| … | |
… | |
| 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 | { |
| … | |
… | |
| 236 | htab[IDX (hval)] = ip; |
246 | htab[IDX (hval)] = ip; |
| 237 | ip++; |
247 | ip++; |
| 238 | } |
248 | } |
| 239 | while (len--); |
249 | while (len--); |
| 240 | #endif |
250 | #endif |
|
|
251 | |
| 241 | lit = 0; op++; /* start run */ |
252 | lit = 0; op++; /* start run */ |
| 242 | continue; |
|
|
| 243 | } |
253 | } |
| 244 | |
254 | else |
|
|
255 | { |
| 245 | /* one more literal byte we must copy */ |
256 | /* one more literal byte we must copy */ |
| 246 | |
|
|
| 247 | if (expect_false (op >= out_end)) |
257 | if (expect_false (op >= out_end)) |
| 248 | return 0; |
258 | return 0; |
| 249 | |
259 | |
| 250 | lit++; |
260 | lit++; *op++ = *ip++; |
|
|
261 | |
|
|
262 | if (expect_false (lit == MAX_LIT)) |
|
|
263 | { |
|
|
264 | op [- lit - 1] = lit - 1; /* stop run */ |
|
|
265 | lit = 0; op++; /* start run */ |
|
|
266 | } |
|
|
267 | } |
|
|
268 | } |
|
|
269 | |
|
|
270 | if (op + 3 > out_end) /* at most 3 bytes can be missing here */ |
|
|
271 | return 0; |
|
|
272 | |
|
|
273 | while (ip < in_end) |
|
|
274 | { |
| 251 | *op++ = *ip++; |
275 | lit++; *op++ = *ip++; |
| 252 | |
276 | |
| 253 | if (expect_false (lit == MAX_LIT)) |
277 | if (expect_false (lit == MAX_LIT)) |
| 254 | { |
278 | { |
| 255 | op [- lit - 1] = lit - 1; /* stop run */ |
279 | op [- lit - 1] = lit - 1; /* stop run */ |
| 256 | lit = 0; op++; /* start run */ |
280 | lit = 0; op++; /* start run */ |
| 257 | } |
281 | } |
| 258 | } |
282 | } |
| 259 | |
283 | |
| 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; |
284 | op [- lit - 1] = lit - 1; /* end run */ |
|
|
285 | op -= !lit; /* undo run if length is zero */ |
| 270 | |
286 | |
| 271 | return op - (u8 *)out_data; |
287 | return op - (u8 *)out_data; |
| 272 | } |
288 | } |
| 273 | |
289 | |
