/*
* lzx-compress.c
- *
- * LZX compression routines
*/
/*
- * Copyright (C) 2012, 2013 Eric Biggers
+ * Copyright (C) 2012, 2013, 2014 Eric Biggers
*
* This file is part of wimlib, a library for working with WIM files.
*
/*
- * This file contains a compressor for the LZX compression format, as used in
- * the WIM file format.
+ * This file contains a compressor for the LZX ("Lempel-Ziv eXtended"?)
+ * compression format, as used in the WIM (Windows IMaging) file format. This
+ * code may need some slight modifications to be used outside of the WIM format.
+ * In particular, in other situations the LZX block header might be slightly
+ * different, and a sliding window rather than a fixed-size window might be
+ * required.
*
- * Format
- * ======
+ * ----------------------------------------------------------------------------
*
- * First, the primary reference for the LZX compression format is the
- * specification released by Microsoft.
+ * Format Overview
*
- * Second, the comments in lzx-decompress.c provide some more information about
- * the LZX compression format, including errors in the Microsoft specification.
+ * The primary reference for LZX is the specification released by Microsoft.
+ * However, the comments in lzx-decompress.c provide more information about LZX
+ * and note some errors in the Microsoft specification.
*
- * Do note that LZX shares many similarities with DEFLATE, the algorithm used by
- * zlib and gzip. Both LZX and DEFLATE use LZ77 matching and Huffman coding,
- * and certain other details are quite similar, such as the method for storing
- * Huffman codes. However, some of the main differences are:
+ * LZX shares many similarities with DEFLATE, the format used by zlib and gzip.
+ * Both LZX and DEFLATE use LZ77 matching and Huffman coding. Certain details
+ * are quite similar, such as the method for storing Huffman codes. However,
+ * the main differences are:
*
* - LZX preprocesses the data to attempt to make x86 machine code slightly more
* compressible before attempting to compress it further.
+ *
* - LZX uses a "main" alphabet which combines literals and matches, with the
* match symbols containing a "length header" (giving all or part of the match
* length) and a "position slot" (giving, roughly speaking, the order of
* magnitude of the match offset).
- * - LZX does not have static Huffman blocks; however it does have two types of
- * dynamic Huffman blocks ("aligned offset" and "verbatim").
+ *
+ * - LZX does not have static Huffman blocks (that is, the kind with preset
+ * Huffman codes); however it does have two types of dynamic Huffman blocks
+ * ("verbatim" and "aligned").
+ *
* - LZX has a minimum match length of 2 rather than 3.
+ *
* - In LZX, match offsets 0 through 2 actually represent entries in an LRU
* queue of match offsets. This is very useful for certain types of files,
* such as binary files that have repeating records.
*
- * Algorithms
- * ==========
+ * ----------------------------------------------------------------------------
+ *
+ * Algorithmic Overview
+ *
+ * At a high level, any implementation of LZX compression must operate as
+ * follows:
*
- * There are actually two distinct overall algorithms implemented here. We
- * shall refer to them as the "slow" algorithm and the "fast" algorithm. The
- * "slow" algorithm spends more time compressing to achieve a higher compression
- * ratio compared to the "fast" algorithm. More details are presented below.
+ * 1. Preprocess the input data to translate the targets of 32-bit x86 call
+ * instructions to absolute offsets. (Actually, this is required for WIM,
+ * but might not be in other places LZX is used.)
*
- * Slow algorithm
- * --------------
+ * 2. Find a sequence of LZ77-style matches and literal bytes that expands to
+ * the preprocessed data.
*
- * The "slow" algorithm to generate LZX-compressed data is roughly as follows:
+ * 3. Divide the match/literal sequence into one or more LZX blocks, each of
+ * which may be "uncompressed", "verbatim", or "aligned".
*
- * 1. Preprocess the input data to translate the targets of x86 call
- * instructions to absolute offsets.
+ * 4. Output each LZX block.
*
- * 2. Build the suffix array and inverse suffix array for the input data. The
- * suffix array contains the indices of all suffixes of the input data,
- * sorted lexcographically by the corresponding suffixes. The "position" of
- * a suffix is the index of that suffix in the original string, whereas the
- * "rank" of a suffix is the index at which that suffix's position is found
- * in the suffix array.
+ * Step (1) is fairly straightforward. It requires looking for 0xe8 bytes in
+ * the input data and performing a translation on the 4 bytes following each
+ * one.
*
- * 3. Build the longest common prefix array corresponding to the suffix array.
+ * Step (4) is complicated, but it is mostly determined by the LZX format. The
+ * only real choice we have is what algorithm to use to build the length-limited
+ * canonical Huffman codes. See lzx_write_all_blocks() for details.
*
- * 4. For each suffix, find the highest lower ranked suffix that has a lower
- * position, the lowest higher ranked suffix that has a lower position, and
- * the length of the common prefix shared between each. This information is
- * later used to link suffix ranks into a doubly-linked list for searching
- * the suffix array.
+ * That leaves steps (2) and (3) as where all the hard stuff happens. Focusing
+ * on step (2), we need to do LZ77-style parsing on the input data, or "window",
+ * to divide it into a sequence of matches and literals. Each position in the
+ * window might have multiple matches associated with it, and we need to choose
+ * which one, if any, to actually use. Therefore, the problem can really be
+ * divided into two areas of concern: (a) finding matches at a given position,
+ * which we shall call "match-finding", and (b) choosing whether to use a
+ * match or a literal at a given position, and if using a match, which one (if
+ * there is more than one available). We shall call this "match-choosing". We
+ * first consider match-finding, then match-choosing.
*
- * 5. Set a default cost model for matches/literals.
+ * ----------------------------------------------------------------------------
*
- * 6. Determine the lowest cost sequence of LZ77 matches ((offset, length)
- * pairs) and literal bytes to divide the input into. Raw match-finding is
- * done by searching the suffix array using a linked list to avoid
- * considering any suffixes that start after the current position. Each run
- * of the match-finder returns the approximate lowest-cost longest match as
- * well as any shorter matches that have even lower approximate costs. Each
- * such run also adds the suffix rank of the current position into the linked
- * list being used to search the suffix array. Parsing, or match-choosing,
- * is solved as a minimum-cost path problem using a forward "optimal parsing"
- * algorithm based on the Deflate encoder from 7-Zip. This algorithm moves
- * forward calculating the minimum cost to reach each byte until either a
- * very long match is found or until a position is found at which no matches
- * start or overlap.
+ * Match-finding
*
- * 7. Build the Huffman codes needed to output the matches/literals.
+ * Given a position in the window, we want to find LZ77-style "matches" with
+ * that position at previous positions in the window. With LZX, the minimum
+ * match length is 2 and the maximum match length is 257. The only restriction
+ * on offsets is that LZX does not allow the last 2 bytes of the window to match
+ * the the beginning of the window.
*
- * 8. Up to a certain number of iterations, use the resulting Huffman codes to
- * refine a cost model and go back to Step #6 to determine an improved
- * sequence of matches and literals.
+ * Depending on how good a compression ratio we want (see the "Match-choosing"
+ * section), we may want to find: (a) all matches, or (b) just the longest
+ * match, or (c) just some "promising" matches that we are able to find quickly,
+ * or (d) just the longest match that we're able to find quickly. Below we
+ * introduce the match-finding methods that the code currently uses or has
+ * previously used:
*
- * 9. Output the resulting block using the match/literal sequences and the
- * Huffman codes that were computed for the block.
+ * - Hash chains. Maintain a table that maps hash codes, computed from
+ * fixed-length byte sequences, to linked lists containing previous window
+ * positions. To search for matches, compute the hash for the current
+ * position in the window and search the appropriate hash chain. When
+ * advancing to the next position, prepend the current position to the
+ * appropriate hash list. This is a good approach for producing matches with
+ * stategy (d) and is useful for fast compression. Therefore, we provide an
+ * option to use this method for LZX compression. See lz_hash.c for the
+ * implementation.
*
- * Note: the algorithm does not yet attempt to split the input into multiple LZX
- * blocks; it instead uses a series of blocks of LZX_DIV_BLOCK_SIZE bytes.
+ * - Binary trees. Similar to hash chains, but each hash bucket contains a
+ * binary tree of previous window positions rather than a linked list. This
+ * is a good approach for producing matches with stategy (c) and is useful for
+ * achieving a good compression ratio. Therefore, we provide an option to use
+ * this method; see lz_bt.c for the implementation.
*
- * Fast algorithm
- * --------------
+ * - Suffix arrays. This code previously used this method to produce matches
+ * with stategy (c), but I've dropped it because it was slower than the binary
+ * trees approach, used more memory, and did not improve the compression ratio
+ * enough to compensate. Download wimlib v1.6.2 if you want the code.
+ * However, the suffix array method was basically as follows. Build the
+ * suffix array for the entire window. The suffix array contains each
+ * possible window position, sorted by the lexicographic order of the strings
+ * that begin at those positions. Find the matches at a given position by
+ * searching the suffix array outwards, in both directions, from the suffix
+ * array slot for that position. This produces the longest matches first, but
+ * "matches" that actually occur at later positions in the window must be
+ * skipped. To do this skipping, use an auxiliary array with dynamically
+ * constructed linked lists. Also, use the inverse suffix array to quickly
+ * find the suffix array slot for a given position without doing a binary
+ * search.
*
- * The fast algorithm (and the only one available in wimlib v1.5.1 and earlier)
- * spends much less time on the main bottlenecks of the compression process ---
- * that is, the match finding and match choosing. Matches are found and chosen
- * with hash chains using a greedy parse with one position of look-ahead. No
- * block splitting is done; only compressing the full input into an aligned
- * offset block is considered.
+ * ----------------------------------------------------------------------------
*
- * Acknowledgments
- * ===============
+ * Match-choosing
*
- * Acknowledgments to several open-source projects and research papers that made
- * it possible to implement this code:
+ * Usually, choosing the longest match is best because it encodes the most data
+ * in that one item. However, sometimes the longest match is not optimal
+ * because (a) choosing a long match now might prevent using an even longer
+ * match later, or (b) more generally, what we actually care about is the number
+ * of bits it will ultimately take to output each match or literal, which is
+ * actually dependent on the entropy encoding using by the underlying
+ * compression format. Consequently, a longer match usually, but not always,
+ * takes fewer bits to encode than multiple shorter matches or literals that
+ * cover the same data.
*
- * - divsufsort (author: Yuta Mori), for the suffix array construction code,
- * located in a separate file (divsufsort.c).
+ * This problem of choosing the truly best match/literal sequence is probably
+ * impossible to solve efficiently when combined with entropy encoding. If we
+ * knew how many bits it takes to output each match/literal, then we could
+ * choose the optimal sequence using shortest-path search a la Dijkstra's
+ * algorithm. However, with entropy encoding, the chosen match/literal sequence
+ * affects its own encoding. Therefore, we can't know how many bits it will
+ * take to actually output any one match or literal until we have actually
+ * chosen the full sequence of matches and literals.
*
- * - "Linear-Time Longest-Common-Prefix Computation in Suffix Arrays and Its
- * Applications" (Kasai et al. 2001), for the LCP array computation.
+ * Notwithstanding the entropy encoding problem, we also aren't guaranteed to
+ * choose the optimal match/literal sequence unless the match-finder (see
+ * section "Match-finder") provides the match-chooser with all possible matches
+ * at each position. However, this is not computationally efficient. For
+ * example, there might be many matches of the same length, and usually (but not
+ * always) the best choice is the one with the smallest offset. So in practice,
+ * it's fine to only consider the smallest offset for a given match length at a
+ * given position. (Actually, for LZX, it's also worth considering repeat
+ * offsets.)
*
- * - "LPF computation revisited" (Crochemore et al. 2009) for the prev and next
- * array computations.
+ * In addition, as mentioned earlier, in LZX we have the choice of using
+ * multiple blocks, each of which resets the Huffman codes. This expands the
+ * search space even further. Therefore, to simplify the problem, we currently
+ * we don't attempt to actually choose the LZX blocks based on the data.
+ * Instead, we just divide the data into fixed-size blocks of LZX_DIV_BLOCK_SIZE
+ * bytes each, and always use verbatim or aligned blocks (never uncompressed).
+ * A previous version of this code recursively split the input data into
+ * equal-sized blocks, up to a maximum depth, and chose the lowest-cost block
+ * divisions. However, this made compression much slower and did not actually
+ * help very much. It remains an open question whether a sufficiently fast and
+ * useful block-splitting algorithm is possible for LZX. Essentially the same
+ * problem also applies to DEFLATE. The Microsoft LZX compressor seemingly does
+ * do block splitting, although I don't know how fast or useful it is,
+ * specifically.
*
- * - 7-Zip (author: Igor Pavlov) for the algorithm for forward optimal parsing
- * (match-choosing).
+ * Now, back to the entropy encoding problem. The "solution" is to use an
+ * iterative approach to compute a good, but not necessarily optimal,
+ * match/literal sequence. Start with a fixed assignment of symbol costs and
+ * choose an "optimal" match/literal sequence based on those costs, using
+ * shortest-path seach a la Dijkstra's algorithm. Then, for each iteration of
+ * the optimization, update the costs based on the entropy encoding of the
+ * current match/literal sequence, then choose a new match/literal sequence
+ * based on the updated costs. Usually, the actual cost to output the current
+ * match/literal sequence will decrease in each iteration until it converges on
+ * a fixed point. This result may not be the truly optimal match/literal
+ * sequence, but it usually is much better than one chosen by doing a "greedy"
+ * parse where we always chooe the longest match.
*
- * - zlib (author: Jean-loup Gailly and Mark Adler), for the hash table
- * match-finding algorithm (used in lz77.c).
+ * An alternative to both greedy parsing and iterative, near-optimal parsing is
+ * "lazy" parsing. Briefly, "lazy" parsing considers just the longest match at
+ * each position, but it waits to choose that match until it has also examined
+ * the next position. This is actually a useful approach; it's used by zlib,
+ * for example. Therefore, for fast compression we combine lazy parsing with
+ * the hash chain max-finder. For normal/high compression we combine
+ * near-optimal parsing with the binary tree match-finder.
*
- * - lzx-compress (author: Matthew T. Russotto), on which some parts of this
- * code were originally based.
+ * Anyway, if you've read through this comment, you hopefully should have a
+ * better idea of why things are done in a certain way in this LZX compressor,
+ * as well as in other compressors for LZ77-based formats (including third-party
+ * ones). In my opinion, the phrase "compression algorithm" is often mis-used
+ * in place of "compression format", since there can be many different
+ * algorithms that all generate compressed data in the same format. The
+ * challenge is to design an algorithm that is efficient but still gives a good
+ * compression ratio.
*/
#ifdef HAVE_CONFIG_H
#include "wimlib/compress_common.h"
#include "wimlib/endianness.h"
#include "wimlib/error.h"
+#include "wimlib/lz.h"
#include "wimlib/lz_hash.h"
-#include "wimlib/lz_sarray.h"
+#include "wimlib/lz_bt.h"
#include "wimlib/lzx.h"
#include "wimlib/util.h"
#include <string.h>
# include "wimlib/decompress_common.h"
#endif
-typedef u32 block_cost_t;
-#define INFINITE_BLOCK_COST (~(block_cost_t)0)
-
#define LZX_OPTIM_ARRAY_SIZE 4096
#define LZX_DIV_BLOCK_SIZE 32768
-#define LZX_MAX_CACHE_PER_POS 10
+#define LZX_CACHE_PER_POS 8
+
+#define LZX_CACHE_LEN (LZX_DIV_BLOCK_SIZE * (LZX_CACHE_PER_POS + 1))
+#define LZX_CACHE_SIZE (LZX_CACHE_LEN * sizeof(struct lz_match))
+#define LZX_MAX_MATCHES_PER_POS (LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1)
/* Codewords for the LZX main, length, and aligned offset Huffman codes */
struct lzx_codewords {
*
* If a codeword has zero frequency, it must still be assigned some nonzero cost
* --- generally a high cost, since even if it gets used in the next iteration,
- * it probably will not be used very times. */
+ * it probably will not be used very many times. */
struct lzx_costs {
u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
u8 len[LZX_LENCODE_NUM_SYMBOLS];
/* Tables for tallying symbol frequencies in the three LZX alphabets */
struct lzx_freqs {
- input_idx_t main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
- input_idx_t len[LZX_LENCODE_NUM_SYMBOLS];
- input_idx_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
+ u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
+ u32 len[LZX_LENCODE_NUM_SYMBOLS];
+ u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
};
/* LZX intermediate match/literal format */
-struct lzx_match {
+struct lzx_item {
/* Bit Description
*
* 31 1 if a match, 0 if a literal.
int block_type;
/* 0-based position in the window at which this block starts. */
- input_idx_t window_pos;
+ u32 window_pos;
/* The number of bytes of uncompressed data this block represents. */
- input_idx_t block_size;
+ u32 block_size;
- /* The position in the 'chosen_matches' array in the `struct
- * lzx_compressor' at which the match/literal specifications for
- * this block begin. */
- input_idx_t chosen_matches_start_pos;
+ /* The match/literal sequence for this block. */
+ struct lzx_item *chosen_items;
- /* The number of match/literal specifications for this block. */
- input_idx_t num_chosen_matches;
+ /* The length of the @chosen_items sequence. */
+ u32 num_chosen_items;
/* Huffman codes for this block. */
struct lzx_codes codes;
};
-/* Include template for the match-choosing algorithm. */
-#define LZ_COMPRESSOR struct lzx_compressor
-#define LZ_ADAPTIVE_STATE struct lzx_lru_queue
-struct lzx_compressor;
-#include "wimlib/lz_optimal.h"
-
/* State of the LZX compressor. */
struct lzx_compressor {
/* Number of bytes of data to be compressed, which is the number of
* bytes of data in @window that are actually valid. */
- input_idx_t window_size;
+ u32 window_size;
/* Allocated size of the @window. */
- input_idx_t max_window_size;
+ u32 max_window_size;
/* Number of symbols in the main alphabet (depends on the
* @max_window_size since it determines the maximum allowed offset). */
/* Space for the sequences of matches/literals that were chosen for each
* block. */
- struct lzx_match *chosen_matches;
+ struct lzx_item *chosen_items;
/* Information about the LZX blocks the preprocessed input was divided
* into. */
struct lzx_costs costs;
/* Fast algorithm only: Array of hash table links. */
- input_idx_t *prev_tab;
+ u32 *prev_tab;
- /* Slow algorithm only: Suffix array match-finder. */
- struct lz_sarray lz_sarray;
+ /* Slow algorithm only: Binary tree match-finder. */
+ struct lz_bt mf;
/* Position in window of next match to return. */
- input_idx_t match_window_pos;
+ u32 match_window_pos;
- /* The match-finder shall ensure the length of matches does not exceed
- * this position in the input. */
- input_idx_t match_window_end;
+ /* The end-of-block position. We can't allow any matches to span this
+ * position. */
+ u32 match_window_end;
/* Matches found by the match-finder are cached in the following array
* to achieve a slight speedup when the same matches are needed on
* subsequent passes. This is suboptimal because different matches may
* be preferred with different cost models, but seems to be a worthwhile
* speedup. */
- struct raw_match *cached_matches;
- unsigned cached_matches_pos;
+ struct lz_match *cached_matches;
+ struct lz_match *cache_ptr;
bool matches_cached;
+ struct lz_match *cache_limit;
+
+ /* Match-chooser state.
+ * When matches have been chosen, optimum_cur_idx is set to the position
+ * in the window of the next match/literal to return and optimum_end_idx
+ * is set to the position in the window at the end of the last
+ * match/literal to return. */
+ struct lzx_mc_pos_data *optimum;
+ unsigned optimum_cur_idx;
+ unsigned optimum_end_idx;
+};
- /* Match chooser. */
- struct lz_match_chooser mc;
+/*
+ * Match chooser position data:
+ *
+ * An array of these structures is used during the match-choosing algorithm.
+ * They correspond to consecutive positions in the window and are used to keep
+ * track of the cost to reach each position, and the match/literal choices that
+ * need to be chosen to reach that position.
+ */
+struct lzx_mc_pos_data {
+ /* The approximate minimum cost, in bits, to reach this position in the
+ * window which has been found so far. */
+ u32 cost;
+#define MC_INFINITE_COST ((u32)~0UL)
+
+ /* The union here is just for clarity, since the fields are used in two
+ * slightly different ways. Initially, the @prev structure is filled in
+ * first, and links go from later in the window to earlier in the
+ * window. Later, @next structure is filled in and links go from
+ * earlier in the window to later in the window. */
+ union {
+ struct {
+ /* Position of the start of the match or literal that
+ * was taken to get to this position in the approximate
+ * minimum-cost parse. */
+ u32 link;
+
+ /* Offset (as in an LZ (length, offset) pair) of the
+ * match or literal that was taken to get to this
+ * position in the approximate minimum-cost parse. */
+ u32 match_offset;
+ } prev;
+ struct {
+ /* Position at which the match or literal starting at
+ * this position ends in the minimum-cost parse. */
+ u32 link;
+
+ /* Offset (as in an LZ (length, offset) pair) of the
+ * match or literal starting at this position in the
+ * approximate minimum-cost parse. */
+ u32 match_offset;
+ } next;
+ };
+
+ /* Adaptive state that exists after an approximate minimum-cost path to
+ * reach this position is taken. */
+ struct lzx_lru_queue queue;
};
/* Returns the LZX position slot that corresponds to a given match offset,
* taking into account the recent offset queue and updating it if the offset is
* found in it. */
static unsigned
-lzx_get_position_slot(unsigned offset, struct lzx_lru_queue *queue)
+lzx_get_position_slot(u32 offset, struct lzx_lru_queue *queue)
{
unsigned position_slot;
/* See if the offset was recently used. */
- for (unsigned i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
+ for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
if (offset == queue->R[i]) {
/* Found it. */
position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET);
/* Bring the new offset to the front of the queue. */
- for (unsigned i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--)
+ for (int i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--)
queue->R[i] = queue->R[i - 1];
queue->R[0] = offset;
* The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or
* LZX_BLOCKTYPE_VERBATIM)
* @match:
- * The match, as a (length, offset) pair.
+ * The match data.
* @codes:
* Pointer to a structure that contains the codewords for the main, length,
* and aligned offset Huffman codes for the current LZX compressed block.
*/
static void
lzx_write_match(struct output_bitstream *out, int block_type,
- struct lzx_match match, const struct lzx_codes *codes)
+ struct lzx_item match, const struct lzx_codes *codes)
{
/* low 8 bits are the match length minus 2 */
unsigned match_len_minus_2 = match.data & 0xff;
lzx_build_precode(const u8 lens[restrict],
const u8 prev_lens[restrict],
const unsigned num_syms,
- input_idx_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
+ u32 precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
u8 output_syms[restrict num_syms],
u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS],
u32 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS],
const u8 prev_lens[restrict],
unsigned num_syms)
{
- input_idx_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
+ u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
u8 output_syms[num_syms];
u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
* @block_type
* The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
* LZX_BLOCKTYPE_VERBATIM).
- * @match_tab
+ * @items
* The array of matches/literals to output.
- * @match_count
- * Number of matches/literals to output (length of @match_tab).
+ * @num_items
+ * Number of matches/literals to output (length of @items).
* @codes
* The main, length, and aligned offset Huffman codes for the current
* LZX compressed block.
*/
static void
-lzx_write_matches_and_literals(struct output_bitstream *ostream,
- int block_type,
- const struct lzx_match match_tab[],
- unsigned match_count,
- const struct lzx_codes *codes)
+lzx_write_items(struct output_bitstream *ostream, int block_type,
+ const struct lzx_item items[], u32 num_items,
+ const struct lzx_codes *codes)
{
- for (unsigned i = 0; i < match_count; i++) {
- struct lzx_match match = match_tab[i];
-
+ for (u32 i = 0; i < num_items; i++) {
/* The high bit of the 32-bit intermediate representation
* indicates whether the item is an actual LZ-style match (1) or
* a literal byte (0). */
- if (match.data & 0x80000000)
- lzx_write_match(ostream, block_type, match, codes);
+ if (items[i].data & 0x80000000)
+ lzx_write_match(ostream, block_type, items[i], codes);
else
- lzx_write_literal(ostream, match.data, codes);
+ lzx_write_literal(ostream, items[i].data, codes);
}
}
unsigned block_size,
unsigned max_window_size,
unsigned num_main_syms,
- struct lzx_match * chosen_matches,
- unsigned num_chosen_matches,
+ struct lzx_item * chosen_items,
+ unsigned num_chosen_items,
const struct lzx_codes * codes,
const struct lzx_codes * prev_codes,
struct output_bitstream * ostream)
LZX_DEBUG("Writing matches and literals...");
/* Write the actual matches and literals. */
- lzx_write_matches_and_literals(ostream, block_type,
- chosen_matches, num_chosen_matches,
- codes);
+ lzx_write_items(ostream, block_type,
+ chosen_items, num_chosen_items, codes);
LZX_DEBUG("Done writing block.");
}
for (unsigned i = 0; i < ctx->num_blocks; i++) {
const struct lzx_block_spec *spec = &ctx->block_specs[i];
- LZX_DEBUG("Writing block %u/%u (type=%d, size=%u, num_chosen_matches=%u)...",
+ LZX_DEBUG("Writing block %u/%u (type=%d, size=%u, num_chosen_items=%u)...",
i + 1, ctx->num_blocks,
spec->block_type, spec->block_size,
- spec->num_chosen_matches);
+ spec->num_chosen_items);
lzx_write_compressed_block(spec->block_type,
spec->block_size,
ctx->max_window_size,
ctx->num_main_syms,
- &ctx->chosen_matches[spec->chosen_matches_start_pos],
- spec->num_chosen_matches,
+ spec->chosen_items,
+ spec->num_chosen_items,
&spec->codes,
prev_codes,
ostream);
/* Constructs an LZX match from a literal byte and updates the main code symbol
* frequencies. */
-static u32
+static inline u32
lzx_tally_literal(u8 lit, struct lzx_freqs *freqs)
{
freqs->main[lit]++;
* queue and the frequency of symbols in the main, length, and aligned offset
* alphabets. The return value is a 32-bit number that provides the match in an
* intermediate representation documented below. */
-static u32
-lzx_tally_match(unsigned match_len, unsigned match_offset,
+static inline u32
+lzx_tally_match(unsigned match_len, u32 match_offset,
struct lzx_freqs *freqs, struct lzx_lru_queue *queue)
{
unsigned position_slot;
freqs->aligned[position_footer & 7]++;
/* Pack the position slot, position footer, and match length into an
- * intermediate representation. See `struct lzx_match' for details.
+ * intermediate representation. See `struct lzx_item' for details.
*/
LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64);
LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17);
struct lzx_record_ctx {
struct lzx_freqs freqs;
struct lzx_lru_queue queue;
- struct lzx_match *matches;
+ struct lzx_item *matches;
};
static void
/* Returns the cost, in bits, to output a literal byte using the specified cost
* model. */
-static unsigned
+static u32
lzx_literal_cost(u8 c, const struct lzx_costs * costs)
{
return costs->main[c];
* well as costs for the codewords in the main, length, and aligned Huffman
* codes, return the approximate number of bits it will take to represent this
* match in the compressed output. Take into account the match offset LRU
- * queue and optionally update it. */
-static unsigned
-lzx_match_cost(unsigned length, unsigned offset, const struct lzx_costs *costs,
+ * queue and also updates it. */
+static u32
+lzx_match_cost(unsigned length, u32 offset, const struct lzx_costs *costs,
struct lzx_lru_queue *queue)
{
unsigned position_slot;
unsigned len_header, main_symbol;
- unsigned cost = 0;
+ unsigned num_extra_bits;
+ u32 cost = 0;
position_slot = lzx_get_position_slot(offset, queue);
cost += costs->main[main_symbol];
/* Account for extra position information. */
- unsigned num_extra_bits = lzx_get_num_extra_bits(position_slot);
+ num_extra_bits = lzx_get_num_extra_bits(position_slot);
if (num_extra_bits >= 3) {
cost += num_extra_bits - 3;
cost += costs->aligned[(offset + LZX_OFFSET_OFFSET) & 7];
}
-/* Fast heuristic cost evaluation to use in the inner loop of the match-finder.
- * Unlike lzx_match_cost() which does a true cost evaluation, this simply
- * prioritize matches based on their offset. */
-static input_idx_t
-lzx_match_cost_fast(input_idx_t length, input_idx_t offset, const void *_queue)
-{
- const struct lzx_lru_queue *queue = _queue;
-
- /* It seems well worth it to take the time to give priority to recently
- * used offsets. */
- for (input_idx_t i = 0; i < LZX_NUM_RECENT_OFFSETS; i++)
- if (offset == queue->R[i])
- return i;
-
- return offset;
-}
-
/* Set the cost model @ctx->costs from the Huffman codeword lengths specified in
* @lens.
*
}
}
-/* Tell the match-finder to skip the specified number of bytes (@n) in the
- * input. */
-static void
-lzx_lz_skip_bytes(struct lzx_compressor *ctx, input_idx_t n)
-{
- LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos);
- if (ctx->matches_cached) {
- ctx->match_window_pos += n;
- while (n--) {
- ctx->cached_matches_pos +=
- ctx->cached_matches[ctx->cached_matches_pos].len + 1;
- }
- } else {
- while (n--) {
- ctx->cached_matches[ctx->cached_matches_pos++].len = 0;
- lz_sarray_skip_position(&ctx->lz_sarray);
- ctx->match_window_pos++;
- }
- LZX_ASSERT(lz_sarray_get_pos(&ctx->lz_sarray) == ctx->match_window_pos);
- }
-}
-
/* Retrieve a list of matches available at the next position in the input.
*
* A pointer to the matches array is written into @matches_ret, and the return
* value is the number of matches found. */
-static u32
-lzx_lz_get_matches_caching(struct lzx_compressor *ctx,
- const struct lzx_lru_queue *queue,
- struct raw_match **matches_ret)
+static unsigned
+lzx_get_matches(struct lzx_compressor *ctx,
+ const struct lz_match **matches_ret)
{
- u32 num_matches;
- struct raw_match *matches;
-
- LZX_ASSERT(ctx->match_window_pos <= ctx->match_window_end);
+ struct lz_match *cache_ptr;
+ struct lz_match *matches;
+ unsigned num_matches;
- matches = &ctx->cached_matches[ctx->cached_matches_pos + 1];
+ LZX_ASSERT(ctx->match_window_pos < ctx->match_window_end);
- if (ctx->matches_cached) {
- num_matches = matches[-1].len;
+ cache_ptr = ctx->cache_ptr;
+ matches = cache_ptr + 1;
+ if (likely(cache_ptr <= ctx->cache_limit)) {
+ if (ctx->matches_cached) {
+ num_matches = cache_ptr->len;
+ } else {
+ num_matches = lz_bt_get_matches(&ctx->mf, matches);
+ cache_ptr->len = num_matches;
+ }
} else {
- LZX_ASSERT(lz_sarray_get_pos(&ctx->lz_sarray) == ctx->match_window_pos);
- num_matches = lz_sarray_get_matches(&ctx->lz_sarray,
- matches,
- lzx_match_cost_fast,
- queue);
- matches[-1].len = num_matches;
+ num_matches = 0;
}
- ctx->cached_matches_pos += num_matches + 1;
- *matches_ret = matches;
- /* Cap the length of returned matches to the number of bytes remaining,
- * if it is not the whole window. */
- if (ctx->match_window_end < ctx->window_size) {
- unsigned maxlen = ctx->match_window_end - ctx->match_window_pos;
- for (u32 i = 0; i < num_matches; i++)
- if (matches[i].len > maxlen)
- matches[i].len = maxlen;
+ /* Don't allow matches to span the end of an LZX block. */
+ if (ctx->match_window_end < ctx->window_size && num_matches != 0) {
+ unsigned limit = ctx->match_window_end - ctx->match_window_pos;
+
+ if (limit >= LZX_MIN_MATCH_LEN) {
+
+ unsigned i = num_matches - 1;
+ do {
+ if (matches[i].len >= limit) {
+ matches[i].len = limit;
+
+ /* Truncation might produce multiple
+ * matches with length 'limit'. Keep at
+ * most 1. */
+ num_matches = i + 1;
+ }
+ } while (i--);
+ } else {
+ num_matches = 0;
+ }
+ cache_ptr->len = num_matches;
}
+
#if 0
fprintf(stderr, "Pos %u/%u: %u matches\n",
- ctx->match_window_pos, ctx->match_window_end, num_matches);
+ ctx->match_window_pos, ctx->window_size, num_matches);
for (unsigned i = 0; i < num_matches; i++)
fprintf(stderr, "\tLen %u Offset %u\n", matches[i].len, matches[i].offset);
#endif
#ifdef ENABLE_LZX_DEBUG
- for (u32 i = 0; i < num_matches; i++) {
+ for (unsigned i = 0; i < num_matches; i++) {
LZX_ASSERT(matches[i].len >= LZX_MIN_MATCH_LEN);
LZX_ASSERT(matches[i].len <= LZX_MAX_MATCH_LEN);
LZX_ASSERT(matches[i].len <= ctx->match_window_end - ctx->match_window_pos);
LZX_ASSERT(!memcmp(&ctx->window[ctx->match_window_pos],
&ctx->window[ctx->match_window_pos - matches[i].offset],
matches[i].len));
+ if (i) {
+ LZX_ASSERT(matches[i].len > matches[i - 1].len);
+ LZX_ASSERT(matches[i].offset > matches[i - 1].offset);
+ }
}
#endif
-
ctx->match_window_pos++;
+ ctx->cache_ptr = matches + num_matches;
+ *matches_ret = matches;
return num_matches;
}
-static u32
-lzx_get_prev_literal_cost(struct lzx_compressor *ctx,
- struct lzx_lru_queue *queue)
+static void
+lzx_skip_bytes(struct lzx_compressor *ctx, unsigned n)
{
- return lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
- &ctx->costs);
+ struct lz_match *cache_ptr;
+
+ LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos);
+
+ cache_ptr = ctx->cache_ptr;
+ ctx->match_window_pos += n;
+ if (ctx->matches_cached) {
+ while (n-- && cache_ptr <= ctx->cache_limit)
+ cache_ptr += 1 + cache_ptr->len;
+ } else {
+ lz_bt_skip_positions(&ctx->mf, n);
+ while (n-- && cache_ptr <= ctx->cache_limit) {
+ cache_ptr->len = 0;
+ cache_ptr += 1;
+ }
+ }
+ ctx->cache_ptr = cache_ptr;
}
-static u32
-lzx_get_match_cost(struct lzx_compressor *ctx,
- struct lzx_lru_queue *queue,
- input_idx_t length, input_idx_t offset)
+/*
+ * Reverse the linked list of near-optimal matches so that they can be returned
+ * in forwards order.
+ *
+ * Returns the first match in the list.
+ */
+static struct lz_match
+lzx_match_chooser_reverse_list(struct lzx_compressor *ctx, unsigned cur_pos)
{
- return lzx_match_cost(length, offset, &ctx->costs, queue);
+ unsigned prev_link, saved_prev_link;
+ unsigned prev_match_offset, saved_prev_match_offset;
+
+ ctx->optimum_end_idx = cur_pos;
+
+ saved_prev_link = ctx->optimum[cur_pos].prev.link;
+ saved_prev_match_offset = ctx->optimum[cur_pos].prev.match_offset;
+
+ do {
+ prev_link = saved_prev_link;
+ prev_match_offset = saved_prev_match_offset;
+
+ saved_prev_link = ctx->optimum[prev_link].prev.link;
+ saved_prev_match_offset = ctx->optimum[prev_link].prev.match_offset;
+
+ ctx->optimum[prev_link].next.link = cur_pos;
+ ctx->optimum[prev_link].next.match_offset = prev_match_offset;
+
+ cur_pos = prev_link;
+ } while (cur_pos != 0);
+
+ ctx->optimum_cur_idx = ctx->optimum[0].next.link;
+
+ return (struct lz_match)
+ { .len = ctx->optimum_cur_idx,
+ .offset = ctx->optimum[0].next.match_offset,
+ };
}
-static struct raw_match
-lzx_lz_get_near_optimal_match(struct lzx_compressor *ctx)
+/*
+ * lzx_get_near_optimal_match() -
+ *
+ * Choose an approximately optimal match or literal to use at the next position
+ * in the string, or "window", being LZ-encoded.
+ *
+ * This is based on algorithms used in 7-Zip, including the DEFLATE encoder
+ * and the LZMA encoder, written by Igor Pavlov.
+ *
+ * Unlike a greedy parser that always takes the longest match, or even a "lazy"
+ * parser with one match/literal look-ahead like zlib, the algorithm used here
+ * may look ahead many matches/literals to determine the approximately optimal
+ * match/literal to code next. The motivation is that the compression ratio is
+ * improved if the compressor can do things like use a shorter-than-possible
+ * match in order to allow a longer match later, and also take into account the
+ * estimated real cost of coding each match/literal based on the underlying
+ * entropy encoding.
+ *
+ * Still, this is not a true optimal parser for several reasons:
+ *
+ * - Real compression formats use entropy encoding of the literal/match
+ * sequence, so the real cost of coding each match or literal is unknown until
+ * the parse is fully determined. It can be approximated based on iterative
+ * parses, but the end result is not guaranteed to be globally optimal.
+ *
+ * - Very long matches are chosen immediately. This is because locations with
+ * long matches are likely to have many possible alternatives that would cause
+ * slow optimal parsing, but also such locations are already highly
+ * compressible so it is not too harmful to just grab the longest match.
+ *
+ * - Not all possible matches at each location are considered because the
+ * underlying match-finder limits the number and type of matches produced at
+ * each position. For example, for a given match length it's usually not
+ * worth it to only consider matches other than the lowest-offset match,
+ * except in the case of a repeat offset.
+ *
+ * - Although we take into account the adaptive state (in LZX, the recent offset
+ * queue), coding decisions made with respect to the adaptive state will be
+ * locally optimal but will not necessarily be globally optimal. This is
+ * because the algorithm only keeps the least-costly path to get to a given
+ * location and does not take into account that a slightly more costly path
+ * could result in a different adaptive state that ultimately results in a
+ * lower global cost.
+ *
+ * - The array space used by this function is bounded, so in degenerate cases it
+ * is forced to start returning matches/literals before the algorithm has
+ * really finished.
+ *
+ * Each call to this function does one of two things:
+ *
+ * 1. Build a sequence of near-optimal matches/literals, up to some point, that
+ * will be returned by subsequent calls to this function, then return the
+ * first one.
+ *
+ * OR
+ *
+ * 2. Return the next match/literal previously computed by a call to this
+ * function.
+ *
+ * The return value is a (length, offset) pair specifying the match or literal
+ * chosen. For literals, the length is 0 or 1 and the offset is meaningless.
+ */
+static struct lz_match
+lzx_get_near_optimal_match(struct lzx_compressor *ctx)
{
- return lz_get_near_optimal_match(&ctx->mc,
- lzx_lz_get_matches_caching,
- lzx_lz_skip_bytes,
- lzx_get_prev_literal_cost,
- lzx_get_match_cost,
- ctx,
- &ctx->queue);
+ unsigned num_matches;
+ const struct lz_match *matches;
+ struct lz_match match;
+ unsigned longest_len;
+ unsigned longest_rep_len;
+ u32 longest_rep_offset;
+ unsigned cur_pos;
+ unsigned end_pos;
+
+ if (ctx->optimum_cur_idx != ctx->optimum_end_idx) {
+ /* Case 2: Return the next match/literal already found. */
+ match.len = ctx->optimum[ctx->optimum_cur_idx].next.link -
+ ctx->optimum_cur_idx;
+ match.offset = ctx->optimum[ctx->optimum_cur_idx].next.match_offset;
+
+ ctx->optimum_cur_idx = ctx->optimum[ctx->optimum_cur_idx].next.link;
+ return match;
+ }
+
+ /* Case 1: Compute a new list of matches/literals to return. */
+
+ ctx->optimum_cur_idx = 0;
+ ctx->optimum_end_idx = 0;
+
+ /* Search for matches at recent offsets. Only keep the one with the
+ * longest match length. */
+ longest_rep_len = LZX_MIN_MATCH_LEN - 1;
+ if (ctx->match_window_pos >= 1) {
+ unsigned limit = min(LZX_MAX_MATCH_LEN,
+ ctx->match_window_end - ctx->match_window_pos);
+ for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
+ u32 offset = ctx->queue.R[i];
+ const u8 *strptr = &ctx->window[ctx->match_window_pos];
+ const u8 *matchptr = strptr - offset;
+ unsigned len = 0;
+ while (len < limit && strptr[len] == matchptr[len])
+ len++;
+ if (len > longest_rep_len) {
+ longest_rep_len = len;
+ longest_rep_offset = offset;
+ }
+ }
+ }
+
+ /* If there's a long match with a recent offset, take it. */
+ if (longest_rep_len >= ctx->params.alg_params.slow.nice_match_length) {
+ lzx_skip_bytes(ctx, longest_rep_len);
+ return (struct lz_match) {
+ .len = longest_rep_len,
+ .offset = longest_rep_offset,
+ };
+ }
+
+ /* Search other matches. */
+ num_matches = lzx_get_matches(ctx, &matches);
+
+ /* If there's a long match, take it. */
+ if (num_matches) {
+ longest_len = matches[num_matches - 1].len;
+ if (longest_len >= ctx->params.alg_params.slow.nice_match_length) {
+ lzx_skip_bytes(ctx, longest_len - 1);
+ return matches[num_matches - 1];
+ }
+ } else {
+ longest_len = 1;
+ }
+
+ /* Calculate the cost to reach the next position by coding a literal.
+ */
+ ctx->optimum[1].queue = ctx->queue;
+ ctx->optimum[1].cost = lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
+ &ctx->costs);
+ ctx->optimum[1].prev.link = 0;
+
+ /* Calculate the cost to reach any position up to and including that
+ * reached by the longest match.
+ *
+ * Note: We consider only the lowest-offset match that reaches each
+ * position.
+ *
+ * Note: Some of the cost calculation stays the same for each offset,
+ * regardless of how many lengths it gets used for. Therefore, to
+ * improve performance, we hand-code the cost calculation instead of
+ * calling lzx_match_cost() to do a from-scratch cost evaluation at each
+ * length. */
+ for (unsigned i = 0, len = 2; i < num_matches; i++) {
+ u32 offset;
+ struct lzx_lru_queue queue;
+ u32 position_cost;
+ unsigned position_slot;
+ unsigned num_extra_bits;
+
+ offset = matches[i].offset;
+ queue = ctx->queue;
+ position_cost = 0;
+
+ position_slot = lzx_get_position_slot(offset, &queue);
+ num_extra_bits = lzx_get_num_extra_bits(position_slot);
+ if (num_extra_bits >= 3) {
+ position_cost += num_extra_bits - 3;
+ position_cost += ctx->costs.aligned[(offset + LZX_OFFSET_OFFSET) & 7];
+ } else {
+ position_cost += num_extra_bits;
+ }
+
+ do {
+ unsigned len_header;
+ unsigned main_symbol;
+ u32 cost;
+
+ cost = position_cost;
+
+ len_header = min(len - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
+ main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
+ cost += ctx->costs.main[main_symbol];
+ if (len_header == LZX_NUM_PRIMARY_LENS)
+ cost += ctx->costs.len[len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS];
+
+ ctx->optimum[len].queue = queue;
+ ctx->optimum[len].prev.link = 0;
+ ctx->optimum[len].prev.match_offset = offset;
+ ctx->optimum[len].cost = cost;
+ } while (++len <= matches[i].len);
+ }
+ end_pos = longest_len;
+
+ if (longest_rep_len >= LZX_MIN_MATCH_LEN) {
+ struct lzx_lru_queue queue;
+ u32 cost;
+
+ while (end_pos < longest_rep_len)
+ ctx->optimum[++end_pos].cost = MC_INFINITE_COST;
+
+ queue = ctx->queue;
+ cost = lzx_match_cost(longest_rep_len, longest_rep_offset,
+ &ctx->costs, &queue);
+ if (cost <= ctx->optimum[longest_rep_len].cost) {
+ ctx->optimum[longest_rep_len].queue = queue;
+ ctx->optimum[longest_rep_len].prev.link = 0;
+ ctx->optimum[longest_rep_len].prev.match_offset = longest_rep_offset;
+ ctx->optimum[longest_rep_len].cost = cost;
+ }
+ }
+
+ /* Step forward, calculating the estimated minimum cost to reach each
+ * position. The algorithm may find multiple paths to reach each
+ * position; only the lowest-cost path is saved.
+ *
+ * The progress of the parse is tracked in the @ctx->optimum array, which
+ * for each position contains the minimum cost to reach that position,
+ * the index of the start of the match/literal taken to reach that
+ * position through the minimum-cost path, the offset of the match taken
+ * (not relevant for literals), and the adaptive state that will exist
+ * at that position after the minimum-cost path is taken. The @cur_pos
+ * variable stores the position at which the algorithm is currently
+ * considering coding choices, and the @end_pos variable stores the
+ * greatest position at which the costs of coding choices have been
+ * saved. (Actually, the algorithm guarantees that all positions up to
+ * and including @end_pos are reachable by at least one path.)
+ *
+ * The loop terminates when any one of the following conditions occurs:
+ *
+ * 1. A match with length greater than or equal to @nice_match_length is
+ * found. When this occurs, the algorithm chooses this match
+ * unconditionally, and consequently the near-optimal match/literal
+ * sequence up to and including that match is fully determined and it
+ * can begin returning the match/literal list.
+ *
+ * 2. @cur_pos reaches a position not overlapped by a preceding match.
+ * In such cases, the near-optimal match/literal sequence up to
+ * @cur_pos is fully determined and it can begin returning the
+ * match/literal list.
+ *
+ * 3. Failing either of the above in a degenerate case, the loop
+ * terminates when space in the @ctx->optimum array is exhausted.
+ * This terminates the algorithm and forces it to start returning
+ * matches/literals even though they may not be globally optimal.
+ *
+ * Upon loop termination, a nonempty list of matches/literals will have
+ * been produced and stored in the @optimum array. These
+ * matches/literals are linked in reverse order, so the last thing this
+ * function does is reverse this list and return the first
+ * match/literal, leaving the rest to be returned immediately by
+ * subsequent calls to this function.
+ */
+ cur_pos = 0;
+ for (;;) {
+ u32 cost;
+
+ /* Advance to next position. */
+ cur_pos++;
+
+ /* Check termination conditions (2) and (3) noted above. */
+ if (cur_pos == end_pos || cur_pos == LZX_OPTIM_ARRAY_SIZE)
+ return lzx_match_chooser_reverse_list(ctx, cur_pos);
+
+ /* Search for matches at recent offsets. */
+ longest_rep_len = LZX_MIN_MATCH_LEN - 1;
+ unsigned limit = min(LZX_MAX_MATCH_LEN,
+ ctx->match_window_end - ctx->match_window_pos);
+ for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
+ u32 offset = ctx->optimum[cur_pos].queue.R[i];
+ const u8 *strptr = &ctx->window[ctx->match_window_pos];
+ const u8 *matchptr = strptr - offset;
+ unsigned len = 0;
+ while (len < limit && strptr[len] == matchptr[len])
+ len++;
+ if (len > longest_rep_len) {
+ longest_rep_len = len;
+ longest_rep_offset = offset;
+ }
+ }
+
+ /* If we found a long match at a recent offset, choose it
+ * immediately. */
+ if (longest_rep_len >= ctx->params.alg_params.slow.nice_match_length) {
+ /* Build the list of matches to return and get
+ * the first one. */
+ match = lzx_match_chooser_reverse_list(ctx, cur_pos);
+
+ /* Append the long match to the end of the list. */
+ ctx->optimum[cur_pos].next.match_offset = longest_rep_offset;
+ ctx->optimum[cur_pos].next.link = cur_pos + longest_rep_len;
+ ctx->optimum_end_idx = cur_pos + longest_rep_len;
+
+ /* Skip over the remaining bytes of the long match. */
+ lzx_skip_bytes(ctx, longest_rep_len);
+
+ /* Return first match in the list. */
+ return match;
+ }
+
+ /* Search other matches. */
+ num_matches = lzx_get_matches(ctx, &matches);
+
+ /* If there's a long match, take it. */
+ if (num_matches) {
+ longest_len = matches[num_matches - 1].len;
+ if (longest_len >= ctx->params.alg_params.slow.nice_match_length) {
+ /* Build the list of matches to return and get
+ * the first one. */
+ match = lzx_match_chooser_reverse_list(ctx, cur_pos);
+
+ /* Append the long match to the end of the list. */
+ ctx->optimum[cur_pos].next.match_offset =
+ matches[num_matches - 1].offset;
+ ctx->optimum[cur_pos].next.link = cur_pos + longest_len;
+ ctx->optimum_end_idx = cur_pos + longest_len;
+
+ /* Skip over the remaining bytes of the long match. */
+ lzx_skip_bytes(ctx, longest_len - 1);
+
+ /* Return first match in the list. */
+ return match;
+ }
+ } else {
+ longest_len = 1;
+ }
+
+ while (end_pos < cur_pos + longest_len)
+ ctx->optimum[++end_pos].cost = MC_INFINITE_COST;
+
+ /* Consider coding a literal. */
+ cost = ctx->optimum[cur_pos].cost +
+ lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
+ &ctx->costs);
+ if (cost < ctx->optimum[cur_pos + 1].cost) {
+ ctx->optimum[cur_pos + 1].queue = ctx->optimum[cur_pos].queue;
+ ctx->optimum[cur_pos + 1].cost = cost;
+ ctx->optimum[cur_pos + 1].prev.link = cur_pos;
+ }
+
+ /* Consider coding a match.
+ *
+ * The hard-coded cost calculation is done for the same reason
+ * stated in the comment for the similar loop earlier.
+ * Actually, it is *this* one that has the biggest effect on
+ * performance; overall LZX compression is > 10% faster with
+ * this code compared to calling lzx_match_cost() with each
+ * length. */
+ for (unsigned i = 0, len = 2; i < num_matches; i++) {
+ u32 offset;
+ struct lzx_lru_queue queue;
+ u32 position_cost;
+ unsigned position_slot;
+ unsigned num_extra_bits;
+
+ offset = matches[i].offset;
+ queue = ctx->optimum[cur_pos].queue;
+ position_cost = ctx->optimum[cur_pos].cost;
+
+ position_slot = lzx_get_position_slot(offset, &queue);
+ num_extra_bits = lzx_get_num_extra_bits(position_slot);
+ if (num_extra_bits >= 3) {
+ position_cost += num_extra_bits - 3;
+ position_cost += ctx->costs.aligned[
+ (offset + LZX_OFFSET_OFFSET) & 7];
+ } else {
+ position_cost += num_extra_bits;
+ }
+
+ do {
+ unsigned len_header;
+ unsigned main_symbol;
+ u32 cost;
+
+ cost = position_cost;
+
+ len_header = min(len - LZX_MIN_MATCH_LEN,
+ LZX_NUM_PRIMARY_LENS);
+ main_symbol = ((position_slot << 3) | len_header) +
+ LZX_NUM_CHARS;
+ cost += ctx->costs.main[main_symbol];
+ if (len_header == LZX_NUM_PRIMARY_LENS) {
+ cost += ctx->costs.len[len -
+ LZX_MIN_MATCH_LEN -
+ LZX_NUM_PRIMARY_LENS];
+ }
+ if (cost < ctx->optimum[cur_pos + len].cost) {
+ ctx->optimum[cur_pos + len].queue = queue;
+ ctx->optimum[cur_pos + len].prev.link = cur_pos;
+ ctx->optimum[cur_pos + len].prev.match_offset = offset;
+ ctx->optimum[cur_pos + len].cost = cost;
+ }
+ } while (++len <= matches[i].len);
+ }
+
+ if (longest_rep_len >= LZX_MIN_MATCH_LEN) {
+ struct lzx_lru_queue queue;
+
+ while (end_pos < cur_pos + longest_rep_len)
+ ctx->optimum[++end_pos].cost = MC_INFINITE_COST;
+
+ queue = ctx->optimum[cur_pos].queue;
+
+ cost = ctx->optimum[cur_pos].cost +
+ lzx_match_cost(longest_rep_len, longest_rep_offset,
+ &ctx->costs, &queue);
+ if (cost <= ctx->optimum[cur_pos + longest_rep_len].cost) {
+ ctx->optimum[cur_pos + longest_rep_len].queue =
+ queue;
+ ctx->optimum[cur_pos + longest_rep_len].prev.link =
+ cur_pos;
+ ctx->optimum[cur_pos + longest_rep_len].prev.match_offset =
+ longest_rep_offset;
+ ctx->optimum[cur_pos + longest_rep_len].cost =
+ cost;
+ }
+ }
+ }
}
/* Set default symbol costs for the LZX Huffman codes. */
unsigned num_passes)
{
const struct lzx_lru_queue orig_queue = ctx->queue;
+ unsigned num_passes_remaining = num_passes;
struct lzx_freqs freqs;
+ const u8 *window_ptr;
+ const u8 *window_end;
+ struct lzx_item *next_chosen_match;
+ struct lz_match lz_match;
+ struct lzx_item lzx_item;
- unsigned orig_window_pos = spec->window_pos;
- unsigned orig_cached_pos = ctx->cached_matches_pos;
-
- LZX_ASSERT(ctx->match_window_pos == spec->window_pos);
+ LZX_ASSERT(num_passes >= 1);
+ LZX_ASSERT(lz_bt_get_position(&ctx->mf) == spec->window_pos);
ctx->match_window_end = spec->window_pos + spec->block_size;
- spec->chosen_matches_start_pos = spec->window_pos;
-
- LZX_ASSERT(num_passes >= 1);
+ ctx->matches_cached = false;
/* The first optimal parsing pass is done using the cost model already
* set in ctx->costs. Each later pass is done using a cost model
- * computed from the previous pass. */
- for (unsigned pass = 0; pass < num_passes; pass++) {
+ * computed from the previous pass.
+ *
+ * To improve performance we only generate the array containing the
+ * matches and literals in intermediate form on the final pass. */
- ctx->match_window_pos = orig_window_pos;
- ctx->cached_matches_pos = orig_cached_pos;
- ctx->queue = orig_queue;
- spec->num_chosen_matches = 0;
+ while (--num_passes_remaining) {
+ ctx->match_window_pos = spec->window_pos;
+ ctx->cache_ptr = ctx->cached_matches;
memset(&freqs, 0, sizeof(freqs));
+ window_ptr = &ctx->window[spec->window_pos];
+ window_end = window_ptr + spec->block_size;
- for (unsigned i = spec->window_pos; i < spec->window_pos + spec->block_size; ) {
- struct raw_match raw_match;
- struct lzx_match lzx_match;
-
- raw_match = lzx_lz_get_near_optimal_match(ctx);
- if (raw_match.len >= LZX_MIN_MATCH_LEN) {
- if (unlikely(raw_match.len == LZX_MIN_MATCH_LEN &&
- raw_match.offset == ctx->max_window_size -
- LZX_MIN_MATCH_LEN))
- {
- /* Degenerate case where the parser
- * generated the minimum match length
- * with the maximum offset. There
- * aren't actually enough position slots
- * to represent this offset, as noted in
- * the comments in
- * lzx_get_num_main_syms(), so we cannot
- * allow it. Use literals instead.
- *
- * Note that this case only occurs if
- * the match-finder can generate matches
- * to the very start of the window. The
- * suffix array match-finder can,
- * although typical hash chain and
- * binary tree match-finders use 0 as a
- * null value and therefore cannot
- * generate such matches. */
- BUILD_BUG_ON(LZX_MIN_MATCH_LEN != 2);
- lzx_match.data = lzx_tally_literal(ctx->window[i],
- &freqs);
- i += 1;
- ctx->chosen_matches[spec->chosen_matches_start_pos +
- spec->num_chosen_matches++]
- = lzx_match;
- lzx_match.data = lzx_tally_literal(ctx->window[i],
- &freqs);
- i += 1;
- } else {
- lzx_match.data = lzx_tally_match(raw_match.len,
- raw_match.offset,
- &freqs,
- &ctx->queue);
- i += raw_match.len;
- }
+ while (window_ptr != window_end) {
+
+ lz_match = lzx_get_near_optimal_match(ctx);
+
+ LZX_ASSERT(!(lz_match.len == LZX_MIN_MATCH_LEN &&
+ lz_match.offset == ctx->max_window_size -
+ LZX_MIN_MATCH_LEN));
+ if (lz_match.len >= LZX_MIN_MATCH_LEN) {
+ lzx_tally_match(lz_match.len, lz_match.offset,
+ &freqs, &ctx->queue);
+ window_ptr += lz_match.len;
} else {
- lzx_match.data = lzx_tally_literal(ctx->window[i], &freqs);
- i += 1;
+ lzx_tally_literal(*window_ptr, &freqs);
+ window_ptr += 1;
}
- ctx->chosen_matches[spec->chosen_matches_start_pos +
- spec->num_chosen_matches++] = lzx_match;
}
-
- lzx_make_huffman_codes(&freqs, &spec->codes,
- ctx->num_main_syms);
- if (pass < num_passes - 1)
- lzx_set_costs(ctx, &spec->codes.lens);
+ lzx_make_huffman_codes(&freqs, &spec->codes, ctx->num_main_syms);
+ lzx_set_costs(ctx, &spec->codes.lens);
+ ctx->queue = orig_queue;
ctx->matches_cached = true;
}
- spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
- ctx->matches_cached = false;
-}
-static void
-lzx_optimize_blocks(struct lzx_compressor *ctx)
-{
- lzx_lru_queue_init(&ctx->queue);
- lz_match_chooser_begin(&ctx->mc);
-
- const unsigned num_passes = ctx->params.alg_params.slow.num_optim_passes;
-
- for (unsigned i = 0; i < ctx->num_blocks; i++)
- lzx_optimize_block(ctx, &ctx->block_specs[i], num_passes);
+ ctx->match_window_pos = spec->window_pos;
+ ctx->cache_ptr = ctx->cached_matches;
+ memset(&freqs, 0, sizeof(freqs));
+ window_ptr = &ctx->window[spec->window_pos];
+ window_end = window_ptr + spec->block_size;
+
+ spec->chosen_items = &ctx->chosen_items[spec->window_pos];
+ next_chosen_match = spec->chosen_items;
+
+ while (window_ptr != window_end) {
+ lz_match = lzx_get_near_optimal_match(ctx);
+
+ LZX_ASSERT(!(lz_match.len == LZX_MIN_MATCH_LEN &&
+ lz_match.offset == ctx->max_window_size -
+ LZX_MIN_MATCH_LEN));
+ if (lz_match.len >= LZX_MIN_MATCH_LEN) {
+ lzx_item.data = lzx_tally_match(lz_match.len,
+ lz_match.offset,
+ &freqs, &ctx->queue);
+ window_ptr += lz_match.len;
+ } else {
+ lzx_item.data = lzx_tally_literal(*window_ptr, &freqs);
+ window_ptr += 1;
+ }
+ *next_chosen_match++ = lzx_item;
+ }
+ spec->num_chosen_items = next_chosen_match - spec->chosen_items;
+ lzx_make_huffman_codes(&freqs, &spec->codes, ctx->num_main_syms);
+ spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
}
/* Prepare the input window into one or more LZX blocks ready to be output. */
static void
lzx_prepare_blocks(struct lzx_compressor * ctx)
{
- /* Initialize the match-finder. */
- lz_sarray_load_window(&ctx->lz_sarray, ctx->window, ctx->window_size);
- ctx->cached_matches_pos = 0;
- ctx->matches_cached = false;
- ctx->match_window_pos = 0;
-
/* Set up a default cost model. */
lzx_set_default_costs(&ctx->costs, ctx->num_main_syms);
- /* TODO: The compression ratio could be slightly improved by performing
+ /* Set up the block specifications.
+ * TODO: The compression ratio could be slightly improved by performing
* data-dependent block splitting instead of using fixed-size blocks.
* Doing so well is a computationally hard problem, however. */
ctx->num_blocks = DIV_ROUND_UP(ctx->window_size, LZX_DIV_BLOCK_SIZE);
for (unsigned i = 0; i < ctx->num_blocks; i++) {
unsigned pos = LZX_DIV_BLOCK_SIZE * i;
ctx->block_specs[i].window_pos = pos;
- ctx->block_specs[i].block_size = min(ctx->window_size - pos, LZX_DIV_BLOCK_SIZE);
+ ctx->block_specs[i].block_size = min(ctx->window_size - pos,
+ LZX_DIV_BLOCK_SIZE);
}
+ /* Load the window into the match-finder. */
+ lz_bt_load_window(&ctx->mf, ctx->window, ctx->window_size);
+
/* Determine sequence of matches/literals to output for each block. */
- lzx_optimize_blocks(ctx);
+ lzx_lru_queue_init(&ctx->queue);
+ ctx->optimum_cur_idx = 0;
+ ctx->optimum_end_idx = 0;
+ for (unsigned i = 0; i < ctx->num_blocks; i++) {
+ lzx_optimize_block(ctx, &ctx->block_specs[i],
+ ctx->params.alg_params.slow.num_optim_passes);
+ }
}
/*
*
* ctx->block_specs[]
* ctx->num_blocks
- * ctx->chosen_matches[]
+ * ctx->chosen_items[]
*/
static void
lzx_prepare_block_fast(struct lzx_compressor * ctx)
/* Initialize symbol frequencies and match offset LRU queue. */
memset(&record_ctx.freqs, 0, sizeof(struct lzx_freqs));
lzx_lru_queue_init(&record_ctx.queue);
- record_ctx.matches = ctx->chosen_matches;
+ record_ctx.matches = ctx->chosen_items;
/* Determine series of matches/literals to output. */
lz_analyze_block(ctx->window,
spec->block_type = LZX_BLOCKTYPE_ALIGNED;
spec->window_pos = 0;
spec->block_size = ctx->window_size;
- spec->num_chosen_matches = (record_ctx.matches - ctx->chosen_matches);
- spec->chosen_matches_start_pos = 0;
+ spec->num_chosen_items = (record_ctx.matches - ctx->chosen_items);
+ spec->chosen_items = ctx->chosen_items;
lzx_make_huffman_codes(&record_ctx.freqs, &spec->codes,
ctx->num_main_syms);
ctx->num_blocks = 1;
}
-static void
-do_call_insn_translation(u32 *call_insn_target, int input_pos,
- s32 file_size)
-{
- s32 abs_offset;
- s32 rel_offset;
-
- rel_offset = le32_to_cpu(*call_insn_target);
- if (rel_offset >= -input_pos && rel_offset < file_size) {
- if (rel_offset < file_size - input_pos) {
- /* "good translation" */
- abs_offset = rel_offset + input_pos;
- } else {
- /* "compensating translation" */
- abs_offset = rel_offset - file_size;
- }
- *call_insn_target = cpu_to_le32(abs_offset);
- }
-}
-
-/* This is the reverse of undo_call_insn_preprocessing() in lzx-decompress.c.
- * See the comment above that function for more information. */
-static void
-do_call_insn_preprocessing(u8 data[], int size)
-{
- for (int i = 0; i < size - 10; i++) {
- if (data[i] == 0xe8) {
- do_call_insn_translation((u32*)&data[i + 1], i,
- LZX_WIM_MAGIC_FILESIZE);
- i += 4;
- }
- }
-}
-
static size_t
lzx_compress(const void *uncompressed_data, size_t uncompressed_size,
void *compressed_data, size_t compressed_size_avail, void *_ctx)
/* Before doing any actual compression, do the call instruction (0xe8
* byte) translation on the uncompressed data. */
- do_call_insn_preprocessing(ctx->window, ctx->window_size);
+ lzx_do_e8_preprocessing(ctx->window, ctx->window_size);
LZX_DEBUG("Preparing blocks...");
LZX_DEBUG("Flushing bitstream...");
compressed_size = flush_output_bitstream(&ostream);
- if (compressed_size == ~(input_idx_t)0) {
+ if (compressed_size == (u32)~0UL) {
LZX_DEBUG("Data did not compress to %zu bytes or less!",
compressed_size_avail);
return 0;
struct lzx_compressor *ctx = _ctx;
if (ctx) {
- FREE(ctx->chosen_matches);
+ FREE(ctx->chosen_items);
FREE(ctx->cached_matches);
- lz_match_chooser_destroy(&ctx->mc);
- lz_sarray_destroy(&ctx->lz_sarray);
+ FREE(ctx->optimum);
+ lz_bt_destroy(&ctx->mf);
FREE(ctx->block_specs);
FREE(ctx->prev_tab);
FREE(ctx->window);
.nice_match_length = 32,
.num_optim_passes = 2,
.max_search_depth = 50,
- .max_matches_per_pos = 3,
.main_nostat_cost = 15,
.len_nostat_cost = 15,
.aligned_nostat_cost = 7,
if (!lzx_window_size_valid(window_size))
return WIMLIB_ERR_INVALID_PARAM;
- LZX_DEBUG("Allocating memory.");
-
ctx = CALLOC(1, sizeof(struct lzx_compressor));
if (ctx == NULL)
goto oom;
if (!params->alg_params.slow.use_len2_matches)
min_match_len = max(min_match_len, 3);
- if (!lz_sarray_init(&ctx->lz_sarray,
- window_size,
- min_match_len,
- LZX_MAX_MATCH_LEN,
- params->alg_params.slow.max_search_depth,
- params->alg_params.slow.max_matches_per_pos))
+ if (!lz_bt_init(&ctx->mf,
+ window_size,
+ min_match_len,
+ LZX_MAX_MATCH_LEN,
+ params->alg_params.slow.nice_match_length,
+ params->alg_params.slow.max_search_depth))
goto oom;
}
if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
- if (!lz_match_chooser_init(&ctx->mc,
- LZX_OPTIM_ARRAY_SIZE,
- params->alg_params.slow.nice_match_length,
- LZX_MAX_MATCH_LEN))
+ ctx->optimum = MALLOC((LZX_OPTIM_ARRAY_SIZE +
+ min(params->alg_params.slow.nice_match_length,
+ LZX_MAX_MATCH_LEN)) *
+ sizeof(ctx->optimum[0]));
+ if (ctx->optimum == NULL)
goto oom;
}
if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
- u32 cache_per_pos;
-
- cache_per_pos = params->alg_params.slow.max_matches_per_pos;
- if (cache_per_pos > LZX_MAX_CACHE_PER_POS)
- cache_per_pos = LZX_MAX_CACHE_PER_POS;
-
- ctx->cached_matches = MALLOC(window_size * (cache_per_pos + 1) *
- sizeof(ctx->cached_matches[0]));
+ ctx->cached_matches = MALLOC(LZX_CACHE_SIZE);
if (ctx->cached_matches == NULL)
goto oom;
+ ctx->cache_limit = ctx->cached_matches +
+ LZX_CACHE_LEN - (LZX_MAX_MATCHES_PER_POS + 1);
}
- ctx->chosen_matches = MALLOC(window_size * sizeof(ctx->chosen_matches[0]));
- if (ctx->chosen_matches == NULL)
+ ctx->chosen_items = MALLOC(window_size * sizeof(ctx->chosen_items[0]));
+ if (ctx->chosen_items == NULL)
goto oom;
memcpy(&ctx->params, params, sizeof(struct wimlib_lzx_compressor_params));
sizeof(((struct lzx_compressor*)0)->block_specs[0]);
if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
- size += max_block_size * sizeof(((struct lzx_compressor*)0)->chosen_matches[0]);
- size += lz_sarray_get_needed_memory(max_block_size);
- size += lz_match_chooser_get_needed_memory(LZX_OPTIM_ARRAY_SIZE,
- params->alg_params.slow.nice_match_length,
- LZX_MAX_MATCH_LEN);
- u32 cache_per_pos;
-
- cache_per_pos = params->alg_params.slow.max_matches_per_pos;
- if (cache_per_pos > LZX_MAX_CACHE_PER_POS)
- cache_per_pos = LZX_MAX_CACHE_PER_POS;
-
- size += max_block_size * (cache_per_pos + 1) *
- sizeof(((struct lzx_compressor*)0)->cached_matches[0]);
+ size += max_block_size * sizeof(((struct lzx_compressor*)0)->chosen_items[0]);
+ size += lz_bt_get_needed_memory(max_block_size);
+ size += (LZX_OPTIM_ARRAY_SIZE +
+ min(params->alg_params.slow.nice_match_length,
+ LZX_MAX_MATCH_LEN)) *
+ sizeof(((struct lzx_compressor *)0)->optimum[0]);
+ size += LZX_CACHE_SIZE;
} else {
size += max_block_size * sizeof(((struct lzx_compressor*)0)->prev_tab[0]);
}