cast5.c

   1 /*
   2  * \file        cast5.c
   3  * \author      Daniel Otte
   4  * \date        26.07.2006
   5  * \par License:
   6  *  GPL
   7  * \brief Implementation of the CAST5 (aka CAST-128) cipher algorithm as described in RFC 2144
   8  *
   9  */
  10
  11  #include <stdint.h>
  12  #include <string.h>
  13  #include "cast5.h"
  14  #include "config.h"
  15  #include "debug.h"
  16
  17  #undef DEBUG
  18
  19  #ifdef DEBUG
  20   #include "uart.h"
  21  #endif
  22
  23 #include "cast5-sbox.h"
  24
  25
  26
  27 #define S5(x) pgm_read_dword(&s5[(x)])
  28 #define S6(x) pgm_read_dword(&s6[(x)])
  29 #define S7(x) pgm_read_dword(&s7[(x)])
  30 #define S8(x) pgm_read_dword(&s8[(x)])
  31
  32
  33 void cast5_init_A(uint8_t *dest, uint8_t *src, bool bmode){
  34         uint8_t mask = bmode?0x8:0;
  35         *((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]);
  36         *((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]);
  37         *((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]);
  38         *((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]);
  39 }
  40
  41 void cast5_init_M(uint8_t *dest, uint8_t *src, bool nmode, bool xmode){
  42         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  43         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  44         #define NMT(x) (src[nmode?nmt[(x)]:(x)])
  45         #define XMT(x) (src[xmt[(xmode<<1) + nmode][(x)]])
  46         *((uint32_t*)(&dest[0x0])) = S5(NMT(0x8)) ^ S6(NMT(0x9)) ^ S7(NMT(0x7)) ^ S8(NMT(0x6)) ^ S5(XMT(0));
  47         *((uint32_t*)(&dest[0x4])) = S5(NMT(0xA)) ^ S6(NMT(0xB)) ^ S7(NMT(0x5)) ^ S8(NMT(0x4)) ^ S6(XMT(1));
  48         *((uint32_t*)(&dest[0x8])) = S5(NMT(0xC)) ^ S6(NMT(0xD)) ^ S7(NMT(0x3)) ^ S8(NMT(0x2)) ^ S7(XMT(2));
  49         *((uint32_t*)(&dest[0xC])) = S5(NMT(0xE)) ^ S6(NMT(0xF)) ^ S7(NMT(0x1)) ^ S8(NMT(0x0)) ^ S8(XMT(3));
  50 }
  51
  52 #define S5B(x) pgm_read_byte(3+(uint8_t*)(&s5[(x)]))
  53 #define S6B(x) pgm_read_byte(3+(uint8_t*)(&s6[(x)]))
  54 #define S7B(x) pgm_read_byte(3+(uint8_t*)(&s7[(x)]))
  55 #define S8B(x) pgm_read_byte(3+(uint8_t*)(&s8[(x)]))
  56
  57 void cast5_init_rM(uint8_t *klo, uint8_t *khi, uint8_t offset, uint8_t *src, bool nmode, bool xmode){
  58         uint8_t nmt[] = {0xB, 0xA, 0x9, 0x8, 0xF, 0xE, 0xD, 0xC, 0x3, 0x2, 0x1, 0x0, 0x7, 0x6, 0x5, 0x4}; /* nmode table */
  59         uint8_t xmt[4][4] = {{0x2, 0x6, 0x9, 0xC}, {0x8, 0xD, 0x3, 0x7}, {0x3, 0x7, 0x8, 0xD}, {0x9, 0xC, 0x2, 0x6}};
  60         uint8_t t, h=0;
  61         t = S5B(NMT(0x8)) ^ S6B(NMT(0x9)) ^ S7B(NMT(0x7)) ^ S8B(NMT(0x6)) ^ S5B(XMT(0));
  62                 klo[offset*2] |= (t & 0x0f);
  63                 h |= (t&0x10); h>>=1;
  64         t = S5B(NMT(0xA)) ^ S6B(NMT(0xB)) ^ S7B(NMT(0x5)) ^ S8B(NMT(0x4)) ^ S6B(XMT(1));
  65                 klo[offset*2] |= (t<<4) & 0xf0;
  66                 h |= t&0x10; h>>=1;
  67         t = S5B(NMT(0xC)) ^ S6B(NMT(0xD)) ^ S7B(NMT(0x3)) ^ S8B(NMT(0x2)) ^ S7B(XMT(2));
  68                 klo[offset*2+1] |= t&0xf;
  69                 h |= t&0x10; h>>=1;
  70         t = S5B(NMT(0xE)) ^ S6B(NMT(0xF)) ^ S7B(NMT(0x1)) ^ S8B(NMT(0x0)) ^ S8B(XMT(3));
  71                 klo[offset*2+1] |= t<<4;
  72                 h |= t&0x10; h >>=1;
  73         #ifdef DEBUG
  74                 uart_putstr("\r\n\t h="); uart_hexdump(&h,1);
  75         #endif
  76         khi[offset>>1] |= h<<((offset&0x1)?4:0);
  77 }
  78
  79 #define S_5X(s) pgm_read_dword(&s5[BPX[(s)]])
  80 #define S_6X(s) pgm_read_dword(&s6[BPX[(s)]])
  81 #define S_7X(s) pgm_read_dword(&s7[BPX[(s)]])
  82 #define S_8X(s) pgm_read_dword(&s8[BPX[(s)]])
  83
  84 #define S_5Z(s) pgm_read_dword(&s5[BPZ[(s)]])
  85 #define S_6Z(s) pgm_read_dword(&s6[BPZ[(s)]])
  86 #define S_7Z(s) pgm_read_dword(&s7[BPZ[(s)]])
  87 #define S_8Z(s) pgm_read_dword(&s8[BPZ[(s)]])
  88
  89
  90
  91 /**
  92  * \brief sets up round keys (context) for cast5 en/decryption.
  93  * @param s Pointer to cast5 context.
  94  * @param key Pointer to binary key.
  95  * @param keylength length of keydata in bits.
  96  */
  97 void cast5_init(cast5_ctx_t* s, uint8_t* key, uint8_t keylength){
  98          /* we migth return if the key is valid and if setup was sucessfull */
  99         uint32_t x[4], z[4];
 100         #define BPX ((uint8_t*)&(x[0]))
 101         #define BPZ ((uint8_t*)&(z[0]))
 102         s->shortkey = (keylength<=80);
 103         /* littel endian only! */
 104         memset(&(x[0]), 0 ,16); /* set x to zero */
 105         if(keylength > 128)
 106                 keylength=128;
 107         memcpy(&(x[0]), key, (keylength+7)/8);
 108
 109
 110         /* todo: merge a and b and compress the whole stuff */
 111         /***** A *****/
 112         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 113         /***** M *****/
 114         cast5_init_M((uint8_t*)(&(s->mask[0])), (uint8_t*)(&z[0]), false, false);
 115         /***** B *****/
 116         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 117         /***** N *****/
 118         cast5_init_M((uint8_t*)(&(s->mask[4])), (uint8_t*)(&x[0]), true, false);
 119         /***** A *****/
 120         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 121         /***** N' *****/
 122         cast5_init_M((uint8_t*)(&(s->mask[8])), (uint8_t*)(&z[0]), true, true);
 123         /***** B *****/
 124         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 125         /***** M' *****/
 126         cast5_init_M((uint8_t*)(&(s->mask[12])), (uint8_t*)(&x[0]), false, true);
 127
 128         /* that were the masking keys, now the rotation keys */
 129         /* set the keys to zero */
 130         memset(&(s->rotl[0]),0,8);
 131         s->roth[0]=s->roth[1]=0;
 132         /***** A *****/
 133         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 134         /***** M *****/
 135         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 0, (uint8_t*)(&z[0]), false, false);
 136         /***** B *****/
 137         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 138         /***** N *****/
 139         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 1, (uint8_t*)(&x[0]), true, false);
 140         /***** A *****/
 141         cast5_init_A((uint8_t*)(&z[0]), (uint8_t*)(&x[0]), false);
 142         /***** N' *****/
 143         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 2, (uint8_t*)(&z[0]), true, true);
 144         /***** B *****/
 145         cast5_init_A((uint8_t*)(&x[0]), (uint8_t*)(&z[0]), true);
 146         /***** M' *****/
 147         cast5_init_rM(&(s->rotl[0]), &(s->roth[0]), 3, (uint8_t*)(&x[0]), false, true);
 148         /* done ;-) */
 149 }
 150
 151
 152
 153 /********************************************************************************************************/
 154
 155 #define ROTL32(a,n) ((a)<<(n) | (a)>>(32-(n)))
 156 #define CHANGE_ENDIAN32(x) ((x)<<24 | (x)>>24 | ((x)&0xff00)<<8 | ((x)&0xff0000)>>8 )
 157
 158 typedef uint32_t cast5_f_t(uint32_t,uint32_t,uint8_t);
 159
 160 #define IA 3
 161 #define IB 2
 162 #define IC 1
 163 #define ID 0
 164
 165
 166 uint32_t cast5_f1(uint32_t d, uint32_t m, uint8_t r){
 167         uint32_t t;
 168         t = ROTL32((d + m),r);
 169 #ifdef DEBUG
 170         uint32_t ia,ib,ic,id;
 171         uart_putstr("\r\n f1("); uart_hexdump(&d, 4); uart_putc(',');
 172                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 173                 uart_hexdump(&t, 4);
 174         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 175         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 176         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 177         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 178         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 179         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 180         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 181         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 182
 183         return (((ia ^ ib) - ic) + id);
 184
 185 #else
 186
 187         return (((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) ^ pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 188                 - pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) + pgm_read_dword(&s4[((uint8_t*)&t)[ID]]));
 189
 190 #endif
 191 }
 192
 193
 194 uint32_t cast5_f2(uint32_t d, uint32_t m, uint8_t r){
 195         uint32_t t;
 196         t = ROTL32((d ^ m),r);
 197 #ifdef DEBUG
 198         uint32_t ia,ib,ic,id;
 199         uart_putstr("\r\n f2("); uart_hexdump(&d, 4); uart_putc(',');
 200                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 201                 uart_hexdump(&t, 4);
 202
 203         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 204         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 205         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 206         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 207
 208         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 209         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 210         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 211         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 212
 213         return (((ia - ib) + ic) ^ id);
 214 #else
 215
 216         return (((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) - pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 217                 + pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) ^ pgm_read_dword(&s4[((uint8_t*)&t)[ID]]));
 218
 219 #endif
 220 }
 221
 222 uint32_t cast5_f3(uint32_t d, uint32_t m, uint8_t r){
 223         uint32_t t;
 224         t = ROTL32((m - d),r);
 225
 226 #ifdef DEBUG
 227         uint32_t ia,ib,ic,id;
 228
 229         uart_putstr("\r\n f3("); uart_hexdump(&d, 4); uart_putc(',');
 230                 uart_hexdump(&m , 4); uart_putc(','); uart_hexdump(&r, 1);uart_putstr("): I=");
 231                 uart_hexdump(&t, 4);
 232
 233         ia = pgm_read_dword(&s1[((uint8_t*)&t)[IA]] );
 234         ib = pgm_read_dword(&s2[((uint8_t*)&t)[IB]] );
 235         ic = pgm_read_dword(&s3[((uint8_t*)&t)[IC]] );
 236         id = pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 237
 238         uart_putstr("\r\n\tIA="); uart_hexdump(&ia, 4);
 239         uart_putstr("\r\n\tIB="); uart_hexdump(&ib, 4);
 240         uart_putstr("\r\n\tIC="); uart_hexdump(&ic, 4);
 241         uart_putstr("\r\n\tID="); uart_hexdump(&id, 4);
 242         return (((ia + ib) ^ ic) - id);
 243 #else
 244         return ((pgm_read_dword(&s1[((uint8_t*)&t)[IA]] ) + pgm_read_dword(&s2[((uint8_t*)&t)[IB]] ))
 245                 ^ pgm_read_dword(&s3[((uint8_t*)&t)[IC]] )) - pgm_read_dword(&s4[((uint8_t*)&t)[ID]] );
 246
 247 #endif
 248 }
 249
 250 /*************************************************************************/
 251
 252 /**
 253  * \brief encrypts a datablock with cast5
 254  * @param s Pointer to cast5 roundkeys (context)
 255  * @param block Pointer to datablock
 256  */
 257 void cast5_enc(cast5_ctx_t *s, void* block){
 258         uint32_t l,r, x, y;
 259         uint8_t i;
 260         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 261         l=((uint32_t*)block)[0];
 262         r=((uint32_t*)block)[1];
 263 //      uart_putstr("\r\n round[-1] = ");
 264 //      uart_hexdump(&r, 4);
 265         for (i=0;i<(s->shortkey?12:16);++i){
 266                 x = r;
 267                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 268                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 269                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 270                 r = l ^ CHANGE_ENDIAN32(y);
 271 //              uart_putstr("\r\n round["); DEBUG_B(i); uart_putstr("] = ");
 272 //              uart_hexdump(&r, 4);
 273                 l = x;
 274         }
 275         ((uint32_t*)block)[0]=r;
 276         ((uint32_t*)block)[1]=l;
 277 }
 278
 279 /*************************************************************************/
 280
 281 /**
 282  * \brief decrypts a datablock with cast5
 283  * @param s Pointer to cast5 roundkeys (context)
 284  * @param block Pointer to datablock
 285  */
 286 void cast5_dec(cast5_ctx_t *s, void* block){
 287         uint32_t l,r, x, y;
 288         int8_t i, rounds;
 289         cast5_f_t* f[]={cast5_f1,cast5_f2,cast5_f3};
 290         l=((uint32_t*)block)[0];
 291         r=((uint32_t*)block)[1];
 292         rounds = (s->shortkey?12:16);
 293         for (i=rounds-1; i>=0 ;--i){
 294                 x = r;
 295                 y = (f[i%3])(CHANGE_ENDIAN32(r), CHANGE_ENDIAN32(s->mask[i]),
 296                         (((s->roth[i>>3]) & (1<<(i&0x7)))?0x10:0x00)
 297                          + ( ((s->rotl[i>>1])>>((i&1)?4:0)) & 0x0f) );
 298                 r = l ^ CHANGE_ENDIAN32(y);
 299                 l = x;
 300         }
 301         ((uint32_t*)block)[0]=r;
 302         ((uint32_t*)block)[1]=l;
 303 }
 304
 305
 306 /*********************************************************************************************************/
 307 /*********************************************************************************************************/
 308 /*********************************************************************************************************/
 309
 310 #if 0
 311
 312 void cast5_old_init(cast5_ctx_t* s, uint8_t* key, uint8_t keylength){
 313          /* we migth return if the key is valid and if setup was sucessfull */
 314         uint32_t x[4], z[4], t;
 315         #define BPX ((uint8_t*)&(x[0]))
 316         #define BPZ ((uint8_t*)&(z[0]))
 317         s->shortkey = (keylength<=80);
 318         /* littel endian only! */
 319         memset(&(x[0]), 0 ,16); /* set x to zero */
 320         memcpy(&(x[0]), key, keylength/8);
 321
 322
 323         /* todo: merge a and b and compress the whole stuff */
 324         /***** A *****/
 325         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 326         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 327         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 328         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 329         /***** M *****/
 330         s->mask[0] = S_5Z(0x8) ^ S_6Z(0x9) ^ S_7Z(0x7) ^ S_8Z(0x6) ^ S_5Z(0x2);
 331         s->mask[1] = S_5Z(0xA) ^ S_6Z(0xB) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_6Z(0x6);
 332         s->mask[2] = S_5Z(0xC) ^ S_6Z(0xD) ^ S_7Z(0x3) ^ S_8Z(0x2) ^ S_7Z(0x9);
 333         s->mask[3] = S_5Z(0xE) ^ S_6Z(0xF) ^ S_7Z(0x1) ^ S_8Z(0x0) ^ S_8Z(0xC);
 334         /***** B *****/
 335         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 336         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 337         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 338         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 339         /***** N *****/
 340         s->mask[4] = S_5X(0x3) ^ S_6X(0x2) ^ S_7X(0xC) ^ S_8X(0xD) ^ S_5X(0x8);
 341         s->mask[5] = S_5X(0x1) ^ S_6X(0x0) ^ S_7X(0xE) ^ S_8X(0xF) ^ S_6X(0xD);
 342         s->mask[6] = S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x8) ^ S_8X(0x9) ^ S_7X(0x3);
 343         s->mask[7] = S_5X(0x5) ^ S_6X(0x4) ^ S_7X(0xA) ^ S_8X(0xB) ^ S_8X(0x7);
 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         /***** N' *****/
 350         s->mask[8] = S_5Z(0x3) ^ S_6Z(0x2) ^ S_7Z(0xC) ^ S_8Z(0xD) ^ S_5Z(0x9);
 351         s->mask[9] = S_5Z(0x1) ^ S_6Z(0x0) ^ S_7Z(0xE) ^ S_8Z(0xF) ^ S_6Z(0xC);
 352         s->mask[10] = S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x8) ^ S_8Z(0x9) ^ S_7Z(0x2);
 353         s->mask[11] = S_5Z(0x5) ^ S_6Z(0x4) ^ S_7Z(0xA) ^ S_8Z(0xB) ^ S_8Z(0x6);
 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         /***** M' *****/
 360         s->mask[12] = S_5X(0x8) ^ S_6X(0x9) ^ S_7X(0x7) ^ S_8X(0x6) ^ S_5X(0x3);
 361         s->mask[13] = S_5X(0xA) ^ S_6X(0xB) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_6X(0x7);
 362         s->mask[14] = S_5X(0xC) ^ S_6X(0xD) ^ S_7X(0x3) ^ S_8X(0x2) ^ S_7X(0x8);
 363         s->mask[15] = S_5X(0xE) ^ S_6X(0xF) ^ S_7X(0x1) ^ S_8X(0x0) ^ S_8X(0xD);
 364
 365         /* that were the masking keys, now the rotation keys */
 366         /* set the keys to zero */
 367         memset(&(s->rotl[0]),0,8);
 368         s->roth[0]=s->roth[1]=0;
 369         /***** A *****/
 370         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 371         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 372         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 373         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 374         /***** M *****/
 375         t = S_5Z(0x8) ^ S_6Z(0x9) ^ S_7Z(0x7) ^ S_8Z(0x6) ^ S_5Z(0x2);
 376         t >>= 24;
 377         s->rotl[0] |= t & 0x0f;
 378         s->roth[0] |= (t >> 4) & (1<<0);
 379         t = S_5Z(0xA) ^ S_6Z(0xB) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_6Z(0x6);
 380         t >>= 24;
 381         s->rotl[0] |= (t<<4) & 0xf0;
 382         s->roth[0] |= (t >> 3) & (1<<1);
 383         t = S_5Z(0xC) ^ S_6Z(0xD) ^ S_7Z(0x3) ^ S_8Z(0x2) ^ S_7Z(0x9);
 384         t >>= 24;
 385         s->rotl[1] |= t & 0x0f;
 386         s->roth[0] |= (t >> 2) & (1<<2);
 387         t = S_5Z(0xE) ^ S_6Z(0xF) ^ S_7Z(0x1) ^ S_8Z(0x0) ^ S_8Z(0xC);
 388         t >>= 24;
 389         s->rotl[1] |= (t<<4) & 0xf0;
 390         s->roth[0] |= (t >> 1) & (1<<3);
 391         /***** B *****/
 392         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 393         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 394         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 395         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 396         /***** N *****/
 397         t = S_5X(0x3) ^ S_6X(0x2) ^ S_7X(0xC) ^ S_8X(0xD) ^ S_5X(0x8);
 398         t >>= 24;
 399         s->rotl[2] |= t & 0x0f;
 400         s->roth[0] |= t & (1<<4);
 401         t = S_5X(0x1) ^ S_6X(0x0) ^ S_7X(0xE) ^ S_8X(0xF) ^ S_6X(0xD);
 402         t >>= 24;
 403         s->rotl[2] |= (t<<4) & 0xf0;
 404         s->roth[0] |= (t<<1) & (1<<5);
 405         t = S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x8) ^ S_8X(0x9) ^ S_7X(0x3);
 406         t >>= 24;
 407         s->rotl[3] |= t & 0x0f;
 408         s->roth[0] |= (t<<2) & (1<<6);
 409         t = S_5X(0x5) ^ S_6X(0x4) ^ S_7X(0xA) ^ S_8X(0xB) ^ S_8X(0x7);
 410         t >>= 24;
 411         s->rotl[3] |= (t<<4) & 0xf0;
 412         s->roth[0] |= (t<<3) & (1<<7);
 413         /***** A *****/
 414         z[0] = x[0] ^ S_5X(0xD) ^ S_6X(0xF) ^ S_7X(0xC) ^ S_8X(0xE) ^ S_7X(0x8);
 415         z[1] = x[2] ^ S_5Z(0x0) ^ S_6Z(0x2) ^ S_7Z(0x1) ^ S_8Z(0x3) ^ S_8X(0xA);
 416         z[2] = x[3] ^ S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x5) ^ S_8Z(0x4) ^ S_5X(0x9);
 417         z[3] = x[1] ^ S_5Z(0xA) ^ S_6Z(0x9) ^ S_7Z(0xB) ^ S_8Z(0x8) ^ S_6X(0xB);
 418         /***** N' *****/
 419         t = S_5Z(0x3) ^ S_6Z(0x2) ^ S_7Z(0xC) ^ S_8Z(0xD) ^ S_5Z(0x9);
 420         t >>= 24;
 421         s->rotl[4] |= t & 0x0f;
 422         s->roth[1] |= (t>>4) & (1<<0);
 423         t = S_5Z(0x1) ^ S_6Z(0x0) ^ S_7Z(0xE) ^ S_8Z(0xF) ^ S_6Z(0xC);
 424         t >>= 24;
 425         s->rotl[4] |= (t<<4) & 0xf0;
 426         s->roth[1] |= (t>>3) & (1<<1);
 427         t = S_5Z(0x7) ^ S_6Z(0x6) ^ S_7Z(0x8) ^ S_8Z(0x9) ^ S_7Z(0x2);
 428         t >>= 24;
 429         s->rotl[5] |= t & 0x0f;
 430         s->roth[1] |= (t>>2) & (1<<2);
 431         t = S_5Z(0x5) ^ S_6Z(0x4) ^ S_7Z(0xA) ^ S_8Z(0xB) ^ S_8Z(0x6);
 432         t >>= 24;
 433         s->rotl[5] |= (t<<4) & 0xf0;
 434         s->roth[1] |= (t>>1) & (1<<3);
 435         /***** B *****/
 436         x[0] = z[2] ^ S_5Z(0x5) ^ S_6Z(0x7) ^ S_7Z(0x4) ^ S_8Z(0x6) ^ S_7Z(0x0);
 437         x[1] = z[0] ^ S_5X(0x0) ^ S_6X(0x2) ^ S_7X(0x1) ^ S_8X(0x3) ^ S_8Z(0x2);
 438         x[2] = z[1] ^ S_5X(0x7) ^ S_6X(0x6) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_5Z(0x1);
 439         x[3] = z[3] ^ S_5X(0xA) ^ S_6X(0x9) ^ S_7X(0xB) ^ S_8X(0x8) ^ S_6Z(0x3);
 440         /***** M' *****/
 441         t = S_5X(0x8) ^ S_6X(0x9) ^ S_7X(0x7) ^ S_8X(0x6) ^ S_5X(0x3);
 442         t >>= 24;
 443         s->rotl[6] |= t & 0x0f;
 444         s->roth[1] |= t & (1<<4);
 445         t = S_5X(0xA) ^ S_6X(0xB) ^ S_7X(0x5) ^ S_8X(0x4) ^ S_6X(0x7);
 446         t >>= 24;
 447         s->rotl[6] |= (t<<4) & 0xf0;
 448         s->roth[1] |= (t<<1) & (1<<5);
 449         t = S_5X(0xC) ^ S_6X(0xD) ^ S_7X(0x3) ^ S_8X(0x2) ^ S_7X(0x8);
 450         t >>= 24;
 451         s->rotl[7] |= t & 0x0f;
 452         s->roth[1] |= (t<<2) & (1<<6);
 453         t = S_5X(0xE) ^ S_6X(0xF) ^ S_7X(0x1) ^ S_8X(0x0) ^ S_8X(0xD);
 454         t >>= 24;
 455         s->rotl[7] |= (t<<4) & 0xf0;
 456         s->roth[1] |= (t<<3) & (1<<7);
 457
 458         /* done ;-) */
 459 }
 460
 461 #endif
 462
 463
 464
 465