CoCalc -- c_skein.c

GitHub Repository: tpruvot/cpuminer-multi
Path: blob/linux/crypto/c_skein.c
¹²⁰¹ views
1
/***********************************************************************
2
**
3
** Implementation of the Skein hash function.
4
**
5
** Source code author: Doug Whiting, 2008.
6
**
7
** This algorithm and source code is released to the public domain.
8
** 
9
************************************************************************/
10

11
#define  SKEIN_PORT_CODE /* instantiate any code in skein_port.h */
12

13
#include <stddef.h>                          /* get size_t definition */
14
#include <string.h>      /* get the memcpy/memset functions */
15
#include "c_skein.h"       /* get the Skein API definitions   */
16

17
#define DISABLE_UNUSED 0
18

19
#ifndef SKEIN_256_NIST_MAX_HASHBITS
20
#define SKEIN_256_NIST_MAX_HASHBITS (0)
21
#endif
22

23
#ifndef SKEIN_512_NIST_MAX_HASHBITS
24
#define SKEIN_512_NIST_MAX_HASHBITS (512)
25
#endif
26

27
#define  SKEIN_MODIFIER_WORDS  ( 2)          /* number of modifier (tweak) words */
28

29
#define  SKEIN_256_STATE_WORDS ( 4)
30
#define  SKEIN_512_STATE_WORDS ( 8)
31
#define  SKEIN1024_STATE_WORDS (16)
32
#define  SKEIN_MAX_STATE_WORDS (16)
33

34
#define  SKEIN_256_STATE_BYTES ( 8*SKEIN_256_STATE_WORDS)
35
#define  SKEIN_512_STATE_BYTES ( 8*SKEIN_512_STATE_WORDS)
36
#define  SKEIN1024_STATE_BYTES ( 8*SKEIN1024_STATE_WORDS)
37

38
#define  SKEIN_256_STATE_BITS  (64*SKEIN_256_STATE_WORDS)
39
#define  SKEIN_512_STATE_BITS  (64*SKEIN_512_STATE_WORDS)
40
#define  SKEIN1024_STATE_BITS  (64*SKEIN1024_STATE_WORDS)
41

42
#define  SKEIN_256_BLOCK_BYTES ( 8*SKEIN_256_STATE_WORDS)
43
#define  SKEIN_512_BLOCK_BYTES ( 8*SKEIN_512_STATE_WORDS)
44
#define  SKEIN1024_BLOCK_BYTES ( 8*SKEIN1024_STATE_WORDS)
45

46
#define SKEIN_RND_SPECIAL       (1000u)
47
#define SKEIN_RND_KEY_INITIAL   (SKEIN_RND_SPECIAL+0u)
48
#define SKEIN_RND_KEY_INJECT    (SKEIN_RND_SPECIAL+1u)
49
#define SKEIN_RND_FEED_FWD      (SKEIN_RND_SPECIAL+2u)
50

51
typedef struct
52
{
53
  size_t  hashBitLen;                      /* size of hash result, in bits */
54
  size_t  bCnt;                            /* current byte count in buffer b[] */
55
  u64b_t  T[SKEIN_MODIFIER_WORDS];         /* tweak words: T[0]=byte cnt, T[1]=flags */
56
} Skein_Ctxt_Hdr_t;
57

58
typedef struct                               /*  256-bit Skein hash context structure */
59
{
60
  Skein_Ctxt_Hdr_t h;                      /* common header context variables */
61
  u64b_t  X[SKEIN_256_STATE_WORDS];        /* chaining variables */
62
  u08b_t  b[SKEIN_256_BLOCK_BYTES];        /* partial block buffer (8-byte aligned) */
63
} Skein_256_Ctxt_t;
64

65
typedef struct                               /*  512-bit Skein hash context structure */
66
{
67
  Skein_Ctxt_Hdr_t h;                      /* common header context variables */
68
  u64b_t  X[SKEIN_512_STATE_WORDS];        /* chaining variables */
69
  u08b_t  b[SKEIN_512_BLOCK_BYTES];        /* partial block buffer (8-byte aligned) */
70
} Skein_512_Ctxt_t;
71

72
typedef struct                               /* 1024-bit Skein hash context structure */
73
{
74
  Skein_Ctxt_Hdr_t h;                      /* common header context variables */
75
  u64b_t  X[SKEIN1024_STATE_WORDS];        /* chaining variables */
76
  u08b_t  b[SKEIN1024_BLOCK_BYTES];        /* partial block buffer (8-byte aligned) */
77
} Skein1024_Ctxt_t;
78

79
/*   Skein APIs for (incremental) "straight hashing" */
80
#if SKEIN_256_NIST_MAX_HASH_BITS
81
static int  Skein_256_Init  (Skein_256_Ctxt_t *ctx, size_t hashBitLen);
82
#endif
83
static int  Skein_512_Init  (Skein_512_Ctxt_t *ctx, size_t hashBitLen);
84
static int  Skein1024_Init  (Skein1024_Ctxt_t *ctx, size_t hashBitLen);
85

86
static int  Skein_256_Update(Skein_256_Ctxt_t *ctx, const u08b_t *msg, size_t msgByteCnt);
87
static int  Skein_512_Update(Skein_512_Ctxt_t *ctx, const u08b_t *msg, size_t msgByteCnt);
88
static int  Skein1024_Update(Skein1024_Ctxt_t *ctx, const u08b_t *msg, size_t msgByteCnt);
89

90
static int  Skein_256_Final (Skein_256_Ctxt_t *ctx, u08b_t * hashVal);
91
static int  Skein_512_Final (Skein_512_Ctxt_t *ctx, u08b_t * hashVal);
92
static int  Skein1024_Final (Skein1024_Ctxt_t *ctx, u08b_t * hashVal);
93

94
/*
95
**   Skein APIs for "extended" initialization: MAC keys, tree hashing.
96
**   After an InitExt() call, just use Update/Final calls as with Init().
97
**
98
**   Notes: Same parameters as _Init() calls, plus treeInfo/key/keyBytes.
99
**          When keyBytes == 0 and treeInfo == SKEIN_SEQUENTIAL, 
100
**              the results of InitExt() are identical to calling Init().
101
**          The function Init() may be called once to "precompute" the IV for
102
**              a given hashBitLen value, then by saving a copy of the context
103
**              the IV computation may be avoided in later calls.
104
**          Similarly, the function InitExt() may be called once per MAC key 
105
**              to precompute the MAC IV, then a copy of the context saved and
106
**              reused for each new MAC computation.
107
**/
108
#if 0
109
static int  Skein_256_InitExt(Skein_256_Ctxt_t *ctx, size_t hashBitLen, u64b_t treeInfo, const u08b_t *key, size_t keyBytes);
110
static int  Skein_512_InitExt(Skein_512_Ctxt_t *ctx, size_t hashBitLen, u64b_t treeInfo, const u08b_t *key, size_t keyBytes);
111
static int  Skein1024_InitExt(Skein1024_Ctxt_t *ctx, size_t hashBitLen, u64b_t treeInfo, const u08b_t *key, size_t keyBytes);
112
#endif
113

114
/*
115
**   Skein APIs for MAC and tree hash:
116
**      Final_Pad:  pad, do final block, but no OUTPUT type
117
**      Output:     do just the output stage
118
*/
119
#if 0
120
static int  Skein_256_Final_Pad(Skein_256_Ctxt_t *ctx, u08b_t * hashVal);
121
static int  Skein_512_Final_Pad(Skein_512_Ctxt_t *ctx, u08b_t * hashVal);
122
static int  Skein1024_Final_Pad(Skein1024_Ctxt_t *ctx, u08b_t * hashVal);
123
#endif
124

125
#ifndef SKEIN_TREE_HASH
126
#define SKEIN_TREE_HASH (1)
127
#endif
128
#if 0
129
#if  SKEIN_TREE_HASH
130
static int  Skein_256_Output   (Skein_256_Ctxt_t *ctx, u08b_t * hashVal);
131
static int  Skein_512_Output   (Skein_512_Ctxt_t *ctx, u08b_t * hashVal);
132
static int  Skein1024_Output   (Skein1024_Ctxt_t *ctx, u08b_t * hashVal);
133
#endif
134
#endif
135

136
/*****************************************************************
137
** "Internal" Skein definitions
138
**    -- not needed for sequential hashing API, but will be 
139
**           helpful for other uses of Skein (e.g., tree hash mode).
140
**    -- included here so that they can be shared between
141
**           reference and optimized code.
142
******************************************************************/
143

144
/* tweak word T[1]: bit field starting positions */
145
#define SKEIN_T1_BIT(BIT)       ((BIT) - 64)            /* offset 64 because it's the second word  */
146

147
#define SKEIN_T1_POS_TREE_LVL   SKEIN_T1_BIT(112)       /* bits 112..118: level in hash tree       */
148
#define SKEIN_T1_POS_BIT_PAD    SKEIN_T1_BIT(119)       /* bit  119     : partial final input byte */
149
#define SKEIN_T1_POS_BLK_TYPE   SKEIN_T1_BIT(120)       /* bits 120..125: type field               */
150
#define SKEIN_T1_POS_FIRST      SKEIN_T1_BIT(126)       /* bits 126     : first block flag         */
151
#define SKEIN_T1_POS_FINAL      SKEIN_T1_BIT(127)       /* bit  127     : final block flag         */
152

153
/* tweak word T[1]: flag bit definition(s) */
154
#define SKEIN_T1_FLAG_FIRST     (((u64b_t)  1 ) << SKEIN_T1_POS_FIRST)
155
#define SKEIN_T1_FLAG_FINAL     (((u64b_t)  1 ) << SKEIN_T1_POS_FINAL)
156
#define SKEIN_T1_FLAG_BIT_PAD   (((u64b_t)  1 ) << SKEIN_T1_POS_BIT_PAD)
157

158
/* tweak word T[1]: tree level bit field mask */
159
#define SKEIN_T1_TREE_LVL_MASK  (((u64b_t)0x7F) << SKEIN_T1_POS_TREE_LVL)
160
#define SKEIN_T1_TREE_LEVEL(n)  (((u64b_t) (n)) << SKEIN_T1_POS_TREE_LVL)
161

162
/* tweak word T[1]: block type field */
163
#define SKEIN_BLK_TYPE_KEY      ( 0)                    /* key, for MAC and KDF */
164
#define SKEIN_BLK_TYPE_CFG      ( 4)                    /* configuration block */
165
#define SKEIN_BLK_TYPE_PERS     ( 8)                    /* personalization string */
166
#define SKEIN_BLK_TYPE_PK       (12)                    /* public key (for digital signature hashing) */
167
#define SKEIN_BLK_TYPE_KDF      (16)                    /* key identifier for KDF */
168
#define SKEIN_BLK_TYPE_NONCE    (20)                    /* nonce for PRNG */
169
#define SKEIN_BLK_TYPE_MSG      (48)                    /* message processing */
170
#define SKEIN_BLK_TYPE_OUT      (63)                    /* output stage */
171
#define SKEIN_BLK_TYPE_MASK     (63)                    /* bit field mask */
172

173
#define SKEIN_T1_BLK_TYPE(T)   (((u64b_t) (SKEIN_BLK_TYPE_##T)) << SKEIN_T1_POS_BLK_TYPE)
174
#define SKEIN_T1_BLK_TYPE_KEY   SKEIN_T1_BLK_TYPE(KEY)  /* key, for MAC and KDF */
175
#define SKEIN_T1_BLK_TYPE_CFG   SKEIN_T1_BLK_TYPE(CFG)  /* configuration block */
176
#define SKEIN_T1_BLK_TYPE_PERS  SKEIN_T1_BLK_TYPE(PERS) /* personalization string */
177
#define SKEIN_T1_BLK_TYPE_PK    SKEIN_T1_BLK_TYPE(PK)   /* public key (for digital signature hashing) */
178
#define SKEIN_T1_BLK_TYPE_KDF   SKEIN_T1_BLK_TYPE(KDF)  /* key identifier for KDF */
179
#define SKEIN_T1_BLK_TYPE_NONCE SKEIN_T1_BLK_TYPE(NONCE)/* nonce for PRNG */
180
#define SKEIN_T1_BLK_TYPE_MSG   SKEIN_T1_BLK_TYPE(MSG)  /* message processing */
181
#define SKEIN_T1_BLK_TYPE_OUT   SKEIN_T1_BLK_TYPE(OUT)  /* output stage */
182
#define SKEIN_T1_BLK_TYPE_MASK  SKEIN_T1_BLK_TYPE(MASK) /* field bit mask */
183

184
#define SKEIN_T1_BLK_TYPE_CFG_FINAL       (SKEIN_T1_BLK_TYPE_CFG | SKEIN_T1_FLAG_FINAL)
185
#define SKEIN_T1_BLK_TYPE_OUT_FINAL       (SKEIN_T1_BLK_TYPE_OUT | SKEIN_T1_FLAG_FINAL)
186

187
#define SKEIN_VERSION           (1)
188

189
#ifndef SKEIN_ID_STRING_LE      /* allow compile-time personalization */
190
#define SKEIN_ID_STRING_LE      (0x33414853)            /* "SHA3" (little-endian)*/
191
#endif
192

193
#define SKEIN_MK_64(hi32,lo32)  ((lo32) + (((u64b_t) (hi32)) << 32))
194
#define SKEIN_SCHEMA_VER        SKEIN_MK_64(SKEIN_VERSION,SKEIN_ID_STRING_LE)
195
#define SKEIN_KS_PARITY         SKEIN_MK_64(0x1BD11BDA,0xA9FC1A22)
196

197
#define SKEIN_CFG_STR_LEN       (4*8)
198

199
/* bit field definitions in config block treeInfo word */
200
#define SKEIN_CFG_TREE_LEAF_SIZE_POS  ( 0)
201
#define SKEIN_CFG_TREE_NODE_SIZE_POS  ( 8)
202
#define SKEIN_CFG_TREE_MAX_LEVEL_POS  (16)
203

204
#define SKEIN_CFG_TREE_LEAF_SIZE_MSK  (((u64b_t) 0xFF) << SKEIN_CFG_TREE_LEAF_SIZE_POS)
205
#define SKEIN_CFG_TREE_NODE_SIZE_MSK  (((u64b_t) 0xFF) << SKEIN_CFG_TREE_NODE_SIZE_POS)
206
#define SKEIN_CFG_TREE_MAX_LEVEL_MSK  (((u64b_t) 0xFF) << SKEIN_CFG_TREE_MAX_LEVEL_POS)
207

208
#define SKEIN_CFG_TREE_INFO(leaf,node,maxLvl)                   \
209
  ( (((u64b_t)(leaf  )) << SKEIN_CFG_TREE_LEAF_SIZE_POS) |    \
210
  (((u64b_t)(node  )) << SKEIN_CFG_TREE_NODE_SIZE_POS) |    \
211
  (((u64b_t)(maxLvl)) << SKEIN_CFG_TREE_MAX_LEVEL_POS) )
212

213
#define SKEIN_CFG_TREE_INFO_SEQUENTIAL SKEIN_CFG_TREE_INFO(0,0,0) /* use as treeInfo in InitExt() call for sequential processing */
214

215
/*
216
**   Skein macros for getting/setting tweak words, etc.
217
**   These are useful for partial input bytes, hash tree init/update, etc.
218
**/
219
#define Skein_Get_Tweak(ctxPtr,TWK_NUM)         ((ctxPtr)->h.T[TWK_NUM])
220
#define Skein_Set_Tweak(ctxPtr,TWK_NUM,tVal)    {(ctxPtr)->h.T[TWK_NUM] = (tVal);}
221

222
#define Skein_Get_T0(ctxPtr)    Skein_Get_Tweak(ctxPtr,0)
223
#define Skein_Get_T1(ctxPtr)    Skein_Get_Tweak(ctxPtr,1)
224
#define Skein_Set_T0(ctxPtr,T0) Skein_Set_Tweak(ctxPtr,0,T0)
225
#define Skein_Set_T1(ctxPtr,T1) Skein_Set_Tweak(ctxPtr,1,T1)
226

227
/* set both tweak words at once */
228
#define Skein_Set_T0_T1(ctxPtr,T0,T1)           \
229
{                                           \
230
  Skein_Set_T0(ctxPtr,(T0));                  \
231
  Skein_Set_T1(ctxPtr,(T1));                  \
232
}
233

234
#define Skein_Set_Type(ctxPtr,BLK_TYPE)         \
235
  Skein_Set_T1(ctxPtr,SKEIN_T1_BLK_TYPE_##BLK_TYPE)
236

237
/* set up for starting with a new type: h.T[0]=0; h.T[1] = NEW_TYPE; h.bCnt=0; */
238
#define Skein_Start_New_Type(ctxPtr,BLK_TYPE)   \
239
{ Skein_Set_T0_T1(ctxPtr,0,SKEIN_T1_FLAG_FIRST | SKEIN_T1_BLK_TYPE_##BLK_TYPE); (ctxPtr)->h.bCnt=0; }
240

241
#define Skein_Clear_First_Flag(hdr)      { (hdr).T[1] &= ~SKEIN_T1_FLAG_FIRST;       }
242
#define Skein_Set_Bit_Pad_Flag(hdr)      { (hdr).T[1] |=  SKEIN_T1_FLAG_BIT_PAD;     }
243

244
#define Skein_Set_Tree_Level(hdr,height) { (hdr).T[1] |= SKEIN_T1_TREE_LEVEL(height);}
245

246
/*****************************************************************
247
** "Internal" Skein definitions for debugging and error checking
248
******************************************************************/
249
#define Skein_Show_Block(bits,ctx,X,blkPtr,wPtr,ksEvenPtr,ksOddPtr)
250
#define Skein_Show_Round(bits,ctx,r,X)
251
#define Skein_Show_R_Ptr(bits,ctx,r,X_ptr)
252
#define Skein_Show_Final(bits,ctx,cnt,outPtr)
253
#define Skein_Show_Key(bits,ctx,key,keyBytes)
254

255

256
#ifndef SKEIN_ERR_CHECK        /* run-time checks (e.g., bad params, uninitialized context)? */
257
#define Skein_Assert(x,retCode)/* default: ignore all Asserts, for performance */
258
#define Skein_assert(x)
259
#elif   defined(SKEIN_ASSERT)
260
#include <assert.h>     
261
#define Skein_Assert(x,retCode) assert(x) 
262
#define Skein_assert(x)         assert(x) 
263
#else
264
#include <assert.h>     
265
#define Skein_Assert(x,retCode) { if (!(x)) return retCode; } /*  caller  error */
266
#define Skein_assert(x)         assert(x)                     /* internal error */
267
#endif
268

269
/*****************************************************************
270
** Skein block function constants (shared across Ref and Opt code)
271
******************************************************************/
272
enum    
273
{   
274
  /* Skein_256 round rotation constants */
275
  R_256_0_0=14, R_256_0_1=16,
276
  R_256_1_0=52, R_256_1_1=57,
277
  R_256_2_0=23, R_256_2_1=40,
278
  R_256_3_0= 5, R_256_3_1=37,
279
  R_256_4_0=25, R_256_4_1=33,
280
  R_256_5_0=46, R_256_5_1=12,
281
  R_256_6_0=58, R_256_6_1=22,
282
  R_256_7_0=32, R_256_7_1=32,
283

284
  /* Skein_512 round rotation constants */
285
  R_512_0_0=46, R_512_0_1=36, R_512_0_2=19, R_512_0_3=37,
286
  R_512_1_0=33, R_512_1_1=27, R_512_1_2=14, R_512_1_3=42,
287
  R_512_2_0=17, R_512_2_1=49, R_512_2_2=36, R_512_2_3=39,
288
  R_512_3_0=44, R_512_3_1= 9, R_512_3_2=54, R_512_3_3=56,
289
  R_512_4_0=39, R_512_4_1=30, R_512_4_2=34, R_512_4_3=24,
290
  R_512_5_0=13, R_512_5_1=50, R_512_5_2=10, R_512_5_3=17,
291
  R_512_6_0=25, R_512_6_1=29, R_512_6_2=39, R_512_6_3=43,
292
  R_512_7_0= 8, R_512_7_1=35, R_512_7_2=56, R_512_7_3=22,
293

294
  /* Skein1024 round rotation constants */
295
  R1024_0_0=24, R1024_0_1=13, R1024_0_2= 8, R1024_0_3=47, R1024_0_4= 8, R1024_0_5=17, R1024_0_6=22, R1024_0_7=37,
296
  R1024_1_0=38, R1024_1_1=19, R1024_1_2=10, R1024_1_3=55, R1024_1_4=49, R1024_1_5=18, R1024_1_6=23, R1024_1_7=52,
297
  R1024_2_0=33, R1024_2_1= 4, R1024_2_2=51, R1024_2_3=13, R1024_2_4=34, R1024_2_5=41, R1024_2_6=59, R1024_2_7=17,
298
  R1024_3_0= 5, R1024_3_1=20, R1024_3_2=48, R1024_3_3=41, R1024_3_4=47, R1024_3_5=28, R1024_3_6=16, R1024_3_7=25,
299
  R1024_4_0=41, R1024_4_1= 9, R1024_4_2=37, R1024_4_3=31, R1024_4_4=12, R1024_4_5=47, R1024_4_6=44, R1024_4_7=30,
300
  R1024_5_0=16, R1024_5_1=34, R1024_5_2=56, R1024_5_3=51, R1024_5_4= 4, R1024_5_5=53, R1024_5_6=42, R1024_5_7=41,
301
  R1024_6_0=31, R1024_6_1=44, R1024_6_2=47, R1024_6_3=46, R1024_6_4=19, R1024_6_5=42, R1024_6_6=44, R1024_6_7=25,
302
  R1024_7_0= 9, R1024_7_1=48, R1024_7_2=35, R1024_7_3=52, R1024_7_4=23, R1024_7_5=31, R1024_7_6=37, R1024_7_7=20
303
};
304

305
#ifndef SKEIN_ROUNDS
306
#define SKEIN_256_ROUNDS_TOTAL (72)          /* number of rounds for the different block sizes */
307
#define SKEIN_512_ROUNDS_TOTAL (72)
308
#define SKEIN1024_ROUNDS_TOTAL (80)
309
#else                                        /* allow command-line define in range 8*(5..14)   */
310
#define SKEIN_256_ROUNDS_TOTAL (8*((((SKEIN_ROUNDS/100) + 5) % 10) + 5))
311
#define SKEIN_512_ROUNDS_TOTAL (8*((((SKEIN_ROUNDS/ 10) + 5) % 10) + 5))
312
#define SKEIN1024_ROUNDS_TOTAL (8*((((SKEIN_ROUNDS    ) + 5) % 10) + 5))
313
#endif
314

315

316
/*
317
***************** Pre-computed Skein IVs *******************
318
**
319
** NOTE: these values are not "magic" constants, but
320
** are generated using the Threefish block function.
321
** They are pre-computed here only for speed; i.e., to
322
** avoid the need for a Threefish call during Init().
323
**
324
** The IV for any fixed hash length may be pre-computed.
325
** Only the most common values are included here.
326
**
327
************************************************************
328
**/
329

330
#define MK_64 SKEIN_MK_64
331

332
/* blkSize =  256 bits. hashSize =  128 bits */
333
const u64b_t SKEIN_256_IV_128[] =
334
    {
335
    MK_64(0xE1111906,0x964D7260),
336
    MK_64(0x883DAAA7,0x7C8D811C),
337
    MK_64(0x10080DF4,0x91960F7A),
338
    MK_64(0xCCF7DDE5,0xB45BC1C2)
339
    };
340

341
/* blkSize =  256 bits. hashSize =  160 bits */
342
const u64b_t SKEIN_256_IV_160[] =
343
    {
344
    MK_64(0x14202314,0x72825E98),
345
    MK_64(0x2AC4E9A2,0x5A77E590),
346
    MK_64(0xD47A5856,0x8838D63E),
347
    MK_64(0x2DD2E496,0x8586AB7D)
348
    };
349

350
/* blkSize =  256 bits. hashSize =  224 bits */
351
const u64b_t SKEIN_256_IV_224[] =
352
    {
353
    MK_64(0xC6098A8C,0x9AE5EA0B),
354
    MK_64(0x876D5686,0x08C5191C),
355
    MK_64(0x99CB88D7,0xD7F53884),
356
    MK_64(0x384BDDB1,0xAEDDB5DE)
357
    };
358

359
/* blkSize =  256 bits. hashSize =  256 bits */
360
const u64b_t SKEIN_256_IV_256[] =
361
    {
362
    MK_64(0xFC9DA860,0xD048B449),
363
    MK_64(0x2FCA6647,0x9FA7D833),
364
    MK_64(0xB33BC389,0x6656840F),
365
    MK_64(0x6A54E920,0xFDE8DA69)
366
    };
367

368
/* blkSize =  512 bits. hashSize =  128 bits */
369
const u64b_t SKEIN_512_IV_128[] =
370
    {
371
    MK_64(0xA8BC7BF3,0x6FBF9F52),
372
    MK_64(0x1E9872CE,0xBD1AF0AA),
373
    MK_64(0x309B1790,0xB32190D3),
374
    MK_64(0xBCFBB854,0x3F94805C),
375
    MK_64(0x0DA61BCD,0x6E31B11B),
376
    MK_64(0x1A18EBEA,0xD46A32E3),
377
    MK_64(0xA2CC5B18,0xCE84AA82),
378
    MK_64(0x6982AB28,0x9D46982D)
379
    };
380

381
/* blkSize =  512 bits. hashSize =  160 bits */
382
const u64b_t SKEIN_512_IV_160[] =
383
    {
384
    MK_64(0x28B81A2A,0xE013BD91),
385
    MK_64(0xC2F11668,0xB5BDF78F),
386
    MK_64(0x1760D8F3,0xF6A56F12),
387
    MK_64(0x4FB74758,0x8239904F),
388
    MK_64(0x21EDE07F,0x7EAF5056),
389
    MK_64(0xD908922E,0x63ED70B8),
390
    MK_64(0xB8EC76FF,0xECCB52FA),
391
    MK_64(0x01A47BB8,0xA3F27A6E)
392
    };
393

394
/* blkSize =  512 bits. hashSize =  224 bits */
395
const u64b_t SKEIN_512_IV_224[] =
396
    {
397
    MK_64(0xCCD06162,0x48677224),
398
    MK_64(0xCBA65CF3,0xA92339EF),
399
    MK_64(0x8CCD69D6,0x52FF4B64),
400
    MK_64(0x398AED7B,0x3AB890B4),
401
    MK_64(0x0F59D1B1,0x457D2BD0),
402
    MK_64(0x6776FE65,0x75D4EB3D),
403
    MK_64(0x99FBC70E,0x997413E9),
404
    MK_64(0x9E2CFCCF,0xE1C41EF7)
405
    };
406

407
/* blkSize =  512 bits. hashSize =  256 bits */
408
const u64b_t SKEIN_512_IV_256[] =
409
    {
410
    MK_64(0xCCD044A1,0x2FDB3E13),
411
    MK_64(0xE8359030,0x1A79A9EB),
412
    MK_64(0x55AEA061,0x4F816E6F),
413
    MK_64(0x2A2767A4,0xAE9B94DB),
414
    MK_64(0xEC06025E,0x74DD7683),
415
    MK_64(0xE7A436CD,0xC4746251),
416
    MK_64(0xC36FBAF9,0x393AD185),
417
    MK_64(0x3EEDBA18,0x33EDFC13)
418
    };
419

420
/* blkSize =  512 bits. hashSize =  384 bits */
421
const u64b_t SKEIN_512_IV_384[] =
422
    {
423
    MK_64(0xA3F6C6BF,0x3A75EF5F),
424
    MK_64(0xB0FEF9CC,0xFD84FAA4),
425
    MK_64(0x9D77DD66,0x3D770CFE),
426
    MK_64(0xD798CBF3,0xB468FDDA),
427
    MK_64(0x1BC4A666,0x8A0E4465),
428
    MK_64(0x7ED7D434,0xE5807407),
429
    MK_64(0x548FC1AC,0xD4EC44D6),
430
    MK_64(0x266E1754,0x6AA18FF8)
431
    };
432

433
/* blkSize =  512 bits. hashSize =  512 bits */
434
const u64b_t SKEIN_512_IV_512[] =
435
    {
436
    MK_64(0x4903ADFF,0x749C51CE),
437
    MK_64(0x0D95DE39,0x9746DF03),
438
    MK_64(0x8FD19341,0x27C79BCE),
439
    MK_64(0x9A255629,0xFF352CB1),
440
    MK_64(0x5DB62599,0xDF6CA7B0),
441
    MK_64(0xEABE394C,0xA9D5C3F4),
442
    MK_64(0x991112C7,0x1A75B523),
443
    MK_64(0xAE18A40B,0x660FCC33)
444
    };
445

446
/* blkSize = 1024 bits. hashSize =  384 bits */
447
const u64b_t SKEIN1024_IV_384[] =
448
    {
449
    MK_64(0x5102B6B8,0xC1894A35),
450
    MK_64(0xFEEBC9E3,0xFE8AF11A),
451
    MK_64(0x0C807F06,0xE32BED71),
452
    MK_64(0x60C13A52,0xB41A91F6),
453
    MK_64(0x9716D35D,0xD4917C38),
454
    MK_64(0xE780DF12,0x6FD31D3A),
455
    MK_64(0x797846B6,0xC898303A),
456
    MK_64(0xB172C2A8,0xB3572A3B),
457
    MK_64(0xC9BC8203,0xA6104A6C),
458
    MK_64(0x65909338,0xD75624F4),
459
    MK_64(0x94BCC568,0x4B3F81A0),
460
    MK_64(0x3EBBF51E,0x10ECFD46),
461
    MK_64(0x2DF50F0B,0xEEB08542),
462
    MK_64(0x3B5A6530,0x0DBC6516),
463
    MK_64(0x484B9CD2,0x167BBCE1),
464
    MK_64(0x2D136947,0xD4CBAFEA)
465
    };
466

467
/* blkSize = 1024 bits. hashSize =  512 bits */
468
const u64b_t SKEIN1024_IV_512[] =
469
    {
470
    MK_64(0xCAEC0E5D,0x7C1B1B18),
471
    MK_64(0xA01B0E04,0x5F03E802),
472
    MK_64(0x33840451,0xED912885),
473
    MK_64(0x374AFB04,0xEAEC2E1C),
474
    MK_64(0xDF25A0E2,0x813581F7),
475
    MK_64(0xE4004093,0x8B12F9D2),
476
    MK_64(0xA662D539,0xC2ED39B6),
477
    MK_64(0xFA8B85CF,0x45D8C75A),
478
    MK_64(0x8316ED8E,0x29EDE796),
479
    MK_64(0x053289C0,0x2E9F91B8),
480
    MK_64(0xC3F8EF1D,0x6D518B73),
481
    MK_64(0xBDCEC3C4,0xD5EF332E),
482
    MK_64(0x549A7E52,0x22974487),
483
    MK_64(0x67070872,0x5B749816),
484
    MK_64(0xB9CD28FB,0xF0581BD1),
485
    MK_64(0x0E2940B8,0x15804974)
486
    };
487

488
/* blkSize = 1024 bits. hashSize = 1024 bits */
489
const u64b_t SKEIN1024_IV_1024[] =
490
    {
491
    MK_64(0xD593DA07,0x41E72355),
492
    MK_64(0x15B5E511,0xAC73E00C),
493
    MK_64(0x5180E5AE,0xBAF2C4F0),
494
    MK_64(0x03BD41D3,0xFCBCAFAF),
495
    MK_64(0x1CAEC6FD,0x1983A898),
496
    MK_64(0x6E510B8B,0xCDD0589F),
497
    MK_64(0x77E2BDFD,0xC6394ADA),
498
    MK_64(0xC11E1DB5,0x24DCB0A3),
499
    MK_64(0xD6D14AF9,0xC6329AB5),
500
    MK_64(0x6A9B0BFC,0x6EB67E0D),
501
    MK_64(0x9243C60D,0xCCFF1332),
502
    MK_64(0x1A1F1DDE,0x743F02D4),
503
    MK_64(0x0996753C,0x10ED0BB8),
504
    MK_64(0x6572DD22,0xF2B4969A),
505
    MK_64(0x61FD3062,0xD00A579A),
506
    MK_64(0x1DE0536E,0x8682E539)
507
    };
508

509

510
#ifndef SKEIN_USE_ASM
511
#define SKEIN_USE_ASM   (0)                     /* default is all C code (no ASM) */
512
#endif
513

514
#ifndef SKEIN_LOOP
515
#define SKEIN_LOOP 001                          /* default: unroll 256 and 512, but not 1024 */
516
#endif
517

518
#define BLK_BITS        (WCNT*64)               /* some useful definitions for code here */
519
#define KW_TWK_BASE     (0)
520
#define KW_KEY_BASE     (3)
521
#define ks              (kw + KW_KEY_BASE)                
522
#define ts              (kw + KW_TWK_BASE)
523

524
#ifdef SKEIN_DEBUG
525
#define DebugSaveTweak(ctx) { ctx->h.T[0] = ts[0]; ctx->h.T[1] = ts[1]; }
526
#else
527
#define DebugSaveTweak(ctx)
528
#endif
529

530
/*****************************  Skein_256 ******************************/
531
#if !(SKEIN_USE_ASM & 256)
532
static void Skein_256_Process_Block(Skein_256_Ctxt_t *ctx,const u08b_t *blkPtr,size_t blkCnt,size_t byteCntAdd)
533
    { /* do it in C */
534
    enum
535
        {
536
        WCNT = SKEIN_256_STATE_WORDS
537
        };
538
#undef  RCNT
539
#define RCNT  (SKEIN_256_ROUNDS_TOTAL/8)
540

541
#ifdef  SKEIN_LOOP                              /* configure how much to unroll the loop */
542
#define SKEIN_UNROLL_256 (((SKEIN_LOOP)/100)%10)
543
#else
544
#define SKEIN_UNROLL_256 (0)
545
#endif
546

547
#if SKEIN_UNROLL_256
548
#if (RCNT % SKEIN_UNROLL_256)
549
#error "Invalid SKEIN_UNROLL_256"               /* sanity check on unroll count */
550
#endif
551
    size_t  r;
552
    u64b_t  kw[WCNT+4+RCNT*2];                  /* key schedule words : chaining vars + tweak + "rotation"*/
553
#else
554
    u64b_t  kw[WCNT+4];                         /* key schedule words : chaining vars + tweak */
555
#endif
556
    u64b_t  X0,X1,X2,X3;                        /* local copy of context vars, for speed */
557
    u64b_t  w [WCNT];                           /* local copy of input block */
558
#ifdef SKEIN_DEBUG
559
    const u64b_t *Xptr[4];                      /* use for debugging (help compiler put Xn in registers) */
560
    Xptr[0] = &X0;  Xptr[1] = &X1;  Xptr[2] = &X2;  Xptr[3] = &X3;
561
#endif
562
    Skein_assert(blkCnt != 0);                  /* never call with blkCnt == 0! */
563
    ts[0] = ctx->h.T[0];
564
    ts[1] = ctx->h.T[1];
565
    do  {
566
        /* this implementation only supports 2**64 input bytes (no carry out here) */
567
        ts[0] += byteCntAdd;                    /* update processed length */
568

569
        /* precompute the key schedule for this block */
570
        ks[0] = ctx->X[0];     
571
        ks[1] = ctx->X[1];
572
        ks[2] = ctx->X[2];
573
        ks[3] = ctx->X[3];
574
        ks[4] = ks[0] ^ ks[1] ^ ks[2] ^ ks[3] ^ SKEIN_KS_PARITY;
575

576
        ts[2] = ts[0] ^ ts[1];
577

578
        Skein_Get64_LSB_First(w,blkPtr,WCNT);   /* get input block in little-endian format */
579
        DebugSaveTweak(ctx);
580
        Skein_Show_Block(BLK_BITS,&ctx->h,ctx->X,blkPtr,w,ks,ts);
581

582
        X0 = w[0] + ks[0];                      /* do the first full key injection */
583
        X1 = w[1] + ks[1] + ts[0];
584
        X2 = w[2] + ks[2] + ts[1];
585
        X3 = w[3] + ks[3];
586

587
        Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INITIAL,Xptr);    /* show starting state values */
588

589
        blkPtr += SKEIN_256_BLOCK_BYTES;
590

591
        /* run the rounds */
592

593
#define Round256(p0,p1,p2,p3,ROT,rNum)                              \
594
    X##p0 += X##p1; X##p1 = RotL_64(X##p1,ROT##_0); X##p1 ^= X##p0; \
595
    X##p2 += X##p3; X##p3 = RotL_64(X##p3,ROT##_1); X##p3 ^= X##p2; \
596

597
#if SKEIN_UNROLL_256 == 0                       
598
#define R256(p0,p1,p2,p3,ROT,rNum)           /* fully unrolled */   \
599
    Round256(p0,p1,p2,p3,ROT,rNum)                                  \
600
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,rNum,Xptr);
601

602
#define I256(R)                                                     \
603
    X0   += ks[((R)+1) % 5];    /* inject the key schedule value */ \
604
    X1   += ks[((R)+2) % 5] + ts[((R)+1) % 3];                      \
605
    X2   += ks[((R)+3) % 5] + ts[((R)+2) % 3];                      \
606
    X3   += ks[((R)+4) % 5] +     (R)+1;                            \
607
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INJECT,Xptr);
608
#else                                       /* looping version */
609
#define R256(p0,p1,p2,p3,ROT,rNum)                                  \
610
    Round256(p0,p1,p2,p3,ROT,rNum)                                  \
611
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,4*(r-1)+rNum,Xptr);
612

613
#define I256(R)                                                     \
614
    X0   += ks[r+(R)+0];        /* inject the key schedule value */ \
615
    X1   += ks[r+(R)+1] + ts[r+(R)+0];                              \
616
    X2   += ks[r+(R)+2] + ts[r+(R)+1];                              \
617
    X3   += ks[r+(R)+3] +    r+(R)   ;                              \
618
    ks[r + (R)+4    ]   = ks[r+(R)-1];     /* rotate key schedule */\
619
    ts[r + (R)+2    ]   = ts[r+(R)-1];                              \
620
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INJECT,Xptr);
621

622
    for (r=1;r < 2*RCNT;r+=2*SKEIN_UNROLL_256)  /* loop thru it */
623
#endif  
624
        {    
625
#define R256_8_rounds(R)                  \
626
        R256(0,1,2,3,R_256_0,8*(R) + 1);  \
627
        R256(0,3,2,1,R_256_1,8*(R) + 2);  \
628
        R256(0,1,2,3,R_256_2,8*(R) + 3);  \
629
        R256(0,3,2,1,R_256_3,8*(R) + 4);  \
630
        I256(2*(R));                      \
631
        R256(0,1,2,3,R_256_4,8*(R) + 5);  \
632
        R256(0,3,2,1,R_256_5,8*(R) + 6);  \
633
        R256(0,1,2,3,R_256_6,8*(R) + 7);  \
634
        R256(0,3,2,1,R_256_7,8*(R) + 8);  \
635
        I256(2*(R)+1);
636

637
        R256_8_rounds( 0);
638

639
#define R256_Unroll_R(NN) ((SKEIN_UNROLL_256 == 0 && SKEIN_256_ROUNDS_TOTAL/8 > (NN)) || (SKEIN_UNROLL_256 > (NN)))
640

641
  #if   R256_Unroll_R( 1)
642
        R256_8_rounds( 1);
643
  #endif
644
  #if   R256_Unroll_R( 2)
645
        R256_8_rounds( 2);
646
  #endif
647
  #if   R256_Unroll_R( 3)
648
        R256_8_rounds( 3);
649
  #endif
650
  #if   R256_Unroll_R( 4)
651
        R256_8_rounds( 4);
652
  #endif
653
  #if   R256_Unroll_R( 5)
654
        R256_8_rounds( 5);
655
  #endif
656
  #if   R256_Unroll_R( 6)
657
        R256_8_rounds( 6);
658
  #endif
659
  #if   R256_Unroll_R( 7)
660
        R256_8_rounds( 7);
661
  #endif
662
  #if   R256_Unroll_R( 8)
663
        R256_8_rounds( 8);
664
  #endif
665
  #if   R256_Unroll_R( 9)
666
        R256_8_rounds( 9);
667
  #endif
668
  #if   R256_Unroll_R(10)
669
        R256_8_rounds(10);
670
  #endif
671
  #if   R256_Unroll_R(11)
672
        R256_8_rounds(11);
673
  #endif
674
  #if   R256_Unroll_R(12)
675
        R256_8_rounds(12);
676
  #endif
677
  #if   R256_Unroll_R(13)
678
        R256_8_rounds(13);
679
  #endif
680
  #if   R256_Unroll_R(14)
681
        R256_8_rounds(14);
682
  #endif
683
  #if  (SKEIN_UNROLL_256 > 14)
684
#error  "need more unrolling in Skein_256_Process_Block"
685
  #endif
686
        }
687
        /* do the final "feedforward" xor, update context chaining vars */
688
        ctx->X[0] = X0 ^ w[0];
689
        ctx->X[1] = X1 ^ w[1];
690
        ctx->X[2] = X2 ^ w[2];
691
        ctx->X[3] = X3 ^ w[3];
692

693
        Skein_Show_Round(BLK_BITS,&ctx->h,SKEIN_RND_FEED_FWD,ctx->X);
694

695
        ts[1] &= ~SKEIN_T1_FLAG_FIRST;
696
        }
697
    while (--blkCnt);
698
    ctx->h.T[0] = ts[0];
699
    ctx->h.T[1] = ts[1];
700
    }
701

702
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
703
static size_t Skein_256_Process_Block_CodeSize(void)
704
    {
705
    return ((u08b_t *) Skein_256_Process_Block_CodeSize) -
706
           ((u08b_t *) Skein_256_Process_Block);
707
    }
708
static uint_t Skein_256_Unroll_Cnt(void)
709
    {
710
    return SKEIN_UNROLL_256;
711
    }
712
#endif
713
#endif
714

715
/*****************************  Skein_512 ******************************/
716
#if !(SKEIN_USE_ASM & 512)
717
static void Skein_512_Process_Block(Skein_512_Ctxt_t *ctx,const u08b_t *blkPtr,size_t blkCnt,size_t byteCntAdd)
718
    { /* do it in C */
719
    enum
720
        {
721
        WCNT = SKEIN_512_STATE_WORDS
722
        };
723
#undef  RCNT
724
#define RCNT  (SKEIN_512_ROUNDS_TOTAL/8)
725

726
#ifdef  SKEIN_LOOP                              /* configure how much to unroll the loop */
727
#define SKEIN_UNROLL_512 (((SKEIN_LOOP)/10)%10)
728
#else
729
#define SKEIN_UNROLL_512 (0)
730
#endif
731

732
#if SKEIN_UNROLL_512
733
#if (RCNT % SKEIN_UNROLL_512)
734
#error "Invalid SKEIN_UNROLL_512"               /* sanity check on unroll count */
735
#endif
736
    size_t  r;
737
    u64b_t  kw[WCNT+4+RCNT*2];                  /* key schedule words : chaining vars + tweak + "rotation"*/
738
#else
739
    u64b_t  kw[WCNT+4];                         /* key schedule words : chaining vars + tweak */
740
#endif
741
    u64b_t  X0,X1,X2,X3,X4,X5,X6,X7;            /* local copy of vars, for speed */
742
    u64b_t  w [WCNT];                           /* local copy of input block */
743
#ifdef SKEIN_DEBUG
744
    const u64b_t *Xptr[8];                      /* use for debugging (help compiler put Xn in registers) */
745
    Xptr[0] = &X0;  Xptr[1] = &X1;  Xptr[2] = &X2;  Xptr[3] = &X3;
746
    Xptr[4] = &X4;  Xptr[5] = &X5;  Xptr[6] = &X6;  Xptr[7] = &X7;
747
#endif
748

749
    Skein_assert(blkCnt != 0);                  /* never call with blkCnt == 0! */
750
    ts[0] = ctx->h.T[0];
751
    ts[1] = ctx->h.T[1];
752
    do  {
753
        /* this implementation only supports 2**64 input bytes (no carry out here) */
754
        ts[0] += byteCntAdd;                    /* update processed length */
755

756
        /* precompute the key schedule for this block */
757
        ks[0] = ctx->X[0];
758
        ks[1] = ctx->X[1];
759
        ks[2] = ctx->X[2];
760
        ks[3] = ctx->X[3];
761
        ks[4] = ctx->X[4];
762
        ks[5] = ctx->X[5];
763
        ks[6] = ctx->X[6];
764
        ks[7] = ctx->X[7];
765
        ks[8] = ks[0] ^ ks[1] ^ ks[2] ^ ks[3] ^ 
766
                ks[4] ^ ks[5] ^ ks[6] ^ ks[7] ^ SKEIN_KS_PARITY;
767

768
        ts[2] = ts[0] ^ ts[1];
769

770
        Skein_Get64_LSB_First(w,blkPtr,WCNT); /* get input block in little-endian format */
771
        DebugSaveTweak(ctx);
772
        Skein_Show_Block(BLK_BITS,&ctx->h,ctx->X,blkPtr,w,ks,ts);
773

774
        X0   = w[0] + ks[0];                    /* do the first full key injection */
775
        X1   = w[1] + ks[1];
776
        X2   = w[2] + ks[2];
777
        X3   = w[3] + ks[3];
778
        X4   = w[4] + ks[4];
779
        X5   = w[5] + ks[5] + ts[0];
780
        X6   = w[6] + ks[6] + ts[1];
781
        X7   = w[7] + ks[7];
782

783
        blkPtr += SKEIN_512_BLOCK_BYTES;
784

785
        Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INITIAL,Xptr);
786
        /* run the rounds */
787
#define Round512(p0,p1,p2,p3,p4,p5,p6,p7,ROT,rNum)                  \
788
    X##p0 += X##p1; X##p1 = RotL_64(X##p1,ROT##_0); X##p1 ^= X##p0; \
789
    X##p2 += X##p3; X##p3 = RotL_64(X##p3,ROT##_1); X##p3 ^= X##p2; \
790
    X##p4 += X##p5; X##p5 = RotL_64(X##p5,ROT##_2); X##p5 ^= X##p4; \
791
    X##p6 += X##p7; X##p7 = RotL_64(X##p7,ROT##_3); X##p7 ^= X##p6; \
792

793
#if SKEIN_UNROLL_512 == 0                       
794
#define R512(p0,p1,p2,p3,p4,p5,p6,p7,ROT,rNum)      /* unrolled */  \
795
    Round512(p0,p1,p2,p3,p4,p5,p6,p7,ROT,rNum)                      \
796
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,rNum,Xptr);
797

798
#define I512(R)                                                     \
799
    X0   += ks[((R)+1) % 9];   /* inject the key schedule value */  \
800
    X1   += ks[((R)+2) % 9];                                        \
801
    X2   += ks[((R)+3) % 9];                                        \
802
    X3   += ks[((R)+4) % 9];                                        \
803
    X4   += ks[((R)+5) % 9];                                        \
804
    X5   += ks[((R)+6) % 9] + ts[((R)+1) % 3];                      \
805
    X6   += ks[((R)+7) % 9] + ts[((R)+2) % 3];                      \
806
    X7   += ks[((R)+8) % 9] +     (R)+1;                            \
807
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INJECT,Xptr);
808
#else                                       /* looping version */
809
#define R512(p0,p1,p2,p3,p4,p5,p6,p7,ROT,rNum)                      \
810
    Round512(p0,p1,p2,p3,p4,p5,p6,p7,ROT,rNum)                      \
811
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,4*(r-1)+rNum,Xptr);
812

813
#define I512(R)                                                     \
814
    X0   += ks[r+(R)+0];        /* inject the key schedule value */ \
815
    X1   += ks[r+(R)+1];                                            \
816
    X2   += ks[r+(R)+2];                                            \
817
    X3   += ks[r+(R)+3];                                            \
818
    X4   += ks[r+(R)+4];                                            \
819
    X5   += ks[r+(R)+5] + ts[r+(R)+0];                              \
820
    X6   += ks[r+(R)+6] + ts[r+(R)+1];                              \
821
    X7   += ks[r+(R)+7] +    r+(R)   ;                              \
822
    ks[r +       (R)+8] = ks[r+(R)-1];  /* rotate key schedule */   \
823
    ts[r +       (R)+2] = ts[r+(R)-1];                              \
824
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INJECT,Xptr);
825

826
    for (r=1;r < 2*RCNT;r+=2*SKEIN_UNROLL_512)   /* loop thru it */
827
#endif                         /* end of looped code definitions */
828
        {
829
#define R512_8_rounds(R)  /* do 8 full rounds */  \
830
        R512(0,1,2,3,4,5,6,7,R_512_0,8*(R)+ 1);   \
831
        R512(2,1,4,7,6,5,0,3,R_512_1,8*(R)+ 2);   \
832
        R512(4,1,6,3,0,5,2,7,R_512_2,8*(R)+ 3);   \
833
        R512(6,1,0,7,2,5,4,3,R_512_3,8*(R)+ 4);   \
834
        I512(2*(R));                              \
835
        R512(0,1,2,3,4,5,6,7,R_512_4,8*(R)+ 5);   \
836
        R512(2,1,4,7,6,5,0,3,R_512_5,8*(R)+ 6);   \
837
        R512(4,1,6,3,0,5,2,7,R_512_6,8*(R)+ 7);   \
838
        R512(6,1,0,7,2,5,4,3,R_512_7,8*(R)+ 8);   \
839
        I512(2*(R)+1);        /* and key injection */
840

841
        R512_8_rounds( 0);
842

843
#define R512_Unroll_R(NN) ((SKEIN_UNROLL_512 == 0 && SKEIN_512_ROUNDS_TOTAL/8 > (NN)) || (SKEIN_UNROLL_512 > (NN)))
844

845
  #if   R512_Unroll_R( 1)
846
        R512_8_rounds( 1);
847
  #endif
848
  #if   R512_Unroll_R( 2)
849
        R512_8_rounds( 2);
850
  #endif
851
  #if   R512_Unroll_R( 3)
852
        R512_8_rounds( 3);
853
  #endif
854
  #if   R512_Unroll_R( 4)
855
        R512_8_rounds( 4);
856
  #endif
857
  #if   R512_Unroll_R( 5)
858
        R512_8_rounds( 5);
859
  #endif
860
  #if   R512_Unroll_R( 6)
861
        R512_8_rounds( 6);
862
  #endif
863
  #if   R512_Unroll_R( 7)
864
        R512_8_rounds( 7);
865
  #endif
866
  #if   R512_Unroll_R( 8)
867
        R512_8_rounds( 8);
868
  #endif
869
  #if   R512_Unroll_R( 9)
870
        R512_8_rounds( 9);
871
  #endif
872
  #if   R512_Unroll_R(10)
873
        R512_8_rounds(10);
874
  #endif
875
  #if   R512_Unroll_R(11)
876
        R512_8_rounds(11);
877
  #endif
878
  #if   R512_Unroll_R(12)
879
        R512_8_rounds(12);
880
  #endif
881
  #if   R512_Unroll_R(13)
882
        R512_8_rounds(13);
883
  #endif
884
  #if   R512_Unroll_R(14)
885
        R512_8_rounds(14);
886
  #endif
887
  #if  (SKEIN_UNROLL_512 > 14)
888
#error  "need more unrolling in Skein_512_Process_Block"
889
  #endif
890
        }
891

892
        /* do the final "feedforward" xor, update context chaining vars */
893
        ctx->X[0] = X0 ^ w[0];
894
        ctx->X[1] = X1 ^ w[1];
895
        ctx->X[2] = X2 ^ w[2];
896
        ctx->X[3] = X3 ^ w[3];
897
        ctx->X[4] = X4 ^ w[4];
898
        ctx->X[5] = X5 ^ w[5];
899
        ctx->X[6] = X6 ^ w[6];
900
        ctx->X[7] = X7 ^ w[7];
901
        Skein_Show_Round(BLK_BITS,&ctx->h,SKEIN_RND_FEED_FWD,ctx->X);
902

903
        ts[1] &= ~SKEIN_T1_FLAG_FIRST;
904
        }
905
    while (--blkCnt);
906
    ctx->h.T[0] = ts[0];
907
    ctx->h.T[1] = ts[1];
908
    }
909

910
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
911
static size_t Skein_512_Process_Block_CodeSize(void)
912
    {
913
    return ((u08b_t *) Skein_512_Process_Block_CodeSize) -
914
           ((u08b_t *) Skein_512_Process_Block);
915
    }
916
static uint_t Skein_512_Unroll_Cnt(void)
917
    {
918
    return SKEIN_UNROLL_512;
919
    }
920
#endif
921
#endif
922

923
/*****************************  Skein1024 ******************************/
924
#if !(SKEIN_USE_ASM & 1024)
925
static void Skein1024_Process_Block(Skein1024_Ctxt_t *ctx,const u08b_t *blkPtr,size_t blkCnt,size_t byteCntAdd)
926
    { /* do it in C, always looping (unrolled is bigger AND slower!) */
927
    enum
928
        {
929
        WCNT = SKEIN1024_STATE_WORDS
930
        };
931
#undef  RCNT
932
#define RCNT  (SKEIN1024_ROUNDS_TOTAL/8)
933

934
#ifdef  SKEIN_LOOP                              /* configure how much to unroll the loop */
935
#define SKEIN_UNROLL_1024 ((SKEIN_LOOP)%10)
936
#else
937
#define SKEIN_UNROLL_1024 (0)
938
#endif
939

940
#if (SKEIN_UNROLL_1024 != 0)
941
#if (RCNT % SKEIN_UNROLL_1024)
942
#error "Invalid SKEIN_UNROLL_1024"              /* sanity check on unroll count */
943
#endif
944
    size_t  r;
945
    u64b_t  kw[WCNT+4+RCNT*2];                  /* key schedule words : chaining vars + tweak + "rotation"*/
946
#else
947
    u64b_t  kw[WCNT+4];                         /* key schedule words : chaining vars + tweak */
948
#endif
949

950
    u64b_t  X00,X01,X02,X03,X04,X05,X06,X07,    /* local copy of vars, for speed */
951
            X08,X09,X10,X11,X12,X13,X14,X15;
952
    u64b_t  w [WCNT];                           /* local copy of input block */
953
#ifdef SKEIN_DEBUG
954
    const u64b_t *Xptr[16];                     /* use for debugging (help compiler put Xn in registers) */
955
    Xptr[ 0] = &X00;  Xptr[ 1] = &X01;  Xptr[ 2] = &X02;  Xptr[ 3] = &X03;
956
    Xptr[ 4] = &X04;  Xptr[ 5] = &X05;  Xptr[ 6] = &X06;  Xptr[ 7] = &X07;
957
    Xptr[ 8] = &X08;  Xptr[ 9] = &X09;  Xptr[10] = &X10;  Xptr[11] = &X11;
958
    Xptr[12] = &X12;  Xptr[13] = &X13;  Xptr[14] = &X14;  Xptr[15] = &X15;
959
#endif
960

961
    Skein_assert(blkCnt != 0);                  /* never call with blkCnt == 0! */
962
    ts[0] = ctx->h.T[0];
963
    ts[1] = ctx->h.T[1];
964
    do  {
965
        /* this implementation only supports 2**64 input bytes (no carry out here) */
966
        ts[0] += byteCntAdd;                    /* update processed length */
967

968
        /* precompute the key schedule for this block */
969
        ks[ 0] = ctx->X[ 0];
970
        ks[ 1] = ctx->X[ 1];
971
        ks[ 2] = ctx->X[ 2];
972
        ks[ 3] = ctx->X[ 3];
973
        ks[ 4] = ctx->X[ 4];
974
        ks[ 5] = ctx->X[ 5];
975
        ks[ 6] = ctx->X[ 6];
976
        ks[ 7] = ctx->X[ 7];
977
        ks[ 8] = ctx->X[ 8];
978
        ks[ 9] = ctx->X[ 9];
979
        ks[10] = ctx->X[10];
980
        ks[11] = ctx->X[11];
981
        ks[12] = ctx->X[12];
982
        ks[13] = ctx->X[13];
983
        ks[14] = ctx->X[14];
984
        ks[15] = ctx->X[15];
985
        ks[16] = ks[ 0] ^ ks[ 1] ^ ks[ 2] ^ ks[ 3] ^
986
                 ks[ 4] ^ ks[ 5] ^ ks[ 6] ^ ks[ 7] ^
987
                 ks[ 8] ^ ks[ 9] ^ ks[10] ^ ks[11] ^
988
                 ks[12] ^ ks[13] ^ ks[14] ^ ks[15] ^ SKEIN_KS_PARITY;
989

990
        ts[2]  = ts[0] ^ ts[1];
991

992
        Skein_Get64_LSB_First(w,blkPtr,WCNT); /* get input block in little-endian format */
993
        DebugSaveTweak(ctx);
994
        Skein_Show_Block(BLK_BITS,&ctx->h,ctx->X,blkPtr,w,ks,ts);
995

996
        X00    = w[ 0] + ks[ 0];                 /* do the first full key injection */
997
        X01    = w[ 1] + ks[ 1];
998
        X02    = w[ 2] + ks[ 2];
999
        X03    = w[ 3] + ks[ 3];
1000
        X04    = w[ 4] + ks[ 4];
1001
        X05    = w[ 5] + ks[ 5];
1002
        X06    = w[ 6] + ks[ 6];
1003
        X07    = w[ 7] + ks[ 7];
1004
        X08    = w[ 8] + ks[ 8];
1005
        X09    = w[ 9] + ks[ 9];
1006
        X10    = w[10] + ks[10];
1007
        X11    = w[11] + ks[11];
1008
        X12    = w[12] + ks[12];
1009
        X13    = w[13] + ks[13] + ts[0];
1010
        X14    = w[14] + ks[14] + ts[1];
1011
        X15    = w[15] + ks[15];
1012

1013
        Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INITIAL,Xptr);
1014

1015
#define Round1024(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,pA,pB,pC,pD,pE,pF,ROT,rNum) \
1016
    X##p0 += X##p1; X##p1 = RotL_64(X##p1,ROT##_0); X##p1 ^= X##p0;   \
1017
    X##p2 += X##p3; X##p3 = RotL_64(X##p3,ROT##_1); X##p3 ^= X##p2;   \
1018
    X##p4 += X##p5; X##p5 = RotL_64(X##p5,ROT##_2); X##p5 ^= X##p4;   \
1019
    X##p6 += X##p7; X##p7 = RotL_64(X##p7,ROT##_3); X##p7 ^= X##p6;   \
1020
    X##p8 += X##p9; X##p9 = RotL_64(X##p9,ROT##_4); X##p9 ^= X##p8;   \
1021
    X##pA += X##pB; X##pB = RotL_64(X##pB,ROT##_5); X##pB ^= X##pA;   \
1022
    X##pC += X##pD; X##pD = RotL_64(X##pD,ROT##_6); X##pD ^= X##pC;   \
1023
    X##pE += X##pF; X##pF = RotL_64(X##pF,ROT##_7); X##pF ^= X##pE;   \
1024

1025
#if SKEIN_UNROLL_1024 == 0                      
1026
#define R1024(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,pA,pB,pC,pD,pE,pF,ROT,rn) \
1027
    Round1024(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,pA,pB,pC,pD,pE,pF,ROT,rn) \
1028
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,rn,Xptr);
1029

1030
#define I1024(R)                                                      \
1031
    X00   += ks[((R)+ 1) % 17]; /* inject the key schedule value */   \
1032
    X01   += ks[((R)+ 2) % 17];                                       \
1033
    X02   += ks[((R)+ 3) % 17];                                       \
1034
    X03   += ks[((R)+ 4) % 17];                                       \
1035
    X04   += ks[((R)+ 5) % 17];                                       \
1036
    X05   += ks[((R)+ 6) % 17];                                       \
1037
    X06   += ks[((R)+ 7) % 17];                                       \
1038
    X07   += ks[((R)+ 8) % 17];                                       \
1039
    X08   += ks[((R)+ 9) % 17];                                       \
1040
    X09   += ks[((R)+10) % 17];                                       \
1041
    X10   += ks[((R)+11) % 17];                                       \
1042
    X11   += ks[((R)+12) % 17];                                       \
1043
    X12   += ks[((R)+13) % 17];                                       \
1044
    X13   += ks[((R)+14) % 17] + ts[((R)+1) % 3];                     \
1045
    X14   += ks[((R)+15) % 17] + ts[((R)+2) % 3];                     \
1046
    X15   += ks[((R)+16) % 17] +     (R)+1;                           \
1047
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INJECT,Xptr); 
1048
#else                                       /* looping version */
1049
#define R1024(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,pA,pB,pC,pD,pE,pF,ROT,rn) \
1050
    Round1024(p0,p1,p2,p3,p4,p5,p6,p7,p8,p9,pA,pB,pC,pD,pE,pF,ROT,rn) \
1051
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,4*(r-1)+rn,Xptr);
1052

1053
#define I1024(R)                                                      \
1054
    X00   += ks[r+(R)+ 0];    /* inject the key schedule value */     \
1055
    X01   += ks[r+(R)+ 1];                                            \
1056
    X02   += ks[r+(R)+ 2];                                            \
1057
    X03   += ks[r+(R)+ 3];                                            \
1058
    X04   += ks[r+(R)+ 4];                                            \
1059
    X05   += ks[r+(R)+ 5];                                            \
1060
    X06   += ks[r+(R)+ 6];                                            \
1061
    X07   += ks[r+(R)+ 7];                                            \
1062
    X08   += ks[r+(R)+ 8];                                            \
1063
    X09   += ks[r+(R)+ 9];                                            \
1064
    X10   += ks[r+(R)+10];                                            \
1065
    X11   += ks[r+(R)+11];                                            \
1066
    X12   += ks[r+(R)+12];                                            \
1067
    X13   += ks[r+(R)+13] + ts[r+(R)+0];                              \
1068
    X14   += ks[r+(R)+14] + ts[r+(R)+1];                              \
1069
    X15   += ks[r+(R)+15] +    r+(R)   ;                              \
1070
    ks[r  +       (R)+16] = ks[r+(R)-1];  /* rotate key schedule */   \
1071
    ts[r  +       (R)+ 2] = ts[r+(R)-1];                              \
1072
    Skein_Show_R_Ptr(BLK_BITS,&ctx->h,SKEIN_RND_KEY_INJECT,Xptr);
1073

1074
    for (r=1;r <= 2*RCNT;r+=2*SKEIN_UNROLL_1024)    /* loop thru it */
1075
#endif  
1076
        {
1077
#define R1024_8_rounds(R)    /* do 8 full rounds */                               \
1078
        R1024(00,01,02,03,04,05,06,07,08,09,10,11,12,13,14,15,R1024_0,8*(R) + 1); \
1079
        R1024(00,09,02,13,06,11,04,15,10,07,12,03,14,05,08,01,R1024_1,8*(R) + 2); \
1080
        R1024(00,07,02,05,04,03,06,01,12,15,14,13,08,11,10,09,R1024_2,8*(R) + 3); \
1081
        R1024(00,15,02,11,06,13,04,09,14,01,08,05,10,03,12,07,R1024_3,8*(R) + 4); \
1082
        I1024(2*(R));                                                             \
1083
        R1024(00,01,02,03,04,05,06,07,08,09,10,11,12,13,14,15,R1024_4,8*(R) + 5); \
1084
        R1024(00,09,02,13,06,11,04,15,10,07,12,03,14,05,08,01,R1024_5,8*(R) + 6); \
1085
        R1024(00,07,02,05,04,03,06,01,12,15,14,13,08,11,10,09,R1024_6,8*(R) + 7); \
1086
        R1024(00,15,02,11,06,13,04,09,14,01,08,05,10,03,12,07,R1024_7,8*(R) + 8); \
1087
        I1024(2*(R)+1);
1088

1089
        R1024_8_rounds( 0);
1090

1091
#define R1024_Unroll_R(NN) ((SKEIN_UNROLL_1024 == 0 && SKEIN1024_ROUNDS_TOTAL/8 > (NN)) || (SKEIN_UNROLL_1024 > (NN)))
1092

1093
  #if   R1024_Unroll_R( 1)
1094
        R1024_8_rounds( 1);
1095
  #endif
1096
  #if   R1024_Unroll_R( 2)
1097
        R1024_8_rounds( 2);
1098
  #endif
1099
  #if   R1024_Unroll_R( 3)
1100
        R1024_8_rounds( 3);
1101
  #endif
1102
  #if   R1024_Unroll_R( 4)
1103
        R1024_8_rounds( 4);
1104
  #endif
1105
  #if   R1024_Unroll_R( 5)
1106
        R1024_8_rounds( 5);
1107
  #endif
1108
  #if   R1024_Unroll_R( 6)
1109
        R1024_8_rounds( 6);
1110
  #endif
1111
  #if   R1024_Unroll_R( 7)
1112
        R1024_8_rounds( 7);
1113
  #endif
1114
  #if   R1024_Unroll_R( 8)
1115
        R1024_8_rounds( 8);
1116
  #endif
1117
  #if   R1024_Unroll_R( 9)
1118
        R1024_8_rounds( 9);
1119
  #endif
1120
  #if   R1024_Unroll_R(10)
1121
        R1024_8_rounds(10);
1122
  #endif
1123
  #if   R1024_Unroll_R(11)
1124
        R1024_8_rounds(11);
1125
  #endif
1126
  #if   R1024_Unroll_R(12)
1127
        R1024_8_rounds(12);
1128
  #endif
1129
  #if   R1024_Unroll_R(13)
1130
        R1024_8_rounds(13);
1131
  #endif
1132
  #if   R1024_Unroll_R(14)
1133
        R1024_8_rounds(14);
1134
  #endif
1135
  #if  (SKEIN_UNROLL_1024 > 14)
1136
#error  "need more unrolling in Skein_1024_Process_Block"
1137
  #endif
1138
        }
1139
        /* do the final "feedforward" xor, update context chaining vars */
1140

1141
        ctx->X[ 0] = X00 ^ w[ 0];
1142
        ctx->X[ 1] = X01 ^ w[ 1];
1143
        ctx->X[ 2] = X02 ^ w[ 2];
1144
        ctx->X[ 3] = X03 ^ w[ 3];
1145
        ctx->X[ 4] = X04 ^ w[ 4];
1146
        ctx->X[ 5] = X05 ^ w[ 5];
1147
        ctx->X[ 6] = X06 ^ w[ 6];
1148
        ctx->X[ 7] = X07 ^ w[ 7];
1149
        ctx->X[ 8] = X08 ^ w[ 8];
1150
        ctx->X[ 9] = X09 ^ w[ 9];
1151
        ctx->X[10] = X10 ^ w[10];
1152
        ctx->X[11] = X11 ^ w[11];
1153
        ctx->X[12] = X12 ^ w[12];
1154
        ctx->X[13] = X13 ^ w[13];
1155
        ctx->X[14] = X14 ^ w[14];
1156
        ctx->X[15] = X15 ^ w[15];
1157

1158
        Skein_Show_Round(BLK_BITS,&ctx->h,SKEIN_RND_FEED_FWD,ctx->X);
1159
        
1160
        ts[1] &= ~SKEIN_T1_FLAG_FIRST;
1161
        blkPtr += SKEIN1024_BLOCK_BYTES;
1162
        }
1163
    while (--blkCnt);
1164
    ctx->h.T[0] = ts[0];
1165
    ctx->h.T[1] = ts[1];
1166
    }
1167

1168
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
1169
static size_t Skein1024_Process_Block_CodeSize(void)
1170
    {
1171
    return ((u08b_t *) Skein1024_Process_Block_CodeSize) -
1172
           ((u08b_t *) Skein1024_Process_Block);
1173
    }
1174
static uint_t Skein1024_Unroll_Cnt(void)
1175
    {
1176
    return SKEIN_UNROLL_1024;
1177
    }
1178
#endif
1179
#endif
1180

1181

1182
#if 0
1183
/*****************************************************************/
1184
/*     256-bit Skein                                             */
1185
/*****************************************************************/
1186

1187
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1188
/* init the context for a straight hashing operation  */
1189
static int Skein_256_Init(Skein_256_Ctxt_t *ctx, size_t hashBitLen)
1190
    {
1191
    union
1192
        {
1193
        u08b_t  b[SKEIN_256_STATE_BYTES];
1194
        u64b_t  w[SKEIN_256_STATE_WORDS];
1195
        } cfg;                              /* config block */
1196
        
1197
    Skein_Assert(hashBitLen > 0,SKEIN_BAD_HASHLEN);
1198
    ctx->h.hashBitLen = hashBitLen;         /* output hash bit count */
1199

1200
    switch (hashBitLen)
1201
        {             /* use pre-computed values, where available */
1202
#ifndef SKEIN_NO_PRECOMP
1203
        case  256: memcpy(ctx->X,SKEIN_256_IV_256,sizeof(ctx->X));  break;
1204
        case  224: memcpy(ctx->X,SKEIN_256_IV_224,sizeof(ctx->X));  break;
1205
        case  160: memcpy(ctx->X,SKEIN_256_IV_160,sizeof(ctx->X));  break;
1206
        case  128: memcpy(ctx->X,SKEIN_256_IV_128,sizeof(ctx->X));  break;
1207
#endif
1208
        default:
1209
            /* here if there is no precomputed IV value available */
1210
            /* build/process the config block, type == CONFIG (could be precomputed) */
1211
            Skein_Start_New_Type(ctx,CFG_FINAL);        /* set tweaks: T0=0; T1=CFG | FINAL */
1212

1213
            cfg.w[0] = Skein_Swap64(SKEIN_SCHEMA_VER);  /* set the schema, version */
1214
            cfg.w[1] = Skein_Swap64(hashBitLen);        /* hash result length in bits */
1215
            cfg.w[2] = Skein_Swap64(SKEIN_CFG_TREE_INFO_SEQUENTIAL);
1216
            memset(&cfg.w[3],0,sizeof(cfg) - 3*sizeof(cfg.w[0])); /* zero pad config block */
1217

1218
            /* compute the initial chaining values from config block */
1219
            memset(ctx->X,0,sizeof(ctx->X));            /* zero the chaining variables */
1220
            Skein_256_Process_Block(ctx,cfg.b,1,SKEIN_CFG_STR_LEN);
1221
            break;
1222
        }
1223
    /* The chaining vars ctx->X are now initialized for the given hashBitLen. */
1224
    /* Set up to process the data message portion of the hash (default) */
1225
    Skein_Start_New_Type(ctx,MSG);              /* T0=0, T1= MSG type */
1226

1227
    return SKEIN_SUCCESS;
1228
    }
1229

1230
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1231
/* init the context for a MAC and/or tree hash operation */
1232
/* [identical to Skein_256_Init() when keyBytes == 0 && treeInfo == SKEIN_CFG_TREE_INFO_SEQUENTIAL] */
1233
static int Skein_256_InitExt(Skein_256_Ctxt_t *ctx,size_t hashBitLen,u64b_t treeInfo, const u08b_t *key, size_t keyBytes)
1234
    {
1235
    union
1236
        {
1237
        u08b_t  b[SKEIN_256_STATE_BYTES];
1238
        u64b_t  w[SKEIN_256_STATE_WORDS];
1239
        } cfg;                              /* config block */
1240
        
1241
    Skein_Assert(hashBitLen > 0,SKEIN_BAD_HASHLEN);
1242
    Skein_Assert(keyBytes == 0 || key != NULL,SKEIN_FAIL);
1243

1244
    /* compute the initial chaining values ctx->X[], based on key */
1245
    if (keyBytes == 0)                          /* is there a key? */
1246
        {                                   
1247
        memset(ctx->X,0,sizeof(ctx->X));        /* no key: use all zeroes as key for config block */
1248
        }
1249
    else                                        /* here to pre-process a key */
1250
        {
1251
        Skein_assert(sizeof(cfg.b) >= sizeof(ctx->X));
1252
        /* do a mini-Init right here */
1253
        ctx->h.hashBitLen=8*sizeof(ctx->X);     /* set output hash bit count = state size */
1254
        Skein_Start_New_Type(ctx,KEY);          /* set tweaks: T0 = 0; T1 = KEY type */
1255
        memset(ctx->X,0,sizeof(ctx->X));        /* zero the initial chaining variables */
1256
        Skein_256_Update(ctx,key,keyBytes);     /* hash the key */
1257
        Skein_256_Final_Pad(ctx,cfg.b);         /* put result into cfg.b[] */
1258
        memcpy(ctx->X,cfg.b,sizeof(cfg.b));     /* copy over into ctx->X[] */
1259
#if SKEIN_NEED_SWAP
1260
        {
1261
        uint_t i;
1262
        for (i=0;i<SKEIN_256_STATE_WORDS;i++)   /* convert key bytes to context words */
1263
            ctx->X[i] = Skein_Swap64(ctx->X[i]);
1264
        }
1265
#endif
1266
        }
1267
    /* build/process the config block, type == CONFIG (could be precomputed for each key) */
1268
    ctx->h.hashBitLen = hashBitLen;             /* output hash bit count */
1269
    Skein_Start_New_Type(ctx,CFG_FINAL);
1270

1271
    memset(&cfg.w,0,sizeof(cfg.w));             /* pre-pad cfg.w[] with zeroes */
1272
    cfg.w[0] = Skein_Swap64(SKEIN_SCHEMA_VER);
1273
    cfg.w[1] = Skein_Swap64(hashBitLen);        /* hash result length in bits */
1274
    cfg.w[2] = Skein_Swap64(treeInfo);          /* tree hash config info (or SKEIN_CFG_TREE_INFO_SEQUENTIAL) */
1275

1276
    Skein_Show_Key(256,&ctx->h,key,keyBytes);
1277

1278
    /* compute the initial chaining values from config block */
1279
    Skein_256_Process_Block(ctx,cfg.b,1,SKEIN_CFG_STR_LEN);
1280

1281
    /* The chaining vars ctx->X are now initialized */
1282
    /* Set up to process the data message portion of the hash (default) */
1283
    ctx->h.bCnt = 0;                            /* buffer b[] starts out empty */
1284
    Skein_Start_New_Type(ctx,MSG);
1285
    
1286
    return SKEIN_SUCCESS;
1287
    }
1288
#endif
1289

1290
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1291
/* process the input bytes */
1292
static int Skein_256_Update(Skein_256_Ctxt_t *ctx, const u08b_t *msg, size_t msgByteCnt)
1293
    {
1294
    size_t n;
1295

1296
    Skein_Assert(ctx->h.bCnt <= SKEIN_256_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1297

1298
    /* process full blocks, if any */
1299
    if (msgByteCnt + ctx->h.bCnt > SKEIN_256_BLOCK_BYTES)
1300
        {
1301
        if (ctx->h.bCnt)                              /* finish up any buffered message data */
1302
            {
1303
            n = SKEIN_256_BLOCK_BYTES - ctx->h.bCnt;  /* # bytes free in buffer b[] */
1304
            if (n)
1305
                {
1306
                Skein_assert(n < msgByteCnt);         /* check on our logic here */
1307
                memcpy(&ctx->b[ctx->h.bCnt],msg,n);
1308
                msgByteCnt  -= n;
1309
                msg         += n;
1310
                ctx->h.bCnt += n;
1311
                }
1312
            Skein_assert(ctx->h.bCnt == SKEIN_256_BLOCK_BYTES);
1313
            Skein_256_Process_Block(ctx,ctx->b,1,SKEIN_256_BLOCK_BYTES);
1314
            ctx->h.bCnt = 0;
1315
            }
1316
        /* now process any remaining full blocks, directly from input message data */
1317
        if (msgByteCnt > SKEIN_256_BLOCK_BYTES)
1318
            {
1319
            n = (msgByteCnt-1) / SKEIN_256_BLOCK_BYTES;   /* number of full blocks to process */
1320
            Skein_256_Process_Block(ctx,msg,n,SKEIN_256_BLOCK_BYTES);
1321
            msgByteCnt -= n * SKEIN_256_BLOCK_BYTES;
1322
            msg        += n * SKEIN_256_BLOCK_BYTES;
1323
            }
1324
        Skein_assert(ctx->h.bCnt == 0);
1325
        }
1326

1327
    /* copy any remaining source message data bytes into b[] */
1328
    if (msgByteCnt)
1329
        {
1330
        Skein_assert(msgByteCnt + ctx->h.bCnt <= SKEIN_256_BLOCK_BYTES);
1331
        memcpy(&ctx->b[ctx->h.bCnt],msg,msgByteCnt);
1332
        ctx->h.bCnt += msgByteCnt;
1333
        }
1334

1335
    return SKEIN_SUCCESS;
1336
    }
1337
   
1338
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1339
/* finalize the hash computation and output the result */
1340
static int Skein_256_Final(Skein_256_Ctxt_t *ctx, u08b_t *hashVal)
1341
    {
1342
    size_t i,n,byteCnt;
1343
    u64b_t X[SKEIN_256_STATE_WORDS];
1344
    Skein_Assert(ctx->h.bCnt <= SKEIN_256_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1345

1346
    ctx->h.T[1] |= SKEIN_T1_FLAG_FINAL;                 /* tag as the final block */
1347
    if (ctx->h.bCnt < SKEIN_256_BLOCK_BYTES)            /* zero pad b[] if necessary */
1348
        memset(&ctx->b[ctx->h.bCnt],0,SKEIN_256_BLOCK_BYTES - ctx->h.bCnt);
1349

1350
    Skein_256_Process_Block(ctx,ctx->b,1,ctx->h.bCnt);  /* process the final block */
1351
    
1352
    /* now output the result */
1353
    byteCnt = (ctx->h.hashBitLen + 7) >> 3;             /* total number of output bytes */
1354

1355
    /* run Threefish in "counter mode" to generate output */
1356
    memset(ctx->b,0,sizeof(ctx->b));  /* zero out b[], so it can hold the counter */
1357
    memcpy(X,ctx->X,sizeof(X));       /* keep a local copy of counter mode "key" */
1358
    for (i=0;i < byteCnt;i += SKEIN_256_BLOCK_BYTES)
1359
        {
1360
        ((u64b_t *)ctx->b)[0]= Skein_Swap64((u64b_t) i); /* build the counter block */
1361
        Skein_Start_New_Type(ctx,OUT_FINAL);
1362
        Skein_256_Process_Block(ctx,ctx->b,1,sizeof(u64b_t)); /* run "counter mode" */
1363
        n = byteCnt - i;   /* number of output bytes left to go */
1364
        if (n >= SKEIN_256_BLOCK_BYTES)
1365
            n  = SKEIN_256_BLOCK_BYTES;
1366
        Skein_Put64_LSB_First(hashVal+i,ctx->X,n);   /* "output" the ctr mode bytes */
1367
        Skein_Show_Final(256,&ctx->h,n,hashVal+i*SKEIN_256_BLOCK_BYTES);
1368
        memcpy(ctx->X,X,sizeof(X));   /* restore the counter mode key for next time */
1369
        }
1370
    return SKEIN_SUCCESS;
1371
    }
1372

1373
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
1374
static size_t Skein_256_API_CodeSize(void)
1375
    {
1376
    return ((u08b_t *) Skein_256_API_CodeSize) -
1377
           ((u08b_t *) Skein_256_Init);
1378
    }
1379
#endif
1380

1381
/*****************************************************************/
1382
/*     512-bit Skein                                             */
1383
/*****************************************************************/
1384

1385
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1386
/* init the context for a straight hashing operation  */
1387
static int Skein_512_Init(Skein_512_Ctxt_t *ctx, size_t hashBitLen)
1388
    {
1389
    union
1390
        {
1391
        u08b_t  b[SKEIN_512_STATE_BYTES];
1392
        u64b_t  w[SKEIN_512_STATE_WORDS];
1393
        } cfg;                              /* config block */
1394
        
1395
    Skein_Assert(hashBitLen > 0,SKEIN_BAD_HASHLEN);
1396
    ctx->h.hashBitLen = hashBitLen;         /* output hash bit count */
1397

1398
    switch (hashBitLen)
1399
        {             /* use pre-computed values, where available */
1400
#ifndef SKEIN_NO_PRECOMP
1401
        case  512: memcpy(ctx->X,SKEIN_512_IV_512,sizeof(ctx->X));  break;
1402
        case  384: memcpy(ctx->X,SKEIN_512_IV_384,sizeof(ctx->X));  break;
1403
        case  256: memcpy(ctx->X,SKEIN_512_IV_256,sizeof(ctx->X));  break;
1404
        case  224: memcpy(ctx->X,SKEIN_512_IV_224,sizeof(ctx->X));  break;
1405
#endif
1406
        default:
1407
            /* here if there is no precomputed IV value available */
1408
            /* build/process the config block, type == CONFIG (could be precomputed) */
1409
            Skein_Start_New_Type(ctx,CFG_FINAL);        /* set tweaks: T0=0; T1=CFG | FINAL */
1410

1411
            cfg.w[0] = Skein_Swap64(SKEIN_SCHEMA_VER);  /* set the schema, version */
1412
            cfg.w[1] = Skein_Swap64(hashBitLen);        /* hash result length in bits */
1413
            cfg.w[2] = Skein_Swap64(SKEIN_CFG_TREE_INFO_SEQUENTIAL);
1414
            memset(&cfg.w[3],0,sizeof(cfg) - 3*sizeof(cfg.w[0])); /* zero pad config block */
1415

1416
            /* compute the initial chaining values from config block */
1417
            memset(ctx->X,0,sizeof(ctx->X));            /* zero the chaining variables */
1418
            Skein_512_Process_Block(ctx,cfg.b,1,SKEIN_CFG_STR_LEN);
1419
            break;
1420
        }
1421

1422
    /* The chaining vars ctx->X are now initialized for the given hashBitLen. */
1423
    /* Set up to process the data message portion of the hash (default) */
1424
    Skein_Start_New_Type(ctx,MSG);              /* T0=0, T1= MSG type */
1425

1426
    return SKEIN_SUCCESS;
1427
    }
1428

1429
#if 0
1430
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1431
/* init the context for a MAC and/or tree hash operation */
1432
/* [identical to Skein_512_Init() when keyBytes == 0 && treeInfo == SKEIN_CFG_TREE_INFO_SEQUENTIAL] */
1433
static int Skein_512_InitExt(Skein_512_Ctxt_t *ctx,size_t hashBitLen,u64b_t treeInfo, const u08b_t *key, size_t keyBytes)
1434
    {
1435
    union
1436
        {
1437
        u08b_t  b[SKEIN_512_STATE_BYTES];
1438
        u64b_t  w[SKEIN_512_STATE_WORDS];
1439
        } cfg;                              /* config block */
1440
        
1441
    Skein_Assert(hashBitLen > 0,SKEIN_BAD_HASHLEN);
1442
    Skein_Assert(keyBytes == 0 || key != NULL,SKEIN_FAIL);
1443

1444
    /* compute the initial chaining values ctx->X[], based on key */
1445
    if (keyBytes == 0)                          /* is there a key? */
1446
        {                                   
1447
        memset(ctx->X,0,sizeof(ctx->X));        /* no key: use all zeroes as key for config block */
1448
        }
1449
    else                                        /* here to pre-process a key */
1450
        {
1451
        Skein_assert(sizeof(cfg.b) >= sizeof(ctx->X));
1452
        /* do a mini-Init right here */
1453
        ctx->h.hashBitLen=8*sizeof(ctx->X);     /* set output hash bit count = state size */
1454
        Skein_Start_New_Type(ctx,KEY);          /* set tweaks: T0 = 0; T1 = KEY type */
1455
        memset(ctx->X,0,sizeof(ctx->X));        /* zero the initial chaining variables */
1456
        Skein_512_Update(ctx,key,keyBytes);     /* hash the key */
1457
        Skein_512_Final_Pad(ctx,cfg.b);         /* put result into cfg.b[] */
1458
        memcpy(ctx->X,cfg.b,sizeof(cfg.b));     /* copy over into ctx->X[] */
1459
#if SKEIN_NEED_SWAP
1460
        {
1461
        uint_t i;
1462
        for (i=0;i<SKEIN_512_STATE_WORDS;i++)   /* convert key bytes to context words */
1463
            ctx->X[i] = Skein_Swap64(ctx->X[i]);
1464
        }
1465
#endif
1466
        }
1467
    /* build/process the config block, type == CONFIG (could be precomputed for each key) */
1468
    ctx->h.hashBitLen = hashBitLen;             /* output hash bit count */
1469
    Skein_Start_New_Type(ctx,CFG_FINAL);
1470

1471
    memset(&cfg.w,0,sizeof(cfg.w));             /* pre-pad cfg.w[] with zeroes */
1472
    cfg.w[0] = Skein_Swap64(SKEIN_SCHEMA_VER);
1473
    cfg.w[1] = Skein_Swap64(hashBitLen);        /* hash result length in bits */
1474
    cfg.w[2] = Skein_Swap64(treeInfo);          /* tree hash config info (or SKEIN_CFG_TREE_INFO_SEQUENTIAL) */
1475

1476
    Skein_Show_Key(512,&ctx->h,key,keyBytes);
1477

1478
    /* compute the initial chaining values from config block */
1479
    Skein_512_Process_Block(ctx,cfg.b,1,SKEIN_CFG_STR_LEN);
1480

1481
    /* The chaining vars ctx->X are now initialized */
1482
    /* Set up to process the data message portion of the hash (default) */
1483
    ctx->h.bCnt = 0;                            /* buffer b[] starts out empty */
1484
    Skein_Start_New_Type(ctx,MSG);
1485
    
1486
    return SKEIN_SUCCESS;
1487
    }
1488
#endif
1489

1490
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1491
/* process the input bytes */
1492
static int Skein_512_Update(Skein_512_Ctxt_t *ctx, const u08b_t *msg, size_t msgByteCnt)
1493
    {
1494
    size_t n;
1495

1496
    Skein_Assert(ctx->h.bCnt <= SKEIN_512_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1497

1498
    /* process full blocks, if any */
1499
    if (msgByteCnt + ctx->h.bCnt > SKEIN_512_BLOCK_BYTES)
1500
        {
1501
        if (ctx->h.bCnt)                              /* finish up any buffered message data */
1502
            {
1503
            n = SKEIN_512_BLOCK_BYTES - ctx->h.bCnt;  /* # bytes free in buffer b[] */
1504
            if (n)
1505
                {
1506
                Skein_assert(n < msgByteCnt);         /* check on our logic here */
1507
                memcpy(&ctx->b[ctx->h.bCnt],msg,n);
1508
                msgByteCnt  -= n;
1509
                msg         += n;
1510
                ctx->h.bCnt += n;
1511
                }
1512
            Skein_assert(ctx->h.bCnt == SKEIN_512_BLOCK_BYTES);
1513
            Skein_512_Process_Block(ctx,ctx->b,1,SKEIN_512_BLOCK_BYTES);
1514
            ctx->h.bCnt = 0;
1515
            }
1516
        /* now process any remaining full blocks, directly from input message data */
1517
        if (msgByteCnt > SKEIN_512_BLOCK_BYTES)
1518
            {
1519
            n = (msgByteCnt-1) / SKEIN_512_BLOCK_BYTES;   /* number of full blocks to process */
1520
            Skein_512_Process_Block(ctx,msg,n,SKEIN_512_BLOCK_BYTES);
1521
            msgByteCnt -= n * SKEIN_512_BLOCK_BYTES;
1522
            msg        += n * SKEIN_512_BLOCK_BYTES;
1523
            }
1524
        Skein_assert(ctx->h.bCnt == 0);
1525
        }
1526

1527
    /* copy any remaining source message data bytes into b[] */
1528
    if (msgByteCnt)
1529
        {
1530
        Skein_assert(msgByteCnt + ctx->h.bCnt <= SKEIN_512_BLOCK_BYTES);
1531
        memcpy(&ctx->b[ctx->h.bCnt],msg,msgByteCnt);
1532
        ctx->h.bCnt += msgByteCnt;
1533
        }
1534

1535
    return SKEIN_SUCCESS;
1536
    }
1537
   
1538
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1539
/* finalize the hash computation and output the result */
1540
static int Skein_512_Final(Skein_512_Ctxt_t *ctx, u08b_t *hashVal)
1541
    {
1542
    size_t i,n,byteCnt;
1543
    u64b_t X[SKEIN_512_STATE_WORDS];
1544
    Skein_Assert(ctx->h.bCnt <= SKEIN_512_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1545

1546
    ctx->h.T[1] |= SKEIN_T1_FLAG_FINAL;                 /* tag as the final block */
1547
    if (ctx->h.bCnt < SKEIN_512_BLOCK_BYTES)            /* zero pad b[] if necessary */
1548
        memset(&ctx->b[ctx->h.bCnt],0,SKEIN_512_BLOCK_BYTES - ctx->h.bCnt);
1549

1550
    Skein_512_Process_Block(ctx,ctx->b,1,ctx->h.bCnt);  /* process the final block */
1551
    
1552
    /* now output the result */
1553
    byteCnt = (ctx->h.hashBitLen + 7) >> 3;             /* total number of output bytes */
1554

1555
    /* run Threefish in "counter mode" to generate output */
1556
    memset(ctx->b,0,sizeof(ctx->b));  /* zero out b[], so it can hold the counter */
1557
    memcpy(X,ctx->X,sizeof(X));       /* keep a local copy of counter mode "key" */
1558
    for (i=0;i*SKEIN_512_BLOCK_BYTES < byteCnt;i++)
1559
        {
1560
        ((u64b_t *)ctx->b)[0]= Skein_Swap64((u64b_t) i); /* build the counter block */
1561
        Skein_Start_New_Type(ctx,OUT_FINAL);
1562
        Skein_512_Process_Block(ctx,ctx->b,1,sizeof(u64b_t)); /* run "counter mode" */
1563
        n = byteCnt - i*SKEIN_512_BLOCK_BYTES;   /* number of output bytes left to go */
1564
        if (n >= SKEIN_512_BLOCK_BYTES)
1565
            n  = SKEIN_512_BLOCK_BYTES;
1566
        Skein_Put64_LSB_First(hashVal+i*SKEIN_512_BLOCK_BYTES,ctx->X,n);   /* "output" the ctr mode bytes */
1567
        Skein_Show_Final(512,&ctx->h,n,hashVal+i*SKEIN_512_BLOCK_BYTES);
1568
        memcpy(ctx->X,X,sizeof(X));   /* restore the counter mode key for next time */
1569
        }
1570
    return SKEIN_SUCCESS;
1571
    }
1572

1573
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
1574
static size_t Skein_512_API_CodeSize(void)
1575
    {
1576
    return ((u08b_t *) Skein_512_API_CodeSize) -
1577
           ((u08b_t *) Skein_512_Init);
1578
    }
1579
#endif
1580

1581
/*****************************************************************/
1582
/*    1024-bit Skein                                             */
1583
/*****************************************************************/
1584
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1585
/* init the context for a straight hashing operation  */
1586
static int Skein1024_Init(Skein1024_Ctxt_t *ctx, size_t hashBitLen)
1587
    {
1588
    union
1589
        {
1590
        u08b_t  b[SKEIN1024_STATE_BYTES];
1591
        u64b_t  w[SKEIN1024_STATE_WORDS];
1592
        } cfg;                              /* config block */
1593
        
1594
    Skein_Assert(hashBitLen > 0,SKEIN_BAD_HASHLEN);
1595
    ctx->h.hashBitLen = hashBitLen;         /* output hash bit count */
1596

1597
    switch (hashBitLen)
1598
        {              /* use pre-computed values, where available */
1599
#ifndef SKEIN_NO_PRECOMP
1600
        case  512: memcpy(ctx->X,SKEIN1024_IV_512 ,sizeof(ctx->X)); break;
1601
        case  384: memcpy(ctx->X,SKEIN1024_IV_384 ,sizeof(ctx->X)); break;
1602
        case 1024: memcpy(ctx->X,SKEIN1024_IV_1024,sizeof(ctx->X)); break;
1603
#endif
1604
        default:
1605
            /* here if there is no precomputed IV value available */
1606
            /* build/process the config block, type == CONFIG (could be precomputed) */
1607
            Skein_Start_New_Type(ctx,CFG_FINAL);        /* set tweaks: T0=0; T1=CFG | FINAL */
1608

1609
            cfg.w[0] = Skein_Swap64(SKEIN_SCHEMA_VER);  /* set the schema, version */
1610
            cfg.w[1] = Skein_Swap64(hashBitLen);        /* hash result length in bits */
1611
            cfg.w[2] = Skein_Swap64(SKEIN_CFG_TREE_INFO_SEQUENTIAL);
1612
            memset(&cfg.w[3],0,sizeof(cfg) - 3*sizeof(cfg.w[0])); /* zero pad config block */
1613

1614
            /* compute the initial chaining values from config block */
1615
            memset(ctx->X,0,sizeof(ctx->X));            /* zero the chaining variables */
1616
            Skein1024_Process_Block(ctx,cfg.b,1,SKEIN_CFG_STR_LEN);
1617
            break;
1618
        }
1619

1620
    /* The chaining vars ctx->X are now initialized for the given hashBitLen. */
1621
    /* Set up to process the data message portion of the hash (default) */
1622
    Skein_Start_New_Type(ctx,MSG);              /* T0=0, T1= MSG type */
1623

1624
    return SKEIN_SUCCESS;
1625
    }
1626

1627
#if 0
1628
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1629
/* init the context for a MAC and/or tree hash operation */
1630
/* [identical to Skein1024_Init() when keyBytes == 0 && treeInfo == SKEIN_CFG_TREE_INFO_SEQUENTIAL] */
1631
static int Skein1024_InitExt(Skein1024_Ctxt_t *ctx,size_t hashBitLen,u64b_t treeInfo, const u08b_t *key, size_t keyBytes)
1632
    {
1633
    union
1634
        {
1635
        u08b_t  b[SKEIN1024_STATE_BYTES];
1636
        u64b_t  w[SKEIN1024_STATE_WORDS];
1637
        } cfg;                              /* config block */
1638
        
1639
    Skein_Assert(hashBitLen > 0,SKEIN_BAD_HASHLEN);
1640
    Skein_Assert(keyBytes == 0 || key != NULL,SKEIN_FAIL);
1641

1642
    /* compute the initial chaining values ctx->X[], based on key */
1643
    if (keyBytes == 0)                          /* is there a key? */
1644
        {                                   
1645
        memset(ctx->X,0,sizeof(ctx->X));        /* no key: use all zeroes as key for config block */
1646
        }
1647
    else                                        /* here to pre-process a key */
1648
        {
1649
        Skein_assert(sizeof(cfg.b) >= sizeof(ctx->X));
1650
        /* do a mini-Init right here */
1651
        ctx->h.hashBitLen=8*sizeof(ctx->X);     /* set output hash bit count = state size */
1652
        Skein_Start_New_Type(ctx,KEY);          /* set tweaks: T0 = 0; T1 = KEY type */
1653
        memset(ctx->X,0,sizeof(ctx->X));        /* zero the initial chaining variables */
1654
        Skein1024_Update(ctx,key,keyBytes);     /* hash the key */
1655
        Skein1024_Final_Pad(ctx,cfg.b);         /* put result into cfg.b[] */
1656
        memcpy(ctx->X,cfg.b,sizeof(cfg.b));     /* copy over into ctx->X[] */
1657
#if SKEIN_NEED_SWAP
1658
        {
1659
        uint_t i;
1660
        for (i=0;i<SKEIN1024_STATE_WORDS;i++)   /* convert key bytes to context words */
1661
            ctx->X[i] = Skein_Swap64(ctx->X[i]);
1662
        }
1663
#endif
1664
        }
1665
    /* build/process the config block, type == CONFIG (could be precomputed for each key) */
1666
    ctx->h.hashBitLen = hashBitLen;             /* output hash bit count */
1667
    Skein_Start_New_Type(ctx,CFG_FINAL);
1668

1669
    memset(&cfg.w,0,sizeof(cfg.w));             /* pre-pad cfg.w[] with zeroes */
1670
    cfg.w[0] = Skein_Swap64(SKEIN_SCHEMA_VER);
1671
    cfg.w[1] = Skein_Swap64(hashBitLen);        /* hash result length in bits */
1672
    cfg.w[2] = Skein_Swap64(treeInfo);          /* tree hash config info (or SKEIN_CFG_TREE_INFO_SEQUENTIAL) */
1673

1674
    Skein_Show_Key(1024,&ctx->h,key,keyBytes);
1675

1676
    /* compute the initial chaining values from config block */
1677
    Skein1024_Process_Block(ctx,cfg.b,1,SKEIN_CFG_STR_LEN);
1678

1679
    /* The chaining vars ctx->X are now initialized */
1680
    /* Set up to process the data message portion of the hash (default) */
1681
    ctx->h.bCnt = 0;                            /* buffer b[] starts out empty */
1682
    Skein_Start_New_Type(ctx,MSG);
1683
    
1684
    return SKEIN_SUCCESS;
1685
    }
1686
#endif
1687

1688
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1689
/* process the input bytes */
1690
static int Skein1024_Update(Skein1024_Ctxt_t *ctx, const u08b_t *msg, size_t msgByteCnt)
1691
    {
1692
    size_t n;
1693

1694
    Skein_Assert(ctx->h.bCnt <= SKEIN1024_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1695

1696
    /* process full blocks, if any */
1697
    if (msgByteCnt + ctx->h.bCnt > SKEIN1024_BLOCK_BYTES)
1698
        {
1699
        if (ctx->h.bCnt)                              /* finish up any buffered message data */
1700
            {
1701
            n = SKEIN1024_BLOCK_BYTES - ctx->h.bCnt;  /* # bytes free in buffer b[] */
1702
            if (n)
1703
                {
1704
                Skein_assert(n < msgByteCnt);         /* check on our logic here */
1705
                memcpy(&ctx->b[ctx->h.bCnt],msg,n);
1706
                msgByteCnt  -= n;
1707
                msg         += n;
1708
                ctx->h.bCnt += n;
1709
                }
1710
            Skein_assert(ctx->h.bCnt == SKEIN1024_BLOCK_BYTES);
1711
            Skein1024_Process_Block(ctx,ctx->b,1,SKEIN1024_BLOCK_BYTES);
1712
            ctx->h.bCnt = 0;
1713
            }
1714
        /* now process any remaining full blocks, directly from input message data */
1715
        if (msgByteCnt > SKEIN1024_BLOCK_BYTES)
1716
            {
1717
            n = (msgByteCnt-1) / SKEIN1024_BLOCK_BYTES;   /* number of full blocks to process */
1718
            Skein1024_Process_Block(ctx,msg,n,SKEIN1024_BLOCK_BYTES);
1719
            msgByteCnt -= n * SKEIN1024_BLOCK_BYTES;
1720
            msg        += n * SKEIN1024_BLOCK_BYTES;
1721
            }
1722
        Skein_assert(ctx->h.bCnt == 0);
1723
        }
1724

1725
    /* copy any remaining source message data bytes into b[] */
1726
    if (msgByteCnt)
1727
        {
1728
        Skein_assert(msgByteCnt + ctx->h.bCnt <= SKEIN1024_BLOCK_BYTES);
1729
        memcpy(&ctx->b[ctx->h.bCnt],msg,msgByteCnt);
1730
        ctx->h.bCnt += msgByteCnt;
1731
        }
1732

1733
    return SKEIN_SUCCESS;
1734
    }
1735
   
1736
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1737
/* finalize the hash computation and output the result */
1738
static int Skein1024_Final(Skein1024_Ctxt_t *ctx, u08b_t *hashVal)
1739
    {
1740
    size_t i,n,byteCnt;
1741
    u64b_t X[SKEIN1024_STATE_WORDS];
1742
    Skein_Assert(ctx->h.bCnt <= SKEIN1024_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1743

1744
    ctx->h.T[1] |= SKEIN_T1_FLAG_FINAL;                 /* tag as the final block */
1745
    if (ctx->h.bCnt < SKEIN1024_BLOCK_BYTES)            /* zero pad b[] if necessary */
1746
        memset(&ctx->b[ctx->h.bCnt],0,SKEIN1024_BLOCK_BYTES - ctx->h.bCnt);
1747

1748
    Skein1024_Process_Block(ctx,ctx->b,1,ctx->h.bCnt);  /* process the final block */
1749
    
1750
    /* now output the result */
1751
    byteCnt = (ctx->h.hashBitLen + 7) >> 3;             /* total number of output bytes */
1752

1753
    /* run Threefish in "counter mode" to generate output */
1754
    memset(ctx->b,0,sizeof(ctx->b));  /* zero out b[], so it can hold the counter */
1755
    memcpy(X,ctx->X,sizeof(X));       /* keep a local copy of counter mode "key" */
1756
    for (i=0;i*SKEIN1024_BLOCK_BYTES < byteCnt;i++)
1757
        {
1758
        ((u64b_t *)ctx->b)[0]= Skein_Swap64((u64b_t) i); /* build the counter block */
1759
        Skein_Start_New_Type(ctx,OUT_FINAL);
1760
        Skein1024_Process_Block(ctx,ctx->b,1,sizeof(u64b_t)); /* run "counter mode" */
1761
        n = byteCnt - i*SKEIN1024_BLOCK_BYTES;   /* number of output bytes left to go */
1762
        if (n >= SKEIN1024_BLOCK_BYTES)
1763
            n  = SKEIN1024_BLOCK_BYTES;
1764
        Skein_Put64_LSB_First(hashVal+i*SKEIN1024_BLOCK_BYTES,ctx->X,n);   /* "output" the ctr mode bytes */
1765
        Skein_Show_Final(1024,&ctx->h,n,hashVal+i*SKEIN1024_BLOCK_BYTES);
1766
        memcpy(ctx->X,X,sizeof(X));   /* restore the counter mode key for next time */
1767
        }
1768
    return SKEIN_SUCCESS;
1769
    }
1770

1771
#if defined(SKEIN_CODE_SIZE) || defined(SKEIN_PERF)
1772
static size_t Skein1024_API_CodeSize(void)
1773
    {
1774
    return ((u08b_t *) Skein1024_API_CodeSize) -
1775
           ((u08b_t *) Skein1024_Init);
1776
    }
1777
#endif
1778

1779
/**************** Functions to support MAC/tree hashing ***************/
1780
/*   (this code is identical for Optimized and Reference versions)    */
1781

1782
#if 0
1783
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1784
/* finalize the hash computation and output the block, no OUTPUT stage */
1785
static int Skein_256_Final_Pad(Skein_256_Ctxt_t *ctx, u08b_t *hashVal)
1786
    {
1787
    Skein_Assert(ctx->h.bCnt <= SKEIN_256_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1788

1789
    ctx->h.T[1] |= SKEIN_T1_FLAG_FINAL;        /* tag as the final block */
1790
    if (ctx->h.bCnt < SKEIN_256_BLOCK_BYTES)   /* zero pad b[] if necessary */
1791
        memset(&ctx->b[ctx->h.bCnt],0,SKEIN_256_BLOCK_BYTES - ctx->h.bCnt);
1792
    Skein_256_Process_Block(ctx,ctx->b,1,ctx->h.bCnt);    /* process the final block */
1793
    
1794
    Skein_Put64_LSB_First(hashVal,ctx->X,SKEIN_256_BLOCK_BYTES);   /* "output" the state bytes */
1795
    
1796
    return SKEIN_SUCCESS;
1797
    }
1798

1799
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1800
/* finalize the hash computation and output the block, no OUTPUT stage */
1801
static int Skein_512_Final_Pad(Skein_512_Ctxt_t *ctx, u08b_t *hashVal)
1802
    {
1803
    Skein_Assert(ctx->h.bCnt <= SKEIN_512_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1804

1805
    ctx->h.T[1] |= SKEIN_T1_FLAG_FINAL;        /* tag as the final block */
1806
    if (ctx->h.bCnt < SKEIN_512_BLOCK_BYTES)   /* zero pad b[] if necessary */
1807
        memset(&ctx->b[ctx->h.bCnt],0,SKEIN_512_BLOCK_BYTES - ctx->h.bCnt);
1808
    Skein_512_Process_Block(ctx,ctx->b,1,ctx->h.bCnt);    /* process the final block */
1809
    
1810
    Skein_Put64_LSB_First(hashVal,ctx->X,SKEIN_512_BLOCK_BYTES);   /* "output" the state bytes */
1811
    
1812
    return SKEIN_SUCCESS;
1813
    }
1814

1815
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1816
/* finalize the hash computation and output the block, no OUTPUT stage */
1817
static int Skein1024_Final_Pad(Skein1024_Ctxt_t *ctx, u08b_t *hashVal)
1818
    {
1819
    Skein_Assert(ctx->h.bCnt <= SKEIN1024_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1820

1821
    ctx->h.T[1] |= SKEIN_T1_FLAG_FINAL;        /* tag as the final block */
1822
    if (ctx->h.bCnt < SKEIN1024_BLOCK_BYTES)   /* zero pad b[] if necessary */
1823
        memset(&ctx->b[ctx->h.bCnt],0,SKEIN1024_BLOCK_BYTES - ctx->h.bCnt);
1824
    Skein1024_Process_Block(ctx,ctx->b,1,ctx->h.bCnt);    /* process the final block */
1825
    
1826
    Skein_Put64_LSB_First(hashVal,ctx->X,SKEIN1024_BLOCK_BYTES);   /* "output" the state bytes */
1827
    
1828
    return SKEIN_SUCCESS;
1829
    }
1830

1831

1832
#if SKEIN_TREE_HASH
1833
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1834
/* just do the OUTPUT stage                                       */
1835
static int Skein_256_Output(Skein_256_Ctxt_t *ctx, u08b_t *hashVal)
1836
    {
1837
    size_t i,n,byteCnt;
1838
    u64b_t X[SKEIN_256_STATE_WORDS];
1839
    Skein_Assert(ctx->h.bCnt <= SKEIN_256_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1840

1841
    /* now output the result */
1842
    byteCnt = (ctx->h.hashBitLen + 7) >> 3;    /* total number of output bytes */
1843

1844
    /* run Threefish in "counter mode" to generate output */
1845
    memset(ctx->b,0,sizeof(ctx->b));  /* zero out b[], so it can hold the counter */
1846
    memcpy(X,ctx->X,sizeof(X));       /* keep a local copy of counter mode "key" */
1847
    for (i=0;i*SKEIN_256_BLOCK_BYTES < byteCnt;i++)
1848
        {
1849
        ((u64b_t *)ctx->b)[0]= Skein_Swap64((u64b_t) i); /* build the counter block */
1850
        Skein_Start_New_Type(ctx,OUT_FINAL);
1851
        Skein_256_Process_Block(ctx,ctx->b,1,sizeof(u64b_t)); /* run "counter mode" */
1852
        n = byteCnt - i*SKEIN_256_BLOCK_BYTES;   /* number of output bytes left to go */
1853
        if (n >= SKEIN_256_BLOCK_BYTES)
1854
            n  = SKEIN_256_BLOCK_BYTES;
1855
        Skein_Put64_LSB_First(hashVal+i*SKEIN_256_BLOCK_BYTES,ctx->X,n);   /* "output" the ctr mode bytes */
1856
        Skein_Show_Final(256,&ctx->h,n,hashVal+i*SKEIN_256_BLOCK_BYTES);
1857
        memcpy(ctx->X,X,sizeof(X));   /* restore the counter mode key for next time */
1858
        }
1859
    return SKEIN_SUCCESS;
1860
    }
1861

1862
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1863
/* just do the OUTPUT stage                                       */
1864
static int Skein_512_Output(Skein_512_Ctxt_t *ctx, u08b_t *hashVal)
1865
    {
1866
    size_t i,n,byteCnt;
1867
    u64b_t X[SKEIN_512_STATE_WORDS];
1868
    Skein_Assert(ctx->h.bCnt <= SKEIN_512_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1869

1870
    /* now output the result */
1871
    byteCnt = (ctx->h.hashBitLen + 7) >> 3;    /* total number of output bytes */
1872

1873
    /* run Threefish in "counter mode" to generate output */
1874
    memset(ctx->b,0,sizeof(ctx->b));  /* zero out b[], so it can hold the counter */
1875
    memcpy(X,ctx->X,sizeof(X));       /* keep a local copy of counter mode "key" */
1876
    for (i=0;i*SKEIN_512_BLOCK_BYTES < byteCnt;i++)
1877
        {
1878
        ((u64b_t *)ctx->b)[0]= Skein_Swap64((u64b_t) i); /* build the counter block */
1879
        Skein_Start_New_Type(ctx,OUT_FINAL);
1880
        Skein_512_Process_Block(ctx,ctx->b,1,sizeof(u64b_t)); /* run "counter mode" */
1881
        n = byteCnt - i*SKEIN_512_BLOCK_BYTES;   /* number of output bytes left to go */
1882
        if (n >= SKEIN_512_BLOCK_BYTES)
1883
            n  = SKEIN_512_BLOCK_BYTES;
1884
        Skein_Put64_LSB_First(hashVal+i*SKEIN_512_BLOCK_BYTES,ctx->X,n);   /* "output" the ctr mode bytes */
1885
        Skein_Show_Final(256,&ctx->h,n,hashVal+i*SKEIN_512_BLOCK_BYTES);
1886
        memcpy(ctx->X,X,sizeof(X));   /* restore the counter mode key for next time */
1887
        }
1888
    return SKEIN_SUCCESS;
1889
    }
1890

1891
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1892
/* just do the OUTPUT stage                                       */
1893
static int Skein1024_Output(Skein1024_Ctxt_t *ctx, u08b_t *hashVal)
1894
    {
1895
    size_t i,n,byteCnt;
1896
    u64b_t X[SKEIN1024_STATE_WORDS];
1897
    Skein_Assert(ctx->h.bCnt <= SKEIN1024_BLOCK_BYTES,SKEIN_FAIL);    /* catch uninitialized context */
1898

1899
    /* now output the result */
1900
    byteCnt = (ctx->h.hashBitLen + 7) >> 3;    /* total number of output bytes */
1901

1902
    /* run Threefish in "counter mode" to generate output */
1903
    memset(ctx->b,0,sizeof(ctx->b));  /* zero out b[], so it can hold the counter */
1904
    memcpy(X,ctx->X,sizeof(X));       /* keep a local copy of counter mode "key" */
1905
    for (i=0;i*SKEIN1024_BLOCK_BYTES < byteCnt;i++)
1906
        {
1907
        ((u64b_t *)ctx->b)[0]= Skein_Swap64((u64b_t) i); /* build the counter block */
1908
        Skein_Start_New_Type(ctx,OUT_FINAL);
1909
        Skein1024_Process_Block(ctx,ctx->b,1,sizeof(u64b_t)); /* run "counter mode" */
1910
        n = byteCnt - i*SKEIN1024_BLOCK_BYTES;   /* number of output bytes left to go */
1911
        if (n >= SKEIN1024_BLOCK_BYTES)
1912
            n  = SKEIN1024_BLOCK_BYTES;
1913
        Skein_Put64_LSB_First(hashVal+i*SKEIN1024_BLOCK_BYTES,ctx->X,n);   /* "output" the ctr mode bytes */
1914
        Skein_Show_Final(256,&ctx->h,n,hashVal+i*SKEIN1024_BLOCK_BYTES);
1915
        memcpy(ctx->X,X,sizeof(X));   /* restore the counter mode key for next time */
1916
        }
1917
    return SKEIN_SUCCESS;
1918
    }
1919
#endif
1920
#endif
1921

1922
typedef struct
1923
{
1924
  uint_t  statebits;                      /* 256, 512, or 1024 */
1925
  union
1926
  {
1927
    Skein_Ctxt_Hdr_t h;                 /* common header "overlay" */
1928
    Skein_256_Ctxt_t ctx_256;
1929
    Skein_512_Ctxt_t ctx_512;
1930
    Skein1024_Ctxt_t ctx1024;
1931
  } u;
1932
}
1933
hashState;
1934

1935
/* "incremental" hashing API */
1936
static SkeinHashReturn Init  (hashState *state, int hashbitlen);
1937
static SkeinHashReturn Update(hashState *state, const SkeinBitSequence *data, SkeinDataLength databitlen);
1938
static SkeinHashReturn Final (hashState *state,       SkeinBitSequence *hashval);
1939

1940
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1941
/* select the context size and init the context */
1942
static SkeinHashReturn Init(hashState *state, int hashbitlen)
1943
{
1944
#if SKEIN_256_NIST_MAX_HASH_BITS
1945
  if (hashbitlen <= SKEIN_256_NIST_MAX_HASHBITS)
1946
  {
1947
    Skein_Assert(hashbitlen > 0,BAD_HASHLEN);
1948
    state->statebits = 64*SKEIN_256_STATE_WORDS;
1949
    return Skein_256_Init(&state->u.ctx_256,(size_t) hashbitlen);
1950
  }
1951
#endif
1952
  if (hashbitlen <= SKEIN_512_NIST_MAX_HASHBITS)
1953
  {
1954
    state->statebits = 64*SKEIN_512_STATE_WORDS;
1955
    return Skein_512_Init(&state->u.ctx_512,(size_t) hashbitlen);
1956
  }
1957
  else
1958
  {
1959
    state->statebits = 64*SKEIN1024_STATE_WORDS;
1960
    return Skein1024_Init(&state->u.ctx1024,(size_t) hashbitlen);
1961
  }
1962
}
1963

1964
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
1965
/* process data to be hashed */
1966
static SkeinHashReturn Update(hashState *state, const SkeinBitSequence *data, SkeinDataLength databitlen)
1967
{
1968
  /* only the final Update() call is allowed do partial bytes, else assert an error */
1969
  Skein_Assert((state->u.h.T[1] & SKEIN_T1_FLAG_BIT_PAD) == 0 || databitlen == 0, SKEIN_FAIL);
1970

1971
  Skein_Assert(state->statebits % 256 == 0 && (state->statebits-256) < 1024,SKEIN_FAIL);
1972
  if ((databitlen & 7) == 0)  /* partial bytes? */
1973
  {
1974
    switch ((state->statebits >> 8) & 3)
1975
    {
1976
    case 2:  return Skein_512_Update(&state->u.ctx_512,data,databitlen >> 3);
1977
    case 1:  return Skein_256_Update(&state->u.ctx_256,data,databitlen >> 3);
1978
    case 0:  return Skein1024_Update(&state->u.ctx1024,data,databitlen >> 3);
1979
    default: return SKEIN_FAIL;
1980
    }
1981
  }
1982
  else
1983
  {   /* handle partial final byte */
1984
    size_t bCnt = (databitlen >> 3) + 1;                  /* number of bytes to handle (nonzero here!) */
1985
    u08b_t b,mask;
1986

1987
    mask = (u08b_t) (1u << (7 - (databitlen & 7)));       /* partial byte bit mask */
1988
    b    = (u08b_t) ((data[bCnt-1] & (0-mask)) | mask);   /* apply bit padding on final byte */
1989

1990
    switch ((state->statebits >> 8) & 3)
1991
    {
1992
    case 2:  Skein_512_Update(&state->u.ctx_512,data,bCnt-1); /* process all but the final byte    */
1993
      Skein_512_Update(&state->u.ctx_512,&b  ,  1   ); /* process the (masked) partial byte */
1994
      break;
1995
    case 1:  Skein_256_Update(&state->u.ctx_256,data,bCnt-1); /* process all but the final byte    */
1996
      Skein_256_Update(&state->u.ctx_256,&b  ,  1   ); /* process the (masked) partial byte */
1997
      break;
1998
    case 0:  Skein1024_Update(&state->u.ctx1024,data,bCnt-1); /* process all but the final byte    */
1999
      Skein1024_Update(&state->u.ctx1024,&b  ,  1   ); /* process the (masked) partial byte */
2000
      break;
2001
    default: return SKEIN_FAIL;
2002
    }
2003
    Skein_Set_Bit_Pad_Flag(state->u.h);                    /* set tweak flag for the final call */
2004

2005
    return SKEIN_SUCCESS;
2006
  }
2007
}
2008

2009
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
2010
/* finalize hash computation and output the result (hashbitlen bits) */
2011
static SkeinHashReturn Final(hashState *state, SkeinBitSequence *hashval)
2012
{
2013
  Skein_Assert(state->statebits % 256 == 0 && (state->statebits-256) < 1024,FAIL);
2014
  switch ((state->statebits >> 8) & 3)
2015
  {
2016
  case 2:  return Skein_512_Final(&state->u.ctx_512,hashval);
2017
  case 1:  return Skein_256_Final(&state->u.ctx_256,hashval);
2018
  case 0:  return Skein1024_Final(&state->u.ctx1024,hashval);
2019
  default: return SKEIN_FAIL;
2020
  }
2021
}
2022

2023
/*++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++*/
2024
/* all-in-one hash function */
2025
SkeinHashReturn skein_hash(int hashbitlen, const SkeinBitSequence *data, /* all-in-one call */
2026
                SkeinDataLength databitlen,SkeinBitSequence *hashval)
2027
{
2028
  hashState  state;
2029
  SkeinHashReturn r = Init(&state,hashbitlen);
2030
  if (r == SKEIN_SUCCESS)
2031
  { /* these calls do not fail when called properly */
2032
    r = Update(&state,data,databitlen);
2033
    Final(&state,hashval);
2034
  }
2035
  return r;
2036
}
2037

2038
Product

Resources

Company
