7 * The author dedicates this file to the public domain.
8 * You can do whatever you want with this file.
10 * ----------------------------------------------------------------------------
12 * This is a Binary Trees (bt) based matchfinder.
14 * The main data structure is a hash table where each hash bucket contains a
15 * binary tree of sequences whose first 4 bytes share the same hash code. Each
16 * sequence is identified by its starting position in the input buffer. Each
17 * binary tree is always sorted such that each left child represents a sequence
18 * lexicographically lesser than its parent and each right child represents a
19 * sequence lexicographically greater than its parent.
21 * The algorithm processes the input buffer sequentially. At each byte
22 * position, the hash code of the first 4 bytes of the sequence beginning at
23 * that position (the sequence being matched against) is computed. This
24 * identifies the hash bucket to use for that position. Then, a new binary tree
25 * node is created to represent the current sequence. Then, in a single tree
26 * traversal, the hash bucket's binary tree is searched for matches and is
27 * re-rooted at the new node.
29 * Compared to the simpler algorithm that uses linked lists instead of binary
30 * trees (see hc_matchfinder.h), the binary tree version gains more information
31 * at each node visitation. Ideally, the binary tree version will examine only
32 * 'log(n)' nodes to find the same matches that the linked list version will
33 * find by examining 'n' nodes. In addition, the binary tree version can
34 * examine fewer bytes at each node by taking advantage of the common prefixes
35 * that result from the sort order, whereas the linked list version may have to
36 * examine up to the full length of the match at each node.
38 * However, it is not always best to use the binary tree version. It requires
39 * nearly twice as much memory as the linked list version, and it takes time to
40 * keep the binary trees sorted, even at positions where the compressor does not
41 * need matches. Generally, when doing fast compression on small buffers,
42 * binary trees are the wrong approach. They are best suited for thorough
43 * compression and/or large buffers.
45 * ----------------------------------------------------------------------------
51 #include "wimlib/lz_extend.h"
52 #include "wimlib/lz_hash.h"
54 #define BT_MATCHFINDER_HASH3_ORDER 15
55 #define BT_MATCHFINDER_HASH3_WAYS 2
56 #define BT_MATCHFINDER_HASH4_ORDER 16
58 /* TEMPLATED functions and structures have MF_SUFFIX appended to their name. */
60 #define TEMPLATED(name) CONCAT(name, MF_SUFFIX)
62 #ifndef _WIMLIB_BT_MATCHFINDER_H
63 #define _WIMLIB_BT_MATCHFINDER_H
65 /* Non-templated definitions */
67 /* Representation of a match found by the bt_matchfinder */
70 /* The number of bytes matched. */
73 /* The offset back from the current position that was matched. */
77 #endif /* _WIMLIB_BT_MATCHFINDER_H */
79 struct TEMPLATED(bt_matchfinder) {
81 /* The hash table for finding length 2 matches, if enabled */
82 #ifdef BT_MATCHFINDER_HASH2_ORDER
83 mf_pos_t hash2_tab[1UL << BT_MATCHFINDER_HASH2_ORDER];
86 /* The hash table for finding length 3 matches */
87 mf_pos_t hash3_tab[1UL << BT_MATCHFINDER_HASH3_ORDER][BT_MATCHFINDER_HASH3_WAYS];
89 /* The hash table which contains the roots of the binary trees for
90 * finding length 4+ matches */
91 mf_pos_t hash4_tab[1UL << BT_MATCHFINDER_HASH4_ORDER];
93 /* The child node references for the binary trees. The left and right
94 * children of the node for the sequence with position 'pos' are
95 * 'child_tab[pos * 2]' and 'child_tab[pos * 2 + 1]', respectively. */
99 /* Return the number of bytes that must be allocated for a 'bt_matchfinder' that
100 * can work with buffers up to the specified size. */
102 TEMPLATED(bt_matchfinder_size)(size_t max_bufsize)
104 return sizeof(struct TEMPLATED(bt_matchfinder)) +
105 (2 * max_bufsize * sizeof(mf_pos_t));
108 /* Prepare the matchfinder for a new input buffer. */
110 TEMPLATED(bt_matchfinder_init)(struct TEMPLATED(bt_matchfinder) *mf)
112 memset(mf, 0, sizeof(*mf));
115 static inline mf_pos_t *
116 TEMPLATED(bt_left_child)(struct TEMPLATED(bt_matchfinder) *mf, u32 node)
118 return &mf->child_tab[(node << 1) + 0];
121 static inline mf_pos_t *
122 TEMPLATED(bt_right_child)(struct TEMPLATED(bt_matchfinder) *mf, u32 node)
124 return &mf->child_tab[(node << 1) + 1];
127 /* The minimum permissible value of 'max_len' for bt_matchfinder_get_matches()
128 * and bt_matchfinder_skip_position(). There must be sufficiently many bytes
129 * remaining to load a 32-bit integer from the *next* position. */
130 #define BT_MATCHFINDER_REQUIRED_NBYTES 5
132 /* Advance the binary tree matchfinder by one byte, optionally recording
133 * matches. @record_matches should be a compile-time constant. */
134 static inline struct lz_match *
135 TEMPLATED(bt_matchfinder_advance_one_byte)(struct TEMPLATED(bt_matchfinder) * const restrict mf,
136 const u8 * const restrict in_begin,
137 const ptrdiff_t cur_pos,
140 const u32 max_search_depth,
141 u32 next_hashes[const restrict static 2],
142 u32 * const restrict best_len_ret,
143 struct lz_match * restrict lz_matchptr,
144 const bool record_matches)
146 const u8 *in_next = in_begin + cur_pos;
147 u32 depth_remaining = max_search_depth;
152 #ifdef BT_MATCHFINDER_HASH2_ORDER
156 STATIC_ASSERT(BT_MATCHFINDER_HASH3_WAYS >= 1 &&
157 BT_MATCHFINDER_HASH3_WAYS <= 2);
159 #if BT_MATCHFINDER_HASH3_WAYS >= 2
163 mf_pos_t *pending_lt_ptr, *pending_gt_ptr;
164 u32 best_lt_len, best_gt_len;
168 next_seq4 = load_u32_unaligned(in_next + 1);
169 next_seq3 = loaded_u32_to_u24(next_seq4);
171 hash3 = next_hashes[0];
172 hash4 = next_hashes[1];
174 next_hashes[0] = lz_hash(next_seq3, BT_MATCHFINDER_HASH3_ORDER);
175 next_hashes[1] = lz_hash(next_seq4, BT_MATCHFINDER_HASH4_ORDER);
176 prefetchw(&mf->hash3_tab[next_hashes[0]]);
177 prefetchw(&mf->hash4_tab[next_hashes[1]]);
179 #ifdef BT_MATCHFINDER_HASH2_ORDER
180 seq2 = load_u16_unaligned(in_next);
181 hash2 = lz_hash(seq2, BT_MATCHFINDER_HASH2_ORDER);
182 cur_node = mf->hash2_tab[hash2];
183 mf->hash2_tab[hash2] = cur_pos;
184 if (record_matches &&
185 seq2 == load_u16_unaligned(&in_begin[cur_node]) &&
186 likely(in_next != in_begin))
188 lz_matchptr->length = 2;
189 lz_matchptr->offset = in_next - &in_begin[cur_node];
194 cur_node = mf->hash3_tab[hash3][0];
195 mf->hash3_tab[hash3][0] = cur_pos;
196 #if BT_MATCHFINDER_HASH3_WAYS >= 2
197 cur_node_2 = mf->hash3_tab[hash3][1];
198 mf->hash3_tab[hash3][1] = cur_node;
200 if (record_matches && likely(in_next != in_begin)) {
201 u32 seq3 = load_u24_unaligned(in_next);
202 if (seq3 == load_u24_unaligned(&in_begin[cur_node])) {
203 lz_matchptr->length = 3;
204 lz_matchptr->offset = in_next - &in_begin[cur_node];
207 #if BT_MATCHFINDER_HASH3_WAYS >= 2
208 else if (seq3 == load_u24_unaligned(&in_begin[cur_node_2])) {
209 lz_matchptr->length = 3;
210 lz_matchptr->offset = in_next - &in_begin[cur_node_2];
216 cur_node = mf->hash4_tab[hash4];
217 mf->hash4_tab[hash4] = cur_pos;
219 pending_lt_ptr = TEMPLATED(bt_left_child)(mf, cur_pos);
220 pending_gt_ptr = TEMPLATED(bt_right_child)(mf, cur_pos);
225 *best_len_ret = best_len;
234 matchptr = &in_begin[cur_node];
236 if (matchptr[len] == in_next[len]) {
237 len = lz_extend(in_next, matchptr, len + 1,
238 (record_matches ? max_len : nice_len));
239 if (!record_matches || len > best_len) {
240 if (record_matches) {
242 lz_matchptr->length = len;
243 lz_matchptr->offset = in_next - matchptr;
246 if (len >= nice_len) {
247 *pending_lt_ptr = *TEMPLATED(bt_left_child)(mf, cur_node);
248 *pending_gt_ptr = *TEMPLATED(bt_right_child)(mf, cur_node);
249 *best_len_ret = best_len;
255 if (matchptr[len] < in_next[len]) {
256 *pending_lt_ptr = cur_node;
257 pending_lt_ptr = TEMPLATED(bt_right_child)(mf, cur_node);
258 cur_node = *pending_lt_ptr;
260 if (best_gt_len < len)
263 *pending_gt_ptr = cur_node;
264 pending_gt_ptr = TEMPLATED(bt_left_child)(mf, cur_node);
265 cur_node = *pending_gt_ptr;
267 if (best_lt_len < len)
271 if (!cur_node || !--depth_remaining) {
274 *best_len_ret = best_len;
281 * Retrieve a list of matches with the current position.
284 * The matchfinder structure.
286 * Pointer to the beginning of the input buffer.
288 * The current position in the input buffer (the position of the sequence
289 * being matched against).
291 * The maximum permissible match length at this position. Must be >=
292 * BT_MATCHFINDER_REQUIRED_NBYTES.
294 * Stop searching if a match of at least this length is found.
295 * Must be <= @max_len.
297 * Limit on the number of potential matches to consider. Must be >= 1.
299 * The precomputed hash codes for the sequence beginning at @in_next.
300 * These will be used and then updated with the precomputed hashcodes for
301 * the sequence beginning at @in_next + 1.
303 * If a match of length >= 4 was found, then the length of the longest such
304 * match is written here; otherwise 3 is written here. (Note: this is
305 * redundant with the 'struct lz_match' array, but this is easier for the
306 * compiler to optimize when inlined and the caller immediately does a
307 * check against 'best_len'.)
309 * An array in which this function will record the matches. The recorded
310 * matches will be sorted by strictly increasing length and (non-strictly)
311 * increasing offset. The maximum number of matches that may be found is
312 * 'nice_len - 1', or one less if length 2 matches are disabled.
314 * The return value is a pointer to the next available slot in the @lz_matchptr
315 * array. (If no matches were found, this will be the same as @lz_matchptr.)
317 static inline struct lz_match *
318 TEMPLATED(bt_matchfinder_get_matches)(struct TEMPLATED(bt_matchfinder) *mf,
323 u32 max_search_depth,
324 u32 next_hashes[static 2],
326 struct lz_match *lz_matchptr)
328 return TEMPLATED(bt_matchfinder_advance_one_byte)(mf,
341 * Advance the matchfinder, but don't record any matches.
343 * This is very similar to bt_matchfinder_get_matches() because both functions
344 * must do hashing and tree re-rooting.
347 TEMPLATED(bt_matchfinder_skip_position)(struct TEMPLATED(bt_matchfinder) *mf,
352 u32 max_search_depth,
353 u32 next_hashes[static 2])
356 TEMPLATED(bt_matchfinder_advance_one_byte)(mf,