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 1.20 by root, Tue Nov 13 08:38:49 2007 UTC vs.
Revision 1.24 by root, Tue Nov 13 10:56:57 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 && !USE_MEMCPY
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 (expect_false (op + 1 + 3 >= out_end))
163 return 0;
164
165 op [- lit - 1] = lit - 1; /* stop run */
166 op -= !lit; /* undo run if length is zero */
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 }
192
193 do
194 len++;
195 while (len < maxlen && ref[len] == ip[len]);
196
197 break;
198 }
199
200 len -= 2;
201 ip++;
202
203 if (len < 7)
204 {
205 *op++ = (off >> 8) + (len << 5);
206 }
207 else
208 {
209 *op++ = (off >> 8) + ( 7 << 5);
210 *op++ = len - 7;
211 }
212
213 *op++ = off;
214
215#if ULTRA_FAST || VERY_FAST
216 ip += len;
217#if VERY_FAST && !ULTRA_FAST
218 --ip;
219#endif
220 hval = FRST (ip);
221
134 hval = NEXT (hval, ip); 222 hval = NEXT (hval, ip);
135 hslot = htab + IDX (hval); 223 htab[IDX (hval)] = ip;
136 ref = *hslot; *hslot = ip; 224 ip++;
137 225
138 if (1 226#if VERY_FAST && !ULTRA_FAST
139#if INIT_HTAB && !USE_MEMCPY 227 hval = NEXT (hval, ip);
140 && ref < ip /* the next test will actually take care of this, but this is faster */ 228 htab[IDX (hval)] = ip;
229 ip++;
141#endif 230#endif
142 && (off = ip - ref - 1) < MAX_OFF
143 && ip + 4 < in_end
144 && ref > (u8 *)in_data
145#if STRICT_ALIGN
146 x x x
147 && ref[0] == ip[0]
148 && ref[1] == ip[1]
149 && ref[2] == ip[2]
150#else 231#else
151 && *(u16 *)ref == *(u16 *)ip 232 do
152 && ref[2] == ip[2]
153#endif
154 ) 233 {
155 {
156 /* match found at *ref++ */
157 unsigned int len = 2;
158 unsigned int maxlen = in_end - ip - len;
159 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
160
161 if (expect_false (op + lit + 1 + 3 >= out_end))
162 return 0;
163
164 if (expect_false (lit))
165 {
166 *op++ = lit - 1;
167 lit = -lit;
168 do
169 *op++ = ip[lit];
170 while (expect_false (++lit));
171 }
172
173 for (;;)
174 {
175 if (expect_true (maxlen > 16))
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 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 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
224 hval = NEXT (hval, ip); 234 hval = NEXT (hval, ip);
225 htab[IDX (hval)] = ip; 235 htab[IDX (hval)] = ip;
226 ip++; 236 ip++;
227
228#if VERY_FAST && !ULTRA_FAST
229 hval = NEXT (hval, ip);
230 htab[IDX (hval)] = ip;
231 ip++;
232#endif
233#else
234 do
235 {
236 hval = NEXT (hval, ip);
237 htab[IDX (hval)] = ip;
238 ip++;
239 } 237 }
240 while (len--); 238 while (len--);
241#endif 239#endif
240 lit = 0; op++; /* start run */
242 continue; 241 continue;
243 }
244 } 242 }
245 else if (expect_false (ip == in_end))
246 break;
247 243
248 /* one more literal byte we must copy */ 244 /* one more literal byte we must copy */
245
246 if (expect_false (op >= out_end))
247 return 0;
248
249 lit++; 249 lit++;
250 ip++; 250 *op++ = *ip++;
251 251
252 if (expect_false (lit == MAX_LIT)) 252 if (expect_false (lit == MAX_LIT))
253 { 253 {
254 if (op + 1 + MAX_LIT >= out_end) 254 op [- lit - 1] = lit - 1; /* stop run */
255 return 0; 255 lit = 0; op++; /* start run */
256
257 *op++ = MAX_LIT - 1;
258
259#ifdef lzf_movsb
260 ip -= MAX_LIT;
261 lzf_movsb (op, ip, lit);
262#else
263 lit = -lit;
264 do
265 *op++ = ip[lit];
266 while (++lit);
267#endif
268 } 256 }
269 } 257 }
270 258
271 if (lit) 259 if (op + 2 >= out_end)
260 return 0;
261
262 while (ip < in_end)
272 { 263 {
273 if (op + lit + 1 >= out_end) 264 lit++;
274 return 0;
275
276 *op++ = lit - 1; 265 *op++ = *ip++;
277#ifdef lzf_movsb
278 ip -= lit;
279 lzf_movsb (op, ip, lit);
280#else
281 lit = -lit;
282 do
283 *op++ = ip[lit];
284 while (++lit);
285#endif
286 } 266 }
287 267
268 op [- lit - 1] = lit - 1; /* end run */
269 op -= !lit; /* undo run if length is zero */
270
288 return op - (u8 *) out_data; 271 return op - (u8 *)out_data;
289} 272}
290 273

Diff Legend

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