drivers/staging/skein/threefish1024Block.c

   1 #include <linux/string.h>
   2 #include <threefishApi.h>
   3
   4
   5 void threefishEncrypt1024(struct threefish_key *keyCtx, u64 *input, u64 *output)
   6 {
   7         u64 b0 = input[0], b1 = input[1],
   8           b2 = input[2], b3 = input[3],
   9           b4 = input[4], b5 = input[5],
  10           b6 = input[6], b7 = input[7],
  11           b8 = input[8], b9 = input[9],
  12           b10 = input[10], b11 = input[11],
  13           b12 = input[12], b13 = input[13],
  14           b14 = input[14], b15 = input[15];
  15         u64 k0 = keyCtx->key[0], k1 = keyCtx->key[1],
  16           k2 = keyCtx->key[2], k3 = keyCtx->key[3],
  17           k4 = keyCtx->key[4], k5 = keyCtx->key[5],
  18           k6 = keyCtx->key[6], k7 = keyCtx->key[7],
  19           k8 = keyCtx->key[8], k9 = keyCtx->key[9],
  20           k10 = keyCtx->key[10], k11 = keyCtx->key[11],
  21           k12 = keyCtx->key[12], k13 = keyCtx->key[13],
  22           k14 = keyCtx->key[14], k15 = keyCtx->key[15],
  23           k16 = keyCtx->key[16];
  24         u64 t0 = keyCtx->tweak[0], t1 = keyCtx->tweak[1],
  25           t2 = keyCtx->tweak[2];
  26
  27         b1 += k1;
  28         b0 += b1 + k0;
  29         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
  30
  31         b3 += k3;
  32         b2 += b3 + k2;
  33         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
  34
  35         b5 += k5;
  36         b4 += b5 + k4;
  37         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
  38
  39         b7 += k7;
  40         b6 += b7 + k6;
  41         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
  42
  43         b9 += k9;
  44         b8 += b9 + k8;
  45         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
  46
  47         b11 += k11;
  48         b10 += b11 + k10;
  49         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
  50
  51         b13 += k13 + t0;
  52         b12 += b13 + k12;
  53         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
  54
  55         b15 += k15;
  56         b14 += b15 + k14 + t1;
  57         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
  58
  59         b0 += b9;
  60         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
  61
  62         b2 += b13;
  63         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
  64
  65         b6 += b11;
  66         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
  67
  68         b4 += b15;
  69         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
  70
  71         b10 += b7;
  72         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
  73
  74         b12 += b3;
  75         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
  76
  77         b14 += b5;
  78         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
  79
  80         b8 += b1;
  81         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
  82
  83         b0 += b7;
  84         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
  85
  86         b2 += b5;
  87         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
  88
  89         b4 += b3;
  90         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
  91
  92         b6 += b1;
  93         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
  94
  95         b12 += b15;
  96         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
  97
  98         b14 += b13;
  99         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
 100
 101         b8 += b11;
 102         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
 103
 104         b10 += b9;
 105         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
 106
 107         b0 += b15;
 108         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
 109
 110         b2 += b11;
 111         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
 112
 113         b6 += b13;
 114         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
 115
 116         b4 += b9;
 117         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
 118
 119         b14 += b1;
 120         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
 121
 122         b8 += b5;
 123         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
 124
 125         b10 += b3;
 126         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
 127
 128         b12 += b7;
 129         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
 130
 131         b1 += k2;
 132         b0 += b1 + k1;
 133         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
 134
 135         b3 += k4;
 136         b2 += b3 + k3;
 137         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
 138
 139         b5 += k6;
 140         b4 += b5 + k5;
 141         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
 142
 143         b7 += k8;
 144         b6 += b7 + k7;
 145         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
 146
 147         b9 += k10;
 148         b8 += b9 + k9;
 149         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
 150
 151         b11 += k12;
 152         b10 += b11 + k11;
 153         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
 154
 155         b13 += k14 + t1;
 156         b12 += b13 + k13;
 157         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
 158
 159         b15 += k16 + 1;
 160         b14 += b15 + k15 + t2;
 161         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
 162
 163         b0 += b9;
 164         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
 165
 166         b2 += b13;
 167         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
 168
 169         b6 += b11;
 170         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
 171
 172         b4 += b15;
 173         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
 174
 175         b10 += b7;
 176         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
 177
 178         b12 += b3;
 179         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
 180
 181         b14 += b5;
 182         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
 183
 184         b8 += b1;
 185         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
 186
 187         b0 += b7;
 188         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
 189
 190         b2 += b5;
 191         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
 192
 193         b4 += b3;
 194         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
 195
 196         b6 += b1;
 197         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
 198
 199         b12 += b15;
 200         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
 201
 202         b14 += b13;
 203         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
 204
 205         b8 += b11;
 206         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
 207
 208         b10 += b9;
 209         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
 210
 211         b0 += b15;
 212         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
 213
 214         b2 += b11;
 215         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
 216
 217         b6 += b13;
 218         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
 219
 220         b4 += b9;
 221         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
 222
 223         b14 += b1;
 224         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
 225
 226         b8 += b5;
 227         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
 228
 229         b10 += b3;
 230         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
 231
 232         b12 += b7;
 233         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
 234
 235         b1 += k3;
 236         b0 += b1 + k2;
 237         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
 238
 239         b3 += k5;
 240         b2 += b3 + k4;
 241         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
 242
 243         b5 += k7;
 244         b4 += b5 + k6;
 245         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
 246
 247         b7 += k9;
 248         b6 += b7 + k8;
 249         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
 250
 251         b9 += k11;
 252         b8 += b9 + k10;
 253         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
 254
 255         b11 += k13;
 256         b10 += b11 + k12;
 257         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
 258
 259         b13 += k15 + t2;
 260         b12 += b13 + k14;
 261         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
 262
 263         b15 += k0 + 2;
 264         b14 += b15 + k16 + t0;
 265         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
 266
 267         b0 += b9;
 268         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
 269
 270         b2 += b13;
 271         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
 272
 273         b6 += b11;
 274         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
 275
 276         b4 += b15;
 277         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
 278
 279         b10 += b7;
 280         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
 281
 282         b12 += b3;
 283         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
 284
 285         b14 += b5;
 286         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
 287
 288         b8 += b1;
 289         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
 290
 291         b0 += b7;
 292         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
 293
 294         b2 += b5;
 295         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
 296
 297         b4 += b3;
 298         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
 299
 300         b6 += b1;
 301         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
 302
 303         b12 += b15;
 304         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
 305
 306         b14 += b13;
 307         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
 308
 309         b8 += b11;
 310         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
 311
 312         b10 += b9;
 313         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
 314
 315         b0 += b15;
 316         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
 317
 318         b2 += b11;
 319         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
 320
 321         b6 += b13;
 322         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
 323
 324         b4 += b9;
 325         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
 326
 327         b14 += b1;
 328         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
 329
 330         b8 += b5;
 331         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
 332
 333         b10 += b3;
 334         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
 335
 336         b12 += b7;
 337         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
 338
 339         b1 += k4;
 340         b0 += b1 + k3;
 341         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
 342
 343         b3 += k6;
 344         b2 += b3 + k5;
 345         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
 346
 347         b5 += k8;
 348         b4 += b5 + k7;
 349         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
 350
 351         b7 += k10;
 352         b6 += b7 + k9;
 353         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
 354
 355         b9 += k12;
 356         b8 += b9 + k11;
 357         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
 358
 359         b11 += k14;
 360         b10 += b11 + k13;
 361         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
 362
 363         b13 += k16 + t0;
 364         b12 += b13 + k15;
 365         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
 366
 367         b15 += k1 + 3;
 368         b14 += b15 + k0 + t1;
 369         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
 370
 371         b0 += b9;
 372         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
 373
 374         b2 += b13;
 375         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
 376
 377         b6 += b11;
 378         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
 379
 380         b4 += b15;
 381         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
 382
 383         b10 += b7;
 384         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
 385
 386         b12 += b3;
 387         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
 388
 389         b14 += b5;
 390         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
 391
 392         b8 += b1;
 393         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
 394
 395         b0 += b7;
 396         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
 397
 398         b2 += b5;
 399         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
 400
 401         b4 += b3;
 402         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
 403
 404         b6 += b1;
 405         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
 406
 407         b12 += b15;
 408         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
 409
 410         b14 += b13;
 411         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
 412
 413         b8 += b11;
 414         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
 415
 416         b10 += b9;
 417         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
 418
 419         b0 += b15;
 420         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
 421
 422         b2 += b11;
 423         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
 424
 425         b6 += b13;
 426         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
 427
 428         b4 += b9;
 429         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
 430
 431         b14 += b1;
 432         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
 433
 434         b8 += b5;
 435         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
 436
 437         b10 += b3;
 438         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
 439
 440         b12 += b7;
 441         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
 442
 443         b1 += k5;
 444         b0 += b1 + k4;
 445         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
 446
 447         b3 += k7;
 448         b2 += b3 + k6;
 449         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
 450
 451         b5 += k9;
 452         b4 += b5 + k8;
 453         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
 454
 455         b7 += k11;
 456         b6 += b7 + k10;
 457         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
 458
 459         b9 += k13;
 460         b8 += b9 + k12;
 461         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
 462
 463         b11 += k15;
 464         b10 += b11 + k14;
 465         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
 466
 467         b13 += k0 + t1;
 468         b12 += b13 + k16;
 469         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
 470
 471         b15 += k2 + 4;
 472         b14 += b15 + k1 + t2;
 473         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
 474
 475         b0 += b9;
 476         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
 477
 478         b2 += b13;
 479         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
 480
 481         b6 += b11;
 482         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
 483
 484         b4 += b15;
 485         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
 486
 487         b10 += b7;
 488         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
 489
 490         b12 += b3;
 491         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
 492
 493         b14 += b5;
 494         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
 495
 496         b8 += b1;
 497         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
 498
 499         b0 += b7;
 500         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
 501
 502         b2 += b5;
 503         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
 504
 505         b4 += b3;
 506         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
 507
 508         b6 += b1;
 509         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
 510
 511         b12 += b15;
 512         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
 513
 514         b14 += b13;
 515         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
 516
 517         b8 += b11;
 518         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
 519
 520         b10 += b9;
 521         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
 522
 523         b0 += b15;
 524         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
 525
 526         b2 += b11;
 527         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
 528
 529         b6 += b13;
 530         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
 531
 532         b4 += b9;
 533         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
 534
 535         b14 += b1;
 536         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
 537
 538         b8 += b5;
 539         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
 540
 541         b10 += b3;
 542         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
 543
 544         b12 += b7;
 545         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
 546
 547         b1 += k6;
 548         b0 += b1 + k5;
 549         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
 550
 551         b3 += k8;
 552         b2 += b3 + k7;
 553         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
 554
 555         b5 += k10;
 556         b4 += b5 + k9;
 557         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
 558
 559         b7 += k12;
 560         b6 += b7 + k11;
 561         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
 562
 563         b9 += k14;
 564         b8 += b9 + k13;
 565         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
 566
 567         b11 += k16;
 568         b10 += b11 + k15;
 569         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
 570
 571         b13 += k1 + t2;
 572         b12 += b13 + k0;
 573         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
 574
 575         b15 += k3 + 5;
 576         b14 += b15 + k2 + t0;
 577         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
 578
 579         b0 += b9;
 580         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
 581
 582         b2 += b13;
 583         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
 584
 585         b6 += b11;
 586         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
 587
 588         b4 += b15;
 589         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
 590
 591         b10 += b7;
 592         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
 593
 594         b12 += b3;
 595         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
 596
 597         b14 += b5;
 598         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
 599
 600         b8 += b1;
 601         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
 602
 603         b0 += b7;
 604         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
 605
 606         b2 += b5;
 607         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
 608
 609         b4 += b3;
 610         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
 611
 612         b6 += b1;
 613         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
 614
 615         b12 += b15;
 616         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
 617
 618         b14 += b13;
 619         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
 620
 621         b8 += b11;
 622         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
 623
 624         b10 += b9;
 625         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
 626
 627         b0 += b15;
 628         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
 629
 630         b2 += b11;
 631         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
 632
 633         b6 += b13;
 634         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
 635
 636         b4 += b9;
 637         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
 638
 639         b14 += b1;
 640         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
 641
 642         b8 += b5;
 643         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
 644
 645         b10 += b3;
 646         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
 647
 648         b12 += b7;
 649         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
 650
 651         b1 += k7;
 652         b0 += b1 + k6;
 653         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
 654
 655         b3 += k9;
 656         b2 += b3 + k8;
 657         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
 658
 659         b5 += k11;
 660         b4 += b5 + k10;
 661         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
 662
 663         b7 += k13;
 664         b6 += b7 + k12;
 665         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
 666
 667         b9 += k15;
 668         b8 += b9 + k14;
 669         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
 670
 671         b11 += k0;
 672         b10 += b11 + k16;
 673         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
 674
 675         b13 += k2 + t0;
 676         b12 += b13 + k1;
 677         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
 678
 679         b15 += k4 + 6;
 680         b14 += b15 + k3 + t1;
 681         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
 682
 683         b0 += b9;
 684         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
 685
 686         b2 += b13;
 687         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
 688
 689         b6 += b11;
 690         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
 691
 692         b4 += b15;
 693         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
 694
 695         b10 += b7;
 696         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
 697
 698         b12 += b3;
 699         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
 700
 701         b14 += b5;
 702         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
 703
 704         b8 += b1;
 705         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
 706
 707         b0 += b7;
 708         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
 709
 710         b2 += b5;
 711         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
 712
 713         b4 += b3;
 714         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
 715
 716         b6 += b1;
 717         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
 718
 719         b12 += b15;
 720         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
 721
 722         b14 += b13;
 723         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
 724
 725         b8 += b11;
 726         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
 727
 728         b10 += b9;
 729         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
 730
 731         b0 += b15;
 732         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
 733
 734         b2 += b11;
 735         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
 736
 737         b6 += b13;
 738         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
 739
 740         b4 += b9;
 741         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
 742
 743         b14 += b1;
 744         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
 745
 746         b8 += b5;
 747         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
 748
 749         b10 += b3;
 750         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
 751
 752         b12 += b7;
 753         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
 754
 755         b1 += k8;
 756         b0 += b1 + k7;
 757         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
 758
 759         b3 += k10;
 760         b2 += b3 + k9;
 761         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
 762
 763         b5 += k12;
 764         b4 += b5 + k11;
 765         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
 766
 767         b7 += k14;
 768         b6 += b7 + k13;
 769         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
 770
 771         b9 += k16;
 772         b8 += b9 + k15;
 773         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
 774
 775         b11 += k1;
 776         b10 += b11 + k0;
 777         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
 778
 779         b13 += k3 + t1;
 780         b12 += b13 + k2;
 781         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
 782
 783         b15 += k5 + 7;
 784         b14 += b15 + k4 + t2;
 785         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
 786
 787         b0 += b9;
 788         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
 789
 790         b2 += b13;
 791         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
 792
 793         b6 += b11;
 794         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
 795
 796         b4 += b15;
 797         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
 798
 799         b10 += b7;
 800         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
 801
 802         b12 += b3;
 803         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
 804
 805         b14 += b5;
 806         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
 807
 808         b8 += b1;
 809         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
 810
 811         b0 += b7;
 812         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
 813
 814         b2 += b5;
 815         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
 816
 817         b4 += b3;
 818         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
 819
 820         b6 += b1;
 821         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
 822
 823         b12 += b15;
 824         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
 825
 826         b14 += b13;
 827         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
 828
 829         b8 += b11;
 830         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
 831
 832         b10 += b9;
 833         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
 834
 835         b0 += b15;
 836         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
 837
 838         b2 += b11;
 839         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
 840
 841         b6 += b13;
 842         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
 843
 844         b4 += b9;
 845         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
 846
 847         b14 += b1;
 848         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
 849
 850         b8 += b5;
 851         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
 852
 853         b10 += b3;
 854         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
 855
 856         b12 += b7;
 857         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
 858
 859         b1 += k9;
 860         b0 += b1 + k8;
 861         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
 862
 863         b3 += k11;
 864         b2 += b3 + k10;
 865         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
 866
 867         b5 += k13;
 868         b4 += b5 + k12;
 869         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
 870
 871         b7 += k15;
 872         b6 += b7 + k14;
 873         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
 874
 875         b9 += k0;
 876         b8 += b9 + k16;
 877         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
 878
 879         b11 += k2;
 880         b10 += b11 + k1;
 881         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
 882
 883         b13 += k4 + t2;
 884         b12 += b13 + k3;
 885         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
 886
 887         b15 += k6 + 8;
 888         b14 += b15 + k5 + t0;
 889         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
 890
 891         b0 += b9;
 892         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
 893
 894         b2 += b13;
 895         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
 896
 897         b6 += b11;
 898         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
 899
 900         b4 += b15;
 901         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
 902
 903         b10 += b7;
 904         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
 905
 906         b12 += b3;
 907         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
 908
 909         b14 += b5;
 910         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
 911
 912         b8 += b1;
 913         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
 914
 915         b0 += b7;
 916         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
 917
 918         b2 += b5;
 919         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
 920
 921         b4 += b3;
 922         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
 923
 924         b6 += b1;
 925         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
 926
 927         b12 += b15;
 928         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
 929
 930         b14 += b13;
 931         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
 932
 933         b8 += b11;
 934         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
 935
 936         b10 += b9;
 937         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
 938
 939         b0 += b15;
 940         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
 941
 942         b2 += b11;
 943         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
 944
 945         b6 += b13;
 946         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
 947
 948         b4 += b9;
 949         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
 950
 951         b14 += b1;
 952         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
 953
 954         b8 += b5;
 955         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
 956
 957         b10 += b3;
 958         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
 959
 960         b12 += b7;
 961         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
 962
 963         b1 += k10;
 964         b0 += b1 + k9;
 965         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
 966
 967         b3 += k12;
 968         b2 += b3 + k11;
 969         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
 970
 971         b5 += k14;
 972         b4 += b5 + k13;
 973         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
 974
 975         b7 += k16;
 976         b6 += b7 + k15;
 977         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
 978
 979         b9 += k1;
 980         b8 += b9 + k0;
 981         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
 982
 983         b11 += k3;
 984         b10 += b11 + k2;
 985         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
 986
 987         b13 += k5 + t0;
 988         b12 += b13 + k4;
 989         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
 990
 991         b15 += k7 + 9;
 992         b14 += b15 + k6 + t1;
 993         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
 994
 995         b0 += b9;
 996         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
 997
 998         b2 += b13;
 999         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
1000
1001         b6 += b11;
1002         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
1003
1004         b4 += b15;
1005         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
1006
1007         b10 += b7;
1008         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
1009
1010         b12 += b3;
1011         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
1012
1013         b14 += b5;
1014         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
1015
1016         b8 += b1;
1017         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
1018
1019         b0 += b7;
1020         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
1021
1022         b2 += b5;
1023         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
1024
1025         b4 += b3;
1026         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
1027
1028         b6 += b1;
1029         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
1030
1031         b12 += b15;
1032         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
1033
1034         b14 += b13;
1035         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
1036
1037         b8 += b11;
1038         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
1039
1040         b10 += b9;
1041         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
1042
1043         b0 += b15;
1044         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
1045
1046         b2 += b11;
1047         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
1048
1049         b6 += b13;
1050         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
1051
1052         b4 += b9;
1053         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
1054
1055         b14 += b1;
1056         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
1057
1058         b8 += b5;
1059         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
1060
1061         b10 += b3;
1062         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
1063
1064         b12 += b7;
1065         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
1066
1067         b1 += k11;
1068         b0 += b1 + k10;
1069         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
1070
1071         b3 += k13;
1072         b2 += b3 + k12;
1073         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
1074
1075         b5 += k15;
1076         b4 += b5 + k14;
1077         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
1078
1079         b7 += k0;
1080         b6 += b7 + k16;
1081         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
1082
1083         b9 += k2;
1084         b8 += b9 + k1;
1085         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
1086
1087         b11 += k4;
1088         b10 += b11 + k3;
1089         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
1090
1091         b13 += k6 + t1;
1092         b12 += b13 + k5;
1093         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
1094
1095         b15 += k8 + 10;
1096         b14 += b15 + k7 + t2;
1097         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
1098
1099         b0 += b9;
1100         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
1101
1102         b2 += b13;
1103         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
1104
1105         b6 += b11;
1106         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
1107
1108         b4 += b15;
1109         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
1110
1111         b10 += b7;
1112         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
1113
1114         b12 += b3;
1115         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
1116
1117         b14 += b5;
1118         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
1119
1120         b8 += b1;
1121         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
1122
1123         b0 += b7;
1124         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
1125
1126         b2 += b5;
1127         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
1128
1129         b4 += b3;
1130         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
1131
1132         b6 += b1;
1133         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
1134
1135         b12 += b15;
1136         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
1137
1138         b14 += b13;
1139         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
1140
1141         b8 += b11;
1142         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
1143
1144         b10 += b9;
1145         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
1146
1147         b0 += b15;
1148         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
1149
1150         b2 += b11;
1151         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
1152
1153         b6 += b13;
1154         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
1155
1156         b4 += b9;
1157         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
1158
1159         b14 += b1;
1160         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
1161
1162         b8 += b5;
1163         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
1164
1165         b10 += b3;
1166         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
1167
1168         b12 += b7;
1169         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
1170
1171         b1 += k12;
1172         b0 += b1 + k11;
1173         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
1174
1175         b3 += k14;
1176         b2 += b3 + k13;
1177         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
1178
1179         b5 += k16;
1180         b4 += b5 + k15;
1181         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
1182
1183         b7 += k1;
1184         b6 += b7 + k0;
1185         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
1186
1187         b9 += k3;
1188         b8 += b9 + k2;
1189         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
1190
1191         b11 += k5;
1192         b10 += b11 + k4;
1193         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
1194
1195         b13 += k7 + t2;
1196         b12 += b13 + k6;
1197         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
1198
1199         b15 += k9 + 11;
1200         b14 += b15 + k8 + t0;
1201         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
1202
1203         b0 += b9;
1204         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
1205
1206         b2 += b13;
1207         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
1208
1209         b6 += b11;
1210         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
1211
1212         b4 += b15;
1213         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
1214
1215         b10 += b7;
1216         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
1217
1218         b12 += b3;
1219         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
1220
1221         b14 += b5;
1222         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
1223
1224         b8 += b1;
1225         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
1226
1227         b0 += b7;
1228         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
1229
1230         b2 += b5;
1231         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
1232
1233         b4 += b3;
1234         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
1235
1236         b6 += b1;
1237         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
1238
1239         b12 += b15;
1240         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
1241
1242         b14 += b13;
1243         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
1244
1245         b8 += b11;
1246         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
1247
1248         b10 += b9;
1249         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
1250
1251         b0 += b15;
1252         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
1253
1254         b2 += b11;
1255         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
1256
1257         b6 += b13;
1258         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
1259
1260         b4 += b9;
1261         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
1262
1263         b14 += b1;
1264         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
1265
1266         b8 += b5;
1267         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
1268
1269         b10 += b3;
1270         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
1271
1272         b12 += b7;
1273         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
1274
1275         b1 += k13;
1276         b0 += b1 + k12;
1277         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
1278
1279         b3 += k15;
1280         b2 += b3 + k14;
1281         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
1282
1283         b5 += k0;
1284         b4 += b5 + k16;
1285         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
1286
1287         b7 += k2;
1288         b6 += b7 + k1;
1289         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
1290
1291         b9 += k4;
1292         b8 += b9 + k3;
1293         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
1294
1295         b11 += k6;
1296         b10 += b11 + k5;
1297         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
1298
1299         b13 += k8 + t0;
1300         b12 += b13 + k7;
1301         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
1302
1303         b15 += k10 + 12;
1304         b14 += b15 + k9 + t1;
1305         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
1306
1307         b0 += b9;
1308         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
1309
1310         b2 += b13;
1311         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
1312
1313         b6 += b11;
1314         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
1315
1316         b4 += b15;
1317         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
1318
1319         b10 += b7;
1320         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
1321
1322         b12 += b3;
1323         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
1324
1325         b14 += b5;
1326         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
1327
1328         b8 += b1;
1329         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
1330
1331         b0 += b7;
1332         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
1333
1334         b2 += b5;
1335         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
1336
1337         b4 += b3;
1338         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
1339
1340         b6 += b1;
1341         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
1342
1343         b12 += b15;
1344         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
1345
1346         b14 += b13;
1347         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
1348
1349         b8 += b11;
1350         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
1351
1352         b10 += b9;
1353         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
1354
1355         b0 += b15;
1356         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
1357
1358         b2 += b11;
1359         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
1360
1361         b6 += b13;
1362         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
1363
1364         b4 += b9;
1365         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
1366
1367         b14 += b1;
1368         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
1369
1370         b8 += b5;
1371         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
1372
1373         b10 += b3;
1374         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
1375
1376         b12 += b7;
1377         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
1378
1379         b1 += k14;
1380         b0 += b1 + k13;
1381         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
1382
1383         b3 += k16;
1384         b2 += b3 + k15;
1385         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
1386
1387         b5 += k1;
1388         b4 += b5 + k0;
1389         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
1390
1391         b7 += k3;
1392         b6 += b7 + k2;
1393         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
1394
1395         b9 += k5;
1396         b8 += b9 + k4;
1397         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
1398
1399         b11 += k7;
1400         b10 += b11 + k6;
1401         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
1402
1403         b13 += k9 + t1;
1404         b12 += b13 + k8;
1405         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
1406
1407         b15 += k11 + 13;
1408         b14 += b15 + k10 + t2;
1409         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
1410
1411         b0 += b9;
1412         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
1413
1414         b2 += b13;
1415         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
1416
1417         b6 += b11;
1418         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
1419
1420         b4 += b15;
1421         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
1422
1423         b10 += b7;
1424         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
1425
1426         b12 += b3;
1427         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
1428
1429         b14 += b5;
1430         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
1431
1432         b8 += b1;
1433         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
1434
1435         b0 += b7;
1436         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
1437
1438         b2 += b5;
1439         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
1440
1441         b4 += b3;
1442         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
1443
1444         b6 += b1;
1445         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
1446
1447         b12 += b15;
1448         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
1449
1450         b14 += b13;
1451         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
1452
1453         b8 += b11;
1454         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
1455
1456         b10 += b9;
1457         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
1458
1459         b0 += b15;
1460         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
1461
1462         b2 += b11;
1463         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
1464
1465         b6 += b13;
1466         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
1467
1468         b4 += b9;
1469         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
1470
1471         b14 += b1;
1472         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
1473
1474         b8 += b5;
1475         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
1476
1477         b10 += b3;
1478         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
1479
1480         b12 += b7;
1481         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
1482
1483         b1 += k15;
1484         b0 += b1 + k14;
1485         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
1486
1487         b3 += k0;
1488         b2 += b3 + k16;
1489         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
1490
1491         b5 += k2;
1492         b4 += b5 + k1;
1493         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
1494
1495         b7 += k4;
1496         b6 += b7 + k3;
1497         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
1498
1499         b9 += k6;
1500         b8 += b9 + k5;
1501         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
1502
1503         b11 += k8;
1504         b10 += b11 + k7;
1505         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
1506
1507         b13 += k10 + t2;
1508         b12 += b13 + k9;
1509         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
1510
1511         b15 += k12 + 14;
1512         b14 += b15 + k11 + t0;
1513         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
1514
1515         b0 += b9;
1516         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
1517
1518         b2 += b13;
1519         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
1520
1521         b6 += b11;
1522         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
1523
1524         b4 += b15;
1525         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
1526
1527         b10 += b7;
1528         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
1529
1530         b12 += b3;
1531         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
1532
1533         b14 += b5;
1534         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
1535
1536         b8 += b1;
1537         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
1538
1539         b0 += b7;
1540         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
1541
1542         b2 += b5;
1543         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
1544
1545         b4 += b3;
1546         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
1547
1548         b6 += b1;
1549         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
1550
1551         b12 += b15;
1552         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
1553
1554         b14 += b13;
1555         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
1556
1557         b8 += b11;
1558         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
1559
1560         b10 += b9;
1561         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
1562
1563         b0 += b15;
1564         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
1565
1566         b2 += b11;
1567         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
1568
1569         b6 += b13;
1570         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
1571
1572         b4 += b9;
1573         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
1574
1575         b14 += b1;
1576         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
1577
1578         b8 += b5;
1579         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
1580
1581         b10 += b3;
1582         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
1583
1584         b12 += b7;
1585         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
1586
1587         b1 += k16;
1588         b0 += b1 + k15;
1589         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
1590
1591         b3 += k1;
1592         b2 += b3 + k0;
1593         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
1594
1595         b5 += k3;
1596         b4 += b5 + k2;
1597         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
1598
1599         b7 += k5;
1600         b6 += b7 + k4;
1601         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
1602
1603         b9 += k7;
1604         b8 += b9 + k6;
1605         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
1606
1607         b11 += k9;
1608         b10 += b11 + k8;
1609         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
1610
1611         b13 += k11 + t0;
1612         b12 += b13 + k10;
1613         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
1614
1615         b15 += k13 + 15;
1616         b14 += b15 + k12 + t1;
1617         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
1618
1619         b0 += b9;
1620         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
1621
1622         b2 += b13;
1623         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
1624
1625         b6 += b11;
1626         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
1627
1628         b4 += b15;
1629         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
1630
1631         b10 += b7;
1632         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
1633
1634         b12 += b3;
1635         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
1636
1637         b14 += b5;
1638         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
1639
1640         b8 += b1;
1641         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
1642
1643         b0 += b7;
1644         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
1645
1646         b2 += b5;
1647         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
1648
1649         b4 += b3;
1650         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
1651
1652         b6 += b1;
1653         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
1654
1655         b12 += b15;
1656         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
1657
1658         b14 += b13;
1659         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
1660
1661         b8 += b11;
1662         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
1663
1664         b10 += b9;
1665         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
1666
1667         b0 += b15;
1668         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
1669
1670         b2 += b11;
1671         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
1672
1673         b6 += b13;
1674         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
1675
1676         b4 += b9;
1677         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
1678
1679         b14 += b1;
1680         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
1681
1682         b8 += b5;
1683         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
1684
1685         b10 += b3;
1686         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
1687
1688         b12 += b7;
1689         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
1690
1691         b1 += k0;
1692         b0 += b1 + k16;
1693         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
1694
1695         b3 += k2;
1696         b2 += b3 + k1;
1697         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
1698
1699         b5 += k4;
1700         b4 += b5 + k3;
1701         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
1702
1703         b7 += k6;
1704         b6 += b7 + k5;
1705         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
1706
1707         b9 += k8;
1708         b8 += b9 + k7;
1709         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
1710
1711         b11 += k10;
1712         b10 += b11 + k9;
1713         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
1714
1715         b13 += k12 + t1;
1716         b12 += b13 + k11;
1717         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
1718
1719         b15 += k14 + 16;
1720         b14 += b15 + k13 + t2;
1721         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
1722
1723         b0 += b9;
1724         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
1725
1726         b2 += b13;
1727         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
1728
1729         b6 += b11;
1730         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
1731
1732         b4 += b15;
1733         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
1734
1735         b10 += b7;
1736         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
1737
1738         b12 += b3;
1739         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
1740
1741         b14 += b5;
1742         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
1743
1744         b8 += b1;
1745         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
1746
1747         b0 += b7;
1748         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
1749
1750         b2 += b5;
1751         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
1752
1753         b4 += b3;
1754         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
1755
1756         b6 += b1;
1757         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
1758
1759         b12 += b15;
1760         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
1761
1762         b14 += b13;
1763         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
1764
1765         b8 += b11;
1766         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
1767
1768         b10 += b9;
1769         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
1770
1771         b0 += b15;
1772         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
1773
1774         b2 += b11;
1775         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
1776
1777         b6 += b13;
1778         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
1779
1780         b4 += b9;
1781         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
1782
1783         b14 += b1;
1784         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
1785
1786         b8 += b5;
1787         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
1788
1789         b10 += b3;
1790         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
1791
1792         b12 += b7;
1793         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
1794
1795         b1 += k1;
1796         b0 += b1 + k0;
1797         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
1798
1799         b3 += k3;
1800         b2 += b3 + k2;
1801         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
1802
1803         b5 += k5;
1804         b4 += b5 + k4;
1805         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
1806
1807         b7 += k7;
1808         b6 += b7 + k6;
1809         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
1810
1811         b9 += k9;
1812         b8 += b9 + k8;
1813         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
1814
1815         b11 += k11;
1816         b10 += b11 + k10;
1817         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
1818
1819         b13 += k13 + t2;
1820         b12 += b13 + k12;
1821         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
1822
1823         b15 += k15 + 17;
1824         b14 += b15 + k14 + t0;
1825         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
1826
1827         b0 += b9;
1828         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
1829
1830         b2 += b13;
1831         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
1832
1833         b6 += b11;
1834         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
1835
1836         b4 += b15;
1837         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
1838
1839         b10 += b7;
1840         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
1841
1842         b12 += b3;
1843         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
1844
1845         b14 += b5;
1846         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
1847
1848         b8 += b1;
1849         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
1850
1851         b0 += b7;
1852         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
1853
1854         b2 += b5;
1855         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
1856
1857         b4 += b3;
1858         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
1859
1860         b6 += b1;
1861         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
1862
1863         b12 += b15;
1864         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
1865
1866         b14 += b13;
1867         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
1868
1869         b8 += b11;
1870         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
1871
1872         b10 += b9;
1873         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
1874
1875         b0 += b15;
1876         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
1877
1878         b2 += b11;
1879         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
1880
1881         b6 += b13;
1882         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
1883
1884         b4 += b9;
1885         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
1886
1887         b14 += b1;
1888         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
1889
1890         b8 += b5;
1891         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
1892
1893         b10 += b3;
1894         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
1895
1896         b12 += b7;
1897         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
1898
1899         b1 += k2;
1900         b0 += b1 + k1;
1901         b1 = ((b1 << 24) | (b1 >> (64 - 24))) ^ b0;
1902
1903         b3 += k4;
1904         b2 += b3 + k3;
1905         b3 = ((b3 << 13) | (b3 >> (64 - 13))) ^ b2;
1906
1907         b5 += k6;
1908         b4 += b5 + k5;
1909         b5 = ((b5 << 8) | (b5 >> (64 - 8))) ^ b4;
1910
1911         b7 += k8;
1912         b6 += b7 + k7;
1913         b7 = ((b7 << 47) | (b7 >> (64 - 47))) ^ b6;
1914
1915         b9 += k10;
1916         b8 += b9 + k9;
1917         b9 = ((b9 << 8) | (b9 >> (64 - 8))) ^ b8;
1918
1919         b11 += k12;
1920         b10 += b11 + k11;
1921         b11 = ((b11 << 17) | (b11 >> (64 - 17))) ^ b10;
1922
1923         b13 += k14 + t0;
1924         b12 += b13 + k13;
1925         b13 = ((b13 << 22) | (b13 >> (64 - 22))) ^ b12;
1926
1927         b15 += k16 + 18;
1928         b14 += b15 + k15 + t1;
1929         b15 = ((b15 << 37) | (b15 >> (64 - 37))) ^ b14;
1930
1931         b0 += b9;
1932         b9 = ((b9 << 38) | (b9 >> (64 - 38))) ^ b0;
1933
1934         b2 += b13;
1935         b13 = ((b13 << 19) | (b13 >> (64 - 19))) ^ b2;
1936
1937         b6 += b11;
1938         b11 = ((b11 << 10) | (b11 >> (64 - 10))) ^ b6;
1939
1940         b4 += b15;
1941         b15 = ((b15 << 55) | (b15 >> (64 - 55))) ^ b4;
1942
1943         b10 += b7;
1944         b7 = ((b7 << 49) | (b7 >> (64 - 49))) ^ b10;
1945
1946         b12 += b3;
1947         b3 = ((b3 << 18) | (b3 >> (64 - 18))) ^ b12;
1948
1949         b14 += b5;
1950         b5 = ((b5 << 23) | (b5 >> (64 - 23))) ^ b14;
1951
1952         b8 += b1;
1953         b1 = ((b1 << 52) | (b1 >> (64 - 52))) ^ b8;
1954
1955         b0 += b7;
1956         b7 = ((b7 << 33) | (b7 >> (64 - 33))) ^ b0;
1957
1958         b2 += b5;
1959         b5 = ((b5 << 4) | (b5 >> (64 - 4))) ^ b2;
1960
1961         b4 += b3;
1962         b3 = ((b3 << 51) | (b3 >> (64 - 51))) ^ b4;
1963
1964         b6 += b1;
1965         b1 = ((b1 << 13) | (b1 >> (64 - 13))) ^ b6;
1966
1967         b12 += b15;
1968         b15 = ((b15 << 34) | (b15 >> (64 - 34))) ^ b12;
1969
1970         b14 += b13;
1971         b13 = ((b13 << 41) | (b13 >> (64 - 41))) ^ b14;
1972
1973         b8 += b11;
1974         b11 = ((b11 << 59) | (b11 >> (64 - 59))) ^ b8;
1975
1976         b10 += b9;
1977         b9 = ((b9 << 17) | (b9 >> (64 - 17))) ^ b10;
1978
1979         b0 += b15;
1980         b15 = ((b15 << 5) | (b15 >> (64 - 5))) ^ b0;
1981
1982         b2 += b11;
1983         b11 = ((b11 << 20) | (b11 >> (64 - 20))) ^ b2;
1984
1985         b6 += b13;
1986         b13 = ((b13 << 48) | (b13 >> (64 - 48))) ^ b6;
1987
1988         b4 += b9;
1989         b9 = ((b9 << 41) | (b9 >> (64 - 41))) ^ b4;
1990
1991         b14 += b1;
1992         b1 = ((b1 << 47) | (b1 >> (64 - 47))) ^ b14;
1993
1994         b8 += b5;
1995         b5 = ((b5 << 28) | (b5 >> (64 - 28))) ^ b8;
1996
1997         b10 += b3;
1998         b3 = ((b3 << 16) | (b3 >> (64 - 16))) ^ b10;
1999
2000         b12 += b7;
2001         b7 = ((b7 << 25) | (b7 >> (64 - 25))) ^ b12;
2002
2003         b1 += k3;
2004         b0 += b1 + k2;
2005         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b0;
2006
2007         b3 += k5;
2008         b2 += b3 + k4;
2009         b3 = ((b3 << 9) | (b3 >> (64 - 9))) ^ b2;
2010
2011         b5 += k7;
2012         b4 += b5 + k6;
2013         b5 = ((b5 << 37) | (b5 >> (64 - 37))) ^ b4;
2014
2015         b7 += k9;
2016         b6 += b7 + k8;
2017         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b6;
2018
2019         b9 += k11;
2020         b8 += b9 + k10;
2021         b9 = ((b9 << 12) | (b9 >> (64 - 12))) ^ b8;
2022
2023         b11 += k13;
2024         b10 += b11 + k12;
2025         b11 = ((b11 << 47) | (b11 >> (64 - 47))) ^ b10;
2026
2027         b13 += k15 + t1;
2028         b12 += b13 + k14;
2029         b13 = ((b13 << 44) | (b13 >> (64 - 44))) ^ b12;
2030
2031         b15 += k0 + 19;
2032         b14 += b15 + k16 + t2;
2033         b15 = ((b15 << 30) | (b15 >> (64 - 30))) ^ b14;
2034
2035         b0 += b9;
2036         b9 = ((b9 << 16) | (b9 >> (64 - 16))) ^ b0;
2037
2038         b2 += b13;
2039         b13 = ((b13 << 34) | (b13 >> (64 - 34))) ^ b2;
2040
2041         b6 += b11;
2042         b11 = ((b11 << 56) | (b11 >> (64 - 56))) ^ b6;
2043
2044         b4 += b15;
2045         b15 = ((b15 << 51) | (b15 >> (64 - 51))) ^ b4;
2046
2047         b10 += b7;
2048         b7 = ((b7 << 4) | (b7 >> (64 - 4))) ^ b10;
2049
2050         b12 += b3;
2051         b3 = ((b3 << 53) | (b3 >> (64 - 53))) ^ b12;
2052
2053         b14 += b5;
2054         b5 = ((b5 << 42) | (b5 >> (64 - 42))) ^ b14;
2055
2056         b8 += b1;
2057         b1 = ((b1 << 41) | (b1 >> (64 - 41))) ^ b8;
2058
2059         b0 += b7;
2060         b7 = ((b7 << 31) | (b7 >> (64 - 31))) ^ b0;
2061
2062         b2 += b5;
2063         b5 = ((b5 << 44) | (b5 >> (64 - 44))) ^ b2;
2064
2065         b4 += b3;
2066         b3 = ((b3 << 47) | (b3 >> (64 - 47))) ^ b4;
2067
2068         b6 += b1;
2069         b1 = ((b1 << 46) | (b1 >> (64 - 46))) ^ b6;
2070
2071         b12 += b15;
2072         b15 = ((b15 << 19) | (b15 >> (64 - 19))) ^ b12;
2073
2074         b14 += b13;
2075         b13 = ((b13 << 42) | (b13 >> (64 - 42))) ^ b14;
2076
2077         b8 += b11;
2078         b11 = ((b11 << 44) | (b11 >> (64 - 44))) ^ b8;
2079
2080         b10 += b9;
2081         b9 = ((b9 << 25) | (b9 >> (64 - 25))) ^ b10;
2082
2083         b0 += b15;
2084         b15 = ((b15 << 9) | (b15 >> (64 - 9))) ^ b0;
2085
2086         b2 += b11;
2087         b11 = ((b11 << 48) | (b11 >> (64 - 48))) ^ b2;
2088
2089         b6 += b13;
2090         b13 = ((b13 << 35) | (b13 >> (64 - 35))) ^ b6;
2091
2092         b4 += b9;
2093         b9 = ((b9 << 52) | (b9 >> (64 - 52))) ^ b4;
2094
2095         b14 += b1;
2096         b1 = ((b1 << 23) | (b1 >> (64 - 23))) ^ b14;
2097
2098         b8 += b5;
2099         b5 = ((b5 << 31) | (b5 >> (64 - 31))) ^ b8;
2100
2101         b10 += b3;
2102         b3 = ((b3 << 37) | (b3 >> (64 - 37))) ^ b10;
2103
2104         b12 += b7;
2105         b7 = ((b7 << 20) | (b7 >> (64 - 20))) ^ b12;
2106
2107         output[0] = b0 + k3;
2108         output[1] = b1 + k4;
2109         output[2] = b2 + k5;
2110         output[3] = b3 + k6;
2111         output[4] = b4 + k7;
2112         output[5] = b5 + k8;
2113         output[6] = b6 + k9;
2114         output[7] = b7 + k10;
2115         output[8] = b8 + k11;
2116         output[9] = b9 + k12;
2117         output[10] = b10 + k13;
2118         output[11] = b11 + k14;
2119         output[12] = b12 + k15;
2120         output[13] = b13 + k16 + t2;
2121         output[14] = b14 + k0 + t0;
2122         output[15] = b15 + k1 + 20;
2123 }
2124
2125 void threefishDecrypt1024(struct threefish_key *keyCtx, u64 *input, u64 *output)
2126 {
2127         u64 b0 = input[0], b1 = input[1],
2128           b2 = input[2], b3 = input[3],
2129           b4 = input[4], b5 = input[5],
2130           b6 = input[6], b7 = input[7],
2131           b8 = input[8], b9 = input[9],
2132           b10 = input[10], b11 = input[11],
2133           b12 = input[12], b13 = input[13],
2134           b14 = input[14], b15 = input[15];
2135         u64 k0 = keyCtx->key[0], k1 = keyCtx->key[1],
2136           k2 = keyCtx->key[2], k3 = keyCtx->key[3],
2137           k4 = keyCtx->key[4], k5 = keyCtx->key[5],
2138           k6 = keyCtx->key[6], k7 = keyCtx->key[7],
2139           k8 = keyCtx->key[8], k9 = keyCtx->key[9],
2140           k10 = keyCtx->key[10], k11 = keyCtx->key[11],
2141           k12 = keyCtx->key[12], k13 = keyCtx->key[13],
2142           k14 = keyCtx->key[14], k15 = keyCtx->key[15],
2143           k16 = keyCtx->key[16];
2144         u64 t0 = keyCtx->tweak[0], t1 = keyCtx->tweak[1],
2145           t2 = keyCtx->tweak[2];
2146         u64 tmp;
2147
2148         b0 -= k3;
2149         b1 -= k4;
2150         b2 -= k5;
2151         b3 -= k6;
2152         b4 -= k7;
2153         b5 -= k8;
2154         b6 -= k9;
2155         b7 -= k10;
2156         b8 -= k11;
2157         b9 -= k12;
2158         b10 -= k13;
2159         b11 -= k14;
2160         b12 -= k15;
2161         b13 -= k16 + t2;
2162         b14 -= k0 + t0;
2163         b15 -= k1 + 20;
2164         tmp = b7 ^ b12;
2165         b7 = (tmp >> 20) | (tmp << (64 - 20));
2166         b12 -= b7;
2167
2168         tmp = b3 ^ b10;
2169         b3 = (tmp >> 37) | (tmp << (64 - 37));
2170         b10 -= b3;
2171
2172         tmp = b5 ^ b8;
2173         b5 = (tmp >> 31) | (tmp << (64 - 31));
2174         b8 -= b5;
2175
2176         tmp = b1 ^ b14;
2177         b1 = (tmp >> 23) | (tmp << (64 - 23));
2178         b14 -= b1;
2179
2180         tmp = b9 ^ b4;
2181         b9 = (tmp >> 52) | (tmp << (64 - 52));
2182         b4 -= b9;
2183
2184         tmp = b13 ^ b6;
2185         b13 = (tmp >> 35) | (tmp << (64 - 35));
2186         b6 -= b13;
2187
2188         tmp = b11 ^ b2;
2189         b11 = (tmp >> 48) | (tmp << (64 - 48));
2190         b2 -= b11;
2191
2192         tmp = b15 ^ b0;
2193         b15 = (tmp >> 9) | (tmp << (64 - 9));
2194         b0 -= b15;
2195
2196         tmp = b9 ^ b10;
2197         b9 = (tmp >> 25) | (tmp << (64 - 25));
2198         b10 -= b9;
2199
2200         tmp = b11 ^ b8;
2201         b11 = (tmp >> 44) | (tmp << (64 - 44));
2202         b8 -= b11;
2203
2204         tmp = b13 ^ b14;
2205         b13 = (tmp >> 42) | (tmp << (64 - 42));
2206         b14 -= b13;
2207
2208         tmp = b15 ^ b12;
2209         b15 = (tmp >> 19) | (tmp << (64 - 19));
2210         b12 -= b15;
2211
2212         tmp = b1 ^ b6;
2213         b1 = (tmp >> 46) | (tmp << (64 - 46));
2214         b6 -= b1;
2215
2216         tmp = b3 ^ b4;
2217         b3 = (tmp >> 47) | (tmp << (64 - 47));
2218         b4 -= b3;
2219
2220         tmp = b5 ^ b2;
2221         b5 = (tmp >> 44) | (tmp << (64 - 44));
2222         b2 -= b5;
2223
2224         tmp = b7 ^ b0;
2225         b7 = (tmp >> 31) | (tmp << (64 - 31));
2226         b0 -= b7;
2227
2228         tmp = b1 ^ b8;
2229         b1 = (tmp >> 41) | (tmp << (64 - 41));
2230         b8 -= b1;
2231
2232         tmp = b5 ^ b14;
2233         b5 = (tmp >> 42) | (tmp << (64 - 42));
2234         b14 -= b5;
2235
2236         tmp = b3 ^ b12;
2237         b3 = (tmp >> 53) | (tmp << (64 - 53));
2238         b12 -= b3;
2239
2240         tmp = b7 ^ b10;
2241         b7 = (tmp >> 4) | (tmp << (64 - 4));
2242         b10 -= b7;
2243
2244         tmp = b15 ^ b4;
2245         b15 = (tmp >> 51) | (tmp << (64 - 51));
2246         b4 -= b15;
2247
2248         tmp = b11 ^ b6;
2249         b11 = (tmp >> 56) | (tmp << (64 - 56));
2250         b6 -= b11;
2251
2252         tmp = b13 ^ b2;
2253         b13 = (tmp >> 34) | (tmp << (64 - 34));
2254         b2 -= b13;
2255
2256         tmp = b9 ^ b0;
2257         b9 = (tmp >> 16) | (tmp << (64 - 16));
2258         b0 -= b9;
2259
2260         tmp = b15 ^ b14;
2261         b15 = (tmp >> 30) | (tmp << (64 - 30));
2262         b14 -= b15 + k16 + t2;
2263         b15 -= k0 + 19;
2264
2265         tmp = b13 ^ b12;
2266         b13 = (tmp >> 44) | (tmp << (64 - 44));
2267         b12 -= b13 + k14;
2268         b13 -= k15 + t1;
2269
2270         tmp = b11 ^ b10;
2271         b11 = (tmp >> 47) | (tmp << (64 - 47));
2272         b10 -= b11 + k12;
2273         b11 -= k13;
2274
2275         tmp = b9 ^ b8;
2276         b9 = (tmp >> 12) | (tmp << (64 - 12));
2277         b8 -= b9 + k10;
2278         b9 -= k11;
2279
2280         tmp = b7 ^ b6;
2281         b7 = (tmp >> 31) | (tmp << (64 - 31));
2282         b6 -= b7 + k8;
2283         b7 -= k9;
2284
2285         tmp = b5 ^ b4;
2286         b5 = (tmp >> 37) | (tmp << (64 - 37));
2287         b4 -= b5 + k6;
2288         b5 -= k7;
2289
2290         tmp = b3 ^ b2;
2291         b3 = (tmp >> 9) | (tmp << (64 - 9));
2292         b2 -= b3 + k4;
2293         b3 -= k5;
2294
2295         tmp = b1 ^ b0;
2296         b1 = (tmp >> 41) | (tmp << (64 - 41));
2297         b0 -= b1 + k2;
2298         b1 -= k3;
2299
2300         tmp = b7 ^ b12;
2301         b7 = (tmp >> 25) | (tmp << (64 - 25));
2302         b12 -= b7;
2303
2304         tmp = b3 ^ b10;
2305         b3 = (tmp >> 16) | (tmp << (64 - 16));
2306         b10 -= b3;
2307
2308         tmp = b5 ^ b8;
2309         b5 = (tmp >> 28) | (tmp << (64 - 28));
2310         b8 -= b5;
2311
2312         tmp = b1 ^ b14;
2313         b1 = (tmp >> 47) | (tmp << (64 - 47));
2314         b14 -= b1;
2315
2316         tmp = b9 ^ b4;
2317         b9 = (tmp >> 41) | (tmp << (64 - 41));
2318         b4 -= b9;
2319
2320         tmp = b13 ^ b6;
2321         b13 = (tmp >> 48) | (tmp << (64 - 48));
2322         b6 -= b13;
2323
2324         tmp = b11 ^ b2;
2325         b11 = (tmp >> 20) | (tmp << (64 - 20));
2326         b2 -= b11;
2327
2328         tmp = b15 ^ b0;
2329         b15 = (tmp >> 5) | (tmp << (64 - 5));
2330         b0 -= b15;
2331
2332         tmp = b9 ^ b10;
2333         b9 = (tmp >> 17) | (tmp << (64 - 17));
2334         b10 -= b9;
2335
2336         tmp = b11 ^ b8;
2337         b11 = (tmp >> 59) | (tmp << (64 - 59));
2338         b8 -= b11;
2339
2340         tmp = b13 ^ b14;
2341         b13 = (tmp >> 41) | (tmp << (64 - 41));
2342         b14 -= b13;
2343
2344         tmp = b15 ^ b12;
2345         b15 = (tmp >> 34) | (tmp << (64 - 34));
2346         b12 -= b15;
2347
2348         tmp = b1 ^ b6;
2349         b1 = (tmp >> 13) | (tmp << (64 - 13));
2350         b6 -= b1;
2351
2352         tmp = b3 ^ b4;
2353         b3 = (tmp >> 51) | (tmp << (64 - 51));
2354         b4 -= b3;
2355
2356         tmp = b5 ^ b2;
2357         b5 = (tmp >> 4) | (tmp << (64 - 4));
2358         b2 -= b5;
2359
2360         tmp = b7 ^ b0;
2361         b7 = (tmp >> 33) | (tmp << (64 - 33));
2362         b0 -= b7;
2363
2364         tmp = b1 ^ b8;
2365         b1 = (tmp >> 52) | (tmp << (64 - 52));
2366         b8 -= b1;
2367
2368         tmp = b5 ^ b14;
2369         b5 = (tmp >> 23) | (tmp << (64 - 23));
2370         b14 -= b5;
2371
2372         tmp = b3 ^ b12;
2373         b3 = (tmp >> 18) | (tmp << (64 - 18));
2374         b12 -= b3;
2375
2376         tmp = b7 ^ b10;
2377         b7 = (tmp >> 49) | (tmp << (64 - 49));
2378         b10 -= b7;
2379
2380         tmp = b15 ^ b4;
2381         b15 = (tmp >> 55) | (tmp << (64 - 55));
2382         b4 -= b15;
2383
2384         tmp = b11 ^ b6;
2385         b11 = (tmp >> 10) | (tmp << (64 - 10));
2386         b6 -= b11;
2387
2388         tmp = b13 ^ b2;
2389         b13 = (tmp >> 19) | (tmp << (64 - 19));
2390         b2 -= b13;
2391
2392         tmp = b9 ^ b0;
2393         b9 = (tmp >> 38) | (tmp << (64 - 38));
2394         b0 -= b9;
2395
2396         tmp = b15 ^ b14;
2397         b15 = (tmp >> 37) | (tmp << (64 - 37));
2398         b14 -= b15 + k15 + t1;
2399         b15 -= k16 + 18;
2400
2401         tmp = b13 ^ b12;
2402         b13 = (tmp >> 22) | (tmp << (64 - 22));
2403         b12 -= b13 + k13;
2404         b13 -= k14 + t0;
2405
2406         tmp = b11 ^ b10;
2407         b11 = (tmp >> 17) | (tmp << (64 - 17));
2408         b10 -= b11 + k11;
2409         b11 -= k12;
2410
2411         tmp = b9 ^ b8;
2412         b9 = (tmp >> 8) | (tmp << (64 - 8));
2413         b8 -= b9 + k9;
2414         b9 -= k10;
2415
2416         tmp = b7 ^ b6;
2417         b7 = (tmp >> 47) | (tmp << (64 - 47));
2418         b6 -= b7 + k7;
2419         b7 -= k8;
2420
2421         tmp = b5 ^ b4;
2422         b5 = (tmp >> 8) | (tmp << (64 - 8));
2423         b4 -= b5 + k5;
2424         b5 -= k6;
2425
2426         tmp = b3 ^ b2;
2427         b3 = (tmp >> 13) | (tmp << (64 - 13));
2428         b2 -= b3 + k3;
2429         b3 -= k4;
2430
2431         tmp = b1 ^ b0;
2432         b1 = (tmp >> 24) | (tmp << (64 - 24));
2433         b0 -= b1 + k1;
2434         b1 -= k2;
2435
2436         tmp = b7 ^ b12;
2437         b7 = (tmp >> 20) | (tmp << (64 - 20));
2438         b12 -= b7;
2439
2440         tmp = b3 ^ b10;
2441         b3 = (tmp >> 37) | (tmp << (64 - 37));
2442         b10 -= b3;
2443
2444         tmp = b5 ^ b8;
2445         b5 = (tmp >> 31) | (tmp << (64 - 31));
2446         b8 -= b5;
2447
2448         tmp = b1 ^ b14;
2449         b1 = (tmp >> 23) | (tmp << (64 - 23));
2450         b14 -= b1;
2451
2452         tmp = b9 ^ b4;
2453         b9 = (tmp >> 52) | (tmp << (64 - 52));
2454         b4 -= b9;
2455
2456         tmp = b13 ^ b6;
2457         b13 = (tmp >> 35) | (tmp << (64 - 35));
2458         b6 -= b13;
2459
2460         tmp = b11 ^ b2;
2461         b11 = (tmp >> 48) | (tmp << (64 - 48));
2462         b2 -= b11;
2463
2464         tmp = b15 ^ b0;
2465         b15 = (tmp >> 9) | (tmp << (64 - 9));
2466         b0 -= b15;
2467
2468         tmp = b9 ^ b10;
2469         b9 = (tmp >> 25) | (tmp << (64 - 25));
2470         b10 -= b9;
2471
2472         tmp = b11 ^ b8;
2473         b11 = (tmp >> 44) | (tmp << (64 - 44));
2474         b8 -= b11;
2475
2476         tmp = b13 ^ b14;
2477         b13 = (tmp >> 42) | (tmp << (64 - 42));
2478         b14 -= b13;
2479
2480         tmp = b15 ^ b12;
2481         b15 = (tmp >> 19) | (tmp << (64 - 19));
2482         b12 -= b15;
2483
2484         tmp = b1 ^ b6;
2485         b1 = (tmp >> 46) | (tmp << (64 - 46));
2486         b6 -= b1;
2487
2488         tmp = b3 ^ b4;
2489         b3 = (tmp >> 47) | (tmp << (64 - 47));
2490         b4 -= b3;
2491
2492         tmp = b5 ^ b2;
2493         b5 = (tmp >> 44) | (tmp << (64 - 44));
2494         b2 -= b5;
2495
2496         tmp = b7 ^ b0;
2497         b7 = (tmp >> 31) | (tmp << (64 - 31));
2498         b0 -= b7;
2499
2500         tmp = b1 ^ b8;
2501         b1 = (tmp >> 41) | (tmp << (64 - 41));
2502         b8 -= b1;
2503
2504         tmp = b5 ^ b14;
2505         b5 = (tmp >> 42) | (tmp << (64 - 42));
2506         b14 -= b5;
2507
2508         tmp = b3 ^ b12;
2509         b3 = (tmp >> 53) | (tmp << (64 - 53));
2510         b12 -= b3;
2511
2512         tmp = b7 ^ b10;
2513         b7 = (tmp >> 4) | (tmp << (64 - 4));
2514         b10 -= b7;
2515
2516         tmp = b15 ^ b4;
2517         b15 = (tmp >> 51) | (tmp << (64 - 51));
2518         b4 -= b15;
2519
2520         tmp = b11 ^ b6;
2521         b11 = (tmp >> 56) | (tmp << (64 - 56));
2522         b6 -= b11;
2523
2524         tmp = b13 ^ b2;
2525         b13 = (tmp >> 34) | (tmp << (64 - 34));
2526         b2 -= b13;
2527
2528         tmp = b9 ^ b0;
2529         b9 = (tmp >> 16) | (tmp << (64 - 16));
2530         b0 -= b9;
2531
2532         tmp = b15 ^ b14;
2533         b15 = (tmp >> 30) | (tmp << (64 - 30));
2534         b14 -= b15 + k14 + t0;
2535         b15 -= k15 + 17;
2536
2537         tmp = b13 ^ b12;
2538         b13 = (tmp >> 44) | (tmp << (64 - 44));
2539         b12 -= b13 + k12;
2540         b13 -= k13 + t2;
2541
2542         tmp = b11 ^ b10;
2543         b11 = (tmp >> 47) | (tmp << (64 - 47));
2544         b10 -= b11 + k10;
2545         b11 -= k11;
2546
2547         tmp = b9 ^ b8;
2548         b9 = (tmp >> 12) | (tmp << (64 - 12));
2549         b8 -= b9 + k8;
2550         b9 -= k9;
2551
2552         tmp = b7 ^ b6;
2553         b7 = (tmp >> 31) | (tmp << (64 - 31));
2554         b6 -= b7 + k6;
2555         b7 -= k7;
2556
2557         tmp = b5 ^ b4;
2558         b5 = (tmp >> 37) | (tmp << (64 - 37));
2559         b4 -= b5 + k4;
2560         b5 -= k5;
2561
2562         tmp = b3 ^ b2;
2563         b3 = (tmp >> 9) | (tmp << (64 - 9));
2564         b2 -= b3 + k2;
2565         b3 -= k3;
2566
2567         tmp = b1 ^ b0;
2568         b1 = (tmp >> 41) | (tmp << (64 - 41));
2569         b0 -= b1 + k0;
2570         b1 -= k1;
2571
2572         tmp = b7 ^ b12;
2573         b7 = (tmp >> 25) | (tmp << (64 - 25));
2574         b12 -= b7;
2575
2576         tmp = b3 ^ b10;
2577         b3 = (tmp >> 16) | (tmp << (64 - 16));
2578         b10 -= b3;
2579
2580         tmp = b5 ^ b8;
2581         b5 = (tmp >> 28) | (tmp << (64 - 28));
2582         b8 -= b5;
2583
2584         tmp = b1 ^ b14;
2585         b1 = (tmp >> 47) | (tmp << (64 - 47));
2586         b14 -= b1;
2587
2588         tmp = b9 ^ b4;
2589         b9 = (tmp >> 41) | (tmp << (64 - 41));
2590         b4 -= b9;
2591
2592         tmp = b13 ^ b6;
2593         b13 = (tmp >> 48) | (tmp << (64 - 48));
2594         b6 -= b13;
2595
2596         tmp = b11 ^ b2;
2597         b11 = (tmp >> 20) | (tmp << (64 - 20));
2598         b2 -= b11;
2599
2600         tmp = b15 ^ b0;
2601         b15 = (tmp >> 5) | (tmp << (64 - 5));
2602         b0 -= b15;
2603
2604         tmp = b9 ^ b10;
2605         b9 = (tmp >> 17) | (tmp << (64 - 17));
2606         b10 -= b9;
2607
2608         tmp = b11 ^ b8;
2609         b11 = (tmp >> 59) | (tmp << (64 - 59));
2610         b8 -= b11;
2611
2612         tmp = b13 ^ b14;
2613         b13 = (tmp >> 41) | (tmp << (64 - 41));
2614         b14 -= b13;
2615
2616         tmp = b15 ^ b12;
2617         b15 = (tmp >> 34) | (tmp << (64 - 34));
2618         b12 -= b15;
2619
2620         tmp = b1 ^ b6;
2621         b1 = (tmp >> 13) | (tmp << (64 - 13));
2622         b6 -= b1;
2623
2624         tmp = b3 ^ b4;
2625         b3 = (tmp >> 51) | (tmp << (64 - 51));
2626         b4 -= b3;
2627
2628         tmp = b5 ^ b2;
2629         b5 = (tmp >> 4) | (tmp << (64 - 4));
2630         b2 -= b5;
2631
2632         tmp = b7 ^ b0;
2633         b7 = (tmp >> 33) | (tmp << (64 - 33));
2634         b0 -= b7;
2635
2636         tmp = b1 ^ b8;
2637         b1 = (tmp >> 52) | (tmp << (64 - 52));
2638         b8 -= b1;
2639
2640         tmp = b5 ^ b14;
2641         b5 = (tmp >> 23) | (tmp << (64 - 23));
2642         b14 -= b5;
2643
2644         tmp = b3 ^ b12;
2645         b3 = (tmp >> 18) | (tmp << (64 - 18));
2646         b12 -= b3;
2647
2648         tmp = b7 ^ b10;
2649         b7 = (tmp >> 49) | (tmp << (64 - 49));
2650         b10 -= b7;
2651
2652         tmp = b15 ^ b4;
2653         b15 = (tmp >> 55) | (tmp << (64 - 55));
2654         b4 -= b15;
2655
2656         tmp = b11 ^ b6;
2657         b11 = (tmp >> 10) | (tmp << (64 - 10));
2658         b6 -= b11;
2659
2660         tmp = b13 ^ b2;
2661         b13 = (tmp >> 19) | (tmp << (64 - 19));
2662         b2 -= b13;
2663
2664         tmp = b9 ^ b0;
2665         b9 = (tmp >> 38) | (tmp << (64 - 38));
2666         b0 -= b9;
2667
2668         tmp = b15 ^ b14;
2669         b15 = (tmp >> 37) | (tmp << (64 - 37));
2670         b14 -= b15 + k13 + t2;
2671         b15 -= k14 + 16;
2672
2673         tmp = b13 ^ b12;
2674         b13 = (tmp >> 22) | (tmp << (64 - 22));
2675         b12 -= b13 + k11;
2676         b13 -= k12 + t1;
2677
2678         tmp = b11 ^ b10;
2679         b11 = (tmp >> 17) | (tmp << (64 - 17));
2680         b10 -= b11 + k9;
2681         b11 -= k10;
2682
2683         tmp = b9 ^ b8;
2684         b9 = (tmp >> 8) | (tmp << (64 - 8));
2685         b8 -= b9 + k7;
2686         b9 -= k8;
2687
2688         tmp = b7 ^ b6;
2689         b7 = (tmp >> 47) | (tmp << (64 - 47));
2690         b6 -= b7 + k5;
2691         b7 -= k6;
2692
2693         tmp = b5 ^ b4;
2694         b5 = (tmp >> 8) | (tmp << (64 - 8));
2695         b4 -= b5 + k3;
2696         b5 -= k4;
2697
2698         tmp = b3 ^ b2;
2699         b3 = (tmp >> 13) | (tmp << (64 - 13));
2700         b2 -= b3 + k1;
2701         b3 -= k2;
2702
2703         tmp = b1 ^ b0;
2704         b1 = (tmp >> 24) | (tmp << (64 - 24));
2705         b0 -= b1 + k16;
2706         b1 -= k0;
2707
2708         tmp = b7 ^ b12;
2709         b7 = (tmp >> 20) | (tmp << (64 - 20));
2710         b12 -= b7;
2711
2712         tmp = b3 ^ b10;
2713         b3 = (tmp >> 37) | (tmp << (64 - 37));
2714         b10 -= b3;
2715
2716         tmp = b5 ^ b8;
2717         b5 = (tmp >> 31) | (tmp << (64 - 31));
2718         b8 -= b5;
2719
2720         tmp = b1 ^ b14;
2721         b1 = (tmp >> 23) | (tmp << (64 - 23));
2722         b14 -= b1;
2723
2724         tmp = b9 ^ b4;
2725         b9 = (tmp >> 52) | (tmp << (64 - 52));
2726         b4 -= b9;
2727
2728         tmp = b13 ^ b6;
2729         b13 = (tmp >> 35) | (tmp << (64 - 35));
2730         b6 -= b13;
2731
2732         tmp = b11 ^ b2;
2733         b11 = (tmp >> 48) | (tmp << (64 - 48));
2734         b2 -= b11;
2735
2736         tmp = b15 ^ b0;
2737         b15 = (tmp >> 9) | (tmp << (64 - 9));
2738         b0 -= b15;
2739
2740         tmp = b9 ^ b10;
2741         b9 = (tmp >> 25) | (tmp << (64 - 25));
2742         b10 -= b9;
2743
2744         tmp = b11 ^ b8;
2745         b11 = (tmp >> 44) | (tmp << (64 - 44));
2746         b8 -= b11;
2747
2748         tmp = b13 ^ b14;
2749         b13 = (tmp >> 42) | (tmp << (64 - 42));
2750         b14 -= b13;
2751
2752         tmp = b15 ^ b12;
2753         b15 = (tmp >> 19) | (tmp << (64 - 19));
2754         b12 -= b15;
2755
2756         tmp = b1 ^ b6;
2757         b1 = (tmp >> 46) | (tmp << (64 - 46));
2758         b6 -= b1;
2759
2760         tmp = b3 ^ b4;
2761         b3 = (tmp >> 47) | (tmp << (64 - 47));
2762         b4 -= b3;
2763
2764         tmp = b5 ^ b2;
2765         b5 = (tmp >> 44) | (tmp << (64 - 44));
2766         b2 -= b5;
2767
2768         tmp = b7 ^ b0;
2769         b7 = (tmp >> 31) | (tmp << (64 - 31));
2770         b0 -= b7;
2771
2772         tmp = b1 ^ b8;
2773         b1 = (tmp >> 41) | (tmp << (64 - 41));
2774         b8 -= b1;
2775
2776         tmp = b5 ^ b14;
2777         b5 = (tmp >> 42) | (tmp << (64 - 42));
2778         b14 -= b5;
2779
2780         tmp = b3 ^ b12;
2781         b3 = (tmp >> 53) | (tmp << (64 - 53));
2782         b12 -= b3;
2783
2784         tmp = b7 ^ b10;
2785         b7 = (tmp >> 4) | (tmp << (64 - 4));
2786         b10 -= b7;
2787
2788         tmp = b15 ^ b4;
2789         b15 = (tmp >> 51) | (tmp << (64 - 51));
2790         b4 -= b15;
2791
2792         tmp = b11 ^ b6;
2793         b11 = (tmp >> 56) | (tmp << (64 - 56));
2794         b6 -= b11;
2795
2796         tmp = b13 ^ b2;
2797         b13 = (tmp >> 34) | (tmp << (64 - 34));
2798         b2 -= b13;
2799
2800         tmp = b9 ^ b0;
2801         b9 = (tmp >> 16) | (tmp << (64 - 16));
2802         b0 -= b9;
2803
2804         tmp = b15 ^ b14;
2805         b15 = (tmp >> 30) | (tmp << (64 - 30));
2806         b14 -= b15 + k12 + t1;
2807         b15 -= k13 + 15;
2808
2809         tmp = b13 ^ b12;
2810         b13 = (tmp >> 44) | (tmp << (64 - 44));
2811         b12 -= b13 + k10;
2812         b13 -= k11 + t0;
2813
2814         tmp = b11 ^ b10;
2815         b11 = (tmp >> 47) | (tmp << (64 - 47));
2816         b10 -= b11 + k8;
2817         b11 -= k9;
2818
2819         tmp = b9 ^ b8;
2820         b9 = (tmp >> 12) | (tmp << (64 - 12));
2821         b8 -= b9 + k6;
2822         b9 -= k7;
2823
2824         tmp = b7 ^ b6;
2825         b7 = (tmp >> 31) | (tmp << (64 - 31));
2826         b6 -= b7 + k4;
2827         b7 -= k5;
2828
2829         tmp = b5 ^ b4;
2830         b5 = (tmp >> 37) | (tmp << (64 - 37));
2831         b4 -= b5 + k2;
2832         b5 -= k3;
2833
2834         tmp = b3 ^ b2;
2835         b3 = (tmp >> 9) | (tmp << (64 - 9));
2836         b2 -= b3 + k0;
2837         b3 -= k1;
2838
2839         tmp = b1 ^ b0;
2840         b1 = (tmp >> 41) | (tmp << (64 - 41));
2841         b0 -= b1 + k15;
2842         b1 -= k16;
2843
2844         tmp = b7 ^ b12;
2845         b7 = (tmp >> 25) | (tmp << (64 - 25));
2846         b12 -= b7;
2847
2848         tmp = b3 ^ b10;
2849         b3 = (tmp >> 16) | (tmp << (64 - 16));
2850         b10 -= b3;
2851
2852         tmp = b5 ^ b8;
2853         b5 = (tmp >> 28) | (tmp << (64 - 28));
2854         b8 -= b5;
2855
2856         tmp = b1 ^ b14;
2857         b1 = (tmp >> 47) | (tmp << (64 - 47));
2858         b14 -= b1;
2859
2860         tmp = b9 ^ b4;
2861         b9 = (tmp >> 41) | (tmp << (64 - 41));
2862         b4 -= b9;
2863
2864         tmp = b13 ^ b6;
2865         b13 = (tmp >> 48) | (tmp << (64 - 48));
2866         b6 -= b13;
2867
2868         tmp = b11 ^ b2;
2869         b11 = (tmp >> 20) | (tmp << (64 - 20));
2870         b2 -= b11;
2871
2872         tmp = b15 ^ b0;
2873         b15 = (tmp >> 5) | (tmp << (64 - 5));
2874         b0 -= b15;
2875
2876         tmp = b9 ^ b10;
2877         b9 = (tmp >> 17) | (tmp << (64 - 17));
2878         b10 -= b9;
2879
2880         tmp = b11 ^ b8;
2881         b11 = (tmp >> 59) | (tmp << (64 - 59));
2882         b8 -= b11;
2883
2884         tmp = b13 ^ b14;
2885         b13 = (tmp >> 41) | (tmp << (64 - 41));
2886         b14 -= b13;
2887
2888         tmp = b15 ^ b12;
2889         b15 = (tmp >> 34) | (tmp << (64 - 34));
2890         b12 -= b15;
2891
2892         tmp = b1 ^ b6;
2893         b1 = (tmp >> 13) | (tmp << (64 - 13));
2894         b6 -= b1;
2895
2896         tmp = b3 ^ b4;
2897         b3 = (tmp >> 51) | (tmp << (64 - 51));
2898         b4 -= b3;
2899
2900         tmp = b5 ^ b2;
2901         b5 = (tmp >> 4) | (tmp << (64 - 4));
2902         b2 -= b5;
2903
2904         tmp = b7 ^ b0;
2905         b7 = (tmp >> 33) | (tmp << (64 - 33));
2906         b0 -= b7;
2907
2908         tmp = b1 ^ b8;
2909         b1 = (tmp >> 52) | (tmp << (64 - 52));
2910         b8 -= b1;
2911
2912         tmp = b5 ^ b14;
2913         b5 = (tmp >> 23) | (tmp << (64 - 23));
2914         b14 -= b5;
2915
2916         tmp = b3 ^ b12;
2917         b3 = (tmp >> 18) | (tmp << (64 - 18));
2918         b12 -= b3;
2919
2920         tmp = b7 ^ b10;
2921         b7 = (tmp >> 49) | (tmp << (64 - 49));
2922         b10 -= b7;
2923
2924         tmp = b15 ^ b4;
2925         b15 = (tmp >> 55) | (tmp << (64 - 55));
2926         b4 -= b15;
2927
2928         tmp = b11 ^ b6;
2929         b11 = (tmp >> 10) | (tmp << (64 - 10));
2930         b6 -= b11;
2931
2932         tmp = b13 ^ b2;
2933         b13 = (tmp >> 19) | (tmp << (64 - 19));
2934         b2 -= b13;
2935
2936         tmp = b9 ^ b0;
2937         b9 = (tmp >> 38) | (tmp << (64 - 38));
2938         b0 -= b9;
2939
2940         tmp = b15 ^ b14;
2941         b15 = (tmp >> 37) | (tmp << (64 - 37));
2942         b14 -= b15 + k11 + t0;
2943         b15 -= k12 + 14;
2944
2945         tmp = b13 ^ b12;
2946         b13 = (tmp >> 22) | (tmp << (64 - 22));
2947         b12 -= b13 + k9;
2948         b13 -= k10 + t2;
2949
2950         tmp = b11 ^ b10;
2951         b11 = (tmp >> 17) | (tmp << (64 - 17));
2952         b10 -= b11 + k7;
2953         b11 -= k8;
2954
2955         tmp = b9 ^ b8;
2956         b9 = (tmp >> 8) | (tmp << (64 - 8));
2957         b8 -= b9 + k5;
2958         b9 -= k6;
2959
2960         tmp = b7 ^ b6;
2961         b7 = (tmp >> 47) | (tmp << (64 - 47));
2962         b6 -= b7 + k3;
2963         b7 -= k4;
2964
2965         tmp = b5 ^ b4;
2966         b5 = (tmp >> 8) | (tmp << (64 - 8));
2967         b4 -= b5 + k1;
2968         b5 -= k2;
2969
2970         tmp = b3 ^ b2;
2971         b3 = (tmp >> 13) | (tmp << (64 - 13));
2972         b2 -= b3 + k16;
2973         b3 -= k0;
2974
2975         tmp = b1 ^ b0;
2976         b1 = (tmp >> 24) | (tmp << (64 - 24));
2977         b0 -= b1 + k14;
2978         b1 -= k15;
2979
2980         tmp = b7 ^ b12;
2981         b7 = (tmp >> 20) | (tmp << (64 - 20));
2982         b12 -= b7;
2983
2984         tmp = b3 ^ b10;
2985         b3 = (tmp >> 37) | (tmp << (64 - 37));
2986         b10 -= b3;
2987
2988         tmp = b5 ^ b8;
2989         b5 = (tmp >> 31) | (tmp << (64 - 31));
2990         b8 -= b5;
2991
2992         tmp = b1 ^ b14;
2993         b1 = (tmp >> 23) | (tmp << (64 - 23));
2994         b14 -= b1;
2995
2996         tmp = b9 ^ b4;
2997         b9 = (tmp >> 52) | (tmp << (64 - 52));
2998         b4 -= b9;
2999
3000         tmp = b13 ^ b6;
3001         b13 = (tmp >> 35) | (tmp << (64 - 35));
3002         b6 -= b13;
3003
3004         tmp = b11 ^ b2;
3005         b11 = (tmp >> 48) | (tmp << (64 - 48));
3006         b2 -= b11;
3007
3008         tmp = b15 ^ b0;
3009         b15 = (tmp >> 9) | (tmp << (64 - 9));
3010         b0 -= b15;
3011
3012         tmp = b9 ^ b10;
3013         b9 = (tmp >> 25) | (tmp << (64 - 25));
3014         b10 -= b9;
3015
3016         tmp = b11 ^ b8;
3017         b11 = (tmp >> 44) | (tmp << (64 - 44));
3018         b8 -= b11;
3019
3020         tmp = b13 ^ b14;
3021         b13 = (tmp >> 42) | (tmp << (64 - 42));
3022         b14 -= b13;
3023
3024         tmp = b15 ^ b12;
3025         b15 = (tmp >> 19) | (tmp << (64 - 19));
3026         b12 -= b15;
3027
3028         tmp = b1 ^ b6;
3029         b1 = (tmp >> 46) | (tmp << (64 - 46));
3030         b6 -= b1;
3031
3032         tmp = b3 ^ b4;
3033         b3 = (tmp >> 47) | (tmp << (64 - 47));
3034         b4 -= b3;
3035
3036         tmp = b5 ^ b2;
3037         b5 = (tmp >> 44) | (tmp << (64 - 44));
3038         b2 -= b5;
3039
3040         tmp = b7 ^ b0;
3041         b7 = (tmp >> 31) | (tmp << (64 - 31));
3042         b0 -= b7;
3043
3044         tmp = b1 ^ b8;
3045         b1 = (tmp >> 41) | (tmp << (64 - 41));
3046         b8 -= b1;
3047
3048         tmp = b5 ^ b14;
3049         b5 = (tmp >> 42) | (tmp << (64 - 42));
3050         b14 -= b5;
3051
3052         tmp = b3 ^ b12;
3053         b3 = (tmp >> 53) | (tmp << (64 - 53));
3054         b12 -= b3;
3055
3056         tmp = b7 ^ b10;
3057         b7 = (tmp >> 4) | (tmp << (64 - 4));
3058         b10 -= b7;
3059
3060         tmp = b15 ^ b4;
3061         b15 = (tmp >> 51) | (tmp << (64 - 51));
3062         b4 -= b15;
3063
3064         tmp = b11 ^ b6;
3065         b11 = (tmp >> 56) | (tmp << (64 - 56));
3066         b6 -= b11;
3067
3068         tmp = b13 ^ b2;
3069         b13 = (tmp >> 34) | (tmp << (64 - 34));
3070         b2 -= b13;
3071
3072         tmp = b9 ^ b0;
3073         b9 = (tmp >> 16) | (tmp << (64 - 16));
3074         b0 -= b9;
3075
3076         tmp = b15 ^ b14;
3077         b15 = (tmp >> 30) | (tmp << (64 - 30));
3078         b14 -= b15 + k10 + t2;
3079         b15 -= k11 + 13;
3080
3081         tmp = b13 ^ b12;
3082         b13 = (tmp >> 44) | (tmp << (64 - 44));
3083         b12 -= b13 + k8;
3084         b13 -= k9 + t1;
3085
3086         tmp = b11 ^ b10;
3087         b11 = (tmp >> 47) | (tmp << (64 - 47));
3088         b10 -= b11 + k6;
3089         b11 -= k7;
3090
3091         tmp = b9 ^ b8;
3092         b9 = (tmp >> 12) | (tmp << (64 - 12));
3093         b8 -= b9 + k4;
3094         b9 -= k5;
3095
3096         tmp = b7 ^ b6;
3097         b7 = (tmp >> 31) | (tmp << (64 - 31));
3098         b6 -= b7 + k2;
3099         b7 -= k3;
3100
3101         tmp = b5 ^ b4;
3102         b5 = (tmp >> 37) | (tmp << (64 - 37));
3103         b4 -= b5 + k0;
3104         b5 -= k1;
3105
3106         tmp = b3 ^ b2;
3107         b3 = (tmp >> 9) | (tmp << (64 - 9));
3108         b2 -= b3 + k15;
3109         b3 -= k16;
3110
3111         tmp = b1 ^ b0;
3112         b1 = (tmp >> 41) | (tmp << (64 - 41));
3113         b0 -= b1 + k13;
3114         b1 -= k14;
3115
3116         tmp = b7 ^ b12;
3117         b7 = (tmp >> 25) | (tmp << (64 - 25));
3118         b12 -= b7;
3119
3120         tmp = b3 ^ b10;
3121         b3 = (tmp >> 16) | (tmp << (64 - 16));
3122         b10 -= b3;
3123
3124         tmp = b5 ^ b8;
3125         b5 = (tmp >> 28) | (tmp << (64 - 28));
3126         b8 -= b5;
3127
3128         tmp = b1 ^ b14;
3129         b1 = (tmp >> 47) | (tmp << (64 - 47));
3130         b14 -= b1;
3131
3132         tmp = b9 ^ b4;
3133         b9 = (tmp >> 41) | (tmp << (64 - 41));
3134         b4 -= b9;
3135
3136         tmp = b13 ^ b6;
3137         b13 = (tmp >> 48) | (tmp << (64 - 48));
3138         b6 -= b13;
3139
3140         tmp = b11 ^ b2;
3141         b11 = (tmp >> 20) | (tmp << (64 - 20));
3142         b2 -= b11;
3143
3144         tmp = b15 ^ b0;
3145         b15 = (tmp >> 5) | (tmp << (64 - 5));
3146         b0 -= b15;
3147
3148         tmp = b9 ^ b10;
3149         b9 = (tmp >> 17) | (tmp << (64 - 17));
3150         b10 -= b9;
3151
3152         tmp = b11 ^ b8;
3153         b11 = (tmp >> 59) | (tmp << (64 - 59));
3154         b8 -= b11;
3155
3156         tmp = b13 ^ b14;
3157         b13 = (tmp >> 41) | (tmp << (64 - 41));
3158         b14 -= b13;
3159
3160         tmp = b15 ^ b12;
3161         b15 = (tmp >> 34) | (tmp << (64 - 34));
3162         b12 -= b15;
3163
3164         tmp = b1 ^ b6;
3165         b1 = (tmp >> 13) | (tmp << (64 - 13));
3166         b6 -= b1;
3167
3168         tmp = b3 ^ b4;
3169         b3 = (tmp >> 51) | (tmp << (64 - 51));
3170         b4 -= b3;
3171
3172         tmp = b5 ^ b2;
3173         b5 = (tmp >> 4) | (tmp << (64 - 4));
3174         b2 -= b5;
3175
3176         tmp = b7 ^ b0;
3177         b7 = (tmp >> 33) | (tmp << (64 - 33));
3178         b0 -= b7;
3179
3180         tmp = b1 ^ b8;
3181         b1 = (tmp >> 52) | (tmp << (64 - 52));
3182         b8 -= b1;
3183
3184         tmp = b5 ^ b14;
3185         b5 = (tmp >> 23) | (tmp << (64 - 23));
3186         b14 -= b5;
3187
3188         tmp = b3 ^ b12;
3189         b3 = (tmp >> 18) | (tmp << (64 - 18));
3190         b12 -= b3;
3191
3192         tmp = b7 ^ b10;
3193         b7 = (tmp >> 49) | (tmp << (64 - 49));
3194         b10 -= b7;
3195
3196         tmp = b15 ^ b4;
3197         b15 = (tmp >> 55) | (tmp << (64 - 55));
3198         b4 -= b15;
3199
3200         tmp = b11 ^ b6;
3201         b11 = (tmp >> 10) | (tmp << (64 - 10));
3202         b6 -= b11;
3203
3204         tmp = b13 ^ b2;
3205         b13 = (tmp >> 19) | (tmp << (64 - 19));
3206         b2 -= b13;
3207
3208         tmp = b9 ^ b0;
3209         b9 = (tmp >> 38) | (tmp << (64 - 38));
3210         b0 -= b9;
3211
3212         tmp = b15 ^ b14;
3213         b15 = (tmp >> 37) | (tmp << (64 - 37));
3214         b14 -= b15 + k9 + t1;
3215         b15 -= k10 + 12;
3216
3217         tmp = b13 ^ b12;
3218         b13 = (tmp >> 22) | (tmp << (64 - 22));
3219         b12 -= b13 + k7;
3220         b13 -= k8 + t0;
3221
3222         tmp = b11 ^ b10;
3223         b11 = (tmp >> 17) | (tmp << (64 - 17));
3224         b10 -= b11 + k5;
3225         b11 -= k6;
3226
3227         tmp = b9 ^ b8;
3228         b9 = (tmp >> 8) | (tmp << (64 - 8));
3229         b8 -= b9 + k3;
3230         b9 -= k4;
3231
3232         tmp = b7 ^ b6;
3233         b7 = (tmp >> 47) | (tmp << (64 - 47));
3234         b6 -= b7 + k1;
3235         b7 -= k2;
3236
3237         tmp = b5 ^ b4;
3238         b5 = (tmp >> 8) | (tmp << (64 - 8));
3239         b4 -= b5 + k16;
3240         b5 -= k0;
3241
3242         tmp = b3 ^ b2;
3243         b3 = (tmp >> 13) | (tmp << (64 - 13));
3244         b2 -= b3 + k14;
3245         b3 -= k15;
3246
3247         tmp = b1 ^ b0;
3248         b1 = (tmp >> 24) | (tmp << (64 - 24));
3249         b0 -= b1 + k12;
3250         b1 -= k13;
3251
3252         tmp = b7 ^ b12;
3253         b7 = (tmp >> 20) | (tmp << (64 - 20));
3254         b12 -= b7;
3255
3256         tmp = b3 ^ b10;
3257         b3 = (tmp >> 37) | (tmp << (64 - 37));
3258         b10 -= b3;
3259
3260         tmp = b5 ^ b8;
3261         b5 = (tmp >> 31) | (tmp << (64 - 31));
3262         b8 -= b5;
3263
3264         tmp = b1 ^ b14;
3265         b1 = (tmp >> 23) | (tmp << (64 - 23));
3266         b14 -= b1;
3267
3268         tmp = b9 ^ b4;
3269         b9 = (tmp >> 52) | (tmp << (64 - 52));
3270         b4 -= b9;
3271
3272         tmp = b13 ^ b6;
3273         b13 = (tmp >> 35) | (tmp << (64 - 35));
3274         b6 -= b13;
3275
3276         tmp = b11 ^ b2;
3277         b11 = (tmp >> 48) | (tmp << (64 - 48));
3278         b2 -= b11;
3279
3280         tmp = b15 ^ b0;
3281         b15 = (tmp >> 9) | (tmp << (64 - 9));
3282         b0 -= b15;
3283
3284         tmp = b9 ^ b10;
3285         b9 = (tmp >> 25) | (tmp << (64 - 25));
3286         b10 -= b9;
3287
3288         tmp = b11 ^ b8;
3289         b11 = (tmp >> 44) | (tmp << (64 - 44));
3290         b8 -= b11;
3291
3292         tmp = b13 ^ b14;
3293         b13 = (tmp >> 42) | (tmp << (64 - 42));
3294         b14 -= b13;
3295
3296         tmp = b15 ^ b12;
3297         b15 = (tmp >> 19) | (tmp << (64 - 19));
3298         b12 -= b15;
3299
3300         tmp = b1 ^ b6;
3301         b1 = (tmp >> 46) | (tmp << (64 - 46));
3302         b6 -= b1;
3303
3304         tmp = b3 ^ b4;
3305         b3 = (tmp >> 47) | (tmp << (64 - 47));
3306         b4 -= b3;
3307
3308         tmp = b5 ^ b2;
3309         b5 = (tmp >> 44) | (tmp << (64 - 44));
3310         b2 -= b5;
3311
3312         tmp = b7 ^ b0;
3313         b7 = (tmp >> 31) | (tmp << (64 - 31));
3314         b0 -= b7;
3315
3316         tmp = b1 ^ b8;
3317         b1 = (tmp >> 41) | (tmp << (64 - 41));
3318         b8 -= b1;
3319
3320         tmp = b5 ^ b14;
3321         b5 = (tmp >> 42) | (tmp << (64 - 42));
3322         b14 -= b5;
3323
3324         tmp = b3 ^ b12;
3325         b3 = (tmp >> 53) | (tmp << (64 - 53));
3326         b12 -= b3;
3327
3328         tmp = b7 ^ b10;
3329         b7 = (tmp >> 4) | (tmp << (64 - 4));
3330         b10 -= b7;
3331
3332         tmp = b15 ^ b4;
3333         b15 = (tmp >> 51) | (tmp << (64 - 51));
3334         b4 -= b15;
3335
3336         tmp = b11 ^ b6;
3337         b11 = (tmp >> 56) | (tmp << (64 - 56));
3338         b6 -= b11;
3339
3340         tmp = b13 ^ b2;
3341         b13 = (tmp >> 34) | (tmp << (64 - 34));
3342         b2 -= b13;
3343
3344         tmp = b9 ^ b0;
3345         b9 = (tmp >> 16) | (tmp << (64 - 16));
3346         b0 -= b9;
3347
3348         tmp = b15 ^ b14;
3349         b15 = (tmp >> 30) | (tmp << (64 - 30));
3350         b14 -= b15 + k8 + t0;
3351         b15 -= k9 + 11;
3352
3353         tmp = b13 ^ b12;
3354         b13 = (tmp >> 44) | (tmp << (64 - 44));
3355         b12 -= b13 + k6;
3356         b13 -= k7 + t2;
3357
3358         tmp = b11 ^ b10;
3359         b11 = (tmp >> 47) | (tmp << (64 - 47));
3360         b10 -= b11 + k4;
3361         b11 -= k5;
3362
3363         tmp = b9 ^ b8;
3364         b9 = (tmp >> 12) | (tmp << (64 - 12));
3365         b8 -= b9 + k2;
3366         b9 -= k3;
3367
3368         tmp = b7 ^ b6;
3369         b7 = (tmp >> 31) | (tmp << (64 - 31));
3370         b6 -= b7 + k0;
3371         b7 -= k1;
3372
3373         tmp = b5 ^ b4;
3374         b5 = (tmp >> 37) | (tmp << (64 - 37));
3375         b4 -= b5 + k15;
3376         b5 -= k16;
3377
3378         tmp = b3 ^ b2;
3379         b3 = (tmp >> 9) | (tmp << (64 - 9));
3380         b2 -= b3 + k13;
3381         b3 -= k14;
3382
3383         tmp = b1 ^ b0;
3384         b1 = (tmp >> 41) | (tmp << (64 - 41));
3385         b0 -= b1 + k11;
3386         b1 -= k12;
3387
3388         tmp = b7 ^ b12;
3389         b7 = (tmp >> 25) | (tmp << (64 - 25));
3390         b12 -= b7;
3391
3392         tmp = b3 ^ b10;
3393         b3 = (tmp >> 16) | (tmp << (64 - 16));
3394         b10 -= b3;
3395
3396         tmp = b5 ^ b8;
3397         b5 = (tmp >> 28) | (tmp << (64 - 28));
3398         b8 -= b5;
3399
3400         tmp = b1 ^ b14;
3401         b1 = (tmp >> 47) | (tmp << (64 - 47));
3402         b14 -= b1;
3403
3404         tmp = b9 ^ b4;
3405         b9 = (tmp >> 41) | (tmp << (64 - 41));
3406         b4 -= b9;
3407
3408         tmp = b13 ^ b6;
3409         b13 = (tmp >> 48) | (tmp << (64 - 48));
3410         b6 -= b13;
3411
3412         tmp = b11 ^ b2;
3413         b11 = (tmp >> 20) | (tmp << (64 - 20));
3414         b2 -= b11;
3415
3416         tmp = b15 ^ b0;
3417         b15 = (tmp >> 5) | (tmp << (64 - 5));
3418         b0 -= b15;
3419
3420         tmp = b9 ^ b10;
3421         b9 = (tmp >> 17) | (tmp << (64 - 17));
3422         b10 -= b9;
3423
3424         tmp = b11 ^ b8;
3425         b11 = (tmp >> 59) | (tmp << (64 - 59));
3426         b8 -= b11;
3427
3428         tmp = b13 ^ b14;
3429         b13 = (tmp >> 41) | (tmp << (64 - 41));
3430         b14 -= b13;
3431
3432         tmp = b15 ^ b12;
3433         b15 = (tmp >> 34) | (tmp << (64 - 34));
3434         b12 -= b15;
3435
3436         tmp = b1 ^ b6;
3437         b1 = (tmp >> 13) | (tmp << (64 - 13));
3438         b6 -= b1;
3439
3440         tmp = b3 ^ b4;
3441         b3 = (tmp >> 51) | (tmp << (64 - 51));
3442         b4 -= b3;
3443
3444         tmp = b5 ^ b2;
3445         b5 = (tmp >> 4) | (tmp << (64 - 4));
3446         b2 -= b5;
3447
3448         tmp = b7 ^ b0;
3449         b7 = (tmp >> 33) | (tmp << (64 - 33));
3450         b0 -= b7;
3451
3452         tmp = b1 ^ b8;
3453         b1 = (tmp >> 52) | (tmp << (64 - 52));
3454         b8 -= b1;
3455
3456         tmp = b5 ^ b14;
3457         b5 = (tmp >> 23) | (tmp << (64 - 23));
3458         b14 -= b5;
3459
3460         tmp = b3 ^ b12;
3461         b3 = (tmp >> 18) | (tmp << (64 - 18));
3462         b12 -= b3;
3463
3464         tmp = b7 ^ b10;
3465         b7 = (tmp >> 49) | (tmp << (64 - 49));
3466         b10 -= b7;
3467
3468         tmp = b15 ^ b4;
3469         b15 = (tmp >> 55) | (tmp << (64 - 55));
3470         b4 -= b15;
3471
3472         tmp = b11 ^ b6;
3473         b11 = (tmp >> 10) | (tmp << (64 - 10));
3474         b6 -= b11;
3475
3476         tmp = b13 ^ b2;
3477         b13 = (tmp >> 19) | (tmp << (64 - 19));
3478         b2 -= b13;
3479
3480         tmp = b9 ^ b0;
3481         b9 = (tmp >> 38) | (tmp << (64 - 38));
3482         b0 -= b9;
3483
3484         tmp = b15 ^ b14;
3485         b15 = (tmp >> 37) | (tmp << (64 - 37));
3486         b14 -= b15 + k7 + t2;
3487         b15 -= k8 + 10;
3488
3489         tmp = b13 ^ b12;
3490         b13 = (tmp >> 22) | (tmp << (64 - 22));
3491         b12 -= b13 + k5;
3492         b13 -= k6 + t1;
3493
3494         tmp = b11 ^ b10;
3495         b11 = (tmp >> 17) | (tmp << (64 - 17));
3496         b10 -= b11 + k3;
3497         b11 -= k4;
3498
3499         tmp = b9 ^ b8;
3500         b9 = (tmp >> 8) | (tmp << (64 - 8));
3501         b8 -= b9 + k1;
3502         b9 -= k2;
3503
3504         tmp = b7 ^ b6;
3505         b7 = (tmp >> 47) | (tmp << (64 - 47));
3506         b6 -= b7 + k16;
3507         b7 -= k0;
3508
3509         tmp = b5 ^ b4;
3510         b5 = (tmp >> 8) | (tmp << (64 - 8));
3511         b4 -= b5 + k14;
3512         b5 -= k15;
3513
3514         tmp = b3 ^ b2;
3515         b3 = (tmp >> 13) | (tmp << (64 - 13));
3516         b2 -= b3 + k12;
3517         b3 -= k13;
3518
3519         tmp = b1 ^ b0;
3520         b1 = (tmp >> 24) | (tmp << (64 - 24));
3521         b0 -= b1 + k10;
3522         b1 -= k11;
3523
3524         tmp = b7 ^ b12;
3525         b7 = (tmp >> 20) | (tmp << (64 - 20));
3526         b12 -= b7;
3527
3528         tmp = b3 ^ b10;
3529         b3 = (tmp >> 37) | (tmp << (64 - 37));
3530         b10 -= b3;
3531
3532         tmp = b5 ^ b8;
3533         b5 = (tmp >> 31) | (tmp << (64 - 31));
3534         b8 -= b5;
3535
3536         tmp = b1 ^ b14;
3537         b1 = (tmp >> 23) | (tmp << (64 - 23));
3538         b14 -= b1;
3539
3540         tmp = b9 ^ b4;
3541         b9 = (tmp >> 52) | (tmp << (64 - 52));
3542         b4 -= b9;
3543
3544         tmp = b13 ^ b6;
3545         b13 = (tmp >> 35) | (tmp << (64 - 35));
3546         b6 -= b13;
3547
3548         tmp = b11 ^ b2;
3549         b11 = (tmp >> 48) | (tmp << (64 - 48));
3550         b2 -= b11;
3551
3552         tmp = b15 ^ b0;
3553         b15 = (tmp >> 9) | (tmp << (64 - 9));
3554         b0 -= b15;
3555
3556         tmp = b9 ^ b10;
3557         b9 = (tmp >> 25) | (tmp << (64 - 25));
3558         b10 -= b9;
3559
3560         tmp = b11 ^ b8;
3561         b11 = (tmp >> 44) | (tmp << (64 - 44));
3562         b8 -= b11;
3563
3564         tmp = b13 ^ b14;
3565         b13 = (tmp >> 42) | (tmp << (64 - 42));
3566         b14 -= b13;
3567
3568         tmp = b15 ^ b12;
3569         b15 = (tmp >> 19) | (tmp << (64 - 19));
3570         b12 -= b15;
3571
3572         tmp = b1 ^ b6;
3573         b1 = (tmp >> 46) | (tmp << (64 - 46));
3574         b6 -= b1;
3575
3576         tmp = b3 ^ b4;
3577         b3 = (tmp >> 47) | (tmp << (64 - 47));
3578         b4 -= b3;
3579
3580         tmp = b5 ^ b2;
3581         b5 = (tmp >> 44) | (tmp << (64 - 44));
3582         b2 -= b5;
3583
3584         tmp = b7 ^ b0;
3585         b7 = (tmp >> 31) | (tmp << (64 - 31));
3586         b0 -= b7;
3587
3588         tmp = b1 ^ b8;
3589         b1 = (tmp >> 41) | (tmp << (64 - 41));
3590         b8 -= b1;
3591
3592         tmp = b5 ^ b14;
3593         b5 = (tmp >> 42) | (tmp << (64 - 42));
3594         b14 -= b5;
3595
3596         tmp = b3 ^ b12;
3597         b3 = (tmp >> 53) | (tmp << (64 - 53));
3598         b12 -= b3;
3599
3600         tmp = b7 ^ b10;
3601         b7 = (tmp >> 4) | (tmp << (64 - 4));
3602         b10 -= b7;
3603
3604         tmp = b15 ^ b4;
3605         b15 = (tmp >> 51) | (tmp << (64 - 51));
3606         b4 -= b15;
3607
3608         tmp = b11 ^ b6;
3609         b11 = (tmp >> 56) | (tmp << (64 - 56));
3610         b6 -= b11;
3611
3612         tmp = b13 ^ b2;
3613         b13 = (tmp >> 34) | (tmp << (64 - 34));
3614         b2 -= b13;
3615
3616         tmp = b9 ^ b0;
3617         b9 = (tmp >> 16) | (tmp << (64 - 16));
3618         b0 -= b9;
3619
3620         tmp = b15 ^ b14;
3621         b15 = (tmp >> 30) | (tmp << (64 - 30));
3622         b14 -= b15 + k6 + t1;
3623         b15 -= k7 + 9;
3624
3625         tmp = b13 ^ b12;
3626         b13 = (tmp >> 44) | (tmp << (64 - 44));
3627         b12 -= b13 + k4;
3628         b13 -= k5 + t0;
3629
3630         tmp = b11 ^ b10;
3631         b11 = (tmp >> 47) | (tmp << (64 - 47));
3632         b10 -= b11 + k2;
3633         b11 -= k3;
3634
3635         tmp = b9 ^ b8;
3636         b9 = (tmp >> 12) | (tmp << (64 - 12));
3637         b8 -= b9 + k0;
3638         b9 -= k1;
3639
3640         tmp = b7 ^ b6;
3641         b7 = (tmp >> 31) | (tmp << (64 - 31));
3642         b6 -= b7 + k15;
3643         b7 -= k16;
3644
3645         tmp = b5 ^ b4;
3646         b5 = (tmp >> 37) | (tmp << (64 - 37));
3647         b4 -= b5 + k13;
3648         b5 -= k14;
3649
3650         tmp = b3 ^ b2;
3651         b3 = (tmp >> 9) | (tmp << (64 - 9));
3652         b2 -= b3 + k11;
3653         b3 -= k12;
3654
3655         tmp = b1 ^ b0;
3656         b1 = (tmp >> 41) | (tmp << (64 - 41));
3657         b0 -= b1 + k9;
3658         b1 -= k10;
3659
3660         tmp = b7 ^ b12;
3661         b7 = (tmp >> 25) | (tmp << (64 - 25));
3662         b12 -= b7;
3663
3664         tmp = b3 ^ b10;
3665         b3 = (tmp >> 16) | (tmp << (64 - 16));
3666         b10 -= b3;
3667
3668         tmp = b5 ^ b8;
3669         b5 = (tmp >> 28) | (tmp << (64 - 28));
3670         b8 -= b5;
3671
3672         tmp = b1 ^ b14;
3673         b1 = (tmp >> 47) | (tmp << (64 - 47));
3674         b14 -= b1;
3675
3676         tmp = b9 ^ b4;
3677         b9 = (tmp >> 41) | (tmp << (64 - 41));
3678         b4 -= b9;
3679
3680         tmp = b13 ^ b6;
3681         b13 = (tmp >> 48) | (tmp << (64 - 48));
3682         b6 -= b13;
3683
3684         tmp = b11 ^ b2;
3685         b11 = (tmp >> 20) | (tmp << (64 - 20));
3686         b2 -= b11;
3687
3688         tmp = b15 ^ b0;
3689         b15 = (tmp >> 5) | (tmp << (64 - 5));
3690         b0 -= b15;
3691
3692         tmp = b9 ^ b10;
3693         b9 = (tmp >> 17) | (tmp << (64 - 17));
3694         b10 -= b9;
3695
3696         tmp = b11 ^ b8;
3697         b11 = (tmp >> 59) | (tmp << (64 - 59));
3698         b8 -= b11;
3699
3700         tmp = b13 ^ b14;
3701         b13 = (tmp >> 41) | (tmp << (64 - 41));
3702         b14 -= b13;
3703
3704         tmp = b15 ^ b12;
3705         b15 = (tmp >> 34) | (tmp << (64 - 34));
3706         b12 -= b15;
3707
3708         tmp = b1 ^ b6;
3709         b1 = (tmp >> 13) | (tmp << (64 - 13));
3710         b6 -= b1;
3711
3712         tmp = b3 ^ b4;
3713         b3 = (tmp >> 51) | (tmp << (64 - 51));
3714         b4 -= b3;
3715
3716         tmp = b5 ^ b2;
3717         b5 = (tmp >> 4) | (tmp << (64 - 4));
3718         b2 -= b5;
3719
3720         tmp = b7 ^ b0;
3721         b7 = (tmp >> 33) | (tmp << (64 - 33));
3722         b0 -= b7;
3723
3724         tmp = b1 ^ b8;
3725         b1 = (tmp >> 52) | (tmp << (64 - 52));
3726         b8 -= b1;
3727
3728         tmp = b5 ^ b14;
3729         b5 = (tmp >> 23) | (tmp << (64 - 23));
3730         b14 -= b5;
3731
3732         tmp = b3 ^ b12;
3733         b3 = (tmp >> 18) | (tmp << (64 - 18));
3734         b12 -= b3;
3735
3736         tmp = b7 ^ b10;
3737         b7 = (tmp >> 49) | (tmp << (64 - 49));
3738         b10 -= b7;
3739
3740         tmp = b15 ^ b4;
3741         b15 = (tmp >> 55) | (tmp << (64 - 55));
3742         b4 -= b15;
3743
3744         tmp = b11 ^ b6;
3745         b11 = (tmp >> 10) | (tmp << (64 - 10));
3746         b6 -= b11;
3747
3748         tmp = b13 ^ b2;
3749         b13 = (tmp >> 19) | (tmp << (64 - 19));
3750         b2 -= b13;
3751
3752         tmp = b9 ^ b0;
3753         b9 = (tmp >> 38) | (tmp << (64 - 38));
3754         b0 -= b9;
3755
3756         tmp = b15 ^ b14;
3757         b15 = (tmp >> 37) | (tmp << (64 - 37));
3758         b14 -= b15 + k5 + t0;
3759         b15 -= k6 + 8;
3760
3761         tmp = b13 ^ b12;
3762         b13 = (tmp >> 22) | (tmp << (64 - 22));
3763         b12 -= b13 + k3;
3764         b13 -= k4 + t2;
3765
3766         tmp = b11 ^ b10;
3767         b11 = (tmp >> 17) | (tmp << (64 - 17));
3768         b10 -= b11 + k1;
3769         b11 -= k2;
3770
3771         tmp = b9 ^ b8;
3772         b9 = (tmp >> 8) | (tmp << (64 - 8));
3773         b8 -= b9 + k16;
3774         b9 -= k0;
3775
3776         tmp = b7 ^ b6;
3777         b7 = (tmp >> 47) | (tmp << (64 - 47));
3778         b6 -= b7 + k14;
3779         b7 -= k15;
3780
3781         tmp = b5 ^ b4;
3782         b5 = (tmp >> 8) | (tmp << (64 - 8));
3783         b4 -= b5 + k12;
3784         b5 -= k13;
3785
3786         tmp = b3 ^ b2;
3787         b3 = (tmp >> 13) | (tmp << (64 - 13));
3788         b2 -= b3 + k10;
3789         b3 -= k11;
3790
3791         tmp = b1 ^ b0;
3792         b1 = (tmp >> 24) | (tmp << (64 - 24));
3793         b0 -= b1 + k8;
3794         b1 -= k9;
3795
3796         tmp = b7 ^ b12;
3797         b7 = (tmp >> 20) | (tmp << (64 - 20));
3798         b12 -= b7;
3799
3800         tmp = b3 ^ b10;
3801         b3 = (tmp >> 37) | (tmp << (64 - 37));
3802         b10 -= b3;
3803
3804         tmp = b5 ^ b8;
3805         b5 = (tmp >> 31) | (tmp << (64 - 31));
3806         b8 -= b5;
3807
3808         tmp = b1 ^ b14;
3809         b1 = (tmp >> 23) | (tmp << (64 - 23));
3810         b14 -= b1;
3811
3812         tmp = b9 ^ b4;
3813         b9 = (tmp >> 52) | (tmp << (64 - 52));
3814         b4 -= b9;
3815
3816         tmp = b13 ^ b6;
3817         b13 = (tmp >> 35) | (tmp << (64 - 35));
3818         b6 -= b13;
3819
3820         tmp = b11 ^ b2;
3821         b11 = (tmp >> 48) | (tmp << (64 - 48));
3822         b2 -= b11;
3823
3824         tmp = b15 ^ b0;
3825         b15 = (tmp >> 9) | (tmp << (64 - 9));
3826         b0 -= b15;
3827
3828         tmp = b9 ^ b10;
3829         b9 = (tmp >> 25) | (tmp << (64 - 25));
3830         b10 -= b9;
3831
3832         tmp = b11 ^ b8;
3833         b11 = (tmp >> 44) | (tmp << (64 - 44));
3834         b8 -= b11;
3835
3836         tmp = b13 ^ b14;
3837         b13 = (tmp >> 42) | (tmp << (64 - 42));
3838         b14 -= b13;
3839
3840         tmp = b15 ^ b12;
3841         b15 = (tmp >> 19) | (tmp << (64 - 19));
3842         b12 -= b15;
3843
3844         tmp = b1 ^ b6;
3845         b1 = (tmp >> 46) | (tmp << (64 - 46));
3846         b6 -= b1;
3847
3848         tmp = b3 ^ b4;
3849         b3 = (tmp >> 47) | (tmp << (64 - 47));
3850         b4 -= b3;
3851
3852         tmp = b5 ^ b2;
3853         b5 = (tmp >> 44) | (tmp << (64 - 44));
3854         b2 -= b5;
3855
3856         tmp = b7 ^ b0;
3857         b7 = (tmp >> 31) | (tmp << (64 - 31));
3858         b0 -= b7;
3859
3860         tmp = b1 ^ b8;
3861         b1 = (tmp >> 41) | (tmp << (64 - 41));
3862         b8 -= b1;
3863
3864         tmp = b5 ^ b14;
3865         b5 = (tmp >> 42) | (tmp << (64 - 42));
3866         b14 -= b5;
3867
3868         tmp = b3 ^ b12;
3869         b3 = (tmp >> 53) | (tmp << (64 - 53));
3870         b12 -= b3;
3871
3872         tmp = b7 ^ b10;
3873         b7 = (tmp >> 4) | (tmp << (64 - 4));
3874         b10 -= b7;
3875
3876         tmp = b15 ^ b4;
3877         b15 = (tmp >> 51) | (tmp << (64 - 51));
3878         b4 -= b15;
3879
3880         tmp = b11 ^ b6;
3881         b11 = (tmp >> 56) | (tmp << (64 - 56));
3882         b6 -= b11;
3883
3884         tmp = b13 ^ b2;
3885         b13 = (tmp >> 34) | (tmp << (64 - 34));
3886         b2 -= b13;
3887
3888         tmp = b9 ^ b0;
3889         b9 = (tmp >> 16) | (tmp << (64 - 16));
3890         b0 -= b9;
3891
3892         tmp = b15 ^ b14;
3893         b15 = (tmp >> 30) | (tmp << (64 - 30));
3894         b14 -= b15 + k4 + t2;
3895         b15 -= k5 + 7;
3896
3897         tmp = b13 ^ b12;
3898         b13 = (tmp >> 44) | (tmp << (64 - 44));
3899         b12 -= b13 + k2;
3900         b13 -= k3 + t1;
3901
3902         tmp = b11 ^ b10;
3903         b11 = (tmp >> 47) | (tmp << (64 - 47));
3904         b10 -= b11 + k0;
3905         b11 -= k1;
3906
3907         tmp = b9 ^ b8;
3908         b9 = (tmp >> 12) | (tmp << (64 - 12));
3909         b8 -= b9 + k15;
3910         b9 -= k16;
3911
3912         tmp = b7 ^ b6;
3913         b7 = (tmp >> 31) | (tmp << (64 - 31));
3914         b6 -= b7 + k13;
3915         b7 -= k14;
3916
3917         tmp = b5 ^ b4;
3918         b5 = (tmp >> 37) | (tmp << (64 - 37));
3919         b4 -= b5 + k11;
3920         b5 -= k12;
3921
3922         tmp = b3 ^ b2;
3923         b3 = (tmp >> 9) | (tmp << (64 - 9));
3924         b2 -= b3 + k9;
3925         b3 -= k10;
3926
3927         tmp = b1 ^ b0;
3928         b1 = (tmp >> 41) | (tmp << (64 - 41));
3929         b0 -= b1 + k7;
3930         b1 -= k8;
3931
3932         tmp = b7 ^ b12;
3933         b7 = (tmp >> 25) | (tmp << (64 - 25));
3934         b12 -= b7;
3935
3936         tmp = b3 ^ b10;
3937         b3 = (tmp >> 16) | (tmp << (64 - 16));
3938         b10 -= b3;
3939
3940         tmp = b5 ^ b8;
3941         b5 = (tmp >> 28) | (tmp << (64 - 28));
3942         b8 -= b5;
3943
3944         tmp = b1 ^ b14;
3945         b1 = (tmp >> 47) | (tmp << (64 - 47));
3946         b14 -= b1;
3947
3948         tmp = b9 ^ b4;
3949         b9 = (tmp >> 41) | (tmp << (64 - 41));
3950         b4 -= b9;
3951
3952         tmp = b13 ^ b6;
3953         b13 = (tmp >> 48) | (tmp << (64 - 48));
3954         b6 -= b13;
3955
3956         tmp = b11 ^ b2;
3957         b11 = (tmp >> 20) | (tmp << (64 - 20));
3958         b2 -= b11;
3959
3960         tmp = b15 ^ b0;
3961         b15 = (tmp >> 5) | (tmp << (64 - 5));
3962         b0 -= b15;
3963
3964         tmp = b9 ^ b10;
3965         b9 = (tmp >> 17) | (tmp << (64 - 17));
3966         b10 -= b9;
3967
3968         tmp = b11 ^ b8;
3969         b11 = (tmp >> 59) | (tmp << (64 - 59));
3970         b8 -= b11;
3971
3972         tmp = b13 ^ b14;
3973         b13 = (tmp >> 41) | (tmp << (64 - 41));
3974         b14 -= b13;
3975
3976         tmp = b15 ^ b12;
3977         b15 = (tmp >> 34) | (tmp << (64 - 34));
3978         b12 -= b15;
3979
3980         tmp = b1 ^ b6;
3981         b1 = (tmp >> 13) | (tmp << (64 - 13));
3982         b6 -= b1;
3983
3984         tmp = b3 ^ b4;
3985         b3 = (tmp >> 51) | (tmp << (64 - 51));
3986         b4 -= b3;
3987
3988         tmp = b5 ^ b2;
3989         b5 = (tmp >> 4) | (tmp << (64 - 4));
3990         b2 -= b5;
3991
3992         tmp = b7 ^ b0;
3993         b7 = (tmp >> 33) | (tmp << (64 - 33));
3994         b0 -= b7;
3995
3996         tmp = b1 ^ b8;
3997         b1 = (tmp >> 52) | (tmp << (64 - 52));
3998         b8 -= b1;
3999
4000         tmp = b5 ^ b14;
4001         b5 = (tmp >> 23) | (tmp << (64 - 23));
4002         b14 -= b5;
4003
4004         tmp = b3 ^ b12;
4005         b3 = (tmp >> 18) | (tmp << (64 - 18));
4006         b12 -= b3;
4007
4008         tmp = b7 ^ b10;
4009         b7 = (tmp >> 49) | (tmp << (64 - 49));
4010         b10 -= b7;
4011
4012         tmp = b15 ^ b4;
4013         b15 = (tmp >> 55) | (tmp << (64 - 55));
4014         b4 -= b15;
4015
4016         tmp = b11 ^ b6;
4017         b11 = (tmp >> 10) | (tmp << (64 - 10));
4018         b6 -= b11;
4019
4020         tmp = b13 ^ b2;
4021         b13 = (tmp >> 19) | (tmp << (64 - 19));
4022         b2 -= b13;
4023
4024         tmp = b9 ^ b0;
4025         b9 = (tmp >> 38) | (tmp << (64 - 38));
4026         b0 -= b9;
4027
4028         tmp = b15 ^ b14;
4029         b15 = (tmp >> 37) | (tmp << (64 - 37));
4030         b14 -= b15 + k3 + t1;
4031         b15 -= k4 + 6;
4032
4033         tmp = b13 ^ b12;
4034         b13 = (tmp >> 22) | (tmp << (64 - 22));
4035         b12 -= b13 + k1;
4036         b13 -= k2 + t0;
4037
4038         tmp = b11 ^ b10;
4039         b11 = (tmp >> 17) | (tmp << (64 - 17));
4040         b10 -= b11 + k16;
4041         b11 -= k0;
4042
4043         tmp = b9 ^ b8;
4044         b9 = (tmp >> 8) | (tmp << (64 - 8));
4045         b8 -= b9 + k14;
4046         b9 -= k15;
4047
4048         tmp = b7 ^ b6;
4049         b7 = (tmp >> 47) | (tmp << (64 - 47));
4050         b6 -= b7 + k12;
4051         b7 -= k13;
4052
4053         tmp = b5 ^ b4;
4054         b5 = (tmp >> 8) | (tmp << (64 - 8));
4055         b4 -= b5 + k10;
4056         b5 -= k11;
4057
4058         tmp = b3 ^ b2;
4059         b3 = (tmp >> 13) | (tmp << (64 - 13));
4060         b2 -= b3 + k8;
4061         b3 -= k9;
4062
4063         tmp = b1 ^ b0;
4064         b1 = (tmp >> 24) | (tmp << (64 - 24));
4065         b0 -= b1 + k6;
4066         b1 -= k7;
4067
4068         tmp = b7 ^ b12;
4069         b7 = (tmp >> 20) | (tmp << (64 - 20));
4070         b12 -= b7;
4071
4072         tmp = b3 ^ b10;
4073         b3 = (tmp >> 37) | (tmp << (64 - 37));
4074         b10 -= b3;
4075
4076         tmp = b5 ^ b8;
4077         b5 = (tmp >> 31) | (tmp << (64 - 31));
4078         b8 -= b5;
4079
4080         tmp = b1 ^ b14;
4081         b1 = (tmp >> 23) | (tmp << (64 - 23));
4082         b14 -= b1;
4083
4084         tmp = b9 ^ b4;
4085         b9 = (tmp >> 52) | (tmp << (64 - 52));
4086         b4 -= b9;
4087
4088         tmp = b13 ^ b6;
4089         b13 = (tmp >> 35) | (tmp << (64 - 35));
4090         b6 -= b13;
4091
4092         tmp = b11 ^ b2;
4093         b11 = (tmp >> 48) | (tmp << (64 - 48));
4094         b2 -= b11;
4095
4096         tmp = b15 ^ b0;
4097         b15 = (tmp >> 9) | (tmp << (64 - 9));
4098         b0 -= b15;
4099
4100         tmp = b9 ^ b10;
4101         b9 = (tmp >> 25) | (tmp << (64 - 25));
4102         b10 -= b9;
4103
4104         tmp = b11 ^ b8;
4105         b11 = (tmp >> 44) | (tmp << (64 - 44));
4106         b8 -= b11;
4107
4108         tmp = b13 ^ b14;
4109         b13 = (tmp >> 42) | (tmp << (64 - 42));
4110         b14 -= b13;
4111
4112         tmp = b15 ^ b12;
4113         b15 = (tmp >> 19) | (tmp << (64 - 19));
4114         b12 -= b15;
4115
4116         tmp = b1 ^ b6;
4117         b1 = (tmp >> 46) | (tmp << (64 - 46));
4118         b6 -= b1;
4119
4120         tmp = b3 ^ b4;
4121         b3 = (tmp >> 47) | (tmp << (64 - 47));
4122         b4 -= b3;
4123
4124         tmp = b5 ^ b2;
4125         b5 = (tmp >> 44) | (tmp << (64 - 44));
4126         b2 -= b5;
4127
4128         tmp = b7 ^ b0;
4129         b7 = (tmp >> 31) | (tmp << (64 - 31));
4130         b0 -= b7;
4131
4132         tmp = b1 ^ b8;
4133         b1 = (tmp >> 41) | (tmp << (64 - 41));
4134         b8 -= b1;
4135
4136         tmp = b5 ^ b14;
4137         b5 = (tmp >> 42) | (tmp << (64 - 42));
4138         b14 -= b5;
4139
4140         tmp = b3 ^ b12;
4141         b3 = (tmp >> 53) | (tmp << (64 - 53));
4142         b12 -= b3;
4143
4144         tmp = b7 ^ b10;
4145         b7 = (tmp >> 4) | (tmp << (64 - 4));
4146         b10 -= b7;
4147
4148         tmp = b15 ^ b4;
4149         b15 = (tmp >> 51) | (tmp << (64 - 51));
4150         b4 -= b15;
4151
4152         tmp = b11 ^ b6;
4153         b11 = (tmp >> 56) | (tmp << (64 - 56));
4154         b6 -= b11;
4155
4156         tmp = b13 ^ b2;
4157         b13 = (tmp >> 34) | (tmp << (64 - 34));
4158         b2 -= b13;
4159
4160         tmp = b9 ^ b0;
4161         b9 = (tmp >> 16) | (tmp << (64 - 16));
4162         b0 -= b9;
4163
4164         tmp = b15 ^ b14;
4165         b15 = (tmp >> 30) | (tmp << (64 - 30));
4166         b14 -= b15 + k2 + t0;
4167         b15 -= k3 + 5;
4168
4169         tmp = b13 ^ b12;
4170         b13 = (tmp >> 44) | (tmp << (64 - 44));
4171         b12 -= b13 + k0;
4172         b13 -= k1 + t2;
4173
4174         tmp = b11 ^ b10;
4175         b11 = (tmp >> 47) | (tmp << (64 - 47));
4176         b10 -= b11 + k15;
4177         b11 -= k16;
4178
4179         tmp = b9 ^ b8;
4180         b9 = (tmp >> 12) | (tmp << (64 - 12));
4181         b8 -= b9 + k13;
4182         b9 -= k14;
4183
4184         tmp = b7 ^ b6;
4185         b7 = (tmp >> 31) | (tmp << (64 - 31));
4186         b6 -= b7 + k11;
4187         b7 -= k12;
4188
4189         tmp = b5 ^ b4;
4190         b5 = (tmp >> 37) | (tmp << (64 - 37));
4191         b4 -= b5 + k9;
4192         b5 -= k10;
4193
4194         tmp = b3 ^ b2;
4195         b3 = (tmp >> 9) | (tmp << (64 - 9));
4196         b2 -= b3 + k7;
4197         b3 -= k8;
4198
4199         tmp = b1 ^ b0;
4200         b1 = (tmp >> 41) | (tmp << (64 - 41));
4201         b0 -= b1 + k5;
4202         b1 -= k6;
4203
4204         tmp = b7 ^ b12;
4205         b7 = (tmp >> 25) | (tmp << (64 - 25));
4206         b12 -= b7;
4207
4208         tmp = b3 ^ b10;
4209         b3 = (tmp >> 16) | (tmp << (64 - 16));
4210         b10 -= b3;
4211
4212         tmp = b5 ^ b8;
4213         b5 = (tmp >> 28) | (tmp << (64 - 28));
4214         b8 -= b5;
4215
4216         tmp = b1 ^ b14;
4217         b1 = (tmp >> 47) | (tmp << (64 - 47));
4218         b14 -= b1;
4219
4220         tmp = b9 ^ b4;
4221         b9 = (tmp >> 41) | (tmp << (64 - 41));
4222         b4 -= b9;
4223
4224         tmp = b13 ^ b6;
4225         b13 = (tmp >> 48) | (tmp << (64 - 48));
4226         b6 -= b13;
4227
4228         tmp = b11 ^ b2;
4229         b11 = (tmp >> 20) | (tmp << (64 - 20));
4230         b2 -= b11;
4231
4232         tmp = b15 ^ b0;
4233         b15 = (tmp >> 5) | (tmp << (64 - 5));
4234         b0 -= b15;
4235
4236         tmp = b9 ^ b10;
4237         b9 = (tmp >> 17) | (tmp << (64 - 17));
4238         b10 -= b9;
4239
4240         tmp = b11 ^ b8;
4241         b11 = (tmp >> 59) | (tmp << (64 - 59));
4242         b8 -= b11;
4243
4244         tmp = b13 ^ b14;
4245         b13 = (tmp >> 41) | (tmp << (64 - 41));
4246         b14 -= b13;
4247
4248         tmp = b15 ^ b12;
4249         b15 = (tmp >> 34) | (tmp << (64 - 34));
4250         b12 -= b15;
4251
4252         tmp = b1 ^ b6;
4253         b1 = (tmp >> 13) | (tmp << (64 - 13));
4254         b6 -= b1;
4255
4256         tmp = b3 ^ b4;
4257         b3 = (tmp >> 51) | (tmp << (64 - 51));
4258         b4 -= b3;
4259
4260         tmp = b5 ^ b2;
4261         b5 = (tmp >> 4) | (tmp << (64 - 4));
4262         b2 -= b5;
4263
4264         tmp = b7 ^ b0;
4265         b7 = (tmp >> 33) | (tmp << (64 - 33));
4266         b0 -= b7;
4267
4268         tmp = b1 ^ b8;
4269         b1 = (tmp >> 52) | (tmp << (64 - 52));
4270         b8 -= b1;
4271
4272         tmp = b5 ^ b14;
4273         b5 = (tmp >> 23) | (tmp << (64 - 23));
4274         b14 -= b5;
4275
4276         tmp = b3 ^ b12;
4277         b3 = (tmp >> 18) | (tmp << (64 - 18));
4278         b12 -= b3;
4279
4280         tmp = b7 ^ b10;
4281         b7 = (tmp >> 49) | (tmp << (64 - 49));
4282         b10 -= b7;
4283
4284         tmp = b15 ^ b4;
4285         b15 = (tmp >> 55) | (tmp << (64 - 55));
4286         b4 -= b15;
4287
4288         tmp = b11 ^ b6;
4289         b11 = (tmp >> 10) | (tmp << (64 - 10));
4290         b6 -= b11;
4291
4292         tmp = b13 ^ b2;
4293         b13 = (tmp >> 19) | (tmp << (64 - 19));
4294         b2 -= b13;
4295
4296         tmp = b9 ^ b0;
4297         b9 = (tmp >> 38) | (tmp << (64 - 38));
4298         b0 -= b9;
4299
4300         tmp = b15 ^ b14;
4301         b15 = (tmp >> 37) | (tmp << (64 - 37));
4302         b14 -= b15 + k1 + t2;
4303         b15 -= k2 + 4;
4304
4305         tmp = b13 ^ b12;
4306         b13 = (tmp >> 22) | (tmp << (64 - 22));
4307         b12 -= b13 + k16;
4308         b13 -= k0 + t1;
4309
4310         tmp = b11 ^ b10;
4311         b11 = (tmp >> 17) | (tmp << (64 - 17));
4312         b10 -= b11 + k14;
4313         b11 -= k15;
4314
4315         tmp = b9 ^ b8;
4316         b9 = (tmp >> 8) | (tmp << (64 - 8));
4317         b8 -= b9 + k12;
4318         b9 -= k13;
4319
4320         tmp = b7 ^ b6;
4321         b7 = (tmp >> 47) | (tmp << (64 - 47));
4322         b6 -= b7 + k10;
4323         b7 -= k11;
4324
4325         tmp = b5 ^ b4;
4326         b5 = (tmp >> 8) | (tmp << (64 - 8));
4327         b4 -= b5 + k8;
4328         b5 -= k9;
4329
4330         tmp = b3 ^ b2;
4331         b3 = (tmp >> 13) | (tmp << (64 - 13));
4332         b2 -= b3 + k6;
4333         b3 -= k7;
4334
4335         tmp = b1 ^ b0;
4336         b1 = (tmp >> 24) | (tmp << (64 - 24));
4337         b0 -= b1 + k4;
4338         b1 -= k5;
4339
4340         tmp = b7 ^ b12;
4341         b7 = (tmp >> 20) | (tmp << (64 - 20));
4342         b12 -= b7;
4343
4344         tmp = b3 ^ b10;
4345         b3 = (tmp >> 37) | (tmp << (64 - 37));
4346         b10 -= b3;
4347
4348         tmp = b5 ^ b8;
4349         b5 = (tmp >> 31) | (tmp << (64 - 31));
4350         b8 -= b5;
4351
4352         tmp = b1 ^ b14;
4353         b1 = (tmp >> 23) | (tmp << (64 - 23));
4354         b14 -= b1;
4355
4356         tmp = b9 ^ b4;
4357         b9 = (tmp >> 52) | (tmp << (64 - 52));
4358         b4 -= b9;
4359
4360         tmp = b13 ^ b6;
4361         b13 = (tmp >> 35) | (tmp << (64 - 35));
4362         b6 -= b13;
4363
4364         tmp = b11 ^ b2;
4365         b11 = (tmp >> 48) | (tmp << (64 - 48));
4366         b2 -= b11;
4367
4368         tmp = b15 ^ b0;
4369         b15 = (tmp >> 9) | (tmp << (64 - 9));
4370         b0 -= b15;
4371
4372         tmp = b9 ^ b10;
4373         b9 = (tmp >> 25) | (tmp << (64 - 25));
4374         b10 -= b9;
4375
4376         tmp = b11 ^ b8;
4377         b11 = (tmp >> 44) | (tmp << (64 - 44));
4378         b8 -= b11;
4379
4380         tmp = b13 ^ b14;
4381         b13 = (tmp >> 42) | (tmp << (64 - 42));
4382         b14 -= b13;
4383
4384         tmp = b15 ^ b12;
4385         b15 = (tmp >> 19) | (tmp << (64 - 19));
4386         b12 -= b15;
4387
4388         tmp = b1 ^ b6;
4389         b1 = (tmp >> 46) | (tmp << (64 - 46));
4390         b6 -= b1;
4391
4392         tmp = b3 ^ b4;
4393         b3 = (tmp >> 47) | (tmp << (64 - 47));
4394         b4 -= b3;
4395
4396         tmp = b5 ^ b2;
4397         b5 = (tmp >> 44) | (tmp << (64 - 44));
4398         b2 -= b5;
4399
4400         tmp = b7 ^ b0;
4401         b7 = (tmp >> 31) | (tmp << (64 - 31));
4402         b0 -= b7;
4403
4404         tmp = b1 ^ b8;
4405         b1 = (tmp >> 41) | (tmp << (64 - 41));
4406         b8 -= b1;
4407
4408         tmp = b5 ^ b14;
4409         b5 = (tmp >> 42) | (tmp << (64 - 42));
4410         b14 -= b5;
4411
4412         tmp = b3 ^ b12;
4413         b3 = (tmp >> 53) | (tmp << (64 - 53));
4414         b12 -= b3;
4415
4416         tmp = b7 ^ b10;
4417         b7 = (tmp >> 4) | (tmp << (64 - 4));
4418         b10 -= b7;
4419
4420         tmp = b15 ^ b4;
4421         b15 = (tmp >> 51) | (tmp << (64 - 51));
4422         b4 -= b15;
4423
4424         tmp = b11 ^ b6;
4425         b11 = (tmp >> 56) | (tmp << (64 - 56));
4426         b6 -= b11;
4427
4428         tmp = b13 ^ b2;
4429         b13 = (tmp >> 34) | (tmp << (64 - 34));
4430         b2 -= b13;
4431
4432         tmp = b9 ^ b0;
4433         b9 = (tmp >> 16) | (tmp << (64 - 16));
4434         b0 -= b9;
4435
4436         tmp = b15 ^ b14;
4437         b15 = (tmp >> 30) | (tmp << (64 - 30));
4438         b14 -= b15 + k0 + t1;
4439         b15 -= k1 + 3;
4440
4441         tmp = b13 ^ b12;
4442         b13 = (tmp >> 44) | (tmp << (64 - 44));
4443         b12 -= b13 + k15;
4444         b13 -= k16 + t0;
4445
4446         tmp = b11 ^ b10;
4447         b11 = (tmp >> 47) | (tmp << (64 - 47));
4448         b10 -= b11 + k13;
4449         b11 -= k14;
4450
4451         tmp = b9 ^ b8;
4452         b9 = (tmp >> 12) | (tmp << (64 - 12));
4453         b8 -= b9 + k11;
4454         b9 -= k12;
4455
4456         tmp = b7 ^ b6;
4457         b7 = (tmp >> 31) | (tmp << (64 - 31));
4458         b6 -= b7 + k9;
4459         b7 -= k10;
4460
4461         tmp = b5 ^ b4;
4462         b5 = (tmp >> 37) | (tmp << (64 - 37));
4463         b4 -= b5 + k7;
4464         b5 -= k8;
4465
4466         tmp = b3 ^ b2;
4467         b3 = (tmp >> 9) | (tmp << (64 - 9));
4468         b2 -= b3 + k5;
4469         b3 -= k6;
4470
4471         tmp = b1 ^ b0;
4472         b1 = (tmp >> 41) | (tmp << (64 - 41));
4473         b0 -= b1 + k3;
4474         b1 -= k4;
4475
4476         tmp = b7 ^ b12;
4477         b7 = (tmp >> 25) | (tmp << (64 - 25));
4478         b12 -= b7;
4479
4480         tmp = b3 ^ b10;
4481         b3 = (tmp >> 16) | (tmp << (64 - 16));
4482         b10 -= b3;
4483
4484         tmp = b5 ^ b8;
4485         b5 = (tmp >> 28) | (tmp << (64 - 28));
4486         b8 -= b5;
4487
4488         tmp = b1 ^ b14;
4489         b1 = (tmp >> 47) | (tmp << (64 - 47));
4490         b14 -= b1;
4491
4492         tmp = b9 ^ b4;
4493         b9 = (tmp >> 41) | (tmp << (64 - 41));
4494         b4 -= b9;
4495
4496         tmp = b13 ^ b6;
4497         b13 = (tmp >> 48) | (tmp << (64 - 48));
4498         b6 -= b13;
4499
4500         tmp = b11 ^ b2;
4501         b11 = (tmp >> 20) | (tmp << (64 - 20));
4502         b2 -= b11;
4503
4504         tmp = b15 ^ b0;
4505         b15 = (tmp >> 5) | (tmp << (64 - 5));
4506         b0 -= b15;
4507
4508         tmp = b9 ^ b10;
4509         b9 = (tmp >> 17) | (tmp << (64 - 17));
4510         b10 -= b9;
4511
4512         tmp = b11 ^ b8;
4513         b11 = (tmp >> 59) | (tmp << (64 - 59));
4514         b8 -= b11;
4515
4516         tmp = b13 ^ b14;
4517         b13 = (tmp >> 41) | (tmp << (64 - 41));
4518         b14 -= b13;
4519
4520         tmp = b15 ^ b12;
4521         b15 = (tmp >> 34) | (tmp << (64 - 34));
4522         b12 -= b15;
4523
4524         tmp = b1 ^ b6;
4525         b1 = (tmp >> 13) | (tmp << (64 - 13));
4526         b6 -= b1;
4527
4528         tmp = b3 ^ b4;
4529         b3 = (tmp >> 51) | (tmp << (64 - 51));
4530         b4 -= b3;
4531
4532         tmp = b5 ^ b2;
4533         b5 = (tmp >> 4) | (tmp << (64 - 4));
4534         b2 -= b5;
4535
4536         tmp = b7 ^ b0;
4537         b7 = (tmp >> 33) | (tmp << (64 - 33));
4538         b0 -= b7;
4539
4540         tmp = b1 ^ b8;
4541         b1 = (tmp >> 52) | (tmp << (64 - 52));
4542         b8 -= b1;
4543
4544         tmp = b5 ^ b14;
4545         b5 = (tmp >> 23) | (tmp << (64 - 23));
4546         b14 -= b5;
4547
4548         tmp = b3 ^ b12;
4549         b3 = (tmp >> 18) | (tmp << (64 - 18));
4550         b12 -= b3;
4551
4552         tmp = b7 ^ b10;
4553         b7 = (tmp >> 49) | (tmp << (64 - 49));
4554         b10 -= b7;
4555
4556         tmp = b15 ^ b4;
4557         b15 = (tmp >> 55) | (tmp << (64 - 55));
4558         b4 -= b15;
4559
4560         tmp = b11 ^ b6;
4561         b11 = (tmp >> 10) | (tmp << (64 - 10));
4562         b6 -= b11;
4563
4564         tmp = b13 ^ b2;
4565         b13 = (tmp >> 19) | (tmp << (64 - 19));
4566         b2 -= b13;
4567
4568         tmp = b9 ^ b0;
4569         b9 = (tmp >> 38) | (tmp << (64 - 38));
4570         b0 -= b9;
4571
4572         tmp = b15 ^ b14;
4573         b15 = (tmp >> 37) | (tmp << (64 - 37));
4574         b14 -= b15 + k16 + t0;
4575         b15 -= k0 + 2;
4576
4577         tmp = b13 ^ b12;
4578         b13 = (tmp >> 22) | (tmp << (64 - 22));
4579         b12 -= b13 + k14;
4580         b13 -= k15 + t2;
4581
4582         tmp = b11 ^ b10;
4583         b11 = (tmp >> 17) | (tmp << (64 - 17));
4584         b10 -= b11 + k12;
4585         b11 -= k13;
4586
4587         tmp = b9 ^ b8;
4588         b9 = (tmp >> 8) | (tmp << (64 - 8));
4589         b8 -= b9 + k10;
4590         b9 -= k11;
4591
4592         tmp = b7 ^ b6;
4593         b7 = (tmp >> 47) | (tmp << (64 - 47));
4594         b6 -= b7 + k8;
4595         b7 -= k9;
4596
4597         tmp = b5 ^ b4;
4598         b5 = (tmp >> 8) | (tmp << (64 - 8));
4599         b4 -= b5 + k6;
4600         b5 -= k7;
4601
4602         tmp = b3 ^ b2;
4603         b3 = (tmp >> 13) | (tmp << (64 - 13));
4604         b2 -= b3 + k4;
4605         b3 -= k5;
4606
4607         tmp = b1 ^ b0;
4608         b1 = (tmp >> 24) | (tmp << (64 - 24));
4609         b0 -= b1 + k2;
4610         b1 -= k3;
4611
4612         tmp = b7 ^ b12;
4613         b7 = (tmp >> 20) | (tmp << (64 - 20));
4614         b12 -= b7;
4615
4616         tmp = b3 ^ b10;
4617         b3 = (tmp >> 37) | (tmp << (64 - 37));
4618         b10 -= b3;
4619
4620         tmp = b5 ^ b8;
4621         b5 = (tmp >> 31) | (tmp << (64 - 31));
4622         b8 -= b5;
4623
4624         tmp = b1 ^ b14;
4625         b1 = (tmp >> 23) | (tmp << (64 - 23));
4626         b14 -= b1;
4627
4628         tmp = b9 ^ b4;
4629         b9 = (tmp >> 52) | (tmp << (64 - 52));
4630         b4 -= b9;
4631
4632         tmp = b13 ^ b6;
4633         b13 = (tmp >> 35) | (tmp << (64 - 35));
4634         b6 -= b13;
4635
4636         tmp = b11 ^ b2;
4637         b11 = (tmp >> 48) | (tmp << (64 - 48));
4638         b2 -= b11;
4639
4640         tmp = b15 ^ b0;
4641         b15 = (tmp >> 9) | (tmp << (64 - 9));
4642         b0 -= b15;
4643
4644         tmp = b9 ^ b10;
4645         b9 = (tmp >> 25) | (tmp << (64 - 25));
4646         b10 -= b9;
4647
4648         tmp = b11 ^ b8;
4649         b11 = (tmp >> 44) | (tmp << (64 - 44));
4650         b8 -= b11;
4651
4652         tmp = b13 ^ b14;
4653         b13 = (tmp >> 42) | (tmp << (64 - 42));
4654         b14 -= b13;
4655
4656         tmp = b15 ^ b12;
4657         b15 = (tmp >> 19) | (tmp << (64 - 19));
4658         b12 -= b15;
4659
4660         tmp = b1 ^ b6;
4661         b1 = (tmp >> 46) | (tmp << (64 - 46));
4662         b6 -= b1;
4663
4664         tmp = b3 ^ b4;
4665         b3 = (tmp >> 47) | (tmp << (64 - 47));
4666         b4 -= b3;
4667
4668         tmp = b5 ^ b2;
4669         b5 = (tmp >> 44) | (tmp << (64 - 44));
4670         b2 -= b5;
4671
4672         tmp = b7 ^ b0;
4673         b7 = (tmp >> 31) | (tmp << (64 - 31));
4674         b0 -= b7;
4675
4676         tmp = b1 ^ b8;
4677         b1 = (tmp >> 41) | (tmp << (64 - 41));
4678         b8 -= b1;
4679
4680         tmp = b5 ^ b14;
4681         b5 = (tmp >> 42) | (tmp << (64 - 42));
4682         b14 -= b5;
4683
4684         tmp = b3 ^ b12;
4685         b3 = (tmp >> 53) | (tmp << (64 - 53));
4686         b12 -= b3;
4687
4688         tmp = b7 ^ b10;
4689         b7 = (tmp >> 4) | (tmp << (64 - 4));
4690         b10 -= b7;
4691
4692         tmp = b15 ^ b4;
4693         b15 = (tmp >> 51) | (tmp << (64 - 51));
4694         b4 -= b15;
4695
4696         tmp = b11 ^ b6;
4697         b11 = (tmp >> 56) | (tmp << (64 - 56));
4698         b6 -= b11;
4699
4700         tmp = b13 ^ b2;
4701         b13 = (tmp >> 34) | (tmp << (64 - 34));
4702         b2 -= b13;
4703
4704         tmp = b9 ^ b0;
4705         b9 = (tmp >> 16) | (tmp << (64 - 16));
4706         b0 -= b9;
4707
4708         tmp = b15 ^ b14;
4709         b15 = (tmp >> 30) | (tmp << (64 - 30));
4710         b14 -= b15 + k15 + t2;
4711         b15 -= k16 + 1;
4712
4713         tmp = b13 ^ b12;
4714         b13 = (tmp >> 44) | (tmp << (64 - 44));
4715         b12 -= b13 + k13;
4716         b13 -= k14 + t1;
4717
4718         tmp = b11 ^ b10;
4719         b11 = (tmp >> 47) | (tmp << (64 - 47));
4720         b10 -= b11 + k11;
4721         b11 -= k12;
4722
4723         tmp = b9 ^ b8;
4724         b9 = (tmp >> 12) | (tmp << (64 - 12));
4725         b8 -= b9 + k9;
4726         b9 -= k10;
4727
4728         tmp = b7 ^ b6;
4729         b7 = (tmp >> 31) | (tmp << (64 - 31));
4730         b6 -= b7 + k7;
4731         b7 -= k8;
4732
4733         tmp = b5 ^ b4;
4734         b5 = (tmp >> 37) | (tmp << (64 - 37));
4735         b4 -= b5 + k5;
4736         b5 -= k6;
4737
4738         tmp = b3 ^ b2;
4739         b3 = (tmp >> 9) | (tmp << (64 - 9));
4740         b2 -= b3 + k3;
4741         b3 -= k4;
4742
4743         tmp = b1 ^ b0;
4744         b1 = (tmp >> 41) | (tmp << (64 - 41));
4745         b0 -= b1 + k1;
4746         b1 -= k2;
4747
4748         tmp = b7 ^ b12;
4749         b7 = (tmp >> 25) | (tmp << (64 - 25));
4750         b12 -= b7;
4751
4752         tmp = b3 ^ b10;
4753         b3 = (tmp >> 16) | (tmp << (64 - 16));
4754         b10 -= b3;
4755
4756         tmp = b5 ^ b8;
4757         b5 = (tmp >> 28) | (tmp << (64 - 28));
4758         b8 -= b5;
4759
4760         tmp = b1 ^ b14;
4761         b1 = (tmp >> 47) | (tmp << (64 - 47));
4762         b14 -= b1;
4763
4764         tmp = b9 ^ b4;
4765         b9 = (tmp >> 41) | (tmp << (64 - 41));
4766         b4 -= b9;
4767
4768         tmp = b13 ^ b6;
4769         b13 = (tmp >> 48) | (tmp << (64 - 48));
4770         b6 -= b13;
4771
4772         tmp = b11 ^ b2;
4773         b11 = (tmp >> 20) | (tmp << (64 - 20));
4774         b2 -= b11;
4775
4776         tmp = b15 ^ b0;
4777         b15 = (tmp >> 5) | (tmp << (64 - 5));
4778         b0 -= b15;
4779
4780         tmp = b9 ^ b10;
4781         b9 = (tmp >> 17) | (tmp << (64 - 17));
4782         b10 -= b9;
4783
4784         tmp = b11 ^ b8;
4785         b11 = (tmp >> 59) | (tmp << (64 - 59));
4786         b8 -= b11;
4787
4788         tmp = b13 ^ b14;
4789         b13 = (tmp >> 41) | (tmp << (64 - 41));
4790         b14 -= b13;
4791
4792         tmp = b15 ^ b12;
4793         b15 = (tmp >> 34) | (tmp << (64 - 34));
4794         b12 -= b15;
4795
4796         tmp = b1 ^ b6;
4797         b1 = (tmp >> 13) | (tmp << (64 - 13));
4798         b6 -= b1;
4799
4800         tmp = b3 ^ b4;
4801         b3 = (tmp >> 51) | (tmp << (64 - 51));
4802         b4 -= b3;
4803
4804         tmp = b5 ^ b2;
4805         b5 = (tmp >> 4) | (tmp << (64 - 4));
4806         b2 -= b5;
4807
4808         tmp = b7 ^ b0;
4809         b7 = (tmp >> 33) | (tmp << (64 - 33));
4810         b0 -= b7;
4811
4812         tmp = b1 ^ b8;
4813         b1 = (tmp >> 52) | (tmp << (64 - 52));
4814         b8 -= b1;
4815
4816         tmp = b5 ^ b14;
4817         b5 = (tmp >> 23) | (tmp << (64 - 23));
4818         b14 -= b5;
4819
4820         tmp = b3 ^ b12;
4821         b3 = (tmp >> 18) | (tmp << (64 - 18));
4822         b12 -= b3;
4823
4824         tmp = b7 ^ b10;
4825         b7 = (tmp >> 49) | (tmp << (64 - 49));
4826         b10 -= b7;
4827
4828         tmp = b15 ^ b4;
4829         b15 = (tmp >> 55) | (tmp << (64 - 55));
4830         b4 -= b15;
4831
4832         tmp = b11 ^ b6;
4833         b11 = (tmp >> 10) | (tmp << (64 - 10));
4834         b6 -= b11;
4835
4836         tmp = b13 ^ b2;
4837         b13 = (tmp >> 19) | (tmp << (64 - 19));
4838         b2 -= b13;
4839
4840         tmp = b9 ^ b0;
4841         b9 = (tmp >> 38) | (tmp << (64 - 38));
4842         b0 -= b9;
4843
4844         tmp = b15 ^ b14;
4845         b15 = (tmp >> 37) | (tmp << (64 - 37));
4846         b14 -= b15 + k14 + t1;
4847         b15 -= k15;
4848
4849         tmp = b13 ^ b12;
4850         b13 = (tmp >> 22) | (tmp << (64 - 22));
4851         b12 -= b13 + k12;
4852         b13 -= k13 + t0;
4853
4854         tmp = b11 ^ b10;
4855         b11 = (tmp >> 17) | (tmp << (64 - 17));
4856         b10 -= b11 + k10;
4857         b11 -= k11;
4858
4859         tmp = b9 ^ b8;
4860         b9 = (tmp >> 8) | (tmp << (64 - 8));
4861         b8 -= b9 + k8;
4862         b9 -= k9;
4863
4864         tmp = b7 ^ b6;
4865         b7 = (tmp >> 47) | (tmp << (64 - 47));
4866         b6 -= b7 + k6;
4867         b7 -= k7;
4868
4869         tmp = b5 ^ b4;
4870         b5 = (tmp >> 8) | (tmp << (64 - 8));
4871         b4 -= b5 + k4;
4872         b5 -= k5;
4873
4874         tmp = b3 ^ b2;
4875         b3 = (tmp >> 13) | (tmp << (64 - 13));
4876         b2 -= b3 + k2;
4877         b3 -= k3;
4878
4879         tmp = b1 ^ b0;
4880         b1 = (tmp >> 24) | (tmp << (64 - 24));
4881         b0 -= b1 + k0;
4882         b1 -= k1;
4883
4884         output[15] = b15;
4885         output[14] = b14;
4886         output[13] = b13;
4887         output[12] = b12;
4888         output[11] = b11;
4889         output[10] = b10;
4890         output[9] = b9;
4891         output[8] = b8;
4892         output[7] = b7;
4893         output[6] = b6;
4894         output[5] = b5;
4895         output[4] = b4;
4896         output[3] = b3;
4897         output[2] = b2;
4898         output[1] = b1;
4899         output[0] = b0;
4900 }