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 1.48 by root, Sat Jun 30 21:39:41 2012 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,
… …
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]
… …
88#define expect_true(expr) expect ((expr) != 0, 1) 87#define expect_true(expr) expect ((expr) != 0, 1)
89 88
90/* 89/*
91 * compressed format 90 * compressed format
92 * 91 *
93 * 000LLLLL <L+1> ; literal 92 * 000LLLLL <L+1> ; literal, L+1=1..33 octets
94 * LLLooooo oooooooo ; backref L 93 * LLLooooo oooooooo ; backref L+1=1..7 octets, o+1=1..4096 offset
95 * 111ooooo LLLLLLLL oooooooo ; backref L+7 94 * 111ooooo LLLLLLLL oooooooo ; backref L+8 octets, o+1=1..4096 offset
96 * 95 *
97 */ 96 */
98 97
99unsigned int 98unsigned int
100lzf_compress (const void *const in_data, unsigned int in_len, 99lzf_compress (const void *const in_data, unsigned int in_len,
… …
105 ) 104 )
106{ 105{
107#if !LZF_STATE_ARG 106#if !LZF_STATE_ARG
108 LZF_STATE htab; 107 LZF_STATE htab;
109#endif 108#endif
110 const u8 **hslot;
111 const u8 *ip = (const u8 *)in_data; 109 const u8 *ip = (const u8 *)in_data;
112 u8 *op = (u8 *)out_data; 110 u8 *op = (u8 *)out_data;
113 const u8 *in_end = ip + in_len; 111 const u8 *in_end = ip + in_len;
114 u8 *out_end = op + out_len; 112 u8 *out_end = op + out_len;
115 const u8 *ref; 113 const u8 *ref;
116 114
115 /* off requires a type wide enough to hold a general pointer difference.
116 * ISO C doesn't have that (size_t might not be enough and ptrdiff_t only
117 * works for differences within a single object). We also assume that
118 * no bit pattern traps. Since the only platform that is both non-POSIX
119 * and fails to support both assumptions is windows 64 bit, we make a
120 * special workaround for it.
121 */
122#if defined (_WIN32) && defined (_M_X64)
123 /* workaround for missing POSIX compliance */
124 #if __GNUC__
125 unsigned long long off;
126 #else
127 unsigned __int64 off;
128 #endif
129#else
130 unsigned long off;
131#endif
117 unsigned int hval; 132 unsigned int hval;
118 unsigned long off;
119 int lit; 133 int lit;
120 134
121 if (!in_len || !out_len) 135 if (!in_len || !out_len)
122 return 0; 136 return 0;
123 137
124#if INIT_HTAB 138#if INIT_HTAB
125 memset (htab, 0, sizeof (htab)); 139 memset (htab, 0, sizeof (htab));
126# if 0
127 for (hslot = htab; hslot < htab + HSIZE; hslot++)
128 *hslot++ = ip;
129# endif
130#endif 140#endif
131 141
132 lit = 0; op++; /* start run */ 142 lit = 0; op++; /* start run */
133 143
134 hval = FRST (ip); 144 hval = FRST (ip);
135 while (ip < in_end - 2) 145 while (ip < in_end - 2)
136 { 146 {
147 LZF_HSLOT *hslot;
148
137 hval = NEXT (hval, ip); 149 hval = NEXT (hval, ip);
138 hslot = htab + IDX (hval); 150 hslot = htab + IDX (hval);
139 ref = *hslot; *hslot = ip; 151 ref = *hslot + LZF_HSLOT_BIAS; *hslot = ip - LZF_HSLOT_BIAS;
140 152
141 if (1 153 if (1
142#if INIT_HTAB && !USE_MEMCPY 154#if INIT_HTAB
143 && ref < ip /* the next test will actually take care of this, but this is faster */ 155 && ref < ip /* the next test will actually take care of this, but this is faster */
144#endif 156#endif
145 && (off = ip - ref - 1) < MAX_OFF 157 && (off = ip - ref - 1) < MAX_OFF
146 && ip + 4 < in_end
147 && ref > (u8 *)in_data 158 && ref > (u8 *)in_data
159 && ref[2] == ip[2]
148#if STRICT_ALIGN 160#if STRICT_ALIGN
149 && ref[0] == ip[0] 161 && ((ref[1] << 8) | ref[0]) == ((ip[1] << 8) | ip[0])
150 && ref[1] == ip[1]
151 && ref[2] == ip[2]
152#else 162#else
153 && *(u16 *)ref == *(u16 *)ip 163 && *(u16 *)ref == *(u16 *)ip
154 && ref[2] == ip[2]
155#endif 164#endif
156 ) 165 )
157 { 166 {
158 /* match found at *ref++ */ 167 /* match found at *ref++ */
159 unsigned int len = 2; 168 unsigned int len = 2;
160 unsigned int maxlen = in_end - ip - len; 169 unsigned int maxlen = in_end - ip - len;
161 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 170 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
162 171
163 if (expect_false (op + 1 + 3 >= out_end)) 172 if (expect_false (op + 3 + 1 >= out_end)) /* first a faster conservative test */
173 if (op - !lit + 3 + 1 >= out_end) /* second the exact but rare test */
164 return 0; 174 return 0;
165 175
166 op [- lit - 1] = lit - 1; /* stop run */ 176 op [- lit - 1] = lit - 1; /* stop run */
167 op -= !lit; /* undo run if length is zero */ 177 op -= !lit; /* undo run if length is zero */
168 178
169 for (;;) 179 for (;;)
… …
196 while (len < maxlen && ref[len] == ip[len]); 206 while (len < maxlen && ref[len] == ip[len]);
197 207
198 break; 208 break;
199 } 209 }
200 210
201 len -= 2; 211 len -= 2; /* len is now #octets - 1 */
202 ip++; 212 ip++;
203 213
204 if (len < 7) 214 if (len < 7)
205 { 215 {
206 *op++ = (off >> 8) + (len << 5); 216 *op++ = (off >> 8) + (len << 5);
… …
211 *op++ = len - 7; 221 *op++ = len - 7;
212 } 222 }
213 223
214 *op++ = off; 224 *op++ = off;
215 225
226 lit = 0; op++; /* start run */
227
228 ip += len + 1;
229
230 if (expect_false (ip >= in_end - 2))
231 break;
232
216#if ULTRA_FAST || VERY_FAST 233#if ULTRA_FAST || VERY_FAST
217 ip += len;
218#if VERY_FAST && !ULTRA_FAST
219 --ip; 234 --ip;
235# if VERY_FAST && !ULTRA_FAST
236 --ip;
220#endif 237# endif
221 hval = FRST (ip); 238 hval = FRST (ip);
222 239
223 hval = NEXT (hval, ip); 240 hval = NEXT (hval, ip);
224 htab[IDX (hval)] = ip; 241 htab[IDX (hval)] = ip - LZF_HSLOT_BIAS;
225 ip++; 242 ip++;
226 243
227#if VERY_FAST && !ULTRA_FAST 244# if VERY_FAST && !ULTRA_FAST
228 hval = NEXT (hval, ip); 245 hval = NEXT (hval, ip);
229 htab[IDX (hval)] = ip; 246 htab[IDX (hval)] = ip - LZF_HSLOT_BIAS;
230 ip++; 247 ip++;
231#endif 248# endif
232#else 249#else
250 ip -= len + 1;
251
233 do 252 do
234 { 253 {
235 hval = NEXT (hval, ip); 254 hval = NEXT (hval, ip);
236 htab[IDX (hval)] = ip; 255 htab[IDX (hval)] = ip - LZF_HSLOT_BIAS;
237 ip++; 256 ip++;
238 } 257 }
239 while (len--); 258 while (len--);
240#endif 259#endif
241 lit = 0; op++; /* start run */
242 continue;
243 } 260 }
244 261 else
262 {
245 /* one more literal byte we must copy */ 263 /* one more literal byte we must copy */
246
247 if (expect_false (op >= out_end)) 264 if (expect_false (op >= out_end))
248 return 0; 265 return 0;
249 266
250 lit++; 267 lit++; *op++ = *ip++;
268
269 if (expect_false (lit == MAX_LIT))
270 {
271 op [- lit - 1] = lit - 1; /* stop run */
272 lit = 0; op++; /* start run */
273 }
274 }
275 }
276
277 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
278 return 0;
279
280 while (ip < in_end)
281 {
251 *op++ = *ip++; 282 lit++; *op++ = *ip++;
252 283
253 if (expect_false (lit == MAX_LIT)) 284 if (expect_false (lit == MAX_LIT))
254 { 285 {
255 op [- lit - 1] = lit - 1; /* stop run */ 286 op [- lit - 1] = lit - 1; /* stop run */
256 lit = 0; op++; /* start run */ 287 lit = 0; op++; /* start run */
257 } 288 }
258 } 289 }
259 290
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; 291 op [- lit - 1] = lit - 1; /* end run */
292 op -= !lit; /* undo run if length is zero */
270 293
271 return op - (u8 *)out_data; 294 return op - (u8 *)out_data;
272} 295}
273 296

Diff Legend

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