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 1.18 by root, Tue Nov 13 08:17:38 2007 UTC vs.
Revision 1.28 by root, Fri Nov 30 10:34:26 2007 UTC

… …
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 unsigned int hval;
119 unsigned long off; 117 unsigned long off;
120 int lit = 0; 118 int lit;
119
120 if (!in_len || !out_len)
121 return 0;
121 122
122#if INIT_HTAB 123#if INIT_HTAB
123 memset (htab, 0, sizeof (htab)); 124 memset (htab, 0, sizeof (htab));
124# if 0 125# if 0
125 for (hslot = htab; hslot < htab + HSIZE; hslot++) 126 for (hslot = htab; hslot < htab + HSIZE; hslot++)
126 *hslot++ = ip; 127 *hslot++ = ip;
127# endif 128# endif
128#endif 129#endif
129 130
130 for (;;) 131 lit = 0; op++; /* start run */
132
133 hval = FRST (ip);
134 while (ip < in_end - 2)
131 { 135 {
132 if (expect_true (ip < in_end - 2)) 136 hval = NEXT (hval, ip);
137 hslot = htab + IDX (hval);
138 ref = *hslot; *hslot = ip;
139
140 if (1
141#if INIT_HTAB
142 && ref < ip /* the next test will actually take care of this, but this is faster */
143#endif
144 && (off = ip - ref - 1) < MAX_OFF
145 && ip + 4 < in_end
146 && ref > (u8 *)in_data
147#if STRICT_ALIGN
148 && ref[0] == ip[0]
149 && ref[1] == ip[1]
150 && ref[2] == ip[2]
151#else
152 && *(u16 *)ref == *(u16 *)ip
153 && ref[2] == ip[2]
154#endif
155 )
133 { 156 {
157 /* match found at *ref++ */
158 unsigned int len = 2;
159 unsigned int maxlen = in_end - ip - len;
160 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
161
162 if (maxlen > 8) maxlen = 8;//D
163
164 op [- lit - 1] = lit - 1; /* stop run */
165 op -= !lit; /* undo run if length is zero */
166
167 if (expect_false (op + 3 + 1 >= out_end))
168 return 0;
169
170 for (;;)
171 {
172 if (expect_true (maxlen > 16))
173 {
174 len++; if (ref [len] != ip [len]) break;
175 len++; if (ref [len] != ip [len]) break;
176 len++; if (ref [len] != ip [len]) break;
177 len++; if (ref [len] != ip [len]) break;
178
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
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
195 do
196 len++;
197 while (len < maxlen && ref[len] == ip[len]);
198
199 break;
200 }
201
202 len -= 2;
203 ip++;
204
205 if (len < 7)
206 {
207 *op++ = (off >> 8) + (len << 5);
208 }
209 else
210 {
211 *op++ = (off >> 8) + ( 7 << 5);
212 *op++ = len - 7;
213 }
214
215 *op++ = off;
216
217#if ULTRA_FAST || VERY_FAST
218 ip += len;
219#if VERY_FAST && !ULTRA_FAST
220 --ip;
221#endif
222 hval = FRST (ip);
223
134 hval = NEXT (hval, ip); 224 hval = NEXT (hval, ip);
135 hslot = htab + IDX (hval); 225 htab[IDX (hval)] = ip;
136 ref = *hslot; *hslot = ip; 226 ip++;
137 227
138 if (1 228#if VERY_FAST && !ULTRA_FAST
139#if INIT_HTAB && !USE_MEMCPY 229 hval = NEXT (hval, ip);
140 && ref < ip /* the next test will actually take care of this, but this is faster */ 230 htab[IDX (hval)] = ip;
231 ip++;
141#endif 232#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 233#else
150 && *(u16 *)ref == *(u16 *)ip 234 do
151 && ref[2] == ip[2]
152#endif
153 ) 235 {
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 (lit)
164 {
165 *op++ = lit - 1;
166 lit = -lit;
167 do
168 *op++ = ip[lit];
169 while (++lit);
170 }
171
172 for (;;)
173 {
174 if (expect_true (ip < in_end - 2 - 8 && maxlen > 8))
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 }
185
186 do
187 len++;
188 while (len < maxlen && ref[len] == ip[len]);
189
190 break;
191 }
192
193 len -= 2;
194 ip++;
195
196 if (len < 7)
197 {
198 *op++ = (off >> 8) + (len << 5);
199 }
200 else
201 {
202 *op++ = (off >> 8) + ( 7 << 5);
203 *op++ = len - 7;
204 }
205
206 *op++ = off;
207
208#if ULTRA_FAST || VERY_FAST
209 ip += len;
210#if VERY_FAST && !ULTRA_FAST
211 --ip;
212#endif
213 hval = FRST (ip);
214
215 hval = NEXT (hval, ip); 236 hval = NEXT (hval, ip);
216 htab[IDX (hval)] = ip; 237 htab[IDX (hval)] = ip;
217 ip++; 238 ip++;
218
219#if VERY_FAST && !ULTRA_FAST
220 hval = NEXT (hval, ip);
221 htab[IDX (hval)] = ip;
222 ip++;
223#endif
224#else
225 do
226 {
227 hval = NEXT (hval, ip);
228 htab[IDX (hval)] = ip;
229 ip++;
230 } 239 }
231 while (len--); 240 while (len--);
232#endif 241#endif
233 continue; 242
234 } 243 lit = 0; op++; /* start run */
235 } 244 }
236 else if (expect_false (ip == in_end)) 245 else
237 break; 246 {
238
239 /* one more literal byte we must copy */ 247 /* one more literal byte we must copy */
240 lit++; 248 if (expect_false (op >= out_end))
241 ip++; 249 return 0;
250
251 lit++; *op++ = *ip++;
252
253 if (expect_false (lit == MAX_LIT))
254 {
255 op [- lit - 1] = lit - 1; /* stop run */
256 lit = 0; op++; /* start run */
257 }
258 }
259 }
260
261 if (op + 2 > out_end) /* at most 2 bytes can be missing here */
262 return 0;
263
264 while (ip < in_end)
265 {
266 lit++; *op++ = *ip++;
242 267
243 if (expect_false (lit == MAX_LIT)) 268 if (expect_false (lit == MAX_LIT))
244 { 269 {
245 if (op + 1 + MAX_LIT >= out_end) 270 op [- lit - 1] = lit - 1; /* stop run */
246 return 0; 271 lit = 0; op++; /* start run */
247
248 *op++ = MAX_LIT - 1;
249
250#ifdef lzf_movsb
251 ip -= lit;
252 lzf_movsb (op, ip, lit);
253#else
254 lit = -lit;
255 do
256 *op++ = ip[lit];
257 while (++lit);
258#endif
259 } 272 }
260 } 273 }
261 274
262 if (lit) 275 op [- lit - 1] = lit - 1; /* end run */
263 { 276 op -= !lit; /* undo run if length is zero */
264 if (op + lit + 1 >= out_end)
265 return 0;
266 277
267 *op++ = lit - 1;
268#ifdef lzf_movsb
269 ip -= lit;
270 lzf_movsb (op, ip, lit);
271#else
272 lit = -lit;
273 do
274 *op++ = ip[lit];
275 while (++lit);
276#endif
277 }
278
279 return op - (u8 *) out_data; 278 return op - (u8 *)out_data;
280} 279}
281 280

Diff Legend

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