| 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 "match_p.h" |
| 14 | #include "functable.h" |
| 15 | |
| 16 | struct match { |
| 17 | unsigned int match_start; |
| 18 | unsigned int match_length; |
| 19 | unsigned int strstart; |
| 20 | unsigned int orgstart; |
| 21 | }; |
| 22 | |
| 23 | #define MAX_DIST2 ((1 << MAX_WBITS) - MIN_LOOKAHEAD) |
| 24 | |
| 25 | static int tr_tally_dist(deflate_state *s, int distance, int length) { |
| 26 | return zng_tr_tally(s, distance, length); |
| 27 | } |
| 28 | |
| 29 | static int tr_tally_lit(deflate_state *s, int c) { |
| 30 | return zng_tr_tally(s, 0, c); |
| 31 | } |
| 32 | |
| 33 | static int emit_match(deflate_state *s, struct match match) { |
| 34 | int flush = 0; |
| 35 | |
| 36 | /* matches that are not long enough we need to emit as literals */ |
| 37 | if (match.match_length < MIN_MATCH) { |
| 38 | while (match.match_length) { |
| 39 | flush += tr_tally_lit(s, s->window[match.strstart]); |
| 40 | s->lookahead--; |
| 41 | match.strstart++; |
| 42 | match.match_length--; |
| 43 | } |
| 44 | return flush; |
| 45 | } |
| 46 | |
| 47 | check_match(s, match.strstart, match.match_start, match.match_length); |
| 48 | |
| 49 | flush += tr_tally_dist(s, match.strstart - match.match_start, match.match_length - MIN_MATCH); |
| 50 | |
| 51 | s->lookahead -= match.match_length; |
| 52 | return flush; |
| 53 | } |
| 54 | |
| 55 | static void insert_match(deflate_state *s, struct match match) { |
| 56 | if (UNLIKELY(s->lookahead <= match.match_length + MIN_MATCH)) |
| 57 | return; |
| 58 | |
| 59 | /* matches that are not long enough we need to emit as literals */ |
| 60 | if (match.match_length < MIN_MATCH) { |
| 61 | #ifdef NOT_TWEAK_COMPILER |
| 62 | while (match.match_length) { |
| 63 | match.strstart++; |
| 64 | match.match_length--; |
| 65 | |
| 66 | if (match.match_length) { |
| 67 | if (match.strstart >= match.orgstart) { |
| 68 | functable.insert_string(s, match.strstart, 1); |
| 69 | } |
| 70 | } |
| 71 | } |
| 72 | #else |
| 73 | match.strstart++; |
| 74 | match.match_length--; |
| 75 | if (match.match_length > 0) { |
| 76 | if (match.strstart >= match.orgstart) { |
| 77 | if (match.strstart + match.match_length - 1 >= match.orgstart) { |
| 78 | functable.insert_string(s, match.strstart, match.match_length); |
| 79 | } else { |
| 80 | functable.insert_string(s, match.strstart, match.orgstart - match.strstart + 1); |
| 81 | } |
| 82 | match.strstart += match.match_length; |
| 83 | match.match_length = 0; |
| 84 | } |
| 85 | } |
| 86 | #endif |
| 87 | return; |
| 88 | } |
| 89 | |
| 90 | /* Insert new strings in the hash table only if the match length |
| 91 | * is not too large. This saves time but degrades compression. |
| 92 | */ |
| 93 | if (match.match_length <= 16* s->max_insert_length && s->lookahead >= MIN_MATCH) { |
| 94 | match.match_length--; /* string at strstart already in table */ |
| 95 | match.strstart++; |
| 96 | #ifdef NOT_TWEAK_COMPILER |
| 97 | do { |
| 98 | if (LIKELY(match.strstart >= match.orgstart)) { |
| 99 | functable.insert_string(s, match.strstart, 1); |
| 100 | } |
| 101 | match.strstart++; |
| 102 | /* strstart never exceeds WSIZE-MAX_MATCH, so there are |
| 103 | * always MIN_MATCH bytes ahead. |
| 104 | */ |
| 105 | } while (--match.match_length != 0); |
| 106 | #else |
| 107 | if (LIKELY(match.strstart >= match.orgstart)) { |
| 108 | if (LIKELY(match.strstart + match.match_length - 1 >= match.orgstart)) { |
| 109 | functable.insert_string(s, match.strstart, match.match_length); |
| 110 | } else { |
| 111 | functable.insert_string(s, match.strstart, match.orgstart - match.strstart + 1); |
| 112 | } |
| 113 | } else if (match.orgstart < match.strstart + match.match_length) { |
| 114 | functable.insert_string(s, match.orgstart, match.strstart + match.match_length - match.orgstart); |
| 115 | } |
| 116 | match.strstart += match.match_length; |
| 117 | match.match_length = 0; |
| 118 | #endif |
| 119 | } else { |
| 120 | match.strstart += match.match_length; |
| 121 | match.match_length = 0; |
| 122 | s->ins_h = s->window[match.strstart]; |
| 123 | if (match.strstart >= (MIN_MATCH - 2)) |
| 124 | #ifndef NOT_TWEAK_COMPILER |
| 125 | functable.insert_string(s, match.strstart + 2 - MIN_MATCH, MIN_MATCH - 2); |
| 126 | #else |
| 127 | functable.insert_string(s, match.strstart + 2 - MIN_MATCH, 1); |
| 128 | #if MIN_MATCH != 3 |
| 129 | #warning Call insert_string() MIN_MATCH-3 more times |
| 130 | #endif |
| 131 | #endif |
| 132 | /* If lookahead < MIN_MATCH, ins_h is garbage, but it does not |
| 133 | * matter since it will be recomputed at next deflate call. |
| 134 | */ |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | static void fizzle_matches(deflate_state *s, struct match *current, struct match *next) { |
| 139 | IPos limit; |
| 140 | unsigned char *match, *orig; |
| 141 | int changed = 0; |
| 142 | struct match c, n; |
| 143 | /* step zero: sanity checks */ |
| 144 | |
| 145 | if (current->match_length <= 1) |
| 146 | return; |
| 147 | |
| 148 | if (UNLIKELY(current->match_length > 1 + next->match_start)) |
| 149 | return; |
| 150 | |
| 151 | if (UNLIKELY(current->match_length > 1 + next->strstart)) |
| 152 | return; |
| 153 | |
| 154 | match = s->window - current->match_length + 1 + next->match_start; |
| 155 | orig = s->window - current->match_length + 1 + next->strstart; |
| 156 | |
| 157 | /* quick exit check.. if this fails then don't bother with anything else */ |
| 158 | if (LIKELY(*match != *orig)) |
| 159 | return; |
| 160 | |
| 161 | c = *current; |
| 162 | n = *next; |
| 163 | |
| 164 | /* step one: try to move the "next" match to the left as much as possible */ |
| 165 | limit = next->strstart > MAX_DIST2 ? next->strstart - MAX_DIST2 : 0; |
| 166 | |
| 167 | match = s->window + n.match_start - 1; |
| 168 | orig = s->window + n.strstart - 1; |
| 169 | |
| 170 | while (*match == *orig) { |
| 171 | if (c.match_length < 1) |
| 172 | break; |
| 173 | if (n.strstart <= limit) |
| 174 | break; |
| 175 | if (n.match_length >= 256) |
| 176 | break; |
| 177 | if (n.match_start <= 1) |
| 178 | break; |
| 179 | |
| 180 | n.strstart--; |
| 181 | n.match_start--; |
| 182 | n.match_length++; |
| 183 | c.match_length--; |
| 184 | match--; |
| 185 | orig--; |
| 186 | changed++; |
| 187 | } |
| 188 | |
| 189 | if (!changed) |
| 190 | return; |
| 191 | |
| 192 | if (c.match_length <= 1 && n.match_length != 2) { |
| 193 | n.orgstart++; |
| 194 | *current = c; |
| 195 | *next = n; |
| 196 | } else { |
| 197 | return; |
| 198 | } |
| 199 | } |
| 200 | |
| 201 | ZLIB_INTERNAL block_state deflate_medium(deflate_state *s, int flush) { |
| 202 | struct match current_match, next_match; |
| 203 | |
| 204 | memset(¤t_match, 0, sizeof(struct match)); |
| 205 | memset(&next_match, 0, sizeof(struct match)); |
| 206 | |
| 207 | for (;;) { |
| 208 | IPos hash_head = 0; /* head of the hash chain */ |
| 209 | int bflush; /* set if current block must be flushed */ |
| 210 | |
| 211 | /* Make sure that we always have enough lookahead, except |
| 212 | * at the end of the input file. We need MAX_MATCH bytes |
| 213 | * for the next match, plus MIN_MATCH bytes to insert the |
| 214 | * string following the next current_match. |
| 215 | */ |
| 216 | if (s->lookahead < MIN_LOOKAHEAD) { |
| 217 | functable.fill_window(s); |
| 218 | if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) { |
| 219 | return need_more; |
| 220 | } |
| 221 | if (s->lookahead == 0) |
| 222 | break; /* flush the current block */ |
| 223 | next_match.match_length = 0; |
| 224 | } |
| 225 | s->prev_length = 2; |
| 226 | |
| 227 | /* Insert the string window[strstart .. strstart+2] in the |
| 228 | * dictionary, and set hash_head to the head of the hash chain: |
| 229 | */ |
| 230 | |
| 231 | /* If we already have a future match from a previous round, just use that */ |
| 232 | if (next_match.match_length > 0) { |
| 233 | current_match = next_match; |
| 234 | next_match.match_length = 0; |
| 235 | |
| 236 | } else { |
| 237 | hash_head = 0; |
| 238 | if (s->lookahead >= MIN_MATCH) { |
| 239 | hash_head = functable.insert_string(s, s->strstart, 1); |
| 240 | } |
| 241 | |
| 242 | /* set up the initial match to be a 1 byte literal */ |
| 243 | current_match.match_start = 0; |
| 244 | current_match.match_length = 1; |
| 245 | current_match.strstart = s->strstart; |
| 246 | current_match.orgstart = current_match.strstart; |
| 247 | |
| 248 | /* Find the longest match, discarding those <= prev_length. |
| 249 | * At this point we have always match_length < MIN_MATCH |
| 250 | */ |
| 251 | |
| 252 | if (hash_head != 0 && s->strstart - hash_head <= MAX_DIST2) { |
| 253 | /* To simplify the code, we prevent matches with the string |
| 254 | * of window index 0 (in particular we have to avoid a match |
| 255 | * of the string with itself at the start of the input file). |
| 256 | */ |
| 257 | current_match.match_length = longest_match(s, hash_head); |
| 258 | current_match.match_start = s->match_start; |
| 259 | if (current_match.match_length < MIN_MATCH) |
| 260 | current_match.match_length = 1; |
| 261 | if (current_match.match_start >= current_match.strstart) { |
| 262 | /* this can happen due to some restarts */ |
| 263 | current_match.match_length = 1; |
| 264 | } |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | insert_match(s, current_match); |
| 269 | |
| 270 | /* now, look ahead one */ |
| 271 | if (s->lookahead > MIN_LOOKAHEAD && (current_match.strstart + current_match.match_length) < (s->window_size - MIN_LOOKAHEAD)) { |
| 272 | s->strstart = current_match.strstart + current_match.match_length; |
| 273 | hash_head = functable.insert_string(s, s->strstart, 1); |
| 274 | |
| 275 | /* set up the initial match to be a 1 byte literal */ |
| 276 | next_match.match_start = 0; |
| 277 | next_match.match_length = 1; |
| 278 | next_match.strstart = s->strstart; |
| 279 | next_match.orgstart = next_match.strstart; |
| 280 | |
| 281 | /* Find the longest match, discarding those <= prev_length. |
| 282 | * At this point we have always match_length < MIN_MATCH |
| 283 | */ |
| 284 | if (hash_head != 0 && s->strstart - hash_head <= MAX_DIST2) { |
| 285 | /* To simplify the code, we prevent matches with the string |
| 286 | * of window index 0 (in particular we have to avoid a match |
| 287 | * of the string with itself at the start of the input file). |
| 288 | */ |
| 289 | next_match.match_length = longest_match(s, hash_head); |
| 290 | next_match.match_start = s->match_start; |
| 291 | if (next_match.match_start >= next_match.strstart) { |
| 292 | /* this can happen due to some restarts */ |
| 293 | next_match.match_length = 1; |
| 294 | } |
| 295 | if (next_match.match_length < MIN_MATCH) |
| 296 | next_match.match_length = 1; |
| 297 | else |
| 298 | fizzle_matches(s, ¤t_match, &next_match); |
| 299 | } |
| 300 | |
| 301 | /* short matches with a very long distance are rarely a good idea encoding wise */ |
| 302 | if (next_match.match_length == 3 && (next_match.strstart - next_match.match_start) > 12000) |
| 303 | next_match.match_length = 1; |
| 304 | s->strstart = current_match.strstart; |
| 305 | |
| 306 | } else { |
| 307 | next_match.match_length = 0; |
| 308 | } |
| 309 | |
| 310 | /* now emit the current match */ |
| 311 | bflush = emit_match(s, current_match); |
| 312 | |
| 313 | /* move the "cursor" forward */ |
| 314 | s->strstart += current_match.match_length; |
| 315 | |
| 316 | if (bflush) |
| 317 | FLUSH_BLOCK(s, 0); |
| 318 | } |
| 319 | s->insert = s->strstart < MIN_MATCH-1 ? s->strstart : MIN_MATCH-1; |
| 320 | if (flush == Z_FINISH) { |
| 321 | FLUSH_BLOCK(s, 1); |
| 322 | return finish_done; |
| 323 | } |
| 324 | if (s->sym_next) |
| 325 | FLUSH_BLOCK(s, 0); |
| 326 | |
| 327 | return block_done; |
| 328 | } |
| 329 | #endif |
| 330 | |