1 ///////////////////////////////////////////////////////////////////////////////
2 //
3 /// \file lz_encoder.c
4 /// \brief LZ in window
5 ///
6 // Authors: Igor Pavlov
7 // Lasse Collin
8 //
9 // This file has been put into the public domain.
10 // You can do whatever you want with this file.
11 //
12 ///////////////////////////////////////////////////////////////////////////////
13
14 #include "lz_encoder.h"
15 #include "lz_encoder_hash.h"
16
17 // See lz_encoder_hash.h. This is a bit hackish but avoids making
18 // endianness a conditional in makefiles.
19 #if defined(WORDS_BIGENDIAN) && !defined(HAVE_SMALL)
20 # include "lz_encoder_hash_table.h"
21 #endif
22
23 #include "memcmplen.h"
24
25
26 typedef struct {
27 /// LZ-based encoder e.g. LZMA
28 lzma_lz_encoder lz;
29
30 /// History buffer and match finder
31 lzma_mf mf;
32
33 /// Next coder in the chain
34 lzma_next_coder next;
35 } lzma_coder;
36
37
38 /// \brief Moves the data in the input window to free space for new data
39 ///
40 /// mf->buffer is a sliding input window, which keeps mf->keep_size_before
41 /// bytes of input history available all the time. Now and then we need to
42 /// "slide" the buffer to make space for the new data to the end of the
43 /// buffer. At the same time, data older than keep_size_before is dropped.
44 ///
45 static void
move_window(lzma_mf * mf)46 move_window(lzma_mf *mf)
47 {
48 // Align the move to a multiple of 16 bytes. Some LZ-based encoders
49 // like LZMA use the lowest bits of mf->read_pos to know the
50 // alignment of the uncompressed data. We also get better speed
51 // for memmove() with aligned buffers.
52 assert(mf->read_pos > mf->keep_size_before);
53 const uint32_t move_offset
54 = (mf->read_pos - mf->keep_size_before) & ~UINT32_C(15);
55
56 assert(mf->write_pos > move_offset);
57 const size_t move_size = mf->write_pos - move_offset;
58
59 assert(move_offset + move_size <= mf->size);
60
61 memmove(mf->buffer, mf->buffer + move_offset, move_size);
62
63 mf->offset += move_offset;
64 mf->read_pos -= move_offset;
65 mf->read_limit -= move_offset;
66 mf->write_pos -= move_offset;
67
68 return;
69 }
70
71
72 /// \brief Tries to fill the input window (mf->buffer)
73 ///
74 /// If we are the last encoder in the chain, our input data is in in[].
75 /// Otherwise we call the next filter in the chain to process in[] and
76 /// write its output to mf->buffer.
77 ///
78 /// This function must not be called once it has returned LZMA_STREAM_END.
79 ///
80 static lzma_ret
fill_window(lzma_coder * coder,const lzma_allocator * allocator,const uint8_t * in,size_t * in_pos,size_t in_size,lzma_action action)81 fill_window(lzma_coder *coder, const lzma_allocator *allocator,
82 const uint8_t *in, size_t *in_pos, size_t in_size,
83 lzma_action action)
84 {
85 assert(coder->mf.read_pos <= coder->mf.write_pos);
86
87 // Move the sliding window if needed.
88 if (coder->mf.read_pos >= coder->mf.size - coder->mf.keep_size_after)
89 move_window(&coder->mf);
90
91 // Maybe this is ugly, but lzma_mf uses uint32_t for most things
92 // (which I find cleanest), but we need size_t here when filling
93 // the history window.
94 size_t write_pos = coder->mf.write_pos;
95 lzma_ret ret;
96 if (coder->next.code == NULL) {
97 // Not using a filter, simply memcpy() as much as possible.
98 lzma_bufcpy(in, in_pos, in_size, coder->mf.buffer,
99 &write_pos, coder->mf.size);
100
101 ret = action != LZMA_RUN && *in_pos == in_size
102 ? LZMA_STREAM_END : LZMA_OK;
103
104 } else {
105 ret = coder->next.code(coder->next.coder, allocator,
106 in, in_pos, in_size,
107 coder->mf.buffer, &write_pos,
108 coder->mf.size, action);
109 }
110
111 coder->mf.write_pos = write_pos;
112
113 // Silence Valgrind. lzma_memcmplen() can read extra bytes
114 // and Valgrind will give warnings if those bytes are uninitialized
115 // because Valgrind cannot see that the values of the uninitialized
116 // bytes are eventually ignored.
117 memzero(coder->mf.buffer + write_pos, LZMA_MEMCMPLEN_EXTRA);
118
119 // If end of stream has been reached or flushing completed, we allow
120 // the encoder to process all the input (that is, read_pos is allowed
121 // to reach write_pos). Otherwise we keep keep_size_after bytes
122 // available as prebuffer.
123 if (ret == LZMA_STREAM_END) {
124 assert(*in_pos == in_size);
125 ret = LZMA_OK;
126 coder->mf.action = action;
127 coder->mf.read_limit = coder->mf.write_pos;
128
129 } else if (coder->mf.write_pos > coder->mf.keep_size_after) {
130 // This needs to be done conditionally, because if we got
131 // only little new input, there may be too little input
132 // to do any encoding yet.
133 coder->mf.read_limit = coder->mf.write_pos
134 - coder->mf.keep_size_after;
135 }
136
137 // Restart the match finder after finished LZMA_SYNC_FLUSH.
138 if (coder->mf.pending > 0
139 && coder->mf.read_pos < coder->mf.read_limit) {
140 // Match finder may update coder->pending and expects it to
141 // start from zero, so use a temporary variable.
142 const uint32_t pending = coder->mf.pending;
143 coder->mf.pending = 0;
144
145 // Rewind read_pos so that the match finder can hash
146 // the pending bytes.
147 assert(coder->mf.read_pos >= pending);
148 coder->mf.read_pos -= pending;
149
150 // Call the skip function directly instead of using
151 // mf_skip(), since we don't want to touch mf->read_ahead.
152 coder->mf.skip(&coder->mf, pending);
153 }
154
155 return ret;
156 }
157
158
159 static lzma_ret
lz_encode(void * coder_ptr,const lzma_allocator * allocator,const uint8_t * restrict in,size_t * restrict in_pos,size_t in_size,uint8_t * restrict out,size_t * restrict out_pos,size_t out_size,lzma_action action)160 lz_encode(void *coder_ptr, const lzma_allocator *allocator,
161 const uint8_t *restrict in, size_t *restrict in_pos,
162 size_t in_size,
163 uint8_t *restrict out, size_t *restrict out_pos,
164 size_t out_size, lzma_action action)
165 {
166 lzma_coder *coder = coder_ptr;
167
168 while (*out_pos < out_size
169 && (*in_pos < in_size || action != LZMA_RUN)) {
170 // Read more data to coder->mf.buffer if needed.
171 if (coder->mf.action == LZMA_RUN && coder->mf.read_pos
172 >= coder->mf.read_limit)
173 return_if_error(fill_window(coder, allocator,
174 in, in_pos, in_size, action));
175
176 // Encode
177 const lzma_ret ret = coder->lz.code(coder->lz.coder,
178 &coder->mf, out, out_pos, out_size);
179 if (ret != LZMA_OK) {
180 // Setting this to LZMA_RUN for cases when we are
181 // flushing. It doesn't matter when finishing or if
182 // an error occurred.
183 coder->mf.action = LZMA_RUN;
184 return ret;
185 }
186 }
187
188 return LZMA_OK;
189 }
190
191
192 static bool
lz_encoder_prepare(lzma_mf * mf,const lzma_allocator * allocator,const lzma_lz_options * lz_options)193 lz_encoder_prepare(lzma_mf *mf, const lzma_allocator *allocator,
194 const lzma_lz_options *lz_options)
195 {
196 // For now, the dictionary size is limited to 1.5 GiB. This may grow
197 // in the future if needed, but it needs a little more work than just
198 // changing this check.
199 if (!IS_ENC_DICT_SIZE_VALID(lz_options->dict_size)
200 || lz_options->nice_len > lz_options->match_len_max)
201 return true;
202
203 mf->keep_size_before = lz_options->before_size + lz_options->dict_size;
204
205 mf->keep_size_after = lz_options->after_size
206 + lz_options->match_len_max;
207
208 // To avoid constant memmove()s, allocate some extra space. Since
209 // memmove()s become more expensive when the size of the buffer
210 // increases, we reserve more space when a large dictionary is
211 // used to make the memmove() calls rarer.
212 //
213 // This works with dictionaries up to about 3 GiB. If bigger
214 // dictionary is wanted, some extra work is needed:
215 // - Several variables in lzma_mf have to be changed from uint32_t
216 // to size_t.
217 // - Memory usage calculation needs something too, e.g. use uint64_t
218 // for mf->size.
219 uint32_t reserve = lz_options->dict_size / 2;
220 if (reserve > (UINT32_C(1) << 30))
221 reserve /= 2;
222
223 reserve += (lz_options->before_size + lz_options->match_len_max
224 + lz_options->after_size) / 2 + (UINT32_C(1) << 19);
225
226 const uint32_t old_size = mf->size;
227 mf->size = mf->keep_size_before + reserve + mf->keep_size_after;
228
229 // Deallocate the old history buffer if it exists but has different
230 // size than what is needed now.
231 if (mf->buffer != NULL && old_size != mf->size) {
232 lzma_free(mf->buffer, allocator);
233 mf->buffer = NULL;
234 }
235
236 // Match finder options
237 mf->match_len_max = lz_options->match_len_max;
238 mf->nice_len = lz_options->nice_len;
239
240 // cyclic_size has to stay smaller than 2 Gi. Note that this doesn't
241 // mean limiting dictionary size to less than 2 GiB. With a match
242 // finder that uses multibyte resolution (hashes start at e.g. every
243 // fourth byte), cyclic_size would stay below 2 Gi even when
244 // dictionary size is greater than 2 GiB.
245 //
246 // It would be possible to allow cyclic_size >= 2 Gi, but then we
247 // would need to be careful to use 64-bit types in various places
248 // (size_t could do since we would need bigger than 32-bit address
249 // space anyway). It would also require either zeroing a multigigabyte
250 // buffer at initialization (waste of time and RAM) or allow
251 // normalization in lz_encoder_mf.c to access uninitialized
252 // memory to keep the code simpler. The current way is simple and
253 // still allows pretty big dictionaries, so I don't expect these
254 // limits to change.
255 mf->cyclic_size = lz_options->dict_size + 1;
256
257 // Validate the match finder ID and setup the function pointers.
258 switch (lz_options->match_finder) {
259 #ifdef HAVE_MF_HC3
260 case LZMA_MF_HC3:
261 mf->find = &lzma_mf_hc3_find;
262 mf->skip = &lzma_mf_hc3_skip;
263 break;
264 #endif
265 #ifdef HAVE_MF_HC4
266 case LZMA_MF_HC4:
267 mf->find = &lzma_mf_hc4_find;
268 mf->skip = &lzma_mf_hc4_skip;
269 break;
270 #endif
271 #ifdef HAVE_MF_BT2
272 case LZMA_MF_BT2:
273 mf->find = &lzma_mf_bt2_find;
274 mf->skip = &lzma_mf_bt2_skip;
275 break;
276 #endif
277 #ifdef HAVE_MF_BT3
278 case LZMA_MF_BT3:
279 mf->find = &lzma_mf_bt3_find;
280 mf->skip = &lzma_mf_bt3_skip;
281 break;
282 #endif
283 #ifdef HAVE_MF_BT4
284 case LZMA_MF_BT4:
285 mf->find = &lzma_mf_bt4_find;
286 mf->skip = &lzma_mf_bt4_skip;
287 break;
288 #endif
289
290 default:
291 return true;
292 }
293
294 // Calculate the sizes of mf->hash and mf->son.
295 //
296 // NOTE: Since 5.3.5beta the LZMA encoder ensures that nice_len
297 // is big enough for the selected match finder. This makes it
298 // easier for applications as nice_len = 2 will always be accepted
299 // even though the effective value can be slightly bigger.
300 const uint32_t hash_bytes
301 = mf_get_hash_bytes(lz_options->match_finder);
302 assert(hash_bytes <= mf->nice_len);
303
304 const bool is_bt = (lz_options->match_finder & 0x10) != 0;
305 uint32_t hs;
306
307 if (hash_bytes == 2) {
308 hs = 0xFFFF;
309 } else {
310 // Round dictionary size up to the next 2^n - 1 so it can
311 // be used as a hash mask.
312 hs = lz_options->dict_size - 1;
313 hs |= hs >> 1;
314 hs |= hs >> 2;
315 hs |= hs >> 4;
316 hs |= hs >> 8;
317 hs >>= 1;
318 hs |= 0xFFFF;
319
320 if (hs > (UINT32_C(1) << 24)) {
321 if (hash_bytes == 3)
322 hs = (UINT32_C(1) << 24) - 1;
323 else
324 hs >>= 1;
325 }
326 }
327
328 mf->hash_mask = hs;
329
330 ++hs;
331 if (hash_bytes > 2)
332 hs += HASH_2_SIZE;
333 if (hash_bytes > 3)
334 hs += HASH_3_SIZE;
335 /*
336 No match finder uses this at the moment.
337 if (mf->hash_bytes > 4)
338 hs += HASH_4_SIZE;
339 */
340
341 const uint32_t old_hash_count = mf->hash_count;
342 const uint32_t old_sons_count = mf->sons_count;
343 mf->hash_count = hs;
344 mf->sons_count = mf->cyclic_size;
345 if (is_bt)
346 mf->sons_count *= 2;
347
348 // Deallocate the old hash array if it exists and has different size
349 // than what is needed now.
350 if (old_hash_count != mf->hash_count
351 || old_sons_count != mf->sons_count) {
352 lzma_free(mf->hash, allocator);
353 mf->hash = NULL;
354
355 lzma_free(mf->son, allocator);
356 mf->son = NULL;
357 }
358
359 // Maximum number of match finder cycles
360 mf->depth = lz_options->depth;
361 if (mf->depth == 0) {
362 if (is_bt)
363 mf->depth = 16 + mf->nice_len / 2;
364 else
365 mf->depth = 4 + mf->nice_len / 4;
366 }
367
368 return false;
369 }
370
371
372 static bool
lz_encoder_init(lzma_mf * mf,const lzma_allocator * allocator,const lzma_lz_options * lz_options)373 lz_encoder_init(lzma_mf *mf, const lzma_allocator *allocator,
374 const lzma_lz_options *lz_options)
375 {
376 // Allocate the history buffer.
377 if (mf->buffer == NULL) {
378 // lzma_memcmplen() is used for the dictionary buffer
379 // so we need to allocate a few extra bytes to prevent
380 // it from reading past the end of the buffer.
381 mf->buffer = lzma_alloc(mf->size + LZMA_MEMCMPLEN_EXTRA,
382 allocator);
383 if (mf->buffer == NULL)
384 return true;
385
386 // Keep Valgrind happy with lzma_memcmplen() and initialize
387 // the extra bytes whose value may get read but which will
388 // effectively get ignored.
389 memzero(mf->buffer + mf->size, LZMA_MEMCMPLEN_EXTRA);
390 }
391
392 // Use cyclic_size as initial mf->offset. This allows
393 // avoiding a few branches in the match finders. The downside is
394 // that match finder needs to be normalized more often, which may
395 // hurt performance with huge dictionaries.
396 mf->offset = mf->cyclic_size;
397 mf->read_pos = 0;
398 mf->read_ahead = 0;
399 mf->read_limit = 0;
400 mf->write_pos = 0;
401 mf->pending = 0;
402
403 #if UINT32_MAX >= SIZE_MAX / 4
404 // Check for integer overflow. (Huge dictionaries are not
405 // possible on 32-bit CPU.)
406 if (mf->hash_count > SIZE_MAX / sizeof(uint32_t)
407 || mf->sons_count > SIZE_MAX / sizeof(uint32_t))
408 return true;
409 #endif
410
411 // Allocate and initialize the hash table. Since EMPTY_HASH_VALUE
412 // is zero, we can use lzma_alloc_zero() or memzero() for mf->hash.
413 //
414 // We don't need to initialize mf->son, but not doing that may
415 // make Valgrind complain in normalization (see normalize() in
416 // lz_encoder_mf.c). Skipping the initialization is *very* good
417 // when big dictionary is used but only small amount of data gets
418 // actually compressed: most of the mf->son won't get actually
419 // allocated by the kernel, so we avoid wasting RAM and improve
420 // initialization speed a lot.
421 if (mf->hash == NULL) {
422 mf->hash = lzma_alloc_zero(mf->hash_count * sizeof(uint32_t),
423 allocator);
424 mf->son = lzma_alloc(mf->sons_count * sizeof(uint32_t),
425 allocator);
426
427 if (mf->hash == NULL || mf->son == NULL) {
428 lzma_free(mf->hash, allocator);
429 mf->hash = NULL;
430
431 lzma_free(mf->son, allocator);
432 mf->son = NULL;
433
434 return true;
435 }
436 } else {
437 /*
438 for (uint32_t i = 0; i < mf->hash_count; ++i)
439 mf->hash[i] = EMPTY_HASH_VALUE;
440 */
441 memzero(mf->hash, mf->hash_count * sizeof(uint32_t));
442 }
443
444 mf->cyclic_pos = 0;
445
446 // Handle preset dictionary.
447 if (lz_options->preset_dict != NULL
448 && lz_options->preset_dict_size > 0) {
449 // If the preset dictionary is bigger than the actual
450 // dictionary, use only the tail.
451 mf->write_pos = my_min(lz_options->preset_dict_size, mf->size);
452 memcpy(mf->buffer, lz_options->preset_dict
453 + lz_options->preset_dict_size - mf->write_pos,
454 mf->write_pos);
455 mf->action = LZMA_SYNC_FLUSH;
456 mf->skip(mf, mf->write_pos);
457 }
458
459 mf->action = LZMA_RUN;
460
461 return false;
462 }
463
464
465 extern uint64_t
lzma_lz_encoder_memusage(const lzma_lz_options * lz_options)466 lzma_lz_encoder_memusage(const lzma_lz_options *lz_options)
467 {
468 // Old buffers must not exist when calling lz_encoder_prepare().
469 lzma_mf mf = {
470 .buffer = NULL,
471 .hash = NULL,
472 .son = NULL,
473 .hash_count = 0,
474 .sons_count = 0,
475 };
476
477 // Setup the size information into mf.
478 if (lz_encoder_prepare(&mf, NULL, lz_options))
479 return UINT64_MAX;
480
481 // Calculate the memory usage.
482 return ((uint64_t)(mf.hash_count) + mf.sons_count) * sizeof(uint32_t)
483 + mf.size + sizeof(lzma_coder);
484 }
485
486
487 static void
lz_encoder_end(void * coder_ptr,const lzma_allocator * allocator)488 lz_encoder_end(void *coder_ptr, const lzma_allocator *allocator)
489 {
490 lzma_coder *coder = coder_ptr;
491
492 lzma_next_end(&coder->next, allocator);
493
494 lzma_free(coder->mf.son, allocator);
495 lzma_free(coder->mf.hash, allocator);
496 lzma_free(coder->mf.buffer, allocator);
497
498 if (coder->lz.end != NULL)
499 coder->lz.end(coder->lz.coder, allocator);
500 else
501 lzma_free(coder->lz.coder, allocator);
502
503 lzma_free(coder, allocator);
504 return;
505 }
506
507
508 static lzma_ret
lz_encoder_update(void * coder_ptr,const lzma_allocator * allocator,const lzma_filter * filters_null lzma_attribute ((__unused__)),const lzma_filter * reversed_filters)509 lz_encoder_update(void *coder_ptr, const lzma_allocator *allocator,
510 const lzma_filter *filters_null lzma_attribute((__unused__)),
511 const lzma_filter *reversed_filters)
512 {
513 lzma_coder *coder = coder_ptr;
514
515 if (coder->lz.options_update == NULL)
516 return LZMA_PROG_ERROR;
517
518 return_if_error(coder->lz.options_update(
519 coder->lz.coder, reversed_filters));
520
521 return lzma_next_filter_update(
522 &coder->next, allocator, reversed_filters + 1);
523 }
524
525
526 static lzma_ret
lz_encoder_set_out_limit(void * coder_ptr,uint64_t * uncomp_size,uint64_t out_limit)527 lz_encoder_set_out_limit(void *coder_ptr, uint64_t *uncomp_size,
528 uint64_t out_limit)
529 {
530 lzma_coder *coder = coder_ptr;
531
532 // This is supported only if there are no other filters chained.
533 if (coder->next.code == NULL && coder->lz.set_out_limit != NULL)
534 return coder->lz.set_out_limit(
535 coder->lz.coder, uncomp_size, out_limit);
536
537 return LZMA_OPTIONS_ERROR;
538 }
539
540
541 extern lzma_ret
lzma_lz_encoder_init(lzma_next_coder * next,const lzma_allocator * allocator,const lzma_filter_info * filters,lzma_ret (* lz_init)(lzma_lz_encoder * lz,const lzma_allocator * allocator,lzma_vli id,const void * options,lzma_lz_options * lz_options))542 lzma_lz_encoder_init(lzma_next_coder *next, const lzma_allocator *allocator,
543 const lzma_filter_info *filters,
544 lzma_ret (*lz_init)(lzma_lz_encoder *lz,
545 const lzma_allocator *allocator,
546 lzma_vli id, const void *options,
547 lzma_lz_options *lz_options))
548 {
549 #if defined(HAVE_SMALL) && !defined(HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR)
550 // We need that the CRC32 table has been initialized.
551 lzma_crc32_init();
552 #endif
553
554 // Allocate and initialize the base data structure.
555 lzma_coder *coder = next->coder;
556 if (coder == NULL) {
557 coder = lzma_alloc(sizeof(lzma_coder), allocator);
558 if (coder == NULL)
559 return LZMA_MEM_ERROR;
560
561 next->coder = coder;
562 next->code = &lz_encode;
563 next->end = &lz_encoder_end;
564 next->update = &lz_encoder_update;
565 next->set_out_limit = &lz_encoder_set_out_limit;
566
567 coder->lz.coder = NULL;
568 coder->lz.code = NULL;
569 coder->lz.end = NULL;
570
571 // mf.size is initialized to silence Valgrind
572 // when used on optimized binaries (GCC may reorder
573 // code in a way that Valgrind gets unhappy).
574 coder->mf.buffer = NULL;
575 coder->mf.size = 0;
576 coder->mf.hash = NULL;
577 coder->mf.son = NULL;
578 coder->mf.hash_count = 0;
579 coder->mf.sons_count = 0;
580
581 coder->next = LZMA_NEXT_CODER_INIT;
582 }
583
584 // Initialize the LZ-based encoder.
585 lzma_lz_options lz_options;
586 return_if_error(lz_init(&coder->lz, allocator,
587 filters[0].id, filters[0].options, &lz_options));
588
589 // Setup the size information into coder->mf and deallocate
590 // old buffers if they have wrong size.
591 if (lz_encoder_prepare(&coder->mf, allocator, &lz_options))
592 return LZMA_OPTIONS_ERROR;
593
594 // Allocate new buffers if needed, and do the rest of
595 // the initialization.
596 if (lz_encoder_init(&coder->mf, allocator, &lz_options))
597 return LZMA_MEM_ERROR;
598
599 // Initialize the next filter in the chain, if any.
600 return lzma_next_filter_init(&coder->next, allocator, filters + 1);
601 }
602
603
604 extern LZMA_API(lzma_bool)
lzma_mf_is_supported(lzma_match_finder mf)605 lzma_mf_is_supported(lzma_match_finder mf)
606 {
607 switch (mf) {
608 #ifdef HAVE_MF_HC3
609 case LZMA_MF_HC3:
610 return true;
611 #endif
612 #ifdef HAVE_MF_HC4
613 case LZMA_MF_HC4:
614 return true;
615 #endif
616 #ifdef HAVE_MF_BT2
617 case LZMA_MF_BT2:
618 return true;
619 #endif
620 #ifdef HAVE_MF_BT3
621 case LZMA_MF_BT3:
622 return true;
623 #endif
624 #ifdef HAVE_MF_BT4
625 case LZMA_MF_BT4:
626 return true;
627 #endif
628 default:
629 return false;
630 }
631 }
632