 451         t = S_5X(0xA) ^ S_6X(0xB) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_6X(0x7);
 452         t >>= 24;
 453         s->rotl[6] |= (t<<4) & 0xf0;
 454         s->roth[1] |= (t<<1) & (1<<5);
 455         t = S_5X(0xC) ^ S_6X(0xD) ^ S_7X(0x3) ^ S_8X(0x2) ^ S_7X(0x8);
 456         t >>= 24;
 457         s->rotl[7] |= t & 0x0f;
 458         s->roth[1] |= (t<<2) & (1<<6);
 459         t = S_5X(0xE) ^ S_6X(0xF) ^ S_7X(0x1) ^ S_8X(0x0) ^ S_8X(0xD);
 460         t >>= 24;
 461         s->rotl[7] |= (t<<4) & 0xf0;
 462         s->roth[1] |= (t<<3) & (1<<7);
 463
 464         /* done ;-) */
 465 }
 466
 467 #endif
 468
 469
 470
 471