1 | /* |
1 | /* |
2 | * Copyright (c) 2000-2007 Marc Alexander Lehmann <schmorp@schmorp.de> |
2 | * Copyright (c) 2000-2010 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, |
… | |
… | |
38 | |
38 | |
39 | #define HSIZE (1 << (HLOG)) |
39 | #define HSIZE (1 << (HLOG)) |
40 | |
40 | |
41 | /* |
41 | /* |
42 | * don't play with this unless you benchmark! |
42 | * don't play with this unless you benchmark! |
43 | * decompression is not dependent on the hash function |
43 | * the data format is not dependent on the hash function. |
44 | * the hashing function might seem strange, just believe me |
44 | * the hash function might seem strange, just believe me, |
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 MULTIPLICATION_IS_SLOW |
50 | # if ULTRA_FAST |
51 | # if ULTRA_FAST |
51 | # define IDX(h) ((( h >> (3*8 - HLOG)) - h ) & (HSIZE - 1)) |
52 | # define IDX(h) ((( h >> (3*8 - HLOG)) - h ) & (HSIZE - 1)) |
52 | # elif VERY_FAST |
53 | # elif VERY_FAST |
53 | # define IDX(h) ((( h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) |
54 | # define IDX(h) ((( h >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) |
|
|
55 | # else |
|
|
56 | # define IDX(h) ((((h ^ (h << 5)) >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) |
|
|
57 | # endif |
54 | # else |
58 | # else |
|
|
59 | /* this one was developed with sesse, |
|
|
60 | * and is very similar to the one in snappy. |
|
|
61 | * it does need a modern enough cpu with a fast multiplication. |
|
|
62 | */ |
55 | # define IDX(h) ((((h ^ (h << 5)) >> (3*8 - HLOG)) - h*5) & (HSIZE - 1)) |
63 | # define IDX(h) (((h * 0x1e35a7bdU) >> (32 - HLOG - 8)) & (HSIZE - 1)) |
56 | # endif |
64 | # endif |
57 | #endif |
65 | #endif |
58 | /* |
|
|
59 | * IDX works because it is very similar to a multiplicative hash, e.g. |
|
|
60 | * ((h * 57321 >> (3*8 - HLOG)) & (HSIZE - 1)) |
|
|
61 | * the latter is also quite fast on newer CPUs, and compresses similarly. |
|
|
62 | * |
|
|
63 | * the next one is also quite good, albeit slow ;) |
|
|
64 | * (int)(cos(h & 0xffffff) * 1e6) |
|
|
65 | */ |
|
|
66 | |
66 | |
67 | #if 0 |
67 | #if 0 |
68 | /* original lzv-like hash function, much worse and thus slower */ |
68 | /* original lzv-like hash function, much worse and thus slower */ |
69 | # define FRST(p) (p[0] << 5) ^ p[1] |
69 | # define FRST(p) (p[0] << 5) ^ p[1] |
70 | # define NEXT(v,p) ((v) << 5) ^ p[2] |
70 | # define NEXT(v,p) ((v) << 5) ^ p[2] |
… | |
… | |
87 | #define expect_true(expr) expect ((expr) != 0, 1) |
87 | #define expect_true(expr) expect ((expr) != 0, 1) |
88 | |
88 | |
89 | /* |
89 | /* |
90 | * compressed format |
90 | * compressed format |
91 | * |
91 | * |
92 | * 000LLLLL <L+1> ; literal |
92 | * 000LLLLL <L+1> ; literal, L+1=1..33 octets |
93 | * LLLooooo oooooooo ; backref L |
93 | * LLLooooo oooooooo ; backref L+1=1..7 octets, o+1=1..4096 offset |
94 | * 111ooooo LLLLLLLL oooooooo ; backref L+7 |
94 | * 111ooooo LLLLLLLL oooooooo ; backref L+8 octets, o+1=1..4096 offset |
95 | * |
95 | * |
96 | */ |
96 | */ |
97 | |
97 | |
98 | unsigned int |
98 | unsigned int |
99 | lzf_compress (const void *const in_data, unsigned int in_len, |
99 | lzf_compress (const void *const in_data, unsigned int in_len, |
… | |
… | |
104 | ) |
104 | ) |
105 | { |
105 | { |
106 | #if !LZF_STATE_ARG |
106 | #if !LZF_STATE_ARG |
107 | LZF_STATE htab; |
107 | LZF_STATE htab; |
108 | #endif |
108 | #endif |
109 | const u8 **hslot; |
|
|
110 | const u8 *ip = (const u8 *)in_data; |
109 | const u8 *ip = (const u8 *)in_data; |
111 | u8 *op = (u8 *)out_data; |
110 | u8 *op = (u8 *)out_data; |
112 | const u8 *in_end = ip + in_len; |
111 | const u8 *in_end = ip + in_len; |
113 | u8 *out_end = op + out_len; |
112 | u8 *out_end = op + out_len; |
114 | const u8 *ref; |
113 | const u8 *ref; |
115 | |
114 | |
|
|
115 | /* off requires a type wide enough to hold a general pointer difference. |
|
|
116 | * ISO C doesn't have that (size_t might not be enough and ptrdiff_t only |
|
|
117 | * works for differences within a single object). We also assume that |
|
|
118 | * no bit pattern traps. Since the only platform that is both non-POSIX |
|
|
119 | * and fails to support both assumptions is windows 64 bit, we make a |
|
|
120 | * special workaround for it. |
|
|
121 | */ |
|
|
122 | #if defined (WIN32) && defined (_M_X64) |
|
|
123 | unsigned _int64 off; /* workaround for missing POSIX compliance */ |
|
|
124 | #else |
|
|
125 | unsigned long off; |
|
|
126 | #endif |
116 | unsigned int hval; |
127 | unsigned int hval; |
117 | unsigned long off; |
|
|
118 | int lit; |
128 | int lit; |
119 | |
129 | |
120 | if (!in_len || !out_len) |
130 | if (!in_len || !out_len) |
121 | return 0; |
131 | return 0; |
122 | |
132 | |
123 | #if INIT_HTAB |
133 | #if INIT_HTAB |
124 | memset (htab, 0, sizeof (htab)); |
134 | memset (htab, 0, sizeof (htab)); |
125 | # if 0 |
|
|
126 | for (hslot = htab; hslot < htab + HSIZE; hslot++) |
|
|
127 | *hslot++ = ip; |
|
|
128 | # endif |
|
|
129 | #endif |
135 | #endif |
130 | |
136 | |
131 | lit = 0; op++; /* start run */ |
137 | lit = 0; op++; /* start run */ |
132 | |
138 | |
133 | hval = FRST (ip); |
139 | hval = FRST (ip); |
134 | while (ip < in_end - 2) |
140 | while (ip < in_end - 2) |
135 | { |
141 | { |
|
|
142 | LZF_HSLOT *hslot; |
|
|
143 | |
136 | hval = NEXT (hval, ip); |
144 | hval = NEXT (hval, ip); |
137 | hslot = htab + IDX (hval); |
145 | hslot = htab + IDX (hval); |
138 | ref = *hslot; *hslot = ip; |
146 | ref = *hslot + LZF_HSLOT_BIAS; *hslot = ip - LZF_HSLOT_BIAS; |
139 | |
147 | |
140 | if (1 |
148 | if (1 |
141 | #if INIT_HTAB && !USE_MEMCPY |
149 | #if INIT_HTAB |
142 | && ref < ip /* the next test will actually take care of this, but this is faster */ |
150 | && ref < ip /* the next test will actually take care of this, but this is faster */ |
143 | #endif |
151 | #endif |
144 | && (off = ip - ref - 1) < MAX_OFF |
152 | && (off = ip - ref - 1) < MAX_OFF |
145 | && ip + 4 < in_end |
|
|
146 | && ref > (u8 *)in_data |
153 | && ref > (u8 *)in_data |
|
|
154 | && ref[2] == ip[2] |
147 | #if STRICT_ALIGN |
155 | #if STRICT_ALIGN |
148 | && ref[0] == ip[0] |
156 | && ((ref[1] << 8) | ref[0]) == ((ip[1] << 8) | ip[0]) |
149 | && ref[1] == ip[1] |
|
|
150 | && ref[2] == ip[2] |
|
|
151 | #else |
157 | #else |
152 | && *(u16 *)ref == *(u16 *)ip |
158 | && *(u16 *)ref == *(u16 *)ip |
153 | && ref[2] == ip[2] |
|
|
154 | #endif |
159 | #endif |
155 | ) |
160 | ) |
156 | { |
161 | { |
157 | /* match found at *ref++ */ |
162 | /* match found at *ref++ */ |
158 | unsigned int len = 2; |
163 | unsigned int len = 2; |
159 | unsigned int maxlen = in_end - ip - len; |
164 | unsigned int maxlen = in_end - ip - len; |
160 | maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; |
165 | maxlen = maxlen > MAX_REF ? MAX_REF : maxlen; |
161 | |
166 | |
162 | if (expect_false (op + 1 + 3 >= out_end)) |
167 | if (expect_false (op + 3 + 1 >= out_end)) /* first a faster conservative test */ |
|
|
168 | if (op - !lit + 3 + 1 >= out_end) /* second the exact but rare test */ |
163 | return 0; |
169 | return 0; |
164 | |
170 | |
165 | op [- lit - 1] = lit - 1; /* stop run */ |
171 | op [- lit - 1] = lit - 1; /* stop run */ |
166 | op -= !lit; /* undo run if length is zero */ |
172 | op -= !lit; /* undo run if length is zero */ |
167 | |
173 | |
168 | for (;;) |
174 | for (;;) |
… | |
… | |
195 | while (len < maxlen && ref[len] == ip[len]); |
201 | while (len < maxlen && ref[len] == ip[len]); |
196 | |
202 | |
197 | break; |
203 | break; |
198 | } |
204 | } |
199 | |
205 | |
200 | len -= 2; |
206 | len -= 2; /* len is now #octets - 1 */ |
201 | ip++; |
207 | ip++; |
202 | |
208 | |
203 | if (len < 7) |
209 | if (len < 7) |
204 | { |
210 | { |
205 | *op++ = (off >> 8) + (len << 5); |
211 | *op++ = (off >> 8) + (len << 5); |
… | |
… | |
210 | *op++ = len - 7; |
216 | *op++ = len - 7; |
211 | } |
217 | } |
212 | |
218 | |
213 | *op++ = off; |
219 | *op++ = off; |
214 | |
220 | |
|
|
221 | lit = 0; op++; /* start run */ |
|
|
222 | |
|
|
223 | ip += len + 1; |
|
|
224 | |
|
|
225 | if (expect_false (ip >= in_end - 2)) |
|
|
226 | break; |
|
|
227 | |
215 | #if ULTRA_FAST || VERY_FAST |
228 | #if ULTRA_FAST || VERY_FAST |
216 | ip += len; |
|
|
217 | #if VERY_FAST && !ULTRA_FAST |
|
|
218 | --ip; |
229 | --ip; |
|
|
230 | # if VERY_FAST && !ULTRA_FAST |
|
|
231 | --ip; |
219 | #endif |
232 | # endif |
220 | hval = FRST (ip); |
233 | hval = FRST (ip); |
221 | |
234 | |
222 | hval = NEXT (hval, ip); |
235 | hval = NEXT (hval, ip); |
223 | htab[IDX (hval)] = ip; |
236 | htab[IDX (hval)] = ip - LZF_HSLOT_BIAS; |
224 | ip++; |
237 | ip++; |
225 | |
238 | |
226 | #if VERY_FAST && !ULTRA_FAST |
239 | # if VERY_FAST && !ULTRA_FAST |
227 | hval = NEXT (hval, ip); |
240 | hval = NEXT (hval, ip); |
228 | htab[IDX (hval)] = ip; |
241 | htab[IDX (hval)] = ip - LZF_HSLOT_BIAS; |
229 | ip++; |
242 | ip++; |
230 | #endif |
243 | # endif |
231 | #else |
244 | #else |
|
|
245 | ip -= len + 1; |
|
|
246 | |
232 | do |
247 | do |
233 | { |
248 | { |
234 | hval = NEXT (hval, ip); |
249 | hval = NEXT (hval, ip); |
235 | htab[IDX (hval)] = ip; |
250 | htab[IDX (hval)] = ip - LZF_HSLOT_BIAS; |
236 | ip++; |
251 | ip++; |
237 | } |
252 | } |
238 | while (len--); |
253 | while (len--); |
239 | #endif |
254 | #endif |
240 | lit = 0; op++; /* start run */ |
|
|
241 | continue; |
|
|
242 | } |
255 | } |
243 | |
256 | else |
|
|
257 | { |
244 | /* one more literal byte we must copy */ |
258 | /* one more literal byte we must copy */ |
245 | |
|
|
246 | if (expect_false (op >= out_end)) |
259 | if (expect_false (op >= out_end)) |
247 | return 0; |
260 | return 0; |
248 | |
261 | |
249 | lit++; |
262 | lit++; *op++ = *ip++; |
|
|
263 | |
|
|
264 | if (expect_false (lit == MAX_LIT)) |
|
|
265 | { |
|
|
266 | op [- lit - 1] = lit - 1; /* stop run */ |
|
|
267 | lit = 0; op++; /* start run */ |
|
|
268 | } |
|
|
269 | } |
|
|
270 | } |
|
|
271 | |
|
|
272 | if (op + 3 > out_end) /* at most 3 bytes can be missing here */ |
|
|
273 | return 0; |
|
|
274 | |
|
|
275 | while (ip < in_end) |
|
|
276 | { |
250 | *op++ = *ip++; |
277 | lit++; *op++ = *ip++; |
251 | |
278 | |
252 | if (expect_false (lit == MAX_LIT)) |
279 | if (expect_false (lit == MAX_LIT)) |
253 | { |
280 | { |
254 | op [- lit - 1] = lit - 1; /* stop run */ |
281 | op [- lit - 1] = lit - 1; /* stop run */ |
255 | lit = 0; op++; /* start run */ |
282 | lit = 0; op++; /* start run */ |
256 | } |
283 | } |
257 | } |
284 | } |
258 | |
285 | |
259 | if (op + lit + 2 >= out_end) |
|
|
260 | return 0; |
|
|
261 | |
|
|
262 | while (ip < in_end) |
|
|
263 | { |
|
|
264 | lit++; |
|
|
265 | *op++ = *ip++; |
|
|
266 | } |
|
|
267 | |
|
|
268 | op [- lit - 1] = lit - 1; |
286 | op [- lit - 1] = lit - 1; /* end run */ |
|
|
287 | op -= !lit; /* undo run if length is zero */ |
269 | |
288 | |
270 | return op - (u8 *)out_data; |
289 | return op - (u8 *)out_data; |
271 | } |
290 | } |
272 | |
291 | |