5e1be485f3bc9ff0590db093edc8f09599c3f0b5
[wimlib] / src / lzx_compress.c
1 /*
2  * lzx_compress.c
3  *
4  * A compressor for the LZX compression format, as used in WIM files.
5  */
6
7 /*
8  * Copyright (C) 2012, 2013, 2014, 2015 Eric Biggers
9  *
10  * This file is free software; you can redistribute it and/or modify it under
11  * the terms of the GNU Lesser General Public License as published by the Free
12  * Software Foundation; either version 3 of the License, or (at your option) any
13  * later version.
14  *
15  * This file is distributed in the hope that it will be useful, but WITHOUT
16  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
17  * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
18  * details.
19  *
20  * You should have received a copy of the GNU Lesser General Public License
21  * along with this file; if not, see http://www.gnu.org/licenses/.
22  */
23
24
25 /*
26  * This file contains a compressor for the LZX ("Lempel-Ziv eXtended")
27  * compression format, as used in the WIM (Windows IMaging) file format.
28  *
29  * Two different parsing algorithms are implemented: "near-optimal" and "lazy".
30  * "Near-optimal" is significantly slower than "lazy", but results in a better
31  * compression ratio.  The "near-optimal" algorithm is used at the default
32  * compression level.
33  *
34  * This file may need some slight modifications to be used outside of the WIM
35  * format.  In particular, in other situations the LZX block header might be
36  * slightly different, and sliding window support might be required.
37  *
38  * Note: LZX is a compression format derived from DEFLATE, the format used by
39  * zlib and gzip.  Both LZX and DEFLATE use LZ77 matching and Huffman coding.
40  * Certain details are quite similar, such as the method for storing Huffman
41  * codes.  However, the main differences are:
42  *
43  * - LZX preprocesses the data to attempt to make x86 machine code slightly more
44  *   compressible before attempting to compress it further.
45  *
46  * - LZX uses a "main" alphabet which combines literals and matches, with the
47  *   match symbols containing a "length header" (giving all or part of the match
48  *   length) and an "offset slot" (giving, roughly speaking, the order of
49  *   magnitude of the match offset).
50  *
51  * - LZX does not have static Huffman blocks (that is, the kind with preset
52  *   Huffman codes); however it does have two types of dynamic Huffman blocks
53  *   ("verbatim" and "aligned").
54  *
55  * - LZX has a minimum match length of 2 rather than 3.  Length 2 matches can be
56  *   useful, but generally only if the parser is smart about choosing them.
57  *
58  * - In LZX, offset slots 0 through 2 actually represent entries in an LRU queue
59  *   of match offsets.  This is very useful for certain types of files, such as
60  *   binary files that have repeating records.
61  */
62
63 #ifdef HAVE_CONFIG_H
64 #  include "config.h"
65 #endif
66
67 /*
68  * Start a new LZX block (with new Huffman codes) after this many bytes.
69  *
70  * Note: actual block sizes may slightly exceed this value.
71  *
72  * TODO: recursive splitting and cost evaluation might be good for an extremely
73  * high compression mode, but otherwise it is almost always far too slow for how
74  * much it helps.  Perhaps some sort of heuristic would be useful?
75  */
76 #define LZX_DIV_BLOCK_SIZE      32768
77
78 /*
79  * LZX_CACHE_PER_POS is the number of lz_match structures to reserve in the
80  * match cache for each byte position.  This value should be high enough so that
81  * nearly the time, all matches found in a given block can fit in the match
82  * cache.  However, fallback behavior (immediately terminating the block) on
83  * cache overflow is still required.
84  */
85 #define LZX_CACHE_PER_POS       7
86
87 /*
88  * LZX_CACHE_LENGTH is the number of lz_match structures in the match cache,
89  * excluding the extra "overflow" entries.  The per-position multiplier is '1 +
90  * LZX_CACHE_PER_POS' instead of 'LZX_CACHE_PER_POS' because there is an
91  * overhead of one lz_match per position, used to hold the match count at that
92  * position.
93  */
94 #define LZX_CACHE_LENGTH        (LZX_DIV_BLOCK_SIZE * (1 + LZX_CACHE_PER_POS))
95
96 /*
97  * LZX_MAX_MATCHES_PER_POS is an upper bound on the number of matches that can
98  * ever be saved in the match cache for a single position.  Since each match we
99  * save for a single position has a distinct length, we can use the number of
100  * possible match lengths in LZX as this bound.  This bound is guaranteed to be
101  * valid in all cases, although if 'nice_match_length < LZX_MAX_MATCH_LEN', then
102  * it will never actually be reached.
103  */
104 #define LZX_MAX_MATCHES_PER_POS LZX_NUM_LENS
105
106 /*
107  * LZX_BIT_COST is a scaling factor that represents the cost to output one bit.
108  * This makes it possible to consider fractional bit costs.
109  *
110  * Note: this is only useful as a statistical trick for when the true costs are
111  * unknown.  In reality, each token in LZX requires a whole number of bits to
112  * output.
113  */
114 #define LZX_BIT_COST            16
115
116 /*
117  * Consideration of aligned offset costs is disabled for now, due to
118  * insufficient benefit gained from the time spent.
119  */
120 #define LZX_CONSIDER_ALIGNED_COSTS      0
121
122 /*
123  * LZX_MAX_FAST_LEVEL is the maximum compression level at which we use the
124  * faster algorithm.
125  */
126 #define LZX_MAX_FAST_LEVEL      34
127
128 /*
129  * LZX_HASH2_ORDER is the log base 2 of the number of entries in the hash table
130  * for finding length 2 matches.  This can be as high as 16 (in which case the
131  * hash function is trivial), but using a smaller hash table speeds up
132  * compression due to reduced cache pressure.
133  */
134 #define LZX_HASH2_ORDER         12
135 #define LZX_HASH2_LENGTH        (1UL << LZX_HASH2_ORDER)
136
137 #include "wimlib/lzx_common.h"
138
139 /*
140  * The maximum allowed window order for the matchfinder.
141  */
142 #define MATCHFINDER_MAX_WINDOW_ORDER    LZX_MAX_WINDOW_ORDER
143
144 #include <string.h>
145
146 #include "wimlib/bt_matchfinder.h"
147 #include "wimlib/compress_common.h"
148 #include "wimlib/compressor_ops.h"
149 #include "wimlib/error.h"
150 #include "wimlib/hc_matchfinder.h"
151 #include "wimlib/lz_extend.h"
152 #include "wimlib/unaligned.h"
153 #include "wimlib/util.h"
154
155 struct lzx_output_bitstream;
156
157 /* Codewords for the LZX Huffman codes.  */
158 struct lzx_codewords {
159         u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
160         u32 len[LZX_LENCODE_NUM_SYMBOLS];
161         u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
162 };
163
164 /* Codeword lengths (in bits) for the LZX Huffman codes.
165  * A zero length means the corresponding codeword has zero frequency.  */
166 struct lzx_lens {
167         u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
168         u8 len[LZX_LENCODE_NUM_SYMBOLS];
169         u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
170 };
171
172 /* Cost model for near-optimal parsing  */
173 struct lzx_costs {
174
175         /* 'match_cost[offset_slot][len - LZX_MIN_MATCH_LEN]' is the cost for a
176          * length 'len' match that has an offset belonging to 'offset_slot'.  */
177         u32 match_cost[LZX_MAX_OFFSET_SLOTS][LZX_NUM_LENS];
178
179         /* Cost for each symbol in the main code  */
180         u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
181
182         /* Cost for each symbol in the length code  */
183         u32 len[LZX_LENCODE_NUM_SYMBOLS];
184
185 #if LZX_CONSIDER_ALIGNED_COSTS
186         /* Cost for each symbol in the aligned code  */
187         u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
188 #endif
189 };
190
191 /* Codewords and lengths for the LZX Huffman codes.  */
192 struct lzx_codes {
193         struct lzx_codewords codewords;
194         struct lzx_lens lens;
195 };
196
197 /* Symbol frequency counters for the LZX Huffman codes.  */
198 struct lzx_freqs {
199         u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
200         u32 len[LZX_LENCODE_NUM_SYMBOLS];
201         u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
202 };
203
204 /* Intermediate LZX match/literal format  */
205 struct lzx_item {
206
207         /* Bits 0  -  9: Main symbol
208          * Bits 10 - 17: Length symbol
209          * Bits 18 - 22: Number of extra offset bits
210          * Bits 23+    : Extra offset bits  */
211         u64 data;
212 };
213
214 /*
215  * This structure represents a byte position in the input buffer and a node in
216  * the graph of possible match/literal choices.
217  *
218  * Logically, each incoming edge to this node is labeled with a literal or a
219  * match that can be taken to reach this position from an earlier position; and
220  * each outgoing edge from this node is labeled with a literal or a match that
221  * can be taken to advance from this position to a later position.
222  */
223 struct lzx_optimum_node {
224
225         /* The cost, in bits, of the lowest-cost path that has been found to
226          * reach this position.  This can change as progressively lower cost
227          * paths are found to reach this position.  */
228         u32 cost;
229
230         /*
231          * The match or literal that was taken to reach this position.  This can
232          * change as progressively lower cost paths are found to reach this
233          * position.
234          *
235          * This variable is divided into two bitfields.
236          *
237          * Literals:
238          *      Low bits are 1, high bits are the literal.
239          *
240          * Explicit offset matches:
241          *      Low bits are the match length, high bits are the offset plus 2.
242          *
243          * Repeat offset matches:
244          *      Low bits are the match length, high bits are the queue index.
245          */
246         u32 item;
247 #define OPTIMUM_OFFSET_SHIFT 9
248 #define OPTIMUM_LEN_MASK ((1 << OPTIMUM_OFFSET_SHIFT) - 1)
249 } _aligned_attribute(8);
250
251 /*
252  * Least-recently-used queue for match offsets.
253  *
254  * This is represented as a 64-bit integer for efficiency.  There are three
255  * offsets of 21 bits each.  Bit 64 is garbage.
256  */
257 struct lzx_lru_queue {
258         u64 R;
259 };
260
261 #define LZX_QUEUE64_OFFSET_SHIFT 21
262 #define LZX_QUEUE64_OFFSET_MASK (((u64)1 << LZX_QUEUE64_OFFSET_SHIFT) - 1)
263
264 #define LZX_QUEUE64_R0_SHIFT (0 * LZX_QUEUE64_OFFSET_SHIFT)
265 #define LZX_QUEUE64_R1_SHIFT (1 * LZX_QUEUE64_OFFSET_SHIFT)
266 #define LZX_QUEUE64_R2_SHIFT (2 * LZX_QUEUE64_OFFSET_SHIFT)
267
268 #define LZX_QUEUE64_R0_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R0_SHIFT)
269 #define LZX_QUEUE64_R1_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R1_SHIFT)
270 #define LZX_QUEUE64_R2_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R2_SHIFT)
271
272 static inline void
273 lzx_lru_queue_init(struct lzx_lru_queue *queue)
274 {
275         queue->R = ((u64)1 << LZX_QUEUE64_R0_SHIFT) |
276                    ((u64)1 << LZX_QUEUE64_R1_SHIFT) |
277                    ((u64)1 << LZX_QUEUE64_R2_SHIFT);
278 }
279
280 static inline u64
281 lzx_lru_queue_R0(struct lzx_lru_queue queue)
282 {
283         return (queue.R >> LZX_QUEUE64_R0_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
284 }
285
286 static inline u64
287 lzx_lru_queue_R1(struct lzx_lru_queue queue)
288 {
289         return (queue.R >> LZX_QUEUE64_R1_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
290 }
291
292 static inline u64
293 lzx_lru_queue_R2(struct lzx_lru_queue queue)
294 {
295         return (queue.R >> LZX_QUEUE64_R2_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
296 }
297
298 /* Push a match offset onto the front (most recently used) end of the queue.  */
299 static inline struct lzx_lru_queue
300 lzx_lru_queue_push(struct lzx_lru_queue queue, u32 offset)
301 {
302         return (struct lzx_lru_queue) {
303                 .R = (queue.R << LZX_QUEUE64_OFFSET_SHIFT) | offset,
304         };
305 }
306
307 /* Pop a match offset off the front (most recently used) end of the queue.  */
308 static inline u32
309 lzx_lru_queue_pop(struct lzx_lru_queue *queue_p)
310 {
311         u32 offset = queue_p->R & LZX_QUEUE64_OFFSET_MASK;
312         queue_p->R >>= LZX_QUEUE64_OFFSET_SHIFT;
313         return offset;
314 }
315
316 /* Swap a match offset to the front of the queue.  */
317 static inline struct lzx_lru_queue
318 lzx_lru_queue_swap(struct lzx_lru_queue queue, unsigned idx)
319 {
320         if (idx == 0)
321                 return queue;
322
323         if (idx == 1)
324                 return (struct lzx_lru_queue) {
325                         .R = (lzx_lru_queue_R1(queue) << LZX_QUEUE64_R0_SHIFT) |
326                              (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R1_SHIFT) |
327                              (queue.R & LZX_QUEUE64_R2_MASK),
328                 };
329
330         return (struct lzx_lru_queue) {
331                 .R = (lzx_lru_queue_R2(queue) << LZX_QUEUE64_R0_SHIFT) |
332                      (queue.R & LZX_QUEUE64_R1_MASK) |
333                      (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R2_SHIFT),
334         };
335 }
336
337 /* The main LZX compressor structure  */
338 struct lzx_compressor {
339
340         /* The "nice" match length: if a match of this length is found, then
341          * choose it immediately without further consideration.  */
342         unsigned nice_match_length;
343
344         /* The maximum search depth: consider at most this many potential
345          * matches at each position.  */
346         unsigned max_search_depth;
347
348         /* The log base 2 of the LZX window size for LZ match offset encoding
349          * purposes.  This will be >= LZX_MIN_WINDOW_ORDER and <=
350          * LZX_MAX_WINDOW_ORDER.  */
351         unsigned window_order;
352
353         /* The number of symbols in the main alphabet.  This depends on
354          * @window_order, since @window_order determines the maximum possible
355          * offset.  */
356         unsigned num_main_syms;
357
358         /* Number of optimization passes per block  */
359         unsigned num_optim_passes;
360
361         /* The preprocessed buffer of data being compressed  */
362         u8 *in_buffer;
363
364         /* The number of bytes of data to be compressed, which is the number of
365          * bytes of data in @in_buffer that are actually valid.  */
366         size_t in_nbytes;
367
368         /* Pointer to the compress() implementation chosen at allocation time */
369         void (*impl)(struct lzx_compressor *, struct lzx_output_bitstream *);
370
371         /* If true, the compressor need not preserve the input buffer if it
372          * compresses the data successfully.  */
373         bool destructive;
374
375         /* The Huffman symbol frequency counters for the current block.  */
376         struct lzx_freqs freqs;
377
378         /* The Huffman codes for the current and previous blocks.  The one with
379          * index 'codes_index' is for the current block, and the other one is
380          * for the previous block.  */
381         struct lzx_codes codes[2];
382         unsigned codes_index;
383
384         /*
385          * The match/literal sequence the algorithm chose for the current block.
386          *
387          * Notes on how large this array actually needs to be:
388          *
389          * - In lzx_compress_near_optimal(), the maximum block size is
390          *   'LZX_DIV_BLOCK_SIZE + LZX_MAX_MATCH_LEN - 1' bytes.  This occurs if
391          *   a match of the maximum length is found on the last byte.  Although
392          *   it is impossible for this particular case to actually result in a
393          *   parse of all literals, we reserve this many spaces anyway.
394          *
395          * - The worst case for lzx_compress_lazy() is a block of almost all
396          *   literals that ends with a series of matches of increasing scores,
397          *   causing a sequence of literals to be chosen before the last match
398          *   is finally chosen.  The number of items actually chosen in this
399          *   scenario is limited by the number of distinct match scores that
400          *   exist for matches shorter than 'nice_match_length'.  Having
401          *   'LZX_MAX_MATCH_LEN - 1' extra spaces is plenty for now.
402          */
403         struct lzx_item chosen_items[LZX_DIV_BLOCK_SIZE + LZX_MAX_MATCH_LEN - 1];
404
405         /* Table mapping match offset => offset slot for small offsets  */
406 #define LZX_NUM_FAST_OFFSETS 32768
407         u8 offset_slot_fast[LZX_NUM_FAST_OFFSETS];
408
409         union {
410                 /* Data for greedy or lazy parsing  */
411                 struct {
412                         /* Hash chains matchfinder (MUST BE LAST!!!)  */
413                         struct hc_matchfinder hc_mf;
414                 };
415
416                 /* Data for near-optimal parsing  */
417                 struct {
418                         /*
419                          * The graph nodes for the current block.
420                          *
421                          * We need at least 'LZX_DIV_BLOCK_SIZE +
422                          * LZX_MAX_MATCH_LEN - 1' nodes because that is the
423                          * maximum block size that may be used.  Add 1 because
424                          * we need a node to represent end-of-block.
425                          *
426                          * It is possible that nodes past end-of-block are
427                          * accessed during match consideration, but this can
428                          * only occur if the block was truncated at
429                          * LZX_DIV_BLOCK_SIZE.  So the same bound still applies.
430                          * Note that since nodes past the end of the block will
431                          * never actually have an effect on the items that are
432                          * chosen for the block, it makes no difference what
433                          * their costs are initialized to (if anything).
434                          */
435                         struct lzx_optimum_node optimum_nodes[LZX_DIV_BLOCK_SIZE +
436                                                               LZX_MAX_MATCH_LEN - 1 + 1];
437
438                         /* The cost model for the current block  */
439                         struct lzx_costs costs;
440
441                         /*
442                          * Cached matches for the current block.  This array
443                          * contains the matches that were found at each position
444                          * in the block.  Specifically, for each position, there
445                          * is a special 'struct lz_match' whose 'length' field
446                          * contains the number of matches that were found at
447                          * that position; this is followed by the matches
448                          * themselves, if any, sorted by strictly increasing
449                          * length.
450                          *
451                          * Note: in rare cases, there will be a very high number
452                          * of matches in the block and this array will overflow.
453                          * If this happens, we force the end of the current
454                          * block.  LZX_CACHE_LENGTH is the length at which we
455                          * actually check for overflow.  The extra slots beyond
456                          * this are enough to absorb the worst case overflow,
457                          * which occurs if starting at
458                          * &match_cache[LZX_CACHE_LENGTH - 1], we write the
459                          * match count header, then write
460                          * LZX_MAX_MATCHES_PER_POS matches, then skip searching
461                          * for matches at 'LZX_MAX_MATCH_LEN - 1' positions and
462                          * write the match count header for each.
463                          */
464                         struct lz_match match_cache[LZX_CACHE_LENGTH +
465                                                     LZX_MAX_MATCHES_PER_POS +
466                                                     LZX_MAX_MATCH_LEN - 1];
467
468                         /* Hash table for finding length 2 matches  */
469                         pos_t hash2_tab[LZX_HASH2_LENGTH];
470
471                         /* Binary trees matchfinder (MUST BE LAST!!!)  */
472                         struct bt_matchfinder bt_mf;
473                 };
474         };
475 };
476
477 /*
478  * Structure to keep track of the current state of sending bits to the
479  * compressed output buffer.
480  *
481  * The LZX bitstream is encoded as a sequence of 16-bit coding units.
482  */
483 struct lzx_output_bitstream {
484
485         /* Bits that haven't yet been written to the output buffer.  */
486         u32 bitbuf;
487
488         /* Number of bits currently held in @bitbuf.  */
489         u32 bitcount;
490
491         /* Pointer to the start of the output buffer.  */
492         u8 *start;
493
494         /* Pointer to the position in the output buffer at which the next coding
495          * unit should be written.  */
496         u8 *next;
497
498         /* Pointer just past the end of the output buffer, rounded down to a
499          * 2-byte boundary.  */
500         u8 *end;
501 };
502
503 /*
504  * Initialize the output bitstream.
505  *
506  * @os
507  *      The output bitstream structure to initialize.
508  * @buffer
509  *      The buffer being written to.
510  * @size
511  *      Size of @buffer, in bytes.
512  */
513 static void
514 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size)
515 {
516         os->bitbuf = 0;
517         os->bitcount = 0;
518         os->start = buffer;
519         os->next = os->start;
520         os->end = os->start + (size & ~1);
521 }
522
523 /*
524  * Write some bits to the output bitstream.
525  *
526  * The bits are given by the low-order @num_bits bits of @bits.  Higher-order
527  * bits in @bits cannot be set.  At most 17 bits can be written at once.
528  *
529  * @max_num_bits is a compile-time constant that specifies the maximum number of
530  * bits that can ever be written at the call site.  It is used to optimize away
531  * the conditional code for writing a second 16-bit coding unit when writing
532  * fewer than 17 bits.
533  *
534  * If the output buffer space is exhausted, then the bits will be ignored, and
535  * lzx_flush_output() will return 0 when it gets called.
536  */
537 static inline void
538 lzx_write_varbits(struct lzx_output_bitstream *os,
539                   const u32 bits, const unsigned num_bits,
540                   const unsigned max_num_bits)
541 {
542         /* This code is optimized for LZX, which never needs to write more than
543          * 17 bits at once.  */
544         LZX_ASSERT(num_bits <= 17);
545         LZX_ASSERT(num_bits <= max_num_bits);
546         LZX_ASSERT(os->bitcount <= 15);
547
548         /* Add the bits to the bit buffer variable.  @bitcount will be at most
549          * 15, so there will be just enough space for the maximum possible
550          * @num_bits of 17.  */
551         os->bitcount += num_bits;
552         os->bitbuf = (os->bitbuf << num_bits) | bits;
553
554         /* Check whether any coding units need to be written.  */
555         if (os->bitcount >= 16) {
556
557                 os->bitcount -= 16;
558
559                 /* Write a coding unit, unless it would overflow the buffer.  */
560                 if (os->next != os->end) {
561                         put_unaligned_u16_le(os->bitbuf >> os->bitcount, os->next);
562                         os->next += 2;
563                 }
564
565                 /* If writing 17 bits, a second coding unit might need to be
566                  * written.  But because 'max_num_bits' is a compile-time
567                  * constant, the compiler will optimize away this code at most
568                  * call sites.  */
569                 if (max_num_bits == 17 && os->bitcount == 16) {
570                         if (os->next != os->end) {
571                                 put_unaligned_u16_le(os->bitbuf, os->next);
572                                 os->next += 2;
573                         }
574                         os->bitcount = 0;
575                 }
576         }
577 }
578
579 /* Use when @num_bits is a compile-time constant.  Otherwise use
580  * lzx_write_varbits().  */
581 static inline void
582 lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
583 {
584         lzx_write_varbits(os, bits, num_bits, num_bits);
585 }
586
587 /*
588  * Flush the last coding unit to the output buffer if needed.  Return the total
589  * number of bytes written to the output buffer, or 0 if an overflow occurred.
590  */
591 static u32
592 lzx_flush_output(struct lzx_output_bitstream *os)
593 {
594         if (os->next == os->end)
595                 return 0;
596
597         if (os->bitcount != 0) {
598                 put_unaligned_u16_le(os->bitbuf << (16 - os->bitcount), os->next);
599                 os->next += 2;
600         }
601
602         return os->next - os->start;
603 }
604
605 /* Build the main, length, and aligned offset Huffman codes used in LZX.
606  *
607  * This takes as input the frequency tables for each code and produces as output
608  * a set of tables that map symbols to codewords and codeword lengths.  */
609 static void
610 lzx_make_huffman_codes(struct lzx_compressor *c)
611 {
612         const struct lzx_freqs *freqs = &c->freqs;
613         struct lzx_codes *codes = &c->codes[c->codes_index];
614
615         make_canonical_huffman_code(c->num_main_syms,
616                                     LZX_MAX_MAIN_CODEWORD_LEN,
617                                     freqs->main,
618                                     codes->lens.main,
619                                     codes->codewords.main);
620
621         make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
622                                     LZX_MAX_LEN_CODEWORD_LEN,
623                                     freqs->len,
624                                     codes->lens.len,
625                                     codes->codewords.len);
626
627         make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
628                                     LZX_MAX_ALIGNED_CODEWORD_LEN,
629                                     freqs->aligned,
630                                     codes->lens.aligned,
631                                     codes->codewords.aligned);
632 }
633
634 /* Reset the symbol frequencies for the LZX Huffman codes.  */
635 static void
636 lzx_reset_symbol_frequencies(struct lzx_compressor *c)
637 {
638         memset(&c->freqs, 0, sizeof(c->freqs));
639 }
640
641 static unsigned
642 lzx_compute_precode_items(const u8 lens[restrict],
643                           const u8 prev_lens[restrict],
644                           const unsigned num_lens,
645                           u32 precode_freqs[restrict],
646                           unsigned precode_items[restrict])
647 {
648         unsigned *itemptr;
649         unsigned run_start;
650         unsigned run_end;
651         unsigned extra_bits;
652         int delta;
653         u8 len;
654
655         itemptr = precode_items;
656         run_start = 0;
657         do {
658                 /* Find the next run of codeword lengths.  */
659
660                 /* len = the length being repeated  */
661                 len = lens[run_start];
662
663                 run_end = run_start + 1;
664
665                 /* Fast case for a single length.  */
666                 if (likely(run_end == num_lens || len != lens[run_end])) {
667                         delta = prev_lens[run_start] - len;
668                         if (delta < 0)
669                                 delta += 17;
670                         precode_freqs[delta]++;
671                         *itemptr++ = delta;
672                         run_start++;
673                         continue;
674                 }
675
676                 /* Extend the run.  */
677                 do {
678                         run_end++;
679                 } while (run_end != num_lens && len == lens[run_end]);
680
681                 if (len == 0) {
682                         /* Run of zeroes.  */
683
684                         /* Symbol 18: RLE 20 to 51 zeroes at a time.  */
685                         while ((run_end - run_start) >= 20) {
686                                 extra_bits = min((run_end - run_start) - 20, 0x1f);
687                                 precode_freqs[18]++;
688                                 *itemptr++ = 18 | (extra_bits << 5);
689                                 run_start += 20 + extra_bits;
690                         }
691
692                         /* Symbol 17: RLE 4 to 19 zeroes at a time.  */
693                         if ((run_end - run_start) >= 4) {
694                                 extra_bits = min((run_end - run_start) - 4, 0xf);
695                                 precode_freqs[17]++;
696                                 *itemptr++ = 17 | (extra_bits << 5);
697                                 run_start += 4 + extra_bits;
698                         }
699                 } else {
700
701                         /* A run of nonzero lengths. */
702
703                         /* Symbol 19: RLE 4 to 5 of any length at a time.  */
704                         while ((run_end - run_start) >= 4) {
705                                 extra_bits = (run_end - run_start) > 4;
706                                 delta = prev_lens[run_start] - len;
707                                 if (delta < 0)
708                                         delta += 17;
709                                 precode_freqs[19]++;
710                                 precode_freqs[delta]++;
711                                 *itemptr++ = 19 | (extra_bits << 5) | (delta << 6);
712                                 run_start += 4 + extra_bits;
713                         }
714                 }
715
716                 /* Output any remaining lengths without RLE.  */
717                 while (run_start != run_end) {
718                         delta = prev_lens[run_start] - len;
719                         if (delta < 0)
720                                 delta += 17;
721                         precode_freqs[delta]++;
722                         *itemptr++ = delta;
723                         run_start++;
724                 }
725         } while (run_start != num_lens);
726
727         return itemptr - precode_items;
728 }
729
730 /*
731  * Output a Huffman code in the compressed form used in LZX.
732  *
733  * The Huffman code is represented in the output as a logical series of codeword
734  * lengths from which the Huffman code, which must be in canonical form, can be
735  * reconstructed.
736  *
737  * The codeword lengths are themselves compressed using a separate Huffman code,
738  * the "precode", which contains a symbol for each possible codeword length in
739  * the larger code as well as several special symbols to represent repeated
740  * codeword lengths (a form of run-length encoding).  The precode is itself
741  * constructed in canonical form, and its codeword lengths are represented
742  * literally in 20 4-bit fields that immediately precede the compressed codeword
743  * lengths of the larger code.
744  *
745  * Furthermore, the codeword lengths of the larger code are actually represented
746  * as deltas from the codeword lengths of the corresponding code in the previous
747  * block.
748  *
749  * @os:
750  *      Bitstream to which to write the compressed Huffman code.
751  * @lens:
752  *      The codeword lengths, indexed by symbol, in the Huffman code.
753  * @prev_lens:
754  *      The codeword lengths, indexed by symbol, in the corresponding Huffman
755  *      code in the previous block, or all zeroes if this is the first block.
756  * @num_lens:
757  *      The number of symbols in the Huffman code.
758  */
759 static void
760 lzx_write_compressed_code(struct lzx_output_bitstream *os,
761                           const u8 lens[restrict],
762                           const u8 prev_lens[restrict],
763                           unsigned num_lens)
764 {
765         u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
766         u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
767         u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
768         unsigned precode_items[num_lens];
769         unsigned num_precode_items;
770         unsigned precode_item;
771         unsigned precode_sym;
772         unsigned i;
773
774         for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
775                 precode_freqs[i] = 0;
776
777         /* Compute the "items" (RLE / literal tokens and extra bits) with which
778          * the codeword lengths in the larger code will be output.  */
779         num_precode_items = lzx_compute_precode_items(lens,
780                                                       prev_lens,
781                                                       num_lens,
782                                                       precode_freqs,
783                                                       precode_items);
784
785         /* Build the precode.  */
786         make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
787                                     LZX_MAX_PRE_CODEWORD_LEN,
788                                     precode_freqs, precode_lens,
789                                     precode_codewords);
790
791         /* Output the lengths of the codewords in the precode.  */
792         for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
793                 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
794
795         /* Output the encoded lengths of the codewords in the larger code.  */
796         for (i = 0; i < num_precode_items; i++) {
797                 precode_item = precode_items[i];
798                 precode_sym = precode_item & 0x1F;
799                 lzx_write_varbits(os, precode_codewords[precode_sym],
800                                   precode_lens[precode_sym],
801                                   LZX_MAX_PRE_CODEWORD_LEN);
802                 if (precode_sym >= 17) {
803                         if (precode_sym == 17) {
804                                 lzx_write_bits(os, precode_item >> 5, 4);
805                         } else if (precode_sym == 18) {
806                                 lzx_write_bits(os, precode_item >> 5, 5);
807                         } else {
808                                 lzx_write_bits(os, (precode_item >> 5) & 1, 1);
809                                 precode_sym = precode_item >> 6;
810                                 lzx_write_varbits(os, precode_codewords[precode_sym],
811                                                   precode_lens[precode_sym],
812                                                   LZX_MAX_PRE_CODEWORD_LEN);
813                         }
814                 }
815         }
816 }
817
818 /* Output a match or literal.  */
819 static inline void
820 lzx_write_item(struct lzx_output_bitstream *os, struct lzx_item item,
821                unsigned ones_if_aligned, const struct lzx_codes *codes)
822 {
823         u64 data = item.data;
824         unsigned main_symbol;
825         unsigned len_symbol;
826         unsigned num_extra_bits;
827         u32 extra_bits;
828
829         main_symbol = data & 0x3FF;
830
831         lzx_write_varbits(os, codes->codewords.main[main_symbol],
832                           codes->lens.main[main_symbol],
833                           LZX_MAX_MAIN_CODEWORD_LEN);
834
835         if (main_symbol < LZX_NUM_CHARS)  /* Literal?  */
836                 return;
837
838         len_symbol = (data >> 10) & 0xFF;
839
840         if (len_symbol != LZX_LENCODE_NUM_SYMBOLS) {
841                 lzx_write_varbits(os, codes->codewords.len[len_symbol],
842                                   codes->lens.len[len_symbol],
843                                   LZX_MAX_LEN_CODEWORD_LEN);
844         }
845
846         num_extra_bits = (data >> 18) & 0x1F;
847         if (num_extra_bits == 0)  /* Small offset or repeat offset match?  */
848                 return;
849
850         extra_bits = data >> 23;
851
852         if ((num_extra_bits & ones_if_aligned) >= LZX_NUM_ALIGNED_OFFSET_BITS) {
853
854                 /* Aligned offset blocks: The low 3 bits of the extra offset
855                  * bits are Huffman-encoded using the aligned offset code.  The
856                  * remaining bits are output literally.  */
857
858                 lzx_write_varbits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS,
859                                   num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS,
860                                   17 - LZX_NUM_ALIGNED_OFFSET_BITS);
861
862                 lzx_write_varbits(os,
863                                   codes->codewords.aligned[extra_bits & LZX_ALIGNED_OFFSET_BITMASK],
864                                   codes->lens.aligned[extra_bits & LZX_ALIGNED_OFFSET_BITMASK],
865                                   LZX_MAX_ALIGNED_CODEWORD_LEN);
866         } else {
867                 /* Verbatim blocks, or fewer than 3 extra bits:  All extra
868                  * offset bits are output literally.  */
869                 lzx_write_varbits(os, extra_bits, num_extra_bits, 17);
870         }
871 }
872
873 /*
874  * Write all matches and literal bytes (which were precomputed) in an LZX
875  * compressed block to the output bitstream in the final compressed
876  * representation.
877  *
878  * @os
879  *      The output bitstream.
880  * @block_type
881  *      The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
882  *      LZX_BLOCKTYPE_VERBATIM).
883  * @items
884  *      The array of matches/literals to output.
885  * @num_items
886  *      Number of matches/literals to output (length of @items).
887  * @codes
888  *      The main, length, and aligned offset Huffman codes for the current
889  *      LZX compressed block.
890  */
891 static void
892 lzx_write_items(struct lzx_output_bitstream *os, int block_type,
893                 const struct lzx_item items[], u32 num_items,
894                 const struct lzx_codes *codes)
895 {
896         unsigned ones_if_aligned = 0U - (block_type == LZX_BLOCKTYPE_ALIGNED);
897
898         for (u32 i = 0; i < num_items; i++)
899                 lzx_write_item(os, items[i], ones_if_aligned, codes);
900 }
901
902 static void
903 lzx_write_compressed_block(int block_type,
904                            u32 block_size,
905                            unsigned window_order,
906                            unsigned num_main_syms,
907                            const struct lzx_item chosen_items[],
908                            u32 num_chosen_items,
909                            const struct lzx_codes * codes,
910                            const struct lzx_lens * prev_lens,
911                            struct lzx_output_bitstream * os)
912 {
913         LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
914                    block_type == LZX_BLOCKTYPE_VERBATIM);
915
916         /* The first three bits indicate the type of block and are one of the
917          * LZX_BLOCKTYPE_* constants.  */
918         lzx_write_bits(os, block_type, 3);
919
920         /* Output the block size.
921          *
922          * The original LZX format seemed to always encode the block size in 3
923          * bytes.  However, the implementation in WIMGAPI, as used in WIM files,
924          * uses the first bit to indicate whether the block is the default size
925          * (32768) or a different size given explicitly by the next 16 bits.
926          *
927          * By default, this compressor uses a window size of 32768 and therefore
928          * follows the WIMGAPI behavior.  However, this compressor also supports
929          * window sizes greater than 32768 bytes, which do not appear to be
930          * supported by WIMGAPI.  In such cases, we retain the default size bit
931          * to mean a size of 32768 bytes but output non-default block size in 24
932          * bits rather than 16.  The compatibility of this behavior is unknown
933          * because WIMs created with chunk size greater than 32768 can seemingly
934          * only be opened by wimlib anyway.  */
935         if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
936                 lzx_write_bits(os, 1, 1);
937         } else {
938                 lzx_write_bits(os, 0, 1);
939
940                 if (window_order >= 16)
941                         lzx_write_bits(os, block_size >> 16, 8);
942
943                 lzx_write_bits(os, block_size & 0xFFFF, 16);
944         }
945
946         /* If it's an aligned offset block, output the aligned offset code.  */
947         if (block_type == LZX_BLOCKTYPE_ALIGNED) {
948                 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
949                         lzx_write_bits(os, codes->lens.aligned[i],
950                                        LZX_ALIGNEDCODE_ELEMENT_SIZE);
951                 }
952         }
953
954         /* Output the main code (two parts).  */
955         lzx_write_compressed_code(os, codes->lens.main,
956                                   prev_lens->main,
957                                   LZX_NUM_CHARS);
958         lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
959                                   prev_lens->main + LZX_NUM_CHARS,
960                                   num_main_syms - LZX_NUM_CHARS);
961
962         /* Output the length code.  */
963         lzx_write_compressed_code(os, codes->lens.len,
964                                   prev_lens->len,
965                                   LZX_LENCODE_NUM_SYMBOLS);
966
967         /* Output the compressed matches and literals.  */
968         lzx_write_items(os, block_type, chosen_items, num_chosen_items, codes);
969 }
970
971 /* Given the frequencies of symbols in an LZX-compressed block and the
972  * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
973  * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
974  * will take fewer bits to output.  */
975 static int
976 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
977                                const struct lzx_codes * codes)
978 {
979         u32 aligned_cost = 0;
980         u32 verbatim_cost = 0;
981
982         /* A verbatim block requires 3 bits in each place that an aligned symbol
983          * would be used in an aligned offset block.  */
984         for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
985                 verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i];
986                 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
987         }
988
989         /* Account for output of the aligned offset code.  */
990         aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
991
992         if (aligned_cost < verbatim_cost)
993                 return LZX_BLOCKTYPE_ALIGNED;
994         else
995                 return LZX_BLOCKTYPE_VERBATIM;
996 }
997
998 /*
999  * Finish an LZX block:
1000  *
1001  * - build the Huffman codes
1002  * - decide whether to output the block as VERBATIM or ALIGNED
1003  * - output the block
1004  * - swap the indices of the current and previous Huffman codes
1005  */
1006 static void
1007 lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1008                  u32 block_size, u32 num_chosen_items)
1009 {
1010         int block_type;
1011
1012         lzx_make_huffman_codes(c);
1013
1014         block_type = lzx_choose_verbatim_or_aligned(&c->freqs,
1015                                                     &c->codes[c->codes_index]);
1016         lzx_write_compressed_block(block_type,
1017                                    block_size,
1018                                    c->window_order,
1019                                    c->num_main_syms,
1020                                    c->chosen_items,
1021                                    num_chosen_items,
1022                                    &c->codes[c->codes_index],
1023                                    &c->codes[c->codes_index ^ 1].lens,
1024                                    os);
1025         c->codes_index ^= 1;
1026 }
1027
1028 /* Return the offset slot for the specified offset, which must be
1029  * less than LZX_NUM_FAST_OFFSETS.  */
1030 static inline unsigned
1031 lzx_get_offset_slot_fast(struct lzx_compressor *c, u32 offset)
1032 {
1033         LZX_ASSERT(offset < LZX_NUM_FAST_OFFSETS);
1034         return c->offset_slot_fast[offset];
1035 }
1036
1037 /* Tally, and optionally record, the specified literal byte.  */
1038 static inline void
1039 lzx_declare_literal(struct lzx_compressor *c, unsigned literal,
1040                     struct lzx_item **next_chosen_item)
1041 {
1042         unsigned main_symbol = lzx_main_symbol_for_literal(literal);
1043
1044         c->freqs.main[main_symbol]++;
1045
1046         if (next_chosen_item) {
1047                 *(*next_chosen_item)++ = (struct lzx_item) {
1048                         .data = main_symbol,
1049                 };
1050         }
1051 }
1052
1053 /* Tally, and optionally record, the specified repeat offset match.  */
1054 static inline void
1055 lzx_declare_repeat_offset_match(struct lzx_compressor *c,
1056                                 unsigned len, unsigned rep_index,
1057                                 struct lzx_item **next_chosen_item)
1058 {
1059         unsigned len_header;
1060         unsigned len_symbol;
1061         unsigned main_symbol;
1062
1063         if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) {
1064                 len_header = len - LZX_MIN_MATCH_LEN;
1065                 len_symbol = LZX_LENCODE_NUM_SYMBOLS;
1066         } else {
1067                 len_header = LZX_NUM_PRIMARY_LENS;
1068                 len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS;
1069                 c->freqs.len[len_symbol]++;
1070         }
1071
1072         main_symbol = lzx_main_symbol_for_match(rep_index, len_header);
1073
1074         c->freqs.main[main_symbol]++;
1075
1076         if (next_chosen_item) {
1077                 *(*next_chosen_item)++ = (struct lzx_item) {
1078                         .data = (u64)main_symbol | ((u64)len_symbol << 10),
1079                 };
1080         }
1081 }
1082
1083 /* Tally, and optionally record, the specified explicit offset match.  */
1084 static inline void
1085 lzx_declare_explicit_offset_match(struct lzx_compressor *c, unsigned len, u32 offset,
1086                                   struct lzx_item **next_chosen_item)
1087 {
1088         unsigned len_header;
1089         unsigned len_symbol;
1090         unsigned main_symbol;
1091         unsigned offset_slot;
1092         unsigned num_extra_bits;
1093         u32 extra_bits;
1094
1095         if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) {
1096                 len_header = len - LZX_MIN_MATCH_LEN;
1097                 len_symbol = LZX_LENCODE_NUM_SYMBOLS;
1098         } else {
1099                 len_header = LZX_NUM_PRIMARY_LENS;
1100                 len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS;
1101                 c->freqs.len[len_symbol]++;
1102         }
1103
1104         offset_slot = (offset < LZX_NUM_FAST_OFFSETS) ?
1105                         lzx_get_offset_slot_fast(c, offset) :
1106                         lzx_get_offset_slot(offset);
1107
1108         main_symbol = lzx_main_symbol_for_match(offset_slot, len_header);
1109
1110         c->freqs.main[main_symbol]++;
1111
1112         num_extra_bits = lzx_extra_offset_bits[offset_slot];
1113
1114         if (num_extra_bits >= LZX_NUM_ALIGNED_OFFSET_BITS)
1115                 c->freqs.aligned[(offset + LZX_OFFSET_ADJUSTMENT) &
1116                                  LZX_ALIGNED_OFFSET_BITMASK]++;
1117
1118         if (next_chosen_item) {
1119
1120                 extra_bits = (offset + LZX_OFFSET_ADJUSTMENT) -
1121                              lzx_offset_slot_base[offset_slot];
1122
1123                 STATIC_ASSERT(LZX_MAINCODE_MAX_NUM_SYMBOLS <= (1 << 10));
1124                 STATIC_ASSERT(LZX_LENCODE_NUM_SYMBOLS <= (1 << 8));
1125                 *(*next_chosen_item)++ = (struct lzx_item) {
1126                         .data = (u64)main_symbol |
1127                                 ((u64)len_symbol << 10) |
1128                                 ((u64)num_extra_bits << 18) |
1129                                 ((u64)extra_bits << 23),
1130                 };
1131         }
1132 }
1133
1134
1135 /* Tally, and optionally record, the specified match or literal.  */
1136 static inline void
1137 lzx_declare_item(struct lzx_compressor *c, u32 item,
1138                  struct lzx_item **next_chosen_item)
1139 {
1140         u32 len = item & OPTIMUM_LEN_MASK;
1141         u32 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1142
1143         if (len == 1)
1144                 lzx_declare_literal(c, offset_data, next_chosen_item);
1145         else if (offset_data < LZX_NUM_RECENT_OFFSETS)
1146                 lzx_declare_repeat_offset_match(c, len, offset_data,
1147                                                 next_chosen_item);
1148         else
1149                 lzx_declare_explicit_offset_match(c, len,
1150                                                   offset_data - LZX_OFFSET_ADJUSTMENT,
1151                                                   next_chosen_item);
1152 }
1153
1154 static inline void
1155 lzx_record_item_list(struct lzx_compressor *c,
1156                      struct lzx_optimum_node *cur_node,
1157                      struct lzx_item **next_chosen_item)
1158 {
1159         struct lzx_optimum_node *end_node;
1160         u32 saved_item;
1161         u32 item;
1162
1163         /* The list is currently in reverse order (last item to first item).
1164          * Reverse it.  */
1165         end_node = cur_node;
1166         saved_item = cur_node->item;
1167         do {
1168                 item = saved_item;
1169                 cur_node -= item & OPTIMUM_LEN_MASK;
1170                 saved_item = cur_node->item;
1171                 cur_node->item = item;
1172         } while (cur_node != c->optimum_nodes);
1173
1174         /* Walk the list of items from beginning to end, tallying and recording
1175          * each item.  */
1176         do {
1177                 lzx_declare_item(c, cur_node->item, next_chosen_item);
1178                 cur_node += (cur_node->item) & OPTIMUM_LEN_MASK;
1179         } while (cur_node != end_node);
1180 }
1181
1182 static inline void
1183 lzx_tally_item_list(struct lzx_compressor *c, struct lzx_optimum_node *cur_node)
1184 {
1185         /* Since we're just tallying the items, we don't need to reverse the
1186          * list.  Processing the items in reverse order is fine.  */
1187         do {
1188                 lzx_declare_item(c, cur_node->item, NULL);
1189                 cur_node -= (cur_node->item & OPTIMUM_LEN_MASK);
1190         } while (cur_node != c->optimum_nodes);
1191 }
1192
1193 /*
1194  * Find an inexpensive path through the graph of possible match/literal choices
1195  * for the current block.  The nodes of the graph are
1196  * c->optimum_nodes[0...block_size].  They correspond directly to the bytes in
1197  * the current block, plus one extra node for end-of-block.  The edges of the
1198  * graph are matches and literals.  The goal is to find the minimum cost path
1199  * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]'.
1200  *
1201  * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and
1202  * proceeding forwards one node at a time.  At each node, a selection of matches
1203  * (len >= 2), as well as the literal byte (len = 1), is considered.  An item of
1204  * length 'len' provides a new path to reach the node 'len' bytes later.  If
1205  * such a path is the lowest cost found so far to reach that later node, then
1206  * that later node is updated with the new path.
1207  *
1208  * Note that although this algorithm is based on minimum cost path search, due
1209  * to various simplifying assumptions the result is not guaranteed to be the
1210  * true minimum cost, or "optimal", path over the graph of all valid LZX
1211  * representations of this block.
1212  *
1213  * Also, note that because of the presence of the recent offsets queue (which is
1214  * a type of adaptive state), the algorithm cannot work backwards and compute
1215  * "cost to end" instead of "cost to beginning".  Furthermore, the way the
1216  * algorithm handles this adaptive state in the "minimum cost" parse is actually
1217  * only an approximation.  It's possible for the globally optimal, minimum cost
1218  * path to contain a prefix, ending at a position, where that path prefix is
1219  * *not* the minimum cost path to that position.  This can happen if such a path
1220  * prefix results in a different adaptive state which results in lower costs
1221  * later.  The algorithm does not solve this problem; it only considers the
1222  * lowest cost to reach each individual position.
1223  */
1224 static struct lzx_lru_queue
1225 lzx_find_min_cost_path(struct lzx_compressor * const restrict c,
1226                        const u8 * const restrict block_begin,
1227                        const u32 block_size,
1228                        const struct lzx_lru_queue initial_queue)
1229 {
1230         struct lzx_optimum_node *cur_node = c->optimum_nodes;
1231         struct lzx_optimum_node * const end_node = &c->optimum_nodes[block_size];
1232         struct lz_match *cache_ptr = c->match_cache;
1233         const u8 *in_next = block_begin;
1234         const u8 * const block_end = block_begin + block_size;
1235
1236         /* Instead of storing the match offset LRU queues in the
1237          * 'lzx_optimum_node' structures, we save memory (and cache lines) by
1238          * storing them in a smaller array.  This works because the algorithm
1239          * only requires a limited history of the adaptive state.  Once a given
1240          * state is more than LZX_MAX_MATCH_LEN bytes behind the current node,
1241          * it is no longer needed.  */
1242         struct lzx_lru_queue queues[512];
1243
1244         STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1);
1245 #define QUEUE(in) (queues[(uintptr_t)(in) % ARRAY_LEN(queues)])
1246
1247         /* Initially, the cost to reach each node is "infinity".  */
1248         memset(c->optimum_nodes, 0xFF,
1249                (block_size + 1) * sizeof(c->optimum_nodes[0]));
1250
1251         QUEUE(block_begin) = initial_queue;
1252
1253         /* The following loop runs 'block_size' iterations, one per node.  */
1254         do {
1255                 unsigned num_matches;
1256                 unsigned literal;
1257                 u32 cost;
1258
1259                 /*
1260                  * A selection of matches for the block was already saved in
1261                  * memory so that we don't have to run the uncompressed data
1262                  * through the matchfinder on every optimization pass.  However,
1263                  * we still search for repeat offset matches during each
1264                  * optimization pass because we cannot predict the state of the
1265                  * recent offsets queue.  But as a heuristic, we don't bother
1266                  * searching for repeat offset matches if the general-purpose
1267                  * matchfinder failed to find any matches.
1268                  *
1269                  * Note that a match of length n at some offset implies there is
1270                  * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at
1271                  * that same offset.  In other words, we don't necessarily need
1272                  * to use the full length of a match.  The key heuristic that
1273                  * saves a significicant amount of time is that for each
1274                  * distinct length, we only consider the smallest offset for
1275                  * which that length is available.  This heuristic also applies
1276                  * to repeat offsets, which we order specially: R0 < R1 < R2 <
1277                  * any explicit offset.  Of course, this heuristic may be
1278                  * produce suboptimal results because offset slots in LZX are
1279                  * subject to entropy encoding, but in practice this is a useful
1280                  * heuristic.
1281                  */
1282
1283                 num_matches = cache_ptr->length;
1284                 cache_ptr++;
1285
1286                 if (num_matches) {
1287                         struct lz_match *end_matches = cache_ptr + num_matches;
1288                         unsigned next_len = LZX_MIN_MATCH_LEN;
1289                         unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN);
1290                         const u8 *matchptr;
1291
1292                         /* Consider R0 match  */
1293                         matchptr = in_next - lzx_lru_queue_R0(QUEUE(in_next));
1294                         if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1295                                 goto R0_done;
1296                         STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2);
1297                         do {
1298                                 u32 cost = cur_node->cost +
1299                                            c->costs.match_cost[0][
1300                                                         next_len - LZX_MIN_MATCH_LEN];
1301                                 if (cost <= (cur_node + next_len)->cost) {
1302                                         (cur_node + next_len)->cost = cost;
1303                                         (cur_node + next_len)->item =
1304                                                 (0 << OPTIMUM_OFFSET_SHIFT) | next_len;
1305                                 }
1306                                 if (unlikely(++next_len > max_len)) {
1307                                         cache_ptr = end_matches;
1308                                         goto done_matches;
1309                                 }
1310                         } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1311
1312                 R0_done:
1313
1314                         /* Consider R1 match  */
1315                         matchptr = in_next - lzx_lru_queue_R1(QUEUE(in_next));
1316                         if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1317                                 goto R1_done;
1318                         if (matchptr[next_len - 1] != in_next[next_len - 1])
1319                                 goto R1_done;
1320                         for (unsigned len = 2; len < next_len - 1; len++)
1321                                 if (matchptr[len] != in_next[len])
1322                                         goto R1_done;
1323                         do {
1324                                 u32 cost = cur_node->cost +
1325                                            c->costs.match_cost[1][
1326                                                         next_len - LZX_MIN_MATCH_LEN];
1327                                 if (cost <= (cur_node + next_len)->cost) {
1328                                         (cur_node + next_len)->cost = cost;
1329                                         (cur_node + next_len)->item =
1330                                                 (1 << OPTIMUM_OFFSET_SHIFT) | next_len;
1331                                 }
1332                                 if (unlikely(++next_len > max_len)) {
1333                                         cache_ptr = end_matches;
1334                                         goto done_matches;
1335                                 }
1336                         } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1337
1338                 R1_done:
1339
1340                         /* Consider R2 match  */
1341                         matchptr = in_next - lzx_lru_queue_R2(QUEUE(in_next));
1342                         if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1343                                 goto R2_done;
1344                         if (matchptr[next_len - 1] != in_next[next_len - 1])
1345                                 goto R2_done;
1346                         for (unsigned len = 2; len < next_len - 1; len++)
1347                                 if (matchptr[len] != in_next[len])
1348                                         goto R2_done;
1349                         do {
1350                                 u32 cost = cur_node->cost +
1351                                            c->costs.match_cost[2][
1352                                                         next_len - LZX_MIN_MATCH_LEN];
1353                                 if (cost <= (cur_node + next_len)->cost) {
1354                                         (cur_node + next_len)->cost = cost;
1355                                         (cur_node + next_len)->item =
1356                                                 (2 << OPTIMUM_OFFSET_SHIFT) | next_len;
1357                                 }
1358                                 if (unlikely(++next_len > max_len)) {
1359                                         cache_ptr = end_matches;
1360                                         goto done_matches;
1361                                 }
1362                         } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1363
1364                 R2_done:
1365
1366                         while (next_len > cache_ptr->length)
1367                                 if (++cache_ptr == end_matches)
1368                                         goto done_matches;
1369
1370                         /* Consider explicit offset matches  */
1371                         do {
1372                                 u32 offset = cache_ptr->offset;
1373                                 u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT;
1374                                 unsigned offset_slot = (offset < LZX_NUM_FAST_OFFSETS) ?
1375                                                 lzx_get_offset_slot_fast(c, offset) :
1376                                                 lzx_get_offset_slot(offset);
1377                                 do {
1378                                         u32 cost = cur_node->cost +
1379                                                    c->costs.match_cost[offset_slot][
1380                                                                 next_len - LZX_MIN_MATCH_LEN];
1381                                 #if LZX_CONSIDER_ALIGNED_COSTS
1382                                         if (lzx_extra_offset_bits[offset_slot] >=
1383                                             LZX_NUM_ALIGNED_OFFSET_BITS)
1384                                                 cost += c->costs.aligned[offset_data &
1385                                                                          LZX_ALIGNED_OFFSET_BITMASK];
1386                                 #endif
1387                                         if (cost < (cur_node + next_len)->cost) {
1388                                                 (cur_node + next_len)->cost = cost;
1389                                                 (cur_node + next_len)->item =
1390                                                         (offset_data << OPTIMUM_OFFSET_SHIFT) | next_len;
1391                                         }
1392                                 } while (++next_len <= cache_ptr->length);
1393                         } while (++cache_ptr != end_matches);
1394                 }
1395
1396         done_matches:
1397
1398                 /* Consider coding a literal.
1399
1400                  * To avoid an extra branch, actually checking the preferability
1401                  * of coding the literal is integrated into the queue update
1402                  * code below.  */
1403                 literal = *in_next++;
1404                 cost = cur_node->cost +
1405                        c->costs.main[lzx_main_symbol_for_literal(literal)];
1406
1407                 /* Advance to the next position.  */
1408                 cur_node++;
1409
1410                 /* The lowest-cost path to the current position is now known.
1411                  * Finalize the recent offsets queue that results from taking
1412                  * this lowest-cost path.  */
1413
1414                 if (cost <= cur_node->cost) {
1415                         /* Literal: queue remains unchanged.  */
1416                         cur_node->cost = cost;
1417                         cur_node->item = (literal << OPTIMUM_OFFSET_SHIFT) | 1;
1418                         QUEUE(in_next) = QUEUE(in_next - 1);
1419                 } else {
1420                         /* Match: queue update is needed.  */
1421                         unsigned len = cur_node->item & OPTIMUM_LEN_MASK;
1422                         u32 offset_data = cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1423                         if (offset_data >= LZX_NUM_RECENT_OFFSETS) {
1424                                 /* Explicit offset match: insert offset at front  */
1425                                 QUEUE(in_next) =
1426                                         lzx_lru_queue_push(QUEUE(in_next - len),
1427                                                            offset_data - LZX_OFFSET_ADJUSTMENT);
1428                         } else {
1429                                 /* Repeat offset match: swap offset to front  */
1430                                 QUEUE(in_next) =
1431                                         lzx_lru_queue_swap(QUEUE(in_next - len),
1432                                                            offset_data);
1433                         }
1434                 }
1435         } while (cur_node != end_node);
1436
1437         /* Return the match offset queue at the end of the minimum cost path. */
1438         return QUEUE(block_end);
1439 }
1440
1441 /* Given the costs for the main and length codewords, compute 'match_costs'.  */
1442 static void
1443 lzx_compute_match_costs(struct lzx_compressor *c)
1444 {
1445         unsigned num_offset_slots = lzx_get_num_offset_slots(c->window_order);
1446         struct lzx_costs *costs = &c->costs;
1447
1448         for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) {
1449
1450                 u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST;
1451                 unsigned main_symbol = lzx_main_symbol_for_match(offset_slot, 0);
1452                 unsigned i;
1453
1454         #if LZX_CONSIDER_ALIGNED_COSTS
1455                 if (lzx_extra_offset_bits[offset_slot] >= LZX_NUM_ALIGNED_OFFSET_BITS)
1456                         extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1457         #endif
1458
1459                 for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++)
1460                         costs->match_cost[offset_slot][i] =
1461                                 costs->main[main_symbol++] + extra_cost;
1462
1463                 extra_cost += costs->main[main_symbol];
1464
1465                 for (; i < LZX_NUM_LENS; i++)
1466                         costs->match_cost[offset_slot][i] =
1467                                 costs->len[i - LZX_NUM_PRIMARY_LENS] + extra_cost;
1468         }
1469 }
1470
1471 /* Set default LZX Huffman symbol costs to bootstrap the iterative optimization
1472  * algorithm.  */
1473 static void
1474 lzx_set_default_costs(struct lzx_compressor *c, const u8 *block, u32 block_size)
1475 {
1476         u32 i;
1477         bool have_byte[256];
1478         unsigned num_used_bytes;
1479
1480         /* The costs below are hard coded to use a scaling factor of 16.  */
1481         STATIC_ASSERT(LZX_BIT_COST == 16);
1482
1483         /*
1484          * Heuristics:
1485          *
1486          * - Use smaller initial costs for literal symbols when the input buffer
1487          *   contains fewer distinct bytes.
1488          *
1489          * - Assume that match symbols are more costly than literal symbols.
1490          *
1491          * - Assume that length symbols for shorter lengths are less costly than
1492          *   length symbols for longer lengths.
1493          */
1494
1495         for (i = 0; i < 256; i++)
1496                 have_byte[i] = false;
1497
1498         for (i = 0; i < block_size; i++)
1499                 have_byte[block[i]] = true;
1500
1501         num_used_bytes = 0;
1502         for (i = 0; i < 256; i++)
1503                 num_used_bytes += have_byte[i];
1504
1505         for (i = 0; i < 256; i++)
1506                 c->costs.main[i] = 140 - (256 - num_used_bytes) / 4;
1507
1508         for (; i < c->num_main_syms; i++)
1509                 c->costs.main[i] = 170;
1510
1511         for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1512                 c->costs.len[i] = 103 + (i / 4);
1513
1514 #if LZX_CONSIDER_ALIGNED_COSTS
1515         for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1516                 c->costs.aligned[i] = LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1517 #endif
1518
1519         lzx_compute_match_costs(c);
1520 }
1521
1522 /* Update the current cost model to reflect the computed Huffman codes.  */
1523 static void
1524 lzx_update_costs(struct lzx_compressor *c)
1525 {
1526         unsigned i;
1527         const struct lzx_lens *lens = &c->codes[c->codes_index].lens;
1528
1529         for (i = 0; i < c->num_main_syms; i++)
1530                 c->costs.main[i] = (lens->main[i] ? lens->main[i] : 15) * LZX_BIT_COST;
1531
1532         for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1533                 c->costs.len[i] = (lens->len[i] ? lens->len[i] : 15) * LZX_BIT_COST;
1534
1535 #if LZX_CONSIDER_ALIGNED_COSTS
1536         for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1537                 c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : 7) * LZX_BIT_COST;
1538 #endif
1539
1540         lzx_compute_match_costs(c);
1541 }
1542
1543 static struct lzx_lru_queue
1544 lzx_optimize_and_write_block(struct lzx_compressor *c,
1545                              struct lzx_output_bitstream *os,
1546                              const u8 *block_begin, const u32 block_size,
1547                              const struct lzx_lru_queue initial_queue)
1548 {
1549         unsigned num_passes_remaining = c->num_optim_passes;
1550         struct lzx_item *next_chosen_item;
1551         struct lzx_lru_queue new_queue;
1552
1553         /* The first optimization pass uses a default cost model.  Each
1554          * additional optimization pass uses a cost model derived from the
1555          * Huffman code computed in the previous pass.  */
1556
1557         lzx_set_default_costs(c, block_begin, block_size);
1558         lzx_reset_symbol_frequencies(c);
1559         do {
1560                 new_queue = lzx_find_min_cost_path(c, block_begin, block_size,
1561                                                    initial_queue);
1562                 if (num_passes_remaining > 1) {
1563                         lzx_tally_item_list(c, c->optimum_nodes + block_size);
1564                         lzx_make_huffman_codes(c);
1565                         lzx_update_costs(c);
1566                         lzx_reset_symbol_frequencies(c);
1567                 }
1568         } while (--num_passes_remaining);
1569
1570         next_chosen_item = c->chosen_items;
1571         lzx_record_item_list(c, c->optimum_nodes + block_size, &next_chosen_item);
1572         lzx_finish_block(c, os, block_size, next_chosen_item - c->chosen_items);
1573         return new_queue;
1574 }
1575
1576 /*
1577  * This is the "near-optimal" LZX compressor.
1578  *
1579  * For each block, it performs a relatively thorough graph search to find an
1580  * inexpensive (in terms of compressed size) way to output that block.
1581  *
1582  * Note: there are actually many things this algorithm leaves on the table in
1583  * terms of compression ratio.  So although it may be "near-optimal", it is
1584  * certainly not "optimal".  The goal is not to produce the optimal compression
1585  * ratio, which for LZX is probably impossible within any practical amount of
1586  * time, but rather to produce a compression ratio significantly better than a
1587  * simpler "greedy" or "lazy" parse while still being relatively fast.
1588  */
1589 static void
1590 lzx_compress_near_optimal(struct lzx_compressor *c,
1591                           struct lzx_output_bitstream *os)
1592 {
1593         const u8 * const in_begin = c->in_buffer;
1594         const u8 *       in_next = in_begin;
1595         const u8 * const in_end  = in_begin + c->in_nbytes;
1596         unsigned max_len = LZX_MAX_MATCH_LEN;
1597         unsigned nice_len = min(c->nice_match_length, max_len);
1598         u32 next_hash;
1599         struct lzx_lru_queue queue;
1600
1601         bt_matchfinder_init(&c->bt_mf);
1602         memset(c->hash2_tab, 0, sizeof(c->hash2_tab));
1603         next_hash = bt_matchfinder_hash_3_bytes(in_next);
1604         lzx_lru_queue_init(&queue);
1605
1606         do {
1607                 /* Starting a new block  */
1608                 const u8 * const in_block_begin = in_next;
1609                 const u8 * const in_block_end =
1610                         in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1611
1612                 /* Run the block through the matchfinder and cache the matches. */
1613                 struct lz_match *cache_ptr = c->match_cache;
1614                 do {
1615                         struct lz_match *lz_matchptr;
1616                         u32 hash2;
1617                         pos_t cur_match;
1618                         unsigned best_len;
1619
1620                         /* If approaching the end of the input buffer, adjust
1621                          * 'max_len' and 'nice_len' accordingly.  */
1622                         if (unlikely(max_len > in_end - in_next)) {
1623                                 max_len = in_end - in_next;
1624                                 nice_len = min(max_len, nice_len);
1625
1626                                 /* This extra check is needed to ensure that we
1627                                  * never output a length 2 match of the very
1628                                  * last two bytes with the very first two bytes,
1629                                  * since such a match has an offset too large to
1630                                  * be represented.  */
1631                                 if (unlikely(max_len < 3)) {
1632                                         in_next++;
1633                                         cache_ptr->length = 0;
1634                                         cache_ptr++;
1635                                         continue;
1636                                 }
1637                         }
1638
1639                         lz_matchptr = cache_ptr + 1;
1640
1641                         /* Check for a length 2 match.  */
1642                         hash2 = lz_hash_2_bytes(in_next, LZX_HASH2_ORDER);
1643                         cur_match = c->hash2_tab[hash2];
1644                         c->hash2_tab[hash2] = in_next - in_begin;
1645                         if (cur_match != 0 &&
1646                             (LZX_HASH2_ORDER == 16 ||
1647                              load_u16_unaligned(&in_begin[cur_match]) ==
1648                              load_u16_unaligned(in_next)))
1649                         {
1650                                 lz_matchptr->length = 2;
1651                                 lz_matchptr->offset = in_next - &in_begin[cur_match];
1652                                 lz_matchptr++;
1653                         }
1654
1655                         /* Check for matches of length >= 3.  */
1656                         lz_matchptr = bt_matchfinder_get_matches(&c->bt_mf,
1657                                                                  in_begin,
1658                                                                  in_next,
1659                                                                  3,
1660                                                                  max_len,
1661                                                                  nice_len,
1662                                                                  c->max_search_depth,
1663                                                                  &next_hash,
1664                                                                  &best_len,
1665                                                                  lz_matchptr);
1666                         in_next++;
1667                         cache_ptr->length = lz_matchptr - (cache_ptr + 1);
1668                         cache_ptr = lz_matchptr;
1669
1670                         /*
1671                          * If there was a very long match found, then don't
1672                          * cache any matches for the bytes covered by that
1673                          * match.  This avoids degenerate behavior when
1674                          * compressing highly redundant data, where the number
1675                          * of matches can be very large.
1676                          *
1677                          * This heuristic doesn't actually hurt the compression
1678                          * ratio very much.  If there's a long match, then the
1679                          * data must be highly compressible, so it doesn't
1680                          * matter as much what we do.
1681                          */
1682                         if (best_len >= nice_len) {
1683                                 --best_len;
1684                                 do {
1685                                         if (unlikely(max_len > in_end - in_next)) {
1686                                                 max_len = in_end - in_next;
1687                                                 nice_len = min(max_len, nice_len);
1688                                                 if (unlikely(max_len < 3)) {
1689                                                         in_next++;
1690                                                         cache_ptr->length = 0;
1691                                                         cache_ptr++;
1692                                                         continue;
1693                                                 }
1694                                         }
1695                                         c->hash2_tab[lz_hash_2_bytes(in_next, LZX_HASH2_ORDER)] =
1696                                                 in_next - in_begin;
1697                                         bt_matchfinder_skip_position(&c->bt_mf,
1698                                                                      in_begin,
1699                                                                      in_next,
1700                                                                      in_end,
1701                                                                      nice_len,
1702                                                                      c->max_search_depth,
1703                                                                      &next_hash);
1704                                         in_next++;
1705                                         cache_ptr->length = 0;
1706                                         cache_ptr++;
1707                                 } while (--best_len);
1708                         }
1709                 } while (in_next < in_block_end &&
1710                          likely(cache_ptr < &c->match_cache[LZX_CACHE_LENGTH]));
1711
1712                 /* We've finished running the block through the matchfinder.
1713                  * Now choose a match/literal sequence and write the block.  */
1714
1715                 queue = lzx_optimize_and_write_block(c, os, in_block_begin,
1716                                                      in_next - in_block_begin,
1717                                                      queue);
1718         } while (in_next != in_end);
1719 }
1720
1721 /*
1722  * Given a pointer to the current byte sequence and the current list of recent
1723  * match offsets, find the longest repeat offset match.
1724  *
1725  * If no match of at least 2 bytes is found, then return 0.
1726  *
1727  * If a match of at least 2 bytes is found, then return its length and set
1728  * *rep_max_idx_ret to the index of its offset in @queue.
1729 */
1730 static unsigned
1731 lzx_find_longest_repeat_offset_match(const u8 * const in_next,
1732                                      const u32 bytes_remaining,
1733                                      struct lzx_lru_queue queue,
1734                                      unsigned *rep_max_idx_ret)
1735 {
1736         STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
1737         LZX_ASSERT(bytes_remaining >= 2);
1738
1739         const unsigned max_len = min(bytes_remaining, LZX_MAX_MATCH_LEN);
1740         const u16 next_2_bytes = load_u16_unaligned(in_next);
1741         const u8 *matchptr;
1742         unsigned rep_max_len;
1743         unsigned rep_max_idx;
1744         unsigned rep_len;
1745
1746         matchptr = in_next - lzx_lru_queue_pop(&queue);
1747         if (load_u16_unaligned(matchptr) == next_2_bytes)
1748                 rep_max_len = lz_extend(in_next, matchptr, 2, max_len);
1749         else
1750                 rep_max_len = 0;
1751         rep_max_idx = 0;
1752
1753         matchptr = in_next - lzx_lru_queue_pop(&queue);
1754         if (load_u16_unaligned(matchptr) == next_2_bytes) {
1755                 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1756                 if (rep_len > rep_max_len) {
1757                         rep_max_len = rep_len;
1758                         rep_max_idx = 1;
1759                 }
1760         }
1761
1762         matchptr = in_next - lzx_lru_queue_pop(&queue);
1763         if (load_u16_unaligned(matchptr) == next_2_bytes) {
1764                 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1765                 if (rep_len > rep_max_len) {
1766                         rep_max_len = rep_len;
1767                         rep_max_idx = 2;
1768                 }
1769         }
1770
1771         *rep_max_idx_ret = rep_max_idx;
1772         return rep_max_len;
1773 }
1774
1775 /* Fast heuristic scoring for lazy parsing: how "good" is this match?  */
1776 static inline unsigned
1777 lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset)
1778 {
1779         unsigned score = len;
1780
1781         if (adjusted_offset < 4096)
1782                 score++;
1783
1784         if (adjusted_offset < 256)
1785                 score++;
1786
1787         return score;
1788 }
1789
1790 static inline unsigned
1791 lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx)
1792 {
1793         return rep_len + 3;
1794 }
1795
1796 /* This is the "lazy" LZX compressor.  */
1797 static void
1798 lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os)
1799 {
1800         const u8 * const in_begin = c->in_buffer;
1801         const u8 *       in_next = in_begin;
1802         const u8 * const in_end  = in_begin + c->in_nbytes;
1803         unsigned max_len = LZX_MAX_MATCH_LEN;
1804         unsigned nice_len = min(c->nice_match_length, max_len);
1805         struct lzx_lru_queue queue;
1806
1807         hc_matchfinder_init(&c->hc_mf);
1808         lzx_lru_queue_init(&queue);
1809
1810         do {
1811                 /* Starting a new block  */
1812
1813                 const u8 * const in_block_begin = in_next;
1814                 const u8 * const in_block_end =
1815                         in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1816                 struct lzx_item *next_chosen_item = c->chosen_items;
1817                 unsigned cur_len;
1818                 u32 cur_offset;
1819                 u32 cur_offset_data;
1820                 unsigned cur_score;
1821                 unsigned next_len;
1822                 u32 next_offset;
1823                 u32 next_offset_data;
1824                 unsigned next_score;
1825                 unsigned rep_max_len;
1826                 unsigned rep_max_idx;
1827                 unsigned rep_score;
1828                 unsigned skip_len;
1829
1830                 lzx_reset_symbol_frequencies(c);
1831
1832                 do {
1833                         if (unlikely(max_len > in_end - in_next)) {
1834                                 max_len = in_end - in_next;
1835                                 nice_len = min(max_len, nice_len);
1836                         }
1837
1838                         /* Find the longest match at the current position.  */
1839
1840                         cur_len = hc_matchfinder_longest_match(&c->hc_mf,
1841                                                                in_begin,
1842                                                                in_next,
1843                                                                2,
1844                                                                max_len,
1845                                                                nice_len,
1846                                                                c->max_search_depth,
1847                                                                &cur_offset);
1848                         if (cur_len < 3 ||
1849                             (cur_len == 3 &&
1850                              cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT &&
1851                              cur_offset != lzx_lru_queue_R0(queue) &&
1852                              cur_offset != lzx_lru_queue_R1(queue) &&
1853                              cur_offset != lzx_lru_queue_R2(queue)))
1854                         {
1855                                 /* There was no match found, or the only match found
1856                                  * was a distant length 3 match.  Output a literal.  */
1857                                 lzx_declare_literal(c, *in_next++,
1858                                                     &next_chosen_item);
1859                                 continue;
1860                         }
1861
1862                         if (cur_offset == lzx_lru_queue_R0(queue)) {
1863                                 in_next++;
1864                                 cur_offset_data = 0;
1865                                 skip_len = cur_len - 1;
1866                                 goto choose_cur_match;
1867                         }
1868
1869                         cur_offset_data = cur_offset + LZX_OFFSET_ADJUSTMENT;
1870                         cur_score = lzx_explicit_offset_match_score(cur_len, cur_offset_data);
1871
1872                         /* Consider a repeat offset match  */
1873                         rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
1874                                                                            in_end - in_next,
1875                                                                            queue,
1876                                                                            &rep_max_idx);
1877                         in_next++;
1878
1879                         if (rep_max_len >= 3 &&
1880                             (rep_score = lzx_repeat_offset_match_score(rep_max_len,
1881                                                                        rep_max_idx)) >= cur_score)
1882                         {
1883                                 cur_len = rep_max_len;
1884                                 cur_offset_data = rep_max_idx;
1885                                 skip_len = rep_max_len - 1;
1886                                 goto choose_cur_match;
1887                         }
1888
1889                 have_cur_match:
1890
1891                         /* We have a match at the current position.  */
1892
1893                         /* If we have a very long match, choose it immediately.  */
1894                         if (cur_len >= nice_len) {
1895                                 skip_len = cur_len - 1;
1896                                 goto choose_cur_match;
1897                         }
1898
1899                         /* See if there's a better match at the next position.  */
1900
1901                         if (unlikely(max_len > in_end - in_next)) {
1902                                 max_len = in_end - in_next;
1903                                 nice_len = min(max_len, nice_len);
1904                         }
1905
1906                         next_len = hc_matchfinder_longest_match(&c->hc_mf,
1907                                                                 in_begin,
1908                                                                 in_next,
1909                                                                 cur_len - 2,
1910                                                                 max_len,
1911                                                                 nice_len,
1912                                                                 c->max_search_depth / 2,
1913                                                                 &next_offset);
1914
1915                         if (next_len <= cur_len - 2) {
1916                                 in_next++;
1917                                 skip_len = cur_len - 2;
1918                                 goto choose_cur_match;
1919                         }
1920
1921                         next_offset_data = next_offset + LZX_OFFSET_ADJUSTMENT;
1922                         next_score = lzx_explicit_offset_match_score(next_len, next_offset_data);
1923
1924                         rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
1925                                                                            in_end - in_next,
1926                                                                            queue,
1927                                                                            &rep_max_idx);
1928                         in_next++;
1929
1930                         if (rep_max_len >= 3 &&
1931                             (rep_score = lzx_repeat_offset_match_score(rep_max_len,
1932                                                                        rep_max_idx)) >= next_score)
1933                         {
1934
1935                                 if (rep_score > cur_score) {
1936                                         /* The next match is better, and it's a
1937                                          * repeat offset match.  */
1938                                         lzx_declare_literal(c, *(in_next - 2),
1939                                                             &next_chosen_item);
1940                                         cur_len = rep_max_len;
1941                                         cur_offset_data = rep_max_idx;
1942                                         skip_len = cur_len - 1;
1943                                         goto choose_cur_match;
1944                                 }
1945                         } else {
1946                                 if (next_score > cur_score) {
1947                                         /* The next match is better, and it's an
1948                                          * explicit offset match.  */
1949                                         lzx_declare_literal(c, *(in_next - 2),
1950                                                             &next_chosen_item);
1951                                         cur_len = next_len;
1952                                         cur_offset_data = next_offset_data;
1953                                         cur_score = next_score;
1954                                         goto have_cur_match;
1955                                 }
1956                         }
1957
1958                         /* The original match was better.  */
1959                         skip_len = cur_len - 2;
1960
1961                 choose_cur_match:
1962                         if (cur_offset_data < LZX_NUM_RECENT_OFFSETS) {
1963                                 lzx_declare_repeat_offset_match(c, cur_len,
1964                                                                 cur_offset_data,
1965                                                                 &next_chosen_item);
1966                                 queue = lzx_lru_queue_swap(queue, cur_offset_data);
1967                         } else {
1968                                 lzx_declare_explicit_offset_match(c, cur_len,
1969                                                                   cur_offset_data - LZX_OFFSET_ADJUSTMENT,
1970                                                                   &next_chosen_item);
1971                                 queue = lzx_lru_queue_push(queue, cur_offset_data - LZX_OFFSET_ADJUSTMENT);
1972                         }
1973
1974                         hc_matchfinder_skip_positions(&c->hc_mf,
1975                                                       in_begin,
1976                                                       in_next,
1977                                                       in_end,
1978                                                       skip_len);
1979                         in_next += skip_len;
1980                 } while (in_next < in_block_end);
1981
1982                 lzx_finish_block(c, os, in_next - in_block_begin,
1983                                  next_chosen_item - c->chosen_items);
1984         } while (in_next != in_end);
1985 }
1986
1987 static void
1988 lzx_init_offset_slot_fast(struct lzx_compressor *c)
1989 {
1990         u8 slot = 0;
1991
1992         for (u32 offset = 0; offset < LZX_NUM_FAST_OFFSETS; offset++) {
1993
1994                 while (offset + LZX_OFFSET_ADJUSTMENT >= lzx_offset_slot_base[slot + 1])
1995                         slot++;
1996
1997                 c->offset_slot_fast[offset] = slot;
1998         }
1999 }
2000
2001 static size_t
2002 lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level)
2003 {
2004         if (compression_level <= LZX_MAX_FAST_LEVEL) {
2005                 return offsetof(struct lzx_compressor, hc_mf) +
2006                         hc_matchfinder_size(max_bufsize);
2007         } else {
2008                 return offsetof(struct lzx_compressor, bt_mf) +
2009                         bt_matchfinder_size(max_bufsize);
2010         }
2011 }
2012
2013 static u64
2014 lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level,
2015                       bool destructive)
2016 {
2017         u64 size = 0;
2018
2019         if (max_bufsize > LZX_MAX_WINDOW_SIZE)
2020                 return 0;
2021
2022         size += lzx_get_compressor_size(max_bufsize, compression_level);
2023         if (!destructive)
2024                 size += max_bufsize; /* in_buffer */
2025         return size;
2026 }
2027
2028 static int
2029 lzx_create_compressor(size_t max_bufsize, unsigned compression_level,
2030                       bool destructive, void **c_ret)
2031 {
2032         unsigned window_order;
2033         struct lzx_compressor *c;
2034
2035         window_order = lzx_get_window_order(max_bufsize);
2036         if (window_order == 0)
2037                 return WIMLIB_ERR_INVALID_PARAM;
2038
2039         c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level));
2040         if (!c)
2041                 goto oom0;
2042
2043         c->destructive = destructive;
2044
2045         c->num_main_syms = lzx_get_num_main_syms(window_order);
2046         c->window_order = window_order;
2047
2048         if (!c->destructive) {
2049                 c->in_buffer = MALLOC(max_bufsize);
2050                 if (!c->in_buffer)
2051                         goto oom1;
2052         }
2053
2054         if (compression_level <= LZX_MAX_FAST_LEVEL) {
2055
2056                 /* Fast compression: Use lazy parsing.  */
2057
2058                 c->impl = lzx_compress_lazy;
2059                 c->max_search_depth = (36 * compression_level) / 20;
2060                 c->nice_match_length = (72 * compression_level) / 20;
2061
2062                 /* lzx_compress_lazy() needs max_search_depth >= 2 because it
2063                  * halves the max_search_depth when attempting a lazy match, and
2064                  * max_search_depth cannot be 0.  */
2065                 if (c->max_search_depth < 2)
2066                         c->max_search_depth = 2;
2067         } else {
2068
2069                 /* Normal / high compression: Use near-optimal parsing.  */
2070
2071                 c->impl = lzx_compress_near_optimal;
2072
2073                 /* Scale nice_match_length and max_search_depth with the
2074                  * compression level.  */
2075                 c->max_search_depth = (24 * compression_level) / 50;
2076                 c->nice_match_length = (32 * compression_level) / 50;
2077
2078                 /* Set a number of optimization passes appropriate for the
2079                  * compression level.  */
2080
2081                 c->num_optim_passes = 1;
2082
2083                 if (compression_level >= 45)
2084                         c->num_optim_passes++;
2085
2086                 /* Use more optimization passes for higher compression levels.
2087                  * But the more passes there are, the less they help --- so
2088                  * don't add them linearly.  */
2089                 if (compression_level >= 70) {
2090                         c->num_optim_passes++;
2091                         if (compression_level >= 100)
2092                                 c->num_optim_passes++;
2093                         if (compression_level >= 150)
2094                                 c->num_optim_passes++;
2095                         if (compression_level >= 200)
2096                                 c->num_optim_passes++;
2097                         if (compression_level >= 300)
2098                                 c->num_optim_passes++;
2099                 }
2100         }
2101
2102         /* max_search_depth == 0 is invalid.  */
2103         if (c->max_search_depth < 1)
2104                 c->max_search_depth = 1;
2105
2106         if (c->nice_match_length > LZX_MAX_MATCH_LEN)
2107                 c->nice_match_length = LZX_MAX_MATCH_LEN;
2108
2109         lzx_init_offset_slot_fast(c);
2110         *c_ret = c;
2111         return 0;
2112
2113 oom1:
2114         FREE(c);
2115 oom0:
2116         return WIMLIB_ERR_NOMEM;
2117 }
2118
2119 static size_t
2120 lzx_compress(const void *restrict in, size_t in_nbytes,
2121              void *restrict out, size_t out_nbytes_avail, void *restrict _c)
2122 {
2123         struct lzx_compressor *c = _c;
2124         struct lzx_output_bitstream os;
2125         size_t result;
2126
2127         /* Don't bother trying to compress very small inputs.  */
2128         if (in_nbytes < 100)
2129                 return 0;
2130
2131         /* Copy the input data into the internal buffer and preprocess it.  */
2132         if (c->destructive)
2133                 c->in_buffer = (void *)in;
2134         else
2135                 memcpy(c->in_buffer, in, in_nbytes);
2136         c->in_nbytes = in_nbytes;
2137         lzx_do_e8_preprocessing(c->in_buffer, in_nbytes);
2138
2139         /* Initially, the previous Huffman codeword lengths are all zeroes.  */
2140         c->codes_index = 0;
2141         memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens));
2142
2143         /* Initialize the output bitstream.  */
2144         lzx_init_output(&os, out, out_nbytes_avail);
2145
2146         /* Call the compression level-specific compress() function.  */
2147         (*c->impl)(c, &os);
2148
2149         /* Flush the output bitstream and return the compressed size or 0.  */
2150         result = lzx_flush_output(&os);
2151         if (!result && c->destructive)
2152                 lzx_undo_e8_preprocessing(c->in_buffer, c->in_nbytes);
2153         return result;
2154 }
2155
2156 static void
2157 lzx_free_compressor(void *_c)
2158 {
2159         struct lzx_compressor *c = _c;
2160
2161         if (!c->destructive)
2162                 FREE(c->in_buffer);
2163         FREE(c);
2164 }
2165
2166 const struct compressor_ops lzx_compressor_ops = {
2167         .get_needed_memory  = lzx_get_needed_memory,
2168         .create_compressor  = lzx_create_compressor,
2169         .compress           = lzx_compress,
2170         .free_compressor    = lzx_free_compressor,
2171 };