CoCalc -- sph

GitHub Repository: tpruvot/cpuminer-multi
Path: blob/linux/sha3/sph_jh.c
¹²⁰¹ views
1
/* $Id: jh.c 255 2011-06-07 19:50:20Z tp $ */
2
/*
3
 * JH implementation.
4
 *
5
 * ==========================(LICENSE BEGIN)============================
6
 *
7
 * Copyright (c) 2007-2010  Projet RNRT SAPHIR
8
 * 
9
 * Permission is hereby granted, free of charge, to any person obtaining
10
 * a copy of this software and associated documentation files (the
11
 * "Software"), to deal in the Software without restriction, including
12
 * without limitation the rights to use, copy, modify, merge, publish,
13
 * distribute, sublicense, and/or sell copies of the Software, and to
14
 * permit persons to whom the Software is furnished to do so, subject to
15
 * the following conditions:
16
 * 
17
 * The above copyright notice and this permission notice shall be
18
 * included in all copies or substantial portions of the Software.
19
 * 
20
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
21
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
22
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
23
 * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
24
 * CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
25
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
26
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27
 *
28
 * ===========================(LICENSE END)=============================
29
 *
30
 * @author   Thomas Pornin <[email protected]>
31
 */
32

33
#include <stddef.h>
34
#include <string.h>
35

36
#include "sph_jh.h"
37

38
#ifdef __cplusplus
39
extern "C"{
40
#endif
41

42

43
#if SPH_SMALL_FOOTPRINT && !defined SPH_SMALL_FOOTPRINT_JH
44
#define SPH_SMALL_FOOTPRINT_JH   1
45
#endif
46

47
#if !defined SPH_JH_64 && SPH_64_TRUE
48
#define SPH_JH_64   1
49
#endif
50

51
#if !SPH_64
52
#undef SPH_JH_64
53
#endif
54

55
#ifdef _MSC_VER
56
#pragma warning (disable: 4146)
57
#endif
58

59
/*
60
 * The internal bitslice representation may use either big-endian or
61
 * little-endian (true bitslice operations do not care about the bit
62
 * ordering, and the bit-swapping linear operations in JH happen to
63
 * be invariant through endianness-swapping). The constants must be
64
 * defined according to the chosen endianness; we use some
65
 * byte-swapping macros for that.
66
 */
67

68
#if SPH_LITTLE_ENDIAN
69

70
#define C32e(x)     ((SPH_C32(x) >> 24) \
71
                    | ((SPH_C32(x) >>  8) & SPH_C32(0x0000FF00)) \
72
                    | ((SPH_C32(x) <<  8) & SPH_C32(0x00FF0000)) \
73
                    | ((SPH_C32(x) << 24) & SPH_C32(0xFF000000)))
74
#define dec32e_aligned   sph_dec32le_aligned
75
#define enc32e           sph_enc32le
76

77
#if SPH_64
78
#define C64e(x)     ((SPH_C64(x) >> 56) \
79
                    | ((SPH_C64(x) >> 40) & SPH_C64(0x000000000000FF00)) \
80
                    | ((SPH_C64(x) >> 24) & SPH_C64(0x0000000000FF0000)) \
81
                    | ((SPH_C64(x) >>  8) & SPH_C64(0x00000000FF000000)) \
82
                    | ((SPH_C64(x) <<  8) & SPH_C64(0x000000FF00000000)) \
83
                    | ((SPH_C64(x) << 24) & SPH_C64(0x0000FF0000000000)) \
84
                    | ((SPH_C64(x) << 40) & SPH_C64(0x00FF000000000000)) \
85
                    | ((SPH_C64(x) << 56) & SPH_C64(0xFF00000000000000)))
86
#define dec64e_aligned   sph_dec64le_aligned
87
#define enc64e           sph_enc64le
88
#endif
89

90
#else
91

92
#define C32e(x)     SPH_C32(x)
93
#define dec32e_aligned   sph_dec32be_aligned
94
#define enc32e           sph_enc32be
95
#if SPH_64
96
#define C64e(x)     SPH_C64(x)
97
#define dec64e_aligned   sph_dec64be_aligned
98
#define enc64e           sph_enc64be
99
#endif
100

101
#endif
102

103
#define Sb(x0, x1, x2, x3, c)   do { \
104
		x3 = ~x3; \
105
		x0 ^= (c) & ~x2; \
106
		tmp = (c) ^ (x0 & x1); \
107
		x0 ^= x2 & x3; \
108
		x3 ^= ~x1 & x2; \
109
		x1 ^= x0 & x2; \
110
		x2 ^= x0 & ~x3; \
111
		x0 ^= x1 | x3; \
112
		x3 ^= x1 & x2; \
113
		x1 ^= tmp & x0; \
114
		x2 ^= tmp; \
115
	} while (0)
116

117
#define Lb(x0, x1, x2, x3, x4, x5, x6, x7)   do { \
118
		x4 ^= x1; \
119
		x5 ^= x2; \
120
		x6 ^= x3 ^ x0; \
121
		x7 ^= x0; \
122
		x0 ^= x5; \
123
		x1 ^= x6; \
124
		x2 ^= x7 ^ x4; \
125
		x3 ^= x4; \
126
	} while (0)
127

128
#if SPH_JH_64
129

130
static const sph_u64 C[] = {
131
	C64e(0x72d5dea2df15f867), C64e(0x7b84150ab7231557),
132
	C64e(0x81abd6904d5a87f6), C64e(0x4e9f4fc5c3d12b40),
133
	C64e(0xea983ae05c45fa9c), C64e(0x03c5d29966b2999a),
134
	C64e(0x660296b4f2bb538a), C64e(0xb556141a88dba231),
135
	C64e(0x03a35a5c9a190edb), C64e(0x403fb20a87c14410),
136
	C64e(0x1c051980849e951d), C64e(0x6f33ebad5ee7cddc),
137
	C64e(0x10ba139202bf6b41), C64e(0xdc786515f7bb27d0),
138
	C64e(0x0a2c813937aa7850), C64e(0x3f1abfd2410091d3),
139
	C64e(0x422d5a0df6cc7e90), C64e(0xdd629f9c92c097ce),
140
	C64e(0x185ca70bc72b44ac), C64e(0xd1df65d663c6fc23),
141
	C64e(0x976e6c039ee0b81a), C64e(0x2105457e446ceca8),
142
	C64e(0xeef103bb5d8e61fa), C64e(0xfd9697b294838197),
143
	C64e(0x4a8e8537db03302f), C64e(0x2a678d2dfb9f6a95),
144
	C64e(0x8afe7381f8b8696c), C64e(0x8ac77246c07f4214),
145
	C64e(0xc5f4158fbdc75ec4), C64e(0x75446fa78f11bb80),
146
	C64e(0x52de75b7aee488bc), C64e(0x82b8001e98a6a3f4),
147
	C64e(0x8ef48f33a9a36315), C64e(0xaa5f5624d5b7f989),
148
	C64e(0xb6f1ed207c5ae0fd), C64e(0x36cae95a06422c36),
149
	C64e(0xce2935434efe983d), C64e(0x533af974739a4ba7),
150
	C64e(0xd0f51f596f4e8186), C64e(0x0e9dad81afd85a9f),
151
	C64e(0xa7050667ee34626a), C64e(0x8b0b28be6eb91727),
152
	C64e(0x47740726c680103f), C64e(0xe0a07e6fc67e487b),
153
	C64e(0x0d550aa54af8a4c0), C64e(0x91e3e79f978ef19e),
154
	C64e(0x8676728150608dd4), C64e(0x7e9e5a41f3e5b062),
155
	C64e(0xfc9f1fec4054207a), C64e(0xe3e41a00cef4c984),
156
	C64e(0x4fd794f59dfa95d8), C64e(0x552e7e1124c354a5),
157
	C64e(0x5bdf7228bdfe6e28), C64e(0x78f57fe20fa5c4b2),
158
	C64e(0x05897cefee49d32e), C64e(0x447e9385eb28597f),
159
	C64e(0x705f6937b324314a), C64e(0x5e8628f11dd6e465),
160
	C64e(0xc71b770451b920e7), C64e(0x74fe43e823d4878a),
161
	C64e(0x7d29e8a3927694f2), C64e(0xddcb7a099b30d9c1),
162
	C64e(0x1d1b30fb5bdc1be0), C64e(0xda24494ff29c82bf),
163
	C64e(0xa4e7ba31b470bfff), C64e(0x0d324405def8bc48),
164
	C64e(0x3baefc3253bbd339), C64e(0x459fc3c1e0298ba0),
165
	C64e(0xe5c905fdf7ae090f), C64e(0x947034124290f134),
166
	C64e(0xa271b701e344ed95), C64e(0xe93b8e364f2f984a),
167
	C64e(0x88401d63a06cf615), C64e(0x47c1444b8752afff),
168
	C64e(0x7ebb4af1e20ac630), C64e(0x4670b6c5cc6e8ce6),
169
	C64e(0xa4d5a456bd4fca00), C64e(0xda9d844bc83e18ae),
170
	C64e(0x7357ce453064d1ad), C64e(0xe8a6ce68145c2567),
171
	C64e(0xa3da8cf2cb0ee116), C64e(0x33e906589a94999a),
172
	C64e(0x1f60b220c26f847b), C64e(0xd1ceac7fa0d18518),
173
	C64e(0x32595ba18ddd19d3), C64e(0x509a1cc0aaa5b446),
174
	C64e(0x9f3d6367e4046bba), C64e(0xf6ca19ab0b56ee7e),
175
	C64e(0x1fb179eaa9282174), C64e(0xe9bdf7353b3651ee),
176
	C64e(0x1d57ac5a7550d376), C64e(0x3a46c2fea37d7001),
177
	C64e(0xf735c1af98a4d842), C64e(0x78edec209e6b6779),
178
	C64e(0x41836315ea3adba8), C64e(0xfac33b4d32832c83),
179
	C64e(0xa7403b1f1c2747f3), C64e(0x5940f034b72d769a),
180
	C64e(0xe73e4e6cd2214ffd), C64e(0xb8fd8d39dc5759ef),
181
	C64e(0x8d9b0c492b49ebda), C64e(0x5ba2d74968f3700d),
182
	C64e(0x7d3baed07a8d5584), C64e(0xf5a5e9f0e4f88e65),
183
	C64e(0xa0b8a2f436103b53), C64e(0x0ca8079e753eec5a),
184
	C64e(0x9168949256e8884f), C64e(0x5bb05c55f8babc4c),
185
	C64e(0xe3bb3b99f387947b), C64e(0x75daf4d6726b1c5d),
186
	C64e(0x64aeac28dc34b36d), C64e(0x6c34a550b828db71),
187
	C64e(0xf861e2f2108d512a), C64e(0xe3db643359dd75fc),
188
	C64e(0x1cacbcf143ce3fa2), C64e(0x67bbd13c02e843b0),
189
	C64e(0x330a5bca8829a175), C64e(0x7f34194db416535c),
190
	C64e(0x923b94c30e794d1e), C64e(0x797475d7b6eeaf3f),
191
	C64e(0xeaa8d4f7be1a3921), C64e(0x5cf47e094c232751),
192
	C64e(0x26a32453ba323cd2), C64e(0x44a3174a6da6d5ad),
193
	C64e(0xb51d3ea6aff2c908), C64e(0x83593d98916b3c56),
194
	C64e(0x4cf87ca17286604d), C64e(0x46e23ecc086ec7f6),
195
	C64e(0x2f9833b3b1bc765e), C64e(0x2bd666a5efc4e62a),
196
	C64e(0x06f4b6e8bec1d436), C64e(0x74ee8215bcef2163),
197
	C64e(0xfdc14e0df453c969), C64e(0xa77d5ac406585826),
198
	C64e(0x7ec1141606e0fa16), C64e(0x7e90af3d28639d3f),
199
	C64e(0xd2c9f2e3009bd20c), C64e(0x5faace30b7d40c30),
200
	C64e(0x742a5116f2e03298), C64e(0x0deb30d8e3cef89a),
201
	C64e(0x4bc59e7bb5f17992), C64e(0xff51e66e048668d3),
202
	C64e(0x9b234d57e6966731), C64e(0xcce6a6f3170a7505),
203
	C64e(0xb17681d913326cce), C64e(0x3c175284f805a262),
204
	C64e(0xf42bcbb378471547), C64e(0xff46548223936a48),
205
	C64e(0x38df58074e5e6565), C64e(0xf2fc7c89fc86508e),
206
	C64e(0x31702e44d00bca86), C64e(0xf04009a23078474e),
207
	C64e(0x65a0ee39d1f73883), C64e(0xf75ee937e42c3abd),
208
	C64e(0x2197b2260113f86f), C64e(0xa344edd1ef9fdee7),
209
	C64e(0x8ba0df15762592d9), C64e(0x3c85f7f612dc42be),
210
	C64e(0xd8a7ec7cab27b07e), C64e(0x538d7ddaaa3ea8de),
211
	C64e(0xaa25ce93bd0269d8), C64e(0x5af643fd1a7308f9),
212
	C64e(0xc05fefda174a19a5), C64e(0x974d66334cfd216a),
213
	C64e(0x35b49831db411570), C64e(0xea1e0fbbedcd549b),
214
	C64e(0x9ad063a151974072), C64e(0xf6759dbf91476fe2)
215
};
216

217
#define Ceven_hi(r)   (C[((r) << 2) + 0])
218
#define Ceven_lo(r)   (C[((r) << 2) + 1])
219
#define Codd_hi(r)    (C[((r) << 2) + 2])
220
#define Codd_lo(r)    (C[((r) << 2) + 3])
221

222
#define S(x0, x1, x2, x3, cb, r)   do { \
223
		Sb(x0 ## h, x1 ## h, x2 ## h, x3 ## h, cb ## hi(r)); \
224
		Sb(x0 ## l, x1 ## l, x2 ## l, x3 ## l, cb ## lo(r)); \
225
	} while (0)
226

227
#define L(x0, x1, x2, x3, x4, x5, x6, x7)   do { \
228
		Lb(x0 ## h, x1 ## h, x2 ## h, x3 ## h, \
229
			x4 ## h, x5 ## h, x6 ## h, x7 ## h); \
230
		Lb(x0 ## l, x1 ## l, x2 ## l, x3 ## l, \
231
			x4 ## l, x5 ## l, x6 ## l, x7 ## l); \
232
	} while (0)
233

234
#define Wz(x, c, n)   do { \
235
		sph_u64 t = (x ## h & (c)) << (n); \
236
		x ## h = ((x ## h >> (n)) & (c)) | t; \
237
		t = (x ## l & (c)) << (n); \
238
		x ## l = ((x ## l >> (n)) & (c)) | t; \
239
	} while (0)
240

241
#define W0(x)   Wz(x, SPH_C64(0x5555555555555555),  1)
242
#define W1(x)   Wz(x, SPH_C64(0x3333333333333333),  2)
243
#define W2(x)   Wz(x, SPH_C64(0x0F0F0F0F0F0F0F0F),  4)
244
#define W3(x)   Wz(x, SPH_C64(0x00FF00FF00FF00FF),  8)
245
#define W4(x)   Wz(x, SPH_C64(0x0000FFFF0000FFFF), 16)
246
#define W5(x)   Wz(x, SPH_C64(0x00000000FFFFFFFF), 32)
247
#define W6(x)   do { \
248
		sph_u64 t = x ## h; \
249
		x ## h = x ## l; \
250
		x ## l = t; \
251
	} while (0)
252

253
#define DECL_STATE \
254
	sph_u64 h0h, h1h, h2h, h3h, h4h, h5h, h6h, h7h; \
255
	sph_u64 h0l, h1l, h2l, h3l, h4l, h5l, h6l, h7l; \
256
	sph_u64 tmp;
257

258
#define READ_STATE(state)   do { \
259
		h0h = (state)->H.wide[ 0]; \
260
		h0l = (state)->H.wide[ 1]; \
261
		h1h = (state)->H.wide[ 2]; \
262
		h1l = (state)->H.wide[ 3]; \
263
		h2h = (state)->H.wide[ 4]; \
264
		h2l = (state)->H.wide[ 5]; \
265
		h3h = (state)->H.wide[ 6]; \
266
		h3l = (state)->H.wide[ 7]; \
267
		h4h = (state)->H.wide[ 8]; \
268
		h4l = (state)->H.wide[ 9]; \
269
		h5h = (state)->H.wide[10]; \
270
		h5l = (state)->H.wide[11]; \
271
		h6h = (state)->H.wide[12]; \
272
		h6l = (state)->H.wide[13]; \
273
		h7h = (state)->H.wide[14]; \
274
		h7l = (state)->H.wide[15]; \
275
	} while (0)
276

277
#define WRITE_STATE(state)   do { \
278
		(state)->H.wide[ 0] = h0h; \
279
		(state)->H.wide[ 1] = h0l; \
280
		(state)->H.wide[ 2] = h1h; \
281
		(state)->H.wide[ 3] = h1l; \
282
		(state)->H.wide[ 4] = h2h; \
283
		(state)->H.wide[ 5] = h2l; \
284
		(state)->H.wide[ 6] = h3h; \
285
		(state)->H.wide[ 7] = h3l; \
286
		(state)->H.wide[ 8] = h4h; \
287
		(state)->H.wide[ 9] = h4l; \
288
		(state)->H.wide[10] = h5h; \
289
		(state)->H.wide[11] = h5l; \
290
		(state)->H.wide[12] = h6h; \
291
		(state)->H.wide[13] = h6l; \
292
		(state)->H.wide[14] = h7h; \
293
		(state)->H.wide[15] = h7l; \
294
	} while (0)
295

296
#define INPUT_BUF1 \
297
	sph_u64 m0h = dec64e_aligned(buf +  0); \
298
	sph_u64 m0l = dec64e_aligned(buf +  8); \
299
	sph_u64 m1h = dec64e_aligned(buf + 16); \
300
	sph_u64 m1l = dec64e_aligned(buf + 24); \
301
	sph_u64 m2h = dec64e_aligned(buf + 32); \
302
	sph_u64 m2l = dec64e_aligned(buf + 40); \
303
	sph_u64 m3h = dec64e_aligned(buf + 48); \
304
	sph_u64 m3l = dec64e_aligned(buf + 56); \
305
	h0h ^= m0h; \
306
	h0l ^= m0l; \
307
	h1h ^= m1h; \
308
	h1l ^= m1l; \
309
	h2h ^= m2h; \
310
	h2l ^= m2l; \
311
	h3h ^= m3h; \
312
	h3l ^= m3l;
313

314
#define INPUT_BUF2 \
315
	h4h ^= m0h; \
316
	h4l ^= m0l; \
317
	h5h ^= m1h; \
318
	h5l ^= m1l; \
319
	h6h ^= m2h; \
320
	h6l ^= m2l; \
321
	h7h ^= m3h; \
322
	h7l ^= m3l;
323

324
static const sph_u64 IV224[] = {
325
	C64e(0x2dfedd62f99a98ac), C64e(0xae7cacd619d634e7),
326
	C64e(0xa4831005bc301216), C64e(0xb86038c6c9661494),
327
	C64e(0x66d9899f2580706f), C64e(0xce9ea31b1d9b1adc),
328
	C64e(0x11e8325f7b366e10), C64e(0xf994857f02fa06c1),
329
	C64e(0x1b4f1b5cd8c840b3), C64e(0x97f6a17f6e738099),
330
	C64e(0xdcdf93a5adeaa3d3), C64e(0xa431e8dec9539a68),
331
	C64e(0x22b4a98aec86a1e4), C64e(0xd574ac959ce56cf0),
332
	C64e(0x15960deab5ab2bbf), C64e(0x9611dcf0dd64ea6e)
333
};
334

335
static const sph_u64 IV256[] = {
336
	C64e(0xeb98a3412c20d3eb), C64e(0x92cdbe7b9cb245c1),
337
	C64e(0x1c93519160d4c7fa), C64e(0x260082d67e508a03),
338
	C64e(0xa4239e267726b945), C64e(0xe0fb1a48d41a9477),
339
	C64e(0xcdb5ab26026b177a), C64e(0x56f024420fff2fa8),
340
	C64e(0x71a396897f2e4d75), C64e(0x1d144908f77de262),
341
	C64e(0x277695f776248f94), C64e(0x87d5b6574780296c),
342
	C64e(0x5c5e272dac8e0d6c), C64e(0x518450c657057a0f),
343
	C64e(0x7be4d367702412ea), C64e(0x89e3ab13d31cd769)
344
};
345

346
static const sph_u64 IV384[] = {
347
	C64e(0x481e3bc6d813398a), C64e(0x6d3b5e894ade879b),
348
	C64e(0x63faea68d480ad2e), C64e(0x332ccb21480f8267),
349
	C64e(0x98aec84d9082b928), C64e(0xd455ea3041114249),
350
	C64e(0x36f555b2924847ec), C64e(0xc7250a93baf43ce1),
351
	C64e(0x569b7f8a27db454c), C64e(0x9efcbd496397af0e),
352
	C64e(0x589fc27d26aa80cd), C64e(0x80c08b8c9deb2eda),
353
	C64e(0x8a7981e8f8d5373a), C64e(0xf43967adddd17a71),
354
	C64e(0xa9b4d3bda475d394), C64e(0x976c3fba9842737f)
355
};
356

357
static const sph_u64 IV512[] = {
358
	C64e(0x6fd14b963e00aa17), C64e(0x636a2e057a15d543),
359
	C64e(0x8a225e8d0c97ef0b), C64e(0xe9341259f2b3c361),
360
	C64e(0x891da0c1536f801e), C64e(0x2aa9056bea2b6d80),
361
	C64e(0x588eccdb2075baa6), C64e(0xa90f3a76baf83bf7),
362
	C64e(0x0169e60541e34a69), C64e(0x46b58a8e2e6fe65a),
363
	C64e(0x1047a7d0c1843c24), C64e(0x3b6e71b12d5ac199),
364
	C64e(0xcf57f6ec9db1f856), C64e(0xa706887c5716b156),
365
	C64e(0xe3c2fcdfe68517fb), C64e(0x545a4678cc8cdd4b)
366
};
367

368
#else
369

370
static const sph_u32 C[] = {
371
	C32e(0x72d5dea2), C32e(0xdf15f867), C32e(0x7b84150a),
372
	C32e(0xb7231557), C32e(0x81abd690), C32e(0x4d5a87f6),
373
	C32e(0x4e9f4fc5), C32e(0xc3d12b40), C32e(0xea983ae0),
374
	C32e(0x5c45fa9c), C32e(0x03c5d299), C32e(0x66b2999a),
375
	C32e(0x660296b4), C32e(0xf2bb538a), C32e(0xb556141a),
376
	C32e(0x88dba231), C32e(0x03a35a5c), C32e(0x9a190edb),
377
	C32e(0x403fb20a), C32e(0x87c14410), C32e(0x1c051980),
378
	C32e(0x849e951d), C32e(0x6f33ebad), C32e(0x5ee7cddc),
379
	C32e(0x10ba1392), C32e(0x02bf6b41), C32e(0xdc786515),
380
	C32e(0xf7bb27d0), C32e(0x0a2c8139), C32e(0x37aa7850),
381
	C32e(0x3f1abfd2), C32e(0x410091d3), C32e(0x422d5a0d),
382
	C32e(0xf6cc7e90), C32e(0xdd629f9c), C32e(0x92c097ce),
383
	C32e(0x185ca70b), C32e(0xc72b44ac), C32e(0xd1df65d6),
384
	C32e(0x63c6fc23), C32e(0x976e6c03), C32e(0x9ee0b81a),
385
	C32e(0x2105457e), C32e(0x446ceca8), C32e(0xeef103bb),
386
	C32e(0x5d8e61fa), C32e(0xfd9697b2), C32e(0x94838197),
387
	C32e(0x4a8e8537), C32e(0xdb03302f), C32e(0x2a678d2d),
388
	C32e(0xfb9f6a95), C32e(0x8afe7381), C32e(0xf8b8696c),
389
	C32e(0x8ac77246), C32e(0xc07f4214), C32e(0xc5f4158f),
390
	C32e(0xbdc75ec4), C32e(0x75446fa7), C32e(0x8f11bb80),
391
	C32e(0x52de75b7), C32e(0xaee488bc), C32e(0x82b8001e),
392
	C32e(0x98a6a3f4), C32e(0x8ef48f33), C32e(0xa9a36315),
393
	C32e(0xaa5f5624), C32e(0xd5b7f989), C32e(0xb6f1ed20),
394
	C32e(0x7c5ae0fd), C32e(0x36cae95a), C32e(0x06422c36),
395
	C32e(0xce293543), C32e(0x4efe983d), C32e(0x533af974),
396
	C32e(0x739a4ba7), C32e(0xd0f51f59), C32e(0x6f4e8186),
397
	C32e(0x0e9dad81), C32e(0xafd85a9f), C32e(0xa7050667),
398
	C32e(0xee34626a), C32e(0x8b0b28be), C32e(0x6eb91727),
399
	C32e(0x47740726), C32e(0xc680103f), C32e(0xe0a07e6f),
400
	C32e(0xc67e487b), C32e(0x0d550aa5), C32e(0x4af8a4c0),
401
	C32e(0x91e3e79f), C32e(0x978ef19e), C32e(0x86767281),
402
	C32e(0x50608dd4), C32e(0x7e9e5a41), C32e(0xf3e5b062),
403
	C32e(0xfc9f1fec), C32e(0x4054207a), C32e(0xe3e41a00),
404
	C32e(0xcef4c984), C32e(0x4fd794f5), C32e(0x9dfa95d8),
405
	C32e(0x552e7e11), C32e(0x24c354a5), C32e(0x5bdf7228),
406
	C32e(0xbdfe6e28), C32e(0x78f57fe2), C32e(0x0fa5c4b2),
407
	C32e(0x05897cef), C32e(0xee49d32e), C32e(0x447e9385),
408
	C32e(0xeb28597f), C32e(0x705f6937), C32e(0xb324314a),
409
	C32e(0x5e8628f1), C32e(0x1dd6e465), C32e(0xc71b7704),
410
	C32e(0x51b920e7), C32e(0x74fe43e8), C32e(0x23d4878a),
411
	C32e(0x7d29e8a3), C32e(0x927694f2), C32e(0xddcb7a09),
412
	C32e(0x9b30d9c1), C32e(0x1d1b30fb), C32e(0x5bdc1be0),
413
	C32e(0xda24494f), C32e(0xf29c82bf), C32e(0xa4e7ba31),
414
	C32e(0xb470bfff), C32e(0x0d324405), C32e(0xdef8bc48),
415
	C32e(0x3baefc32), C32e(0x53bbd339), C32e(0x459fc3c1),
416
	C32e(0xe0298ba0), C32e(0xe5c905fd), C32e(0xf7ae090f),
417
	C32e(0x94703412), C32e(0x4290f134), C32e(0xa271b701),
418
	C32e(0xe344ed95), C32e(0xe93b8e36), C32e(0x4f2f984a),
419
	C32e(0x88401d63), C32e(0xa06cf615), C32e(0x47c1444b),
420
	C32e(0x8752afff), C32e(0x7ebb4af1), C32e(0xe20ac630),
421
	C32e(0x4670b6c5), C32e(0xcc6e8ce6), C32e(0xa4d5a456),
422
	C32e(0xbd4fca00), C32e(0xda9d844b), C32e(0xc83e18ae),
423
	C32e(0x7357ce45), C32e(0x3064d1ad), C32e(0xe8a6ce68),
424
	C32e(0x145c2567), C32e(0xa3da8cf2), C32e(0xcb0ee116),
425
	C32e(0x33e90658), C32e(0x9a94999a), C32e(0x1f60b220),
426
	C32e(0xc26f847b), C32e(0xd1ceac7f), C32e(0xa0d18518),
427
	C32e(0x32595ba1), C32e(0x8ddd19d3), C32e(0x509a1cc0),
428
	C32e(0xaaa5b446), C32e(0x9f3d6367), C32e(0xe4046bba),
429
	C32e(0xf6ca19ab), C32e(0x0b56ee7e), C32e(0x1fb179ea),
430
	C32e(0xa9282174), C32e(0xe9bdf735), C32e(0x3b3651ee),
431
	C32e(0x1d57ac5a), C32e(0x7550d376), C32e(0x3a46c2fe),
432
	C32e(0xa37d7001), C32e(0xf735c1af), C32e(0x98a4d842),
433
	C32e(0x78edec20), C32e(0x9e6b6779), C32e(0x41836315),
434
	C32e(0xea3adba8), C32e(0xfac33b4d), C32e(0x32832c83),
435
	C32e(0xa7403b1f), C32e(0x1c2747f3), C32e(0x5940f034),
436
	C32e(0xb72d769a), C32e(0xe73e4e6c), C32e(0xd2214ffd),
437
	C32e(0xb8fd8d39), C32e(0xdc5759ef), C32e(0x8d9b0c49),
438
	C32e(0x2b49ebda), C32e(0x5ba2d749), C32e(0x68f3700d),
439
	C32e(0x7d3baed0), C32e(0x7a8d5584), C32e(0xf5a5e9f0),
440
	C32e(0xe4f88e65), C32e(0xa0b8a2f4), C32e(0x36103b53),
441
	C32e(0x0ca8079e), C32e(0x753eec5a), C32e(0x91689492),
442
	C32e(0x56e8884f), C32e(0x5bb05c55), C32e(0xf8babc4c),
443
	C32e(0xe3bb3b99), C32e(0xf387947b), C32e(0x75daf4d6),
444
	C32e(0x726b1c5d), C32e(0x64aeac28), C32e(0xdc34b36d),
445
	C32e(0x6c34a550), C32e(0xb828db71), C32e(0xf861e2f2),
446
	C32e(0x108d512a), C32e(0xe3db6433), C32e(0x59dd75fc),
447
	C32e(0x1cacbcf1), C32e(0x43ce3fa2), C32e(0x67bbd13c),
448
	C32e(0x02e843b0), C32e(0x330a5bca), C32e(0x8829a175),
449
	C32e(0x7f34194d), C32e(0xb416535c), C32e(0x923b94c3),
450
	C32e(0x0e794d1e), C32e(0x797475d7), C32e(0xb6eeaf3f),
451
	C32e(0xeaa8d4f7), C32e(0xbe1a3921), C32e(0x5cf47e09),
452
	C32e(0x4c232751), C32e(0x26a32453), C32e(0xba323cd2),
453
	C32e(0x44a3174a), C32e(0x6da6d5ad), C32e(0xb51d3ea6),
454
	C32e(0xaff2c908), C32e(0x83593d98), C32e(0x916b3c56),
455
	C32e(0x4cf87ca1), C32e(0x7286604d), C32e(0x46e23ecc),
456
	C32e(0x086ec7f6), C32e(0x2f9833b3), C32e(0xb1bc765e),
457
	C32e(0x2bd666a5), C32e(0xefc4e62a), C32e(0x06f4b6e8),
458
	C32e(0xbec1d436), C32e(0x74ee8215), C32e(0xbcef2163),
459
	C32e(0xfdc14e0d), C32e(0xf453c969), C32e(0xa77d5ac4),
460
	C32e(0x06585826), C32e(0x7ec11416), C32e(0x06e0fa16),
461
	C32e(0x7e90af3d), C32e(0x28639d3f), C32e(0xd2c9f2e3),
462
	C32e(0x009bd20c), C32e(0x5faace30), C32e(0xb7d40c30),
463
	C32e(0x742a5116), C32e(0xf2e03298), C32e(0x0deb30d8),
464
	C32e(0xe3cef89a), C32e(0x4bc59e7b), C32e(0xb5f17992),
465
	C32e(0xff51e66e), C32e(0x048668d3), C32e(0x9b234d57),
466
	C32e(0xe6966731), C32e(0xcce6a6f3), C32e(0x170a7505),
467
	C32e(0xb17681d9), C32e(0x13326cce), C32e(0x3c175284),
468
	C32e(0xf805a262), C32e(0xf42bcbb3), C32e(0x78471547),
469
	C32e(0xff465482), C32e(0x23936a48), C32e(0x38df5807),
470
	C32e(0x4e5e6565), C32e(0xf2fc7c89), C32e(0xfc86508e),
471
	C32e(0x31702e44), C32e(0xd00bca86), C32e(0xf04009a2),
472
	C32e(0x3078474e), C32e(0x65a0ee39), C32e(0xd1f73883),
473
	C32e(0xf75ee937), C32e(0xe42c3abd), C32e(0x2197b226),
474
	C32e(0x0113f86f), C32e(0xa344edd1), C32e(0xef9fdee7),
475
	C32e(0x8ba0df15), C32e(0x762592d9), C32e(0x3c85f7f6),
476
	C32e(0x12dc42be), C32e(0xd8a7ec7c), C32e(0xab27b07e),
477
	C32e(0x538d7dda), C32e(0xaa3ea8de), C32e(0xaa25ce93),
478
	C32e(0xbd0269d8), C32e(0x5af643fd), C32e(0x1a7308f9),
479
	C32e(0xc05fefda), C32e(0x174a19a5), C32e(0x974d6633),
480
	C32e(0x4cfd216a), C32e(0x35b49831), C32e(0xdb411570),
481
	C32e(0xea1e0fbb), C32e(0xedcd549b), C32e(0x9ad063a1),
482
	C32e(0x51974072), C32e(0xf6759dbf), C32e(0x91476fe2)
483
};
484

485
#define Ceven_w3(r)   (C[((r) << 3) + 0])
486
#define Ceven_w2(r)   (C[((r) << 3) + 1])
487
#define Ceven_w1(r)   (C[((r) << 3) + 2])
488
#define Ceven_w0(r)   (C[((r) << 3) + 3])
489
#define Codd_w3(r)    (C[((r) << 3) + 4])
490
#define Codd_w2(r)    (C[((r) << 3) + 5])
491
#define Codd_w1(r)    (C[((r) << 3) + 6])
492
#define Codd_w0(r)    (C[((r) << 3) + 7])
493

494
#define S(x0, x1, x2, x3, cb, r)   do { \
495
		Sb(x0 ## 3, x1 ## 3, x2 ## 3, x3 ## 3, cb ## w3(r)); \
496
		Sb(x0 ## 2, x1 ## 2, x2 ## 2, x3 ## 2, cb ## w2(r)); \
497
		Sb(x0 ## 1, x1 ## 1, x2 ## 1, x3 ## 1, cb ## w1(r)); \
498
		Sb(x0 ## 0, x1 ## 0, x2 ## 0, x3 ## 0, cb ## w0(r)); \
499
	} while (0)
500

501
#define L(x0, x1, x2, x3, x4, x5, x6, x7)   do { \
502
		Lb(x0 ## 3, x1 ## 3, x2 ## 3, x3 ## 3, \
503
			x4 ## 3, x5 ## 3, x6 ## 3, x7 ## 3); \
504
		Lb(x0 ## 2, x1 ## 2, x2 ## 2, x3 ## 2, \
505
			x4 ## 2, x5 ## 2, x6 ## 2, x7 ## 2); \
506
		Lb(x0 ## 1, x1 ## 1, x2 ## 1, x3 ## 1, \
507
			x4 ## 1, x5 ## 1, x6 ## 1, x7 ## 1); \
508
		Lb(x0 ## 0, x1 ## 0, x2 ## 0, x3 ## 0, \
509
			x4 ## 0, x5 ## 0, x6 ## 0, x7 ## 0); \
510
	} while (0)
511

512
#define Wz(x, c, n)   do { \
513
		sph_u32 t = (x ## 3 & (c)) << (n); \
514
		x ## 3 = ((x ## 3 >> (n)) & (c)) | t; \
515
		t = (x ## 2 & (c)) << (n); \
516
		x ## 2 = ((x ## 2 >> (n)) & (c)) | t; \
517
		t = (x ## 1 & (c)) << (n); \
518
		x ## 1 = ((x ## 1 >> (n)) & (c)) | t; \
519
		t = (x ## 0 & (c)) << (n); \
520
		x ## 0 = ((x ## 0 >> (n)) & (c)) | t; \
521
	} while (0)
522

523
#define W0(x)   Wz(x, SPH_C32(0x55555555),  1)
524
#define W1(x)   Wz(x, SPH_C32(0x33333333),  2)
525
#define W2(x)   Wz(x, SPH_C32(0x0F0F0F0F),  4)
526
#define W3(x)   Wz(x, SPH_C32(0x00FF00FF),  8)
527
#define W4(x)   Wz(x, SPH_C32(0x0000FFFF), 16)
528
#define W5(x)   do { \
529
		sph_u32 t = x ## 3; \
530
		x ## 3 = x ## 2; \
531
		x ## 2 = t; \
532
		t = x ## 1; \
533
		x ## 1 = x ## 0; \
534
		x ## 0 = t; \
535
	} while (0)
536
#define W6(x)   do { \
537
		sph_u32 t = x ## 3; \
538
		x ## 3 = x ## 1; \
539
		x ## 1 = t; \
540
		t = x ## 2; \
541
		x ## 2 = x ## 0; \
542
		x ## 0 = t; \
543
	} while (0)
544

545
#define DECL_STATE \
546
	sph_u32 h03, h02, h01, h00, h13, h12, h11, h10; \
547
	sph_u32 h23, h22, h21, h20, h33, h32, h31, h30; \
548
	sph_u32 h43, h42, h41, h40, h53, h52, h51, h50; \
549
	sph_u32 h63, h62, h61, h60, h73, h72, h71, h70; \
550
	sph_u32 tmp;
551

552
#define READ_STATE(state)   do { \
553
		h03 = (state)->H.narrow[ 0]; \
554
		h02 = (state)->H.narrow[ 1]; \
555
		h01 = (state)->H.narrow[ 2]; \
556
		h00 = (state)->H.narrow[ 3]; \
557
		h13 = (state)->H.narrow[ 4]; \
558
		h12 = (state)->H.narrow[ 5]; \
559
		h11 = (state)->H.narrow[ 6]; \
560
		h10 = (state)->H.narrow[ 7]; \
561
		h23 = (state)->H.narrow[ 8]; \
562
		h22 = (state)->H.narrow[ 9]; \
563
		h21 = (state)->H.narrow[10]; \
564
		h20 = (state)->H.narrow[11]; \
565
		h33 = (state)->H.narrow[12]; \
566
		h32 = (state)->H.narrow[13]; \
567
		h31 = (state)->H.narrow[14]; \
568
		h30 = (state)->H.narrow[15]; \
569
		h43 = (state)->H.narrow[16]; \
570
		h42 = (state)->H.narrow[17]; \
571
		h41 = (state)->H.narrow[18]; \
572
		h40 = (state)->H.narrow[19]; \
573
		h53 = (state)->H.narrow[20]; \
574
		h52 = (state)->H.narrow[21]; \
575
		h51 = (state)->H.narrow[22]; \
576
		h50 = (state)->H.narrow[23]; \
577
		h63 = (state)->H.narrow[24]; \
578
		h62 = (state)->H.narrow[25]; \
579
		h61 = (state)->H.narrow[26]; \
580
		h60 = (state)->H.narrow[27]; \
581
		h73 = (state)->H.narrow[28]; \
582
		h72 = (state)->H.narrow[29]; \
583
		h71 = (state)->H.narrow[30]; \
584
		h70 = (state)->H.narrow[31]; \
585
	} while (0)
586

587
#define WRITE_STATE(state)   do { \
588
		(state)->H.narrow[ 0] = h03; \
589
		(state)->H.narrow[ 1] = h02; \
590
		(state)->H.narrow[ 2] = h01; \
591
		(state)->H.narrow[ 3] = h00; \
592
		(state)->H.narrow[ 4] = h13; \
593
		(state)->H.narrow[ 5] = h12; \
594
		(state)->H.narrow[ 6] = h11; \
595
		(state)->H.narrow[ 7] = h10; \
596
		(state)->H.narrow[ 8] = h23; \
597
		(state)->H.narrow[ 9] = h22; \
598
		(state)->H.narrow[10] = h21; \
599
		(state)->H.narrow[11] = h20; \
600
		(state)->H.narrow[12] = h33; \
601
		(state)->H.narrow[13] = h32; \
602
		(state)->H.narrow[14] = h31; \
603
		(state)->H.narrow[15] = h30; \
604
		(state)->H.narrow[16] = h43; \
605
		(state)->H.narrow[17] = h42; \
606
		(state)->H.narrow[18] = h41; \
607
		(state)->H.narrow[19] = h40; \
608
		(state)->H.narrow[20] = h53; \
609
		(state)->H.narrow[21] = h52; \
610
		(state)->H.narrow[22] = h51; \
611
		(state)->H.narrow[23] = h50; \
612
		(state)->H.narrow[24] = h63; \
613
		(state)->H.narrow[25] = h62; \
614
		(state)->H.narrow[26] = h61; \
615
		(state)->H.narrow[27] = h60; \
616
		(state)->H.narrow[28] = h73; \
617
		(state)->H.narrow[29] = h72; \
618
		(state)->H.narrow[30] = h71; \
619
		(state)->H.narrow[31] = h70; \
620
	} while (0)
621

622
#define INPUT_BUF1 \
623
	sph_u32 m03 = dec32e_aligned(buf +  0); \
624
	sph_u32 m02 = dec32e_aligned(buf +  4); \
625
	sph_u32 m01 = dec32e_aligned(buf +  8); \
626
	sph_u32 m00 = dec32e_aligned(buf + 12); \
627
	sph_u32 m13 = dec32e_aligned(buf + 16); \
628
	sph_u32 m12 = dec32e_aligned(buf + 20); \
629
	sph_u32 m11 = dec32e_aligned(buf + 24); \
630
	sph_u32 m10 = dec32e_aligned(buf + 28); \
631
	sph_u32 m23 = dec32e_aligned(buf + 32); \
632
	sph_u32 m22 = dec32e_aligned(buf + 36); \
633
	sph_u32 m21 = dec32e_aligned(buf + 40); \
634
	sph_u32 m20 = dec32e_aligned(buf + 44); \
635
	sph_u32 m33 = dec32e_aligned(buf + 48); \
636
	sph_u32 m32 = dec32e_aligned(buf + 52); \
637
	sph_u32 m31 = dec32e_aligned(buf + 56); \
638
	sph_u32 m30 = dec32e_aligned(buf + 60); \
639
	h03 ^= m03; \
640
	h02 ^= m02; \
641
	h01 ^= m01; \
642
	h00 ^= m00; \
643
	h13 ^= m13; \
644
	h12 ^= m12; \
645
	h11 ^= m11; \
646
	h10 ^= m10; \
647
	h23 ^= m23; \
648
	h22 ^= m22; \
649
	h21 ^= m21; \
650
	h20 ^= m20; \
651
	h33 ^= m33; \
652
	h32 ^= m32; \
653
	h31 ^= m31; \
654
	h30 ^= m30;
655

656
#define INPUT_BUF2 \
657
	h43 ^= m03; \
658
	h42 ^= m02; \
659
	h41 ^= m01; \
660
	h40 ^= m00; \
661
	h53 ^= m13; \
662
	h52 ^= m12; \
663
	h51 ^= m11; \
664
	h50 ^= m10; \
665
	h63 ^= m23; \
666
	h62 ^= m22; \
667
	h61 ^= m21; \
668
	h60 ^= m20; \
669
	h73 ^= m33; \
670
	h72 ^= m32; \
671
	h71 ^= m31; \
672
	h70 ^= m30;
673

674
static const sph_u32 IV224[] = {
675
	C32e(0x2dfedd62), C32e(0xf99a98ac), C32e(0xae7cacd6), C32e(0x19d634e7),
676
	C32e(0xa4831005), C32e(0xbc301216), C32e(0xb86038c6), C32e(0xc9661494),
677
	C32e(0x66d9899f), C32e(0x2580706f), C32e(0xce9ea31b), C32e(0x1d9b1adc),
678
	C32e(0x11e8325f), C32e(0x7b366e10), C32e(0xf994857f), C32e(0x02fa06c1),
679
	C32e(0x1b4f1b5c), C32e(0xd8c840b3), C32e(0x97f6a17f), C32e(0x6e738099),
680
	C32e(0xdcdf93a5), C32e(0xadeaa3d3), C32e(0xa431e8de), C32e(0xc9539a68),
681
	C32e(0x22b4a98a), C32e(0xec86a1e4), C32e(0xd574ac95), C32e(0x9ce56cf0),
682
	C32e(0x15960dea), C32e(0xb5ab2bbf), C32e(0x9611dcf0), C32e(0xdd64ea6e)
683
};
684

685
static const sph_u32 IV256[] = {
686
	C32e(0xeb98a341), C32e(0x2c20d3eb), C32e(0x92cdbe7b), C32e(0x9cb245c1),
687
	C32e(0x1c935191), C32e(0x60d4c7fa), C32e(0x260082d6), C32e(0x7e508a03),
688
	C32e(0xa4239e26), C32e(0x7726b945), C32e(0xe0fb1a48), C32e(0xd41a9477),
689
	C32e(0xcdb5ab26), C32e(0x026b177a), C32e(0x56f02442), C32e(0x0fff2fa8),
690
	C32e(0x71a39689), C32e(0x7f2e4d75), C32e(0x1d144908), C32e(0xf77de262),
691
	C32e(0x277695f7), C32e(0x76248f94), C32e(0x87d5b657), C32e(0x4780296c),
692
	C32e(0x5c5e272d), C32e(0xac8e0d6c), C32e(0x518450c6), C32e(0x57057a0f),
693
	C32e(0x7be4d367), C32e(0x702412ea), C32e(0x89e3ab13), C32e(0xd31cd769)
694
};
695

696
static const sph_u32 IV384[] = {
697
	C32e(0x481e3bc6), C32e(0xd813398a), C32e(0x6d3b5e89), C32e(0x4ade879b),
698
	C32e(0x63faea68), C32e(0xd480ad2e), C32e(0x332ccb21), C32e(0x480f8267),
699
	C32e(0x98aec84d), C32e(0x9082b928), C32e(0xd455ea30), C32e(0x41114249),
700
	C32e(0x36f555b2), C32e(0x924847ec), C32e(0xc7250a93), C32e(0xbaf43ce1),
701
	C32e(0x569b7f8a), C32e(0x27db454c), C32e(0x9efcbd49), C32e(0x6397af0e),
702
	C32e(0x589fc27d), C32e(0x26aa80cd), C32e(0x80c08b8c), C32e(0x9deb2eda),
703
	C32e(0x8a7981e8), C32e(0xf8d5373a), C32e(0xf43967ad), C32e(0xddd17a71),
704
	C32e(0xa9b4d3bd), C32e(0xa475d394), C32e(0x976c3fba), C32e(0x9842737f)
705
};
706

707
static const sph_u32 IV512[] = {
708
	C32e(0x6fd14b96), C32e(0x3e00aa17), C32e(0x636a2e05), C32e(0x7a15d543),
709
	C32e(0x8a225e8d), C32e(0x0c97ef0b), C32e(0xe9341259), C32e(0xf2b3c361),
710
	C32e(0x891da0c1), C32e(0x536f801e), C32e(0x2aa9056b), C32e(0xea2b6d80),
711
	C32e(0x588eccdb), C32e(0x2075baa6), C32e(0xa90f3a76), C32e(0xbaf83bf7),
712
	C32e(0x0169e605), C32e(0x41e34a69), C32e(0x46b58a8e), C32e(0x2e6fe65a),
713
	C32e(0x1047a7d0), C32e(0xc1843c24), C32e(0x3b6e71b1), C32e(0x2d5ac199),
714
	C32e(0xcf57f6ec), C32e(0x9db1f856), C32e(0xa706887c), C32e(0x5716b156),
715
	C32e(0xe3c2fcdf), C32e(0xe68517fb), C32e(0x545a4678), C32e(0xcc8cdd4b)
716
};
717

718
#endif
719

720
#define SL(ro)   SLu(r + ro, ro)
721

722
#define SLu(r, ro)   do { \
723
		S(h0, h2, h4, h6, Ceven_, r); \
724
		S(h1, h3, h5, h7, Codd_, r); \
725
		L(h0, h2, h4, h6, h1, h3, h5, h7); \
726
		W ## ro(h1); \
727
		W ## ro(h3); \
728
		W ## ro(h5); \
729
		W ## ro(h7); \
730
	} while (0)
731

732
#if SPH_SMALL_FOOTPRINT_JH
733

734
#if SPH_JH_64
735

736
/*
737
 * The "small footprint" 64-bit version just uses a partially unrolled
738
 * loop.
739
 */
740

741
#define E8   do { \
742
		unsigned r; \
743
		for (r = 0; r < 42; r += 7) { \
744
			SL(0); \
745
			SL(1); \
746
			SL(2); \
747
			SL(3); \
748
			SL(4); \
749
			SL(5); \
750
			SL(6); \
751
		} \
752
	} while (0)
753

754
#else
755

756
#define E8   do { \
757
		unsigned r, g; \
758
		for (r = g = 0; r < 42; r ++) { \
759
			S(h0, h2, h4, h6, Ceven_, r); \
760
			S(h1, h3, h5, h7, Codd_, r); \
761
			L(h0, h2, h4, h6, h1, h3, h5, h7); \
762
			switch (g) { \
763
			case 0: \
764
				W0(h1); \
765
				W0(h3); \
766
				W0(h5); \
767
				W0(h7); \
768
				break; \
769
			case 1: \
770
				W1(h1); \
771
				W1(h3); \
772
				W1(h5); \
773
				W1(h7); \
774
				break; \
775
			case 2: \
776
				W2(h1); \
777
				W2(h3); \
778
				W2(h5); \
779
				W2(h7); \
780
				break; \
781
			case 3: \
782
				W3(h1); \
783
				W3(h3); \
784
				W3(h5); \
785
				W3(h7); \
786
				break; \
787
			case 4: \
788
				W4(h1); \
789
				W4(h3); \
790
				W4(h5); \
791
				W4(h7); \
792
				break; \
793
			case 5: \
794
				W5(h1); \
795
				W5(h3); \
796
				W5(h5); \
797
				W5(h7); \
798
				break; \
799
			case 6: \
800
				W6(h1); \
801
				W6(h3); \
802
				W6(h5); \
803
				W6(h7); \
804
				break; \
805
			} \
806
			if (++ g == 7) \
807
				g = 0; \
808
		} \
809
	} while (0)
810

811
#endif
812

813
#else
814

815
#if SPH_JH_64
816

817
/*
818
 * On a "true 64-bit" architecture, we can unroll at will.
819
 */
820

821
#define E8   do { \
822
		SLu( 0, 0); \
823
		SLu( 1, 1); \
824
		SLu( 2, 2); \
825
		SLu( 3, 3); \
826
		SLu( 4, 4); \
827
		SLu( 5, 5); \
828
		SLu( 6, 6); \
829
		SLu( 7, 0); \
830
		SLu( 8, 1); \
831
		SLu( 9, 2); \
832
		SLu(10, 3); \
833
		SLu(11, 4); \
834
		SLu(12, 5); \
835
		SLu(13, 6); \
836
		SLu(14, 0); \
837
		SLu(15, 1); \
838
		SLu(16, 2); \
839
		SLu(17, 3); \
840
		SLu(18, 4); \
841
		SLu(19, 5); \
842
		SLu(20, 6); \
843
		SLu(21, 0); \
844
		SLu(22, 1); \
845
		SLu(23, 2); \
846
		SLu(24, 3); \
847
		SLu(25, 4); \
848
		SLu(26, 5); \
849
		SLu(27, 6); \
850
		SLu(28, 0); \
851
		SLu(29, 1); \
852
		SLu(30, 2); \
853
		SLu(31, 3); \
854
		SLu(32, 4); \
855
		SLu(33, 5); \
856
		SLu(34, 6); \
857
		SLu(35, 0); \
858
		SLu(36, 1); \
859
		SLu(37, 2); \
860
		SLu(38, 3); \
861
		SLu(39, 4); \
862
		SLu(40, 5); \
863
		SLu(41, 6); \
864
	} while (0)
865

866
#else
867

868
/*
869
 * We are not aiming at a small footprint, but we are still using a
870
 * 32-bit implementation. Full loop unrolling would smash the L1
871
 * cache on some "big" architectures (32 kB L1 cache).
872
 */
873

874
#define E8   do { \
875
		unsigned r; \
876
		for (r = 0; r < 42; r += 7) { \
877
			SL(0); \
878
			SL(1); \
879
			SL(2); \
880
			SL(3); \
881
			SL(4); \
882
			SL(5); \
883
			SL(6); \
884
		} \
885
	} while (0)
886

887
#endif
888

889
#endif
890

891
static void
892
jh_init(sph_jh_context *sc, const void *iv)
893
{
894
	sc->ptr = 0;
895
#if SPH_JH_64
896
	memcpy(sc->H.wide, iv, sizeof sc->H.wide);
897
#else
898
	memcpy(sc->H.narrow, iv, sizeof sc->H.narrow);
899
#endif
900
#if SPH_64
901
	sc->block_count = 0;
902
#else
903
	sc->block_count_high = 0;
904
	sc->block_count_low = 0;
905
#endif
906
}
907

908
static void
909
jh_core(sph_jh_context *sc, const void *data, size_t len)
910
{
911
	unsigned char *buf;
912
	size_t ptr;
913
	DECL_STATE
914

915
	buf = sc->buf;
916
	ptr = sc->ptr;
917
	if (len < (sizeof sc->buf) - ptr) {
918
		memcpy(buf + ptr, data, len);
919
		ptr += len;
920
		sc->ptr = ptr;
921
		return;
922
	}
923

924
	READ_STATE(sc);
925
	while (len > 0) {
926
		size_t clen;
927

928
		clen = (sizeof sc->buf) - ptr;
929
		if (clen > len)
930
			clen = len;
931
		memcpy(buf + ptr, data, clen);
932
		ptr += clen;
933
		data = (const unsigned char *)data + clen;
934
		len -= clen;
935
		if (ptr == sizeof sc->buf) {
936
			INPUT_BUF1;
937
			E8;
938
			INPUT_BUF2;
939
#if SPH_64
940
			sc->block_count ++;
941
#else
942
			if ((sc->block_count_low = SPH_T32(
943
				sc->block_count_low + 1)) == 0)
944
				sc->block_count_high ++;
945
#endif
946
			ptr = 0;
947
		}
948
	}
949
	WRITE_STATE(sc);
950
	sc->ptr = ptr;
951
}
952

953
static void
954
jh_close(sph_jh_context *sc, unsigned ub, unsigned n,
955
	void *dst, size_t out_size_w32, const void *iv)
956
{
957
	unsigned z;
958
	unsigned char buf[128];
959
	size_t numz, u;
960
#if SPH_64
961
	sph_u64 l0, l1;
962
#else
963
	sph_u32 l0, l1, l2, l3;
964
#endif
965

966
	z = 0x80 >> n;
967
	buf[0] = ((ub & -z) | z) & 0xFF;
968
	if (sc->ptr == 0 && n == 0) {
969
		numz = 47;
970
	} else {
971
		numz = 111 - sc->ptr;
972
	}
973
	memset(buf + 1, 0, numz);
974
#if SPH_64
975
	l0 = SPH_T64(sc->block_count << 9) + (sc->ptr << 3) + n;
976
	l1 = SPH_T64(sc->block_count >> 55);
977
	sph_enc64be(buf + numz + 1, l1);
978
	sph_enc64be(buf + numz + 9, l0);
979
#else
980
	l0 = SPH_T32(sc->block_count_low << 9) + (sc->ptr << 3) + n;
981
	l1 = SPH_T32(sc->block_count_low >> 23)
982
		+ SPH_T32(sc->block_count_high << 9);
983
	l2 = SPH_T32(sc->block_count_high >> 23);
984
	l3 = 0;
985
	sph_enc32be(buf + numz +  1, l3);
986
	sph_enc32be(buf + numz +  5, l2);
987
	sph_enc32be(buf + numz +  9, l1);
988
	sph_enc32be(buf + numz + 13, l0);
989
#endif
990
	jh_core(sc, buf, numz + 17);
991
#if SPH_JH_64
992
	for (u = 0; u < 8; u ++)
993
		enc64e(buf + (u << 3), sc->H.wide[u + 8]);
994
#else
995
	for (u = 0; u < 16; u ++)
996
		enc32e(buf + (u << 2), sc->H.narrow[u + 16]);
997
#endif
998
	memcpy(dst, buf + ((16 - out_size_w32) << 2), out_size_w32 << 2);
999
	jh_init(sc, iv);
1000
}
1001

1002
/* see sph_jh.h */
1003
void
1004
sph_jh224_init(void *cc)
1005
{
1006
	jh_init(cc, IV224);
1007
}
1008

1009
/* see sph_jh.h */
1010
void
1011
sph_jh224(void *cc, const void *data, size_t len)
1012
{
1013
	jh_core(cc, data, len);
1014
}
1015

1016
/* see sph_jh.h */
1017
void
1018
sph_jh224_close(void *cc, void *dst)
1019
{
1020
	jh_close(cc, 0, 0, dst, 7, IV224);
1021
}
1022

1023
/* see sph_jh.h */
1024
void
1025
sph_jh224_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
1026
{
1027
	jh_close(cc, ub, n, dst, 7, IV224);
1028
}
1029

1030
/* see sph_jh.h */
1031
void
1032
sph_jh256_init(void *cc)
1033
{
1034
	jh_init(cc, IV256);
1035
}
1036

1037
/* see sph_jh.h */
1038
void
1039
sph_jh256(void *cc, const void *data, size_t len)
1040
{
1041
	jh_core(cc, data, len);
1042
}
1043

1044
/* see sph_jh.h */
1045
void
1046
sph_jh256_close(void *cc, void *dst)
1047
{
1048
	jh_close(cc, 0, 0, dst, 8, IV256);
1049
}
1050

1051
/* see sph_jh.h */
1052
void
1053
sph_jh256_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
1054
{
1055
	jh_close(cc, ub, n, dst, 8, IV256);
1056
}
1057

1058
/* see sph_jh.h */
1059
void
1060
sph_jh384_init(void *cc)
1061
{
1062
	jh_init(cc, IV384);
1063
}
1064

1065
/* see sph_jh.h */
1066
void
1067
sph_jh384(void *cc, const void *data, size_t len)
1068
{
1069
	jh_core(cc, data, len);
1070
}
1071

1072
/* see sph_jh.h */
1073
void
1074
sph_jh384_close(void *cc, void *dst)
1075
{
1076
	jh_close(cc, 0, 0, dst, 12, IV384);
1077
}
1078

1079
/* see sph_jh.h */
1080
void
1081
sph_jh384_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
1082
{
1083
	jh_close(cc, ub, n, dst, 12, IV384);
1084
}
1085

1086
/* see sph_jh.h */
1087
void
1088
sph_jh512_init(void *cc)
1089
{
1090
	jh_init(cc, IV512);
1091
}
1092

1093
/* see sph_jh.h */
1094
void
1095
sph_jh512(void *cc, const void *data, size_t len)
1096
{
1097
	jh_core(cc, data, len);
1098
}
1099

1100
/* see sph_jh.h */
1101
void
1102
sph_jh512_close(void *cc, void *dst)
1103
{
1104
	jh_close(cc, 0, 0, dst, 16, IV512);
1105
}
1106

1107
/* see sph_jh.h */
1108
void
1109
sph_jh512_addbits_and_close(void *cc, unsigned ub, unsigned n, void *dst)
1110
{
1111
	jh_close(cc, ub, n, dst, 16, IV512);
1112
}
1113

1114
#ifdef __cplusplus
1115
}
1116
#endif
1117

1118
Product

Resources

Company
