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 1.31 by root, Tue May 6 22:49:18 2008 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;
117#if WIN32
118 unsigned _int64 off; /* workaround for microsoft bug (they claim to support POSIX) */
119#else
119 unsigned long off; 120 unsigned long off;
120 int lit = 0; 121#endif
122 int lit;
123
124 if (!in_len || !out_len)
125 return 0;
121 126
122#if INIT_HTAB 127#if INIT_HTAB
123 memset (htab, 0, sizeof (htab)); 128 memset (htab, 0, sizeof (htab));
124# if 0 129# if 0
125 for (hslot = htab; hslot < htab + HSIZE; hslot++) 130 for (hslot = htab; hslot < htab + HSIZE; hslot++)
126 *hslot++ = ip; 131 *hslot++ = ip;
127# endif 132# endif
128#endif 133#endif
129 134
130 for (;;) 135 lit = 0; op++; /* start run */
136
137 hval = FRST (ip);
138 while (ip < in_end - 2)
131 { 139 {
132 if (expect_true (ip < in_end - 2)) 140 hval = NEXT (hval, ip);
141 hslot = htab + IDX (hval);
142 ref = *hslot; *hslot = ip;
143
144 if (1
145#if INIT_HTAB
146 && ref < ip /* the next test will actually take care of this, but this is faster */
147#endif
148 && (off = ip - ref - 1) < MAX_OFF
149 && ip + 4 < in_end
150 && ref > (u8 *)in_data
151#if STRICT_ALIGN
152 && ref[0] == ip[0]
153 && ref[1] == ip[1]
154 && ref[2] == ip[2]
155#else
156 && *(u16 *)ref == *(u16 *)ip
157 && ref[2] == ip[2]
158#endif
159 )
133 { 160 {
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++ */ 161 /* match found at *ref++ */
156 unsigned int len = 2; 162 unsigned int len = 2;
157 unsigned int maxlen = in_end - ip - len; 163 unsigned int maxlen = in_end - ip - len;
158 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 164 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
159 165
166 op [- lit - 1] = lit - 1; /* stop run */
167 op -= !lit; /* undo run if length is zero */
168
160 if (expect_false (op + lit + 1 + 3 >= out_end)) 169 if (expect_false (op + 3 + 1 >= out_end))
161 return 0; 170 return 0;
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
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
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
191 len++; if (ref [len] != ip [len]) break;
192 len++; if (ref [len] != ip [len]) break;
193 len++; if (ref [len] != ip [len]) break;
194 len++; if (ref [len] != ip [len]) break;
195 }
162 196
163 do 197 do
164 len++; 198 len++;
165 while (len < maxlen && ref[len] == ip[len]); 199 while (len < maxlen && ref[len] == ip[len]);
166 200
167 if (lit)
168 { 201 break;
169 *op++ = lit - 1;
170 lit = -lit;
171 do
172 *op++ = ip[lit];
173 while (++lit);
174 } 202 }
175 203
176 len -= 2; 204 len -= 2;
177 ip++; 205 ip++;
178 206
179 if (len < 7) 207 if (len < 7)
180 { 208 {
181 *op++ = (off >> 8) + (len << 5); 209 *op++ = (off >> 8) + (len << 5);
182 } 210 }
183 else 211 else
184 { 212 {
185 *op++ = (off >> 8) + ( 7 << 5); 213 *op++ = (off >> 8) + ( 7 << 5);
186 *op++ = len - 7; 214 *op++ = len - 7;
187 } 215 }
188 216
189 *op++ = off; 217 *op++ = off;
190 218
191#if ULTRA_FAST || VERY_FAST 219#if ULTRA_FAST || VERY_FAST
192 ip += len; 220 ip += len;
193#if VERY_FAST && !ULTRA_FAST 221#if VERY_FAST && !ULTRA_FAST
194 --ip; 222 --ip;
195#endif 223#endif
196 hval = FRST (ip); 224 hval = FRST (ip);
197 225
226 hval = NEXT (hval, ip);
227 htab[IDX (hval)] = ip;
228 ip++;
229
230#if VERY_FAST && !ULTRA_FAST
231 hval = NEXT (hval, ip);
232 htab[IDX (hval)] = ip;
233 ip++;
234#endif
235#else
236 do
237 {
198 hval = NEXT (hval, ip); 238 hval = NEXT (hval, ip);
199 htab[IDX (hval)] = ip; 239 htab[IDX (hval)] = ip;
200 ip++; 240 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 } 241 }
214 while (len--); 242 while (len--);
215#endif 243#endif
216 continue; 244
217 } 245 lit = 0; op++; /* start run */
218 } 246 }
219 else if (expect_false (ip == in_end)) 247 else
220 break; 248 {
221
222 /* one more literal byte we must copy */ 249 /* one more literal byte we must copy */
223 lit++; 250 if (expect_false (op >= out_end))
224 ip++; 251 return 0;
252
253 lit++; *op++ = *ip++;
254
255 if (expect_false (lit == MAX_LIT))
256 {
257 op [- lit - 1] = lit - 1; /* stop run */
258 lit = 0; op++; /* start run */
259 }
260 }
261 }
262
263 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
264 return 0;
265
266 while (ip < in_end)
267 {
268 lit++; *op++ = *ip++;
225 269
226 if (expect_false (lit == MAX_LIT)) 270 if (expect_false (lit == MAX_LIT))
227 { 271 {
228 if (op + 1 + MAX_LIT >= out_end) 272 op [- lit - 1] = lit - 1; /* stop run */
229 return 0; 273 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 } 274 }
243 } 275 }
244 276
245 if (lit) 277 op [- lit - 1] = lit - 1; /* end run */
246 { 278 op -= !lit; /* undo run if length is zero */
247 if (op + lit + 1 >= out_end)
248 return 0;
249 279
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; 280 return op - (u8 *)out_data;
263} 281}
264 282

Diff Legend

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