ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/liblzf/lzf_c.c
(Generate patch)

Comparing liblzf/lzf_c.c (file contents):
Revision 1.22 by root, Tue Nov 13 10:41:52 2007 UTC vs.
Revision 1.51 by root, Wed May 24 18:37:18 2017 UTC

1/* 1/*
2 * Copyright (c) 2000-2007 Marc Alexander Lehmann <schmorp@schmorp.de> 2 * Copyright (c) 2000-2010,2012 Marc Alexander Lehmann <schmorp@schmorp.de>
3 * 3 *
4 * Redistribution and use in source and binary forms, with or without modifica- 4 * Redistribution and use in source and binary forms, with or without modifica-
5 * tion, are permitted provided that the following conditions are met: 5 * tion, are permitted provided that the following conditions are met:
6 * 6 *
7 * 1. Redistributions of source code must retain the above copyright notice, 7 * 1. Redistributions of source code must retain the above copyright notice,
8 * this list of conditions and the following disclaimer. 8 * this list of conditions and the following disclaimer.
9 * 9 *
10 * 2. Redistributions in binary form must reproduce the above copyright 10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the 11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution. 12 * documentation and/or other materials provided with the distribution.
13 * 13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
15 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER- 15 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER-
16 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO 16 * CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
17 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE- 17 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE-
18 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 18 * CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
… …
38 38
39#define HSIZE (1 << (HLOG)) 39#define HSIZE (1 << (HLOG))
40 40
41/* 41/*
42 * don't play with this unless you benchmark! 42 * don't play with this unless you benchmark!
43 * decompression is not dependent on the hash function 43 * the data format is not dependent on the hash function.
44 * the hashing function might seem strange, just believe me 44 * the hash function might seem strange, just believe me,
45 * it works ;) 45 * it works ;)
46 */ 46 */
47#ifndef FRST 47#ifndef FRST
48# define FRST(p) (((p[0]) << 8) | p[1]) 48# define FRST(p) (((p[0]) << 8) | p[1])
49# define NEXT(v,p) (((v) << 8) | p[2]) 49# define NEXT(v,p) (((v) << 8) | p[2])
50# if MULTIPLICATION_IS_SLOW
50# if ULTRA_FAST 51# if ULTRA_FAST
51# define IDX(h) (((h >> (3*8 - HLOG)) - h ) & (HSIZE - 1)) 52# define IDX(h) ((( h >> (3*8 - HLOG)) - h ) & (HSIZE - 1))
52# elif VERY_FAST 53# elif VERY_FAST
53# define IDX(h) (((h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) 54# define IDX(h) ((( h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1))
55# else
56# define IDX(h) ((((h ^ (h << 5)) >> (3*8 - HLOG)) - h*5) & (HSIZE - 1))
57# endif
54# else 58# else
59/* this one was developed with sesse,
60 * and is very similar to the one in snappy.
61 * it does need a modern enough cpu with a fast multiplication.
62 */
55# define IDX(h) ((((h ^ (h << 5)) >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) 63# define IDX(h) (((h * 0x1e35a7bdU) >> (32 - HLOG - 8)) & (HSIZE - 1))
56# endif 64# endif
57/*# define IDX(h) ((ip[0] * 121 ^ ip[1] * 33 ^ ip[2] * 1) & (HSIZE-1))*/
58#endif 65#endif
59/*
60 * IDX works because it is very similar to a multiplicative hash, e.g.
61 * ((h * 57321 >> (3*8 - HLOG)) & (HSIZE - 1))
62 * the latter is also quite fast on newer CPUs, and compresses similarly.
63 *
64 * the next one is also quite good, albeit slow ;)
65 * (int)(cos(h & 0xffffff) * 1e6)
66 */
67 66
68#if 0 67#if 0
69/* original lzv-like hash function, much worse and thus slower */ 68/* original lzv-like hash function, much worse and thus slower */
70# define FRST(p) (p[0] << 5) ^ p[1] 69# define FRST(p) (p[0] << 5) ^ p[1]
71# define NEXT(v,p) ((v) << 5) ^ p[2] 70# define NEXT(v,p) ((v) << 5) ^ p[2]
72# define IDX(h) ((h) & (HSIZE - 1)) 71# define IDX(h) ((h) & (HSIZE - 1))
73#endif 72#endif
74 73
75#define MAX_LIT (1 << 5)
76#define MAX_OFF (1 << 13)
77#define MAX_REF ((1 << 8) + (1 << 3))
78
79#if __GNUC__ >= 3 74#if __GNUC__ >= 3
80# define expect(expr,value) __builtin_expect ((expr),(value)) 75# define expect(expr,value) __builtin_expect ((expr),(value))
81# define inline inline 76# define inline inline
82#else 77#else
83# define expect(expr,value) (expr) 78# define expect(expr,value) (expr)
… …
88#define expect_true(expr) expect ((expr) != 0, 1) 83#define expect_true(expr) expect ((expr) != 0, 1)
89 84
90/* 85/*
91 * compressed format 86 * compressed format
92 * 87 *
93 * 000LLLLL <L+1> ; literal 88 * 000LLLLL <L+1> ; literal, L+1=1..33 octets
94 * LLLooooo oooooooo ; backref L 89 * LLLooooo oooooooo ; backref L+1=1..7 octets, o+1=1..4096 offset
95 * 111ooooo LLLLLLLL oooooooo ; backref L+7 90 * 111ooooo LLLLLLLL oooooooo ; backref L+8 octets, o+1=1..4096 offset
96 * 91 *
97 */ 92 */
98 93
99unsigned int 94unsigned int
100lzf_compress (const void *const in_data, unsigned int in_len, 95lzf_compress (const void *const in_data, unsigned int in_len,
… …
105 ) 100 )
106{ 101{
107#if !LZF_STATE_ARG 102#if !LZF_STATE_ARG
108 LZF_STATE htab; 103 LZF_STATE htab;
109#endif 104#endif
110 const u8 **hslot;
111 const u8 *ip = (const u8 *)in_data; 105 const u8 *ip = (const u8 *)in_data;
112 u8 *op = (u8 *)out_data; 106 u8 *op = (u8 *)out_data;
113 const u8 *in_end = ip + in_len; 107 const u8 *in_end = ip + in_len;
114 u8 *out_end = op + out_len; 108 u8 *out_end = op + out_len;
115 const u8 *ref; 109 const u8 *ref;
116 110
111 /* off requires a type wide enough to hold a general pointer difference.
112 * ISO C doesn't have that (size_t might not be enough and ptrdiff_t only
113 * works for differences within a single object). We also assume that
114 * no bit pattern traps. Since the only platform that is both non-POSIX
115 * and fails to support both assumptions is windows 64 bit, we make a
116 * special workaround for it.
117 */
118#if defined (_WIN32) && defined (_M_X64)
119 /* workaround for missing POSIX compliance */
120 #if __GNUC__
121 unsigned long long off;
122 #else
123 unsigned __int64 off;
124 #endif
125#else
126 unsigned long off;
127#endif
117 unsigned int hval; 128 unsigned int hval;
118 unsigned long off;
119 int lit; 129 int lit;
120 130
121 if (!in_len || !out_len) 131 if (!in_len || !out_len)
122 return 0; 132 return 0;
123 133
124#if INIT_HTAB 134#if INIT_HTAB
125 memset (htab, 0, sizeof (htab)); 135 memset (htab, 0, sizeof (LZF_STATE));
126# if 0
127 for (hslot = htab; hslot < htab + HSIZE; hslot++)
128 *hslot++ = ip;
129# endif
130#endif 136#endif
131 137
132 lit = 0; op++; /* start run */ 138 lit = 0; op++; /* start run */
133 139
134 hval = FRST (ip); 140 hval = FRST (ip);
135 while (ip < in_end - 2) 141 while (ip < in_end - 2)
136 { 142 {
143 LZF_HSLOT *hslot;
144
137 hval = NEXT (hval, ip); 145 hval = NEXT (hval, ip);
138 hslot = htab + IDX (hval); 146 hslot = htab + IDX (hval);
139 ref = *hslot; *hslot = ip; 147 ref = *hslot + LZF_HSLOT_BIAS; *hslot = ip - LZF_HSLOT_BIAS;
140 148
141 if (1 149 if (1
142#if INIT_HTAB && !USE_MEMCPY 150#if INIT_HTAB
143 && ref < ip /* the next test will actually take care of this, but this is faster */ 151 && ref < ip /* the next test will actually take care of this, but this is faster */
144#endif 152#endif
145 && (off = ip - ref - 1) < MAX_OFF 153 && (off = ip - ref - 1) < LZF_MAX_OFF
146 && ip + 4 < in_end
147 && ref > (u8 *)in_data 154 && ref > (u8 *)in_data
155 && ref[2] == ip[2]
148#if STRICT_ALIGN 156#if STRICT_ALIGN
149 && ref[0] == ip[0] 157 && ((ref[1] << 8) | ref[0]) == ((ip[1] << 8) | ip[0])
150 && ref[1] == ip[1]
151 && ref[2] == ip[2]
152#else 158#else
153 && *(u16 *)ref == *(u16 *)ip 159 && *(u16 *)ref == *(u16 *)ip
154 && ref[2] == ip[2]
155#endif 160#endif
156 ) 161 )
157 { 162 {
158 /* match found at *ref++ */ 163 /* match found at *ref++ */
159 unsigned int len = 2; 164 unsigned int len = 2;
160 unsigned int maxlen = in_end - ip - len; 165 unsigned int maxlen = in_end - ip - len;
161 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 166 maxlen = maxlen > LZF_MAX_REF ? LZF_MAX_REF : maxlen;
162 167
163 if (expect_false (op + 1 + 3 >= out_end)) 168 if (expect_false (op + 3 + 1 >= out_end)) /* first a faster conservative test */
169 if (op - !lit + 3 + 1 >= out_end) /* second the exact but rare test */
164 return 0; 170 return 0;
165 171
166 op [- lit - 1] = lit - 1; /* stop run */ 172 op [- lit - 1] = lit - 1; /* stop run */
167 op -= !lit; /* undo run if length is zero */ 173 op -= !lit; /* undo run if length is zero */
168 174
169 for (;;) 175 for (;;)
… …
196 while (len < maxlen && ref[len] == ip[len]); 202 while (len < maxlen && ref[len] == ip[len]);
197 203
198 break; 204 break;
199 } 205 }
200 206
201 len -= 2; 207 len -= 2; /* len is now #octets - 1 */
202 ip++; 208 ip++;
203 209
204 if (len < 7) 210 if (len < 7)
205 { 211 {
206 *op++ = (off >> 8) + (len << 5); 212 *op++ = (off >> 8) + (len << 5);
… …
211 *op++ = len - 7; 217 *op++ = len - 7;
212 } 218 }
213 219
214 *op++ = off; 220 *op++ = off;
215 221
222 lit = 0; op++; /* start run */
223
224 ip += len + 1;
225
226 if (expect_false (ip >= in_end - 2))
227 break;
228
216#if ULTRA_FAST || VERY_FAST 229#if ULTRA_FAST || VERY_FAST
217 ip += len;
218#if VERY_FAST && !ULTRA_FAST
219 --ip; 230 --ip;
231# if VERY_FAST && !ULTRA_FAST
232 --ip;
220#endif 233# endif
221 hval = FRST (ip); 234 hval = FRST (ip);
222 235
223 hval = NEXT (hval, ip); 236 hval = NEXT (hval, ip);
224 htab[IDX (hval)] = ip; 237 htab[IDX (hval)] = ip - LZF_HSLOT_BIAS;
225 ip++; 238 ip++;
226 239
227#if VERY_FAST && !ULTRA_FAST 240# if VERY_FAST && !ULTRA_FAST
228 hval = NEXT (hval, ip); 241 hval = NEXT (hval, ip);
229 htab[IDX (hval)] = ip; 242 htab[IDX (hval)] = ip - LZF_HSLOT_BIAS;
230 ip++; 243 ip++;
231#endif 244# endif
232#else 245#else
246 ip -= len + 1;
247
233 do 248 do
234 { 249 {
235 hval = NEXT (hval, ip); 250 hval = NEXT (hval, ip);
236 htab[IDX (hval)] = ip; 251 htab[IDX (hval)] = ip - LZF_HSLOT_BIAS;
237 ip++; 252 ip++;
238 } 253 }
239 while (len--); 254 while (len--);
240#endif 255#endif
241 lit = 0; op++; /* start run */
242 continue;
243 } 256 }
244 257 else
258 {
245 /* one more literal byte we must copy */ 259 /* one more literal byte we must copy */
246
247 if (expect_false (op >= out_end)) 260 if (expect_false (op >= out_end))
248 return 0; 261 return 0;
249 262
250 lit++; 263 lit++; *op++ = *ip++;
264
265 if (expect_false (lit == LZF_MAX_LIT))
266 {
267 op [- lit - 1] = lit - 1; /* stop run */
268 lit = 0; op++; /* start run */
269 }
270 }
271 }
272
273 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
274 return 0;
275
276 while (ip < in_end)
277 {
251 *op++ = *ip++; 278 lit++; *op++ = *ip++;
252 279
253 if (expect_false (lit == MAX_LIT)) 280 if (expect_false (lit == LZF_MAX_LIT))
254 { 281 {
255 op [- lit - 1] = lit - 1; /* stop run */ 282 op [- lit - 1] = lit - 1; /* stop run */
256 lit = 0; op++; /* start run */ 283 lit = 0; op++; /* start run */
257 } 284 }
258 } 285 }
259 286
260 if (op + lit + 2 >= out_end)
261 return 0;
262
263 while (ip < in_end)
264 {
265 lit++;
266 *op++ = *ip++;
267 }
268
269 op [- lit - 1] = lit - 1; 287 op [- lit - 1] = lit - 1; /* end run */
288 op -= !lit; /* undo run if length is zero */
270 289
271 return op - (u8 *)out_data; 290 return op - (u8 *)out_data;
272} 291}
273 292

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines