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 1.30 by root, Fri Nov 30 10:36:48 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 {
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++ */ 157 /* match found at *ref++ */
156 unsigned int len = 2; 158 unsigned int len = 2;
157 unsigned int maxlen = in_end - ip - len; 159 unsigned int maxlen = in_end - ip - len;
158 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 160 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
159 161
162 op [- lit - 1] = lit - 1; /* stop run */
163 op -= !lit; /* undo run if length is zero */
164
160 if (expect_false (op + lit + 1 + 3 >= out_end)) 165 if (expect_false (op + 3 + 1 >= out_end))
161 return 0; 166 return 0;
167
168 for (;;)
169 {
170 if (expect_true (maxlen > 16))
171 {
172 len++; if (ref [len] != ip [len]) break;
173 len++; if (ref [len] != ip [len]) break;
174 len++; if (ref [len] != ip [len]) break;
175 len++; if (ref [len] != ip [len]) break;
176
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
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
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 }
162 192
163 do 193 do
164 len++; 194 len++;
165 while (len < maxlen && ref[len] == ip[len]); 195 while (len < maxlen && ref[len] == ip[len]);
166 196
167 if (lit)
168 { 197 break;
169 *op++ = lit - 1;
170 lit = -lit;
171 do
172 *op++ = ip[lit];
173 while (++lit);
174 } 198 }
175 199
176 len -= 2; 200 len -= 2;
177 ip++; 201 ip++;
178 202
179 if (len < 7) 203 if (len < 7)
180 { 204 {
181 *op++ = (off >> 8) + (len << 5); 205 *op++ = (off >> 8) + (len << 5);
182 } 206 }
183 else 207 else
184 { 208 {
185 *op++ = (off >> 8) + ( 7 << 5); 209 *op++ = (off >> 8) + ( 7 << 5);
186 *op++ = len - 7; 210 *op++ = len - 7;
187 } 211 }
188 212
189 *op++ = off; 213 *op++ = off;
190 214
191#if ULTRA_FAST || VERY_FAST 215#if ULTRA_FAST || VERY_FAST
192 ip += len; 216 ip += len;
193#if VERY_FAST && !ULTRA_FAST 217#if VERY_FAST && !ULTRA_FAST
194 --ip; 218 --ip;
195#endif 219#endif
196 hval = FRST (ip); 220 hval = FRST (ip);
197 221
222 hval = NEXT (hval, ip);
223 htab[IDX (hval)] = ip;
224 ip++;
225
226#if VERY_FAST && !ULTRA_FAST
227 hval = NEXT (hval, ip);
228 htab[IDX (hval)] = ip;
229 ip++;
230#endif
231#else
232 do
233 {
198 hval = NEXT (hval, ip); 234 hval = NEXT (hval, ip);
199 htab[IDX (hval)] = ip; 235 htab[IDX (hval)] = ip;
200 ip++; 236 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 } 237 }
214 while (len--); 238 while (len--);
215#endif 239#endif
216 continue; 240
217 } 241 lit = 0; op++; /* start run */
218 } 242 }
219 else if (expect_false (ip == in_end)) 243 else
220 break; 244 {
221
222 /* one more literal byte we must copy */ 245 /* one more literal byte we must copy */
223 lit++; 246 if (expect_false (op >= out_end))
224 ip++; 247 return 0;
248
249 lit++; *op++ = *ip++;
250
251 if (expect_false (lit == MAX_LIT))
252 {
253 op [- lit - 1] = lit - 1; /* stop run */
254 lit = 0; op++; /* start run */
255 }
256 }
257 }
258
259 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
260 return 0;
261
262 while (ip < in_end)
263 {
264 lit++; *op++ = *ip++;
225 265
226 if (expect_false (lit == MAX_LIT)) 266 if (expect_false (lit == MAX_LIT))
227 { 267 {
228 if (op + 1 + MAX_LIT >= out_end) 268 op [- lit - 1] = lit - 1; /* stop run */
229 return 0; 269 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 } 270 }
243 } 271 }
244 272
245 if (lit) 273 op [- lit - 1] = lit - 1; /* end run */
246 { 274 op -= !lit; /* undo run if length is zero */
247 if (op + lit + 1 >= out_end)
248 return 0;
249 275
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; 276 return op - (u8 *)out_data;
263} 277}
264 278

Diff Legend

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