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 1.17 by root, Tue Nov 13 07:54:18 2007 UTC vs.
Revision 1.39 by root, Sun May 30 05:52:57 2010 UTC

1/* 1/*
2 * Copyright (c) 2000-2007 Marc Alexander Lehmann <schmorp@schmorp.de> 2 * Copyright (c) 2000-2010 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,
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 ULTRA_FAST
51# define IDX(h) ((( h >> (3*8 - HLOG)) - h ) & (HSIZE - 1))
52# elif VERY_FAST
53# define IDX(h) ((( h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1))
54# else
50# define IDX(h) ((((h ^ (h << 5)) >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) 55# define IDX(h) ((((h ^ (h << 5)) >> (3*8 - HLOG)) - h*5) & (HSIZE - 1))
51/*# define IDX(h) ((ip[0] * 121 ^ ip[1] * 33 ^ ip[2] * 1) & (HSIZE-1))*/ 56# endif
52#endif 57#endif
53/* 58/*
54 * IDX works because it is very similar to a multiplicative hash, e.g. 59 * IDX works because it is very similar to a multiplicative hash, e.g.
55 * ((h * 57321 >> (3*8 - HLOG)) & (HSIZE - 1)) 60 * ((h * 57321 >> (3*8 - HLOG)) & (HSIZE - 1))
56 * the latter is also quite fast on newer CPUs, and compresses similarly. 61 * the latter is also quite fast on newer CPUs, and compresses similarly.
68 73
69#define MAX_LIT (1 << 5) 74#define MAX_LIT (1 << 5)
70#define MAX_OFF (1 << 13) 75#define MAX_OFF (1 << 13)
71#define MAX_REF ((1 << 8) + (1 << 3)) 76#define MAX_REF ((1 << 8) + (1 << 3))
72 77
73#if (__i386 || __amd64) && __GNUC__ >= 3
74# define lzf_movsb(dst, src, len) \
75 asm ("rep movsb" \
76 : "=D" (dst), "=S" (src), "=c" (len) \
77 : "0" (dst), "1" (src), "2" (len));
78#endif
79
80#if __GNUC__ >= 3 78#if __GNUC__ >= 3
81# define expect(expr,value) __builtin_expect ((expr),(value)) 79# define expect(expr,value) __builtin_expect ((expr),(value))
82# define inline inline 80# define inline inline
83#else 81#else
84# define expect(expr,value) (expr) 82# define expect(expr,value) (expr)
89#define expect_true(expr) expect ((expr) != 0, 1) 87#define expect_true(expr) expect ((expr) != 0, 1)
90 88
91/* 89/*
92 * compressed format 90 * compressed format
93 * 91 *
94 * 000LLLLL <L+1> ; literal 92 * 000LLLLL <L+1> ; literal, L+1=1..33 octets
95 * LLLooooo oooooooo ; backref L 93 * LLLooooo oooooooo ; backref L+1=1..7 octets, o+1=1..4096 offset
96 * 111ooooo LLLLLLLL oooooooo ; backref L+7 94 * 111ooooo LLLLLLLL oooooooo ; backref L+8 octets, o+1=1..4096 offset
97 * 95 *
98 */ 96 */
99 97
100unsigned int 98unsigned int
101lzf_compress (const void *const in_data, unsigned int in_len, 99lzf_compress (const void *const in_data, unsigned int in_len,
113 u8 *op = (u8 *)out_data; 111 u8 *op = (u8 *)out_data;
114 const u8 *in_end = ip + in_len; 112 const u8 *in_end = ip + in_len;
115 u8 *out_end = op + out_len; 113 u8 *out_end = op + out_len;
116 const u8 *ref; 114 const u8 *ref;
117 115
118 unsigned int hval = FRST (ip); 116 /* off requires a type wide enough to hold a general pointer difference.
117 * ISO C doesn't have that (size_t might not be enough and ptrdiff_t only
118 * works for differences within a single object). We also assume that no
119 * no bit pattern traps. Since the only platform that is both non-POSIX
120 * and fails to support both assumptions is windows 64 bit, we make a
121 * special workaround for it.
122 */
123#if defined (WIN32) && defined (_M_X64)
124 unsigned _int64 off; /* workaround for missing POSIX compliance */
125#else
119 unsigned long off; 126 unsigned long off;
120 int lit = 0; 127#endif
128 unsigned int hval;
129 int lit;
130
131 if (!in_len || !out_len)
132 return 0;
121 133
122#if INIT_HTAB 134#if INIT_HTAB
123 memset (htab, 0, sizeof (htab)); 135 memset (htab, 0, sizeof (htab));
124# if 0 136# if 0
125 for (hslot = htab; hslot < htab + HSIZE; hslot++) 137 for (hslot = htab; hslot < htab + HSIZE; hslot++)
126 *hslot++ = ip; 138 *hslot++ = ip;
127# endif 139# endif
128#endif 140#endif
129 141
130 for (;;) 142 lit = 0; op++; /* start run */
143
144 hval = FRST (ip);
145 while (ip < in_end - 2)
131 { 146 {
132 if (expect_true (ip < in_end - 2)) 147 hval = NEXT (hval, ip);
148 hslot = htab + IDX (hval);
149 ref = *hslot; *hslot = ip;
150
151 if (1
152#if INIT_HTAB
153 && ref < ip /* the next test will actually take care of this, but this is faster */
154#endif
155 && (off = ip - ref - 1) < MAX_OFF
156 && ip + 4 < in_end
157 && ref > (u8 *)in_data
158#if STRICT_ALIGN
159 && ref[0] == ip[0]
160 && ref[1] == ip[1]
161 && ref[2] == ip[2]
162#else
163 && *(u16 *)ref == *(u16 *)ip
164 && ref[2] == ip[2]
165#endif
166 )
133 { 167 {
134 hval = NEXT (hval, ip);
135 hslot = htab + IDX (hval);
136 ref = *hslot; *hslot = ip;
137
138 if (1
139#if INIT_HTAB && !USE_MEMCPY
140 && ref < ip /* the next test will actually take care of this, but this is faster */
141#endif
142 && (off = ip - ref - 1) < MAX_OFF
143 && ip + 4 < in_end
144 && ref > (u8 *)in_data
145#if STRICT_ALIGN
146 && ref[0] == ip[0]
147 && ref[1] == ip[1]
148 && ref[2] == ip[2]
149#else
150 && *(u16 *)ref == *(u16 *)ip
151 && ref[2] == ip[2]
152#endif
153 )
154 {
155 /* match found at *ref++ */ 168 /* match found at *ref++ */
156 unsigned int len = 2; 169 unsigned int len = 2;
157 unsigned int maxlen = in_end - ip - len; 170 unsigned int maxlen = in_end - ip - len;
158 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 171 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
159 172
160 if (expect_false (op + lit + 1 + 3 >= out_end)) 173 if (expect_false (op + 3 + 1 >= out_end)) /* first a faster conservative test */
174 if (op - !lit + 3 + 1 >= out_end) /* second the exact but rare test */
161 return 0; 175 return 0;
176
177 op [- lit - 1] = lit - 1; /* stop run */
178 op -= !lit; /* undo run if length is zero */
179
180 for (;;)
181 {
182 if (expect_true (maxlen > 16))
183 {
184 len++; if (ref [len] != ip [len]) break;
185 len++; if (ref [len] != ip [len]) break;
186 len++; if (ref [len] != ip [len]) break;
187 len++; if (ref [len] != ip [len]) break;
188
189 len++; if (ref [len] != ip [len]) break;
190 len++; if (ref [len] != ip [len]) break;
191 len++; if (ref [len] != ip [len]) break;
192 len++; if (ref [len] != ip [len]) break;
193
194 len++; if (ref [len] != ip [len]) break;
195 len++; if (ref [len] != ip [len]) break;
196 len++; if (ref [len] != ip [len]) break;
197 len++; if (ref [len] != ip [len]) break;
198
199 len++; if (ref [len] != ip [len]) break;
200 len++; if (ref [len] != ip [len]) break;
201 len++; if (ref [len] != ip [len]) break;
202 len++; if (ref [len] != ip [len]) break;
203 }
162 204
163 do 205 do
164 len++; 206 len++;
165 while (len < maxlen && ref[len] == ip[len]); 207 while (len < maxlen && ref[len] == ip[len]);
166 208
167 if (lit)
168 { 209 break;
169 *op++ = lit - 1;
170 lit = -lit;
171 do
172 *op++ = ip[lit];
173 while (++lit);
174 } 210 }
175 211
176 len -= 2; 212 len -= 2; /* len is now #octets - 1 */
177 ip++; 213 ip++;
178 214
179 if (len < 7) 215 if (len < 7)
180 { 216 {
181 *op++ = (off >> 8) + (len << 5); 217 *op++ = (off >> 8) + (len << 5);
182 } 218 }
183 else 219 else
184 { 220 {
185 *op++ = (off >> 8) + ( 7 << 5); 221 *op++ = (off >> 8) + ( 7 << 5);
186 *op++ = len - 7; 222 *op++ = len - 7;
187 } 223 }
188 224
189 *op++ = off; 225 *op++ = off;
226 lit = 0; op++; /* start run */
227
228 ip += len + 1;
229
230 if (expect_false (ip >= in_end - 2))
231 break;
190 232
191#if ULTRA_FAST || VERY_FAST 233#if ULTRA_FAST || VERY_FAST
192 ip += len; 234 --ip;
193#if VERY_FAST && !ULTRA_FAST 235# if VERY_FAST && !ULTRA_FAST
194 --ip; 236 --ip;
195#endif 237# endif
196 hval = FRST (ip); 238 hval = FRST (ip);
197 239
240 hval = NEXT (hval, ip);
241 htab[IDX (hval)] = ip;
242 ip++;
243
244# if VERY_FAST && !ULTRA_FAST
245 hval = NEXT (hval, ip);
246 htab[IDX (hval)] = ip;
247 ip++;
248# endif
249#else
250 ip -= len + 1;
251
252 do
253 {
198 hval = NEXT (hval, ip); 254 hval = NEXT (hval, ip);
199 htab[IDX (hval)] = ip; 255 htab[IDX (hval)] = ip;
200 ip++; 256 ip++;
201
202#if VERY_FAST && !ULTRA_FAST
203 hval = NEXT (hval, ip);
204 htab[IDX (hval)] = ip;
205 ip++;
206#endif
207#else
208 do
209 {
210 hval = NEXT (hval, ip);
211 htab[IDX (hval)] = ip;
212 ip++;
213 } 257 }
214 while (len--); 258 while (len--);
215#endif 259#endif
216 continue;
217 }
218 } 260 }
219 else if (expect_false (ip == in_end)) 261 else
220 break; 262 {
221
222 /* one more literal byte we must copy */ 263 /* one more literal byte we must copy */
223 lit++; 264 if (expect_false (op >= out_end))
224 ip++; 265 return 0;
266
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 {
282 lit++; *op++ = *ip++;
225 283
226 if (expect_false (lit == MAX_LIT)) 284 if (expect_false (lit == MAX_LIT))
227 { 285 {
228 if (op + 1 + MAX_LIT >= out_end) 286 op [- lit - 1] = lit - 1; /* stop run */
229 return 0; 287 lit = 0; op++; /* start run */
230
231 *op++ = MAX_LIT - 1;
232
233#ifdef lzf_movsb
234 ip -= lit;
235 lzf_movsb (op, ip, lit);
236#else
237 lit = -lit;
238 do
239 *op++ = ip[lit];
240 while (++lit);
241#endif
242 } 288 }
243 } 289 }
244 290
245 if (lit) 291 op [- lit - 1] = lit - 1; /* end run */
246 { 292 op -= !lit; /* undo run if length is zero */
247 if (op + lit + 1 >= out_end)
248 return 0;
249 293
250 *op++ = lit - 1;
251#ifdef lzf_movsb
252 ip -= lit;
253 lzf_movsb (op, ip, lit);
254#else
255 lit = -lit;
256 do
257 *op++ = ip[lit];
258 while (++lit);
259#endif
260 }
261
262 return op - (u8 *) out_data; 294 return op - (u8 *)out_data;
263} 295}
264 296

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines