src/rp_kernel_on_cpu.c

   1 /**
   2  * Author......: Jens Steube <jens.steube@gmail.com>
   3  * License.....: MIT
   4  */
   5
   6 #pragma GCC diagnostic ignored "-Wunused-parameter"
   7 #pragma GCC diagnostic ignored "-Wunused-function"
   8
   9 #include <rp_kernel_on_cpu.h>
  10
  11 static u32 generate_cmask (u32 buf)
  12 {
  13   const u32 rmask =  ((buf & 0x40404040) >> 1)
  14                   & ~((buf & 0x80808080) >> 2);
  15
  16   const u32 hmask = (buf & 0x1f1f1f1f) + 0x05050505;
  17   const u32 lmask = (buf & 0x1f1f1f1f) + 0x1f1f1f1f;
  18
  19   return rmask & ~hmask & lmask;
  20 }
  21
  22 static void truncate_right (u32 w0[4], u32 w1[4], const u32 len)
  23 {
  24   const u32 tmp = (1 << ((len % 4) * 8)) - 1;
  25
  26   switch (len / 4)
  27   {
  28     case  0:  w0[0] &= tmp;
  29               w0[1]  = 0;
  30               w0[2]  = 0;
  31               w0[3]  = 0;
  32               w1[0]  = 0;
  33               w1[1]  = 0;
  34               w1[2]  = 0;
  35               w1[3]  = 0;
  36               break;
  37     case  1:  w0[1] &= tmp;
  38               w0[2]  = 0;
  39               w0[3]  = 0;
  40               w1[0]  = 0;
  41               w1[1]  = 0;
  42               w1[2]  = 0;
  43               w1[3]  = 0;
  44               break;
  45     case  2:  w0[2] &= tmp;
  46               w0[3]  = 0;
  47               w1[0]  = 0;
  48               w1[1]  = 0;
  49               w1[2]  = 0;
  50               w1[3]  = 0;
  51               break;
  52     case  3:  w0[3] &= tmp;
  53               w1[0]  = 0;
  54               w1[1]  = 0;
  55               w1[2]  = 0;
  56               w1[3]  = 0;
  57               break;
  58     case  4:  w1[0] &= tmp;
  59               w1[1]  = 0;
  60               w1[2]  = 0;
  61               w1[3]  = 0;
  62               break;
  63     case  5:  w1[1] &= tmp;
  64               w1[2]  = 0;
  65               w1[3]  = 0;
  66               break;
  67     case  6:  w1[2] &= tmp;
  68               w1[3]  = 0;
  69               break;
  70     case  7:  w1[3] &= tmp;
  71               break;
  72   }
  73 }
  74
  75 static void truncate_left (u32 w0[4], u32 w1[4], const u32 len)
  76 {
  77   const u32 tmp = ~((1 << ((len % 4) * 8)) - 1);
  78
  79   switch (len / 4)
  80   {
  81     case  0:  w0[0] &= tmp;
  82               break;
  83     case  1:  w0[0]  = 0;
  84               w0[1] &= tmp;
  85               break;
  86     case  2:  w0[0]  = 0;
  87               w0[1]  = 0;
  88               w0[2] &= tmp;
  89               break;
  90     case  3:  w0[0]  = 0;
  91               w0[1]  = 0;
  92               w0[2]  = 0;
  93               w0[3] &= tmp;
  94               break;
  95     case  4:  w0[0]  = 0;
  96               w0[1]  = 0;
  97               w0[2]  = 0;
  98               w0[3]  = 0;
  99               w1[0] &= tmp;
 100               break;
 101     case  5:  w0[0]  = 0;
 102               w0[1]  = 0;
 103               w0[2]  = 0;
 104               w0[3]  = 0;
 105               w1[0]  = 0;
 106               w1[1] &= tmp;
 107               break;
 108     case  6:  w0[0]  = 0;
 109               w0[1]  = 0;
 110               w0[2]  = 0;
 111               w0[3]  = 0;
 112               w1[0]  = 0;
 113               w1[1]  = 0;
 114               w1[2] &= tmp;
 115               break;
 116     case  7:  w0[0]  = 0;
 117               w0[1]  = 0;
 118               w0[2]  = 0;
 119               w0[3]  = 0;
 120               w1[0]  = 0;
 121               w1[1]  = 0;
 122               w1[2]  = 0;
 123               w1[3] &= tmp;
 124               break;
 125   }
 126 }
 127
 128 static void lshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
 129 {
 130   out0[0] = in0[0] >>  8 | in0[1] << 24;
 131   out0[1] = in0[1] >>  8 | in0[2] << 24;
 132   out0[2] = in0[2] >>  8 | in0[3] << 24;
 133   out0[3] = in0[3] >>  8 | in1[0] << 24;
 134   out1[0] = in1[0] >>  8 | in1[1] << 24;
 135   out1[1] = in1[1] >>  8 | in1[2] << 24;
 136   out1[2] = in1[2] >>  8 | in1[3] << 24;
 137   out1[3] = in1[3] >>  8;
 138 }
 139
 140 static void rshift_block (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4])
 141 {
 142   out1[3] = in1[3] <<  8 | in1[2] >> 24;
 143   out1[2] = in1[2] <<  8 | in1[1] >> 24;
 144   out1[1] = in1[1] <<  8 | in1[0] >> 24;
 145   out1[0] = in1[0] <<  8 | in0[3] >> 24;
 146   out0[3] = in0[3] <<  8 | in0[2] >> 24;
 147   out0[2] = in0[2] <<  8 | in0[1] >> 24;
 148   out0[1] = in0[1] <<  8 | in0[0] >> 24;
 149   out0[0] = in0[0] <<  8;
 150 }
 151
 152 static void rshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4], const u32 num)
 153 {
 154   switch (num)
 155   {
 156     case  0:  out1[3] = in1[3];
 157               out1[2] = in1[2];
 158               out1[1] = in1[1];
 159               out1[0] = in1[0];
 160               out0[3] = in0[3];
 161               out0[2] = in0[2];
 162               out0[1] = in0[1];
 163               out0[0] = in0[0];
 164               break;
 165     case  1:  out1[3] = in1[3] <<  8 | in1[2] >> 24;
 166               out1[2] = in1[2] <<  8 | in1[1] >> 24;
 167               out1[1] = in1[1] <<  8 | in1[0] >> 24;
 168               out1[0] = in1[0] <<  8 | in0[3] >> 24;
 169               out0[3] = in0[3] <<  8 | in0[2] >> 24;
 170               out0[2] = in0[2] <<  8 | in0[1] >> 24;
 171               out0[1] = in0[1] <<  8 | in0[0] >> 24;
 172               out0[0] = in0[0] <<  8;
 173               break;
 174     case  2:  out1[3] = in1[3] << 16 | in1[2] >> 16;
 175               out1[2] = in1[2] << 16 | in1[1] >> 16;
 176               out1[1] = in1[1] << 16 | in1[0] >> 16;
 177               out1[0] = in1[0] << 16 | in0[3] >> 16;
 178               out0[3] = in0[3] << 16 | in0[2] >> 16;
 179               out0[2] = in0[2] << 16 | in0[1] >> 16;
 180               out0[1] = in0[1] << 16 | in0[0] >> 16;
 181               out0[0] = in0[0] << 16;
 182               break;
 183     case  3:  out1[3] = in1[3] << 24 | in1[2] >>  8;
 184               out1[2] = in1[2] << 24 | in1[1] >>  8;
 185               out1[1] = in1[1] << 24 | in1[0] >>  8;
 186               out1[0] = in1[0] << 24 | in0[3] >>  8;
 187               out0[3] = in0[3] << 24 | in0[2] >>  8;
 188               out0[2] = in0[2] << 24 | in0[1] >>  8;
 189               out0[1] = in0[1] << 24 | in0[0] >>  8;
 190               out0[0] = in0[0] << 24;
 191               break;
 192     case  4:  out1[3] = in1[2];
 193               out1[2] = in1[1];
 194               out1[1] = in1[0];
 195               out1[0] = in0[3];
 196               out0[3] = in0[2];
 197               out0[2] = in0[1];
 198               out0[1] = in0[0];
 199               out0[0] = 0;
 200               break;
 201     case  5:  out1[3] = in1[2] <<  8 | in1[1] >> 24;
 202               out1[2] = in1[1] <<  8 | in1[0] >> 24;
 203               out1[1] = in1[0] <<  8 | in0[3] >> 24;
 204               out1[0] = in0[3] <<  8 | in0[2] >> 24;
 205               out0[3] = in0[2] <<  8 | in0[1] >> 24;
 206               out0[2] = in0[1] <<  8 | in0[0] >> 24;
 207               out0[1] = in0[0] <<  8;
 208               out0[0] = 0;
 209               break;
 210     case  6:  out1[3] = in1[2] << 16 | in1[1] >> 16;
 211               out1[2] = in1[1] << 16 | in1[0] >> 16;
 212               out1[1] = in1[0] << 16 | in0[3] >> 16;
 213               out1[0] = in0[3] << 16 | in0[2] >> 16;
 214               out0[3] = in0[2] << 16 | in0[1] >> 16;
 215               out0[2] = in0[1] << 16 | in0[0] >> 16;
 216               out0[1] = in0[0] << 16;
 217               out0[0] = 0;
 218               break;
 219     case  7:  out1[3] = in1[2] << 24 | in1[1] >>  8;
 220               out1[2] = in1[1] << 24 | in1[0] >>  8;
 221               out1[1] = in1[0] << 24 | in0[3] >>  8;
 222               out1[0] = in0[3] << 24 | in0[2] >>  8;
 223               out0[3] = in0[2] << 24 | in0[1] >>  8;
 224               out0[2] = in0[1] << 24 | in0[0] >>  8;
 225               out0[1] = in0[0] << 24;
 226               out0[0] = 0;
 227               break;
 228     case  8:  out1[3] = in1[1];
 229               out1[2] = in1[0];
 230               out1[1] = in0[3];
 231               out1[0] = in0[2];
 232               out0[3] = in0[1];
 233               out0[2] = in0[0];
 234               out0[1] = 0;
 235               out0[0] = 0;
 236               break;
 237     case  9:  out1[3] = in1[1] <<  8 | in1[0] >> 24;
 238               out1[2] = in1[0] <<  8 | in0[3] >> 24;
 239               out1[1] = in0[3] <<  8 | in0[2] >> 24;
 240               out1[0] = in0[2] <<  8 | in0[1] >> 24;
 241               out0[3] = in0[1] <<  8 | in0[0] >> 24;
 242               out0[2] = in0[0] <<  8;
 243               out0[1] = 0;
 244               out0[0] = 0;
 245               break;
 246     case 10:  out1[3] = in1[1] << 16 | in1[0] >> 16;
 247               out1[2] = in1[0] << 16 | in0[3] >> 16;
 248               out1[1] = in0[3] << 16 | in0[2] >> 16;
 249               out1[0] = in0[2] << 16 | in0[1] >> 16;
 250               out0[3] = in0[1] << 16 | in0[0] >> 16;
 251               out0[2] = in0[0] << 16;
 252               out0[1] = 0;
 253               out0[0] = 0;
 254               break;
 255     case 11:  out1[3] = in1[1] << 24 | in1[0] >>  8;
 256               out1[2] = in1[0] << 24 | in0[3] >>  8;
 257               out1[1] = in0[3] << 24 | in0[2] >>  8;
 258               out1[0] = in0[2] << 24 | in0[1] >>  8;
 259               out0[3] = in0[1] << 24 | in0[0] >>  8;
 260               out0[2] = in0[0] << 24;
 261               out0[1] = 0;
 262               out0[0] = 0;
 263               break;
 264     case 12:  out1[3] = in1[0];
 265               out1[2] = in0[3];
 266               out1[1] = in0[2];
 267               out1[0] = in0[1];
 268               out0[3] = in0[0];
 269               out0[2] = 0;
 270               out0[1] = 0;
 271               out0[0] = 0;
 272               break;
 273     case 13:  out1[3] = in1[0] <<  8 | in0[3] >> 24;
 274               out1[2] = in0[3] <<  8 | in0[2] >> 24;
 275               out1[1] = in0[2] <<  8 | in0[1] >> 24;
 276               out1[0] = in0[1] <<  8 | in0[0] >> 24;
 277               out0[3] = in0[0] <<  8;
 278               out0[2] = 0;
 279               out0[1] = 0;
 280               out0[0] = 0;
 281               break;
 282     case 14:  out1[3] = in1[0] << 16 | in0[3] >> 16;
 283               out1[2] = in0[3] << 16 | in0[2] >> 16;
 284               out1[1] = in0[2] << 16 | in0[1] >> 16;
 285               out1[0] = in0[1] << 16 | in0[0] >> 16;
 286               out0[3] = in0[0] << 16;
 287               out0[2] = 0;
 288               out0[1] = 0;
 289               out0[0] = 0;
 290               break;
 291     case 15:  out1[3] = in1[0] << 24 | in0[3] >>  8;
 292               out1[2] = in0[3] << 24 | in0[2] >>  8;
 293               out1[1] = in0[2] << 24 | in0[1] >>  8;
 294               out1[0] = in0[1] << 24 | in0[0] >>  8;
 295               out0[3] = in0[0] << 24;
 296               out0[2] = 0;
 297               out0[1] = 0;
 298               out0[0] = 0;
 299               break;
 300     case 16:  out1[3] = in0[3];
 301               out1[2] = in0[2];
 302               out1[1] = in0[1];
 303               out1[0] = in0[0];
 304               out0[3] = 0;
 305               out0[2] = 0;
 306               out0[1] = 0;
 307               out0[0] = 0;
 308               break;
 309     case 17:  out1[3] = in0[3] <<  8 | in0[2] >> 24;
 310               out1[2] = in0[2] <<  8 | in0[1] >> 24;
 311               out1[1] = in0[1] <<  8 | in0[0] >> 24;
 312               out1[0] = in0[0] <<  8;
 313               out0[3] = 0;
 314               out0[2] = 0;
 315               out0[1] = 0;
 316               out0[0] = 0;
 317               break;
 318     case 18:  out1[3] = in0[3] << 16 | in0[2] >> 16;
 319               out1[2] = in0[2] << 16 | in0[1] >> 16;
 320               out1[1] = in0[1] << 16 | in0[0] >> 16;
 321               out1[0] = in0[0] << 16;
 322               out0[3] = 0;
 323               out0[2] = 0;
 324               out0[1] = 0;
 325               out0[0] = 0;
 326               break;
 327     case 19:  out1[3] = in0[3] << 24 | in0[2] >>  8;
 328               out1[2] = in0[2] << 24 | in0[1] >>  8;
 329               out1[1] = in0[1] << 24 | in0[0] >>  8;
 330               out1[0] = in0[0] << 24;
 331               out0[3] = 0;
 332               out0[2] = 0;
 333               out0[1] = 0;
 334               out0[0] = 0;
 335               break;
 336     case 20:  out1[3] = in0[2];
 337               out1[2] = in0[1];
 338               out1[1] = in0[0];
 339               out1[0] = 0;
 340               out0[3] = 0;
 341               out0[2] = 0;
 342               out0[1] = 0;
 343               out0[0] = 0;
 344               break;
 345     case 21:  out1[3] = in0[2] <<  8 | in0[1] >> 24;
 346               out1[2] = in0[1] <<  8 | in0[0] >> 24;
 347               out1[1] = in0[0] <<  8;
 348               out1[0] = 0;
 349               out0[3] = 0;
 350               out0[2] = 0;
 351               out0[1] = 0;
 352               out0[0] = 0;
 353               break;
 354     case 22:  out1[3] = in0[2] << 16 | in0[1] >> 16;
 355               out1[2] = in0[1] << 16 | in0[0] >> 16;
 356               out1[1] = in0[0] << 16;
 357               out1[0] = 0;
 358               out0[3] = 0;
 359               out0[2] = 0;
 360               out0[1] = 0;
 361               out0[0] = 0;
 362               break;
 363     case 23:  out1[3] = in0[2] << 24 | in0[1] >>  8;
 364               out1[2] = in0[1] << 24 | in0[0] >>  8;
 365               out1[1] = in0[0] << 24;
 366               out1[0] = 0;
 367               out0[3] = 0;
 368               out0[2] = 0;
 369               out0[1] = 0;
 370               out0[0] = 0;
 371               break;
 372     case 24:  out1[3] = in0[1];
 373               out1[2] = in0[0];
 374               out1[1] = 0;
 375               out1[0] = 0;
 376               out0[3] = 0;
 377               out0[2] = 0;
 378               out0[1] = 0;
 379               out0[0] = 0;
 380               break;
 381     case 25:  out1[3] = in0[1] <<  8 | in0[0] >> 24;
 382               out1[2] = in0[0] <<  8;
 383               out1[1] = 0;
 384               out1[0] = 0;
 385               out0[3] = 0;
 386               out0[2] = 0;
 387               out0[1] = 0;
 388               out0[0] = 0;
 389               break;
 390     case 26:  out1[3] = in0[1] << 16 | in0[0] >> 16;
 391               out1[2] = in0[0] << 16;
 392               out1[1] = 0;
 393               out1[0] = 0;
 394               out0[3] = 0;
 395               out0[2] = 0;
 396               out0[1] = 0;
 397               out0[0] = 0;
 398               break;
 399     case 27:  out1[3] = in0[1] << 24 | in0[0] >>  8;
 400               out1[2] = in0[0] << 24;
 401               out1[1] = 0;
 402               out1[0] = 0;
 403               out0[3] = 0;
 404               out0[2] = 0;
 405               out0[1] = 0;
 406               out0[0] = 0;
 407               break;
 408     case 28:  out1[3] = in0[0];
 409               out1[2] = 0;
 410               out1[1] = 0;
 411               out1[0] = 0;
 412               out0[3] = 0;
 413               out0[2] = 0;
 414               out0[1] = 0;
 415               out0[0] = 0;
 416               break;
 417     case 29:  out1[3] = in0[0] <<  8;
 418               out1[2] = 0;
 419               out1[1] = 0;
 420               out1[0] = 0;
 421               out0[3] = 0;
 422               out0[2] = 0;
 423               out0[1] = 0;
 424               out0[0] = 0;
 425               break;
 426     case 30:  out1[3] = in0[0] << 16;
 427               out1[2] = 0;
 428               out1[1] = 0;
 429               out1[0] = 0;
 430               out0[3] = 0;
 431               out0[2] = 0;
 432               out0[1] = 0;
 433               out0[0] = 0;
 434               break;
 435     case 31:  out1[3] = in0[0] << 24;
 436               out1[2] = 0;
 437               out1[1] = 0;
 438               out1[0] = 0;
 439               out0[3] = 0;
 440               out0[2] = 0;
 441               out0[1] = 0;
 442               out0[0] = 0;
 443               break;
 444   }
 445 }
 446
 447 static void lshift_block_N (const u32 in0[4], const u32 in1[4], u32 out0[4], u32 out1[4], const u32 num)
 448 {
 449   switch (num)
 450   {
 451     case  0:  out0[0] = in0[0];
 452               out0[1] = in0[1];
 453               out0[2] = in0[2];
 454               out0[3] = in0[3];
 455               out1[0] = in1[0];
 456               out1[1] = in1[1];
 457               out1[2] = in1[2];
 458               out1[3] = in1[3];
 459               break;
 460     case  1:  out0[0] = in0[0] >>  8 | in0[1] << 24;
 461               out0[1] = in0[1] >>  8 | in0[2] << 24;
 462               out0[2] = in0[2] >>  8 | in0[3] << 24;
 463               out0[3] = in0[3] >>  8 | in1[0] << 24;
 464               out1[0] = in1[0] >>  8 | in1[1] << 24;
 465               out1[1] = in1[1] >>  8 | in1[2] << 24;
 466               out1[2] = in1[2] >>  8 | in1[3] << 24;
 467               out1[3] = in1[3] >>  8;
 468               break;
 469     case  2:  out0[0] = in0[0] >> 16 | in0[1] << 16;
 470               out0[1] = in0[1] >> 16 | in0[2] << 16;
 471               out0[2] = in0[2] >> 16 | in0[3] << 16;
 472               out0[3] = in0[3] >> 16 | in1[0] << 16;
 473               out1[0] = in1[0] >> 16 | in1[1] << 16;
 474               out1[1] = in1[1] >> 16 | in1[2] << 16;
 475               out1[2] = in1[2] >> 16 | in1[3] << 16;
 476               out1[3] = in1[3] >> 16;
 477               break;
 478     case  3:  out0[0] = in0[0] >> 24 | in0[1] <<  8;
 479               out0[1] = in0[1] >> 24 | in0[2] <<  8;
 480               out0[2] = in0[2] >> 24 | in0[3] <<  8;
 481               out0[3] = in0[3] >> 24 | in1[0] <<  8;
 482               out1[0] = in1[0] >> 24 | in1[1] <<  8;
 483               out1[1] = in1[1] >> 24 | in1[2] <<  8;
 484               out1[2] = in1[2] >> 24 | in1[3] <<  8;
 485               out1[3] = in1[3] >> 24;
 486               break;
 487     case  4:  out0[0] = in0[1];
 488               out0[1] = in0[2];
 489               out0[2] = in0[3];
 490               out0[3] = in1[0];
 491               out1[0] = in1[1];
 492               out1[1] = in1[2];
 493               out1[2] = in1[3];
 494               out1[3] = 0;
 495               break;
 496     case  5:  out0[0] = in0[1] >>  8 | in0[2] << 24;
 497               out0[1] = in0[2] >>  8 | in0[3] << 24;
 498               out0[2] = in0[3] >>  8 | in1[0] << 24;
 499               out0[3] = in1[0] >>  8 | in1[1] << 24;
 500               out1[0] = in1[1] >>  8 | in1[2] << 24;
 501               out1[1] = in1[2] >>  8 | in1[3] << 24;
 502               out1[2] = in1[3] >>  8;
 503               out1[3] = 0;
 504               break;
 505     case  6:  out0[0] = in0[1] >> 16 | in0[2] << 16;
 506               out0[1] = in0[2] >> 16 | in0[3] << 16;
 507               out0[2] = in0[3] >> 16 | in1[0] << 16;
 508               out0[3] = in1[0] >> 16 | in1[1] << 16;
 509               out1[0] = in1[1] >> 16 | in1[2] << 16;
 510               out1[1] = in1[2] >> 16 | in1[3] << 16;
 511               out1[2] = in1[3] >> 16;
 512               out1[3] = 0;
 513               break;
 514     case  7:  out0[0] = in0[1] >> 24 | in0[2] <<  8;
 515               out0[1] = in0[2] >> 24 | in0[3] <<  8;
 516               out0[2] = in0[3] >> 24 | in1[0] <<  8;
 517               out0[3] = in1[0] >> 24 | in1[1] <<  8;
 518               out1[0] = in1[1] >> 24 | in1[2] <<  8;
 519               out1[1] = in1[2] >> 24 | in1[3] <<  8;
 520               out1[2] = in1[3] >> 24;
 521               out1[3] = 0;
 522               break;
 523     case  8:  out0[0] = in0[2];
 524               out0[1] = in0[3];
 525               out0[2] = in1[0];
 526               out0[3] = in1[1];
 527               out1[0] = in1[2];
 528               out1[1] = in1[3];
 529               out1[2] = 0;
 530               out1[3] = 0;
 531               break;
 532     case  9:  out0[0] = in0[2] >>  8 | in0[3] << 24;
 533               out0[1] = in0[3] >>  8 | in1[0] << 24;
 534               out0[2] = in1[0] >>  8 | in1[1] << 24;
 535               out0[3] = in1[1] >>  8 | in1[2] << 24;
 536               out1[0] = in1[2] >>  8 | in1[3] << 24;
 537               out1[1] = in1[3] >>  8;
 538               out1[2] = 0;
 539               out1[3] = 0;
 540               break;
 541     case 10:  out0[0] = in0[2] >> 16 | in0[3] << 16;
 542               out0[1] = in0[3] >> 16 | in1[0] << 16;
 543               out0[2] = in1[0] >> 16 | in1[1] << 16;
 544               out0[3] = in1[1] >> 16 | in1[2] << 16;
 545               out1[0] = in1[2] >> 16 | in1[3] << 16;
 546               out1[1] = in1[3] >> 16;
 547               out1[2] = 0;
 548               out1[3] = 0;
 549               break;
 550     case 11:  out0[0] = in0[2] >> 24 | in0[3] <<  8;
 551               out0[1] = in0[3] >> 24 | in1[0] <<  8;
 552               out0[2] = in1[0] >> 24 | in1[1] <<  8;
 553               out0[3] = in1[1] >> 24 | in1[2] <<  8;
 554               out1[0] = in1[2] >> 24 | in1[3] <<  8;
 555               out1[1] = in1[3] >> 24;
 556               out1[2] = 0;
 557               out1[3] = 0;
 558               break;
 559     case 12:  out0[0] = in0[3];
 560               out0[1] = in1[0];
 561               out0[2] = in1[1];
 562               out0[3] = in1[2];
 563               out1[0] = in1[3];
 564               out1[1] = 0;
 565               out1[2] = 0;
 566               out1[3] = 0;
 567               break;
 568     case 13:
 569               out0[0] = in0[3] >>  8 | in1[0] << 24;
 570               out0[1] = in1[0] >>  8 | in1[1] << 24;
 571               out0[2] = in1[1] >>  8 | in1[2] << 24;
 572               out0[3] = in1[2] >>  8 | in1[3] << 24;
 573               out1[0] = in1[3] >>  8;
 574               out1[1] = 0;
 575               out1[2] = 0;
 576               out1[3] = 0;
 577               break;
 578     case 14:  out0[0] = in0[3] >> 16 | in1[0] << 16;
 579               out0[1] = in1[0] >> 16 | in1[1] << 16;
 580               out0[2] = in1[1] >> 16 | in1[2] << 16;
 581               out0[3] = in1[2] >> 16 | in1[3] << 16;
 582               out1[0] = in1[3] >> 16;
 583               out1[1] = 0;
 584               out1[2] = 0;
 585               out1[3] = 0;
 586               break;
 587     case 15:  out0[0] = in0[3] >> 24 | in1[0] <<  8;
 588               out0[1] = in1[0] >> 24 | in1[1] <<  8;
 589               out0[2] = in1[1] >> 24 | in1[2] <<  8;
 590               out0[3] = in1[2] >> 24 | in1[3] <<  8;
 591               out1[0] = in1[3] >> 24;
 592               out1[1] = 0;
 593               out1[2] = 0;
 594               out1[3] = 0;
 595               break;
 596     case 16:  out0[0] = in1[0];
 597               out0[1] = in1[1];
 598               out0[2] = in1[2];
 599               out0[3] = in1[3];
 600               out1[0] = 0;
 601               out1[1] = 0;
 602               out1[2] = 0;
 603               out1[3] = 0;
 604               break;
 605     case 17:  out0[0] = in1[0] >>  8 | in1[1] << 24;
 606               out0[1] = in1[1] >>  8 | in1[2] << 24;
 607               out0[2] = in1[2] >>  8 | in1[3] << 24;
 608               out0[3] = in1[3] >>  8;
 609               out1[0] = 0;
 610               out1[1] = 0;
 611               out1[2] = 0;
 612               out1[3] = 0;
 613               break;
 614     case 18:  out0[0] = in1[0] >> 16 | in1[1] << 16;
 615               out0[1] = in1[1] >> 16 | in1[2] << 16;
 616               out0[2] = in1[2] >> 16 | in1[3] << 16;
 617               out0[3] = in1[3] >> 16;
 618               out1[0] = 0;
 619               out1[1] = 0;
 620               out1[2] = 0;
 621               out1[3] = 0;
 622               break;
 623     case 19:  out0[0] = in1[0] >> 24 | in1[1] <<  8;
 624               out0[1] = in1[1] >> 24 | in1[2] <<  8;
 625               out0[2] = in1[2] >> 24 | in1[3] <<  8;
 626               out0[3] = in1[3] >> 24;
 627               out1[0] = 0;
 628               out1[1] = 0;
 629               out1[2] = 0;
 630               out1[3] = 0;
 631               break;
 632     case 20:  out0[0] = in1[1];
 633               out0[1] = in1[2];
 634               out0[2] = in1[3];
 635               out0[3] = 0;
 636               out1[0] = 0;
 637               out1[1] = 0;
 638               out1[2] = 0;
 639               out1[3] = 0;
 640               break;
 641     case 21:  out0[0] = in1[1] >>  8 | in1[2] << 24;
 642               out0[1] = in1[2] >>  8 | in1[3] << 24;
 643               out0[2] = in1[3] >>  8;
 644               out0[3] = 0;
 645               out1[0] = 0;
 646               out1[1] = 0;
 647               out1[2] = 0;
 648               out1[3] = 0;
 649               break;
 650     case 22:  out0[0] = in1[1] >> 16 | in1[2] << 16;
 651               out0[1] = in1[2] >> 16 | in1[3] << 16;
 652               out0[2] = in1[3] >> 16;
 653               out0[3] = 0;
 654               out1[0] = 0;
 655               out1[1] = 0;
 656               out1[2] = 0;
 657               out1[3] = 0;
 658               break;
 659     case 23:  out0[0] = in1[1] >> 24 | in1[2] <<  8;
 660               out0[1] = in1[2] >> 24 | in1[3] <<  8;
 661               out0[2] = in1[3] >> 24;
 662               out0[3] = 0;
 663               out1[0] = 0;
 664               out1[1] = 0;
 665               out1[2] = 0;
 666               out1[3] = 0;
 667               break;
 668     case 24:  out0[0] = in1[2];
 669               out0[1] = in1[3];
 670               out0[2] = 0;
 671               out0[3] = 0;
 672               out1[0] = 0;
 673               out1[1] = 0;
 674               out1[2] = 0;
 675               out1[3] = 0;
 676               break;
 677     case 25:  out0[0] = in1[2] >>  8 | in1[3] << 24;
 678               out0[1] = in1[3] >>  8;
 679               out0[2] = 0;
 680               out0[3] = 0;
 681               out1[0] = 0;
 682               out1[1] = 0;
 683               out1[2] = 0;
 684               out1[3] = 0;
 685               break;
 686     case 26:  out0[0] = in1[2] >> 16 | in1[3] << 16;
 687               out0[1] = in1[3] >> 16;
 688               out0[2] = 0;
 689               out0[3] = 0;
 690               out1[0] = 0;
 691               out1[1] = 0;
 692               out1[2] = 0;
 693               out1[3] = 0;
 694               break;
 695     case 27:  out0[0] = in1[2] >> 24 | in1[3] <<  8;
 696               out0[1] = in1[3] >> 24;
 697               out0[2] = 0;
 698               out0[3] = 0;
 699               out1[0] = 0;
 700               out1[1] = 0;
 701               out1[2] = 0;
 702               out1[3] = 0;
 703               break;
 704     case 28:  out0[0] = in1[3];
 705               out0[1] = 0;
 706               out0[2] = 0;
 707               out0[3] = 0;
 708               out1[0] = 0;
 709               out1[1] = 0;
 710               out1[2] = 0;
 711               out1[3] = 0;
 712               break;
 713     case 29:  out0[0] = in1[3] >>  8;
 714               out0[1] = 0;
 715               out0[2] = 0;
 716               out0[3] = 0;
 717               out1[0] = 0;
 718               out1[1] = 0;
 719               out1[2] = 0;
 720               out1[3] = 0;
 721               break;
 722     case 30:  out0[0] = in1[3] >> 16;
 723               out0[1] = 0;
 724               out0[2] = 0;
 725               out0[3] = 0;
 726               out1[0] = 0;
 727               out1[1] = 0;
 728               out1[2] = 0;
 729               out1[3] = 0;
 730               break;
 731     case 31:  out0[0] = in1[3] >> 24;
 732               out0[1] = 0;
 733               out0[2] = 0;
 734               out0[3] = 0;
 735               out1[0] = 0;
 736               out1[1] = 0;
 737               out1[2] = 0;
 738               out1[3] = 0;
 739               break;
 740   }
 741 }
 742
 743 static void append_block1 (const u32 offset, u32 dst0[4], u32 dst1[4], const u32 src_r0)
 744 {
 745   u32 tmp[2] = { 0 };
 746
 747   switch (offset & 3)
 748   {
 749     case  0:  tmp[0] = src_r0;
 750               tmp[1] = 0;
 751               break;
 752     case  1:  tmp[0] = src_r0 <<  8;
 753               tmp[1] = src_r0 >> 24;
 754               break;
 755     case  2:  tmp[0] = src_r0 << 16;
 756               tmp[1] = src_r0 >> 16;
 757               break;
 758     case  3:  tmp[0] = src_r0 << 24;
 759               tmp[1] = src_r0 >>  8;
 760               break;
 761   }
 762
 763   switch (offset / 4)
 764   {
 765     case  0:  dst0[0] |= tmp[0];
 766               dst0[1]  = tmp[1];
 767               break;
 768     case  1:  dst0[1] |= tmp[0];
 769               dst0[2]  = tmp[1];
 770               break;
 771     case  2:  dst0[2] |= tmp[0];
 772               dst0[3]  = tmp[1];
 773               break;
 774     case  3:  dst0[3] |= tmp[0];
 775               dst1[0]  = tmp[1];
 776               break;
 777     case  4:  dst1[0] |= tmp[0];
 778               dst1[1]  = tmp[1];
 779               break;
 780     case  5:  dst1[1] |= tmp[0];
 781               dst1[2]  = tmp[1];
 782               break;
 783     case  6:  dst1[2] |= tmp[0];
 784               dst1[3]  = tmp[1];
 785               break;
 786     case  7:  dst1[3] |= tmp[0];
 787               break;
 788   }
 789 }
 790
 791 static void append_block8 (const u32 offset, u32 dst0[4], u32 dst1[4], const u32 src_l0[4], const u32 src_l1[4], const u32 src_r0[4], const u32 src_r1[4])
 792 {
 793   switch (offset)
 794   {
 795     case 0:
 796       dst0[0] = src_r0[0];
 797       dst0[1] = src_r0[1];
 798       dst0[2] = src_r0[2];
 799       dst0[3] = src_r0[3];
 800       dst1[0] = src_r1[0];
 801       dst1[1] = src_r1[1];
 802       dst1[2] = src_r1[2];
 803       dst1[3] = src_r1[3];
 804       break;
 805
 806     case 1:
 807       dst0[0] = src_l0[0]       | src_r0[0] <<  8;
 808       dst0[1] = src_r0[0] >> 24 | src_r0[1] <<  8;
 809       dst0[2] = src_r0[1] >> 24 | src_r0[2] <<  8;
 810       dst0[3] = src_r0[2] >> 24 | src_r0[3] <<  8;
 811       dst1[0] = src_r0[3] >> 24 | src_r1[0] <<  8;
 812       dst1[1] = src_r1[0] >> 24 | src_r1[1] <<  8;
 813       dst1[2] = src_r1[1] >> 24 | src_r1[2] <<  8;
 814       dst1[3] = src_r1[2] >> 24 | src_r1[3] <<  8;
 815       break;
 816
 817     case 2:
 818       dst0[0] = src_l0[0]       | src_r0[0] << 16;
 819       dst0[1] = src_r0[0] >> 16 | src_r0[1] << 16;
 820       dst0[2] = src_r0[1] >> 16 | src_r0[2] << 16;
 821       dst0[3] = src_r0[2] >> 16 | src_r0[3] << 16;
 822       dst1[0] = src_r0[3] >> 16 | src_r1[0] << 16;
 823       dst1[1] = src_r1[0] >> 16 | src_r1[1] << 16;
 824       dst1[2] = src_r1[1] >> 16 | src_r1[2] << 16;
 825       dst1[3] = src_r1[2] >> 16 | src_r1[3] << 16;
 826       break;
 827
 828     case 3:
 829       dst0[0] = src_l0[0]       | src_r0[0] << 24;
 830       dst0[1] = src_r0[0] >>  8 | src_r0[1] << 24;
 831       dst0[2] = src_r0[1] >>  8 | src_r0[2] << 24;
 832       dst0[3] = src_r0[2] >>  8 | src_r0[3] << 24;
 833       dst1[0] = src_r0[3] >>  8 | src_r1[0] << 24;
 834       dst1[1] = src_r1[0] >>  8 | src_r1[1] << 24;
 835       dst1[2] = src_r1[1] >>  8 | src_r1[2] << 24;
 836       dst1[3] = src_r1[2] >>  8 | src_r1[3] << 24;
 837       break;
 838
 839     case 4:
 840       dst0[1] = src_r0[0];
 841       dst0[2] = src_r0[1];
 842       dst0[3] = src_r0[2];
 843       dst1[0] = src_r0[3];
 844       dst1[1] = src_r1[0];
 845       dst1[2] = src_r1[1];
 846       dst1[3] = src_r1[2];
 847       break;
 848
 849     case 5:
 850       dst0[1] = src_l0[1]       | src_r0[0] <<  8;
 851       dst0[2] = src_r0[0] >> 24 | src_r0[1] <<  8;
 852       dst0[3] = src_r0[1] >> 24 | src_r0[2] <<  8;
 853       dst1[0] = src_r0[2] >> 24 | src_r0[3] <<  8;
 854       dst1[1] = src_r0[3] >> 24 | src_r1[0] <<  8;
 855       dst1[2] = src_r1[0] >> 24 | src_r1[1] <<  8;
 856       dst1[3] = src_r1[1] >> 24 | src_r1[2] <<  8;
 857       break;
 858
 859     case 6:
 860       dst0[1] = src_l0[1]       | src_r0[0] << 16;
 861       dst0[2] = src_r0[0] >> 16 | src_r0[1] << 16;
 862       dst0[3] = src_r0[1] >> 16 | src_r0[2] << 16;
 863       dst1[0] = src_r0[2] >> 16 | src_r0[3] << 16;
 864       dst1[1] = src_r0[3] >> 16 | src_r1[0] << 16;
 865       dst1[2] = src_r1[0] >> 16 | src_r1[1] << 16;
 866       dst1[3] = src_r1[1] >> 16 | src_r1[2] << 16;
 867       break;
 868
 869     case 7:
 870       dst0[1] = src_l0[1]       | src_r0[0] << 24;
 871       dst0[2] = src_r0[0] >>  8 | src_r0[1] << 24;
 872       dst0[3] = src_r0[1] >>  8 | src_r0[2] << 24;
 873       dst1[0] = src_r0[2] >>  8 | src_r0[3] << 24;
 874       dst1[1] = src_r0[3] >>  8 | src_r1[0] << 24;
 875       dst1[2] = src_r1[0] >>  8 | src_r1[1] << 24;
 876       dst1[3] = src_r1[1] >>  8 | src_r1[2] << 24;
 877       break;
 878
 879     case 8:
 880       dst0[2] = src_r0[0];
 881       dst0[3] = src_r0[1];
 882       dst1[0] = src_r0[2];
 883       dst1[1] = src_r0[3];
 884       dst1[2] = src_r1[0];
 885       dst1[3] = src_r1[1];
 886       break;
 887
 888     case 9:
 889       dst0[2] = src_l0[2]       | src_r0[0] <<  8;
 890       dst0[3] = src_r0[0] >> 24 | src_r0[1] <<  8;
 891       dst1[0] = src_r0[1] >> 24 | src_r0[2] <<  8;
 892       dst1[1] = src_r0[2] >> 24 | src_r0[3] <<  8;
 893       dst1[2] = src_r0[3] >> 24 | src_r1[0] <<  8;
 894       dst1[3] = src_r1[0] >> 24 | src_r1[1] <<  8;
 895       break;
 896
 897     case 10:
 898       dst0[2] = src_l0[2]       | src_r0[0] << 16;
 899       dst0[3] = src_r0[0] >> 16 | src_r0[1] << 16;
 900       dst1[0] = src_r0[1] >> 16 | src_r0[2] << 16;
 901       dst1[1] = src_r0[2] >> 16 | src_r0[3] << 16;
 902       dst1[2] = src_r0[3] >> 16 | src_r1[0] << 16;
 903       dst1[3] = src_r1[0] >> 16 | src_r1[1] << 16;
 904       break;
 905
 906     case 11:
 907       dst0[2] = src_l0[2]       | src_r0[0] << 24;
 908       dst0[3] = src_r0[0] >>  8 | src_r0[1] << 24;
 909       dst1[0] = src_r0[1] >>  8 | src_r0[2] << 24;
 910       dst1[1] = src_r0[2] >>  8 | src_r0[3] << 24;
 911       dst1[2] = src_r0[3] >>  8 | src_r1[0] << 24;
 912       dst1[3] = src_r1[0] >>  8 | src_r1[1] << 24;
 913       break;
 914
 915     case 12:
 916       dst0[3] = src_r0[0];
 917       dst1[0] = src_r0[1];
 918       dst1[1] = src_r0[2];
 919       dst1[2] = src_r0[3];
 920       dst1[3] = src_r1[0];
 921       break;
 922
 923     case 13:
 924       dst0[3] = src_l0[3]       | src_r0[0] <<  8;
 925       dst1[0] = src_r0[0] >> 24 | src_r0[1] <<  8;
 926       dst1[1] = src_r0[1] >> 24 | src_r0[2] <<  8;
 927       dst1[2] = src_r0[2] >> 24 | src_r0[3] <<  8;
 928       dst1[3] = src_r0[3] >> 24 | src_r1[0] <<  8;
 929       break;
 930
 931     case 14:
 932       dst0[3] = src_l0[3]       | src_r0[0] << 16;
 933       dst1[0] = src_r0[0] >> 16 | src_r0[1] << 16;
 934       dst1[1] = src_r0[1] >> 16 | src_r0[2] << 16;
 935       dst1[2] = src_r0[2] >> 16 | src_r0[3] << 16;
 936       dst1[3] = src_r0[3] >> 16 | src_r1[0] << 16;
 937       break;
 938
 939     case 15:
 940       dst0[3] = src_l0[3]       | src_r0[0] << 24;
 941       dst1[0] = src_r0[0] >>  8 | src_r0[1] << 24;
 942       dst1[1] = src_r0[1] >>  8 | src_r0[2] << 24;
 943       dst1[2] = src_r0[2] >>  8 | src_r0[3] << 24;
 944       dst1[3] = src_r0[3] >>  8 | src_r1[0] << 24;
 945       break;
 946
 947     case 16:
 948       dst1[0] = src_r0[0];
 949       dst1[1] = src_r0[1];
 950       dst1[2] = src_r0[2];
 951       dst1[3] = src_r0[3];
 952       break;
 953
 954     case 17:
 955       dst1[0] = src_l1[0]       | src_r0[0] <<  8;
 956       dst1[1] = src_r0[0] >> 24 | src_r0[1] <<  8;
 957       dst1[2] = src_r0[1] >> 24 | src_r0[2] <<  8;
 958       dst1[3] = src_r0[2] >> 24 | src_r0[3] <<  8;
 959       break;
 960
 961     case 18:
 962       dst1[0] = src_l1[0]       | src_r0[0] << 16;
 963       dst1[1] = src_r0[0] >> 16 | src_r0[1] << 16;
 964       dst1[2] = src_r0[1] >> 16 | src_r0[2] << 16;
 965       dst1[3] = src_r0[2] >> 16 | src_r0[3] << 16;
 966       break;
 967
 968     case 19:
 969       dst1[0] = src_l1[0]       | src_r0[0] << 24;
 970       dst1[1] = src_r0[0] >>  8 | src_r0[1] << 24;
 971       dst1[2] = src_r0[1] >>  8 | src_r0[2] << 24;
 972       dst1[3] = src_r0[2] >>  8 | src_r0[3] << 24;
 973       break;
 974
 975     case 20:
 976       dst1[1] = src_r0[0];
 977       dst1[2] = src_r0[1];
 978       dst1[3] = src_r0[2];
 979       break;
 980
 981     case 21:
 982       dst1[1] = src_l1[1]       | src_r0[0] <<  8;
 983       dst1[2] = src_r0[0] >> 24 | src_r0[1] <<  8;
 984       dst1[3] = src_r0[1] >> 24 | src_r0[2] <<  8;
 985       break;
 986
 987     case 22:
 988       dst1[1] = src_l1[1]       | src_r0[0] << 16;
 989       dst1[2] = src_r0[0] >> 16 | src_r0[1] << 16;
 990       dst1[3] = src_r0[1] >> 16 | src_r0[2] << 16;
 991       break;
 992
 993     case 23:
 994       dst1[1] = src_l1[1]       | src_r0[0] << 24;
 995       dst1[2] = src_r0[0] >>  8 | src_r0[1] << 24;
 996       dst1[3] = src_r0[1] >>  8 | src_r0[2] << 24;
 997       break;
 998
 999     case 24:
1000       dst1[2] = src_r0[0];
1001       dst1[3] = src_r0[1];
1002       break;
1003
1004     case 25:
1005       dst1[2] = src_l1[2]       | src_r0[0] <<  8;
1006       dst1[3] = src_r0[0] >> 24 | src_r0[1] <<  8;
1007       break;
1008
1009     case 26:
1010       dst1[2] = src_l1[2]       | src_r0[0] << 16;
1011       dst1[3] = src_r0[0] >> 16 | src_r0[1] << 16;
1012       break;
1013
1014     case 27:
1015       dst1[2] = src_l1[2]       | src_r0[0] << 24;
1016       dst1[3] = src_r0[0] >>  8 | src_r0[1] << 24;
1017       break;
1018
1019     case 28:
1020       dst1[3] = src_r0[0];
1021       break;
1022
1023     case 29:
1024       dst1[3] = src_l1[3]       | src_r0[0] <<  8;
1025       break;
1026
1027     case 30:
1028       dst1[3] = src_l1[3]       | src_r0[0] << 16;
1029       break;
1030
1031     case 31:
1032       dst1[3] = src_l1[3]       | src_r0[0] << 24;
1033       break;
1034   }
1035 }
1036
1037 static void reverse_block (u32 in0[4], u32 in1[4], u32 out0[4], u32 out1[4], const u32 len)
1038 {
1039   rshift_block_N (in0, in1, out0, out1, 32 - len);
1040
1041   u32 tib40[4];
1042   u32 tib41[4];
1043
1044   tib40[0] = out1[3];
1045   tib40[1] = out1[2];
1046   tib40[2] = out1[1];
1047   tib40[3] = out1[0];
1048   tib41[0] = out0[3];
1049   tib41[1] = out0[2];
1050   tib41[2] = out0[1];
1051   tib41[3] = out0[0];
1052
1053   out0[0] = swap_workaround (tib40[0]);
1054   out0[1] = swap_workaround (tib40[1]);
1055   out0[2] = swap_workaround (tib40[2]);
1056   out0[3] = swap_workaround (tib40[3]);
1057   out1[0] = swap_workaround (tib41[0]);
1058   out1[1] = swap_workaround (tib41[1]);
1059   out1[2] = swap_workaround (tib41[2]);
1060   out1[3] = swap_workaround (tib41[3]);
1061 }
1062
1063 static u32 rule_op_mangle_lrest (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1064 {
1065   buf0[0] |= (generate_cmask (buf0[0]));
1066   buf0[1] |= (generate_cmask (buf0[1]));
1067   buf0[2] |= (generate_cmask (buf0[2]));
1068   buf0[3] |= (generate_cmask (buf0[3]));
1069   buf1[0] |= (generate_cmask (buf1[0]));
1070   buf1[1] |= (generate_cmask (buf1[1]));
1071   buf1[2] |= (generate_cmask (buf1[2]));
1072   buf1[3] |= (generate_cmask (buf1[3]));
1073
1074   return in_len;
1075 }
1076
1077 static u32 rule_op_mangle_urest (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1078 {
1079   buf0[0] &= ~(generate_cmask (buf0[0]));
1080   buf0[1] &= ~(generate_cmask (buf0[1]));
1081   buf0[2] &= ~(generate_cmask (buf0[2]));
1082   buf0[3] &= ~(generate_cmask (buf0[3]));
1083   buf1[0] &= ~(generate_cmask (buf1[0]));
1084   buf1[1] &= ~(generate_cmask (buf1[1]));
1085   buf1[2] &= ~(generate_cmask (buf1[2]));
1086   buf1[3] &= ~(generate_cmask (buf1[3]));
1087
1088   return in_len;
1089 }
1090
1091 static u32 rule_op_mangle_lrest_ufirst (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1092 {
1093   rule_op_mangle_lrest (p0, p1, buf0, buf1, in_len);
1094
1095   buf0[0] &= ~(0x00000020 & generate_cmask (buf0[0]));
1096
1097   return in_len;
1098 }
1099
1100 static u32 rule_op_mangle_urest_lfirst (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1101 {
1102   rule_op_mangle_urest (p0, p1, buf0, buf1, in_len);
1103
1104   buf0[0] |= (0x00000020 & generate_cmask (buf0[0]));
1105
1106   return in_len;
1107 }
1108
1109 static u32 rule_op_mangle_trest (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1110 {
1111   buf0[0] ^= (generate_cmask (buf0[0]));
1112   buf0[1] ^= (generate_cmask (buf0[1]));
1113   buf0[2] ^= (generate_cmask (buf0[2]));
1114   buf0[3] ^= (generate_cmask (buf0[3]));
1115   buf1[0] ^= (generate_cmask (buf1[0]));
1116   buf1[1] ^= (generate_cmask (buf1[1]));
1117   buf1[2] ^= (generate_cmask (buf1[2]));
1118   buf1[3] ^= (generate_cmask (buf1[3]));
1119
1120   return in_len;
1121 }
1122
1123 static u32 rule_op_mangle_toggle_at (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1124 {
1125   if (p0 >= in_len) return (in_len);
1126
1127   const u32 tmp = 0x20u << ((p0 & 3) * 8);
1128
1129   switch (p0 / 4)
1130   {
1131     case  0:  buf0[0] ^= (tmp & generate_cmask (buf0[0])); break;
1132     case  1:  buf0[1] ^= (tmp & generate_cmask (buf0[1])); break;
1133     case  2:  buf0[2] ^= (tmp & generate_cmask (buf0[2])); break;
1134     case  3:  buf0[3] ^= (tmp & generate_cmask (buf0[3])); break;
1135     case  4:  buf1[0] ^= (tmp & generate_cmask (buf1[0])); break;
1136     case  5:  buf1[1] ^= (tmp & generate_cmask (buf1[1])); break;
1137     case  6:  buf1[2] ^= (tmp & generate_cmask (buf1[2])); break;
1138     case  7:  buf1[3] ^= (tmp & generate_cmask (buf1[3])); break;
1139   }
1140
1141   return in_len;
1142 }
1143
1144 static u32 rule_op_mangle_reverse (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1145 {
1146   reverse_block (buf0, buf1, buf0, buf1, in_len);
1147
1148   return in_len;
1149 }
1150
1151 static u32 rule_op_mangle_dupeword (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1152 {
1153   if ((in_len + in_len) >= 32) return (in_len);
1154
1155   u32 out_len = in_len;
1156
1157   u32 tib40[4];
1158   u32 tib41[4];
1159
1160   tib40[0] = buf0[0];
1161   tib40[1] = buf0[1];
1162   tib40[2] = buf0[2];
1163   tib40[3] = buf0[3];
1164   tib41[0] = buf1[0];
1165   tib41[1] = buf1[1];
1166   tib41[2] = buf1[2];
1167   tib41[3] = buf1[3];
1168
1169   append_block8 (out_len, buf0, buf1, buf0, buf1, tib40, tib41);
1170
1171   out_len += in_len;
1172
1173   return out_len;
1174 }
1175
1176 static u32 rule_op_mangle_dupeword_times (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1177 {
1178   if (((in_len * p0) + in_len) >= 32) return (in_len);
1179
1180   u32 out_len = in_len;
1181
1182   u32 tib40[4];
1183   u32 tib41[4];
1184
1185   tib40[0] = buf0[0];
1186   tib40[1] = buf0[1];
1187   tib40[2] = buf0[2];
1188   tib40[3] = buf0[3];
1189   tib41[0] = buf1[0];
1190   tib41[1] = buf1[1];
1191   tib41[2] = buf1[2];
1192   tib41[3] = buf1[3];
1193
1194   for (u32 i = 0; i < p0; i++)
1195   {
1196     append_block8 (out_len, buf0, buf1, buf0, buf1, tib40, tib41);
1197
1198     out_len += in_len;
1199   }
1200
1201   return out_len;
1202 }
1203
1204 static u32 rule_op_mangle_reflect (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1205 {
1206   if ((in_len + in_len) >= 32) return (in_len);
1207
1208   u32 out_len = in_len;
1209
1210   u32 tib40[4];
1211   u32 tib41[4];
1212
1213   reverse_block (buf0, buf1, tib40, tib41, out_len);
1214
1215   append_block8 (out_len, buf0, buf1, buf0, buf1, tib40, tib41);
1216
1217   out_len += in_len;
1218
1219   return out_len;
1220 }
1221
1222 static u32 rule_op_mangle_append (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1223 {
1224   if ((in_len + 1) >= 32) return (in_len);
1225
1226   u32 out_len = in_len;
1227
1228   append_block1 (out_len, buf0, buf1, p0);
1229
1230   out_len++;
1231
1232   return out_len;
1233 }
1234
1235 static u32 rule_op_mangle_prepend (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1236 {
1237   if ((in_len + 1) >= 32) return (in_len);
1238
1239   u32 out_len = in_len;
1240
1241   rshift_block (buf0, buf1, buf0, buf1);
1242
1243   buf0[0] = buf0[0] | p0;
1244
1245   out_len++;
1246
1247   return out_len;
1248 }
1249
1250 static u32 rule_op_mangle_rotate_left (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1251 {
1252   if (in_len == 0) return (in_len);
1253
1254   const u32 in_len1 = in_len - 1;
1255
1256   const u32 sh = (in_len1 & 3) * 8;
1257
1258   const u32 tmp = (buf0[0] & 0xff) << sh;
1259
1260   lshift_block (buf0, buf1, buf0, buf1);
1261
1262   switch (in_len1 / 4)
1263   {
1264     case  0:  buf0[0] |= tmp; break;
1265     case  1:  buf0[1] |= tmp; break;
1266     case  2:  buf0[2] |= tmp; break;
1267     case  3:  buf0[3] |= tmp; break;
1268     case  4:  buf1[0] |= tmp; break;
1269     case  5:  buf1[1] |= tmp; break;
1270     case  6:  buf1[2] |= tmp; break;
1271     case  7:  buf1[3] |= tmp; break;
1272   }
1273
1274   return in_len;
1275 }
1276
1277 static u32 rule_op_mangle_rotate_right (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1278 {
1279   if (in_len == 0) return (in_len);
1280
1281   const u32 in_len1 = in_len - 1;
1282
1283   const u32 sh = (in_len1 & 3) * 8;
1284
1285   u32 tmp = 0;
1286
1287   switch (in_len1 / 4)
1288   {
1289     case  0:  tmp = (buf0[0] >> sh) & 0xff; break;
1290     case  1:  tmp = (buf0[1] >> sh) & 0xff; break;
1291     case  2:  tmp = (buf0[2] >> sh) & 0xff; break;
1292     case  3:  tmp = (buf0[3] >> sh) & 0xff; break;
1293     case  4:  tmp = (buf1[0] >> sh) & 0xff; break;
1294     case  5:  tmp = (buf1[1] >> sh) & 0xff; break;
1295     case  6:  tmp = (buf1[2] >> sh) & 0xff; break;
1296     case  7:  tmp = (buf1[3] >> sh) & 0xff; break;
1297   }
1298
1299   rshift_block (buf0, buf1, buf0, buf1);
1300
1301   buf0[0] |= tmp;
1302
1303   truncate_right (buf0, buf1, in_len);
1304
1305   return in_len;
1306 }
1307
1308 static u32 rule_op_mangle_delete_first (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1309 {
1310   if (in_len == 0) return (in_len);
1311
1312   const u32 in_len1 = in_len - 1;
1313
1314   lshift_block (buf0, buf1, buf0, buf1);
1315
1316   return in_len1;
1317 }
1318
1319 static u32 rule_op_mangle_delete_last (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1320 {
1321   if (in_len == 0) return (in_len);
1322
1323   const u32 in_len1 = in_len - 1;
1324
1325   const u32 tmp = (1 << ((in_len1 & 3) * 8)) - 1;
1326
1327   switch (in_len1 / 4)
1328   {
1329     case  0:  buf0[0] &= tmp; break;
1330     case  1:  buf0[1] &= tmp; break;
1331     case  2:  buf0[2] &= tmp; break;
1332     case  3:  buf0[3] &= tmp; break;
1333     case  4:  buf1[0] &= tmp; break;
1334     case  5:  buf1[1] &= tmp; break;
1335     case  6:  buf1[2] &= tmp; break;
1336     case  7:  buf1[3] &= tmp; break;
1337   }
1338
1339   return in_len1;
1340 }
1341
1342 static u32 rule_op_mangle_delete_at (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1343 {
1344   if (p0 >= in_len) return (in_len);
1345
1346   u32 out_len = in_len;
1347
1348   u32 tib40[4];
1349   u32 tib41[4];
1350
1351   lshift_block (buf0, buf1, tib40, tib41);
1352
1353   const u32 ml = (1 << ((p0 & 3) * 8)) - 1;
1354   const u32 mr = ~ml;
1355
1356   switch (p0 / 4)
1357   {
1358     case  0:  buf0[0] =  (buf0[0] & ml)
1359                       | (tib40[0] & mr);
1360               buf0[1] =  tib40[1];
1361               buf0[2] =  tib40[2];
1362               buf0[3] =  tib40[3];
1363               buf1[0] =  tib41[0];
1364               buf1[1] =  tib41[1];
1365               buf1[2] =  tib41[2];
1366               buf1[3] =  tib41[3];
1367               break;
1368     case  1:  buf0[1] =  (buf0[1] & ml)
1369                       | (tib40[1] & mr);
1370               buf0[2] =  tib40[2];
1371               buf0[3] =  tib40[3];
1372               buf1[0] =  tib41[0];
1373               buf1[1] =  tib41[1];
1374               buf1[2] =  tib41[2];
1375               buf1[3] =  tib41[3];
1376               break;
1377     case  2:  buf0[2] =  (buf0[2] & ml)
1378                       | (tib40[2] & mr);
1379               buf0[3] =  tib40[3];
1380               buf1[0] =  tib41[0];
1381               buf1[1] =  tib41[1];
1382               buf1[2] =  tib41[2];
1383               buf1[3] =  tib41[3];
1384               break;
1385     case  3:  buf0[3] =  (buf0[3] & ml)
1386                       | (tib40[3] & mr);
1387               buf1[0] =  tib41[0];
1388               buf1[1] =  tib41[1];
1389               buf1[2] =  tib41[2];
1390               buf1[3] =  tib41[3];
1391               break;
1392     case  4:  buf1[0] =  (buf1[0] & ml)
1393                       | (tib41[0] & mr);
1394               buf1[1] =  tib41[1];
1395               buf1[2] =  tib41[2];
1396               buf1[3] =  tib41[3];
1397               break;
1398     case  5:  buf1[1] =  (buf1[1] & ml)
1399                       | (tib41[1] & mr);
1400               buf1[2] =  tib41[2];
1401               buf1[3] =  tib41[3];
1402               break;
1403     case  6:  buf1[2] =  (buf1[2] & ml)
1404                       | (tib41[2] & mr);
1405               buf1[3] =  tib41[3];
1406               break;
1407     case  7:  buf1[3] =  (buf1[3] & ml)
1408                       | (tib41[3] & mr);
1409               break;
1410   }
1411
1412   out_len--;
1413
1414   return out_len;
1415 }
1416
1417 static u32 rule_op_mangle_extract (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1418 {
1419   if (p0 >= in_len) return (in_len);
1420
1421   if ((p0 + p1) > in_len) return (in_len);
1422
1423   u32 out_len = p1;
1424
1425   lshift_block_N (buf0, buf1, buf0, buf1, p0);
1426
1427   truncate_right (buf0, buf1, out_len);
1428
1429   return out_len;
1430 }
1431
1432 static u32 rule_op_mangle_omit (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1433 {
1434   if (p0 >= in_len) return (in_len);
1435
1436   if ((p0 + p1) > in_len) return (in_len);
1437
1438   u32 out_len = in_len;
1439
1440   u32 tib40[4];
1441   u32 tib41[4];
1442
1443   tib40[0] = 0;
1444   tib40[1] = 0;
1445   tib40[2] = 0;
1446   tib40[3] = 0;
1447   tib41[0] = 0;
1448   tib41[1] = 0;
1449   tib41[2] = 0;
1450   tib41[3] = 0;
1451
1452   lshift_block_N (buf0, buf1, tib40, tib41, p1);
1453
1454   const u32 ml = (1 << ((p0 & 3) * 8)) - 1;
1455   const u32 mr = ~ml;
1456
1457   switch (p0 / 4)
1458   {
1459     case  0:  buf0[0] =  (buf0[0] & ml)
1460                       | (tib40[0] & mr);
1461               buf0[1] =  tib40[1];
1462               buf0[2] =  tib40[2];
1463               buf0[3] =  tib40[3];
1464               buf1[0] =  tib41[0];
1465               buf1[1] =  tib41[1];
1466               buf1[2] =  tib41[2];
1467               buf1[3] =  tib41[3];
1468               break;
1469     case  1:  buf0[1] =  (buf0[1] & ml)
1470                       | (tib40[1] & mr);
1471               buf0[2] =  tib40[2];
1472               buf0[3] =  tib40[3];
1473               buf1[0] =  tib41[0];
1474               buf1[1] =  tib41[1];
1475               buf1[2] =  tib41[2];
1476               buf1[3] =  tib41[3];
1477               break;
1478     case  2:  buf0[2] =  (buf0[2] & ml)
1479                       | (tib40[2] & mr);
1480               buf0[3] =  tib40[3];
1481               buf1[0] =  tib41[0];
1482               buf1[1] =  tib41[1];
1483               buf1[2] =  tib41[2];
1484               buf1[3] =  tib41[3];
1485               break;
1486     case  3:  buf0[3] =  (buf0[3] & ml)
1487                       | (tib40[3] & mr);
1488               buf1[0] =  tib41[0];
1489               buf1[1] =  tib41[1];
1490               buf1[2] =  tib41[2];
1491               buf1[3] =  tib41[3];
1492               break;
1493     case  4:  buf1[0] =  (buf1[0] & ml)
1494                       | (tib41[0] & mr);
1495               buf1[1] =  tib41[1];
1496               buf1[2] =  tib41[2];
1497               buf1[3] =  tib41[3];
1498               break;
1499     case  5:  buf1[1] =  (buf1[1] & ml)
1500                       | (tib41[1] & mr);
1501               buf1[2] =  tib41[2];
1502               buf1[3] =  tib41[3];
1503               break;
1504     case  6:  buf1[2] =  (buf1[2] & ml)
1505                       | (tib41[2] & mr);
1506               buf1[3] =  tib41[3];
1507               break;
1508     case  7:  buf1[3] =  (buf1[3] & ml)
1509                       | (tib41[3] & mr);
1510               break;
1511   }
1512
1513   out_len -= p1;
1514
1515   return out_len;
1516 }
1517
1518 static u32 rule_op_mangle_insert (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1519 {
1520   if (p0 > in_len) return (in_len);
1521
1522   if ((in_len + 1) >= 32) return (in_len);
1523
1524   u32 out_len = in_len;
1525
1526   u32 tib40[4];
1527   u32 tib41[4];
1528
1529   rshift_block (buf0, buf1, tib40, tib41);
1530
1531   const u32 p1n = p1 << ((p0 & 3) * 8);
1532
1533   const u32 ml = (1 << ((p0 & 3) * 8)) - 1;
1534
1535   const u32 mr = 0xffffff00 << ((p0 & 3) * 8);
1536
1537   switch (p0 / 4)
1538   {
1539     case  0:  buf0[0] =  (buf0[0] & ml) | p1n | (tib40[0] & mr);
1540               buf0[1] =  tib40[1];
1541               buf0[2] =  tib40[2];
1542               buf0[3] =  tib40[3];
1543               buf1[0] =  tib41[0];
1544               buf1[1] =  tib41[1];
1545               buf1[2] =  tib41[2];
1546               buf1[3] =  tib41[3];
1547               break;
1548     case  1:  buf0[1] =  (buf0[1] & ml) | p1n | (tib40[1] & mr);
1549               buf0[2] =  tib40[2];
1550               buf0[3] =  tib40[3];
1551               buf1[0] =  tib41[0];
1552               buf1[1] =  tib41[1];
1553               buf1[2] =  tib41[2];
1554               buf1[3] =  tib41[3];
1555               break;
1556     case  2:  buf0[2] =  (buf0[2] & ml) | p1n | (tib40[2] & mr);
1557               buf0[3] =  tib40[3];
1558               buf1[0] =  tib41[0];
1559               buf1[1] =  tib41[1];
1560               buf1[2] =  tib41[2];
1561               buf1[3] =  tib41[3];
1562               break;
1563     case  3:  buf0[3] =  (buf0[3] & ml) | p1n | (tib40[3] & mr);
1564               buf1[0] =  tib41[0];
1565               buf1[1] =  tib41[1];
1566               buf1[2] =  tib41[2];
1567               buf1[3] =  tib41[3];
1568               break;
1569     case  4:  buf1[0] =  (buf1[0] & ml) | p1n | (tib41[0] & mr);
1570               buf1[1] =  tib41[1];
1571               buf1[2] =  tib41[2];
1572               buf1[3] =  tib41[3];
1573               break;
1574     case  5:  buf1[1] =  (buf1[1] & ml) | p1n | (tib41[1] & mr);
1575               buf1[2] =  tib41[2];
1576               buf1[3] =  tib41[3];
1577               break;
1578     case  6:  buf1[2] =  (buf1[2] & ml) | p1n | (tib41[2] & mr);
1579               buf1[3] =  tib41[3];
1580               break;
1581     case  7:  buf1[3] =  (buf1[3] & ml) | p1n | (tib41[3] & mr);
1582               break;
1583   }
1584
1585   out_len++;
1586
1587   return out_len;
1588 }
1589
1590 static u32 rule_op_mangle_overstrike (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1591 {
1592   if (p0 >= in_len) return (in_len);
1593
1594   const u32 p1n = p1 << ((p0 & 3) * 8);
1595
1596   const u32 m = ~(0xffu << ((p0 & 3) * 8));
1597
1598   switch (p0 / 4)
1599   {
1600     case  0: buf0[0] = (buf0[0] & m) | p1n; break;
1601     case  1: buf0[1] = (buf0[1] & m) | p1n; break;
1602     case  2: buf0[2] = (buf0[2] & m) | p1n; break;
1603     case  3: buf0[3] = (buf0[3] & m) | p1n; break;
1604     case  4: buf1[0] = (buf1[0] & m) | p1n; break;
1605     case  5: buf1[1] = (buf1[1] & m) | p1n; break;
1606     case  6: buf1[2] = (buf1[2] & m) | p1n; break;
1607     case  7: buf1[3] = (buf1[3] & m) | p1n; break;
1608   }
1609
1610   return in_len;
1611 }
1612
1613 static u32 rule_op_mangle_truncate_at (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1614 {
1615   if (p0 >= in_len) return (in_len);
1616
1617   truncate_right (buf0, buf1, p0);
1618
1619   return p0;
1620 }
1621
1622 static u32 rule_op_mangle_replace (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1623 {
1624   for (u32 i = 0; i < in_len; i++)
1625   {
1626     switch (i)
1627     {
1628       case  0: if (((buf0[0] >>  0) & 0xFF) == p0) buf0[0] = (buf0[0] & 0xFFFFFF00) | p1 <<  0; break;
1629       case  1: if (((buf0[0] >>  8) & 0xFF) == p0) buf0[0] = (buf0[0] & 0xFFFF00FF) | p1 <<  8; break;
1630       case  2: if (((buf0[0] >> 16) & 0xFF) == p0) buf0[0] = (buf0[0] & 0xFF00FFFF) | p1 << 16; break;
1631       case  3: if (((buf0[0] >> 24) & 0xFF) == p0) buf0[0] = (buf0[0] & 0x00FFFFFF) | p1 << 24; break;
1632       case  4: if (((buf0[1] >>  0) & 0xFF) == p0) buf0[1] = (buf0[1] & 0xFFFFFF00) | p1 <<  0; break;
1633       case  5: if (((buf0[1] >>  8) & 0xFF) == p0) buf0[1] = (buf0[1] & 0xFFFF00FF) | p1 <<  8; break;
1634       case  6: if (((buf0[1] >> 16) & 0xFF) == p0) buf0[1] = (buf0[1] & 0xFF00FFFF) | p1 << 16; break;
1635       case  7: if (((buf0[1] >> 24) & 0xFF) == p0) buf0[1] = (buf0[1] & 0x00FFFFFF) | p1 << 24; break;
1636       case  8: if (((buf0[2] >>  0) & 0xFF) == p0) buf0[2] = (buf0[2] & 0xFFFFFF00) | p1 <<  0; break;
1637       case  9: if (((buf0[2] >>  8) & 0xFF) == p0) buf0[2] = (buf0[2] & 0xFFFF00FF) | p1 <<  8; break;
1638       case 10: if (((buf0[2] >> 16) & 0xFF) == p0) buf0[2] = (buf0[2] & 0xFF00FFFF) | p1 << 16; break;
1639       case 11: if (((buf0[2] >> 24) & 0xFF) == p0) buf0[2] = (buf0[2] & 0x00FFFFFF) | p1 << 24; break;
1640       case 12: if (((buf0[3] >>  0) & 0xFF) == p0) buf0[3] = (buf0[3] & 0xFFFFFF00) | p1 <<  0; break;
1641       case 13: if (((buf0[3] >>  8) & 0xFF) == p0) buf0[3] = (buf0[3] & 0xFFFF00FF) | p1 <<  8; break;
1642       case 14: if (((buf0[3] >> 16) & 0xFF) == p0) buf0[3] = (buf0[3] & 0xFF00FFFF) | p1 << 16; break;
1643       case 15: if (((buf0[3] >> 24) & 0xFF) == p0) buf0[3] = (buf0[3] & 0x00FFFFFF) | p1 << 24; break;
1644       case 16: if (((buf1[0] >>  0) & 0xFF) == p0) buf1[0] = (buf1[0] & 0xFFFFFF00) | p1 <<  0; break;
1645       case 17: if (((buf1[0] >>  8) & 0xFF) == p0) buf1[0] = (buf1[0] & 0xFFFF00FF) | p1 <<  8; break;
1646       case 18: if (((buf1[0] >> 16) & 0xFF) == p0) buf1[0] = (buf1[0] & 0xFF00FFFF) | p1 << 16; break;
1647       case 19: if (((buf1[0] >> 24) & 0xFF) == p0) buf1[0] = (buf1[0] & 0x00FFFFFF) | p1 << 24; break;
1648       case 20: if (((buf1[1] >>  0) & 0xFF) == p0) buf1[1] = (buf1[1] & 0xFFFFFF00) | p1 <<  0; break;
1649       case 21: if (((buf1[1] >>  8) & 0xFF) == p0) buf1[1] = (buf1[1] & 0xFFFF00FF) | p1 <<  8; break;
1650       case 22: if (((buf1[1] >> 16) & 0xFF) == p0) buf1[1] = (buf1[1] & 0xFF00FFFF) | p1 << 16; break;
1651       case 23: if (((buf1[1] >> 24) & 0xFF) == p0) buf1[1] = (buf1[1] & 0x00FFFFFF) | p1 << 24; break;
1652       case 24: if (((buf1[2] >>  0) & 0xFF) == p0) buf1[2] = (buf1[2] & 0xFFFFFF00) | p1 <<  0; break;
1653       case 25: if (((buf1[2] >>  8) & 0xFF) == p0) buf1[2] = (buf1[2] & 0xFFFF00FF) | p1 <<  8; break;
1654       case 26: if (((buf1[2] >> 16) & 0xFF) == p0) buf1[2] = (buf1[2] & 0xFF00FFFF) | p1 << 16; break;
1655       case 27: if (((buf1[2] >> 24) & 0xFF) == p0) buf1[2] = (buf1[2] & 0x00FFFFFF) | p1 << 24; break;
1656       case 28: if (((buf1[3] >>  0) & 0xFF) == p0) buf1[3] = (buf1[3] & 0xFFFFFF00) | p1 <<  0; break;
1657       case 29: if (((buf1[3] >>  8) & 0xFF) == p0) buf1[3] = (buf1[3] & 0xFFFF00FF) | p1 <<  8; break;
1658       case 30: if (((buf1[3] >> 16) & 0xFF) == p0) buf1[3] = (buf1[3] & 0xFF00FFFF) | p1 << 16; break;
1659       case 31: if (((buf1[3] >> 24) & 0xFF) == p0) buf1[3] = (buf1[3] & 0x00FFFFFF) | p1 << 24; break;
1660     }
1661   }
1662
1663   return in_len;
1664 }
1665
1666 /*
1667 static u32 rule_op_mangle_purgechar (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1668 {
1669   // TODO
1670   return in_len;
1671 }
1672
1673 static u32 rule_op_mangle_togglecase_rec (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1674 {
1675   // TODO
1676   return in_len;
1677 }
1678 */
1679
1680 static u32 rule_op_mangle_dupechar_first (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1681 {
1682   if ( in_len       ==  0) return (in_len);
1683   if ((in_len + p0) >= 32) return (in_len);
1684
1685   u32 out_len = in_len;
1686
1687   const u32 tmp = buf0[0] & 0xFF;
1688
1689   rshift_block_N (buf0, buf1, buf0, buf1, p0);
1690
1691   switch (p0)
1692   {
1693     case  1:  buf0[0] |= tmp <<  0;
1694               break;
1695     case  2:  buf0[0] |= tmp <<  0 | tmp << 8;
1696               break;
1697     case  3:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16;
1698               break;
1699     case  4:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1700               break;
1701     case  5:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1702               buf0[1] |= tmp <<  0;
1703               break;
1704     case  6:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1705               buf0[1] |= tmp <<  0 | tmp << 8;
1706               break;
1707     case  7:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1708               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16;
1709               break;
1710     case  8:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1711               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1712               break;
1713     case  9:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1714               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1715               buf0[2] |= tmp <<  0;
1716               break;
1717     case 10:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1718               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1719               buf0[2] |= tmp <<  0 | tmp << 8;
1720               break;
1721     case 11:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1722               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1723               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16;
1724               break;
1725     case 12:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1726               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1727               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1728               break;
1729     case 13:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1730               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1731               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1732               buf0[3] |= tmp <<  0;
1733               break;
1734     case 14:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1735               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1736               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1737               buf0[3] |= tmp <<  0 | tmp << 8;
1738               break;
1739     case 15:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1740               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1741               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1742               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16;
1743               break;
1744     case 16:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1745               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1746               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1747               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1748               break;
1749     case 17:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1750               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1751               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1752               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1753               buf1[0] |= tmp <<  0;
1754               break;
1755     case 18:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1756               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1757               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1758               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1759               buf1[0] |= tmp <<  0 | tmp << 8;
1760               break;
1761     case 19:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1762               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1763               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1764               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1765               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16;
1766               break;
1767     case 20:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1768               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1769               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1770               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1771               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1772               break;
1773     case 21:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1774               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1775               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1776               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1777               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1778               buf1[1] |= tmp <<  0;
1779               break;
1780     case 22:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1781               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1782               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1783               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1784               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1785               buf1[1] |= tmp <<  0 | tmp << 8;
1786               break;
1787     case 23:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1788               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1789               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1790               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1791               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1792               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16;
1793               break;
1794     case 24:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1795               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1796               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1797               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1798               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1799               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1800               break;
1801     case 25:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1802               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1803               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1804               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1805               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1806               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1807               buf1[2] |= tmp <<  0;
1808               break;
1809     case 26:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1810               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1811               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1812               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1813               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1814               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1815               buf1[2] |= tmp <<  0 | tmp << 8;
1816               break;
1817     case 27:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1818               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1819               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1820               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1821               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1822               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1823               buf1[2] |= tmp <<  0 | tmp << 8 | tmp << 16;
1824               break;
1825     case 28:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1826               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1827               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1828               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1829               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1830               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1831               buf1[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1832               break;
1833     case 29:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1834               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1835               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1836               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1837               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1838               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1839               buf1[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1840               buf1[3] |= tmp <<  0;
1841               break;
1842     case 30:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1843               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1844               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1845               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1846               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1847               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1848               buf1[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1849               buf1[3] |= tmp <<  0 | tmp << 8;
1850               break;
1851     case 31:  buf0[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1852               buf0[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1853               buf0[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1854               buf0[3] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1855               buf1[0] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1856               buf1[1] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1857               buf1[2] |= tmp <<  0 | tmp << 8 | tmp << 16 | tmp << 24;
1858               buf1[3] |= tmp <<  0 | tmp << 8 | tmp << 16;
1859               break;
1860   }
1861
1862   out_len += p0;
1863
1864   return out_len;
1865 }
1866
1867 static u32 rule_op_mangle_dupechar_last (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1868 {
1869   if ( in_len       ==  0) return (in_len);
1870   if ((in_len + p0) >= 32) return (in_len);
1871
1872   const u32 in_len1 = in_len - 1;
1873
1874   const u32 sh = (in_len1 & 3) * 8;
1875
1876   u32 tmp = 0;
1877
1878   switch (in_len1 / 4)
1879   {
1880     case  0:  tmp = (buf0[0] >> sh) & 0xff; break;
1881     case  1:  tmp = (buf0[1] >> sh) & 0xff; break;
1882     case  2:  tmp = (buf0[2] >> sh) & 0xff; break;
1883     case  3:  tmp = (buf0[3] >> sh) & 0xff; break;
1884     case  4:  tmp = (buf1[0] >> sh) & 0xff; break;
1885     case  5:  tmp = (buf1[1] >> sh) & 0xff; break;
1886     case  6:  tmp = (buf1[2] >> sh) & 0xff; break;
1887     case  7:  tmp = (buf1[3] >> sh) & 0xff; break;
1888   }
1889
1890   u32 out_len = in_len;
1891
1892   for (u32 i = 0; i < p0; i++)
1893   {
1894     append_block1 (out_len, buf0, buf1, tmp);
1895
1896     out_len++;
1897   }
1898
1899   return out_len;
1900 }
1901
1902 static u32 rule_op_mangle_dupechar_all (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1903 {
1904   if ( in_len           ==  0) return (in_len);
1905   if ((in_len + in_len) >= 32) return (in_len);
1906
1907   u32 out_len = in_len;
1908
1909   u32 tib40[4];
1910   u32 tib41[4];
1911
1912   tib40[0] = ((buf0[0] & 0x000000FF) <<  0) | ((buf0[0] & 0x0000FF00) <<  8);
1913   tib40[1] = ((buf0[0] & 0x00FF0000) >> 16) | ((buf0[0] & 0xFF000000) >>  8);
1914   tib40[2] = ((buf0[1] & 0x000000FF) <<  0) | ((buf0[1] & 0x0000FF00) <<  8);
1915   tib40[3] = ((buf0[1] & 0x00FF0000) >> 16) | ((buf0[1] & 0xFF000000) >>  8);
1916   tib41[0] = ((buf0[2] & 0x000000FF) <<  0) | ((buf0[2] & 0x0000FF00) <<  8);
1917   tib41[1] = ((buf0[2] & 0x00FF0000) >> 16) | ((buf0[2] & 0xFF000000) >>  8);
1918   tib41[2] = ((buf0[3] & 0x000000FF) <<  0) | ((buf0[3] & 0x0000FF00) <<  8);
1919   tib41[3] = ((buf0[3] & 0x00FF0000) >> 16) | ((buf0[3] & 0xFF000000) >>  8);
1920
1921   buf0[0] = tib40[0] | (tib40[0] <<  8);
1922   buf0[1] = tib40[1] | (tib40[1] <<  8);
1923   buf0[2] = tib40[2] | (tib40[2] <<  8);
1924   buf0[3] = tib40[3] | (tib40[3] <<  8);
1925   buf1[0] = tib41[0] | (tib41[0] <<  8);
1926   buf1[1] = tib41[1] | (tib41[1] <<  8);
1927   buf1[2] = tib41[2] | (tib41[2] <<  8);
1928   buf1[3] = tib41[3] | (tib41[3] <<  8);
1929
1930   out_len = out_len + out_len;
1931
1932   return out_len;
1933 }
1934
1935 static u32 rule_op_mangle_switch_first (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1936 {
1937   if (in_len < 2) return (in_len);
1938
1939   buf0[0] = (buf0[0] & 0xFFFF0000) | ((buf0[0] << 8) & 0x0000FF00) | ((buf0[0] >> 8) & 0x000000FF);
1940
1941   return in_len;
1942 }
1943
1944 static u32 rule_op_mangle_switch_last (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
1945 {
1946   if (in_len < 2) return (in_len);
1947
1948   switch (in_len)
1949   {
1950     case  2:  buf0[0] = ((buf0[0] << 8) & 0x0000FF00) | ((buf0[0] >> 8) & 0x000000FF);
1951               break;
1952     case  3:  buf0[0] =  (buf0[0] & 0x000000FF) | ((buf0[0] <<  8) & 0x00FF0000) | ((buf0[0] >>  8) & 0x0000FF00);
1953               break;
1954     case  4:  buf0[0] =  (buf0[0] & 0x0000FFFF) | ((buf0[0] <<  8) & 0xFF000000) | ((buf0[0] >>  8) & 0x00FF0000);
1955               break;
1956     case  5:  buf0[1] =  (buf0[0] & 0xFF000000) |   buf0[1];
1957               buf0[0] =  (buf0[0] & 0x00FFFFFF) |  (buf0[1] << 24);
1958               buf0[1] =  (buf0[1] >> 24);
1959               break;
1960     case  6:  buf0[1] = ((buf0[1] << 8) & 0x0000FF00) | ((buf0[1] >> 8) & 0x000000FF);
1961               break;
1962     case  7:  buf0[1] =  (buf0[1] & 0x000000FF) | ((buf0[1] <<  8) & 0x00FF0000) | ((buf0[1] >>  8) & 0x0000FF00);
1963               break;
1964     case  8:  buf0[1] =  (buf0[1] & 0x0000FFFF) | ((buf0[1] <<  8) & 0xFF000000) | ((buf0[1] >>  8) & 0x00FF0000);
1965               break;
1966     case  9:  buf0[2] =  (buf0[1] & 0xFF000000) |   buf0[2];
1967               buf0[1] =  (buf0[1] & 0x00FFFFFF) |  (buf0[2] << 24);
1968               buf0[2] =  (buf0[2] >> 24);
1969               break;
1970     case 10:  buf0[2] = ((buf0[2] << 8) & 0x0000FF00) | ((buf0[2] >> 8) & 0x000000FF);
1971               break;
1972     case 11:  buf0[2] =  (buf0[2] & 0x000000FF) | ((buf0[2] <<  8) & 0x00FF0000) | ((buf0[2] >>  8) & 0x0000FF00);
1973               break;
1974     case 12:  buf0[2] =  (buf0[2] & 0x0000FFFF) | ((buf0[2] <<  8) & 0xFF000000) | ((buf0[2] >>  8) & 0x00FF0000);
1975               break;
1976     case 13:  buf0[3] =  (buf0[2] & 0xFF000000) |   buf0[3];
1977               buf0[2] =  (buf0[2] & 0x00FFFFFF) |  (buf0[3] << 24);
1978               buf0[3] =  (buf0[3] >> 24);
1979               break;
1980     case 14:  buf0[3] = ((buf0[3] << 8) & 0x0000FF00) | ((buf0[3] >> 8) & 0x000000FF);
1981               break;
1982     case 15:  buf0[3] =  (buf0[3] & 0x000000FF) |  ((buf0[3] <<  8) & 0x00FF0000) | ((buf0[3] >>  8) & 0x0000FF00);
1983               break;
1984     case 16:  buf0[3] =  (buf0[3] & 0x0000FFFF) | ((buf0[3] <<  8) & 0xFF000000) | ((buf0[3] >>  8) & 0x00FF0000);
1985               break;
1986     case 17:  buf1[0] =  (buf0[3] & 0xFF000000) |   buf1[0];
1987               buf0[3] =  (buf0[3] & 0x00FFFFFF) |  (buf1[0] << 24);
1988               buf1[0] =  (buf1[0] >> 24);
1989               break;
1990     case 18:  buf1[0] = ((buf1[0] << 8) & 0x0000FF00) | ((buf1[0] >> 8) & 0x000000FF);
1991               break;
1992     case 19:  buf1[0] =  (buf1[0] & 0x000000FF) | ((buf1[0] <<  8) & 0x00FF0000) | ((buf1[0] >>  8) & 0x0000FF00);
1993               break;
1994     case 20:  buf1[0] =  (buf1[0] & 0x0000FFFF) | ((buf1[0] <<  8) & 0xFF000000) | ((buf1[0] >>  8) & 0x00FF0000);
1995               break;
1996     case 21:  buf1[1] =  (buf1[0] & 0xFF000000) |   buf1[1];
1997               buf1[0] =  (buf1[0] & 0x00FFFFFF) |  (buf1[1] << 24);
1998               buf1[1] =  (buf1[1] >> 24);
1999               break;
2000     case 22:  buf1[1] = ((buf1[1] << 8) & 0x0000FF00) | ((buf1[1] >> 8) & 0x000000FF);
2001               break;
2002     case 23:  buf1[1] =  (buf1[1] & 0x000000FF) | ((buf1[1] <<  8) & 0x00FF0000) | ((buf1[1] >>  8) & 0x0000FF00);
2003               break;
2004     case 24:  buf1[1] =  (buf1[1] & 0x0000FFFF) | ((buf1[1] <<  8) & 0xFF000000) | ((buf1[1] >>  8) & 0x00FF0000);
2005               break;
2006     case 25:  buf1[2] =  (buf1[1] & 0xFF000000) |   buf1[2];
2007               buf1[1] =  (buf1[1] & 0x00FFFFFF) |  (buf1[2] << 24);
2008               buf1[2] =  (buf1[2] >> 24);
2009               break;
2010     case 26:  buf1[2] = ((buf1[2] << 8) & 0x0000FF00) | ((buf1[2] >> 8) & 0x000000FF);
2011               break;
2012     case 27:  buf1[2] =  (buf1[2] & 0x000000FF) | ((buf1[2] <<  8) & 0x00FF0000) | ((buf1[2] >>  8) & 0x0000FF00);
2013               break;
2014     case 28:  buf1[2] =  (buf1[2] & 0x0000FFFF) | ((buf1[2] <<  8) & 0xFF000000) | ((buf1[2] >>  8) & 0x00FF0000);
2015               break;
2016     case 29:  buf1[3] =  (buf1[2] & 0xFF000000) |   buf1[3];
2017               buf1[2] =  (buf1[2] & 0x00FFFFFF) |  (buf1[3] << 24);
2018               buf1[3] =  (buf1[3] >> 24);
2019               break;
2020     case 30:  buf1[3] = ((buf1[3] << 8) & 0x0000FF00) | ((buf1[3] >> 8) & 0x000000FF);
2021               break;
2022     case 31:  buf1[3] =  (buf1[3] & 0x000000FF) |  ((buf1[3] <<  8) & 0x00FF0000) | ((buf1[3] >>  8) & 0x0000FF00);
2023               break;
2024   }
2025
2026   return in_len;
2027 }
2028
2029 static u32 rule_op_mangle_switch_at (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2030 {
2031   if (p0 >= in_len) return (in_len);
2032   if (p1 >= in_len) return (in_len);
2033
2034   u32 tmp0 = 0;
2035   u32 tmp1 = 0;
2036
2037   switch (p0)
2038   {
2039     case  0:  tmp0 = (buf0[0] >>  0) & 0xFF;
2040               break;
2041     case  1:  tmp0 = (buf0[0] >>  8) & 0xFF;
2042               break;
2043     case  2:  tmp0 = (buf0[0] >> 16) & 0xFF;
2044               break;
2045     case  3:  tmp0 = (buf0[0] >> 24) & 0xFF;
2046               break;
2047     case  4:  tmp0 = (buf0[1] >>  0) & 0xFF;
2048               break;
2049     case  5:  tmp0 = (buf0[1] >>  8) & 0xFF;
2050               break;
2051     case  6:  tmp0 = (buf0[1] >> 16) & 0xFF;
2052               break;
2053     case  7:  tmp0 = (buf0[1] >> 24) & 0xFF;
2054               break;
2055     case  8:  tmp0 = (buf0[2] >>  0) & 0xFF;
2056               break;
2057     case  9:  tmp0 = (buf0[2] >>  8) & 0xFF;
2058               break;
2059     case 10:  tmp0 = (buf0[2] >> 16) & 0xFF;
2060               break;
2061     case 11:  tmp0 = (buf0[2] >> 24) & 0xFF;
2062               break;
2063     case 12:  tmp0 = (buf0[3] >>  0) & 0xFF;
2064               break;
2065     case 13:  tmp0 = (buf0[3] >>  8) & 0xFF;
2066               break;
2067     case 14:  tmp0 = (buf0[3] >> 16) & 0xFF;
2068               break;
2069     case 15:  tmp0 = (buf0[3] >> 24) & 0xFF;
2070               break;
2071     case 16:  tmp0 = (buf1[0] >>  0) & 0xFF;
2072               break;
2073     case 17:  tmp0 = (buf1[0] >>  8) & 0xFF;
2074               break;
2075     case 18:  tmp0 = (buf1[0] >> 16) & 0xFF;
2076               break;
2077     case 19:  tmp0 = (buf1[0] >> 24) & 0xFF;
2078               break;
2079     case 20:  tmp0 = (buf1[1] >>  0) & 0xFF;
2080               break;
2081     case 21:  tmp0 = (buf1[1] >>  8) & 0xFF;
2082               break;
2083     case 22:  tmp0 = (buf1[1] >> 16) & 0xFF;
2084               break;
2085     case 23:  tmp0 = (buf1[1] >> 24) & 0xFF;
2086               break;
2087     case 24:  tmp0 = (buf1[2] >>  0) & 0xFF;
2088               break;
2089     case 25:  tmp0 = (buf1[2] >>  8) & 0xFF;
2090               break;
2091     case 26:  tmp0 = (buf1[2] >> 16) & 0xFF;
2092               break;
2093     case 27:  tmp0 = (buf1[2] >> 24) & 0xFF;
2094               break;
2095     case 28:  tmp0 = (buf1[3] >>  0) & 0xFF;
2096               break;
2097     case 29:  tmp0 = (buf1[3] >>  8) & 0xFF;
2098               break;
2099     case 30:  tmp0 = (buf1[3] >> 16) & 0xFF;
2100               break;
2101     case 31:  tmp0 = (buf1[3] >> 24) & 0xFF;
2102               break;
2103   }
2104
2105   switch (p1)
2106   {
2107     case  0:  tmp1 = (buf0[0] >>  0) & 0xff;
2108               buf0[0]  = (buf0[0] & 0xffffff00) | tmp0 <<  0;
2109               break;
2110     case  1:  tmp1 = (buf0[0] >>  8) & 0xff;
2111               buf0[0]  = (buf0[0] & 0xffff00ff) | tmp0 <<  8;
2112               break;
2113     case  2:  tmp1 = (buf0[0] >> 16) & 0xff;
2114               buf0[0]  = (buf0[0] & 0xff00ffff) | tmp0 << 16;
2115               break;
2116     case  3:  tmp1 = (buf0[0] >> 24) & 0xff;
2117               buf0[0]  = (buf0[0] & 0x00ffffff) | tmp0 << 24;
2118               break;
2119     case  4:  tmp1 = (buf0[1] >>  0) & 0xff;
2120               buf0[1]  = (buf0[1] & 0xffffff00) | tmp0 <<  0;
2121               break;
2122     case  5:  tmp1 = (buf0[1] >>  8) & 0xff;
2123               buf0[1]  = (buf0[1] & 0xffff00ff) | tmp0 <<  8;
2124               break;
2125     case  6:  tmp1 = (buf0[1] >> 16) & 0xff;
2126               buf0[1]  = (buf0[1] & 0xff00ffff) | tmp0 << 16;
2127               break;
2128     case  7:  tmp1 = (buf0[1] >> 24) & 0xff;
2129               buf0[1]  = (buf0[1] & 0x00ffffff) | tmp0 << 24;
2130               break;
2131     case  8:  tmp1 = (buf0[2] >>  0) & 0xff;
2132               buf0[2]  = (buf0[2] & 0xffffff00) | tmp0 <<  0;
2133               break;
2134     case  9:  tmp1 = (buf0[2] >>  8) & 0xff;
2135               buf0[2]  = (buf0[2] & 0xffff00ff) | tmp0 <<  8;
2136               break;
2137     case 10:  tmp1 = (buf0[2] >> 16) & 0xff;
2138               buf0[2]  = (buf0[2] & 0xff00ffff) | tmp0 << 16;
2139               break;
2140     case 11:  tmp1 = (buf0[2] >> 24) & 0xff;
2141               buf0[2]  = (buf0[2] & 0x00ffffff) | tmp0 << 24;
2142               break;
2143     case 12:  tmp1 = (buf0[3] >>  0) & 0xff;
2144               buf0[3]  = (buf0[3] & 0xffffff00) | tmp0 <<  0;
2145               break;
2146     case 13:  tmp1 = (buf0[3] >>  8) & 0xff;
2147               buf0[3]  = (buf0[3] & 0xffff00ff) | tmp0 <<  8;
2148               break;
2149     case 14:  tmp1 = (buf0[3] >> 16) & 0xff;
2150               buf0[3]  = (buf0[3] & 0xff00ffff) | tmp0 << 16;
2151               break;
2152     case 15:  tmp1 = (buf0[3] >> 24) & 0xff;
2153               buf0[3]  = (buf0[3] & 0x00ffffff) | tmp0 << 24;
2154               break;
2155     case 16:  tmp1 = (buf1[0] >>  0) & 0xff;
2156               buf1[0]  = (buf1[0] & 0xffffff00) | tmp0 <<  0;
2157               break;
2158     case 17:  tmp1 = (buf1[0] >>  8) & 0xff;
2159               buf1[0]  = (buf1[0] & 0xffff00ff) | tmp0 <<  8;
2160               break;
2161     case 18:  tmp1 = (buf1[0] >> 16) & 0xff;
2162               buf1[0]  = (buf1[0] & 0xff00ffff) | tmp0 << 16;
2163               break;
2164     case 19:  tmp1 = (buf1[0] >> 24) & 0xff;
2165               buf1[0]  = (buf1[0] & 0x00ffffff) | tmp0 << 24;
2166               break;
2167     case 20:  tmp1 = (buf1[1] >>  0) & 0xff;
2168               buf1[1]  = (buf1[1] & 0xffffff00) | tmp0 <<  0;
2169               break;
2170     case 21:  tmp1 = (buf1[1] >>  8) & 0xff;
2171               buf1[1]  = (buf1[1] & 0xffff00ff) | tmp0 <<  8;
2172               break;
2173     case 22:  tmp1 = (buf1[1] >> 16) & 0xff;
2174               buf1[1]  = (buf1[1] & 0xff00ffff) | tmp0 << 16;
2175               break;
2176     case 23:  tmp1 = (buf1[1] >> 24) & 0xff;
2177               buf1[1]  = (buf1[1] & 0x00ffffff) | tmp0 << 24;
2178               break;
2179     case 24:  tmp1 = (buf1[2] >>  0) & 0xff;
2180               buf1[2]  = (buf1[2] & 0xffffff00) | tmp0 <<  0;
2181               break;
2182     case 25:  tmp1 = (buf1[2] >>  8) & 0xff;
2183               buf1[2]  = (buf1[2] & 0xffff00ff) | tmp0 <<  8;
2184               break;
2185     case 26:  tmp1 = (buf1[2] >> 16) & 0xff;
2186               buf1[2]  = (buf1[2] & 0xff00ffff) | tmp0 << 16;
2187               break;
2188     case 27:  tmp1 = (buf1[2] >> 24) & 0xff;
2189               buf1[2]  = (buf1[2] & 0x00ffffff) | tmp0 << 24;
2190               break;
2191     case 28:  tmp1 = (buf1[3] >>  0) & 0xff;
2192               buf1[3]  = (buf1[3] & 0xffffff00) | tmp0 <<  0;
2193               break;
2194     case 29:  tmp1 = (buf1[3] >>  8) & 0xff;
2195               buf1[3]  = (buf1[3] & 0xffff00ff) | tmp0 <<  8;
2196               break;
2197     case 30:  tmp1 = (buf1[3] >> 16) & 0xff;
2198               buf1[3]  = (buf1[3] & 0xff00ffff) | tmp0 << 16;
2199               break;
2200     case 31:  tmp1 = (buf1[3] >> 24) & 0xff;
2201               buf1[3]  = (buf1[3] & 0x00ffffff) | tmp0 << 24;
2202               break;
2203   }
2204
2205   switch (p0)
2206   {
2207     case  0:  buf0[0]  = (buf0[0] & 0xffffff00) | tmp1 <<  0;
2208               break;
2209     case  1:  buf0[0]  = (buf0[0] & 0xffff00ff) | tmp1 <<  8;
2210               break;
2211     case  2:  buf0[0]  = (buf0[0] & 0xff00ffff) | tmp1 << 16;
2212               break;
2213     case  3:  buf0[0]  = (buf0[0] & 0x00ffffff) | tmp1 << 24;
2214               break;
2215     case  4:  buf0[1]  = (buf0[1] & 0xffffff00) | tmp1 <<  0;
2216               break;
2217     case  5:  buf0[1]  = (buf0[1] & 0xffff00ff) | tmp1 <<  8;
2218               break;
2219     case  6:  buf0[1]  = (buf0[1] & 0xff00ffff) | tmp1 << 16;
2220               break;
2221     case  7:  buf0[1]  = (buf0[1] & 0x00ffffff) | tmp1 << 24;
2222               break;
2223     case  8:  buf0[2]  = (buf0[2] & 0xffffff00) | tmp1 <<  0;
2224               break;
2225     case  9:  buf0[2]  = (buf0[2] & 0xffff00ff) | tmp1 <<  8;
2226               break;
2227     case 10:  buf0[2]  = (buf0[2] & 0xff00ffff) | tmp1 << 16;
2228               break;
2229     case 11:  buf0[2]  = (buf0[2] & 0x00ffffff) | tmp1 << 24;
2230               break;
2231     case 12:  buf0[3]  = (buf0[3] & 0xffffff00) | tmp1 <<  0;
2232               break;
2233     case 13:  buf0[3]  = (buf0[3] & 0xffff00ff) | tmp1 <<  8;
2234               break;
2235     case 14:  buf0[3]  = (buf0[3] & 0xff00ffff) | tmp1 << 16;
2236               break;
2237     case 15:  buf0[3]  = (buf0[3] & 0x00ffffff) | tmp1 << 24;
2238               break;
2239     case 16:  buf1[0]  = (buf1[0] & 0xffffff00) | tmp1 <<  0;
2240               break;
2241     case 17:  buf1[0]  = (buf1[0] & 0xffff00ff) | tmp1 <<  8;
2242               break;
2243     case 18:  buf1[0]  = (buf1[0] & 0xff00ffff) | tmp1 << 16;
2244               break;
2245     case 19:  buf1[0]  = (buf1[0] & 0x00ffffff) | tmp1 << 24;
2246               break;
2247     case 20:  buf1[1]  = (buf1[1] & 0xffffff00) | tmp1 <<  0;
2248               break;
2249     case 21:  buf1[1]  = (buf1[1] & 0xffff00ff) | tmp1 <<  8;
2250               break;
2251     case 22:  buf1[1]  = (buf1[1] & 0xff00ffff) | tmp1 << 16;
2252               break;
2253     case 23:  buf1[1]  = (buf1[1] & 0x00ffffff) | tmp1 << 24;
2254               break;
2255     case 24:  buf1[2]  = (buf1[2] & 0xffffff00) | tmp1 <<  0;
2256               break;
2257     case 25:  buf1[2]  = (buf1[2] & 0xffff00ff) | tmp1 <<  8;
2258               break;
2259     case 26:  buf1[2]  = (buf1[2] & 0xff00ffff) | tmp1 << 16;
2260               break;
2261     case 27:  buf1[2]  = (buf1[2] & 0x00ffffff) | tmp1 << 24;
2262               break;
2263     case 28:  buf1[3]  = (buf1[3] & 0xffffff00) | tmp1 <<  0;
2264               break;
2265     case 29:  buf1[3]  = (buf1[3] & 0xffff00ff) | tmp1 <<  8;
2266               break;
2267     case 30:  buf1[3]  = (buf1[3] & 0xff00ffff) | tmp1 << 16;
2268               break;
2269     case 31:  buf1[3]  = (buf1[3] & 0x00ffffff) | tmp1 << 24;
2270               break;
2271   }
2272
2273   return in_len;
2274 }
2275
2276 static u32 rule_op_mangle_chr_shiftl (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2277 {
2278   if (p0 >= in_len) return (in_len);
2279
2280   const u32 mr = 0xffu << ((p0 & 3) * 8);
2281   const u32 ml = ~mr;
2282
2283   switch (p0 / 4)
2284   {
2285     case  0:  buf0[0] = (buf0[0] & ml) | (((buf0[0] & mr) << 1) & mr); break;
2286     case  1:  buf0[1] = (buf0[1] & ml) | (((buf0[1] & mr) << 1) & mr); break;
2287     case  2:  buf0[2] = (buf0[2] & ml) | (((buf0[2] & mr) << 1) & mr); break;
2288     case  3:  buf0[3] = (buf0[3] & ml) | (((buf0[3] & mr) << 1) & mr); break;
2289     case  4:  buf1[0] = (buf1[0] & ml) | (((buf1[0] & mr) << 1) & mr); break;
2290     case  5:  buf1[1] = (buf1[1] & ml) | (((buf1[1] & mr) << 1) & mr); break;
2291     case  6:  buf1[2] = (buf1[2] & ml) | (((buf1[2] & mr) << 1) & mr); break;
2292     case  7:  buf1[3] = (buf1[3] & ml) | (((buf1[3] & mr) << 1) & mr); break;
2293   }
2294
2295   return in_len;
2296 }
2297
2298 static u32 rule_op_mangle_chr_shiftr (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2299 {
2300   if (p0 >= in_len) return (in_len);
2301
2302   const u32 mr = 0xffu << ((p0 & 3) * 8);
2303   const u32 ml = ~mr;
2304
2305   switch (p0 / 4)
2306   {
2307     case  0:  buf0[0] = (buf0[0] & ml) | (((buf0[0] & mr) >> 1) & mr); break;
2308     case  1:  buf0[1] = (buf0[1] & ml) | (((buf0[1] & mr) >> 1) & mr); break;
2309     case  2:  buf0[2] = (buf0[2] & ml) | (((buf0[2] & mr) >> 1) & mr); break;
2310     case  3:  buf0[3] = (buf0[3] & ml) | (((buf0[3] & mr) >> 1) & mr); break;
2311     case  4:  buf1[0] = (buf1[0] & ml) | (((buf1[0] & mr) >> 1) & mr); break;
2312     case  5:  buf1[1] = (buf1[1] & ml) | (((buf1[1] & mr) >> 1) & mr); break;
2313     case  6:  buf1[2] = (buf1[2] & ml) | (((buf1[2] & mr) >> 1) & mr); break;
2314     case  7:  buf1[3] = (buf1[3] & ml) | (((buf1[3] & mr) >> 1) & mr); break;
2315   }
2316
2317   return in_len;
2318 }
2319
2320 static u32 rule_op_mangle_chr_incr (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2321 {
2322   if (p0 >= in_len) return (in_len);
2323
2324   const u32 mr = 0xffu << ((p0 & 3) * 8);
2325   const u32 ml = ~mr;
2326
2327   const u32 n = 0x01010101 & mr;
2328
2329   switch (p0 / 4)
2330   {
2331     case  0:  buf0[0] = (buf0[0] & ml) | (((buf0[0] & mr) + n) & mr); break;
2332     case  1:  buf0[1] = (buf0[1] & ml) | (((buf0[1] & mr) + n) & mr); break;
2333     case  2:  buf0[2] = (buf0[2] & ml) | (((buf0[2] & mr) + n) & mr); break;
2334     case  3:  buf0[3] = (buf0[3] & ml) | (((buf0[3] & mr) + n) & mr); break;
2335     case  4:  buf1[0] = (buf1[0] & ml) | (((buf1[0] & mr) + n) & mr); break;
2336     case  5:  buf1[1] = (buf1[1] & ml) | (((buf1[1] & mr) + n) & mr); break;
2337     case  6:  buf1[2] = (buf1[2] & ml) | (((buf1[2] & mr) + n) & mr); break;
2338     case  7:  buf1[3] = (buf1[3] & ml) | (((buf1[3] & mr) + n) & mr); break;
2339   }
2340
2341   return in_len;
2342 }
2343
2344 static u32 rule_op_mangle_chr_decr (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2345 {
2346   if (p0 >= in_len) return (in_len);
2347
2348   const u32 mr = 0xffu << ((p0 & 3) * 8);
2349   const u32 ml = ~mr;
2350
2351   const u32 n = 0x01010101 & mr;
2352
2353   switch (p0 / 4)
2354   {
2355     case  0:  buf0[0] = (buf0[0] & ml) | (((buf0[0] & mr) - n) & mr); break;
2356     case  1:  buf0[1] = (buf0[1] & ml) | (((buf0[1] & mr) - n) & mr); break;
2357     case  2:  buf0[2] = (buf0[2] & ml) | (((buf0[2] & mr) - n) & mr); break;
2358     case  3:  buf0[3] = (buf0[3] & ml) | (((buf0[3] & mr) - n) & mr); break;
2359     case  4:  buf1[0] = (buf1[0] & ml) | (((buf1[0] & mr) - n) & mr); break;
2360     case  5:  buf1[1] = (buf1[1] & ml) | (((buf1[1] & mr) - n) & mr); break;
2361     case  6:  buf1[2] = (buf1[2] & ml) | (((buf1[2] & mr) - n) & mr); break;
2362     case  7:  buf1[3] = (buf1[3] & ml) | (((buf1[3] & mr) - n) & mr); break;
2363   }
2364
2365   return in_len;
2366 }
2367
2368 static u32 rule_op_mangle_replace_np1 (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2369 {
2370   if ((p0 + 1) >= in_len) return (in_len);
2371
2372   u32 tib40[4];
2373   u32 tib41[4];
2374
2375   lshift_block (buf0, buf1, tib40, tib41);
2376
2377   const u32 mr = 0xffu << ((p0 & 3) * 8);
2378   const u32 ml = ~mr;
2379
2380   switch (p0 / 4)
2381   {
2382     case  0:  buf0[0] = (buf0[0] & ml) | (tib40[0] & mr); break;
2383     case  1:  buf0[1] = (buf0[1] & ml) | (tib40[1] & mr); break;
2384     case  2:  buf0[2] = (buf0[2] & ml) | (tib40[2] & mr); break;
2385     case  3:  buf0[3] = (buf0[3] & ml) | (tib40[3] & mr); break;
2386     case  4:  buf1[0] = (buf1[0] & ml) | (tib41[0] & mr); break;
2387     case  5:  buf1[1] = (buf1[1] & ml) | (tib41[1] & mr); break;
2388     case  6:  buf1[2] = (buf1[2] & ml) | (tib41[2] & mr); break;
2389     case  7:  buf1[3] = (buf1[3] & ml) | (tib41[3] & mr); break;
2390   }
2391
2392   return in_len;
2393 }
2394
2395 static u32 rule_op_mangle_replace_nm1 (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2396 {
2397   if (p0 == 0) return (in_len);
2398
2399   if (p0 >= in_len) return (in_len);
2400
2401   u32 tib40[4];
2402   u32 tib41[4];
2403
2404   rshift_block (buf0, buf1, tib40, tib41);
2405
2406   const u32 mr = 0xffu << ((p0 & 3) * 8);
2407   const u32 ml = ~mr;
2408
2409   switch (p0 / 4)
2410   {
2411     case  0:  buf0[0] = (buf0[0] & ml) | (tib40[0] & mr); break;
2412     case  1:  buf0[1] = (buf0[1] & ml) | (tib40[1] & mr); break;
2413     case  2:  buf0[2] = (buf0[2] & ml) | (tib40[2] & mr); break;
2414     case  3:  buf0[3] = (buf0[3] & ml) | (tib40[3] & mr); break;
2415     case  4:  buf1[0] = (buf1[0] & ml) | (tib41[0] & mr); break;
2416     case  5:  buf1[1] = (buf1[1] & ml) | (tib41[1] & mr); break;
2417     case  6:  buf1[2] = (buf1[2] & ml) | (tib41[2] & mr); break;
2418     case  7:  buf1[3] = (buf1[3] & ml) | (tib41[3] & mr); break;
2419   }
2420
2421   return in_len;
2422 }
2423
2424 static u32 rule_op_mangle_dupeblock_first (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2425 {
2426   if (p0 > in_len) return (in_len);
2427
2428   if ((in_len + p0) >= 32) return (in_len);
2429
2430   u32 out_len = in_len;
2431
2432   u32 tib40[4];
2433   u32 tib41[4];
2434
2435   tib40[0] = buf0[0];
2436   tib40[1] = buf0[1];
2437   tib40[2] = buf0[2];
2438   tib40[3] = buf0[3];
2439   tib41[0] = buf1[0];
2440   tib41[1] = buf1[1];
2441   tib41[2] = buf1[2];
2442   tib41[3] = buf1[3];
2443
2444   truncate_right (tib40, tib41, p0);
2445
2446   rshift_block_N (buf0, buf1, buf0, buf1, p0);
2447
2448   buf0[0] |= tib40[0];
2449   buf0[1] |= tib40[1];
2450   buf0[2] |= tib40[2];
2451   buf0[3] |= tib40[3];
2452   buf1[0] |= tib41[0];
2453   buf1[1] |= tib41[1];
2454   buf1[2] |= tib41[2];
2455   buf1[3] |= tib41[3];
2456
2457   out_len += p0;
2458
2459   return out_len;
2460 }
2461
2462 static u32 rule_op_mangle_dupeblock_last (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2463 {
2464   if (p0 > in_len) return (in_len);
2465
2466   if ((in_len + p0) >= 32) return (in_len);
2467
2468   u32 out_len = in_len;
2469
2470   u32 tib40[4];
2471   u32 tib41[4];
2472
2473   rshift_block_N (buf0, buf1, tib40, tib41, p0);
2474
2475   truncate_left (tib40, tib41, out_len);
2476
2477   buf0[0] |= tib40[0];
2478   buf0[1] |= tib40[1];
2479   buf0[2] |= tib40[2];
2480   buf0[3] |= tib40[3];
2481   buf1[0] |= tib41[0];
2482   buf1[1] |= tib41[1];
2483   buf1[2] |= tib41[2];
2484   buf1[3] |= tib41[3];
2485
2486   out_len += p0;
2487
2488   return out_len;
2489 }
2490
2491 static u32 rule_op_mangle_title (const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2492 {
2493   buf0[0] |= (generate_cmask (buf0[0]));
2494   buf0[1] |= (generate_cmask (buf0[1]));
2495   buf0[2] |= (generate_cmask (buf0[2]));
2496   buf0[3] |= (generate_cmask (buf0[3]));
2497   buf1[0] |= (generate_cmask (buf1[0]));
2498   buf1[1] |= (generate_cmask (buf1[1]));
2499   buf1[2] |= (generate_cmask (buf1[2]));
2500   buf1[3] |= (generate_cmask (buf1[3]));
2501
2502   buf0[0] &= ~(0x00000020 & generate_cmask (buf0[0]));
2503
2504   for (u32 i = 0; i < in_len; i++)
2505   {
2506     u32 tmp0 = 0;
2507     u32 tmp1 = 0;
2508
2509     switch (i)
2510     {
2511       case  0:  tmp0 = (buf0[0] >>  0) & 0xFF;
2512                 tmp1 = ~(0x00002000 & generate_cmask (buf0[0])); break;
2513       case  1:  tmp0 = (buf0[0] >>  8) & 0xFF;
2514                 tmp1 = ~(0x00200000 & generate_cmask (buf0[0])); break;
2515       case  2:  tmp0 = (buf0[0] >> 16) & 0xFF;
2516                 tmp1 = ~(0x20000000 & generate_cmask (buf0[0])); break;
2517       case  3:  tmp0 = (buf0[0] >> 24) & 0xFF;
2518                 tmp1 = ~(0x00000020 & generate_cmask (buf0[1])); break;
2519       case  4:  tmp0 = (buf0[1] >>  0) & 0xFF;
2520                 tmp1 = ~(0x00002000 & generate_cmask (buf0[1])); break;
2521       case  5:  tmp0 = (buf0[1] >>  8) & 0xFF;
2522                 tmp1 = ~(0x00200000 & generate_cmask (buf0[1])); break;
2523       case  6:  tmp0 = (buf0[1] >> 16) & 0xFF;
2524                 tmp1 = ~(0x20000000 & generate_cmask (buf0[1])); break;
2525       case  7:  tmp0 = (buf0[1] >> 24) & 0xFF;
2526                 tmp1 = ~(0x00000020 & generate_cmask (buf0[2])); break;
2527       case  8:  tmp0 = (buf0[2] >>  0) & 0xFF;
2528                 tmp1 = ~(0x00002000 & generate_cmask (buf0[2])); break;
2529       case  9:  tmp0 = (buf0[2] >>  8) & 0xFF;
2530                 tmp1 = ~(0x00200000 & generate_cmask (buf0[2])); break;
2531       case 10:  tmp0 = (buf0[2] >> 16) & 0xFF;
2532                 tmp1 = ~(0x20000000 & generate_cmask (buf0[2])); break;
2533       case 11:  tmp0 = (buf0[2] >> 24) & 0xFF;
2534                 tmp1 = ~(0x00000020 & generate_cmask (buf0[3])); break;
2535       case 12:  tmp0 = (buf0[3] >>  0) & 0xFF;
2536                 tmp1 = ~(0x00002000 & generate_cmask (buf0[3])); break;
2537       case 13:  tmp0 = (buf0[3] >>  8) & 0xFF;
2538                 tmp1 = ~(0x00200000 & generate_cmask (buf0[3])); break;
2539       case 14:  tmp0 = (buf0[3] >> 16) & 0xFF;
2540                 tmp1 = ~(0x20000000 & generate_cmask (buf0[3])); break;
2541       case 15:  tmp0 = (buf0[3] >> 24) & 0xFF;
2542                 tmp1 = ~(0x00000020 & generate_cmask (buf1[0])); break;
2543       case 16:  tmp0 = (buf1[0] >>  0) & 0xFF;
2544                 tmp1 = ~(0x00002000 & generate_cmask (buf1[0])); break;
2545       case 17:  tmp0 = (buf1[0] >>  8) & 0xFF;
2546                 tmp1 = ~(0x00200000 & generate_cmask (buf1[0])); break;
2547       case 18:  tmp0 = (buf1[0] >> 16) & 0xFF;
2548                 tmp1 = ~(0x20000000 & generate_cmask (buf1[0])); break;
2549       case 19:  tmp0 = (buf1[0] >> 24) & 0xFF;
2550                 tmp1 = ~(0x00000020 & generate_cmask (buf1[1])); break;
2551       case 20:  tmp0 = (buf1[1] >>  0) & 0xFF;
2552                 tmp1 = ~(0x00002000 & generate_cmask (buf1[1])); break;
2553       case 21:  tmp0 = (buf1[1] >>  8) & 0xFF;
2554                 tmp1 = ~(0x00200000 & generate_cmask (buf1[1])); break;
2555       case 22:  tmp0 = (buf1[1] >> 16) & 0xFF;
2556                 tmp1 = ~(0x20000000 & generate_cmask (buf1[1])); break;
2557       case 23:  tmp0 = (buf1[1] >> 24) & 0xFF;
2558                 tmp1 = ~(0x00000020 & generate_cmask (buf1[2])); break;
2559       case 24:  tmp0 = (buf1[2] >>  0) & 0xFF;
2560                 tmp1 = ~(0x00002000 & generate_cmask (buf1[2])); break;
2561       case 25:  tmp0 = (buf1[2] >>  8) & 0xFF;
2562                 tmp1 = ~(0x00200000 & generate_cmask (buf1[2])); break;
2563       case 26:  tmp0 = (buf1[2] >> 16) & 0xFF;
2564                 tmp1 = ~(0x20000000 & generate_cmask (buf1[2])); break;
2565       case 27:  tmp0 = (buf1[2] >> 24) & 0xFF;
2566                 tmp1 = ~(0x00000020 & generate_cmask (buf1[3])); break;
2567       case 28:  tmp0 = (buf1[3] >>  0) & 0xFF;
2568                 tmp1 = ~(0x00002000 & generate_cmask (buf1[3])); break;
2569       case 29:  tmp0 = (buf1[3] >>  8) & 0xFF;
2570                 tmp1 = ~(0x00200000 & generate_cmask (buf1[3])); break;
2571       case 30:  tmp0 = (buf1[3] >> 16) & 0xFF;
2572                 tmp1 = ~(0x20000000 & generate_cmask (buf1[3])); break;
2573     }
2574
2575     if (i < 3)
2576     {
2577       if (tmp0   == ' ') buf0[0]   &= tmp1  ;
2578     }
2579     else if (i < 7)
2580     {
2581       if (tmp0   == ' ') buf0[1]   &= tmp1  ;
2582     }
2583     else if (i < 11)
2584     {
2585       if (tmp0   == ' ') buf0[2]   &= tmp1  ;
2586     }
2587     else if (i < 15)
2588     {
2589       if (tmp0   == ' ') buf0[3]   &= tmp1  ;
2590     }
2591     else if (i < 19)
2592     {
2593       if (tmp0   == ' ') buf1[0]   &= tmp1  ;
2594     }
2595     else if (i < 23)
2596     {
2597       if (tmp0   == ' ') buf1[1]   &= tmp1  ;
2598     }
2599     else if (i < 27)
2600     {
2601       if (tmp0   == ' ') buf1[2]   &= tmp1  ;
2602     }
2603     else if (i < 31)
2604     {
2605       if (tmp0   == ' ') buf1[3]   &= tmp1  ;
2606     }
2607   }
2608
2609   return in_len;
2610 }
2611
2612 u32 apply_rule (const u32 name, const u32 p0, const u32 p1, u32 buf0[4], u32 buf1[4], const u32 in_len)
2613 {
2614   u32 out_len = in_len;
2615
2616   switch (name)
2617   {
2618     case RULE_OP_MANGLE_LREST:            out_len = rule_op_mangle_lrest            (p0, p1, buf0, buf1, out_len); break;
2619     case RULE_OP_MANGLE_UREST:            out_len = rule_op_mangle_urest            (p0, p1, buf0, buf1, out_len); break;
2620     case RULE_OP_MANGLE_LREST_UFIRST:     out_len = rule_op_mangle_lrest_ufirst     (p0, p1, buf0, buf1, out_len); break;
2621     case RULE_OP_MANGLE_UREST_LFIRST:     out_len = rule_op_mangle_urest_lfirst     (p0, p1, buf0, buf1, out_len); break;
2622     case RULE_OP_MANGLE_TREST:            out_len = rule_op_mangle_trest            (p0, p1, buf0, buf1, out_len); break;
2623     case RULE_OP_MANGLE_TOGGLE_AT:        out_len = rule_op_mangle_toggle_at        (p0, p1, buf0, buf1, out_len); break;
2624     case RULE_OP_MANGLE_REVERSE:          out_len = rule_op_mangle_reverse          (p0, p1, buf0, buf1, out_len); break;
2625     case RULE_OP_MANGLE_DUPEWORD:         out_len = rule_op_mangle_dupeword         (p0, p1, buf0, buf1, out_len); break;
2626     case RULE_OP_MANGLE_DUPEWORD_TIMES:   out_len = rule_op_mangle_dupeword_times   (p0, p1, buf0, buf1, out_len); break;
2627     case RULE_OP_MANGLE_REFLECT:          out_len = rule_op_mangle_reflect          (p0, p1, buf0, buf1, out_len); break;
2628     case RULE_OP_MANGLE_APPEND:           out_len = rule_op_mangle_append           (p0, p1, buf0, buf1, out_len); break;
2629     case RULE_OP_MANGLE_PREPEND:          out_len = rule_op_mangle_prepend          (p0, p1, buf0, buf1, out_len); break;
2630     case RULE_OP_MANGLE_ROTATE_LEFT:      out_len = rule_op_mangle_rotate_left      (p0, p1, buf0, buf1, out_len); break;
2631     case RULE_OP_MANGLE_ROTATE_RIGHT:     out_len = rule_op_mangle_rotate_right     (p0, p1, buf0, buf1, out_len); break;
2632     case RULE_OP_MANGLE_DELETE_FIRST:     out_len = rule_op_mangle_delete_first     (p0, p1, buf0, buf1, out_len); break;
2633     case RULE_OP_MANGLE_DELETE_LAST:      out_len = rule_op_mangle_delete_last      (p0, p1, buf0, buf1, out_len); break;
2634     case RULE_OP_MANGLE_DELETE_AT:        out_len = rule_op_mangle_delete_at        (p0, p1, buf0, buf1, out_len); break;
2635     case RULE_OP_MANGLE_EXTRACT:          out_len = rule_op_mangle_extract          (p0, p1, buf0, buf1, out_len); break;
2636     case RULE_OP_MANGLE_OMIT:             out_len = rule_op_mangle_omit             (p0, p1, buf0, buf1, out_len); break;
2637     case RULE_OP_MANGLE_INSERT:           out_len = rule_op_mangle_insert           (p0, p1, buf0, buf1, out_len); break;
2638     case RULE_OP_MANGLE_OVERSTRIKE:       out_len = rule_op_mangle_overstrike       (p0, p1, buf0, buf1, out_len); break;
2639     case RULE_OP_MANGLE_TRUNCATE_AT:      out_len = rule_op_mangle_truncate_at      (p0, p1, buf0, buf1, out_len); break;
2640     case RULE_OP_MANGLE_REPLACE:          out_len = rule_op_mangle_replace          (p0, p1, buf0, buf1, out_len); break;
2641     //case RULE_OP_MANGLE_PURGECHAR:        out_len = rule_op_mangle_purgechar        (p0, p1, buf0, buf1, out_len); break;
2642     //case RULE_OP_MANGLE_TOGGLECASE_REC:   out_len = rule_op_mangle_togglecase_rec   (p0, p1, buf0, buf1, out_len); break;
2643     case RULE_OP_MANGLE_DUPECHAR_FIRST:   out_len = rule_op_mangle_dupechar_first   (p0, p1, buf0, buf1, out_len); break;
2644     case RULE_OP_MANGLE_DUPECHAR_LAST:    out_len = rule_op_mangle_dupechar_last    (p0, p1, buf0, buf1, out_len); break;
2645     case RULE_OP_MANGLE_DUPECHAR_ALL:     out_len = rule_op_mangle_dupechar_all     (p0, p1, buf0, buf1, out_len); break;
2646     case RULE_OP_MANGLE_SWITCH_FIRST:     out_len = rule_op_mangle_switch_first     (p0, p1, buf0, buf1, out_len); break;
2647     case RULE_OP_MANGLE_SWITCH_LAST:      out_len = rule_op_mangle_switch_last      (p0, p1, buf0, buf1, out_len); break;
2648     case RULE_OP_MANGLE_SWITCH_AT:        out_len = rule_op_mangle_switch_at        (p0, p1, buf0, buf1, out_len); break;
2649     case RULE_OP_MANGLE_CHR_SHIFTL:       out_len = rule_op_mangle_chr_shiftl       (p0, p1, buf0, buf1, out_len); break;
2650     case RULE_OP_MANGLE_CHR_SHIFTR:       out_len = rule_op_mangle_chr_shiftr       (p0, p1, buf0, buf1, out_len); break;
2651     case RULE_OP_MANGLE_CHR_INCR:         out_len = rule_op_mangle_chr_incr         (p0, p1, buf0, buf1, out_len); break;
2652     case RULE_OP_MANGLE_CHR_DECR:         out_len = rule_op_mangle_chr_decr         (p0, p1, buf0, buf1, out_len); break;
2653     case RULE_OP_MANGLE_REPLACE_NP1:      out_len = rule_op_mangle_replace_np1      (p0, p1, buf0, buf1, out_len); break;
2654     case RULE_OP_MANGLE_REPLACE_NM1:      out_len = rule_op_mangle_replace_nm1      (p0, p1, buf0, buf1, out_len); break;
2655     case RULE_OP_MANGLE_DUPEBLOCK_FIRST:  out_len = rule_op_mangle_dupeblock_first  (p0, p1, buf0, buf1, out_len); break;
2656     case RULE_OP_MANGLE_DUPEBLOCK_LAST:   out_len = rule_op_mangle_dupeblock_last   (p0, p1, buf0, buf1, out_len); break;
2657     case RULE_OP_MANGLE_TITLE:            out_len = rule_op_mangle_title            (p0, p1, buf0, buf1, out_len); break;
2658   }
2659
2660   return out_len;
2661 }
2662
2663 u32 apply_rules (u32 *cmds, u32 buf0[4], u32 buf1[4], const u32 len)
2664 {
2665   u32 out_len = len;
2666
2667   for (u32 i = 0; cmds[i] != 0; i++)
2668   {
2669     const u32 cmd = cmds[i];
2670
2671     const u32 name = (cmd >>  0) & 0xff;
2672     const u32 p0   = (cmd >>  8) & 0xff;
2673     const u32 p1   = (cmd >> 16) & 0xff;
2674
2675     out_len = apply_rule (name, p0, p1, buf0, buf1, out_len);
2676   }
2677
2678   return out_len;
2679 }