| 1 | /* |
1 | /* |
| 2 | * Copyright (c) 2000-2007 Marc Alexander Lehmann <schmorp@schmorp.de> |
2 | * Copyright (c) 2000-2008 Marc Alexander Lehmann <schmorp@schmorp.de> |
| 3 | * |
3 | * |
| 4 | * Redistribution and use in source and binary forms, with or without modifica- |
4 | * Redistribution and use in source and binary forms, with or without modifica- |
| 5 | * tion, are permitted provided that the following conditions are met: |
5 | * tion, are permitted provided that the following conditions are met: |
| 6 | * |
6 | * |
| 7 | * 1. Redistributions of source code must retain the above copyright notice, |
7 | * 1. Redistributions of source code must retain the above copyright notice, |
| … | |
… | |
| 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 |
50 | # if ULTRA_FAST |
| 51 | # define IDX(h) (((h >> (3*8 - HLOG)) - h ) & (HSIZE - 1)) |
51 | # define IDX(h) ((( h >> (3*8 - HLOG)) - h ) & (HSIZE - 1)) |
| 52 | # elif VERY_FAST |
52 | # elif VERY_FAST |
| 53 | # define IDX(h) (((h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) |
53 | # define IDX(h) ((( h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) |
| 54 | # else |
54 | # else |
| 55 | # 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)) |
| 56 | # endif |
56 | # endif |
| 57 | /*# define IDX(h) ((ip[0] * 121 ^ ip[1] * 33 ^ ip[2] * 1) & (HSIZE-1))*/ |
|
|
| 58 | #endif |
57 | #endif |
| 59 | /* |
58 | /* |
| 60 | * 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. |
| 61 | * ((h * 57321 >> (3*8 - HLOG)) & (HSIZE - 1)) |
60 | * ((h * 57321 >> (3*8 - HLOG)) & (HSIZE - 1)) |
| 62 | * 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. |
| … | |
… | |
| 112 | u8 *op = (u8 *)out_data; |
111 | u8 *op = (u8 *)out_data; |
| 113 | const u8 *in_end = ip + in_len; |
112 | const u8 *in_end = ip + in_len; |
| 114 | u8 *out_end = op + out_len; |
113 | u8 *out_end = op + out_len; |
| 115 | const u8 *ref; |
114 | const u8 *ref; |
| 116 | |
115 | |
|
|
116 | /* off requires a type wide enough to hold a general pointer difference. |
|
|
117 | * ISO C doesn't have that (size_t might not be enough and ptrdiff_t only |
|
|
118 | * works for differences within a single object). We also assume that no |
|
|
119 | * no bit pattern traps. Since the only platform that is both non-POSIX |
|
|
120 | * and fails to support both assumptions is windows 64 bit, we make a |
|
|
121 | * special workaround for it. |
|
|
122 | */ |
|
|
123 | #if defined (WIN32) && defined (_M_X64) |
|
|
124 | unsigned _int64 off; /* workaround for missing POSIX compliance */ |
|
|
125 | #else |
|
|
126 | unsigned long off; |
|
|
127 | #endif |
| 117 | unsigned int hval; |
128 | unsigned int hval; |
| 118 | unsigned long off; |
|
|
| 119 | int lit; |
129 | int lit; |
| 120 | |
130 | |
| 121 | if (!in_len || !out_len) |
131 | if (!in_len || !out_len) |
| 122 | return 0; |
132 | return 0; |
| 123 | |
133 | |
| … | |
… | |
| 137 | hval = NEXT (hval, ip); |
147 | hval = NEXT (hval, ip); |
| 138 | hslot = htab + IDX (hval); |
148 | hslot = htab + IDX (hval); |
| 139 | ref = *hslot; *hslot = ip; |
149 | ref = *hslot; *hslot = ip; |
| 140 | |
150 | |
| 141 | if (1 |
151 | if (1 |
| 142 | #if INIT_HTAB && !USE_MEMCPY |
152 | #if INIT_HTAB |
| 143 | && ref < ip /* the next test will actually take care of this, but this is faster */ |
153 | && ref < ip /* the next test will actually take care of this, but this is faster */ |
| 144 | #endif |
154 | #endif |
| 145 | && (off = ip - ref - 1) < MAX_OFF |
155 | && (off = ip - ref - 1) < MAX_OFF |
| 146 | && ip + 4 < in_end |
156 | && ip + 4 < in_end |
| 147 | && ref > (u8 *)in_data |
157 | && ref > (u8 *)in_data |
| … | |
… | |
| 158 | /* match found at *ref++ */ |
168 | /* match found at *ref++ */ |
| 159 | unsigned int len = 2; |
169 | unsigned int len = 2; |
| 160 | unsigned int maxlen = in_end - ip - len; |
170 | unsigned int maxlen = in_end - ip - len; |
| 161 | maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; |
171 | maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; |
| 162 | |
172 | |
| 163 | if (expect_false (op + 1 + 3 >= out_end)) |
|
|
| 164 | return 0; |
|
|
| 165 | |
|
|
| 166 | op [- lit - 1] = lit - 1; /* stop run */ |
173 | op [- lit - 1] = lit - 1; /* stop run */ |
| 167 | op -= !lit; /* undo run if length is zero */ |
174 | op -= !lit; /* undo run if length is zero */ |
|
|
175 | |
|
|
176 | if (expect_false (op + 3 + 1 >= out_end)) |
|
|
177 | return 0; |
| 168 | |
178 | |
| 169 | for (;;) |
179 | for (;;) |
| 170 | { |
180 | { |
| 171 | if (expect_true (maxlen > 16)) |
181 | if (expect_true (maxlen > 16)) |
| 172 | { |
182 | { |
| … | |
… | |
| 196 | while (len < maxlen && ref[len] == ip[len]); |
206 | while (len < maxlen && ref[len] == ip[len]); |
| 197 | |
207 | |
| 198 | break; |
208 | break; |
| 199 | } |
209 | } |
| 200 | |
210 | |
| 201 | len -= 2; |
211 | len -= 2; /* len is now #octets - 1 */ |
| 202 | ip++; |
212 | ip++; |
| 203 | |
213 | |
| 204 | if (len < 7) |
214 | if (len < 7) |
| 205 | { |
215 | { |
| 206 | *op++ = (off >> 8) + (len << 5); |
216 | *op++ = (off >> 8) + (len << 5); |
| … | |
… | |
| 210 | *op++ = (off >> 8) + ( 7 << 5); |
220 | *op++ = (off >> 8) + ( 7 << 5); |
| 211 | *op++ = len - 7; |
221 | *op++ = len - 7; |
| 212 | } |
222 | } |
| 213 | |
223 | |
| 214 | *op++ = off; |
224 | *op++ = off; |
|
|
225 | lit = 0; op++; /* start run */ |
|
|
226 | |
|
|
227 | ip += len + 1; |
|
|
228 | |
|
|
229 | if (expect_false (ip > in_end - 2)) |
|
|
230 | break; |
| 215 | |
231 | |
| 216 | #if ULTRA_FAST || VERY_FAST |
232 | #if ULTRA_FAST || VERY_FAST |
| 217 | ip += len; |
|
|
| 218 | #if VERY_FAST && !ULTRA_FAST |
|
|
| 219 | --ip; |
233 | --ip; |
|
|
234 | # if VERY_FAST && !ULTRA_FAST |
|
|
235 | --ip; |
| 220 | #endif |
236 | # endif |
| 221 | hval = FRST (ip); |
237 | hval = FRST (ip); |
| 222 | |
238 | |
| 223 | hval = NEXT (hval, ip); |
239 | hval = NEXT (hval, ip); |
| 224 | htab[IDX (hval)] = ip; |
240 | htab[IDX (hval)] = ip; |
| 225 | ip++; |
241 | ip++; |
| 226 | |
242 | |
| 227 | #if VERY_FAST && !ULTRA_FAST |
243 | # if VERY_FAST && !ULTRA_FAST |
| 228 | hval = NEXT (hval, ip); |
244 | hval = NEXT (hval, ip); |
| 229 | htab[IDX (hval)] = ip; |
245 | htab[IDX (hval)] = ip; |
| 230 | ip++; |
246 | ip++; |
| 231 | #endif |
247 | # endif |
| 232 | #else |
248 | #else |
|
|
249 | ip -= len + 1; |
|
|
250 | |
| 233 | do |
251 | do |
| 234 | { |
252 | { |
| 235 | hval = NEXT (hval, ip); |
253 | hval = NEXT (hval, ip); |
| 236 | htab[IDX (hval)] = ip; |
254 | htab[IDX (hval)] = ip; |
| 237 | ip++; |
255 | ip++; |
| 238 | } |
256 | } |
| 239 | while (len--); |
257 | while (len--); |
| 240 | #endif |
258 | #endif |
| 241 | lit = 0; op++; /* start run */ |
|
|
| 242 | continue; |
|
|
| 243 | } |
259 | } |
| 244 | |
260 | else |
|
|
261 | { |
| 245 | /* one more literal byte we must copy */ |
262 | /* one more literal byte we must copy */ |
| 246 | |
|
|
| 247 | if (expect_false (op >= out_end)) |
263 | if (expect_false (op >= out_end)) |
| 248 | return 0; |
264 | return 0; |
| 249 | |
265 | |
| 250 | lit++; |
266 | lit++; *op++ = *ip++; |
|
|
267 | |
|
|
268 | if (expect_false (lit == MAX_LIT)) |
|
|
269 | { |
|
|
270 | op [- lit - 1] = lit - 1; /* stop run */ |
|
|
271 | lit = 0; op++; /* start run */ |
|
|
272 | } |
|
|
273 | } |
|
|
274 | } |
|
|
275 | |
|
|
276 | if (op + 3 > out_end) /* at most 3 bytes can be missing here */ |
|
|
277 | return 0; |
|
|
278 | |
|
|
279 | while (ip < in_end) |
|
|
280 | { |
| 251 | *op++ = *ip++; |
281 | lit++; *op++ = *ip++; |
| 252 | |
282 | |
| 253 | if (expect_false (lit == MAX_LIT)) |
283 | if (expect_false (lit == MAX_LIT)) |
| 254 | { |
284 | { |
| 255 | op [- lit - 1] = lit - 1; /* stop run */ |
285 | op [- lit - 1] = lit - 1; /* stop run */ |
| 256 | lit = 0; op++; /* start run */ |
286 | lit = 0; op++; /* start run */ |
| 257 | } |
287 | } |
| 258 | } |
288 | } |
| 259 | |
289 | |
| 260 | if (op + lit + 2 >= out_end) |
|
|
| 261 | return 0; |
|
|
| 262 | |
|
|
| 263 | while (ip < in_end) |
|
|
| 264 | { |
|
|
| 265 | lit++; |
|
|
| 266 | *op++ = *ip++; |
|
|
| 267 | } |
|
|
| 268 | |
|
|
| 269 | op [- lit - 1] = lit - 1; |
290 | op [- lit - 1] = lit - 1; /* end run */ |
|
|
291 | op -= !lit; /* undo run if length is zero */ |
| 270 | |
292 | |
| 271 | return op - (u8 *)out_data; |
293 | return op - (u8 *)out_data; |
| 272 | } |
294 | } |
| 273 | |
295 | |
