1 /*
2  * hc_matchfinder.h - Lempel-Ziv matchfinding with a hash table of linked lists
3  *
4  * Originally public domain; changes after 2016-09-07 are copyrighted.
5  *
6  * Copyright 2016 Eric Biggers
7  *
8  * Permission is hereby granted, free of charge, to any person
9  * obtaining a copy of this software and associated documentation
10  * files (the "Software"), to deal in the Software without
11  * restriction, including without limitation the rights to use,
12  * copy, modify, merge, publish, distribute, sublicense, and/or sell
13  * copies of the Software, and to permit persons to whom the
14  * Software is furnished to do so, subject to the following
15  * conditions:
16  *
17  * The above copyright notice and this permission notice shall be
18  * included in all copies or substantial portions of the Software.
19  *
20  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
21  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
22  * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
23  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
24  * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
25  * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
26  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
27  * OTHER DEALINGS IN THE SOFTWARE.
28  *
29  * ---------------------------------------------------------------------------
30  *
31  *				   Algorithm
32  *
33  * This is a Hash Chains (hc) based matchfinder.
34  *
35  * The main data structure is a hash table where each hash bucket contains a
36  * linked list (or "chain") of sequences whose first 4 bytes share the same hash
37  * code.  Each sequence is identified by its starting position in the input
38  * buffer.
39  *
40  * The algorithm processes the input buffer sequentially.  At each byte
41  * position, the hash code of the first 4 bytes of the sequence beginning at
42  * that position (the sequence being matched against) is computed.  This
43  * identifies the hash bucket to use for that position.  Then, this hash
44  * bucket's linked list is searched for matches.  Then, a new linked list node
45  * is created to represent the current sequence and is prepended to the list.
46  *
47  * This algorithm has several useful properties:
48  *
49  * - It only finds true Lempel-Ziv matches; i.e., those where the matching
50  *   sequence occurs prior to the sequence being matched against.
51  *
52  * - The sequences in each linked list are always sorted by decreasing starting
53  *   position.  Therefore, the closest (smallest offset) matches are found
54  *   first, which in many compression formats tend to be the cheapest to encode.
55  *
56  * - Although fast running time is not guaranteed due to the possibility of the
57  *   lists getting very long, the worst degenerate behavior can be easily
58  *   prevented by capping the number of nodes searched at each position.
59  *
60  * - If the compressor decides not to search for matches at a certain position,
61  *   then that position can be quickly inserted without searching the list.
62  *
63  * - The algorithm is adaptable to sliding windows: just store the positions
64  *   relative to a "base" value that is updated from time to time, and stop
65  *   searching each list when the sequences get too far away.
66  *
67  * ----------------------------------------------------------------------------
68  *
69  *				 Optimizations
70  *
71  * The main hash table and chains handle length 4+ matches.  Length 3 matches
72  * are handled by a separate hash table with no chains.  This works well for
73  * typical "greedy" or "lazy"-style compressors, where length 3 matches are
74  * often only helpful if they have small offsets.  Instead of searching a full
75  * chain for length 3+ matches, the algorithm just checks for one close length 3
76  * match, then focuses on finding length 4+ matches.
77  *
78  * The longest_match() and skip_positions() functions are inlined into the
79  * compressors that use them.  This isn't just about saving the overhead of a
80  * function call.  These functions are intended to be called from the inner
81  * loops of compressors, where giving the compiler more control over register
82  * allocation is very helpful.  There is also significant benefit to be gained
83  * from allowing the CPU to predict branches independently at each call site.
84  * For example, "lazy"-style compressors can be written with two calls to
85  * longest_match(), each of which starts with a different 'best_len' and
86  * therefore has significantly different performance characteristics.
87  *
88  * Although any hash function can be used, a multiplicative hash is fast and
89  * works well.
90  *
91  * On some processors, it is significantly faster to extend matches by whole
92  * words (32 or 64 bits) instead of by individual bytes.  For this to be the
93  * case, the processor must implement unaligned memory accesses efficiently and
94  * must have either a fast "find first set bit" instruction or a fast "find last
95  * set bit" instruction, depending on the processor's endianness.
96  *
97  * The code uses one loop for finding the first match and one loop for finding a
98  * longer match.  Each of these loops is tuned for its respective task and in
99  * combination are faster than a single generalized loop that handles both
100  * tasks.
101  *
102  * The code also uses a tight inner loop that only compares the last and first
103  * bytes of a potential match.  It is only when these bytes match that a full
104  * match extension is attempted.
105  *
106  * ----------------------------------------------------------------------------
107  */
108 
109 #include "matchfinder_common.h"
110 
111 #define HC_MATCHFINDER_HASH3_ORDER	15
112 #define HC_MATCHFINDER_HASH4_ORDER	16
113 
114 #define HC_MATCHFINDER_TOTAL_HASH_LENGTH		\
115 	((1UL << HC_MATCHFINDER_HASH3_ORDER) +		\
116 	 (1UL << HC_MATCHFINDER_HASH4_ORDER))
117 
118 struct hc_matchfinder {
119 
120 	/* The hash table for finding length 3 matches  */
121 	mf_pos_t hash3_tab[1UL << HC_MATCHFINDER_HASH3_ORDER];
122 
123 	/* The hash table which contains the first nodes of the linked lists for
124 	 * finding length 4+ matches  */
125 	mf_pos_t hash4_tab[1UL << HC_MATCHFINDER_HASH4_ORDER];
126 
127 	/* The "next node" references for the linked lists.  The "next node" of
128 	 * the node for the sequence with position 'pos' is 'next_tab[pos]'.  */
129 	mf_pos_t next_tab[MATCHFINDER_WINDOW_SIZE];
130 
131 }
132 #ifdef _aligned_attribute
133   _aligned_attribute(MATCHFINDER_ALIGNMENT)
134 #endif
135 ;
136 
137 /* Prepare the matchfinder for a new input buffer.  */
138 static forceinline void
hc_matchfinder_init(struct hc_matchfinder * mf)139 hc_matchfinder_init(struct hc_matchfinder *mf)
140 {
141 	matchfinder_init((mf_pos_t *)mf, HC_MATCHFINDER_TOTAL_HASH_LENGTH);
142 }
143 
144 static forceinline void
hc_matchfinder_slide_window(struct hc_matchfinder * mf)145 hc_matchfinder_slide_window(struct hc_matchfinder *mf)
146 {
147 	matchfinder_rebase((mf_pos_t *)mf,
148 			   sizeof(struct hc_matchfinder) / sizeof(mf_pos_t));
149 }
150 
151 /*
152  * Find the longest match longer than 'best_len' bytes.
153  *
154  * @mf
155  *	The matchfinder structure.
156  * @in_base_p
157  *	Location of a pointer which points to the place in the input data the
158  *	matchfinder currently stores positions relative to.  This may be updated
159  *	by this function.
160  * @cur_pos
161  *	The current position in the input buffer relative to @in_base (the
162  *	position of the sequence being matched against).
163  * @best_len
164  *	Require a match longer than this length.
165  * @max_len
166  *	The maximum permissible match length at this position.
167  * @nice_len
168  *	Stop searching if a match of at least this length is found.
169  *	Must be <= @max_len.
170  * @max_search_depth
171  *	Limit on the number of potential matches to consider.  Must be >= 1.
172  * @next_hashes
173  *	The precomputed hash codes for the sequence beginning at @in_next.
174  *	These will be used and then updated with the precomputed hashcodes for
175  *	the sequence beginning at @in_next + 1.
176  * @offset_ret
177  *	If a match is found, its offset is returned in this location.
178  *
179  * Return the length of the match found, or 'best_len' if no match longer than
180  * 'best_len' was found.
181  */
182 static forceinline u32
hc_matchfinder_longest_match(struct hc_matchfinder * const restrict mf,const u8 ** const restrict in_base_p,const u8 * const restrict in_next,u32 best_len,const u32 max_len,const u32 nice_len,const u32 max_search_depth,u32 * const restrict next_hashes,u32 * const restrict offset_ret)183 hc_matchfinder_longest_match(struct hc_matchfinder * const restrict mf,
184 			     const u8 ** const restrict in_base_p,
185 			     const u8 * const restrict in_next,
186 			     u32 best_len,
187 			     const u32 max_len,
188 			     const u32 nice_len,
189 			     const u32 max_search_depth,
190 			     u32 * const restrict next_hashes,
191 			     u32 * const restrict offset_ret)
192 {
193 	u32 depth_remaining = max_search_depth;
194 	const u8 *best_matchptr = in_next;
195 	mf_pos_t cur_node3, cur_node4;
196 	u32 hash3, hash4;
197 	u32 next_hashseq;
198 	u32 seq4;
199 	const u8 *matchptr;
200 	u32 len;
201 	u32 cur_pos = in_next - *in_base_p;
202 	const u8 *in_base;
203 	mf_pos_t cutoff;
204 
205 	if (cur_pos == MATCHFINDER_WINDOW_SIZE) {
206 		hc_matchfinder_slide_window(mf);
207 		*in_base_p += MATCHFINDER_WINDOW_SIZE;
208 		cur_pos = 0;
209 	}
210 
211 	in_base = *in_base_p;
212 	cutoff = cur_pos - MATCHFINDER_WINDOW_SIZE;
213 
214 	if (unlikely(max_len < 5)) /* can we read 4 bytes from 'in_next + 1'? */
215 		goto out;
216 
217 	/* Get the precomputed hash codes.  */
218 	hash3 = next_hashes[0];
219 	hash4 = next_hashes[1];
220 
221 	/* From the hash buckets, get the first node of each linked list.  */
222 	cur_node3 = mf->hash3_tab[hash3];
223 	cur_node4 = mf->hash4_tab[hash4];
224 
225 	/* Update for length 3 matches.  This replaces the singleton node in the
226 	 * 'hash3' bucket with the node for the current sequence.  */
227 	mf->hash3_tab[hash3] = cur_pos;
228 
229 	/* Update for length 4 matches.  This prepends the node for the current
230 	 * sequence to the linked list in the 'hash4' bucket.  */
231 	mf->hash4_tab[hash4] = cur_pos;
232 	mf->next_tab[cur_pos] = cur_node4;
233 
234 	/* Compute the next hash codes.  */
235 	next_hashseq = get_unaligned_le32(in_next + 1);
236 	next_hashes[0] = lz_hash(next_hashseq & 0xFFFFFF, HC_MATCHFINDER_HASH3_ORDER);
237 	next_hashes[1] = lz_hash(next_hashseq, HC_MATCHFINDER_HASH4_ORDER);
238 	prefetchw(&mf->hash3_tab[next_hashes[0]]);
239 	prefetchw(&mf->hash4_tab[next_hashes[1]]);
240 
241 	if (best_len < 4) {  /* No match of length >= 4 found yet?  */
242 
243 		/* Check for a length 3 match if needed.  */
244 
245 		if (cur_node3 <= cutoff)
246 			goto out;
247 
248 		seq4 = load_u32_unaligned(in_next);
249 
250 		if (best_len < 3) {
251 			matchptr = &in_base[cur_node3];
252 			if (load_u24_unaligned(matchptr) == loaded_u32_to_u24(seq4)) {
253 				best_len = 3;
254 				best_matchptr = matchptr;
255 			}
256 		}
257 
258 		/* Check for a length 4 match.  */
259 
260 		if (cur_node4 <= cutoff)
261 			goto out;
262 
263 		for (;;) {
264 			/* No length 4 match found yet.  Check the first 4 bytes.  */
265 			matchptr = &in_base[cur_node4];
266 
267 			if (load_u32_unaligned(matchptr) == seq4)
268 				break;
269 
270 			/* The first 4 bytes did not match.  Keep trying.  */
271 			cur_node4 = mf->next_tab[cur_node4 & (MATCHFINDER_WINDOW_SIZE - 1)];
272 			if (cur_node4 <= cutoff || !--depth_remaining)
273 				goto out;
274 		}
275 
276 		/* Found a match of length >= 4.  Extend it to its full length.  */
277 		best_matchptr = matchptr;
278 		best_len = lz_extend(in_next, best_matchptr, 4, max_len);
279 		if (best_len >= nice_len)
280 			goto out;
281 		cur_node4 = mf->next_tab[cur_node4 & (MATCHFINDER_WINDOW_SIZE - 1)];
282 		if (cur_node4 <= cutoff || !--depth_remaining)
283 			goto out;
284 	} else {
285 		if (cur_node4 <= cutoff || best_len >= nice_len)
286 			goto out;
287 	}
288 
289 	/* Check for matches of length >= 5.  */
290 
291 	for (;;) {
292 		for (;;) {
293 			matchptr = &in_base[cur_node4];
294 
295 			/* Already found a length 4 match.  Try for a longer
296 			 * match; start by checking either the last 4 bytes and
297 			 * the first 4 bytes, or the last byte.  (The last byte,
298 			 * the one which would extend the match length by 1, is
299 			 * the most important.)  */
300 		#if UNALIGNED_ACCESS_IS_FAST
301 			if ((load_u32_unaligned(matchptr + best_len - 3) ==
302 			     load_u32_unaligned(in_next + best_len - 3)) &&
303 			    (load_u32_unaligned(matchptr) ==
304 			     load_u32_unaligned(in_next)))
305 		#else
306 			if (matchptr[best_len] == in_next[best_len])
307 		#endif
308 				break;
309 
310 			/* Continue to the next node in the list.  */
311 			cur_node4 = mf->next_tab[cur_node4 & (MATCHFINDER_WINDOW_SIZE - 1)];
312 			if (cur_node4 <= cutoff || !--depth_remaining)
313 				goto out;
314 		}
315 
316 	#if UNALIGNED_ACCESS_IS_FAST
317 		len = 4;
318 	#else
319 		len = 0;
320 	#endif
321 		len = lz_extend(in_next, matchptr, len, max_len);
322 		if (len > best_len) {
323 			/* This is the new longest match.  */
324 			best_len = len;
325 			best_matchptr = matchptr;
326 			if (best_len >= nice_len)
327 				goto out;
328 		}
329 
330 		/* Continue to the next node in the list.  */
331 		cur_node4 = mf->next_tab[cur_node4 & (MATCHFINDER_WINDOW_SIZE - 1)];
332 		if (cur_node4 <= cutoff || !--depth_remaining)
333 			goto out;
334 	}
335 out:
336 	*offset_ret = in_next - best_matchptr;
337 	return best_len;
338 }
339 
340 /*
341  * Advance the matchfinder, but don't search for matches.
342  *
343  * @mf
344  *	The matchfinder structure.
345  * @in_base_p
346  *	Location of a pointer which points to the place in the input data the
347  *	matchfinder currently stores positions relative to.  This may be updated
348  *	by this function.
349  * @cur_pos
350  *	The current position in the input buffer relative to @in_base.
351  * @end_pos
352  *	The end position of the input buffer, relative to @in_base.
353  * @next_hashes
354  *	The precomputed hash codes for the sequence beginning at @in_next.
355  *	These will be used and then updated with the precomputed hashcodes for
356  *	the sequence beginning at @in_next + @count.
357  * @count
358  *	The number of bytes to advance.  Must be > 0.
359  *
360  * Returns @in_next + @count.
361  */
362 static forceinline const u8 *
hc_matchfinder_skip_positions(struct hc_matchfinder * const restrict mf,const u8 ** const restrict in_base_p,const u8 * in_next,const u8 * const in_end,const u32 count,u32 * const restrict next_hashes)363 hc_matchfinder_skip_positions(struct hc_matchfinder * const restrict mf,
364 			      const u8 ** const restrict in_base_p,
365 			      const u8 *in_next,
366 			      const u8 * const in_end,
367 			      const u32 count,
368 			      u32 * const restrict next_hashes)
369 {
370 	u32 cur_pos;
371 	u32 hash3, hash4;
372 	u32 next_hashseq;
373 	u32 remaining = count;
374 
375 	if (unlikely(count + 5 > in_end - in_next))
376 		return &in_next[count];
377 
378 	cur_pos = in_next - *in_base_p;
379 	hash3 = next_hashes[0];
380 	hash4 = next_hashes[1];
381 	do {
382 		if (cur_pos == MATCHFINDER_WINDOW_SIZE) {
383 			hc_matchfinder_slide_window(mf);
384 			*in_base_p += MATCHFINDER_WINDOW_SIZE;
385 			cur_pos = 0;
386 		}
387 		mf->hash3_tab[hash3] = cur_pos;
388 		mf->next_tab[cur_pos] = mf->hash4_tab[hash4];
389 		mf->hash4_tab[hash4] = cur_pos;
390 
391 		next_hashseq = get_unaligned_le32(++in_next);
392 		hash3 = lz_hash(next_hashseq & 0xFFFFFF, HC_MATCHFINDER_HASH3_ORDER);
393 		hash4 = lz_hash(next_hashseq, HC_MATCHFINDER_HASH4_ORDER);
394 		cur_pos++;
395 	} while (--remaining);
396 
397 	prefetchw(&mf->hash3_tab[hash3]);
398 	prefetchw(&mf->hash4_tab[hash4]);
399 	next_hashes[0] = hash3;
400 	next_hashes[1] = hash4;
401 
402 	return in_next;
403 }
404