4434cde34a02bd13a8b1685765207c2d8b15bee1
[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       6
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  * The maximum compression level at which we use the faster algorithm.
124  */
125 #define LZX_MAX_FAST_LEVEL      34
126
127 /*
128  * LZX_HASH2_ORDER is the log base 2 of the number of entries in the hash table
129  * for finding length 2 matches.  This can be as high as 16 (in which case the
130  * hash function is trivial), but using a smaller hash table actually speeds up
131  * compression due to reduced cache pressure.
132  */
133 #define LZX_HASH2_ORDER         12
134 #define LZX_HASH2_LENGTH        (1UL << LZX_HASH2_ORDER)
135
136 #include "wimlib/lzx_common.h"
137
138 /*
139  * The maximum allowed window order for the matchfinder.
140  */
141 #define MATCHFINDER_MAX_WINDOW_ORDER    LZX_MAX_WINDOW_ORDER
142
143 #include <string.h>
144
145 #include "wimlib/bt_matchfinder.h"
146 #include "wimlib/compress_common.h"
147 #include "wimlib/compressor_ops.h"
148 #include "wimlib/endianness.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         /* The Huffman symbol frequency counters for the current block.  */
372         struct lzx_freqs freqs;
373
374         /* The Huffman codes for the current and previous blocks.  The one with
375          * index 'codes_index' is for the current block, and the other one is
376          * for the previous block.  */
377         struct lzx_codes codes[2];
378         unsigned codes_index;
379
380         /*
381          * The match/literal sequence the algorithm chose for the current block.
382          *
383          * Notes on how large this array actually needs to be:
384          *
385          * - In lzx_compress_near_optimal(), the maximum block size is
386          *   'LZX_DIV_BLOCK_SIZE + LZX_MAX_MATCH_LEN - 1' bytes.  This occurs if
387          *   a match of the maximum length is found on the last byte.  Although
388          *   it is impossible for this particular case to actually result in a
389          *   parse of all literals, we reserve this many spaces anyway.
390          *
391          * - The worst case for lzx_compress_lazy() is a block of almost all
392          *   literals that ends with a series of matches of increasing scores,
393          *   causing a sequence of literals to be chosen before the last match
394          *   is finally chosen.  The number of items actually chosen in this
395          *   scenario is limited by the number of distinct match scores that
396          *   exist for matches shorter than 'nice_match_length'.  Having
397          *   'LZX_MAX_MATCH_LEN - 1' extra spaces is plenty for now.
398          */
399         struct lzx_item chosen_items[LZX_DIV_BLOCK_SIZE + LZX_MAX_MATCH_LEN - 1];
400
401         /* Table mapping match offset => offset slot for small offsets  */
402 #define LZX_NUM_FAST_OFFSETS 32768
403         u8 offset_slot_fast[LZX_NUM_FAST_OFFSETS];
404
405         union {
406                 /* Data for greedy or lazy parsing  */
407                 struct {
408                         /* Hash chains matchfinder (MUST BE LAST!!!)  */
409                         struct hc_matchfinder hc_mf;
410                 };
411
412                 /* Data for near-optimal parsing  */
413                 struct {
414                         /*
415                          * The graph nodes for the current block.
416                          *
417                          * We need at least 'LZX_DIV_BLOCK_SIZE +
418                          * LZX_MAX_MATCH_LEN - 1' nodes because that is the
419                          * maximum block size that may be used.  Add 1 because
420                          * we need a node to represent end-of-block.
421                          *
422                          * It is possible that nodes past end-of-block are
423                          * accessed during match consideration, but this can
424                          * only occur if the block was truncated at
425                          * LZX_DIV_BLOCK_SIZE.  So the same bound still applies.
426                          * Note that since nodes past the end of the block will
427                          * never actually have an effect on the items that are
428                          * chosen for the block, it makes no difference what
429                          * their costs are initialized to (if anything).
430                          */
431                         struct lzx_optimum_node optimum_nodes[LZX_DIV_BLOCK_SIZE +
432                                                               LZX_MAX_MATCH_LEN - 1 + 1];
433
434                         /* The cost model for the current block  */
435                         struct lzx_costs costs;
436
437                         /*
438                          * Cached matches for the current block.  This array
439                          * contains the matches that were found at each position
440                          * in the block.  Specifically, for each position, there
441                          * is a special 'struct lz_match' whose 'length' field
442                          * contains the number of matches that were found at
443                          * that position; this is followed by the matches
444                          * themselves, if any, sorted by strictly increasing
445                          * length and strictly increasing offset.
446                          *
447                          * Note: in rare cases, there will be a very high number
448                          * of matches in the block and this array will overflow.
449                          * If this happens, we force the end of the current
450                          * block.  LZX_CACHE_LENGTH is the length at which we
451                          * actually check for overflow.  The extra slots beyond
452                          * this are enough to absorb the worst case overflow,
453                          * which occurs if starting at
454                          * &match_cache[LZX_CACHE_LENGTH - 1], we write the
455                          * match count header, then write
456                          * LZX_MAX_MATCHES_PER_POS matches, then skip searching
457                          * for matches at 'LZX_MAX_MATCH_LEN - 1' positions and
458                          * write the match count header for each.
459                          */
460                         struct lz_match match_cache[LZX_CACHE_LENGTH +
461                                                     LZX_MAX_MATCHES_PER_POS +
462                                                     LZX_MAX_MATCH_LEN - 1];
463
464                         /* Hash table for finding length 2 matches  */
465                         pos_t hash2_tab[LZX_HASH2_LENGTH]
466                                 _aligned_attribute(MATCHFINDER_ALIGNMENT);
467
468                         /* Binary trees matchfinder (MUST BE LAST!!!)  */
469                         struct bt_matchfinder bt_mf;
470                 };
471         };
472 };
473
474 /*
475  * Structure to keep track of the current state of sending bits to the
476  * compressed output buffer.
477  *
478  * The LZX bitstream is encoded as a sequence of 16-bit coding units.
479  */
480 struct lzx_output_bitstream {
481
482         /* Bits that haven't yet been written to the output buffer.  */
483         u32 bitbuf;
484
485         /* Number of bits currently held in @bitbuf.  */
486         u32 bitcount;
487
488         /* Pointer to the start of the output buffer.  */
489         le16 *start;
490
491         /* Pointer to the position in the output buffer at which the next coding
492          * unit should be written.  */
493         le16 *next;
494
495         /* Pointer past the end of the output buffer.  */
496         le16 *end;
497 };
498
499 /*
500  * Initialize the output bitstream.
501  *
502  * @os
503  *      The output bitstream structure to initialize.
504  * @buffer
505  *      The buffer being written to.
506  * @size
507  *      Size of @buffer, in bytes.
508  */
509 static void
510 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size)
511 {
512         os->bitbuf = 0;
513         os->bitcount = 0;
514         os->start = buffer;
515         os->next = os->start;
516         os->end = os->start + size / sizeof(le16);
517 }
518
519 /*
520  * Write some bits to the output bitstream.
521  *
522  * The bits are given by the low-order @num_bits bits of @bits.  Higher-order
523  * bits in @bits cannot be set.  At most 17 bits can be written at once.
524  *
525  * @max_num_bits is a compile-time constant that specifies the maximum number of
526  * bits that can ever be written at the call site.  Currently, it is used to
527  * optimize away the conditional code for writing a second 16-bit coding unit
528  * when writing fewer than 17 bits.
529  *
530  * If the output buffer space is exhausted, then the bits will be ignored, and
531  * lzx_flush_output() will return 0 when it gets called.
532  */
533 static inline void
534 lzx_write_varbits(struct lzx_output_bitstream *os,
535                   const u32 bits, const unsigned num_bits,
536                   const unsigned max_num_bits)
537 {
538         /* This code is optimized for LZX, which never needs to write more than
539          * 17 bits at once.  */
540         LZX_ASSERT(num_bits <= 17);
541         LZX_ASSERT(num_bits <= max_num_bits);
542         LZX_ASSERT(os->bitcount <= 15);
543
544         /* Add the bits to the bit buffer variable.  @bitcount will be at most
545          * 15, so there will be just enough space for the maximum possible
546          * @num_bits of 17.  */
547         os->bitcount += num_bits;
548         os->bitbuf = (os->bitbuf << num_bits) | bits;
549
550         /* Check whether any coding units need to be written.  */
551         if (os->bitcount >= 16) {
552
553                 os->bitcount -= 16;
554
555                 /* Write a coding unit, unless it would overflow the buffer.  */
556                 if (os->next != os->end)
557                         put_unaligned_u16_le(os->bitbuf >> os->bitcount, os->next++);
558
559                 /* If writing 17 bits, a second coding unit might need to be
560                  * written.  But because 'max_num_bits' is a compile-time
561                  * constant, the compiler will optimize away this code at most
562                  * call sites.  */
563                 if (max_num_bits == 17 && os->bitcount == 16) {
564                         if (os->next != os->end)
565                                 put_unaligned_u16_le(os->bitbuf, os->next++);
566                         os->bitcount = 0;
567                 }
568         }
569 }
570
571 /* Use when @num_bits is a compile-time constant.  Otherwise use
572  * lzx_write_varbits().  */
573 static inline void
574 lzx_write_bits(struct lzx_output_bitstream *os,
575                const u32 bits, const unsigned num_bits)
576 {
577         lzx_write_varbits(os, bits, num_bits, num_bits);
578 }
579
580 /*
581  * Flush the last coding unit to the output buffer if needed.  Return the total
582  * number of bytes written to the output buffer, or 0 if an overflow occurred.
583  */
584 static u32
585 lzx_flush_output(struct lzx_output_bitstream *os)
586 {
587         if (os->next == os->end)
588                 return 0;
589
590         if (os->bitcount != 0)
591                 put_unaligned_u16_le(os->bitbuf << (16 - os->bitcount), os->next++);
592
593         return (const u8 *)os->next - (const u8 *)os->start;
594 }
595
596 /* Build the main, length, and aligned offset Huffman codes used in LZX.
597  *
598  * This takes as input the frequency tables for each code and produces as output
599  * a set of tables that map symbols to codewords and codeword lengths.  */
600 static void
601 lzx_make_huffman_codes(struct lzx_compressor *c)
602 {
603         const struct lzx_freqs *freqs = &c->freqs;
604         struct lzx_codes *codes = &c->codes[c->codes_index];
605
606         make_canonical_huffman_code(c->num_main_syms,
607                                     LZX_MAX_MAIN_CODEWORD_LEN,
608                                     freqs->main,
609                                     codes->lens.main,
610                                     codes->codewords.main);
611
612         make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
613                                     LZX_MAX_LEN_CODEWORD_LEN,
614                                     freqs->len,
615                                     codes->lens.len,
616                                     codes->codewords.len);
617
618         make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
619                                     LZX_MAX_ALIGNED_CODEWORD_LEN,
620                                     freqs->aligned,
621                                     codes->lens.aligned,
622                                     codes->codewords.aligned);
623 }
624
625 /* Reset the symbol frequencies for the LZX Huffman codes.  */
626 static void
627 lzx_reset_symbol_frequencies(struct lzx_compressor *c)
628 {
629         memset(&c->freqs, 0, sizeof(c->freqs));
630 }
631
632 static unsigned
633 lzx_compute_precode_items(const u8 lens[restrict],
634                           const u8 prev_lens[restrict],
635                           const unsigned num_lens,
636                           u32 precode_freqs[restrict],
637                           unsigned precode_items[restrict])
638 {
639         unsigned *itemptr;
640         unsigned run_start;
641         unsigned run_end;
642         unsigned extra_bits;
643         int delta;
644         u8 len;
645
646         itemptr = precode_items;
647         run_start = 0;
648         do {
649                 /* Find the next run of codeword lengths.  */
650
651                 /* len = the length being repeated  */
652                 len = lens[run_start];
653
654                 run_end = run_start + 1;
655
656                 /* Fast case for a single length.  */
657                 if (likely(run_end == num_lens || len != lens[run_end])) {
658                         delta = prev_lens[run_start] - len;
659                         if (delta < 0)
660                                 delta += 17;
661                         precode_freqs[delta]++;
662                         *itemptr++ = delta;
663                         run_start++;
664                         continue;
665                 }
666
667                 /* Extend the run.  */
668                 do {
669                         run_end++;
670                 } while (run_end != num_lens && len == lens[run_end]);
671
672                 if (len == 0) {
673                         /* Run of zeroes.  */
674
675                         /* Symbol 18: RLE 20 to 51 zeroes at a time.  */
676                         while ((run_end - run_start) >= 20) {
677                                 extra_bits = min((run_end - run_start) - 20, 0x1f);
678                                 precode_freqs[18]++;
679                                 *itemptr++ = 18 | (extra_bits << 5);
680                                 run_start += 20 + extra_bits;
681                         }
682
683                         /* Symbol 17: RLE 4 to 19 zeroes at a time.  */
684                         if ((run_end - run_start) >= 4) {
685                                 extra_bits = min((run_end - run_start) - 4, 0xf);
686                                 precode_freqs[17]++;
687                                 *itemptr++ = 17 | (extra_bits << 5);
688                                 run_start += 4 + extra_bits;
689                         }
690                 } else {
691
692                         /* A run of nonzero lengths. */
693
694                         /* Symbol 19: RLE 4 to 5 of any length at a time.  */
695                         while ((run_end - run_start) >= 4) {
696                                 extra_bits = (run_end - run_start) > 4;
697                                 delta = prev_lens[run_start] - len;
698                                 if (delta < 0)
699                                         delta += 17;
700                                 precode_freqs[19]++;
701                                 precode_freqs[delta]++;
702                                 *itemptr++ = 19 | (extra_bits << 5) | (delta << 6);
703                                 run_start += 4 + extra_bits;
704                         }
705                 }
706
707                 /* Output any remaining lengths without RLE.  */
708                 while (run_start != run_end) {
709                         delta = prev_lens[run_start] - len;
710                         if (delta < 0)
711                                 delta += 17;
712                         precode_freqs[delta]++;
713                         *itemptr++ = delta;
714                         run_start++;
715                 }
716         } while (run_start != num_lens);
717
718         return itemptr - precode_items;
719 }
720
721 /*
722  * Output a Huffman code in the compressed form used in LZX.
723  *
724  * The Huffman code is represented in the output as a logical series of codeword
725  * lengths from which the Huffman code, which must be in canonical form, can be
726  * reconstructed.
727  *
728  * The codeword lengths are themselves compressed using a separate Huffman code,
729  * the "precode", which contains a symbol for each possible codeword length in
730  * the larger code as well as several special symbols to represent repeated
731  * codeword lengths (a form of run-length encoding).  The precode is itself
732  * constructed in canonical form, and its codeword lengths are represented
733  * literally in 20 4-bit fields that immediately precede the compressed codeword
734  * lengths of the larger code.
735  *
736  * Furthermore, the codeword lengths of the larger code are actually represented
737  * as deltas from the codeword lengths of the corresponding code in the previous
738  * block.
739  *
740  * @os:
741  *      Bitstream to which to write the compressed Huffman code.
742  * @lens:
743  *      The codeword lengths, indexed by symbol, in the Huffman code.
744  * @prev_lens:
745  *      The codeword lengths, indexed by symbol, in the corresponding Huffman
746  *      code in the previous block, or all zeroes if this is the first block.
747  * @num_lens:
748  *      The number of symbols in the Huffman code.
749  */
750 static void
751 lzx_write_compressed_code(struct lzx_output_bitstream *os,
752                           const u8 lens[restrict],
753                           const u8 prev_lens[restrict],
754                           unsigned num_lens)
755 {
756         u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
757         u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
758         u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
759         unsigned precode_items[num_lens];
760         unsigned num_precode_items;
761         unsigned precode_item;
762         unsigned precode_sym;
763         unsigned i;
764
765         for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
766                 precode_freqs[i] = 0;
767
768         /* Compute the "items" (RLE / literal tokens and extra bits) with which
769          * the codeword lengths in the larger code will be output.  */
770         num_precode_items = lzx_compute_precode_items(lens,
771                                                       prev_lens,
772                                                       num_lens,
773                                                       precode_freqs,
774                                                       precode_items);
775
776         /* Build the precode.  */
777         make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
778                                     LZX_MAX_PRE_CODEWORD_LEN,
779                                     precode_freqs, precode_lens,
780                                     precode_codewords);
781
782         /* Output the lengths of the codewords in the precode.  */
783         for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
784                 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
785
786         /* Output the encoded lengths of the codewords in the larger code.  */
787         for (i = 0; i < num_precode_items; i++) {
788                 precode_item = precode_items[i];
789                 precode_sym = precode_item & 0x1F;
790                 lzx_write_varbits(os, precode_codewords[precode_sym],
791                                   precode_lens[precode_sym],
792                                   LZX_MAX_PRE_CODEWORD_LEN);
793                 if (precode_sym >= 17) {
794                         if (precode_sym == 17) {
795                                 lzx_write_bits(os, precode_item >> 5, 4);
796                         } else if (precode_sym == 18) {
797                                 lzx_write_bits(os, precode_item >> 5, 5);
798                         } else {
799                                 lzx_write_bits(os, (precode_item >> 5) & 1, 1);
800                                 precode_sym = precode_item >> 6;
801                                 lzx_write_varbits(os, precode_codewords[precode_sym],
802                                                   precode_lens[precode_sym],
803                                                   LZX_MAX_PRE_CODEWORD_LEN);
804                         }
805                 }
806         }
807 }
808
809 /* Output a match or literal.  */
810 static inline void
811 lzx_write_item(struct lzx_output_bitstream *os, struct lzx_item item,
812                unsigned ones_if_aligned, const struct lzx_codes *codes)
813 {
814         u64 data = item.data;
815         unsigned main_symbol;
816         unsigned len_symbol;
817         unsigned num_extra_bits;
818         u32 extra_bits;
819
820         main_symbol = data & 0x3FF;
821
822         lzx_write_varbits(os, codes->codewords.main[main_symbol],
823                           codes->lens.main[main_symbol],
824                           LZX_MAX_MAIN_CODEWORD_LEN);
825
826         if (main_symbol < LZX_NUM_CHARS)  /* Literal?  */
827                 return;
828
829         len_symbol = (data >> 10) & 0xFF;
830
831         if (len_symbol != LZX_LENCODE_NUM_SYMBOLS) {
832                 lzx_write_varbits(os, codes->codewords.len[len_symbol],
833                                   codes->lens.len[len_symbol],
834                                   LZX_MAX_LEN_CODEWORD_LEN);
835         }
836
837         num_extra_bits = (data >> 18) & 0x1F;
838         if (num_extra_bits == 0)  /* Small offset or repeat offset match?  */
839                 return;
840
841         extra_bits = data >> 23;
842
843         if ((num_extra_bits & ones_if_aligned) >= LZX_NUM_ALIGNED_OFFSET_BITS) {
844
845                 /* Aligned offset blocks: The low 3 bits of the extra offset
846                  * bits are Huffman-encoded using the aligned offset code.  The
847                  * remaining bits are output literally.  */
848
849                 lzx_write_varbits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS,
850                                   num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS,
851                                   17 - LZX_NUM_ALIGNED_OFFSET_BITS);
852
853                 lzx_write_varbits(os,
854                                   codes->codewords.aligned[extra_bits & LZX_ALIGNED_OFFSET_BITMASK],
855                                   codes->lens.aligned[extra_bits & LZX_ALIGNED_OFFSET_BITMASK],
856                                   LZX_MAX_ALIGNED_CODEWORD_LEN);
857         } else {
858                 /* Verbatim blocks, or fewer than 3 extra bits:  All extra
859                  * offset bits are output literally.  */
860                 lzx_write_varbits(os, extra_bits, num_extra_bits, 17);
861         }
862 }
863
864 /*
865  * Write all matches and literal bytes (which were precomputed) in an LZX
866  * compressed block to the output bitstream in the final compressed
867  * representation.
868  *
869  * @os
870  *      The output bitstream.
871  * @block_type
872  *      The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
873  *      LZX_BLOCKTYPE_VERBATIM).
874  * @items
875  *      The array of matches/literals to output.
876  * @num_items
877  *      Number of matches/literals to output (length of @items).
878  * @codes
879  *      The main, length, and aligned offset Huffman codes for the current
880  *      LZX compressed block.
881  */
882 static void
883 lzx_write_items(struct lzx_output_bitstream *os, int block_type,
884                 const struct lzx_item items[], u32 num_items,
885                 const struct lzx_codes *codes)
886 {
887         unsigned ones_if_aligned = 0U - (block_type == LZX_BLOCKTYPE_ALIGNED);
888
889         for (u32 i = 0; i < num_items; i++)
890                 lzx_write_item(os, items[i], ones_if_aligned, codes);
891 }
892
893 static void
894 lzx_write_compressed_block(int block_type,
895                            u32 block_size,
896                            unsigned window_order,
897                            unsigned num_main_syms,
898                            const struct lzx_item chosen_items[],
899                            u32 num_chosen_items,
900                            const struct lzx_codes * codes,
901                            const struct lzx_lens * prev_lens,
902                            struct lzx_output_bitstream * os)
903 {
904         LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
905                    block_type == LZX_BLOCKTYPE_VERBATIM);
906
907         /* The first three bits indicate the type of block and are one of the
908          * LZX_BLOCKTYPE_* constants.  */
909         lzx_write_bits(os, block_type, 3);
910
911         /* Output the block size.
912          *
913          * The original LZX format seemed to always encode the block size in 3
914          * bytes.  However, the implementation in WIMGAPI, as used in WIM files,
915          * uses the first bit to indicate whether the block is the default size
916          * (32768) or a different size given explicitly by the next 16 bits.
917          *
918          * By default, this compressor uses a window size of 32768 and therefore
919          * follows the WIMGAPI behavior.  However, this compressor also supports
920          * window sizes greater than 32768 bytes, which do not appear to be
921          * supported by WIMGAPI.  In such cases, we retain the default size bit
922          * to mean a size of 32768 bytes but output non-default block size in 24
923          * bits rather than 16.  The compatibility of this behavior is unknown
924          * because WIMs created with chunk size greater than 32768 can seemingly
925          * only be opened by wimlib anyway.  */
926         if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
927                 lzx_write_bits(os, 1, 1);
928         } else {
929                 lzx_write_bits(os, 0, 1);
930
931                 if (window_order >= 16)
932                         lzx_write_bits(os, block_size >> 16, 8);
933
934                 lzx_write_bits(os, block_size & 0xFFFF, 16);
935         }
936
937         /* If it's an aligned offset block, output the aligned offset code.  */
938         if (block_type == LZX_BLOCKTYPE_ALIGNED) {
939                 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
940                         lzx_write_bits(os, codes->lens.aligned[i],
941                                        LZX_ALIGNEDCODE_ELEMENT_SIZE);
942                 }
943         }
944
945         /* Output the main code (two parts).  */
946         lzx_write_compressed_code(os, codes->lens.main,
947                                   prev_lens->main,
948                                   LZX_NUM_CHARS);
949         lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
950                                   prev_lens->main + LZX_NUM_CHARS,
951                                   num_main_syms - LZX_NUM_CHARS);
952
953         /* Output the length code.  */
954         lzx_write_compressed_code(os, codes->lens.len,
955                                   prev_lens->len,
956                                   LZX_LENCODE_NUM_SYMBOLS);
957
958         /* Output the compressed matches and literals.  */
959         lzx_write_items(os, block_type, chosen_items, num_chosen_items, codes);
960 }
961
962 /* Given the frequencies of symbols in an LZX-compressed block and the
963  * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
964  * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
965  * will take fewer bits to output.  */
966 static int
967 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
968                                const struct lzx_codes * codes)
969 {
970         u32 aligned_cost = 0;
971         u32 verbatim_cost = 0;
972
973         /* A verbatim block requires 3 bits in each place that an aligned symbol
974          * would be used in an aligned offset block.  */
975         for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
976                 verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i];
977                 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
978         }
979
980         /* Account for output of the aligned offset code.  */
981         aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
982
983         if (aligned_cost < verbatim_cost)
984                 return LZX_BLOCKTYPE_ALIGNED;
985         else
986                 return LZX_BLOCKTYPE_VERBATIM;
987 }
988
989 /*
990  * Finish an LZX block:
991  *
992  * - build the Huffman codes
993  * - decide whether to output the block as VERBATIM or ALIGNED
994  * - output the block
995  * - swap the indices of the current and previous Huffman codes
996  */
997 static void
998 lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os,
999                  u32 block_size, u32 num_chosen_items)
1000 {
1001         int block_type;
1002
1003         lzx_make_huffman_codes(c);
1004
1005         block_type = lzx_choose_verbatim_or_aligned(&c->freqs,
1006                                                     &c->codes[c->codes_index]);
1007         lzx_write_compressed_block(block_type,
1008                                    block_size,
1009                                    c->window_order,
1010                                    c->num_main_syms,
1011                                    c->chosen_items,
1012                                    num_chosen_items,
1013                                    &c->codes[c->codes_index],
1014                                    &c->codes[c->codes_index ^ 1].lens,
1015                                    os);
1016         c->codes_index ^= 1;
1017 }
1018
1019 /* Return the offset slot for the specified offset, which must be
1020  * less than LZX_NUM_FAST_OFFSETS.  */
1021 static inline unsigned
1022 lzx_get_offset_slot_fast(struct lzx_compressor *c, u32 offset)
1023 {
1024         LZX_ASSERT(offset < LZX_NUM_FAST_OFFSETS);
1025         return c->offset_slot_fast[offset];
1026 }
1027
1028 /* Tally, and optionally record, the specified literal byte.  */
1029 static inline void
1030 lzx_declare_literal(struct lzx_compressor *c, unsigned literal,
1031                     struct lzx_item **next_chosen_item)
1032 {
1033         unsigned main_symbol = lzx_main_symbol_for_literal(literal);
1034
1035         c->freqs.main[main_symbol]++;
1036
1037         if (next_chosen_item) {
1038                 *(*next_chosen_item)++ = (struct lzx_item) {
1039                         .data = main_symbol,
1040                 };
1041         }
1042 }
1043
1044 /* Tally, and optionally record, the specified repeat offset match.  */
1045 static inline void
1046 lzx_declare_repeat_offset_match(struct lzx_compressor *c,
1047                                 unsigned len, unsigned rep_index,
1048                                 struct lzx_item **next_chosen_item)
1049 {
1050         unsigned len_header;
1051         unsigned len_symbol;
1052         unsigned main_symbol;
1053
1054         if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) {
1055                 len_header = len - LZX_MIN_MATCH_LEN;
1056                 len_symbol = LZX_LENCODE_NUM_SYMBOLS;
1057         } else {
1058                 len_header = LZX_NUM_PRIMARY_LENS;
1059                 len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS;
1060                 c->freqs.len[len_symbol]++;
1061         }
1062
1063         main_symbol = lzx_main_symbol_for_match(rep_index, len_header);
1064
1065         c->freqs.main[main_symbol]++;
1066
1067         if (next_chosen_item) {
1068                 *(*next_chosen_item)++ = (struct lzx_item) {
1069                         .data = (u64)main_symbol | ((u64)len_symbol << 10),
1070                 };
1071         }
1072 }
1073
1074 /* Tally, and optionally record, the specified explicit offset match.  */
1075 static inline void
1076 lzx_declare_explicit_offset_match(struct lzx_compressor *c, unsigned len, u32 offset,
1077                                   struct lzx_item **next_chosen_item)
1078 {
1079         unsigned len_header;
1080         unsigned len_symbol;
1081         unsigned main_symbol;
1082         unsigned offset_slot;
1083         unsigned num_extra_bits;
1084         u32 extra_bits;
1085
1086         if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) {
1087                 len_header = len - LZX_MIN_MATCH_LEN;
1088                 len_symbol = LZX_LENCODE_NUM_SYMBOLS;
1089         } else {
1090                 len_header = LZX_NUM_PRIMARY_LENS;
1091                 len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS;
1092                 c->freqs.len[len_symbol]++;
1093         }
1094
1095         offset_slot = (offset < LZX_NUM_FAST_OFFSETS) ?
1096                         lzx_get_offset_slot_fast(c, offset) :
1097                         lzx_get_offset_slot(offset);
1098
1099         main_symbol = lzx_main_symbol_for_match(offset_slot, len_header);
1100
1101         c->freqs.main[main_symbol]++;
1102
1103         num_extra_bits = lzx_extra_offset_bits[offset_slot];
1104
1105         if (num_extra_bits >= LZX_NUM_ALIGNED_OFFSET_BITS)
1106                 c->freqs.aligned[(offset + LZX_OFFSET_ADJUSTMENT) &
1107                                  LZX_ALIGNED_OFFSET_BITMASK]++;
1108
1109         if (next_chosen_item) {
1110
1111                 extra_bits = (offset + LZX_OFFSET_ADJUSTMENT) -
1112                              lzx_offset_slot_base[offset_slot];
1113
1114                 BUILD_BUG_ON(LZX_MAINCODE_MAX_NUM_SYMBOLS > (1 << 10));
1115                 BUILD_BUG_ON(LZX_LENCODE_NUM_SYMBOLS > (1 << 8));
1116                 *(*next_chosen_item)++ = (struct lzx_item) {
1117                         .data = (u64)main_symbol |
1118                                 ((u64)len_symbol << 10) |
1119                                 ((u64)num_extra_bits << 18) |
1120                                 ((u64)extra_bits << 23),
1121                 };
1122         }
1123 }
1124
1125
1126 /* Tally, and optionally record, the specified match or literal.  */
1127 static inline void
1128 lzx_declare_item(struct lzx_compressor *c, u32 item,
1129                  struct lzx_item **next_chosen_item)
1130 {
1131         u32 len = item & OPTIMUM_LEN_MASK;
1132         u32 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1133
1134         if (len == 1)
1135                 lzx_declare_literal(c, offset_data, next_chosen_item);
1136         else if (offset_data < LZX_NUM_RECENT_OFFSETS)
1137                 lzx_declare_repeat_offset_match(c, len, offset_data,
1138                                                 next_chosen_item);
1139         else
1140                 lzx_declare_explicit_offset_match(c, len,
1141                                                   offset_data - LZX_OFFSET_ADJUSTMENT,
1142                                                   next_chosen_item);
1143 }
1144
1145 static inline void
1146 lzx_record_item_list(struct lzx_compressor *c,
1147                      struct lzx_optimum_node *cur_node,
1148                      struct lzx_item **next_chosen_item)
1149 {
1150         struct lzx_optimum_node *end_node;
1151         u32 saved_item;
1152         u32 item;
1153
1154         /* The list is currently in reverse order (last item to first item).
1155          * Reverse it.  */
1156         end_node = cur_node;
1157         saved_item = cur_node->item;
1158         do {
1159                 item = saved_item;
1160                 cur_node -= item & OPTIMUM_LEN_MASK;
1161                 saved_item = cur_node->item;
1162                 cur_node->item = item;
1163         } while (cur_node != c->optimum_nodes);
1164
1165         /* Walk the list of items from beginning to end, tallying and recording
1166          * each item.  */
1167         do {
1168                 lzx_declare_item(c, cur_node->item, next_chosen_item);
1169                 cur_node += (cur_node->item) & OPTIMUM_LEN_MASK;
1170         } while (cur_node != end_node);
1171 }
1172
1173 static inline void
1174 lzx_tally_item_list(struct lzx_compressor *c, struct lzx_optimum_node *cur_node)
1175 {
1176         /* Since we're just tallying the items, we don't need to reverse the
1177          * list.  Processing the items in reverse order is fine.  */
1178         do {
1179                 lzx_declare_item(c, cur_node->item, NULL);
1180                 cur_node -= (cur_node->item & OPTIMUM_LEN_MASK);
1181         } while (cur_node != c->optimum_nodes);
1182 }
1183
1184 /*
1185  * Find an inexpensive path through the graph of possible match/literal choices
1186  * for the current block.  The nodes of the graph are
1187  * c->optimum_nodes[0...block_size].  They correspond directly to the bytes in
1188  * the current block, plus one extra node for end-of-block.  The edges of the
1189  * graph are matches and literals.  The goal is to find the minimum cost path
1190  * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]'.
1191  *
1192  * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and
1193  * proceeding forwards one node at a time.  At each node, a selection of matches
1194  * (len >= 2), as well as the literal byte (len = 1), is considered.  An item of
1195  * length 'len' provides a new path to reach the node 'len' bytes later.  If
1196  * such a path is the lowest cost found so far to reach that later node, then
1197  * that later node is updated with the new path.
1198  *
1199  * Note that although this algorithm is based on minimum cost path search, due
1200  * to various simplifying assumptions the result is not guaranteed to be the
1201  * true minimum cost, or "optimal", path over the graph of all valid LZX
1202  * representations of this block.
1203  *
1204  * Also, note that because of the presence of the recent offsets queue (which is
1205  * a type of adaptive state), the algorithm cannot work backwards and compute
1206  * "cost to end" instead of "cost to beginning".  Furthermore, the way the
1207  * algorithm handles this adaptive state in the "minimum-cost" parse is actually
1208  * only an approximation.  It's possible for the globally optimal, minimum cost
1209  * path to contain a prefix, ending at a position, where that path prefix is
1210  * *not* the minimum cost path to that position.  This can happen if such a path
1211  * prefix results in a different adaptive state which results in lower costs
1212  * later.  The algorithm does not solve this problem; it only considers the
1213  * lowest cost to reach each individual position.
1214  */
1215 static struct lzx_lru_queue
1216 lzx_find_min_cost_path(struct lzx_compressor * const restrict c,
1217                        const u8 * const restrict block_begin,
1218                        const u32 block_size,
1219                        const struct lzx_lru_queue initial_queue)
1220 {
1221         struct lzx_optimum_node *cur_node = c->optimum_nodes;
1222         struct lzx_optimum_node * const end_node = &c->optimum_nodes[block_size];
1223         struct lz_match *cache_ptr = c->match_cache;
1224         const u8 *in_next = block_begin;
1225         const u8 * const block_end = block_begin + block_size;
1226
1227         /* Instead of storing the match offset LRU queues in the
1228          * 'lzx_optimum_node' structures, we save memory (and cache lines) by
1229          * storing them in a smaller array.  This works because the algorithm
1230          * only requires a limited history of the adaptive state.  Once a given
1231          * state is more than LZX_MAX_MATCH_LEN bytes behind the current node,
1232          * it is no longer needed.  */
1233         struct lzx_lru_queue queues[512];
1234
1235         BUILD_BUG_ON(ARRAY_LEN(queues) < LZX_MAX_MATCH_LEN + 1);
1236 #define QUEUE(in) (queues[(uintptr_t)(in) % ARRAY_LEN(queues)])
1237
1238         /* Initially, the cost to reach each node is "infinity".  */
1239         memset(c->optimum_nodes, 0xFF,
1240                (block_size + 1) * sizeof(c->optimum_nodes[0]));
1241
1242         QUEUE(block_begin) = initial_queue;
1243
1244         /* The following loop runs 'block_size' iterations, one per node.  */
1245         do {
1246                 unsigned num_matches;
1247                 unsigned literal;
1248                 u32 cost;
1249
1250                 /*
1251                  * A selection of matches for the block was already saved in
1252                  * memory so that we don't have to run the uncompressed data
1253                  * through the matchfinder on every optimization pass.  However,
1254                  * we still search for repeat offset matches during each
1255                  * optimization pass because we cannot predict the state of the
1256                  * recent offsets queue.  But as a heuristic, we don't bother
1257                  * searching for repeat offset matches if the general-purpose
1258                  * matchfinder failed to find any matches.
1259                  *
1260                  * Note that a match of length n at some offset implies there is
1261                  * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at
1262                  * that same offset.  In other words, we don't necessarily need
1263                  * to use the full length of a match.  The key heuristic that
1264                  * saves a significicant amount of time is that for each
1265                  * distinct length, we only consider the smallest offset for
1266                  * which that length is available.  This heuristic also applies
1267                  * to repeat offsets, which we order specially: R0 < R1 < R2 <
1268                  * any explicit offset.  Of course, this heuristic may be
1269                  * produce suboptimal results because offset slots in LZX are
1270                  * subject to entropy encoding, but in practice this is a useful
1271                  * heuristic.
1272                  */
1273
1274                 num_matches = cache_ptr->length;
1275                 cache_ptr++;
1276
1277                 if (num_matches) {
1278                         struct lz_match *end_matches = cache_ptr + num_matches;
1279                         unsigned next_len = LZX_MIN_MATCH_LEN;
1280                         unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN);
1281                         const u8 *matchptr;
1282
1283                         /* Consider R0 match  */
1284                         matchptr = in_next - lzx_lru_queue_R0(QUEUE(in_next));
1285                         if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1286                                 goto R0_done;
1287                         BUILD_BUG_ON(LZX_MIN_MATCH_LEN != 2);
1288                         do {
1289                                 u32 cost = cur_node->cost +
1290                                            c->costs.match_cost[0][
1291                                                         next_len - LZX_MIN_MATCH_LEN];
1292                                 if (cost <= (cur_node + next_len)->cost) {
1293                                         (cur_node + next_len)->cost = cost;
1294                                         (cur_node + next_len)->item =
1295                                                 (0 << OPTIMUM_OFFSET_SHIFT) | next_len;
1296                                 }
1297                                 if (unlikely(++next_len > max_len)) {
1298                                         cache_ptr = end_matches;
1299                                         goto done_matches;
1300                                 }
1301                         } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1302
1303                 R0_done:
1304
1305                         /* Consider R1 match  */
1306                         matchptr = in_next - lzx_lru_queue_R1(QUEUE(in_next));
1307                         if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1308                                 goto R1_done;
1309                         if (matchptr[next_len - 1] != in_next[next_len - 1])
1310                                 goto R1_done;
1311                         for (unsigned len = 2; len < next_len - 1; len++)
1312                                 if (matchptr[len] != in_next[len])
1313                                         goto R1_done;
1314                         do {
1315                                 u32 cost = cur_node->cost +
1316                                            c->costs.match_cost[1][
1317                                                         next_len - LZX_MIN_MATCH_LEN];
1318                                 if (cost <= (cur_node + next_len)->cost) {
1319                                         (cur_node + next_len)->cost = cost;
1320                                         (cur_node + next_len)->item =
1321                                                 (1 << OPTIMUM_OFFSET_SHIFT) | next_len;
1322                                 }
1323                                 if (unlikely(++next_len > max_len)) {
1324                                         cache_ptr = end_matches;
1325                                         goto done_matches;
1326                                 }
1327                         } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1328
1329                 R1_done:
1330
1331                         /* Consider R2 match  */
1332                         matchptr = in_next - lzx_lru_queue_R2(QUEUE(in_next));
1333                         if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1334                                 goto R2_done;
1335                         if (matchptr[next_len - 1] != in_next[next_len - 1])
1336                                 goto R2_done;
1337                         for (unsigned len = 2; len < next_len - 1; len++)
1338                                 if (matchptr[len] != in_next[len])
1339                                         goto R2_done;
1340                         do {
1341                                 u32 cost = cur_node->cost +
1342                                            c->costs.match_cost[2][
1343                                                         next_len - LZX_MIN_MATCH_LEN];
1344                                 if (cost <= (cur_node + next_len)->cost) {
1345                                         (cur_node + next_len)->cost = cost;
1346                                         (cur_node + next_len)->item =
1347                                                 (2 << OPTIMUM_OFFSET_SHIFT) | next_len;
1348                                 }
1349                                 if (unlikely(++next_len > max_len)) {
1350                                         cache_ptr = end_matches;
1351                                         goto done_matches;
1352                                 }
1353                         } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1354
1355                 R2_done:
1356
1357                         while (next_len > cache_ptr->length)
1358                                 if (++cache_ptr == end_matches)
1359                                         goto done_matches;
1360
1361                         /* Consider explicit offset matches  */
1362                         do {
1363                                 u32 offset = cache_ptr->offset;
1364                                 u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT;
1365                                 unsigned offset_slot = (offset < LZX_NUM_FAST_OFFSETS) ?
1366                                                 lzx_get_offset_slot_fast(c, offset) :
1367                                                 lzx_get_offset_slot(offset);
1368                                 do {
1369                                         u32 cost = cur_node->cost +
1370                                                    c->costs.match_cost[offset_slot][
1371                                                                 next_len - LZX_MIN_MATCH_LEN];
1372                                 #if LZX_CONSIDER_ALIGNED_COSTS
1373                                         if (lzx_extra_offset_bits[offset_slot] >=
1374                                             LZX_NUM_ALIGNED_OFFSET_BITS)
1375                                                 cost += c->costs.aligned[offset_data &
1376                                                                          LZX_ALIGNED_OFFSET_BITMASK];
1377                                 #endif
1378                                         if (cost < (cur_node + next_len)->cost) {
1379                                                 (cur_node + next_len)->cost = cost;
1380                                                 (cur_node + next_len)->item =
1381                                                         (offset_data << OPTIMUM_OFFSET_SHIFT) | next_len;
1382                                         }
1383                                 } while (++next_len <= cache_ptr->length);
1384                         } while (++cache_ptr != end_matches);
1385                 }
1386
1387         done_matches:
1388
1389                 /* Consider coding a literal.
1390
1391                  * To avoid an extra branch, actually checking the preferability
1392                  * of coding the literal is integrated into the queue update
1393                  * code below.  */
1394                 literal = *in_next++;
1395                 cost = cur_node->cost +
1396                        c->costs.main[lzx_main_symbol_for_literal(literal)];
1397
1398                 /* Advance to the next position.  */
1399                 cur_node++;
1400
1401                 /* The lowest-cost path to the current position is now known.
1402                  * Finalize the recent offsets queue that results from taking
1403                  * this lowest-cost path.  */
1404
1405                 if (cost <= cur_node->cost) {
1406                         /* Literal: queue remains unchanged.  */
1407                         cur_node->cost = cost;
1408                         cur_node->item = (literal << OPTIMUM_OFFSET_SHIFT) | 1;
1409                         QUEUE(in_next) = QUEUE(in_next - 1);
1410                 } else {
1411                         /* Match: queue update is needed.  */
1412                         unsigned len = cur_node->item & OPTIMUM_LEN_MASK;
1413                         u32 offset_data = cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1414                         if (offset_data >= LZX_NUM_RECENT_OFFSETS) {
1415                                 /* Explicit offset match: insert offset at front  */
1416                                 QUEUE(in_next) =
1417                                         lzx_lru_queue_push(QUEUE(in_next - len),
1418                                                            offset_data - LZX_OFFSET_ADJUSTMENT);
1419                         } else {
1420                                 /* Repeat offset match: swap offset to front  */
1421                                 QUEUE(in_next) =
1422                                         lzx_lru_queue_swap(QUEUE(in_next - len),
1423                                                            offset_data);
1424                         }
1425                 }
1426         } while (cur_node != end_node);
1427
1428         /* Return the match offset queue at the end of the minimum-cost path. */
1429         return QUEUE(block_end);
1430 }
1431
1432 /* Given the costs for the main and length codewords, compute 'match_costs'.  */
1433 static void
1434 lzx_compute_match_costs(struct lzx_compressor *c)
1435 {
1436         unsigned num_offset_slots = lzx_get_num_offset_slots(c->window_order);
1437         struct lzx_costs *costs = &c->costs;
1438
1439         for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) {
1440
1441                 u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST;
1442                 unsigned main_symbol = lzx_main_symbol_for_match(offset_slot, 0);
1443                 unsigned i;
1444
1445         #if LZX_CONSIDER_ALIGNED_COSTS
1446                 if (lzx_extra_offset_bits[offset_slot] >= LZX_NUM_ALIGNED_OFFSET_BITS)
1447                         extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1448         #endif
1449
1450                 for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++)
1451                         costs->match_cost[offset_slot][i] =
1452                                 costs->main[main_symbol++] + extra_cost;
1453
1454                 extra_cost += costs->main[main_symbol];
1455
1456                 for (; i < LZX_NUM_LENS; i++)
1457                         costs->match_cost[offset_slot][i] =
1458                                 costs->len[i - LZX_NUM_PRIMARY_LENS] + extra_cost;
1459         }
1460 }
1461
1462 /* Set default LZX Huffman symbol costs to bootstrap the iterative optimization
1463  * algorithm.  */
1464 static void
1465 lzx_set_default_costs(struct lzx_compressor *c, const u8 *block, u32 block_size)
1466 {
1467         u32 i;
1468         bool have_byte[256];
1469         unsigned num_used_bytes;
1470
1471         /* The costs below are hard coded to use a scaling factor of 16.  */
1472         BUILD_BUG_ON(LZX_BIT_COST != 16);
1473
1474         /*
1475          * Heuristics:
1476          *
1477          * - Use smaller initial costs for literal symbols when the input buffer
1478          *   contains fewer distinct bytes.
1479          *
1480          * - Assume that match symbols are more costly than literal symbols.
1481          *
1482          * - Assume that length symbols for shorter lengths are less costly than
1483          *   length symbols for longer lengths.
1484          */
1485
1486         for (i = 0; i < 256; i++)
1487                 have_byte[i] = false;
1488
1489         for (i = 0; i < block_size; i++)
1490                 have_byte[block[i]] = true;
1491
1492         num_used_bytes = 0;
1493         for (i = 0; i < 256; i++)
1494                 num_used_bytes += have_byte[i];
1495
1496         for (i = 0; i < 256; i++)
1497                 c->costs.main[i] = 140 - (256 - num_used_bytes) / 4;
1498
1499         for (; i < c->num_main_syms; i++)
1500                 c->costs.main[i] = 170;
1501
1502         for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1503                 c->costs.len[i] = 103 + (i / 4);
1504
1505 #if LZX_CONSIDER_ALIGNED_COSTS
1506         for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1507                 c->costs.aligned[i] = LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1508 #endif
1509
1510         lzx_compute_match_costs(c);
1511 }
1512
1513 /* Update the current cost model to reflect the computed Huffman codes.  */
1514 static void
1515 lzx_update_costs(struct lzx_compressor *c)
1516 {
1517         unsigned i;
1518         const struct lzx_lens *lens = &c->codes[c->codes_index].lens;
1519
1520         for (i = 0; i < c->num_main_syms; i++)
1521                 c->costs.main[i] = (lens->main[i] ? lens->main[i] : 15) * LZX_BIT_COST;
1522
1523         for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1524                 c->costs.len[i] = (lens->len[i] ? lens->len[i] : 15) * LZX_BIT_COST;
1525
1526 #if LZX_CONSIDER_ALIGNED_COSTS
1527         for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1528                 c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : 7) * LZX_BIT_COST;
1529 #endif
1530
1531         lzx_compute_match_costs(c);
1532 }
1533
1534 static struct lzx_lru_queue
1535 lzx_optimize_and_write_block(struct lzx_compressor *c,
1536                              struct lzx_output_bitstream *os,
1537                              const u8 *block_begin, const u32 block_size,
1538                              const struct lzx_lru_queue initial_queue)
1539 {
1540         unsigned num_passes_remaining = c->num_optim_passes;
1541         struct lzx_item *next_chosen_item;
1542         struct lzx_lru_queue new_queue;
1543
1544         /* The first optimization pass uses a default cost model.  Each
1545          * additional optimization pass uses a cost model derived from the
1546          * Huffman code computed in the previous pass.  */
1547
1548         lzx_set_default_costs(c, block_begin, block_size);
1549         lzx_reset_symbol_frequencies(c);
1550         do {
1551                 new_queue = lzx_find_min_cost_path(c, block_begin, block_size,
1552                                                    initial_queue);
1553                 if (num_passes_remaining > 1) {
1554                         lzx_tally_item_list(c, c->optimum_nodes + block_size);
1555                         lzx_make_huffman_codes(c);
1556                         lzx_update_costs(c);
1557                         lzx_reset_symbol_frequencies(c);
1558                 }
1559         } while (--num_passes_remaining);
1560
1561         next_chosen_item = c->chosen_items;
1562         lzx_record_item_list(c, c->optimum_nodes + block_size, &next_chosen_item);
1563         lzx_finish_block(c, os, block_size, next_chosen_item - c->chosen_items);
1564         return new_queue;
1565 }
1566
1567 /*
1568  * This is the "near-optimal" LZX compressor.
1569  *
1570  * For each block, it performs a relatively thorough graph search to find an
1571  * inexpensive (in terms of compressed size) way to output that block.
1572  *
1573  * Note: there are actually many things this algorithm leaves on the table in
1574  * terms of compression ratio.  So although it may be "near-optimal", it is
1575  * certainly not "optimal".  The goal is not to produce the optimal compression
1576  * ratio, which for LZX is probably impossible within any practical amount of
1577  * time, but rather to produce a compression ratio significantly better than a
1578  * simpler "greedy" or "lazy" parse while still being relatively fast.
1579  */
1580 static void
1581 lzx_compress_near_optimal(struct lzx_compressor *c,
1582                           struct lzx_output_bitstream *os)
1583 {
1584         const u8 * const in_begin = c->in_buffer;
1585         const u8 *       in_next = in_begin;
1586         const u8 * const in_end  = in_begin + c->in_nbytes;
1587         unsigned max_len = LZX_MAX_MATCH_LEN;
1588         unsigned nice_len = min(c->nice_match_length, max_len);
1589         u32 next_hash;
1590         struct lzx_lru_queue queue;
1591
1592         bt_matchfinder_init(&c->bt_mf);
1593         matchfinder_init(c->hash2_tab, LZX_HASH2_LENGTH);
1594         next_hash = bt_matchfinder_hash_3_bytes(in_next);
1595         lzx_lru_queue_init(&queue);
1596
1597         do {
1598                 /* Starting a new block  */
1599                 const u8 * const in_block_begin = in_next;
1600                 const u8 * const in_block_end =
1601                         in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1602
1603                 /* Run the block through the matchfinder and cache the matches. */
1604                 struct lz_match *cache_ptr = c->match_cache;
1605                 do {
1606                         struct lz_match *lz_matchptr;
1607                         u32 hash2;
1608                         pos_t cur_match;
1609                         unsigned best_len;
1610
1611                         /* If approaching the end of the input buffer, adjust
1612                          * 'max_len' and 'nice_len' accordingly.  */
1613                         if (unlikely(max_len > in_end - in_next)) {
1614                                 max_len = in_end - in_next;
1615                                 nice_len = min(max_len, nice_len);
1616
1617                                 /* This extra check is needed to ensure that
1618                                  * reading the next 3 bytes when looking for a
1619                                  * length 2 match is valid.  In addition, we
1620                                  * cannot allow ourselves to find a length 2
1621                                  * match of the very last two bytes with the
1622                                  * very first two bytes, since such a match has
1623                                  * an offset too large to be represented.  */
1624                                 if (unlikely(max_len < 3)) {
1625                                         in_next++;
1626                                         cache_ptr->length = 0;
1627                                         cache_ptr++;
1628                                         continue;
1629                                 }
1630                         }
1631
1632                         lz_matchptr = cache_ptr + 1;
1633
1634                         /* Check for a length 2 match.  */
1635                         hash2 = lz_hash_2_bytes(in_next, LZX_HASH2_ORDER);
1636                         cur_match = c->hash2_tab[hash2];
1637                         c->hash2_tab[hash2] = in_next - in_begin;
1638                         if (matchfinder_node_valid(cur_match) &&
1639                             (LZX_HASH2_ORDER == 16 ||
1640                              load_u16_unaligned(&in_begin[cur_match]) ==
1641                              load_u16_unaligned(in_next)) &&
1642                             in_begin[cur_match + 2] != in_next[2])
1643                         {
1644                                 lz_matchptr->length = 2;
1645                                 lz_matchptr->offset = in_next - &in_begin[cur_match];
1646                                 lz_matchptr++;
1647                         }
1648
1649                         /* Check for matches of length >= 3.  */
1650                         lz_matchptr = bt_matchfinder_get_matches(&c->bt_mf,
1651                                                                  in_begin,
1652                                                                  in_next,
1653                                                                  3,
1654                                                                  max_len,
1655                                                                  nice_len,
1656                                                                  c->max_search_depth,
1657                                                                  &next_hash,
1658                                                                  &best_len,
1659                                                                  lz_matchptr);
1660                         in_next++;
1661                         cache_ptr->length = lz_matchptr - (cache_ptr + 1);
1662                         cache_ptr = lz_matchptr;
1663
1664                         /*
1665                          * If there was a very long match found, then don't
1666                          * cache any matches for the bytes covered by that
1667                          * match.  This avoids degenerate behavior when
1668                          * compressing highly redundant data, where the number
1669                          * of matches can be very large.
1670                          *
1671                          * This heuristic doesn't actually hurt the compression
1672                          * ratio very much.  If there's a long match, then the
1673                          * data must be highly compressible, so it doesn't
1674                          * matter as much what we do.
1675                          */
1676                         if (best_len >= nice_len) {
1677                                 --best_len;
1678                                 do {
1679                                         if (unlikely(max_len > in_end - in_next)) {
1680                                                 max_len = in_end - in_next;
1681                                                 nice_len = min(max_len, nice_len);
1682                                                 if (unlikely(max_len < 3)) {
1683                                                         in_next++;
1684                                                         cache_ptr->length = 0;
1685                                                         cache_ptr++;
1686                                                         continue;
1687                                                 }
1688                                         }
1689                                         c->hash2_tab[lz_hash_2_bytes(in_next, LZX_HASH2_ORDER)] =
1690                                                 in_next - in_begin;
1691                                         bt_matchfinder_skip_position(&c->bt_mf,
1692                                                                      in_begin,
1693                                                                      in_next,
1694                                                                      in_end,
1695                                                                      nice_len,
1696                                                                      c->max_search_depth,
1697                                                                      &next_hash);
1698                                         in_next++;
1699                                         cache_ptr->length = 0;
1700                                         cache_ptr++;
1701                                 } while (--best_len);
1702                         }
1703                 } while (in_next < in_block_end &&
1704                          likely(cache_ptr < &c->match_cache[LZX_CACHE_LENGTH]));
1705
1706                 /* We've finished running the block through the matchfinder.
1707                  * Now choose a match/literal sequence and write the block.  */
1708
1709                 queue = lzx_optimize_and_write_block(c, os, in_block_begin,
1710                                                      in_next - in_block_begin,
1711                                                      queue);
1712         } while (in_next != in_end);
1713 }
1714
1715 /*
1716  * Given a pointer to the current byte sequence and the current list of recent
1717  * match offsets, find the longest repeat offset match.
1718  *
1719  * If no match of at least 2 bytes is found, then return 0.
1720  *
1721  * If a match of at least 2 bytes is found, then return its length and set
1722  * *rep_max_idx_ret to the index of its offset in @queue.
1723 */
1724 static unsigned
1725 lzx_find_longest_repeat_offset_match(const u8 * const in_next,
1726                                      const u32 bytes_remaining,
1727                                      struct lzx_lru_queue queue,
1728                                      unsigned *rep_max_idx_ret)
1729 {
1730         BUILD_BUG_ON(LZX_NUM_RECENT_OFFSETS != 3);
1731         LZX_ASSERT(bytes_remaining >= 2);
1732
1733         const unsigned max_len = min(bytes_remaining, LZX_MAX_MATCH_LEN);
1734         const u16 next_2_bytes = load_u16_unaligned(in_next);
1735         const u8 *matchptr;
1736         unsigned rep_max_len;
1737         unsigned rep_max_idx;
1738         unsigned rep_len;
1739
1740         matchptr = in_next - lzx_lru_queue_pop(&queue);
1741         if (load_u16_unaligned(matchptr) == next_2_bytes)
1742                 rep_max_len = lz_extend(in_next, matchptr, 2, max_len);
1743         else
1744                 rep_max_len = 0;
1745         rep_max_idx = 0;
1746
1747         matchptr = in_next - lzx_lru_queue_pop(&queue);
1748         if (load_u16_unaligned(matchptr) == next_2_bytes) {
1749                 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1750                 if (rep_len > rep_max_len) {
1751                         rep_max_len = rep_len;
1752                         rep_max_idx = 1;
1753                 }
1754         }
1755
1756         matchptr = in_next - lzx_lru_queue_pop(&queue);
1757         if (load_u16_unaligned(matchptr) == next_2_bytes) {
1758                 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1759                 if (rep_len > rep_max_len) {
1760                         rep_max_len = rep_len;
1761                         rep_max_idx = 2;
1762                 }
1763         }
1764
1765         *rep_max_idx_ret = rep_max_idx;
1766         return rep_max_len;
1767 }
1768
1769 /* Fast heuristic scoring for lazy parsing: how "good" is this match?  */
1770 static inline unsigned
1771 lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset)
1772 {
1773         unsigned score = len;
1774
1775         if (adjusted_offset < 4096)
1776                 score++;
1777
1778         if (adjusted_offset < 256)
1779                 score++;
1780
1781         return score;
1782 }
1783
1784 static inline unsigned
1785 lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx)
1786 {
1787         return rep_len + 3;
1788 }
1789
1790 /* This is the "lazy" LZX compressor.  */
1791 static void
1792 lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os)
1793 {
1794         const u8 * const in_begin = c->in_buffer;
1795         const u8 *       in_next = in_begin;
1796         const u8 * const in_end  = in_begin + c->in_nbytes;
1797         unsigned max_len = LZX_MAX_MATCH_LEN;
1798         unsigned nice_len = min(c->nice_match_length, max_len);
1799         struct lzx_lru_queue queue;
1800
1801         hc_matchfinder_init(&c->hc_mf);
1802         lzx_lru_queue_init(&queue);
1803
1804         do {
1805                 /* Starting a new block  */
1806
1807                 const u8 * const in_block_begin = in_next;
1808                 const u8 * const in_block_end =
1809                         in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1810                 struct lzx_item *next_chosen_item = c->chosen_items;
1811                 unsigned cur_len;
1812                 u32 cur_offset;
1813                 u32 cur_offset_data;
1814                 unsigned cur_score;
1815                 unsigned next_len;
1816                 u32 next_offset;
1817                 u32 next_offset_data;
1818                 unsigned next_score;
1819                 unsigned rep_max_len;
1820                 unsigned rep_max_idx;
1821                 unsigned rep_score;
1822                 unsigned skip_len;
1823
1824                 lzx_reset_symbol_frequencies(c);
1825
1826                 do {
1827                         if (unlikely(max_len > in_end - in_next)) {
1828                                 max_len = in_end - in_next;
1829                                 nice_len = min(max_len, nice_len);
1830                         }
1831
1832                         /* Find the longest match at the current position.  */
1833
1834                         cur_len = hc_matchfinder_longest_match(&c->hc_mf,
1835                                                                in_begin,
1836                                                                in_next,
1837                                                                2,
1838                                                                max_len,
1839                                                                nice_len,
1840                                                                c->max_search_depth,
1841                                                                &cur_offset);
1842                         if (cur_len < 3 ||
1843                             (cur_len == 3 &&
1844                              cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT &&
1845                              cur_offset != lzx_lru_queue_R0(queue) &&
1846                              cur_offset != lzx_lru_queue_R1(queue) &&
1847                              cur_offset != lzx_lru_queue_R2(queue)))
1848                         {
1849                                 /* There was no match found, or the only match found
1850                                  * was a distant length 3 match.  Output a literal.  */
1851                                 lzx_declare_literal(c, *in_next++,
1852                                                     &next_chosen_item);
1853                                 continue;
1854                         }
1855
1856                         if (cur_offset == lzx_lru_queue_R0(queue)) {
1857                                 in_next++;
1858                                 cur_offset_data = 0;
1859                                 skip_len = cur_len - 1;
1860                                 goto choose_cur_match;
1861                         }
1862
1863                         cur_offset_data = cur_offset + LZX_OFFSET_ADJUSTMENT;
1864                         cur_score = lzx_explicit_offset_match_score(cur_len, cur_offset_data);
1865
1866                         /* Consider a repeat offset match  */
1867                         rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
1868                                                                            in_end - in_next,
1869                                                                            queue,
1870                                                                            &rep_max_idx);
1871                         in_next++;
1872
1873                         if (rep_max_len >= 3 &&
1874                             (rep_score = lzx_repeat_offset_match_score(rep_max_len,
1875                                                                        rep_max_idx)) >= cur_score)
1876                         {
1877                                 cur_len = rep_max_len;
1878                                 cur_offset_data = rep_max_idx;
1879                                 skip_len = rep_max_len - 1;
1880                                 goto choose_cur_match;
1881                         }
1882
1883                 have_cur_match:
1884
1885                         /* We have a match at the current position.  */
1886
1887                         /* If we have a very long match, choose it immediately.  */
1888                         if (cur_len >= nice_len) {
1889                                 skip_len = cur_len - 1;
1890                                 goto choose_cur_match;
1891                         }
1892
1893                         /* See if there's a better match at the next position.  */
1894
1895                         if (unlikely(max_len > in_end - in_next)) {
1896                                 max_len = in_end - in_next;
1897                                 nice_len = min(max_len, nice_len);
1898                         }
1899
1900                         next_len = hc_matchfinder_longest_match(&c->hc_mf,
1901                                                                 in_begin,
1902                                                                 in_next,
1903                                                                 cur_len - 2,
1904                                                                 max_len,
1905                                                                 nice_len,
1906                                                                 c->max_search_depth / 2,
1907                                                                 &next_offset);
1908
1909                         if (next_len <= cur_len - 2) {
1910                                 in_next++;
1911                                 skip_len = cur_len - 2;
1912                                 goto choose_cur_match;
1913                         }
1914
1915                         next_offset_data = next_offset + LZX_OFFSET_ADJUSTMENT;
1916                         next_score = lzx_explicit_offset_match_score(next_len, next_offset_data);
1917
1918                         rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
1919                                                                            in_end - in_next,
1920                                                                            queue,
1921                                                                            &rep_max_idx);
1922                         in_next++;
1923
1924                         if (rep_max_len >= 3 &&
1925                             (rep_score = lzx_repeat_offset_match_score(rep_max_len,
1926                                                                        rep_max_idx)) >= next_score)
1927                         {
1928
1929                                 if (rep_score > cur_score) {
1930                                         /* The next match is better, and it's a
1931                                          * repeat offset match.  */
1932                                         lzx_declare_literal(c, *(in_next - 2),
1933                                                             &next_chosen_item);
1934                                         cur_len = rep_max_len;
1935                                         cur_offset_data = rep_max_idx;
1936                                         skip_len = cur_len - 1;
1937                                         goto choose_cur_match;
1938                                 }
1939                         } else {
1940                                 if (next_score > cur_score) {
1941                                         /* The next match is better, and it's an
1942                                          * explicit offset match.  */
1943                                         lzx_declare_literal(c, *(in_next - 2),
1944                                                             &next_chosen_item);
1945                                         cur_len = next_len;
1946                                         cur_offset_data = next_offset_data;
1947                                         cur_score = next_score;
1948                                         goto have_cur_match;
1949                                 }
1950                         }
1951
1952                         /* The original match was better.  */
1953                         skip_len = cur_len - 2;
1954
1955                 choose_cur_match:
1956                         if (cur_offset_data < LZX_NUM_RECENT_OFFSETS) {
1957                                 lzx_declare_repeat_offset_match(c, cur_len,
1958                                                                 cur_offset_data,
1959                                                                 &next_chosen_item);
1960                                 queue = lzx_lru_queue_swap(queue, cur_offset_data);
1961                         } else {
1962                                 lzx_declare_explicit_offset_match(c, cur_len,
1963                                                                   cur_offset_data - LZX_OFFSET_ADJUSTMENT,
1964                                                                   &next_chosen_item);
1965                                 queue = lzx_lru_queue_push(queue, cur_offset_data - LZX_OFFSET_ADJUSTMENT);
1966                         }
1967
1968                         hc_matchfinder_skip_positions(&c->hc_mf,
1969                                                       in_begin,
1970                                                       in_next,
1971                                                       in_end,
1972                                                       skip_len);
1973                         in_next += skip_len;
1974                 } while (in_next < in_block_end);
1975
1976                 lzx_finish_block(c, os, in_next - in_block_begin,
1977                                  next_chosen_item - c->chosen_items);
1978         } while (in_next != in_end);
1979 }
1980
1981 static void
1982 lzx_init_offset_slot_fast(struct lzx_compressor *c)
1983 {
1984         u8 slot = 0;
1985
1986         for (u32 offset = 0; offset < LZX_NUM_FAST_OFFSETS; offset++) {
1987
1988                 while (offset + LZX_OFFSET_ADJUSTMENT >= lzx_offset_slot_base[slot + 1])
1989                         slot++;
1990
1991                 c->offset_slot_fast[offset] = slot;
1992         }
1993 }
1994
1995 static size_t
1996 lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level)
1997 {
1998         if (compression_level <= LZX_MAX_FAST_LEVEL) {
1999                 return offsetof(struct lzx_compressor, hc_mf) +
2000                         hc_matchfinder_size(max_bufsize);
2001         } else {
2002                 return offsetof(struct lzx_compressor, bt_mf) +
2003                         bt_matchfinder_size(max_bufsize);
2004         }
2005 }
2006
2007 static u64
2008 lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level)
2009 {
2010         u64 size = 0;
2011
2012         if (max_bufsize > LZX_MAX_WINDOW_SIZE)
2013                 return 0;
2014
2015         size += lzx_get_compressor_size(max_bufsize, compression_level);
2016         size += max_bufsize; /* in_buffer */
2017         return size;
2018 }
2019
2020 static int
2021 lzx_create_compressor(size_t max_bufsize, unsigned compression_level,
2022                       void **c_ret)
2023 {
2024         unsigned window_order;
2025         struct lzx_compressor *c;
2026
2027         window_order = lzx_get_window_order(max_bufsize);
2028         if (window_order == 0)
2029                 return WIMLIB_ERR_INVALID_PARAM;
2030
2031         c = ALIGNED_MALLOC(lzx_get_compressor_size(max_bufsize,
2032                                                    compression_level),
2033                            MATCHFINDER_ALIGNMENT);
2034         if (!c)
2035                 goto oom0;
2036
2037         c->num_main_syms = lzx_get_num_main_syms(window_order);
2038         c->window_order = window_order;
2039
2040         c->in_buffer = MALLOC(max_bufsize);
2041         if (!c->in_buffer)
2042                 goto oom1;
2043
2044         if (compression_level <= LZX_MAX_FAST_LEVEL) {
2045
2046                 /* Fast compression: Use lazy parsing.  */
2047
2048                 c->impl = lzx_compress_lazy;
2049                 c->max_search_depth = (36 * compression_level) / 20;
2050                 c->nice_match_length = (72 * compression_level) / 20;
2051
2052                 /* lzx_compress_lazy() needs max_search_depth >= 2 because it
2053                  * halves the max_search_depth when attempting a lazy match, and
2054                  * max_search_depth cannot be 0.  */
2055                 if (c->max_search_depth < 2)
2056                         c->max_search_depth = 2;
2057         } else {
2058
2059                 /* Normal / high compression: Use near-optimal parsing.  */
2060
2061                 c->impl = lzx_compress_near_optimal;
2062
2063                 /* Scale nice_match_length and max_search_depth with the
2064                  * compression level.  */
2065                 c->max_search_depth = (24 * compression_level) / 50;
2066                 c->nice_match_length = (32 * compression_level) / 50;
2067
2068                 /* Set a number of optimization passes appropriate for the
2069                  * compression level.  */
2070
2071                 c->num_optim_passes = 1;
2072
2073                 if (compression_level >= 45)
2074                         c->num_optim_passes++;
2075
2076                 /* Use more optimization passes for higher compression levels.
2077                  * But the more passes there are, the less they help --- so
2078                  * don't add them linearly.  */
2079                 if (compression_level >= 70) {
2080                         c->num_optim_passes++;
2081                         if (compression_level >= 100)
2082                                 c->num_optim_passes++;
2083                         if (compression_level >= 150)
2084                                 c->num_optim_passes++;
2085                         if (compression_level >= 200)
2086                                 c->num_optim_passes++;
2087                         if (compression_level >= 300)
2088                                 c->num_optim_passes++;
2089                 }
2090         }
2091
2092         /* max_search_depth == 0 is invalid.  */
2093         if (c->max_search_depth < 1)
2094                 c->max_search_depth = 1;
2095
2096         if (c->nice_match_length > LZX_MAX_MATCH_LEN)
2097                 c->nice_match_length = LZX_MAX_MATCH_LEN;
2098
2099         lzx_init_offset_slot_fast(c);
2100         *c_ret = c;
2101         return 0;
2102
2103 oom1:
2104         ALIGNED_FREE(c);
2105 oom0:
2106         return WIMLIB_ERR_NOMEM;
2107 }
2108
2109 static size_t
2110 lzx_compress(const void *in, size_t in_nbytes,
2111              void *out, size_t out_nbytes_avail, void *_c)
2112 {
2113         struct lzx_compressor *c = _c;
2114         struct lzx_output_bitstream os;
2115
2116         /* Don't bother trying to compress very small inputs.  */
2117         if (in_nbytes < 100)
2118                 return 0;
2119
2120         /* Copy the input data into the internal buffer and preprocess it.  */
2121         memcpy(c->in_buffer, in, in_nbytes);
2122         c->in_nbytes = in_nbytes;
2123         lzx_do_e8_preprocessing(c->in_buffer, in_nbytes);
2124
2125         /* Initially, the previous Huffman codeword lengths are all zeroes.  */
2126         c->codes_index = 0;
2127         memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens));
2128
2129         /* Initialize the output bitstream.  */
2130         lzx_init_output(&os, out, out_nbytes_avail);
2131
2132         /* Call the compression level-specific compress() function.  */
2133         (*c->impl)(c, &os);
2134
2135         /* Flush the output bitstream and return the compressed size or 0.  */
2136         return lzx_flush_output(&os);
2137 }
2138
2139 static void
2140 lzx_free_compressor(void *_c)
2141 {
2142         struct lzx_compressor *c = _c;
2143
2144         FREE(c->in_buffer);
2145         ALIGNED_FREE(c);
2146 }
2147
2148 const struct compressor_ops lzx_compressor_ops = {
2149         .get_needed_memory  = lzx_get_needed_memory,
2150         .create_compressor  = lzx_create_compressor,
2151         .compress           = lzx_compress,
2152         .free_compressor    = lzx_free_compressor,
2153 };