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  * \date        26.07.2006
  23  * \par License:
  24  *  GPL
  25  * \brief Implementation of the CAST5 (aka CAST-128) cipher algorithm as described in RFC 2144
  26  *
  27  */
  28
  29  #include <stdint.h>
  30  #include <string.h>
  31  #include "cast5.h"
  32  #include "config.h"
  33  #include "debug.h"
  34
  35  #undef DEBUG
  36
  37  #ifdef DEBUG
  38   #include "uart.h"
  39  #endif
  40
  41 #include "cast5-sbox.h"
  42
  43
  44
  45 #define S5(x) pgm_read_dword(&s5[(x)])
  46 #define S6(x) pgm_read_dword(&s6[(x)])
  47 #define S7(x) pgm_read_dword(&s7[(x)])
  48 #define S8(x) pgm_read_dword(&s8[(x)])
  49
  50
  51 void cast5_init_A(uint8_t *dest, uint8_t *src, bool bmode){
  52         uint8_t mask = bmode?0x8:0;
  53         *((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]);
  54         *((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]);
  55         *((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]);
  56         *((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]);
  57 }
  58
  59 void cast5_init_M(uint8_t *dest, uint8_t *src, bool nmode, bool xmode){
  60         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  61         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  62         #define NMT(x) (src[nmode?nmt[(x)]:(x)])
  63         #define XMT(x) (src[xmt[(xmode<<1) + nmode][(x)]])
  64         *((uint32_t*)(&dest[0x0])) = S5(NMT(0x8)) ^ S6(NMT(0x9)) ^ S7(NMT(0x7)) ^ S8(NMT(0x6)) ^ S5(XMT(0));
  65         *((uint32_t*)(&dest[0x4])) = S5(NMT(0xA)) ^ S6(NMT(0xB)) ^ S7(NMT(0x5)) ^ S8(NMT(0x4)) ^ S6(XMT(1));
  66         *((uint32_t*)(&dest[0x8])) = S5(NMT(0xC)) ^ S6(NMT(0xD)) ^ S7(NMT(0x3)) ^ S8(NMT(0x2)) ^ S7(XMT(2));
  67         *((uint32_t*)(&dest[0xC])) = S5(NMT(0xE)) ^ S6(NMT(0xF)) ^ S7(NMT(0x1)) ^ S8(NMT(0x0)) ^ S8(XMT(3));
  68 }
  69
  70 #define S5B(x) pgm_read_byte(3+(uint8_t*)(&s5[(x)]))
  71 #define S6B(x) pgm_read_byte(3+(uint8_t*)(&s6[(x)]))
  72 #define S7B(x) pgm_read_byte(3+(uint8_t*)(&s7[(x)]))
  73 #define S8B(x) pgm_read_byte(3+(uint8_t*)(&s8[(x)]))
  74
  75 void cast5_init_rM(uint8_t *klo, uint8_t *khi, uint8_t offset, uint8_t *src, bool nmode, bool xmode){
  76         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  77         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  78         uint8_t t, h=0;
  79         t = S5B(NMT(0x8)) ^ S6B(NMT(0x9)) ^ S7B(NMT(0x7)) ^ S8B(NMT(0x6)) ^ S5B(XMT(0));
  80                 klo[offset*2] |= (t & 0x0f);
  81                 h |= (t&0x10); h>>=1;
  82         t = S5B(NMT(0xA)) ^ S6B(NMT(0xB)) ^ S7B(NMT(0x5)) ^ S8B(NMT(0x4)) ^ S6B(XMT(1));
  83                 klo[offset*2] |= (t<<4) & 0xf0;
  84                 h |= t&0x10; h>>=1;
  85         t = S5B(NMT(0xC)) ^ S6B(NMT(0xD)) ^ S7B(NMT(0x3)) ^ S8B(NMT(0x2)) ^ S7B(XMT(2));
  86                 klo[offset*2+1] |= t&0xf;
  87                 h |= t&0x10; h>>=1;
  88         t = S5B(NMT(0xE)) ^ S6B(NMT(0xF)) ^ S7B(NMT(0x1)) ^ S8B(NMT(0x0)) ^ S8B(XMT(3));
  89                 klo[offset*2+1] |= t<<4;
  90                 h |= t&0x10; h >>=1;
  91         #ifdef DEBUG
  92                 uart_putstr("\r\n\t h="); uart_hexdump(&h,1);
  93         #endif
  94         khi[offset>>1] |= h<<((offset&0x1)?4:0);
  95 }
  96
  97 #define S_5X(s) pgm_read_dword(&s5[BPX[(s)]])
  98 #define S_6X(s) pgm_read_dword(&s6[BPX[(s)]])
  99 #define S_7X(s) pgm_read_dword(&s7[BPX[(s)]])
 100 #define S_8X(s) pgm_read_dword(&s8[BPX[(s)]])
 101
 102 #define S_5Z(s) pgm_read_dword(&s5[BPZ[(s)]])
 103 #define S_6Z(s) pgm_read_dword(&s6[BPZ[(s)]])
 104 #define S_7Z(s) pgm_read_dword(&s7[BPZ[(s)]])
 105 #define S_8Z(s) pgm_read_dword(&s8[BPZ[(s)]])
 106
 107
 108
 109 /**
 110  * \brief sets up round keys (context) for cast5 en/decryption.
 111  * @param s Pointer to cast5 context.
 112  * @param key Pointer to binary key.
 113  * @param keylength length of keydata in bits.
 114  */
 115 void cast5_init(cast5_ctx_t* s, uint8_t* key, uint8_t keylength){
 116          /* we migth return if the key is valid and if setup was sucessfull */
 117         uint32_t x[4], z[4];
 118         #define BPX ((uint8_t*)&(x[0]))
 119         #define BPZ ((uint8_t*)&(z[0]))
 120         s->shortkey = (keylength<=80);
 121         /* littel endian only! */
 122         memset(&(x[0]), 0 ,16); /* set x to zero */
 123         if(keylength > 128)
 124                 keylength=128;
 125         memcpy(&(x[0]), key, (keylength+7)/8);
 126
 127
 128         /* todo: merge a and b and compress the whole stuff */
 129         /***** A *****/
 130         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 131         /***** M *****/
 132         cast5_init_M((uint8_t*)(&(s->mask[0])), (uint8_t*)(&z[0]), false, false);
 133         /***** B *****/
 134         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 135         /***** N *****/
 136         cast5_init_M((uint8_t*)(&(s->mask[4])), (uint8_t*)(&x[0]), true, false);
 137         /***** A *****/
 138         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 139         /***** N' *****/
 140         cast5_init_M((uint8_t*)(&(s->mask[8])), (uint8_t*)(&z[0]), true, true);
 141         /***** B *****/
 142         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 143         /***** M' *****/
 144         cast5_init_M((uint8_t*)(&(s->mask[12])), (uint8_t*)(&x[0]), false, true);
 145
 146         /* that were the masking keys, now the rotation keys */
 147         /* set the keys to zero */
 148         memset(&(s->rotl[0]),0,8);
 149         s->roth[0]=s->roth[1]=0;
 150         /***** A *****/
 151         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 152         /***** M *****/
 153         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 0, (uint8_t*)(&z[0]), false, false);
 154         /***** B *****/
 155         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 156         /***** N *****/
 157         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 1, (uint8_t*)(&x[0]), true, false);
 158         /***** A *****/
 159         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 160         /***** N' *****/
 161         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 2, (uint8_t*)(&z[0]), true, true);
 162         /***** B *****/
 163         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 164         /***** M' *****/
 165         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 3, (uint8_t*)(&x[0]), false, true);
 166         /* done ;-) */
 167 }
 168
 169
 170
 171 /********************************************************************************************************/
 172
 173 #define ROTL32(a,n) ((a)<<(n) | (a)>>(32-(n)))
 174 #define CHANGE_ENDIAN32(x) ((x)<<24 | (x)>>24 | ((x)&0xff00)<<8 | ((x)&0xff0000)>>8 )
 175
 176 typedef uint32_t cast5_f_t(uint32_t,uint32_t,uint8_t);
 177
 178 #define IA 3
 179 #define IB 2
 180 #define IC 1
 181 #define ID 0
 182
 183
 184 uint32_t cast5_f1(uint32_t d, uint32_t m, uint8_t r){
 185         uint32_t t;
 186         t = ROTL32((d + m),r);
 187 #ifdef DEBUG
 188         uint32_t ia,ib,ic,id;
 189         uart_putstr("\r\n f1("); uart_hexdump(&d, 4); uart_putc(',');
 190                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 191                 uart_hexdump(&t, 4);
 192         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 193         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 194         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 195         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 196         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 197         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 198         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 199         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 200
 201         return (((ia ^ ib) - ic) + id);
 202
 203 #else
 204
 205         return (((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) ^ pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 206                 - pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) + pgm_read_dword(&s4[((uint8_t*)&t)[ID]]));
 207
 208 #endif
 209 }
 210
 211
 212 uint32_t cast5_f2(uint32_t d, uint32_t m, uint8_t r){
 213         uint32_t t;
 214         t = ROTL32((d ^ m),r);
 215 #ifdef DEBUG
 216         uint32_t ia,ib,ic,id;
 217         uart_putstr("\r\n f2("); uart_hexdump(&d, 4); uart_putc(',');
 218                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 219                 uart_hexdump(&t, 4);
 220
 221         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 222         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 223         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 224         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 225
 226         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 227         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 228         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 229         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 230
 231         return (((ia - ib) + ic) ^ id);
 232 #else
 233
 234         return (((pgm_read_dword(&s1[((uint8_t*)&t)[IA]])
 235                 - pgm_read_dword(&s2[((uint8_t*)&t)[IB]]) )
 236                     + pgm_read_dword(&s3[((uint8_t*)&t)[IC]]) )
 237                     ^ pgm_read_dword(&s4[((uint8_t*)&t)[ID]]) );
 238
 239 #endif
 240 }
 241
 242 uint32_t cast5_f3(uint32_t d, uint32_t m, uint8_t r){
 243         uint32_t t;
 244         t = ROTL32((m - d),r);
 245
 246 #ifdef DEBUG
 247         uint32_t ia,ib,ic,id;
 248
 249         uart_putstr("\r\n f3("); uart_hexdump(&d, 4); uart_putc(',');
 250                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 251                 uart_hexdump(&t, 4);
 252
 253         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 254         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 255         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 256         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 257
 258         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 259         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 260         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 261         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 262         return (((ia + ib) ^ ic) - id);
 263 #else
 264         return ((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) + pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 265                 ^ pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) - pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 266
 267 #endif
 268 }
 269
 270 /*************************************************************************/
 271
 272 /**
 273  * \brief encrypts a datablock with cast5
 274  * @param s Pointer to cast5 roundkeys (context)
 275  * @param block Pointer to datablock
 276  */
 277 void cast5_enc(cast5_ctx_t *s, void* block){
 278         uint32_t l,r, x, y;
 279         uint8_t i;
 280         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 281         l=((uint32_t*)block)[0];
 282         r=((uint32_t*)block)[1];
 283 //      uart_putstr("\r\n round[-1] = ");
 284 //      uart_hexdump(&r, 4);
 285         for (i=0;i<(s->shortkey?12:16);++i){
 286                 x = r;
 287                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 288                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 289                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 290                 r = l ^ CHANGE_ENDIAN32(y);
 291 //              uart_putstr("\r\n round["); DEBUG_B(i); uart_putstr("] = ");
 292 //              uart_hexdump(&r, 4);
 293                 l = x;
 294         }
 295         ((uint32_t*)block)[0]=r;
 296         ((uint32_t*)block)[1]=l;
 297 }
 298
 299 /*************************************************************************/
 300
 301 /**
 302  * \brief decrypts a datablock with cast5
 303  * @param s Pointer to cast5 roundkeys (context)
 304  * @param block Pointer to datablock
 305  */
 306 void cast5_dec(cast5_ctx_t *s, void* block){
 307         uint32_t l,r, x, y;
 308         int8_t i, rounds;
 309         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 310         l=((uint32_t*)block)[0];
 311         r=((uint32_t*)block)[1];
 312         rounds = (s->shortkey?12:16);
 313         for (i=rounds-1; i>=0 ;--i){
 314                 x = r;
 315                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 316                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 317                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 318                 r = l ^ CHANGE_ENDIAN32(y);
 319                 l = x;
 320         }
 321         ((uint32_t*)block)[0]=r;
 322         ((uint32_t*)block)[1]=l;
 323 }
 324
 325
 326 /*********************************************************************************************************/
 327 /*********************************************************************************************************/
 328 /*********************************************************************************************************/
 329
 330 #if 0
 331
 332 void cast5_old_init(cast5_ctx_t* s, uint8_t* key, uint8_t keylength){
 333          /* we migth return if the key is valid and if setup was sucessfull */
 334         uint32_t x[4], z[4], t;
 335         #define BPX ((uint8_t*)&(x[0]))
 336         #define BPZ ((uint8_t*)&(z[0]))
 337         s->shortkey = (keylength<=80);
 338         /* littel endian only! */
 339         memset(&(x[0]), 0 ,16); /* set x to zero */
 340         memcpy(&(x[0]), key, keylength/8);
 341
 342
 343         /* todo: merge a and b and compress the whole stuff */
 344         /***** A *****/
 345         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 346         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 347         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 348         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 349         /***** M *****/
 350         s->mask[0] = S_5Z(0x8) ^ S_6Z(0x9) ^ S_7Z(0x7) ^ S_8Z(0x6) ^ S_5Z(0x2);
 351         s->mask[1] = S_5Z(0xA) ^ S_6Z(0xB) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_6Z(0x6);
 352         s->mask[2] = S_5Z(0xC) ^ S_6Z(0xD) ^ S_7Z(0x3) ^ S_8Z(0x2) ^ S_7Z(0x9);
 353         s->mask[3] = S_5Z(0xE) ^ S_6Z(0xF) ^ S_7Z(0x1) ^ S_8Z(0x0) ^ S_8Z(0xC);
 354         /***** B *****/
 355         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 356         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 357         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 358         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 359         /***** N *****/
 360         s->mask[4] = S_5X(0x3) ^ S_6X(0x2) ^ S_7X(0xC) ^ S_8X(0xD) ^ S_5X(0x8);
 361         s->mask[5] = S_5X(0x1) ^ S_6X(0x0) ^ S_7X(0xE) ^ S_8X(0xF) ^ S_6X(0xD);
 362         s->mask[6] = S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x8) ^ S_8X(0x9) ^ S_7X(0x3);
 363         s->mask[7] = S_5X(0x5) ^ S_6X(0x4) ^ S_7X(0xA) ^ S_8X(0xB) ^ S_8X(0x7);
 364         /***** A *****/
 365         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 366         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 367         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 368         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 369         /***** N' *****/
 370         s->mask[8] = S_5Z(0x3) ^ S_6Z(0x2) ^ S_7Z(0xC) ^ S_8Z(0xD) ^ S_5Z(0x9);
 371         s->mask[9] = S_5Z(0x1) ^ S_6Z(0x0) ^ S_7Z(0xE) ^ S_8Z(0xF) ^ S_6Z(0xC);
 372         s->mask[10] = S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x8) ^ S_8Z(0x9) ^ S_7Z(0x2);
 373         s->mask[11] = S_5Z(0x5) ^ S_6Z(0x4) ^ S_7Z(0xA) ^ S_8Z(0xB) ^ S_8Z(0x6);
 374         /***** B *****/
 375         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 376         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 377         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 378         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 379         /***** M' *****/
 380         s->mask[12] = S_5X(0x8) ^ S_6X(0x9) ^ S_7X(0x7) ^ S_8X(0x6) ^ S_5X(0x3);
 381         s->mask[13] = S_5X(0xA) ^ S_6X(0xB) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_6X(0x7);
 382         s->mask[14] = S_5X(0xC) ^ S_6X(0xD) ^ S_7X(0x3) ^ S_8X(0x2) ^ S_7X(0x8);
 383         s->mask[15] = S_5X(0xE) ^ S_6X(0xF) ^ S_7X(0x1) ^ S_8X(0x0) ^ S_8X(0xD);
 384
 385         /* that were the masking keys, now the rotation keys */
 386         /* set the keys to zero */
 387         memset(&(s->rotl[0]),0,8);
 388         s->roth[0]=s->roth[1]=0;
 389         /***** A *****/
 390         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 391         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 392         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 393         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 394         /***** M *****/
 395         t = S_5Z(0x8) ^ S_6Z(0x9) ^ S_7Z(0x7) ^ S_8Z(0x6) ^ S_5Z(0x2);
 396         t >>= 24;
 397         s->rotl[0] |= t & 0x0f;
 398         s->roth[0] |= (t >> 4) & (1<<0);
 399         t = S_5Z(0xA) ^ S_6Z(0xB) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_6Z(0x6);
 400         t >>= 24;
 401         s->rotl[0] |= (t<<4) & 0xf0;
 402         s->roth[0] |= (t >> 3) & (1<<1);
 403         t = S_5Z(0xC) ^ S_6Z(0xD) ^ S_7Z(0x3) ^ S_8Z(0x2) ^ S_7Z(0x9);
 404         t >>= 24;
 405         s->rotl[1] |= t & 0x0f;
 406         s->roth[0] |= (t >> 2) & (1<<2);
 407         t = S_5Z(0xE) ^ S_6Z(0xF) ^ S_7Z(0x1) ^ S_8Z(0x0) ^ S_8Z(0xC);
 408         t >>= 24;
 409         s->rotl[1] |= (t<<4) & 0xf0;
 410         s->roth[0] |= (t >> 1) & (1<<3);
 411         /***** B *****/
 412         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 413         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 414         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 415         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 416         /***** N *****/
 417         t = S_5X(0x3) ^ S_6X(0x2) ^ S_7X(0xC) ^ S_8X(0xD) ^ S_5X(0x8);
 418         t >>= 24;
 419         s->rotl[2] |= t & 0x0f;
 420         s->roth[0] |= t & (1<<4);
 421         t = S_5X(0x1) ^ S_6X(0x0) ^ S_7X(0xE) ^ S_8X(0xF) ^ S_6X(0xD);
 422         t >>= 24;
 423         s->rotl[2] |= (t<<4) & 0xf0;
 424         s->roth[0] |= (t<<1) & (1<<5);
 425         t = S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x8) ^ S_8X(0x9) ^ S_7X(0x3);
 426         t >>= 24;
 427         s->rotl[3] |= t & 0x0f;
 428         s->roth[0] |= (t<<2) & (1<<6);
 429         t = S_5X(0x5) ^ S_6X(0x4) ^ S_7X(0xA) ^ S_8X(0xB) ^ S_8X(0x7);
 430         t >>= 24;
 431         s->rotl[3] |= (t<<4) & 0xf0;
 432         s->roth[0] |= (t<<3) & (1<<7);
 433         /***** A *****/
 434         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 435         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 436         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 437         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 438         /***** N' *****/
 439         t = S_5Z(0x3) ^ S_6Z(0x2) ^ S_7Z(0xC) ^ S_8Z(0xD) ^ S_5Z(0x9);
 440         t >>= 24;
 441         s->rotl[4] |= t & 0x0f;
 442         s->roth[1] |= (t>>4) & (1<<0);
 443         t = S_5Z(0x1) ^ S_6Z(0x0) ^ S_7Z(0xE) ^ S_8Z(0xF) ^ S_6Z(0xC);
 444         t >>= 24;
 445         s->rotl[4] |= (t<<4) & 0xf0;
 446         s->roth[1] |= (t>>3) & (1<<1);
 447         t = S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x8) ^ S_8Z(0x9) ^ S_7Z(0x2);
 448         t >>= 24;
 449         s->rotl[5] |= t & 0x0f;
 450         s->roth[1] |= (t>>2) & (1<<2);
 451         t = S_5Z(0x5) ^ S_6Z(0x4) ^ S_7Z(0xA) ^ S_8Z(0xB) ^ S_8Z(0x6);
 452         t >>= 24;
 453         s->rotl[5] |= (t<<4) & 0xf0;
 454         s->roth[1] |= (t>>1) & (1<<3);
 455         /***** B *****/
 456         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 457         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 458         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 459         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 460         /***** M' *****/
 461         t = S_5X(0x8) ^ S_6X(0x9) ^ S_7X(0x7) ^ S_8X(0x6) ^ S_5X(0x3);
 462         t >>= 24;
 463         s->rotl[6] |= t & 0x0f;
 464         s->roth[1] |= t & (1<<4);
 465         t = S_5X(0xA) ^ S_6X(0xB) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_6X(0x7);
 466         t >>= 24;
 467         s->rotl[6] |= (t<<4) & 0xf0;
 468         s->roth[1] |= (t<<1) & (1<<5);
 469         t = S_5X(0xC) ^ S_6X(0xD) ^ S_7X(0x3) ^ S_8X(0x2) ^ S_7X(0x8);
 470         t >>= 24;
 471         s->rotl[7] |= t & 0x0f;
 472         s->roth[1] |= (t<<2) & (1<<6);
 473         t = S_5X(0xE) ^ S_6X(0xF) ^ S_7X(0x1) ^ S_8X(0x0) ^ S_8X(0xD);
 474         t >>= 24;
 475         s->rotl[7] |= (t<<4) & 0xf0;
 476         s->roth[1] |= (t<<3) & (1<<7);
 477
 478         /* done ;-) */
 479 }
 480
 481 #endif
 482
 483
 484
 485