/src/zlib-ng/deflate_medium.c
Line | Count | Source |
1 | | /* deflate_medium.c -- The deflate_medium deflate strategy |
2 | | * |
3 | | * Copyright (C) 2013 Intel Corporation. All rights reserved. |
4 | | * Authors: |
5 | | * Arjan van de Ven <arjan@linux.intel.com> |
6 | | * |
7 | | * For conditions of distribution and use, see copyright notice in zlib.h |
8 | | */ |
9 | | #ifndef NO_MEDIUM_STRATEGY |
10 | | #include "zbuild.h" |
11 | | #include "deflate.h" |
12 | | #include "deflate_p.h" |
13 | | #include "functable.h" |
14 | | |
15 | | struct match { |
16 | | uint16_t match_start; |
17 | | uint16_t match_length; |
18 | | uint16_t strstart; |
19 | | uint16_t orgstart; |
20 | | }; |
21 | | |
22 | 0 | static int emit_match(deflate_state *s, struct match match) { |
23 | 0 | int bflush = 0; |
24 | | |
25 | | /* matches that are not long enough we need to emit as literals */ |
26 | 0 | if (match.match_length < WANT_MIN_MATCH) { |
27 | 0 | while (match.match_length) { |
28 | 0 | bflush += zng_tr_tally_lit(s, s->window[match.strstart]); |
29 | 0 | s->lookahead--; |
30 | 0 | match.strstart++; |
31 | 0 | match.match_length--; |
32 | 0 | } |
33 | 0 | return bflush; |
34 | 0 | } |
35 | | |
36 | 0 | check_match(s, match.strstart, match.match_start, match.match_length); |
37 | |
|
38 | 0 | bflush += zng_tr_tally_dist(s, match.strstart - match.match_start, match.match_length - STD_MIN_MATCH); |
39 | |
|
40 | 0 | s->lookahead -= match.match_length; |
41 | 0 | return bflush; |
42 | 0 | } |
43 | | |
44 | 0 | static void insert_match(deflate_state *s, struct match match) { |
45 | 0 | if (UNLIKELY(s->lookahead <= (unsigned int)(match.match_length + WANT_MIN_MATCH))) |
46 | 0 | return; |
47 | | |
48 | | /* matches that are not long enough we need to emit as literals */ |
49 | 0 | if (LIKELY(match.match_length < WANT_MIN_MATCH)) { |
50 | 0 | match.strstart++; |
51 | 0 | match.match_length--; |
52 | 0 | if (UNLIKELY(match.match_length > 0)) { |
53 | 0 | if (match.strstart >= match.orgstart) { |
54 | 0 | if (match.strstart + match.match_length - 1 >= match.orgstart) { |
55 | 0 | insert_string(s, match.strstart, match.match_length); |
56 | 0 | } else { |
57 | 0 | insert_string(s, match.strstart, match.orgstart - match.strstart + 1); |
58 | 0 | } |
59 | 0 | match.strstart += match.match_length; |
60 | 0 | match.match_length = 0; |
61 | 0 | } |
62 | 0 | } |
63 | 0 | return; |
64 | 0 | } |
65 | | |
66 | | /* Insert new strings in the hash table only if the match length |
67 | | * is not too large. This saves time but degrades compression. |
68 | | */ |
69 | 0 | if (match.match_length <= 16 * s->max_insert_length && s->lookahead >= WANT_MIN_MATCH) { |
70 | 0 | match.match_length--; /* string at strstart already in table */ |
71 | 0 | match.strstart++; |
72 | |
|
73 | 0 | if (LIKELY(match.strstart >= match.orgstart)) { |
74 | 0 | if (LIKELY(match.strstart + match.match_length - 1 >= match.orgstart)) { |
75 | 0 | insert_string(s, match.strstart, match.match_length); |
76 | 0 | } else { |
77 | 0 | insert_string(s, match.strstart, match.orgstart - match.strstart + 1); |
78 | 0 | } |
79 | 0 | } else if (match.orgstart < match.strstart + match.match_length) { |
80 | 0 | insert_string(s, match.orgstart, match.strstart + match.match_length - match.orgstart); |
81 | 0 | } |
82 | 0 | match.strstart += match.match_length; |
83 | 0 | match.match_length = 0; |
84 | 0 | } else { |
85 | 0 | match.strstart += match.match_length; |
86 | 0 | match.match_length = 0; |
87 | |
|
88 | 0 | if (match.strstart >= (STD_MIN_MATCH - 2)) |
89 | 0 | quick_insert_string(s, match.strstart + 2 - STD_MIN_MATCH); |
90 | | |
91 | | /* If lookahead < WANT_MIN_MATCH, ins_h is garbage, but it does not |
92 | | * matter since it will be recomputed at next deflate call. |
93 | | */ |
94 | 0 | } |
95 | 0 | } |
96 | | |
97 | 0 | static void fizzle_matches(deflate_state *s, struct match *current, struct match *next) { |
98 | 0 | Pos limit; |
99 | 0 | unsigned char *match, *orig; |
100 | 0 | int changed = 0; |
101 | 0 | struct match c, n; |
102 | | /* step zero: sanity checks */ |
103 | |
|
104 | 0 | if (current->match_length <= 1) |
105 | 0 | return; |
106 | | |
107 | 0 | if (UNLIKELY(current->match_length > 1 + next->match_start)) |
108 | 0 | return; |
109 | | |
110 | 0 | if (UNLIKELY(current->match_length > 1 + next->strstart)) |
111 | 0 | return; |
112 | | |
113 | 0 | match = s->window - current->match_length + 1 + next->match_start; |
114 | 0 | orig = s->window - current->match_length + 1 + next->strstart; |
115 | | |
116 | | /* quick exit check.. if this fails then don't bother with anything else */ |
117 | 0 | if (LIKELY(*match != *orig)) |
118 | 0 | return; |
119 | | |
120 | 0 | c = *current; |
121 | 0 | n = *next; |
122 | | |
123 | | /* step one: try to move the "next" match to the left as much as possible */ |
124 | 0 | limit = next->strstart > MAX_DIST(s) ? next->strstart - (Pos)MAX_DIST(s) : 0; |
125 | |
|
126 | 0 | match = s->window + n.match_start - 1; |
127 | 0 | orig = s->window + n.strstart - 1; |
128 | |
|
129 | 0 | while (*match == *orig) { |
130 | 0 | if (UNLIKELY(c.match_length < 1)) |
131 | 0 | break; |
132 | 0 | if (UNLIKELY(n.strstart <= limit)) |
133 | 0 | break; |
134 | 0 | if (UNLIKELY(n.match_length >= 256)) |
135 | 0 | break; |
136 | 0 | if (UNLIKELY(n.match_start <= 1)) |
137 | 0 | break; |
138 | | |
139 | 0 | n.strstart--; |
140 | 0 | n.match_start--; |
141 | 0 | n.match_length++; |
142 | 0 | c.match_length--; |
143 | 0 | match--; |
144 | 0 | orig--; |
145 | 0 | changed++; |
146 | 0 | } |
147 | |
|
148 | 0 | if (!changed) |
149 | 0 | return; |
150 | | |
151 | 0 | if (c.match_length <= 1 && n.match_length != 2) { |
152 | 0 | n.orgstart++; |
153 | 0 | *current = c; |
154 | 0 | *next = n; |
155 | 0 | } else { |
156 | 0 | return; |
157 | 0 | } |
158 | 0 | } |
159 | | |
160 | 0 | Z_INTERNAL block_state deflate_medium(deflate_state *s, int flush) { |
161 | | /* Align the first struct to start on a new cacheline, this allows us to fit both structs in one cacheline */ |
162 | 0 | ALIGNED_(16) struct match current_match; |
163 | 0 | struct match next_match; |
164 | | |
165 | | /* For levels below 5, don't check the next position for a better match */ |
166 | 0 | int early_exit = s->level < 5; |
167 | |
|
168 | 0 | memset(¤t_match, 0, sizeof(struct match)); |
169 | 0 | memset(&next_match, 0, sizeof(struct match)); |
170 | |
|
171 | 0 | for (;;) { |
172 | 0 | Pos hash_head = 0; /* head of the hash chain */ |
173 | 0 | int bflush = 0; /* set if current block must be flushed */ |
174 | 0 | int64_t dist; |
175 | | |
176 | | /* Make sure that we always have enough lookahead, except |
177 | | * at the end of the input file. We need STD_MAX_MATCH bytes |
178 | | * for the next match, plus WANT_MIN_MATCH bytes to insert the |
179 | | * string following the next current_match. |
180 | | */ |
181 | 0 | if (s->lookahead < MIN_LOOKAHEAD) { |
182 | 0 | PREFIX(fill_window)(s); |
183 | 0 | if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) { |
184 | 0 | return need_more; |
185 | 0 | } |
186 | 0 | if (UNLIKELY(s->lookahead == 0)) |
187 | 0 | break; /* flush the current block */ |
188 | 0 | next_match.match_length = 0; |
189 | 0 | } |
190 | | |
191 | | /* Insert the string window[strstart .. strstart+2] in the |
192 | | * dictionary, and set hash_head to the head of the hash chain: |
193 | | */ |
194 | | |
195 | | /* If we already have a future match from a previous round, just use that */ |
196 | 0 | if (!early_exit && next_match.match_length > 0) { |
197 | 0 | current_match = next_match; |
198 | 0 | next_match.match_length = 0; |
199 | 0 | } else { |
200 | 0 | hash_head = 0; |
201 | 0 | if (s->lookahead >= WANT_MIN_MATCH) { |
202 | 0 | hash_head = quick_insert_string(s, s->strstart); |
203 | 0 | } |
204 | |
|
205 | 0 | current_match.strstart = (uint16_t)s->strstart; |
206 | 0 | current_match.orgstart = current_match.strstart; |
207 | | |
208 | | /* Find the longest match, discarding those <= prev_length. |
209 | | * At this point we have always match_length < WANT_MIN_MATCH |
210 | | */ |
211 | |
|
212 | 0 | dist = (int64_t)s->strstart - hash_head; |
213 | 0 | if (dist <= MAX_DIST(s) && dist > 0 && hash_head != 0) { |
214 | | /* To simplify the code, we prevent matches with the string |
215 | | * of window index 0 (in particular we have to avoid a match |
216 | | * of the string with itself at the start of the input file). |
217 | | */ |
218 | 0 | current_match.match_length = (uint16_t)FUNCTABLE_CALL(longest_match)(s, hash_head); |
219 | 0 | current_match.match_start = (uint16_t)s->match_start; |
220 | 0 | if (UNLIKELY(current_match.match_length < WANT_MIN_MATCH)) |
221 | 0 | current_match.match_length = 1; |
222 | 0 | if (UNLIKELY(current_match.match_start >= current_match.strstart)) { |
223 | | /* this can happen due to some restarts */ |
224 | 0 | current_match.match_length = 1; |
225 | 0 | } |
226 | 0 | } else { |
227 | | /* Set up the match to be a 1 byte literal */ |
228 | 0 | current_match.match_start = 0; |
229 | 0 | current_match.match_length = 1; |
230 | 0 | } |
231 | 0 | } |
232 | |
|
233 | 0 | insert_match(s, current_match); |
234 | | |
235 | | /* now, look ahead one */ |
236 | 0 | if (LIKELY(!early_exit && s->lookahead > MIN_LOOKAHEAD && (uint32_t)(current_match.strstart + current_match.match_length) < (s->window_size - MIN_LOOKAHEAD))) { |
237 | 0 | s->strstart = current_match.strstart + current_match.match_length; |
238 | 0 | hash_head = quick_insert_string(s, s->strstart); |
239 | |
|
240 | 0 | next_match.strstart = (uint16_t)s->strstart; |
241 | 0 | next_match.orgstart = next_match.strstart; |
242 | | |
243 | | /* Find the longest match, discarding those <= prev_length. |
244 | | * At this point we have always match_length < WANT_MIN_MATCH |
245 | | */ |
246 | |
|
247 | 0 | dist = (int64_t)s->strstart - hash_head; |
248 | 0 | if (dist <= MAX_DIST(s) && dist > 0 && hash_head != 0) { |
249 | | /* To simplify the code, we prevent matches with the string |
250 | | * of window index 0 (in particular we have to avoid a match |
251 | | * of the string with itself at the start of the input file). |
252 | | */ |
253 | 0 | next_match.match_length = (uint16_t)FUNCTABLE_CALL(longest_match)(s, hash_head); |
254 | 0 | next_match.match_start = (uint16_t)s->match_start; |
255 | 0 | if (UNLIKELY(next_match.match_start >= next_match.strstart)) { |
256 | | /* this can happen due to some restarts */ |
257 | 0 | next_match.match_length = 1; |
258 | 0 | } |
259 | 0 | if (next_match.match_length < WANT_MIN_MATCH) |
260 | 0 | next_match.match_length = 1; |
261 | 0 | else |
262 | 0 | fizzle_matches(s, ¤t_match, &next_match); |
263 | 0 | } else { |
264 | | /* Set up the match to be a 1 byte literal */ |
265 | 0 | next_match.match_start = 0; |
266 | 0 | next_match.match_length = 1; |
267 | 0 | } |
268 | |
|
269 | 0 | s->strstart = current_match.strstart; |
270 | 0 | } else { |
271 | 0 | next_match.match_length = 0; |
272 | 0 | } |
273 | | |
274 | | /* now emit the current match */ |
275 | 0 | bflush = emit_match(s, current_match); |
276 | | |
277 | | /* move the "cursor" forward */ |
278 | 0 | s->strstart += current_match.match_length; |
279 | |
|
280 | 0 | if (UNLIKELY(bflush)) |
281 | 0 | FLUSH_BLOCK(s, 0); |
282 | 0 | } |
283 | 0 | s->insert = s->strstart < (STD_MIN_MATCH - 1) ? s->strstart : (STD_MIN_MATCH - 1); |
284 | 0 | if (flush == Z_FINISH) { |
285 | 0 | FLUSH_BLOCK(s, 1); |
286 | 0 | return finish_done; |
287 | 0 | } |
288 | 0 | if (UNLIKELY(s->sym_next)) |
289 | 0 | FLUSH_BLOCK(s, 0); |
290 | |
|
291 | 0 | return block_done; |
292 | 0 | } |
293 | | #endif |