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 1.16 by root, Mon Nov 5 23:28:40 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.
… …
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 77
73#if (__i386 || __amd64) && __GNUC__ >= 3 78#if __GNUC__ >= 3
74# define lzf_movsb(dst, src, len) \ 79# define expect(expr,value) __builtin_expect ((expr),(value))
75 asm ("rep movsb" \ 80# define inline inline
76 : "=D" (dst), "=S" (src), "=c" (len) \ 81#else
77 : "0" (dst), "1" (src), "2" (len)); 82# define expect(expr,value) (expr)
83# define inline static
78#endif 84#endif
85
86#define expect_false(expr) expect ((expr) != 0, 0)
87#define expect_true(expr) expect ((expr) != 0, 1)
79 88
80/* 89/*
81 * compressed format 90 * compressed format
82 * 91 *
83 * 000LLLLL <L+1> ; literal 92 * 000LLLLL <L+1> ; literal
… …
102 u8 *op = (u8 *)out_data; 111 u8 *op = (u8 *)out_data;
103 const u8 *in_end = ip + in_len; 112 const u8 *in_end = ip + in_len;
104 u8 *out_end = op + out_len; 113 u8 *out_end = op + out_len;
105 const u8 *ref; 114 const u8 *ref;
106 115
107 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
108 unsigned long off; 120 unsigned long off;
109 int lit = 0; 121#endif
122 int lit;
123
124 if (!in_len || !out_len)
125 return 0;
110 126
111#if INIT_HTAB 127#if INIT_HTAB
112 memset (htab, 0, sizeof (htab)); 128 memset (htab, 0, sizeof (htab));
113# if 0 129# if 0
114 for (hslot = htab; hslot < htab + HSIZE; hslot++) 130 for (hslot = htab; hslot < htab + HSIZE; hslot++)
115 *hslot++ = ip; 131 *hslot++ = ip;
116# endif 132# endif
117#endif 133#endif
118 134
119 for (;;) 135 lit = 0; op++; /* start run */
136
137 hval = FRST (ip);
138 while (ip < in_end - 2)
120 { 139 {
121 if (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 )
122 { 160 {
123 hval = NEXT (hval, ip);
124 hslot = htab + IDX (hval);
125 ref = *hslot; *hslot = ip;
126
127 if (1
128#if INIT_HTAB && !USE_MEMCPY
129 && ref < ip /* the next test will actually take care of this, but this is faster */
130#endif
131 && (off = ip - ref - 1) < MAX_OFF
132 && ip + 4 < in_end
133 && ref > (u8 *)in_data
134#if STRICT_ALIGN
135 && ref[0] == ip[0]
136 && ref[1] == ip[1]
137 && ref[2] == ip[2]
138#else
139 && *(u16 *)ref == *(u16 *)ip
140 && ref[2] == ip[2]
141#endif
142 )
143 {
144 /* match found at *ref++ */ 161 /* match found at *ref++ */
145 unsigned int len = 2; 162 unsigned int len = 2;
146 unsigned int maxlen = in_end - ip - len; 163 unsigned int maxlen = in_end - ip - len;
147 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; 164 maxlen = maxlen > MAX_REF ? MAX_REF : maxlen;
148 165
149 if (op + lit + 1 + 3 >= out_end) 166 op [- lit - 1] = lit - 1; /* stop run */
167 op -= !lit; /* undo run if length is zero */
168
169 if (expect_false (op + 3 + 1 >= out_end))
150 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 }
151 196
152 do 197 do
153 len++; 198 len++;
154 while (len < maxlen && ref[len] == ip[len]); 199 while (len < maxlen && ref[len] == ip[len]);
155 200
156 if (lit)
157 { 201 break;
158 *op++ = lit - 1;
159 lit = -lit;
160 do
161 *op++ = ip[lit];
162 while (++lit);
163 } 202 }
164 203
165 len -= 2; 204 len -= 2;
166 ip++; 205 ip++;
167 206
168 if (len < 7) 207 if (len < 7)
169 { 208 {
170 *op++ = (off >> 8) + (len << 5); 209 *op++ = (off >> 8) + (len << 5);
171 } 210 }
172 else 211 else
173 { 212 {
174 *op++ = (off >> 8) + ( 7 << 5); 213 *op++ = (off >> 8) + ( 7 << 5);
175 *op++ = len - 7; 214 *op++ = len - 7;
176 } 215 }
177 216
178 *op++ = off; 217 *op++ = off;
179 218
180#if ULTRA_FAST || VERY_FAST 219#if ULTRA_FAST || VERY_FAST
181 ip += len; 220 ip += len;
182#if VERY_FAST && !ULTRA_FAST 221#if VERY_FAST && !ULTRA_FAST
183 --ip; 222 --ip;
184#endif 223#endif
185 hval = FRST (ip); 224 hval = FRST (ip);
186 225
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 {
187 hval = NEXT (hval, ip); 238 hval = NEXT (hval, ip);
188 htab[IDX (hval)] = ip; 239 htab[IDX (hval)] = ip;
189 ip++; 240 ip++;
190
191#if VERY_FAST && !ULTRA_FAST
192 hval = NEXT (hval, ip);
193 htab[IDX (hval)] = ip;
194 ip++;
195#endif
196#else
197 do
198 {
199 hval = NEXT (hval, ip);
200 htab[IDX (hval)] = ip;
201 ip++;
202 } 241 }
203 while (len--); 242 while (len--);
204#endif 243#endif
205 continue; 244
206 } 245 lit = 0; op++; /* start run */
207 } 246 }
208 else if (ip == in_end) 247 else
209 break;
210
211 /* one more literal byte we must copy */
212 lit++;
213 ip++;
214
215 if (lit == MAX_LIT)
216 { 248 {
217 if (op + 1 + MAX_LIT >= out_end) 249 /* one more literal byte we must copy */
250 if (expect_false (op >= out_end))
218 return 0; 251 return 0;
219 252
220 *op++ = MAX_LIT - 1; 253 lit++; *op++ = *ip++;
221 254
222#ifdef lzf_movsb 255 if (expect_false (lit == MAX_LIT))
223 ip -= lit;
224 lzf_movsb (op, ip, lit);
225#else
226 lit = -lit;
227 do 256 {
228 *op++ = ip[lit]; 257 op [- lit - 1] = lit - 1; /* stop run */
229 while (++lit); 258 lit = 0; op++; /* start run */
230#endif 259 }
231 } 260 }
232 } 261 }
233 262
234 if (lit) 263 if (op + 3 > out_end) /* at most 3 bytes can be missing here */
264 return 0;
265
266 while (ip < in_end)
235 { 267 {
236 if (op + lit + 1 >= out_end) 268 lit++; *op++ = *ip++;
237 return 0;
238 269
239 *op++ = lit - 1; 270 if (expect_false (lit == MAX_LIT))
240#ifdef lzf_movsb 271 {
241 ip -= lit; 272 op [- lit - 1] = lit - 1; /* stop run */
242 lzf_movsb (op, ip, lit); 273 lit = 0; op++; /* start run */
243#else 274 }
244 lit = -lit;
245 do
246 *op++ = ip[lit];
247 while (++lit);
248#endif
249 } 275 }
250 276
277 op [- lit - 1] = lit - 1; /* end run */
278 op -= !lit; /* undo run if length is zero */
279
251 return op - (u8 *) out_data; 280 return op - (u8 *)out_data;
252} 281}
253 282

Diff Legend

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