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 1.19 by root, Tue Nov 13 08:38:35 2007 UTC vs.
Revision 1.35 by root, Sat Aug 30 06:43:38 2008 UTC

1/* 1/*
2 * Copyright (c) 2000-2007 Marc Alexander Lehmann <schmorp@schmorp.de> 2 * Copyright (c) 2000-2008 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.
67#endif 72#endif
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
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 77
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
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 {
168 /* match found at *ref++ */
169 unsigned int len = 2;
170 unsigned int maxlen = in_end - ip - len;
171 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
172
173 op [- lit - 1] = lit - 1; /* stop run */
174 op -= !lit; /* undo run if length is zero */
175
176 if (expect_false (op + 3 + 1 >= out_end))
177 return 0;
178
179 for (;;)
180 {
181 if (expect_true (maxlen > 16))
182 {
183 len++; if (ref [len] != ip [len]) break;
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
188 len++; if (ref [len] != ip [len]) break;
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
193 len++; if (ref [len] != ip [len]) break;
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
198 len++; if (ref [len] != ip [len]) break;
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 }
203
204 do
205 len++;
206 while (len < maxlen && ref[len] == ip[len]);
207
208 break;
209 }
210
211 len -= 2; /* len is now #octets - 1 */
212 ip++;
213
214 if (len < 7)
215 {
216 *op++ = (off >> 8) + (len << 5);
217 }
218 else
219 {
220 *op++ = (off >> 8) + ( 7 << 5);
221 *op++ = len - 7;
222 }
223
224 *op++ = off;
225 lit = 0; op++; /* start run */
226
227 ip += len + 1;
228
229 if (expect_false (ip >= in_end - 2))
230 break;
231
232#if ULTRA_FAST || VERY_FAST
233 --ip;
234# if VERY_FAST && !ULTRA_FAST
235 --ip;
236# endif
237 hval = FRST (ip);
238
134 hval = NEXT (hval, ip); 239 hval = NEXT (hval, ip);
135 hslot = htab + IDX (hval); 240 htab[IDX (hval)] = ip;
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++ */
156 unsigned int len = 2;
157 unsigned int maxlen = in_end - ip - len;
158 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
159
160 if (expect_false (op + lit + 1 + 3 >= out_end))
161 return 0;
162
163 if (expect_false (lit))
164 {
165 *op++ = lit - 1;
166 lit = -lit;
167 do
168 *op++ = ip[lit];
169 while (expect_false (++lit));
170 }
171
172 for (;;)
173 {
174 if (expect_true (maxlen > 16))
175 {
176 len++; if (ref [len] != ip [len]) break;
177 len++; if (ref [len] != ip [len]) break;
178 len++; if (ref [len] != ip [len]) break;
179 len++; if (ref [len] != ip [len]) break;
180 len++; if (ref [len] != ip [len]) break;
181 len++; if (ref [len] != ip [len]) break;
182 len++; if (ref [len] != ip [len]) break;
183 len++; if (ref [len] != ip [len]) break;
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 len++; if (ref [len] != ip [len]) break;
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 }
193
194 do
195 len++;
196 while (len < maxlen && ref[len] == ip[len]);
197
198 break;
199 }
200
201 len -= 2;
202 ip++; 241 ip++;
203 242
204 if (len < 7)
205 {
206 *op++ = (off >> 8) + (len << 5);
207 }
208 else
209 {
210 *op++ = (off >> 8) + ( 7 << 5);
211 *op++ = len - 7;
212 }
213
214 *op++ = off;
215
216#if ULTRA_FAST || VERY_FAST
217 ip += len;
218#if VERY_FAST && !ULTRA_FAST 243# if VERY_FAST && !ULTRA_FAST
244 hval = NEXT (hval, ip);
245 htab[IDX (hval)] = ip;
219 --ip; 246 ip++;
220#endif 247# endif
221 hval = FRST (ip); 248#else
249 ip -= len + 1;
222 250
251 do
252 {
223 hval = NEXT (hval, ip); 253 hval = NEXT (hval, ip);
224 htab[IDX (hval)] = ip; 254 htab[IDX (hval)] = ip;
225 ip++; 255 ip++;
226
227#if VERY_FAST && !ULTRA_FAST
228 hval = NEXT (hval, ip);
229 htab[IDX (hval)] = ip;
230 ip++;
231#endif
232#else
233 do
234 {
235 hval = NEXT (hval, ip);
236 htab[IDX (hval)] = ip;
237 ip++;
238 } 256 }
239 while (len--); 257 while (len--);
240#endif 258#endif
241 continue;
242 }
243 } 259 }
244 else if (expect_false (ip == in_end)) 260 else
245 break; 261 {
246
247 /* one more literal byte we must copy */ 262 /* one more literal byte we must copy */
248 lit++; 263 if (expect_false (op >= out_end))
249 ip++; 264 return 0;
265
266 lit++; *op++ = *ip++;
267
268 if (expect_false (lit == MAX_LIT))
269 {
270 op [- lit - 1] = lit - 1; /* stop run */
271 lit = 0; op++; /* start run */
272 }
273 }
274 }
275
276 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
277 return 0;
278
279 while (ip < in_end)
280 {
281 lit++; *op++ = *ip++;
250 282
251 if (expect_false (lit == MAX_LIT)) 283 if (expect_false (lit == MAX_LIT))
252 { 284 {
253 if (op + 1 + MAX_LIT >= out_end) 285 op [- lit - 1] = lit - 1; /* stop run */
254 return 0; 286 lit = 0; op++; /* start run */
255
256 *op++ = MAX_LIT - 1;
257
258#ifdef lzf_movsb
259 ip -= MAX_LIT;
260 lzf_movsb (op, ip, lit);
261#else
262 lit = -lit;
263 do
264 *op++ = ip[lit];
265 while (++lit);
266#endif
267 } 287 }
268 } 288 }
269 289
270 if (lit) 290 op [- lit - 1] = lit - 1; /* end run */
271 { 291 op -= !lit; /* undo run if length is zero */
272 if (op + lit + 1 >= out_end)
273 return 0;
274 292
275 *op++ = lit - 1;
276#ifdef lzf_movsb
277 ip -= lit;
278 lzf_movsb (op, ip, lit);
279#else
280 lit = -lit;
281 do
282 *op++ = ip[lit];
283 while (++lit);
284#endif
285 }
286
287 return op - (u8 *) out_data; 293 return op - (u8 *)out_data;
288} 294}
289 295

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines