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 1.16 by root, Mon Nov 5 23:28:40 2007 UTC vs.
Revision 1.38 by root, Sun May 30 05:46:56 2010 UTC

1/* 1/*
2 * Copyright (c) 2000-2007 Marc Alexander Lehmann <schmorp@schmorp.de> 2 * Copyright (c) 2000-2009 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 78#if __GNUC__ >= 3
74# define lzf_movsb(dst, src, len) \ 79# define expect(expr,value) __builtin_expect ((expr),(value))
75 asm ("rep movsb" \ 80# define inline inline
76 : "=D" (dst), "=S" (src), "=c" (len) \ 81#else
77 : "0" (dst), "1" (src), "2" (len)); 82# define expect(expr,value) (expr)
83# define inline static
78#endif 84#endif
85
86#define expect_false(expr) expect ((expr) != 0, 0)
87#define expect_true(expr) expect ((expr) != 0, 1)
79 88
80/* 89/*
81 * compressed format 90 * compressed format
82 * 91 *
83 * 000LLLLL <L+1> ; literal 92 * 000LLLLL <L+1> ; literal, L+1=1..33 octets
84 * LLLooooo oooooooo ; backref L 93 * LLLooooo oooooooo ; backref L+1=1..7 octets, o+1=1..4096 offset
85 * 111ooooo LLLLLLLL oooooooo ; backref L+7 94 * 111ooooo LLLLLLLL oooooooo ; backref L+8 octets, o+1=1..4096 offset
86 * 95 *
87 */ 96 */
88 97
89unsigned int 98unsigned int
90lzf_compress (const void *const in_data, unsigned int in_len, 99lzf_compress (const void *const in_data, unsigned int in_len,
102 u8 *op = (u8 *)out_data; 111 u8 *op = (u8 *)out_data;
103 const u8 *in_end = ip + in_len; 112 const u8 *in_end = ip + in_len;
104 u8 *out_end = op + out_len; 113 u8 *out_end = op + out_len;
105 const u8 *ref; 114 const u8 *ref;
106 115
107 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
108 unsigned long off; 126 unsigned long off;
109 int lit = 0; 127#endif
128 unsigned int hval;
129 int lit;
130
131 if (!in_len || !out_len)
132 return 0;
110 133
111#if INIT_HTAB 134#if INIT_HTAB
112 memset (htab, 0, sizeof (htab)); 135 memset (htab, 0, sizeof (htab));
113# if 0 136# if 0
114 for (hslot = htab; hslot < htab + HSIZE; hslot++) 137 for (hslot = htab; hslot < htab + HSIZE; hslot++)
115 *hslot++ = ip; 138 *hslot++ = ip;
116# endif 139# endif
117#endif 140#endif
118 141
119 for (;;) 142 lit = 0; op++; /* start run */
143
144 hval = FRST (ip);
145 while (ip < in_end - 2)
120 { 146 {
121 if (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 )
122 { 167 {
123 hval = NEXT (hval, ip);
124 hslot = htab + IDX (hval);
125 ref = *hslot; *hslot = ip;
126
127 if (1
128#if INIT_HTAB && !USE_MEMCPY
129 && ref < ip /* the next test will actually take care of this, but this is faster */
130#endif
131 && (off = ip - ref - 1) < MAX_OFF
132 && ip + 4 < in_end
133 && ref > (u8 *)in_data
134#if STRICT_ALIGN
135 && ref[0] == ip[0]
136 && ref[1] == ip[1]
137 && ref[2] == ip[2]
138#else
139 && *(u16 *)ref == *(u16 *)ip
140 && ref[2] == ip[2]
141#endif
142 )
143 {
144 /* match found at *ref++ */ 168 /* match found at *ref++ */
145 unsigned int len = 2; 169 unsigned int len = 2;
146 unsigned int maxlen = in_end - ip - len; 170 unsigned int maxlen = in_end - ip - len;
147 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 171 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
148 172
149 if (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 */
150 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 }
151 204
152 do 205 do
153 len++; 206 len++;
154 while (len < maxlen && ref[len] == ip[len]); 207 while (len < maxlen && ref[len] == ip[len]);
155 208
156 if (lit)
157 { 209 break;
158 *op++ = lit - 1;
159 lit = -lit;
160 do
161 *op++ = ip[lit];
162 while (++lit);
163 } 210 }
164 211
165 len -= 2; 212 len -= 2; /* len is now #octets - 1 */
166 ip++; 213 ip++;
167 214
168 if (len < 7) 215 if (len < 7)
169 { 216 {
170 *op++ = (off >> 8) + (len << 5); 217 *op++ = (off >> 8) + (len << 5);
171 } 218 }
172 else 219 else
173 { 220 {
174 *op++ = (off >> 8) + ( 7 << 5); 221 *op++ = (off >> 8) + ( 7 << 5);
175 *op++ = len - 7; 222 *op++ = len - 7;
176 } 223 }
177 224
178 *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;
179 232
180#if ULTRA_FAST || VERY_FAST 233#if ULTRA_FAST || VERY_FAST
181 ip += len; 234 --ip;
182#if VERY_FAST && !ULTRA_FAST 235# if VERY_FAST && !ULTRA_FAST
183 --ip; 236 --ip;
184#endif 237# endif
185 hval = FRST (ip); 238 hval = FRST (ip);
186 239
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 {
187 hval = NEXT (hval, ip); 254 hval = NEXT (hval, ip);
188 htab[IDX (hval)] = ip; 255 htab[IDX (hval)] = ip;
189 ip++; 256 ip++;
190
191#if VERY_FAST && !ULTRA_FAST
192 hval = NEXT (hval, ip);
193 htab[IDX (hval)] = ip;
194 ip++;
195#endif
196#else
197 do
198 {
199 hval = NEXT (hval, ip);
200 htab[IDX (hval)] = ip;
201 ip++;
202 } 257 }
203 while (len--); 258 while (len--);
204#endif 259#endif
205 continue;
206 }
207 } 260 }
208 else if (ip == in_end) 261 else
209 break;
210
211 /* one more literal byte we must copy */
212 lit++;
213 ip++;
214
215 if (lit == MAX_LIT)
216 { 262 {
217 if (op + 1 + MAX_LIT >= out_end) 263 /* one more literal byte we must copy */
264 if (expect_false (op >= out_end))
218 return 0; 265 return 0;
219 266
220 *op++ = MAX_LIT - 1; 267 lit++; *op++ = *ip++;
221 268
222#ifdef lzf_movsb 269 if (expect_false (lit == MAX_LIT))
223 ip -= lit;
224 lzf_movsb (op, ip, lit);
225#else
226 lit = -lit;
227 do 270 {
228 *op++ = ip[lit]; 271 op [- lit - 1] = lit - 1; /* stop run */
229 while (++lit); 272 lit = 0; op++; /* start run */
230#endif 273 }
231 } 274 }
232 } 275 }
233 276
234 if (lit) 277 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
278 return 0;
279
280 while (ip < in_end)
235 { 281 {
236 if (op + lit + 1 >= out_end) 282 lit++; *op++ = *ip++;
237 return 0;
238 283
239 *op++ = lit - 1; 284 if (expect_false (lit == MAX_LIT))
240#ifdef lzf_movsb 285 {
241 ip -= lit; 286 op [- lit - 1] = lit - 1; /* stop run */
242 lzf_movsb (op, ip, lit); 287 lit = 0; op++; /* start run */
243#else 288 }
244 lit = -lit;
245 do
246 *op++ = ip[lit];
247 while (++lit);
248#endif
249 } 289 }
250 290
291 op [- lit - 1] = lit - 1; /* end run */
292 op -= !lit; /* undo run if length is zero */
293
251 return op - (u8 *) out_data; 294 return op - (u8 *)out_data;
252} 295}
253 296

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines