serpent-sboxes-bitslice.c

   1 /* serpent-sboxes.c
   2  * a non-bitsliced implementation of the serpent sboxes
   3  * author: Daniel Otte
   4  * email:  daniel.otte@rub.de
   5  * license: GPLv3
   6  */
   7
   8 #include <stdint.h>
   9 #include <string.h>
  10 #include <avr/pgmspace.h>
  11 #include "serpent-sboxes.h"
  12
  13
  14
  15
  16 /* S0:   3  8 15  1 10  6  5 11 14 13  4  2  7  0  9 12 */
  17
  18 /* depth = 5,7,4,2, Total gates=18 */
  19 static
  20 void sb0(uint8_t* out, const uint8_t* in, uint8_t i){
  21 // (a,b,c,d,w,x,y,z)
  22         uint8_t t01, t02, t03, t05, t06, t07, t08, t09, t11, t12, t13, t14, t15, t17;
  23         t01          = in[4*1+i]  ^ in[4*2+i];
  24         t02          = in[4*0+i]  | in[4*3+i];
  25         t03          = in[4*0+i]  ^ in[4*1+i];
  26         out[4*3+i]   = t02        ^ t01;
  27         t05          = in[4*2+i]  | out[4*3+i];
  28         t06          = in[4*0+i]  ^ in[4*3+i];
  29         t07          = in[4*1+i]  | in[4*2+i];
  30         t08          = in[4*3+i]  & t05;
  31         t09          = t03        & t07;
  32         out[4*2+i]   = t09        ^ t08;
  33         t11          = t09        & out[4*2+i];
  34         t12          = in[4*2+i]  ^ in[4*3+i];
  35         t13          = t07        ^ t11;
  36         t14          = in[4*1+i]  & t06;
  37         t15          = t06        ^ t13;
  38         out[4*0+i]   =     ~ t15;
  39         t17          = out[4*0+i] ^ t14;
  40         out[4*1+i]   = t12        ^ t17;
  41 }
  42
  43 /* InvS0:  13  3 11  0 10  6  5 12  1 14  4  7 15  9  8  2 */
  44
  45 /* depth = 8,4,3,6, Total gates=19 */
  46 static
  47 void sb0_inv(uint8_t* out, const uint8_t* in, uint8_t i){
  48         uint8_t t02, t03, t04, t05, t06, t08, t09, t10, t12, t13, t14, t15, t17, t18, t01;
  49         t01          = in[4*2+i] ^ in[4*3+i];
  50         t02          = in[4*0+i] | in[4*1+i];
  51         t03          = in[4*1+i] | in[4*2+i];
  52         t04          = in[4*2+i] & t01;
  53         t05          = t02       ^ t01;
  54         t06          = in[4*0+i] | t04;
  55         out[4*2+i]   =     ~ t05;
  56         t08          = in[4*1+i]   ^ in[4*3+i];
  57         t09          = t03       & t08;
  58         t10          = in[4*3+i]   | out[4*2+i];
  59         out[4*1+i]   = t09       ^ t06;
  60         t12          = in[4*0+i]   | t05;
  61         t13          = out[4*1+i]   ^ t12;
  62         t14          = t03       ^ t10;
  63         t15          = in[4*0+i]   ^ in[4*2+i];
  64         out[4*3+i]   = t14       ^ t13;
  65         t17          = t05       & t13;
  66         t18          = t14       | t17;
  67         out[4*0+i]   = t15       ^ t18;
  68 }
  69
  70 /* S1:  15 12  2  7  9  0  5 10  1 11 14  8  6 13  3  4 */
  71
  72 /* depth = 10,7,3,5, Total gates=18 */
  73 static
  74 void sb1(uint8_t* out, const uint8_t* in, uint8_t i){
  75         uint8_t t02, t03, t04, t05, t06, t07, t08, t10, t11, t12, t13, t16, t17, t01;
  76         t01          = in[4*0+i]   | in[4*3+i];
  77         t02          = in[4*2+i]   ^ in[4*3+i];
  78         t03          =     ~ in[4*1+i];
  79         t04          = in[4*0+i]   ^ in[4*2+i];
  80         t05          = in[4*0+i]   | t03;
  81         t06          = in[4*3+i]   & t04;
  82         t07          = t01       & t02;
  83         t08          = in[4*1+i]   | t06;
  84         out[4*2+i]   = t02       ^ t05;
  85         t10          = t07       ^ t08;
  86         t11          = t01       ^ t10;
  87         t12          = out[4*2+i]   ^ t11;
  88         t13          = in[4*1+i]   & in[4*3+i];
  89         out[4*3+i]   =     ~ t10;
  90         out[4*1+i]   = t13       ^ t12;
  91         t16          = t10       | out[4*1+i];
  92         t17          = t05       & t16;
  93         out[4*0+i]   = in[4*2+i]   ^ t17;
  94 }
  95
  96 /* InvS1:   5  8  2 14 15  6 12  3 11  4  7  9  1 13 10  0 */
  97
  98 /* depth = 7,4,5,3, Total gates=18 */
  99 static void sb1_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 100         uint8_t t02, t03, t04, t05, t06, t07, t08, t09, t10, t11, t14, t15, t17, t01;
 101         t01          = in[4*0+i]   ^ in[4*1+i];
 102         t02          = in[4*1+i]   | in[4*3+i];
 103         t03          = in[4*0+i]   & in[4*2+i];
 104         t04          = in[4*2+i]   ^ t02;
 105         t05          = in[4*0+i]   | t04;
 106         t06          = t01       & t05;
 107         t07          = in[4*3+i]   | t03;
 108         t08          = in[4*1+i]   ^ t06;
 109         t09          = t07       ^ t06;
 110         t10          = t04       | t03;
 111         t11          = in[4*3+i]   & t08;
 112         out[4*2+i]   =     ~ t09;
 113         out[4*1+i]   = t10       ^ t11;
 114         t14          = in[4*0+i]   | out[4*2+i];
 115         t15          = t06       ^ out[4*1+i];
 116         out[4*3+i]   = t01       ^ t04;
 117         t17          = in[4*2+i]   ^ t15;
 118         out[4*0+i]   = t14       ^ t17;
 119 }
 120
 121 /* S2:   8  6  7  9  3 12 10 15 13  1 14  4  0 11  5  2 */
 122
 123 /* depth = 3,8,11,7, Total gates=16 */
 124 static void sb2(uint8_t* out, const uint8_t* in, uint8_t i){
 125         uint8_t t02, t03, t05, t06, t07, t08, t09, t10, t12, t13, t14, t01;
 126         t01          = in[4*0+i]   | in[4*2+i];
 127         t02          = in[4*0+i]   ^ in[4*1+i];
 128         t03          = in[4*3+i]   ^ t01;
 129         out[4*0+i]   = t02       ^ t03;
 130         t05          = in[4*2+i]   ^ out[4*0+i];
 131         t06          = in[4*1+i]   ^ t05;
 132         t07          = in[4*1+i]   | t05;
 133         t08          = t01       & t06;
 134         t09          = t03       ^ t07;
 135         t10          = t02       | t09;
 136         out[4*1+i]   = t10       ^ t08;
 137         t12          = in[4*0+i]   | in[4*3+i];
 138         t13          = t09       ^ out[4*1+i];
 139         t14          = in[4*1+i]   ^ t13;
 140         out[4*3+i]   =     ~ t09;
 141         out[4*2+i]   = t12       ^ t14;
 142 }
 143
 144 /* InvS2:  12  9 15  4 11 14  1  2  0  3  6 13  5  8 10  7 */
 145
 146 /* depth = 3,6,8,3, Total gates=18 */
 147 static void sb2_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 148         uint8_t t02, t03, t04, t06, t07, t08, t09, t10, t11, t12, t15, t16, t17, t01;
 149         t01          = in[4*0+i]   ^ in[4*3+i];
 150         t02          = in[4*2+i]   ^ in[4*3+i];
 151         t03          = in[4*0+i]   & in[4*2+i];
 152         t04          = in[4*1+i]   | t02;
 153         out[4*0+i]   = t01       ^ t04;
 154         t06          = in[4*0+i]   | in[4*2+i];
 155         t07          = in[4*3+i]   | out[4*0+i];
 156         t08          =     ~ in[4*3+i];
 157         t09          = in[4*1+i]   & t06;
 158         t10          = t08       | t03;
 159         t11          = in[4*1+i]   & t07;
 160         t12          = t06       & t02;
 161         out[4*3+i]   = t09       ^ t10;
 162         out[4*1+i]   = t12       ^ t11;
 163         t15          = in[4*2+i]   & out[4*3+i];
 164         t16          = out[4*0+i]   ^ out[4*1+i];
 165         t17          = t10       ^ t15;
 166         out[4*2+i]   = t16       ^ t17;
 167 }
 168
 169 /* S3:   0 15 11  8 12  9  6  3 13  1  2  4 10  7  5 14 */
 170
 171 /* depth = 8,3,5,5, Total gates=18 */
 172 static void sb3(uint8_t* out, const uint8_t* in, uint8_t i){
 173         uint8_t t02, t03, t04, t05, t06, t07, t08, t09, t10, t11, t13, t14, t15, t01;
 174         t01          = in[4*0+i]   ^ in[4*2+i];
 175         t02          = in[4*0+i]   | in[4*3+i];
 176         t03          = in[4*0+i]   & in[4*3+i];
 177         t04          = t01       & t02;
 178         t05          = in[4*1+i]   | t03;
 179         t06          = in[4*0+i]   & in[4*1+i];
 180         t07          = in[4*3+i]   ^ t04;
 181         t08          = in[4*2+i]   | t06;
 182         t09          = in[4*1+i]   ^ t07;
 183         t10          = in[4*3+i]   & t05;
 184         t11          = t02       ^ t10;
 185         out[4*3+i]   = t08       ^ t09;
 186         t13          = in[4*3+i]   | out[4*3+i];
 187         t14          = in[4*0+i]   | t07;
 188         t15          = in[4*1+i]   & t13;
 189         out[4*2+i]   = t08       ^ t11;
 190         out[4*0+i]   = t14       ^ t15;
 191         out[4*1+i]   = t05       ^ t04;
 192 }
 193
 194 /* InvS3:   0  9 10  7 11 14  6 13  3  5 12  2  4  8 15  1 */
 195
 196 /* depth = 3,6,4,4, Total gates=17 */
 197 static void sb3_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 198         uint8_t t02, t03, t04, t05, t06, t07, t09, t11, t12, t13, t14, t16, t01;
 199         t01          = in[4*2+i]   | in[4*3+i];
 200         t02          = in[4*0+i]   | in[4*3+i];
 201         t03          = in[4*2+i]   ^ t02;
 202         t04          = in[4*1+i]   ^ t02;
 203         t05          = in[4*0+i]   ^ in[4*3+i];
 204         t06          = t04       & t03;
 205         t07          = in[4*1+i]   & t01;
 206         out[4*2+i]   = t05       ^ t06;
 207         t09          = in[4*0+i]   ^ t03;
 208         out[4*0+i]   = t07       ^ t03;
 209         t11          = out[4*0+i]   | t05;
 210         t12          = t09       & t11;
 211         t13          = in[4*0+i]   & out[4*2+i];
 212         t14          = t01       ^ t05;
 213         out[4*1+i]   = in[4*1+i]   ^ t12;
 214         t16          = in[4*1+i]   | t13;
 215         out[4*3+i]   = t14       ^ t16;
 216 }
 217
 218 /* S4:   1 15  8  3 12  0 11  6  2  5  4 10  9 14  7 13 */
 219
 220 /* depth = 6,7,5,3, Total gates=19 */
 221 static void sb4(uint8_t* out, const uint8_t* in, uint8_t i){
 222         uint8_t t02, t03, t04, t05, t06, t08, t09, t10, t11, t12, t13, t14, t15, t16, t01;
 223         t01          = in[4*0+i]   | in[4*1+i];
 224         t02          = in[4*1+i]   | in[4*2+i];
 225         t03          = in[4*0+i]   ^ t02;
 226         t04          = in[4*1+i]   ^ in[4*3+i];
 227         t05          = in[4*3+i]   | t03;
 228         t06          = in[4*3+i]   & t01;
 229         out[4*3+i]   = t03       ^ t06;
 230         t08          = out[4*3+i]   & t04;
 231         t09          = t04       & t05;
 232         t10          = in[4*2+i]   ^ t06;
 233         t11          = in[4*1+i]   & in[4*2+i];
 234         t12          = t04       ^ t08;
 235         t13          = t11       | t03;
 236         t14          = t10       ^ t09;
 237         t15          = in[4*0+i]   & t05;
 238         t16          = t11       | t12;
 239         out[4*2+i]   = t13       ^ t08;
 240         out[4*1+i]   = t15       ^ t16;
 241         out[4*0+i]   =     ~ t14;
 242 }
 243
 244 /* InvS4:   5  0  8  3 10  9  7 14  2 12 11  6  4 15 13  1 */
 245
 246 /* depth = 6,4,7,3, Total gates=17 */
 247 static void sb4_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 248         uint8_t t02, t03, t04, t05, t06, t07, t09, t10, t11, t12, t13, t15, t01;
 249         t01          = in[4*1+i]   | in[4*3+i];
 250         t02          = in[4*2+i]   | in[4*3+i];
 251         t03          = in[4*0+i]   & t01;
 252         t04          = in[4*1+i]   ^ t02;
 253         t05          = in[4*2+i]   ^ in[4*3+i];
 254         t06          =     ~ t03;
 255         t07          = in[4*0+i]   & t04;
 256         out[4*1+i]   = t05       ^ t07;
 257         t09          = out[4*1+i]   | t06;
 258         t10          = in[4*0+i]   ^ t07;
 259         t11          = t01       ^ t09;
 260         t12          = in[4*3+i]   ^ t04;
 261         t13          = in[4*2+i]   | t10;
 262         out[4*3+i]   = t03       ^ t12;
 263         t15          = in[4*0+i]   ^ t04;
 264         out[4*2+i]   = t11       ^ t13;
 265         out[4*0+i]   = t15       ^ t09;
 266 }
 267
 268 /* S5:  15  5  2 11  4 10  9 12  0  3 14  8 13  6  7  1 */
 269
 270 /* depth = 4,6,8,6, Total gates=17 */
 271 static void sb5(uint8_t* out, const uint8_t* in, uint8_t i){
 272         uint8_t t02, t03, t04, t05, t07, t08, t09, t10, t11, t12, t13, t14, t01;
 273         t01          = in[4*1+i]   ^ in[4*3+i];
 274         t02          = in[4*1+i]   | in[4*3+i];
 275         t03          = in[4*0+i]   & t01;
 276         t04          = in[4*2+i]   ^ t02;
 277         t05          = t03       ^ t04;
 278         out[4*0+i]   =     ~ t05;
 279         t07          = in[4*0+i]   ^ t01;
 280         t08          = in[4*3+i]   | out[4*0+i];
 281         t09          = in[4*1+i]   | t05;
 282         t10          = in[4*3+i]   ^ t08;
 283         t11          = in[4*1+i]   | t07;
 284         t12          = t03       | out[4*0+i];
 285         t13          = t07       | t10;
 286         t14          = t01       ^ t11;
 287         out[4*2+i]   = t09       ^ t13;
 288         out[4*1+i]   = t07       ^ t08;
 289         out[4*3+i]   = t12       ^ t14;
 290 }
 291
 292 /* InvS5:   8 15  2  9  4  1 13 14 11  6  5  3  7 12 10  0 */
 293
 294 /* depth = 4,6,9,7, Total gates=17 */
 295 static void sb5_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 296         uint8_t t02, t03, t04, t05, t07, t08, t09, t10, t12, t13, t15, t16, t01;
 297         t01          = in[4*0+i]   & in[4*3+i];
 298         t02          = in[4*2+i]   ^ t01;
 299         t03          = in[4*0+i]   ^ in[4*3+i];
 300         t04          = in[4*1+i]   & t02;
 301         t05          = in[4*0+i]   & in[4*2+i];
 302         out[4*0+i]   = t03       ^ t04;
 303         t07          = in[4*0+i]   & out[4*0+i];
 304         t08          = t01       ^ out[4*0+i];
 305         t09          = in[4*1+i]   | t05;
 306         t10          =     ~ in[4*1+i];
 307         out[4*1+i]   = t08       ^ t09;
 308         t12          = t10       | t07;
 309         t13          = out[4*0+i]   | out[4*1+i];
 310         out[4*3+i]   = t02       ^ t12;
 311         t15          = t02       ^ t13;
 312         t16          = in[4*1+i]   ^ in[4*3+i];
 313         out[4*2+i]   = t16       ^ t15;
 314 }
 315
 316 /* S6:   7  2 12  5  8  4  6 11 14  9  1 15 13  3 10  0 */
 317
 318 /* depth = 8,3,6,3, Total gates=19 */
 319 static void sb6(uint8_t* out, const uint8_t* in, uint8_t i){
 320         uint8_t t02, t03, t04, t05, t07, t08, t09, t10, t11, t12, t13, t15, t17, t18, t01;
 321         t01          = in[4*0+i]   & in[4*3+i];
 322         t02          = in[4*1+i]   ^ in[4*2+i];
 323         t03          = in[4*0+i]   ^ in[4*3+i];
 324         t04          = t01       ^ t02;
 325         t05          = in[4*1+i]   | in[4*2+i];
 326         out[4*1+i]   =     ~ t04;
 327         t07          = t03       & t05;
 328         t08          = in[4*1+i]   & out[4*1+i];
 329         t09          = in[4*0+i]   | in[4*2+i];
 330         t10          = t07       ^ t08;
 331         t11          = in[4*1+i]   | in[4*3+i];
 332         t12          = in[4*2+i]   ^ t11;
 333         t13          = t09       ^ t10;
 334         out[4*2+i]   =     ~ t13;
 335         t15          = out[4*1+i]   & t03;
 336         out[4*3+i]   = t12       ^ t07;
 337         t17          = in[4*0+i]   ^ in[4*1+i];
 338         t18          = out[4*2+i]   ^ t15;
 339         out[4*0+i]   = t17       ^ t18;
 340 }
 341
 342 /* InvS6:  15 10  1 13  5  3  6  0  4  9 14  7  2 12  8 11 */
 343
 344 /* depth = 5,3,8,6, Total gates=19 */
 345 static void sb6_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 346         uint8_t t02, t03, t04, t05, t06, t07, t08, t09, t12, t13, t14, t15, t16, t17, t01;
 347         t01          = in[4*0+i]   ^ in[4*2+i];
 348         t02          =     ~ in[4*2+i];
 349         t03          = in[4*1+i]   & t01;
 350         t04          = in[4*1+i]   | t02;
 351         t05          = in[4*3+i]   | t03;
 352         t06          = in[4*1+i]   ^ in[4*3+i];
 353         t07          = in[4*0+i]   & t04;
 354         t08          = in[4*0+i]   | t02;
 355         t09          = t07       ^ t05;
 356         out[4*1+i]   = t06       ^ t08;
 357         out[4*0+i]   =     ~ t09;
 358         t12          = in[4*1+i]   & out[4*0+i];
 359         t13          = t01       & t05;
 360         t14          = t01       ^ t12;
 361         t15          = t07       ^ t13;
 362         t16          = in[4*3+i]   | t02;
 363         t17          = in[4*0+i]   ^ out[4*1+i];
 364         out[4*3+i]   = t17       ^ t15;
 365         out[4*2+i]   = t16       ^ t14;
 366 }
 367
 368 /* S7:   1 13 15  0 14  8  2 11  7  4 12 10  9  3  5  6 */
 369
 370 /* depth = 10,7,10,4, Total gates=19 */
 371 static void sb7(uint8_t* out, const uint8_t* in, uint8_t i){
 372         uint8_t t02, t03, t04, t05, t06, t08, t09, t10, t11, t13, t14, t15, t16, t17, t01;
 373         t01          = in[4*0+i]   & in[4*2+i];
 374         t02          =     ~ in[4*3+i];
 375         t03          = in[4*0+i]   & t02;
 376         t04          = in[4*1+i]   | t01;
 377         t05          = in[4*0+i]   & in[4*1+i];
 378         t06          = in[4*2+i]   ^ t04;
 379         out[4*3+i]   = t03       ^ t06;
 380         t08          = in[4*2+i]   | out[4*3+i];
 381         t09          = in[4*3+i]   | t05;
 382         t10          = in[4*0+i]   ^ t08;
 383         t11          = t04       & out[4*3+i];
 384         out[4*1+i]   = t09       ^ t10;
 385         t13          = in[4*1+i]   ^ out[4*1+i];
 386         t14          = t01       ^ out[4*1+i];
 387         t15          = in[4*2+i]   ^ t05;
 388         t16          = t11       | t13;
 389         t17          = t02       | t14;
 390         out[4*0+i]   = t15       ^ t17;
 391         out[4*2+i]   = in[4*0+i]   ^ t16;
 392 }
 393
 394 /* InvS7:   3  0  6 13  9 14 15  8  5 12 11  7 10  1  4  2 */
 395
 396 /* depth = 9,7,3,3, Total gates=18 */
 397 static void sb7_inv(uint8_t* out, const uint8_t* in, uint8_t i){
 398         uint8_t t02, t03, t04, t06, t07, t08, t09, t10, t11, t13, t14, t15, t16, t01;
 399         t01          = in[4*0+i]   & in[4*1+i];
 400         t02          = in[4*0+i]   | in[4*1+i];
 401         t03          = in[4*2+i]   | t01;
 402         t04          = in[4*3+i]   & t02;
 403         out[4*3+i]   = t03       ^ t04;
 404         t06          = in[4*1+i]   ^ t04;
 405         t07          = in[4*3+i]   ^ out[4*3+i];
 406         t08          =     ~ t07;
 407         t09          = t06       | t08;
 408         t10          = in[4*1+i]   ^ in[4*3+i];
 409         t11          = in[4*0+i]   | in[4*3+i];
 410         out[4*1+i]   = in[4*0+i]   ^ t09;
 411         t13          = in[4*2+i]   ^ t06;
 412         t14          = in[4*2+i]   & t11;
 413         t15          = in[4*3+i]   | out[4*1+i];
 414         t16          = t01       | t10;
 415         out[4*0+i]   = t13       ^ t15;
 416         out[4*2+i]   = t14       ^ t16;
 417 }
 418
 419 typedef void(*sb_fpt)(uint8_t*, const uint8_t*, uint8_t i);
 420
 421 sb_fpt sf_tab[] PROGMEM = {
 422         sb0, sb1, sb2, sb3,
 423         sb4, sb5, sb6, sb7
 424 };
 425
 426 sb_fpt sinvf_tab[] PROGMEM = {
 427         sb0_inv, sb1_inv, sb2_inv, sb3_inv,
 428         sb4_inv, sb5_inv, sb6_inv, sb7_inv
 429 };
 430
 431 void sbox128(void * w, uint8_t box){
 432         uint8_t i, buffer[16];
 433         box &= 0x7;
 434
 435         sb_fpt fp;
 436         fp = (sb_fpt)pgm_read_word(&(sf_tab[box]));
 437         for(i=0; i<4; ++i){
 438                 fp(buffer, (uint8_t*)w, i);
 439         }
 440         memcpy(w, buffer, 16);
 441 }
 442
 443 void inv_sbox128(void * w, uint8_t box){
 444         uint8_t i, buffer[16];
 445         box &= 0x7;
 446
 447         sb_fpt fp;
 448         fp = (sb_fpt)pgm_read_word(&(sinvf_tab[box]));
 449         for(i=0; i<4; ++i){
 450                 fp(buffer, (uint8_t*)w, i);
 451         }
 452         memcpy(w, buffer, 16);
 453 }
 454
 455
 456
 457
 458
 459
 460
 461