1 ///////////////////////////////////////////////////////////////////////////////
3 /// \file lzma_encoder.c
4 /// \brief LZMA encoder
6 // Authors: Igor Pavlov
9 // This file has been put into the public domain.
10 // You can do whatever you want with this file.
12 ///////////////////////////////////////////////////////////////////////////////
14 #include "lzma2_encoder.h"
15 #include "lzma_encoder_private.h"
24 literal_matched(lzma_range_encoder
*rc
, probability
*subcoder
,
25 uint32_t match_byte
, uint32_t symbol
)
27 uint32_t offset
= 0x100;
28 symbol
+= UINT32_C(1) << 8;
32 const uint32_t match_bit
= match_byte
& offset
;
33 const uint32_t subcoder_index
34 = offset
+ match_bit
+ (symbol
>> 8);
35 const uint32_t bit
= (symbol
>> 7) & 1;
36 rc_bit(rc
, &subcoder
[subcoder_index
], bit
);
39 offset
&= ~(match_byte
^ symbol
);
41 } while (symbol
< (UINT32_C(1) << 16));
46 literal(lzma_lzma1_encoder
*coder
, lzma_mf
*mf
, uint32_t position
)
48 // Locate the literal byte to be encoded and the subcoder.
49 const uint8_t cur_byte
= mf
->buffer
[
50 mf
->read_pos
- mf
->read_ahead
];
51 probability
*subcoder
= literal_subcoder(coder
->literal
,
52 coder
->literal_context_bits
, coder
->literal_pos_mask
,
53 position
, mf
->buffer
[mf
->read_pos
- mf
->read_ahead
- 1]);
55 if (is_literal_state(coder
->state
)) {
56 // Previous LZMA-symbol was a literal. Encode a normal
57 // literal without a match byte.
58 rc_bittree(&coder
->rc
, subcoder
, 8, cur_byte
);
60 // Previous LZMA-symbol was a match. Use the last byte of
61 // the match as a "match byte". That is, compare the bits
62 // of the current literal and the match byte.
63 const uint8_t match_byte
= mf
->buffer
[
64 mf
->read_pos
- coder
->reps
[0] - 1
66 literal_matched(&coder
->rc
, subcoder
, match_byte
, cur_byte
);
69 update_literal(coder
->state
);
78 length_update_prices(lzma_length_encoder
*lc
, const uint32_t pos_state
)
80 const uint32_t table_size
= lc
->table_size
;
81 lc
->counters
[pos_state
] = table_size
;
83 const uint32_t a0
= rc_bit_0_price(lc
->choice
);
84 const uint32_t a1
= rc_bit_1_price(lc
->choice
);
85 const uint32_t b0
= a1
+ rc_bit_0_price(lc
->choice2
);
86 const uint32_t b1
= a1
+ rc_bit_1_price(lc
->choice2
);
87 uint32_t *const prices
= lc
->prices
[pos_state
];
90 for (i
= 0; i
< table_size
&& i
< LEN_LOW_SYMBOLS
; ++i
)
91 prices
[i
] = a0
+ rc_bittree_price(lc
->low
[pos_state
],
94 for (; i
< table_size
&& i
< LEN_LOW_SYMBOLS
+ LEN_MID_SYMBOLS
; ++i
)
95 prices
[i
] = b0
+ rc_bittree_price(lc
->mid
[pos_state
],
96 LEN_MID_BITS
, i
- LEN_LOW_SYMBOLS
);
98 for (; i
< table_size
; ++i
)
99 prices
[i
] = b1
+ rc_bittree_price(lc
->high
, LEN_HIGH_BITS
,
100 i
- LEN_LOW_SYMBOLS
- LEN_MID_SYMBOLS
);
107 length(lzma_range_encoder
*rc
, lzma_length_encoder
*lc
,
108 const uint32_t pos_state
, uint32_t len
, const bool fast_mode
)
110 assert(len
<= MATCH_LEN_MAX
);
111 len
-= MATCH_LEN_MIN
;
113 if (len
< LEN_LOW_SYMBOLS
) {
114 rc_bit(rc
, &lc
->choice
, 0);
115 rc_bittree(rc
, lc
->low
[pos_state
], LEN_LOW_BITS
, len
);
117 rc_bit(rc
, &lc
->choice
, 1);
118 len
-= LEN_LOW_SYMBOLS
;
120 if (len
< LEN_MID_SYMBOLS
) {
121 rc_bit(rc
, &lc
->choice2
, 0);
122 rc_bittree(rc
, lc
->mid
[pos_state
], LEN_MID_BITS
, len
);
124 rc_bit(rc
, &lc
->choice2
, 1);
125 len
-= LEN_MID_SYMBOLS
;
126 rc_bittree(rc
, lc
->high
, LEN_HIGH_BITS
, len
);
130 // Only getoptimum uses the prices so don't update the table when
133 if (--lc
->counters
[pos_state
] == 0)
134 length_update_prices(lc
, pos_state
);
143 match(lzma_lzma1_encoder
*coder
, const uint32_t pos_state
,
144 const uint32_t distance
, const uint32_t len
)
146 update_match(coder
->state
);
148 length(&coder
->rc
, &coder
->match_len_encoder
, pos_state
, len
,
151 const uint32_t dist_slot
= get_dist_slot(distance
);
152 const uint32_t dist_state
= get_dist_state(len
);
153 rc_bittree(&coder
->rc
, coder
->dist_slot
[dist_state
],
154 DIST_SLOT_BITS
, dist_slot
);
156 if (dist_slot
>= DIST_MODEL_START
) {
157 const uint32_t footer_bits
= (dist_slot
>> 1) - 1;
158 const uint32_t base
= (2 | (dist_slot
& 1)) << footer_bits
;
159 const uint32_t dist_reduced
= distance
- base
;
161 if (dist_slot
< DIST_MODEL_END
) {
162 // Careful here: base - dist_slot - 1 can be -1, but
163 // rc_bittree_reverse starts at probs[1], not probs[0].
164 rc_bittree_reverse(&coder
->rc
,
165 coder
->dist_special
+ base
- dist_slot
- 1,
166 footer_bits
, dist_reduced
);
168 rc_direct(&coder
->rc
, dist_reduced
>> ALIGN_BITS
,
169 footer_bits
- ALIGN_BITS
);
171 &coder
->rc
, coder
->dist_align
,
172 ALIGN_BITS
, dist_reduced
& ALIGN_MASK
);
173 ++coder
->align_price_count
;
177 coder
->reps
[3] = coder
->reps
[2];
178 coder
->reps
[2] = coder
->reps
[1];
179 coder
->reps
[1] = coder
->reps
[0];
180 coder
->reps
[0] = distance
;
181 ++coder
->match_price_count
;
190 rep_match(lzma_lzma1_encoder
*coder
, const uint32_t pos_state
,
191 const uint32_t rep
, const uint32_t len
)
194 rc_bit(&coder
->rc
, &coder
->is_rep0
[coder
->state
], 0);
196 &coder
->is_rep0_long
[coder
->state
][pos_state
],
199 const uint32_t distance
= coder
->reps
[rep
];
200 rc_bit(&coder
->rc
, &coder
->is_rep0
[coder
->state
], 1);
203 rc_bit(&coder
->rc
, &coder
->is_rep1
[coder
->state
], 0);
205 rc_bit(&coder
->rc
, &coder
->is_rep1
[coder
->state
], 1);
206 rc_bit(&coder
->rc
, &coder
->is_rep2
[coder
->state
],
210 coder
->reps
[3] = coder
->reps
[2];
212 coder
->reps
[2] = coder
->reps
[1];
215 coder
->reps
[1] = coder
->reps
[0];
216 coder
->reps
[0] = distance
;
220 update_short_rep(coder
->state
);
222 length(&coder
->rc
, &coder
->rep_len_encoder
, pos_state
, len
,
224 update_long_rep(coder
->state
);
234 encode_symbol(lzma_lzma1_encoder
*coder
, lzma_mf
*mf
,
235 uint32_t back
, uint32_t len
, uint32_t position
)
237 const uint32_t pos_state
= position
& coder
->pos_mask
;
239 if (back
== UINT32_MAX
) {
240 // Literal i.e. eight-bit byte
243 &coder
->is_match
[coder
->state
][pos_state
], 0);
244 literal(coder
, mf
, position
);
246 // Some type of match
248 &coder
->is_match
[coder
->state
][pos_state
], 1);
251 // It's a repeated match i.e. the same distance
252 // has been used earlier.
253 rc_bit(&coder
->rc
, &coder
->is_rep
[coder
->state
], 1);
254 rep_match(coder
, pos_state
, back
, len
);
257 rc_bit(&coder
->rc
, &coder
->is_rep
[coder
->state
], 0);
258 match(coder
, pos_state
, back
- REPS
, len
);
262 assert(mf
->read_ahead
>= len
);
263 mf
->read_ahead
-= len
;
268 encode_init(lzma_lzma1_encoder
*coder
, lzma_mf
*mf
)
270 assert(mf_position(mf
) == 0);
272 if (mf
->read_pos
== mf
->read_limit
) {
273 if (mf
->action
== LZMA_RUN
)
274 return false; // We cannot do anything.
276 // We are finishing (we cannot get here when flushing).
277 assert(mf
->write_pos
== mf
->read_pos
);
278 assert(mf
->action
== LZMA_FINISH
);
280 // Do the actual initialization. The first LZMA symbol must
281 // always be a literal.
284 rc_bit(&coder
->rc
, &coder
->is_match
[0][0], 0);
285 rc_bittree(&coder
->rc
, coder
->literal
[0], 8, mf
->buffer
[0]);
288 // Initialization is done (except if empty file).
289 coder
->is_initialized
= true;
296 encode_eopm(lzma_lzma1_encoder
*coder
, uint32_t position
)
298 const uint32_t pos_state
= position
& coder
->pos_mask
;
299 rc_bit(&coder
->rc
, &coder
->is_match
[coder
->state
][pos_state
], 1);
300 rc_bit(&coder
->rc
, &coder
->is_rep
[coder
->state
], 0);
301 match(coder
, pos_state
, UINT32_MAX
, MATCH_LEN_MIN
);
305 /// Number of bytes that a single encoding loop in lzma_lzma_encode() can
306 /// consume from the dictionary. This limit comes from lzma_lzma_optimum()
307 /// and may need to be updated if that function is significantly modified.
308 #define LOOP_INPUT_MAX (OPTS + 1)
312 lzma_lzma_encode(lzma_lzma1_encoder
*restrict coder
, lzma_mf
*restrict mf
,
313 uint8_t *restrict out
, size_t *restrict out_pos
,
314 size_t out_size
, uint32_t limit
)
316 // Initialize the stream if no data has been encoded yet.
317 if (!coder
->is_initialized
&& !encode_init(coder
, mf
))
320 // Get the lowest bits of the uncompressed offset from the LZ layer.
321 uint32_t position
= mf_position(mf
);
324 // Encode pending bits, if any. Calling this before encoding
325 // the next symbol is needed only with plain LZMA, since
326 // LZMA2 always provides big enough buffer to flush
327 // everything out from the range encoder. For the same reason,
328 // rc_encode() never returns true when this function is used
329 // as part of LZMA2 encoder.
330 if (rc_encode(&coder
->rc
, out
, out_pos
, out_size
)) {
331 assert(limit
== UINT32_MAX
);
335 // With LZMA2 we need to take care that compressed size of
336 // a chunk doesn't get too big.
337 // FIXME? Check if this could be improved.
338 if (limit
!= UINT32_MAX
339 && (mf
->read_pos
- mf
->read_ahead
>= limit
340 || *out_pos
+ rc_pending(&coder
->rc
)
345 // Check that there is some input to process.
346 if (mf
->read_pos
>= mf
->read_limit
) {
347 if (mf
->action
== LZMA_RUN
)
350 if (mf
->read_ahead
== 0)
354 // Get optimal match (repeat position and length).
355 // Value ranges for pos:
356 // - [0, REPS): repeated match
357 // - [REPS, UINT32_MAX):
358 // match at (pos - REPS)
359 // - UINT32_MAX: not a match but a literal
360 // Value ranges for len:
361 // - [MATCH_LEN_MIN, MATCH_LEN_MAX]
365 if (coder
->fast_mode
)
366 lzma_lzma_optimum_fast(coder
, mf
, &back
, &len
);
368 lzma_lzma_optimum_normal(
369 coder
, mf
, &back
, &len
, position
);
371 encode_symbol(coder
, mf
, back
, len
, position
);
376 if (!coder
->is_flushed
) {
377 coder
->is_flushed
= true;
379 // We don't support encoding plain LZMA streams without EOPM,
380 // and LZMA2 doesn't use EOPM at LZMA level.
381 if (limit
== UINT32_MAX
)
382 encode_eopm(coder
, position
);
384 // Flush the remaining bytes from the range encoder.
385 rc_flush(&coder
->rc
);
387 // Copy the remaining bytes to the output buffer. If there
388 // isn't enough output space, we will copy out the remaining
389 // bytes on the next call to this function by using
390 // the rc_encode() call in the encoding loop above.
391 if (rc_encode(&coder
->rc
, out
, out_pos
, out_size
)) {
392 assert(limit
== UINT32_MAX
);
397 // Make it ready for the next LZMA2 chunk.
398 coder
->is_flushed
= false;
400 return LZMA_STREAM_END
;
405 lzma_encode(void *coder
, lzma_mf
*restrict mf
,
406 uint8_t *restrict out
, size_t *restrict out_pos
,
409 // Plain LZMA has no support for sync-flushing.
410 if (unlikely(mf
->action
== LZMA_SYNC_FLUSH
))
411 return LZMA_OPTIONS_ERROR
;
413 return lzma_lzma_encode(coder
, mf
, out
, out_pos
, out_size
, UINT32_MAX
);
422 is_options_valid(const lzma_options_lzma
*options
)
424 // Validate some of the options. LZ encoder validates nice_len too
425 // but we need a valid value here earlier.
426 return is_lclppb_valid(options
)
427 && options
->nice_len
>= MATCH_LEN_MIN
428 && options
->nice_len
<= MATCH_LEN_MAX
429 && (options
->mode
== LZMA_MODE_FAST
430 || options
->mode
== LZMA_MODE_NORMAL
);
435 set_lz_options(lzma_lz_options
*lz_options
, const lzma_options_lzma
*options
)
437 // LZ encoder initialization does the validation for these so we
438 // don't need to validate here.
439 lz_options
->before_size
= OPTS
;
440 lz_options
->dict_size
= options
->dict_size
;
441 lz_options
->after_size
= LOOP_INPUT_MAX
;
442 lz_options
->match_len_max
= MATCH_LEN_MAX
;
443 lz_options
->nice_len
= options
->nice_len
;
444 lz_options
->match_finder
= options
->mf
;
445 lz_options
->depth
= options
->depth
;
446 lz_options
->preset_dict
= options
->preset_dict
;
447 lz_options
->preset_dict_size
= options
->preset_dict_size
;
453 length_encoder_reset(lzma_length_encoder
*lencoder
,
454 const uint32_t num_pos_states
, const bool fast_mode
)
456 bit_reset(lencoder
->choice
);
457 bit_reset(lencoder
->choice2
);
459 for (size_t pos_state
= 0; pos_state
< num_pos_states
; ++pos_state
) {
460 bittree_reset(lencoder
->low
[pos_state
], LEN_LOW_BITS
);
461 bittree_reset(lencoder
->mid
[pos_state
], LEN_MID_BITS
);
464 bittree_reset(lencoder
->high
, LEN_HIGH_BITS
);
467 for (uint32_t pos_state
= 0; pos_state
< num_pos_states
;
469 length_update_prices(lencoder
, pos_state
);
476 lzma_lzma_encoder_reset(lzma_lzma1_encoder
*coder
,
477 const lzma_options_lzma
*options
)
479 if (!is_options_valid(options
))
480 return LZMA_OPTIONS_ERROR
;
482 coder
->pos_mask
= (1U << options
->pb
) - 1;
483 coder
->literal_context_bits
= options
->lc
;
484 coder
->literal_pos_mask
= (1U << options
->lp
) - 1;
487 rc_reset(&coder
->rc
);
490 coder
->state
= STATE_LIT_LIT
;
491 for (size_t i
= 0; i
< REPS
; ++i
)
494 literal_init(coder
->literal
, options
->lc
, options
->lp
);
497 for (size_t i
= 0; i
< STATES
; ++i
) {
498 for (size_t j
= 0; j
<= coder
->pos_mask
; ++j
) {
499 bit_reset(coder
->is_match
[i
][j
]);
500 bit_reset(coder
->is_rep0_long
[i
][j
]);
503 bit_reset(coder
->is_rep
[i
]);
504 bit_reset(coder
->is_rep0
[i
]);
505 bit_reset(coder
->is_rep1
[i
]);
506 bit_reset(coder
->is_rep2
[i
]);
509 for (size_t i
= 0; i
< FULL_DISTANCES
- DIST_MODEL_END
; ++i
)
510 bit_reset(coder
->dist_special
[i
]);
513 for (size_t i
= 0; i
< DIST_STATES
; ++i
)
514 bittree_reset(coder
->dist_slot
[i
], DIST_SLOT_BITS
);
516 bittree_reset(coder
->dist_align
, ALIGN_BITS
);
519 length_encoder_reset(&coder
->match_len_encoder
,
520 1U << options
->pb
, coder
->fast_mode
);
522 length_encoder_reset(&coder
->rep_len_encoder
,
523 1U << options
->pb
, coder
->fast_mode
);
525 // Price counts are incremented every time appropriate probabilities
526 // are changed. price counts are set to zero when the price tables
527 // are updated, which is done when the appropriate price counts have
528 // big enough value, and lzma_mf.read_ahead == 0 which happens at
529 // least every OPTS (a few thousand) possible price count increments.
531 // By resetting price counts to UINT32_MAX / 2, we make sure that the
532 // price tables will be initialized before they will be used (since
533 // the value is definitely big enough), and that it is OK to increment
534 // price counts without risk of integer overflow (since UINT32_MAX / 2
535 // is small enough). The current code doesn't increment price counts
536 // before initializing price tables, but it maybe done in future if
537 // we add support for saving the state between LZMA2 chunks.
538 coder
->match_price_count
= UINT32_MAX
/ 2;
539 coder
->align_price_count
= UINT32_MAX
/ 2;
541 coder
->opts_end_index
= 0;
542 coder
->opts_current_index
= 0;
549 lzma_lzma_encoder_create(void **coder_ptr
,
550 const lzma_allocator
*allocator
,
551 const lzma_options_lzma
*options
, lzma_lz_options
*lz_options
)
553 // Allocate lzma_lzma1_encoder if it wasn't already allocated.
554 if (*coder_ptr
== NULL
) {
555 *coder_ptr
= lzma_alloc(sizeof(lzma_lzma1_encoder
), allocator
);
556 if (*coder_ptr
== NULL
)
557 return LZMA_MEM_ERROR
;
560 lzma_lzma1_encoder
*coder
= *coder_ptr
;
562 // Set compression mode. We haven't validates the options yet,
563 // but it's OK here, since nothing bad happens with invalid
564 // options in the code below, and they will get rejected by
565 // lzma_lzma_encoder_reset() call at the end of this function.
566 switch (options
->mode
) {
568 coder
->fast_mode
= true;
571 case LZMA_MODE_NORMAL
: {
572 coder
->fast_mode
= false;
574 // Set dist_table_size.
575 // Round the dictionary size up to next 2^n.
576 uint32_t log_size
= 0;
577 while ((UINT32_C(1) << log_size
) < options
->dict_size
)
580 coder
->dist_table_size
= log_size
* 2;
582 // Length encoders' price table size
583 coder
->match_len_encoder
.table_size
584 = options
->nice_len
+ 1 - MATCH_LEN_MIN
;
585 coder
->rep_len_encoder
.table_size
586 = options
->nice_len
+ 1 - MATCH_LEN_MIN
;
591 return LZMA_OPTIONS_ERROR
;
594 // We don't need to write the first byte as literal if there is
595 // a non-empty preset dictionary. encode_init() wouldn't even work
596 // if there is a non-empty preset dictionary, because encode_init()
597 // assumes that position is zero and previous byte is also zero.
598 coder
->is_initialized
= options
->preset_dict
!= NULL
599 && options
->preset_dict_size
> 0;
600 coder
->is_flushed
= false;
602 set_lz_options(lz_options
, options
);
604 return lzma_lzma_encoder_reset(coder
, options
);
609 lzma_encoder_init(lzma_lz_encoder
*lz
, const lzma_allocator
*allocator
,
610 const void *options
, lzma_lz_options
*lz_options
)
612 lz
->code
= &lzma_encode
;
613 return lzma_lzma_encoder_create(
614 &lz
->coder
, allocator
, options
, lz_options
);
619 lzma_lzma_encoder_init(lzma_next_coder
*next
, const lzma_allocator
*allocator
,
620 const lzma_filter_info
*filters
)
622 return lzma_lz_encoder_init(
623 next
, allocator
, filters
, &lzma_encoder_init
);
628 lzma_lzma_encoder_memusage(const void *options
)
630 if (!is_options_valid(options
))
633 lzma_lz_options lz_options
;
634 set_lz_options(&lz_options
, options
);
636 const uint64_t lz_memusage
= lzma_lz_encoder_memusage(&lz_options
);
637 if (lz_memusage
== UINT64_MAX
)
640 return (uint64_t)(sizeof(lzma_lzma1_encoder
)) + lz_memusage
;
645 lzma_lzma_lclppb_encode(const lzma_options_lzma
*options
, uint8_t *byte
)
647 if (!is_lclppb_valid(options
))
650 *byte
= (options
->pb
* 5 + options
->lp
) * 9 + options
->lc
;
651 assert(*byte
<= (4 * 5 + 4) * 9 + 8);
657 #ifdef HAVE_ENCODER_LZMA1
659 lzma_lzma_props_encode(const void *options
, uint8_t *out
)
661 const lzma_options_lzma
*const opt
= options
;
663 if (lzma_lzma_lclppb_encode(opt
, out
))
664 return LZMA_PROG_ERROR
;
666 unaligned_write32le(out
+ 1, opt
->dict_size
);
673 extern LZMA_API(lzma_bool
)
674 lzma_mode_is_supported(lzma_mode mode
)
676 return mode
== LZMA_MODE_FAST
|| mode
== LZMA_MODE_NORMAL
;