X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=src%2Flzx-compress.c;h=a972cc35bf3c24722372db9538c3b9a72f21f1c8;hp=cdbb63699004edbed7066f9344f6d1c6f30207d2;hb=7953f731d41d728a8881872bcf82fd8f9d1f7ee8;hpb=10be6c712f4f5f066458dbab11e81f8e64cd2324 diff --git a/src/lzx-compress.c b/src/lzx-compress.c index cdbb6369..a972cc35 100644 --- a/src/lzx-compress.c +++ b/src/lzx-compress.c @@ -1,5 +1,7 @@ /* * lzx-compress.c + * + * A compressor that produces output compatible with the LZX compression format. */ /* @@ -23,494 +25,377 @@ /* - * 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. - * - * ---------------------------------------------------------------------------- + * This file contains a compressor for the LZX ("Lempel-Ziv eXtended") + * compression format, as used in the WIM (Windows IMaging) file format. * - * Format Overview + * Two different parsing algorithms are implemented: "near-optimal" and "lazy". + * "Near-optimal" is significantly slower than "lazy", but results in a better + * compression ratio. The "near-optimal" algorithm is used at the default + * compression level. * - * 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. + * This file 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. * - * 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: + * Note: LZX is a compression format derived from 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 + * length) and an "offset slot" (giving, roughly speaking, the order of * magnitude of the match offset). * * - 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. - * - * ---------------------------------------------------------------------------- - * - * Algorithmic Overview - * - * At a high level, any implementation of LZX compression must operate as - * follows: - * - * 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.) - * - * 2. Find a sequence of LZ77-style matches and literal bytes that expands to - * the preprocessed data. - * - * 3. Divide the match/literal sequence into one or more LZX blocks, each of - * which may be "uncompressed", "verbatim", or "aligned". - * - * 4. Output each LZX block. - * - * 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. - * - * 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. - * - * 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. - * - * ---------------------------------------------------------------------------- - * - * Match-finding - * - * 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. + * - LZX has a minimum match length of 2 rather than 3. Length 2 matches can be + * useful, but generally only if the parser is smart about choosing them. * - * 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: - * - * - 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. - * - * - 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. - * - * - 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. - * - * ---------------------------------------------------------------------------- - * - * Match-choosing - * - * 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. - * - * 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. - * - * 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.) - * - * 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. - * - * 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. - * - * 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. - * - * 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. + * - In LZX, offset slots 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. */ #ifdef HAVE_CONFIG_H # include "config.h" #endif -#include "wimlib.h" -#include "wimlib/compressor_ops.h" #include "wimlib/compress_common.h" +#include "wimlib/compressor_ops.h" #include "wimlib/endianness.h" #include "wimlib/error.h" -#include "wimlib/lz.h" -#include "wimlib/lz_hash.h" -#include "wimlib/lz_bt.h" +#include "wimlib/lz_mf.h" +#include "wimlib/lz_repsearch.h" #include "wimlib/lzx.h" #include "wimlib/util.h" -#include -#ifdef ENABLE_LZX_DEBUG -# include "wimlib/decompress_common.h" -#endif +#include +#include -#define LZX_OPTIM_ARRAY_SIZE 4096 +#define LZX_OPTIM_ARRAY_LENGTH 4096 #define LZX_DIV_BLOCK_SIZE 32768 #define LZX_CACHE_PER_POS 8 +#define LZX_MAX_MATCHES_PER_POS (LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1) + #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_compressor; + +/* Codewords for the LZX Huffman codes. */ struct lzx_codewords { u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; u32 len[LZX_LENCODE_NUM_SYMBOLS]; u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* Codeword lengths (in bits) for the LZX main, length, and aligned offset - * Huffman codes. - * - * A 0 length means the codeword has zero frequency. - */ +/* Codeword lengths (in bits) for the LZX Huffman codes. + * A zero length means the corresponding codeword has zero frequency. */ struct lzx_lens { u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; u8 len[LZX_LENCODE_NUM_SYMBOLS]; u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* Costs for the LZX main, length, and aligned offset Huffman symbols. - * - * 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 many times. */ +/* Estimated cost, in bits, to output each symbol in the LZX Huffman codes. */ struct lzx_costs { u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; u8 len[LZX_LENCODE_NUM_SYMBOLS]; u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* The LZX main, length, and aligned offset Huffman codes */ +/* Codewords and lengths for the LZX Huffman codes. */ struct lzx_codes { struct lzx_codewords codewords; struct lzx_lens lens; }; -/* Tables for tallying symbol frequencies in the three LZX alphabets */ +/* Symbol frequency counters for the LZX Huffman codes. */ struct lzx_freqs { u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; u32 len[LZX_LENCODE_NUM_SYMBOLS]; u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* LZX intermediate match/literal format */ +/* Intermediate LZX match/literal format */ struct lzx_item { - /* Bit Description - * - * 31 1 if a match, 0 if a literal. - * - * 30-25 position slot. This can be at most 50, so it will fit in 6 - * bits. - * - * 8-24 position footer. This is the offset of the real formatted - * offset from the position base. This can be at most 17 bits - * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17). - * - * 0-7 length of match, minus 2. This can be at most - * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */ - u32 data; -}; -/* Specification for an LZX block. */ -struct lzx_block_spec { + /* Bits 0 - 9: Main symbol + * Bits 10 - 17: Length symbol + * Bits 18 - 22: Number of extra offset bits + * Bits 23+ : Extra offset bits */ + u64 data; +}; - /* One of the LZX_BLOCKTYPE_* constants indicating which type of this - * block. */ - int block_type; +/* Internal compression parameters */ +struct lzx_compressor_params { + u32 (*choose_items_for_block)(struct lzx_compressor *, u32, u32); + u32 num_optim_passes; + enum lz_mf_algo mf_algo; + u32 min_match_length; + u32 nice_match_length; + u32 max_search_depth; +}; - /* 0-based position in the window at which this block starts. */ - u32 window_pos; +/* + * Match chooser position data: + * + * An array of these structures is used during the near-optimal 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 number of bytes of uncompressed data this block represents. */ - u32 block_size; + /* The cost, in bits, of the lowest-cost path that has been found to + * reach this position. This can change as progressively lower cost + * paths are found to reach this position. */ + u32 cost; +#define MC_INFINITE_COST UINT32_MAX - /* The match/literal sequence for this block. */ - struct lzx_item *chosen_items; + /* The match or literal that was taken to reach this position. This can + * change as progressively lower cost paths are found to reach this + * position. + * + * This variable is divided into two bitfields. + * + * Literals: + * Low bits are 1, high bits are the literal. + * + * Explicit offset matches: + * Low bits are the match length, high bits are the offset plus 2. + * + * Repeat offset matches: + * Low bits are the match length, high bits are the queue index. + */ + u32 mc_item_data; +#define MC_OFFSET_SHIFT 9 +#define MC_LEN_MASK ((1 << MC_OFFSET_SHIFT) - 1) - /* The length of the @chosen_items sequence. */ - u32 num_chosen_items; + /* The state of the LZX recent match offsets queue at this position. + * This is filled in lazily, only after the minimum-cost path to this + * position is found. + * + * Note: the way we handle this adaptive state in the "minimum-cost" + * parse is actually only an approximation. It's possible for the + * globally optimal, minimum cost path to contain a prefix, ending at a + * position, where that path prefix is *not* the minimum cost path to + * that position. This can happen if such a path prefix results in a + * different adaptive state which results in lower costs later. We do + * not solve this problem; we only consider the lowest cost to reach + * each position, which seems to be an acceptable approximation. */ + struct lzx_lru_queue queue _aligned_attribute(16); + +} _aligned_attribute(16); + +/* State of the LZX compressor */ +struct lzx_compressor { - /* Huffman codes for this block. */ - struct lzx_codes codes; -}; + /* Internal compression parameters */ + struct lzx_compressor_params params; -/* State of the LZX compressor. */ -struct lzx_compressor { + /* The preprocessed buffer of data being compressed */ + u8 *cur_window; - /* The parameters that were used to create the compressor. */ - struct wimlib_lzx_compressor_params params; + /* Number of bytes of data to be compressed, which is the number of + * bytes of data in @cur_window that are actually valid. */ + u32 cur_window_size; - /* The buffer of data to be compressed. - * - * 0xe8 byte preprocessing is done directly on the data here before - * further compression. + /* log2 order of the LZX window size for LZ match offset encoding + * purposes. Will be >= LZX_MIN_WINDOW_ORDER and <= + * LZX_MAX_WINDOW_ORDER. * - * Note that this compressor does *not* use a real sliding window!!!! - * It's not needed in the WIM format, since every chunk is compressed - * independently. This is by design, to allow random access to the - * chunks. - * - * We reserve a few extra bytes to potentially allow reading off the end - * of the array in the match-finding code for optimization purposes - * (currently only needed for the hash chain match-finder). */ - u8 *window; + * Note: 1 << @window_order is normally equal to @max_window_size, + * a.k.a. the allocated size of @cur_window, but it will be greater than + * @max_window_size in the event that the compressor was created with a + * non-power-of-2 block size. (See lzx_get_window_order().) */ + unsigned window_order; + + /* Number of symbols in the main alphabet. This depends on + * @window_order, since @window_order determines the maximum possible + * offset. It does not, however, depend on the *actual* size of the + * current data buffer being processed, which might be less than 1 << + * @window_order. */ + unsigned num_main_syms; - /* Number of bytes of data to be compressed, which is the number of - * bytes of data in @window that are actually valid. */ - u32 window_size; + /* Lempel-Ziv match-finder */ + struct lz_mf *mf; - /* Allocated size of the @window. */ - u32 max_window_size; + /* Match-finder wrapper functions and data for near-optimal parsing. + * + * When doing more than one match-choosing pass over the data, matches + * found by the match-finder are cached 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 it is a very worthwhile speedup. */ + unsigned (*get_matches_func)(struct lzx_compressor *, const struct lz_match **); + void (*skip_bytes_func)(struct lzx_compressor *, unsigned n); + u32 match_window_pos; + u32 match_window_end; + struct lz_match *cached_matches; + struct lz_match *cache_ptr; + struct lz_match *cache_limit; - /* Number of symbols in the main alphabet (depends on the - * @max_window_size since it determines the maximum allowed offset). */ - unsigned num_main_syms; + /* Position data for near-optimal parsing. */ + struct lzx_mc_pos_data optimum[LZX_OPTIM_ARRAY_LENGTH + LZX_MAX_MATCH_LEN]; + + /* The cost model currently being used for near-optimal parsing. */ + struct lzx_costs costs; /* The current match offset LRU queue. */ struct lzx_lru_queue queue; - /* Space for the sequences of matches/literals that were chosen for each - * block. */ - struct lzx_item *chosen_items; + /* Frequency counters for the current block. */ + struct lzx_freqs freqs; + + /* The Huffman codes for the current and previous blocks. */ + struct lzx_codes codes[2]; - /* Information about the LZX blocks the preprocessed input was divided - * into. */ - struct lzx_block_spec *block_specs; + /* Which 'struct lzx_codes' is being used for the current block. The + * other was used for the previous block (if this isn't the first + * block). */ + unsigned int codes_index; - /* Number of LZX blocks the input was divided into; a.k.a. the number of - * elements of @block_specs that are valid. */ - unsigned num_blocks; + /* Dummy lengths that are always 0. */ + struct lzx_lens zero_lens; - /* This is simply filled in with zeroes and used to avoid special-casing - * the output of the first compressed Huffman code, which conceptually - * has a delta taken from a code with all symbols having zero-length - * codewords. */ - struct lzx_codes zero_codes; + /* Matches/literals that were chosen for the current block. */ + struct lzx_item chosen_items[LZX_DIV_BLOCK_SIZE]; - /* The current cost model. */ - struct lzx_costs costs; + /* Table mapping match offset => offset slot for small offsets */ +#define LZX_NUM_FAST_OFFSETS 32768 + u8 offset_slot_fast[LZX_NUM_FAST_OFFSETS]; +}; - /* Fast algorithm only: Array of hash table links. */ - u32 *prev_tab; +/* + * Structure to keep track of the current state of sending bits to the + * compressed output buffer. + * + * The LZX bitstream is encoded as a sequence of 16-bit coding units. + */ +struct lzx_output_bitstream { - /* Slow algorithm only: Binary tree match-finder. */ - struct lz_bt mf; + /* Bits that haven't yet been written to the output buffer. */ + u32 bitbuf; - /* Position in window of next match to return. */ - u32 match_window_pos; + /* Number of bits currently held in @bitbuf. */ + u32 bitcount; - /* The end-of-block position. We can't allow any matches to span this - * position. */ - u32 match_window_end; + /* Pointer to the start of the output buffer. */ + le16 *start; - /* 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 lz_match *cached_matches; - struct lz_match *cache_ptr; - bool matches_cached; - struct lz_match *cache_limit; + /* Pointer to the position in the output buffer at which the next coding + * unit should be written. */ + le16 *next; - /* 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; + /* Pointer past the end of the output buffer. */ + le16 *end; }; /* - * 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. + * Initialize the output bitstream. + * + * @os + * The output bitstream structure to initialize. + * @buffer + * The buffer being written to. + * @size + * Size of @buffer, in bytes. */ -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; -}; +static void +lzx_init_output(struct lzx_output_bitstream *os, void *buffer, u32 size) +{ + os->bitbuf = 0; + os->bitcount = 0; + os->start = buffer; + os->next = os->start; + os->end = os->start + size / sizeof(le16); +} -/* 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(u32 offset, struct lzx_lru_queue *queue) +/* + * Write some bits to the output bitstream. + * + * The bits are given by the low-order @num_bits bits of @bits. Higher-order + * bits in @bits cannot be set. At most 17 bits can be written at once. + * + * @max_num_bits is a compile-time constant that specifies the maximum number of + * bits that can ever be written at the call site. Currently, it is used to + * optimize away the conditional code for writing a second 16-bit coding unit + * when writing fewer than 17 bits. + * + * If the output buffer space is exhausted, then the bits will be ignored, and + * lzx_flush_output() will return 0 when it gets called. + */ +static inline void +lzx_write_varbits(struct lzx_output_bitstream *os, + const u32 bits, const unsigned int num_bits, + const unsigned int max_num_bits) { - unsigned position_slot; - - /* See if the offset was recently used. */ - for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) { - if (offset == queue->R[i]) { - /* Found it. */ - - /* Bring the repeat offset to the front of the - * queue. Note: this is, in fact, not a real - * LRU queue because repeat matches are simply - * swapped to the front. */ - swap(queue->R[0], queue->R[i]); - - /* The resulting position slot is simply the first index - * at which the offset was found in the queue. */ - return i; + /* This code is optimized for LZX, which never needs to write more than + * 17 bits at once. */ + LZX_ASSERT(num_bits <= 17); + LZX_ASSERT(num_bits <= max_num_bits); + LZX_ASSERT(os->bitcount <= 15); + + /* Add the bits to the bit buffer variable. @bitcount will be at most + * 15, so there will be just enough space for the maximum possible + * @num_bits of 17. */ + os->bitcount += num_bits; + os->bitbuf = (os->bitbuf << num_bits) | bits; + + /* Check whether any coding units need to be written. */ + if (os->bitcount >= 16) { + + os->bitcount -= 16; + + /* Write a coding unit, unless it would overflow the buffer. */ + if (os->next != os->end) + *os->next++ = cpu_to_le16(os->bitbuf >> os->bitcount); + + /* If writing 17 bits, a second coding unit might need to be + * written. But because 'max_num_bits' is a compile-time + * constant, the compiler will optimize away this code at most + * call sites. */ + if (max_num_bits == 17 && os->bitcount == 16) { + if (os->next != os->end) + *os->next++ = cpu_to_le16(os->bitbuf); + os->bitcount = 0; } } +} + +/* Use when @num_bits is a compile-time constant. Otherwise use + * lzx_write_varbits(). */ +static inline void +lzx_write_bits(struct lzx_output_bitstream *os, + const u32 bits, const unsigned int num_bits) +{ + lzx_write_varbits(os, bits, num_bits, num_bits); +} - /* The offset was not recently used; look up its real position slot. */ - position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET); +/* + * Flush the last coding unit to the output buffer if needed. Return the total + * number of bytes written to the output buffer, or 0 if an overflow occurred. + */ +static u32 +lzx_flush_output(struct lzx_output_bitstream *os) +{ + if (os->next == os->end) + return 0; - /* Bring the new offset to the front of the queue. */ - for (int i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--) - queue->R[i] = queue->R[i - 1]; - queue->R[0] = offset; + if (os->bitcount != 0) + *os->next++ = cpu_to_le16(os->bitbuf << (16 - os->bitcount)); - return position_slot; + return (const u8 *)os->next - (const u8 *)os->start; } /* Build the main, length, and aligned offset Huffman codes used in LZX. @@ -518,8 +403,7 @@ lzx_get_position_slot(u32 offset, struct lzx_lru_queue *queue) * This takes as input the frequency tables for each code and produces as output * a set of tables that map symbols to codewords and codeword lengths. */ static void -lzx_make_huffman_codes(const struct lzx_freqs *freqs, - struct lzx_codes *codes, +lzx_make_huffman_codes(const struct lzx_freqs *freqs, struct lzx_codes *codes, unsigned num_main_syms) { make_canonical_huffman_code(num_main_syms, @@ -541,240 +425,93 @@ lzx_make_huffman_codes(const struct lzx_freqs *freqs, codes->codewords.aligned); } -/* - * Output a precomputed LZX match. - * - * @out: - * The bitstream to which to write the match. - * @block_type: - * The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or - * LZX_BLOCKTYPE_VERBATIM) - * @match: - * The match, as a (length, offset) pair. - * @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_item match, const struct lzx_codes *codes) +static unsigned +lzx_compute_precode_items(const u8 lens[restrict], + const u8 prev_lens[restrict], + const unsigned num_lens, + u32 precode_freqs[restrict], + unsigned precode_items[restrict]) { - /* low 8 bits are the match length minus 2 */ - unsigned match_len_minus_2 = match.data & 0xff; - /* Next 17 bits are the position footer */ - unsigned position_footer = (match.data >> 8) & 0x1ffff; /* 17 bits */ - /* Next 6 bits are the position slot. */ - unsigned position_slot = (match.data >> 25) & 0x3f; /* 6 bits */ - unsigned len_header; - unsigned len_footer; - unsigned main_symbol; - unsigned num_extra_bits; - unsigned verbatim_bits; - unsigned aligned_bits; - - /* If the match length is less than MIN_MATCH_LEN (= 2) + - * NUM_PRIMARY_LENS (= 7), the length header contains - * the match length minus MIN_MATCH_LEN, and there is no - * length footer. - * - * Otherwise, the length header contains - * NUM_PRIMARY_LENS, and the length footer contains - * the match length minus NUM_PRIMARY_LENS minus - * MIN_MATCH_LEN. */ - if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) { - len_header = match_len_minus_2; - } else { - len_header = LZX_NUM_PRIMARY_LENS; - len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS; - } + unsigned *itemptr; + unsigned run_start; + unsigned run_end; + unsigned extra_bits; + int delta; + u8 len; + + itemptr = precode_items; + run_start = 0; + do { + /* Find the next run of codeword lengths. */ - /* Combine the position slot with the length header into a single symbol - * that will be encoded with the main code. - * - * The actual main symbol is offset by LZX_NUM_CHARS because values - * under LZX_NUM_CHARS are used to indicate a literal byte rather than a - * match. */ - main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS; - - /* Output main symbol. */ - bitstream_put_bits(out, codes->codewords.main[main_symbol], - codes->lens.main[main_symbol]); - - /* If there is a length footer, output it using the - * length Huffman code. */ - if (len_header == LZX_NUM_PRIMARY_LENS) - bitstream_put_bits(out, codes->codewords.len[len_footer], - codes->lens.len[len_footer]); - - num_extra_bits = lzx_get_num_extra_bits(position_slot); - - /* For aligned offset blocks with at least 3 extra bits, output the - * verbatim bits literally, then the aligned bits encoded using the - * aligned offset code. Otherwise, only the verbatim bits need to be - * output. */ - if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) { - - verbatim_bits = position_footer >> 3; - bitstream_put_bits(out, verbatim_bits, - num_extra_bits - 3); - - aligned_bits = (position_footer & 7); - bitstream_put_bits(out, - codes->codewords.aligned[aligned_bits], - codes->lens.aligned[aligned_bits]); - } else { - /* verbatim bits is the same as the position - * footer, in this case. */ - bitstream_put_bits(out, position_footer, num_extra_bits); - } -} + /* len = the length being repeated */ + len = lens[run_start]; -/* Output an LZX literal (encoded with the main Huffman code). */ -static void -lzx_write_literal(struct output_bitstream *out, u8 literal, - const struct lzx_codes *codes) -{ - bitstream_put_bits(out, - codes->codewords.main[literal], - codes->lens.main[literal]); -} + run_end = run_start + 1; -static unsigned -lzx_build_precode(const u8 lens[restrict], - const u8 prev_lens[restrict], - const unsigned num_syms, - 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], - unsigned *num_additional_bits_ret) -{ - memset(precode_freqs, 0, - LZX_PRECODE_NUM_SYMBOLS * sizeof(precode_freqs[0])); - - /* Since the code word lengths use a form of RLE encoding, the goal here - * is to find each run of identical lengths when going through them in - * symbol order (including runs of length 1). For each run, as many - * lengths are encoded using RLE as possible, and the rest are output - * literally. - * - * output_syms[] will be filled in with the length symbols that will be - * output, including RLE codes, not yet encoded using the precode. - * - * cur_run_len keeps track of how many code word lengths are in the - * current run of identical lengths. */ - unsigned output_syms_idx = 0; - unsigned cur_run_len = 1; - unsigned num_additional_bits = 0; - for (unsigned i = 1; i <= num_syms; i++) { - - if (i != num_syms && lens[i] == lens[i - 1]) { - /* Still in a run--- keep going. */ - cur_run_len++; + /* Fast case for a single length. */ + if (likely(run_end == num_lens || len != lens[run_end])) { + delta = prev_lens[run_start] - len; + if (delta < 0) + delta += 17; + precode_freqs[delta]++; + *itemptr++ = delta; + run_start++; continue; } - /* Run ended! Check if it is a run of zeroes or a run of - * nonzeroes. */ - - /* The symbol that was repeated in the run--- not to be confused - * with the length *of* the run (cur_run_len) */ - unsigned len_in_run = lens[i - 1]; - - if (len_in_run == 0) { - /* A run of 0's. Encode it in as few length - * codes as we can. */ + /* Extend the run. */ + do { + run_end++; + } while (run_end != num_lens && len == lens[run_end]); - /* The magic length 18 indicates a run of 20 + n zeroes, - * where n is an uncompressed literal 5-bit integer that - * follows the magic length. */ - while (cur_run_len >= 20) { - unsigned additional_bits; + if (len == 0) { + /* Run of zeroes. */ - additional_bits = min(cur_run_len - 20, 0x1f); - num_additional_bits += 5; + /* Symbol 18: RLE 20 to 51 zeroes at a time. */ + while ((run_end - run_start) >= 20) { + extra_bits = min((run_end - run_start) - 20, 0x1f); precode_freqs[18]++; - output_syms[output_syms_idx++] = 18; - output_syms[output_syms_idx++] = additional_bits; - cur_run_len -= 20 + additional_bits; + *itemptr++ = 18 | (extra_bits << 5); + run_start += 20 + extra_bits; } - /* The magic length 17 indicates a run of 4 + n zeroes, - * where n is an uncompressed literal 4-bit integer that - * follows the magic length. */ - while (cur_run_len >= 4) { - unsigned additional_bits; - - additional_bits = min(cur_run_len - 4, 0xf); - num_additional_bits += 4; + /* Symbol 17: RLE 4 to 19 zeroes at a time. */ + if ((run_end - run_start) >= 4) { + extra_bits = min((run_end - run_start) - 4, 0xf); precode_freqs[17]++; - output_syms[output_syms_idx++] = 17; - output_syms[output_syms_idx++] = additional_bits; - cur_run_len -= 4 + additional_bits; + *itemptr++ = 17 | (extra_bits << 5); + run_start += 4 + extra_bits; } - } else { /* A run of nonzero lengths. */ - /* The magic length 19 indicates a run of 4 + n - * nonzeroes, where n is a literal bit that follows the - * magic length, and where the value of the lengths in - * the run is given by an extra length symbol, encoded - * with the precode, that follows the literal bit. - * - * The extra length symbol is encoded as a difference - * from the length of the codeword for the first symbol - * in the run in the previous code. - * */ - while (cur_run_len >= 4) { - unsigned additional_bits; - signed char delta; - - additional_bits = (cur_run_len > 4); - num_additional_bits += 1; - delta = (signed char)prev_lens[i - cur_run_len] - - (signed char)len_in_run; + /* Symbol 19: RLE 4 to 5 of any length at a time. */ + while ((run_end - run_start) >= 4) { + extra_bits = (run_end - run_start) > 4; + delta = prev_lens[run_start] - len; if (delta < 0) delta += 17; precode_freqs[19]++; - precode_freqs[(unsigned char)delta]++; - output_syms[output_syms_idx++] = 19; - output_syms[output_syms_idx++] = additional_bits; - output_syms[output_syms_idx++] = delta; - cur_run_len -= 4 + additional_bits; + precode_freqs[delta]++; + *itemptr++ = 19 | (extra_bits << 5) | (delta << 6); + run_start += 4 + extra_bits; } } - /* Any remaining lengths in the run are outputted without RLE, - * as a difference from the length of that codeword in the - * previous code. */ - while (cur_run_len > 0) { - signed char delta; - - delta = (signed char)prev_lens[i - cur_run_len] - - (signed char)len_in_run; + /* Output any remaining lengths without RLE. */ + while (run_start != run_end) { + delta = prev_lens[run_start] - len; if (delta < 0) delta += 17; - - precode_freqs[(unsigned char)delta]++; - output_syms[output_syms_idx++] = delta; - cur_run_len--; + precode_freqs[delta]++; + *itemptr++ = delta; + run_start++; } + } while (run_start != num_lens); - cur_run_len = 1; - } - - /* Build the precode from the frequencies of the length symbols. */ - - make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, - LZX_MAX_PRE_CODEWORD_LEN, - precode_freqs, precode_lens, - precode_codewords); - - *num_additional_bits_ret = num_additional_bits; - - return output_syms_idx; + return itemptr - precode_items; } /* @@ -796,168 +533,175 @@ lzx_build_precode(const u8 lens[restrict], * as deltas from the codeword lengths of the corresponding code in the previous * block. * - * @out: + * @os: * Bitstream to which to write the compressed Huffman code. * @lens: * The codeword lengths, indexed by symbol, in the Huffman code. * @prev_lens: * The codeword lengths, indexed by symbol, in the corresponding Huffman * code in the previous block, or all zeroes if this is the first block. - * @num_syms: + * @num_lens: * The number of symbols in the Huffman code. */ static void -lzx_write_compressed_code(struct output_bitstream *out, +lzx_write_compressed_code(struct lzx_output_bitstream *os, const u8 lens[restrict], const u8 prev_lens[restrict], - unsigned num_syms) + unsigned num_lens) { 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]; + unsigned precode_items[num_lens]; + unsigned num_precode_items; + unsigned precode_item; + unsigned precode_sym; unsigned i; - unsigned num_output_syms; - u8 precode_sym; - unsigned dummy; - - num_output_syms = lzx_build_precode(lens, - prev_lens, - num_syms, - precode_freqs, - output_syms, - precode_lens, - precode_codewords, - &dummy); - - /* Write the lengths of the precode codes to the output. */ + for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) - bitstream_put_bits(out, precode_lens[i], - LZX_PRECODE_ELEMENT_SIZE); + precode_freqs[i] = 0; - /* Write the length symbols, encoded with the precode, to the output. */ + /* Compute the "items" (RLE / literal tokens and extra bits) with which + * the codeword lengths in the larger code will be output. */ + num_precode_items = lzx_compute_precode_items(lens, + prev_lens, + num_lens, + precode_freqs, + precode_items); - for (i = 0; i < num_output_syms; ) { - precode_sym = output_syms[i++]; + /* Build the precode. */ + make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, + LZX_MAX_PRE_CODEWORD_LEN, + precode_freqs, precode_lens, + precode_codewords); - bitstream_put_bits(out, precode_codewords[precode_sym], - precode_lens[precode_sym]); - switch (precode_sym) { - case 17: - bitstream_put_bits(out, output_syms[i++], 4); - break; - case 18: - bitstream_put_bits(out, output_syms[i++], 5); - break; - case 19: - bitstream_put_bits(out, output_syms[i++], 1); - bitstream_put_bits(out, - precode_codewords[output_syms[i]], - precode_lens[output_syms[i]]); - i++; - break; - default: - break; + /* Output the lengths of the codewords in the precode. */ + for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) + lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE); + + /* Output the encoded lengths of the codewords in the larger code. */ + for (i = 0; i < num_precode_items; i++) { + precode_item = precode_items[i]; + precode_sym = precode_item & 0x1F; + lzx_write_varbits(os, precode_codewords[precode_sym], + precode_lens[precode_sym], + LZX_MAX_PRE_CODEWORD_LEN); + if (precode_sym >= 17) { + if (precode_sym == 17) { + lzx_write_bits(os, precode_item >> 5, 4); + } else if (precode_sym == 18) { + lzx_write_bits(os, precode_item >> 5, 5); + } else { + lzx_write_bits(os, (precode_item >> 5) & 1, 1); + precode_sym = precode_item >> 6; + lzx_write_varbits(os, precode_codewords[precode_sym], + precode_lens[precode_sym], + LZX_MAX_PRE_CODEWORD_LEN); + } } } } +/* Output a match or literal. */ +static inline void +lzx_write_item(struct lzx_output_bitstream *os, struct lzx_item item, + unsigned ones_if_aligned, const struct lzx_codes *codes) +{ + u64 data = item.data; + unsigned main_symbol; + unsigned len_symbol; + unsigned num_extra_bits; + u32 extra_bits; + + main_symbol = data & 0x3FF; + + lzx_write_varbits(os, codes->codewords.main[main_symbol], + codes->lens.main[main_symbol], + LZX_MAX_MAIN_CODEWORD_LEN); + + if (main_symbol < LZX_NUM_CHARS) /* Literal? */ + return; + + len_symbol = (data >> 10) & 0xFF; + + if (len_symbol != LZX_LENCODE_NUM_SYMBOLS) { + lzx_write_varbits(os, codes->codewords.len[len_symbol], + codes->lens.len[len_symbol], + LZX_MAX_LEN_CODEWORD_LEN); + } + + num_extra_bits = (data >> 18) & 0x1F; + if (num_extra_bits == 0) /* Small offset or repeat offset match? */ + return; + + extra_bits = data >> 23; + + /*if (block_type == LZX_BLOCKTYPE_ALIGNED && num_extra_bits >= 3) {*/ + if ((num_extra_bits & ones_if_aligned) >= 3) { + + /* Aligned offset blocks: The low 3 bits of the extra offset + * bits are Huffman-encoded using the aligned offset code. The + * remaining bits are output literally. */ + + lzx_write_varbits(os, extra_bits >> 3, num_extra_bits - 3, 14); + + lzx_write_varbits(os, codes->codewords.aligned[extra_bits & 7], + codes->lens.aligned[extra_bits & 7], + LZX_MAX_ALIGNED_CODEWORD_LEN); + } else { + /* Verbatim blocks, or fewer than 3 extra bits: All extra + * offset bits are output literally. */ + lzx_write_varbits(os, extra_bits, num_extra_bits, 17); + } +} + /* * Write all matches and literal bytes (which were precomputed) in an LZX * compressed block to the output bitstream in the final compressed * representation. * - * @ostream + * @os * The output bitstream. * @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_item match_tab[], - unsigned match_count, - const struct lzx_codes *codes) -{ - for (unsigned i = 0; i < match_count; i++) { - struct lzx_item match = match_tab[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); - else - lzx_write_literal(ostream, match.data, codes); - } -} - -static void -lzx_assert_codes_valid(const struct lzx_codes * codes, unsigned num_main_syms) +lzx_write_items(struct lzx_output_bitstream *os, int block_type, + const struct lzx_item items[], u32 num_items, + const struct lzx_codes *codes) { -#ifdef ENABLE_LZX_DEBUG - unsigned i; - - for (i = 0; i < num_main_syms; i++) - LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_MAIN_CODEWORD_LEN); - - for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) - LZX_ASSERT(codes->lens.len[i] <= LZX_MAX_LEN_CODEWORD_LEN); + unsigned ones_if_aligned = 0U - (block_type == LZX_BLOCKTYPE_ALIGNED); - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) - LZX_ASSERT(codes->lens.aligned[i] <= LZX_MAX_ALIGNED_CODEWORD_LEN); - - const unsigned tablebits = 10; - u16 decode_table[(1 << tablebits) + - (2 * max(num_main_syms, LZX_LENCODE_NUM_SYMBOLS))] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); - LZX_ASSERT(0 == make_huffman_decode_table(decode_table, - num_main_syms, - min(tablebits, LZX_MAINCODE_TABLEBITS), - codes->lens.main, - LZX_MAX_MAIN_CODEWORD_LEN)); - LZX_ASSERT(0 == make_huffman_decode_table(decode_table, - LZX_LENCODE_NUM_SYMBOLS, - min(tablebits, LZX_LENCODE_TABLEBITS), - codes->lens.len, - LZX_MAX_LEN_CODEWORD_LEN)); - LZX_ASSERT(0 == make_huffman_decode_table(decode_table, - LZX_ALIGNEDCODE_NUM_SYMBOLS, - min(tablebits, LZX_ALIGNEDCODE_TABLEBITS), - codes->lens.aligned, - LZX_MAX_ALIGNED_CODEWORD_LEN)); -#endif /* ENABLE_LZX_DEBUG */ + for (u32 i = 0; i < num_items; i++) + lzx_write_item(os, items[i], ones_if_aligned, codes); } -/* Write an LZX aligned offset or verbatim block to the output. */ +/* Write an LZX aligned offset or verbatim block to the output bitstream. */ static void lzx_write_compressed_block(int block_type, - unsigned block_size, - unsigned max_window_size, + u32 block_size, + unsigned window_order, unsigned num_main_syms, struct lzx_item * chosen_items, - unsigned num_chosen_items, + u32 num_chosen_items, const struct lzx_codes * codes, - const struct lzx_codes * prev_codes, - struct output_bitstream * ostream) + const struct lzx_lens * prev_lens, + struct lzx_output_bitstream * os) { - unsigned i; - LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED || block_type == LZX_BLOCKTYPE_VERBATIM); - lzx_assert_codes_valid(codes, num_main_syms); /* The first three bits indicate the type of block and are one of the * LZX_BLOCKTYPE_* constants. */ - bitstream_put_bits(ostream, block_type, 3); + lzx_write_bits(os, block_type, 3); /* Output the block size. * @@ -975,237 +719,423 @@ lzx_write_compressed_block(int block_type, * because WIMs created with chunk size greater than 32768 can seemingly * only be opened by wimlib anyway. */ if (block_size == LZX_DEFAULT_BLOCK_SIZE) { - bitstream_put_bits(ostream, 1, 1); + lzx_write_bits(os, 1, 1); } else { - bitstream_put_bits(ostream, 0, 1); + lzx_write_bits(os, 0, 1); - if (max_window_size >= 65536) - bitstream_put_bits(ostream, block_size >> 16, 8); + if (window_order >= 16) + lzx_write_bits(os, block_size >> 16, 8); - bitstream_put_bits(ostream, block_size, 16); + lzx_write_bits(os, block_size & 0xFFFF, 16); } - /* Write out lengths of the main code. Note that the LZX specification - * incorrectly states that the aligned offset code comes after the - * length code, but in fact it is the very first code to be written - * (before the main code). */ - if (block_type == LZX_BLOCKTYPE_ALIGNED) - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) - bitstream_put_bits(ostream, codes->lens.aligned[i], - LZX_ALIGNEDCODE_ELEMENT_SIZE); - - LZX_DEBUG("Writing main code..."); - - /* Write the precode and lengths for the first LZX_NUM_CHARS symbols in - * the main code, which are the codewords for literal bytes. */ - lzx_write_compressed_code(ostream, - codes->lens.main, - prev_codes->lens.main, - LZX_NUM_CHARS); + /* If it's an aligned offset block, output the aligned offset code. */ + if (block_type == LZX_BLOCKTYPE_ALIGNED) { + for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + lzx_write_bits(os, codes->lens.aligned[i], + LZX_ALIGNEDCODE_ELEMENT_SIZE); + } + } - /* Write the precode and lengths for the rest of the main code, which - * are the codewords for match headers. */ - lzx_write_compressed_code(ostream, - codes->lens.main + LZX_NUM_CHARS, - prev_codes->lens.main + LZX_NUM_CHARS, + /* Output the main code (two parts). */ + lzx_write_compressed_code(os, codes->lens.main, + prev_lens->main, + LZX_NUM_CHARS); + lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS, + prev_lens->main + LZX_NUM_CHARS, num_main_syms - LZX_NUM_CHARS); - LZX_DEBUG("Writing length code..."); - - /* Write the precode and lengths for the length code. */ - lzx_write_compressed_code(ostream, - codes->lens.len, - prev_codes->lens.len, + /* Output the length code. */ + lzx_write_compressed_code(os, codes->lens.len, + prev_lens->len, LZX_LENCODE_NUM_SYMBOLS); - LZX_DEBUG("Writing matches and literals..."); - - /* Write the actual matches and literals. */ - lzx_write_matches_and_literals(ostream, block_type, - chosen_items, num_chosen_items, - codes); - - LZX_DEBUG("Done writing block."); + /* Output the compressed matches and literals. */ + lzx_write_items(os, block_type, chosen_items, num_chosen_items, codes); } -/* Write out the LZX blocks that were computed. */ -static void -lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream) +/* Don't allow matches to span the end of an LZX block. */ +static inline unsigned +maybe_truncate_matches(struct lz_match matches[], unsigned num_matches, + struct lzx_compressor *c) { + if (c->match_window_end < c->cur_window_size && num_matches != 0) { + u32 limit = c->match_window_end - c->match_window_pos; + + if (limit >= LZX_MIN_MATCH_LEN) { + + unsigned i = num_matches - 1; + do { + if (matches[i].len >= limit) { + matches[i].len = limit; - const struct lzx_codes *prev_codes = &ctx->zero_codes; - 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_items=%u)...", - i + 1, ctx->num_blocks, - spec->block_type, spec->block_size, - spec->num_chosen_items); - - lzx_write_compressed_block(spec->block_type, - spec->block_size, - ctx->max_window_size, - ctx->num_main_syms, - spec->chosen_items, - spec->num_chosen_items, - &spec->codes, - prev_codes, - ostream); - - prev_codes = &spec->codes; + /* Truncation might produce multiple + * matches with length 'limit'. Keep at + * most 1. */ + num_matches = i + 1; + } + } while (i--); + } else { + num_matches = 0; + } } + return num_matches; } -/* Constructs an LZX match from a literal byte and updates the main code symbol - * frequencies. */ -static inline u32 -lzx_tally_literal(u8 lit, struct lzx_freqs *freqs) +static unsigned +lzx_get_matches_fillcache_singleblock(struct lzx_compressor *c, + const struct lz_match **matches_ret) { - freqs->main[lit]++; - return (u32)lit; + struct lz_match *cache_ptr; + struct lz_match *matches; + unsigned num_matches; + + cache_ptr = c->cache_ptr; + matches = cache_ptr + 1; + if (likely(cache_ptr <= c->cache_limit)) { + num_matches = lz_mf_get_matches(c->mf, matches); + cache_ptr->len = num_matches; + c->cache_ptr = matches + num_matches; + } else { + num_matches = 0; + } + c->match_window_pos++; + *matches_ret = matches; + return num_matches; } -/* Constructs an LZX match from an offset and a length, and updates the LRU - * 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 inline u32 -lzx_tally_match(unsigned match_len, u32 match_offset, - struct lzx_freqs *freqs, struct lzx_lru_queue *queue) +static unsigned +lzx_get_matches_fillcache_multiblock(struct lzx_compressor *c, + const struct lz_match **matches_ret) { - unsigned position_slot; - unsigned position_footer; - u32 len_header; - unsigned main_symbol; - unsigned len_footer; - unsigned adjusted_match_len; - - LZX_ASSERT(match_len >= LZX_MIN_MATCH_LEN && match_len <= LZX_MAX_MATCH_LEN); - - /* The match offset shall be encoded as a position slot (itself encoded - * as part of the main symbol) and a position footer. */ - position_slot = lzx_get_position_slot(match_offset, queue); - position_footer = (match_offset + LZX_OFFSET_OFFSET) & - ((1U << lzx_get_num_extra_bits(position_slot)) - 1); - - /* The match length shall be encoded as a length header (itself encoded - * as part of the main symbol) and an optional length footer. */ - adjusted_match_len = match_len - LZX_MIN_MATCH_LEN; - if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) { - /* No length footer needed. */ - len_header = adjusted_match_len; + struct lz_match *cache_ptr; + struct lz_match *matches; + unsigned num_matches; + + cache_ptr = c->cache_ptr; + matches = cache_ptr + 1; + if (likely(cache_ptr <= c->cache_limit)) { + num_matches = lz_mf_get_matches(c->mf, matches); + num_matches = maybe_truncate_matches(matches, num_matches, c); + cache_ptr->len = num_matches; + c->cache_ptr = matches + num_matches; } else { - /* Length footer needed. It will be encoded using the length - * code. */ - len_header = LZX_NUM_PRIMARY_LENS; - len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS; - freqs->len[len_footer]++; + num_matches = 0; } + c->match_window_pos++; + *matches_ret = matches; + return num_matches; +} - /* Account for the main symbol. */ - main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS; +static unsigned +lzx_get_matches_usecache(struct lzx_compressor *c, + const struct lz_match **matches_ret) +{ + struct lz_match *cache_ptr; + struct lz_match *matches; + unsigned num_matches; - freqs->main[main_symbol]++; + cache_ptr = c->cache_ptr; + matches = cache_ptr + 1; + if (cache_ptr <= c->cache_limit) { + num_matches = cache_ptr->len; + c->cache_ptr = matches + num_matches; + } else { + num_matches = 0; + } + c->match_window_pos++; + *matches_ret = matches; + return num_matches; +} - /* In an aligned offset block, 3 bits of the position footer are output - * as an aligned offset symbol. Account for this, although we may - * ultimately decide to output the block as verbatim. */ +static unsigned +lzx_get_matches_usecache_nocheck(struct lzx_compressor *c, + const struct lz_match **matches_ret) +{ + struct lz_match *cache_ptr; + struct lz_match *matches; + unsigned num_matches; - /* The following check is equivalent to: - * - * if (lzx_extra_bits[position_slot] >= 3) - * - * Note that this correctly excludes position slots that correspond to - * recent offsets. */ - if (position_slot >= 8) - freqs->aligned[position_footer & 7]++; + cache_ptr = c->cache_ptr; + matches = cache_ptr + 1; + num_matches = cache_ptr->len; + c->cache_ptr = matches + num_matches; + c->match_window_pos++; + *matches_ret = matches; + return num_matches; +} - /* Pack the position slot, position footer, and match length into an - * 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); - LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256); - - LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1); - LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1); - LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1); - return 0x80000000 | - (position_slot << 25) | - (position_footer << 8) | - (adjusted_match_len); +static unsigned +lzx_get_matches_nocache_singleblock(struct lzx_compressor *c, + const struct lz_match **matches_ret) +{ + struct lz_match *matches; + unsigned num_matches; + + matches = c->cache_ptr; + num_matches = lz_mf_get_matches(c->mf, matches); + c->match_window_pos++; + *matches_ret = matches; + return num_matches; } -struct lzx_record_ctx { - struct lzx_freqs freqs; - struct lzx_lru_queue queue; - struct lzx_item *matches; -}; +static unsigned +lzx_get_matches_nocache_multiblock(struct lzx_compressor *c, + const struct lz_match **matches_ret) +{ + struct lz_match *matches; + unsigned num_matches; + + matches = c->cache_ptr; + num_matches = lz_mf_get_matches(c->mf, matches); + num_matches = maybe_truncate_matches(matches, num_matches, c); + c->match_window_pos++; + *matches_ret = matches; + return num_matches; +} + +/* + * Find matches at the next position in the window. + * + * This uses a wrapper function around the underlying match-finder. + * + * Returns the number of matches found and sets *matches_ret to point to the + * matches array. The matches will be sorted by strictly increasing length and + * offset. + */ +static inline unsigned +lzx_get_matches(struct lzx_compressor *c, const struct lz_match **matches_ret) +{ + return (*c->get_matches_func)(c, matches_ret); +} static void -lzx_record_match(unsigned len, unsigned offset, void *_ctx) +lzx_skip_bytes_fillcache(struct lzx_compressor *c, unsigned n) { - struct lzx_record_ctx *ctx = _ctx; + struct lz_match *cache_ptr; - (ctx->matches++)->data = lzx_tally_match(len, offset, &ctx->freqs, &ctx->queue); + cache_ptr = c->cache_ptr; + c->match_window_pos += n; + lz_mf_skip_positions(c->mf, n); + if (cache_ptr <= c->cache_limit) { + do { + cache_ptr->len = 0; + cache_ptr += 1; + } while (--n && cache_ptr <= c->cache_limit); + } + c->cache_ptr = cache_ptr; } static void -lzx_record_literal(u8 lit, void *_ctx) +lzx_skip_bytes_usecache(struct lzx_compressor *c, unsigned n) { - struct lzx_record_ctx *ctx = _ctx; + struct lz_match *cache_ptr; - (ctx->matches++)->data = lzx_tally_literal(lit, &ctx->freqs); + cache_ptr = c->cache_ptr; + c->match_window_pos += n; + if (cache_ptr <= c->cache_limit) { + do { + cache_ptr += 1 + cache_ptr->len; + } while (--n && cache_ptr <= c->cache_limit); + } + c->cache_ptr = cache_ptr; } -/* Returns the cost, in bits, to output a literal byte using the specified cost - * model. */ -static u32 -lzx_literal_cost(u8 c, const struct lzx_costs * costs) +static void +lzx_skip_bytes_usecache_nocheck(struct lzx_compressor *c, unsigned n) { - return costs->main[c]; + struct lz_match *cache_ptr; + + cache_ptr = c->cache_ptr; + c->match_window_pos += n; + do { + cache_ptr += 1 + cache_ptr->len; + } while (--n); + c->cache_ptr = cache_ptr; } -/* Given a (length, offset) pair that could be turned into a valid LZX match as - * 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 also updates it. */ -static u32 -lzx_match_cost(unsigned length, u32 offset, const struct lzx_costs *costs, - struct lzx_lru_queue *queue) +static void +lzx_skip_bytes_nocache(struct lzx_compressor *c, unsigned n) { - unsigned position_slot; - unsigned len_header, main_symbol; - unsigned num_extra_bits; - u32 cost = 0; + c->match_window_pos += n; + lz_mf_skip_positions(c->mf, n); +} - position_slot = lzx_get_position_slot(offset, queue); +/* + * Skip the specified number of positions in the window (don't search for + * matches at them). + * + * This uses a wrapper function around the underlying match-finder. + */ +static inline void +lzx_skip_bytes(struct lzx_compressor *c, unsigned n) +{ + return (*c->skip_bytes_func)(c, n); +} - len_header = min(length - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS); - main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS; +/* Tally, and optionally record, the specified literal byte. */ +static inline void +lzx_declare_literal(struct lzx_compressor *c, unsigned literal, + struct lzx_item **next_chosen_item) +{ + unsigned main_symbol = literal; - /* Account for main symbol. */ - cost += costs->main[main_symbol]; + c->freqs.main[main_symbol]++; + + if (next_chosen_item) { + *(*next_chosen_item)++ = (struct lzx_item) { + .data = main_symbol, + }; + } +} + +/* Tally, and optionally record, the specified repeat offset match. */ +static inline void +lzx_declare_repeat_offset_match(struct lzx_compressor *c, + unsigned len, unsigned rep_index, + struct lzx_item **next_chosen_item) +{ + unsigned len_header; + unsigned main_symbol; + unsigned len_symbol; - /* Account for extra position information. */ - 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]; + if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) { + len_header = len - LZX_MIN_MATCH_LEN; + len_symbol = LZX_LENCODE_NUM_SYMBOLS; } else { - cost += num_extra_bits; + len_header = LZX_NUM_PRIMARY_LENS; + len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS; + c->freqs.len[len_symbol]++; } - /* Account for extra length information. */ - if (len_header == LZX_NUM_PRIMARY_LENS) - cost += costs->len[length - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS]; + main_symbol = LZX_NUM_CHARS + ((rep_index << 3) | len_header); - return cost; + c->freqs.main[main_symbol]++; + + if (next_chosen_item) { + *(*next_chosen_item)++ = (struct lzx_item) { + .data = (u64)main_symbol | ((u64)len_symbol << 10), + }; + } +} + +/* Tally, and optionally record, the specified explicit offset match. */ +static inline void +lzx_declare_explicit_offset_match(struct lzx_compressor *c, unsigned len, u32 offset, + struct lzx_item **next_chosen_item) +{ + unsigned len_header; + unsigned main_symbol; + unsigned len_symbol; + unsigned offset_slot; + unsigned num_extra_bits; + u32 extra_bits; + + if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) { + len_header = len - LZX_MIN_MATCH_LEN; + len_symbol = LZX_LENCODE_NUM_SYMBOLS; + } else { + len_header = LZX_NUM_PRIMARY_LENS; + len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS; + c->freqs.len[len_symbol]++; + } + + offset_slot = lzx_get_offset_slot_raw(offset + LZX_OFFSET_OFFSET); + + main_symbol = LZX_NUM_CHARS + ((offset_slot << 3) | len_header); + + c->freqs.main[main_symbol]++; + + if (offset_slot >= 8) + c->freqs.aligned[(offset + LZX_OFFSET_OFFSET) & 7]++; + + if (next_chosen_item) { + + num_extra_bits = lzx_extra_offset_bits[offset_slot]; + + extra_bits = (offset + LZX_OFFSET_OFFSET) - + lzx_offset_slot_base[offset_slot]; + + *(*next_chosen_item)++ = (struct lzx_item) { + .data = (u64)main_symbol | + ((u64)len_symbol << 10) | + ((u64)num_extra_bits << 18) | + ((u64)extra_bits << 23), + }; + } +} + +/* Tally, and optionally record, the specified match or literal. */ +static inline void +lzx_declare_item(struct lzx_compressor *c, u32 mc_item_data, + struct lzx_item **next_chosen_item) +{ + u32 len = mc_item_data & MC_LEN_MASK; + u32 offset_data = mc_item_data >> MC_OFFSET_SHIFT; + + if (len == 1) + lzx_declare_literal(c, offset_data, next_chosen_item); + else if (offset_data < LZX_NUM_RECENT_OFFSETS) + lzx_declare_repeat_offset_match(c, len, offset_data, + next_chosen_item); + else + lzx_declare_explicit_offset_match(c, len, + offset_data - LZX_OFFSET_OFFSET, + next_chosen_item); +} + +static inline void +lzx_record_item_list(struct lzx_compressor *c, + struct lzx_mc_pos_data *cur_optimum_ptr, + struct lzx_item **next_chosen_item) +{ + struct lzx_mc_pos_data *end_optimum_ptr; + u32 saved_item; + u32 item; + + /* The list is currently in reverse order (last item to first item). + * Reverse it. */ + end_optimum_ptr = cur_optimum_ptr; + saved_item = cur_optimum_ptr->mc_item_data; + do { + item = saved_item; + cur_optimum_ptr -= item & MC_LEN_MASK; + saved_item = cur_optimum_ptr->mc_item_data; + cur_optimum_ptr->mc_item_data = item; + } while (cur_optimum_ptr != c->optimum); + + /* Walk the list of items from beginning to end, tallying and recording + * each item. */ + do { + lzx_declare_item(c, cur_optimum_ptr->mc_item_data, next_chosen_item); + cur_optimum_ptr += (cur_optimum_ptr->mc_item_data) & MC_LEN_MASK; + } while (cur_optimum_ptr != end_optimum_ptr); +} +static inline void +lzx_tally_item_list(struct lzx_compressor *c, struct lzx_mc_pos_data *cur_optimum_ptr) +{ + /* Since we're just tallying the items, we don't need to reverse the + * list. Processing the items in reverse order is fine. */ + do { + lzx_declare_item(c, cur_optimum_ptr->mc_item_data, NULL); + cur_optimum_ptr -= (cur_optimum_ptr->mc_item_data & MC_LEN_MASK); + } while (cur_optimum_ptr != c->optimum); } -/* Set the cost model @ctx->costs from the Huffman codeword lengths specified in +/* Tally, and optionally (if next_chosen_item != NULL) record, in order, all + * items in the current list of items found by the match-chooser. */ +static void +lzx_declare_item_list(struct lzx_compressor *c, struct lzx_mc_pos_data *cur_optimum_ptr, + struct lzx_item **next_chosen_item) +{ + if (next_chosen_item) + lzx_record_item_list(c, cur_optimum_ptr, next_chosen_item); + else + lzx_tally_item_list(c, cur_optimum_ptr); +} + +/* Set the cost model @c->costs from the Huffman codeword lengths specified in * @lens. * * The cost model and codeword lengths are almost the same thing, but the @@ -1215,610 +1145,742 @@ lzx_match_cost(unsigned length, u32 offset, const struct lzx_costs *costs, * length) to take into account the fact that uses of these symbols are expected * to be rare. */ static void -lzx_set_costs(struct lzx_compressor * ctx, const struct lzx_lens * lens) +lzx_set_costs(struct lzx_compressor *c, const struct lzx_lens * lens) { unsigned i; - unsigned num_main_syms = ctx->num_main_syms; /* Main code */ - for (i = 0; i < num_main_syms; i++) { - ctx->costs.main[i] = lens->main[i]; - if (ctx->costs.main[i] == 0) - ctx->costs.main[i] = ctx->params.alg_params.slow.main_nostat_cost; - } + for (i = 0; i < c->num_main_syms; i++) + c->costs.main[i] = lens->main[i] ? lens->main[i] : 15; /* Length code */ - for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) { - ctx->costs.len[i] = lens->len[i]; - if (ctx->costs.len[i] == 0) - ctx->costs.len[i] = ctx->params.alg_params.slow.len_nostat_cost; - } + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) + c->costs.len[i] = lens->len[i] ? lens->len[i] : 15; /* Aligned offset code */ - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { - ctx->costs.aligned[i] = lens->aligned[i]; - if (ctx->costs.aligned[i] == 0) - ctx->costs.aligned[i] = ctx->params.alg_params.slow.aligned_nostat_cost; - } + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) + c->costs.aligned[i] = lens->aligned[i] ? lens->aligned[i] : 7; } -/* 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 unsigned -lzx_get_matches(struct lzx_compressor *ctx, - const struct lz_match **matches_ret) +/* Set default LZX Huffman symbol costs to bootstrap the iterative optimization + * algorithm. */ +static void +lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms) { - struct lz_match *cache_ptr; - struct lz_match *matches; - unsigned num_matches; + unsigned i; - LZX_ASSERT(ctx->match_window_pos < ctx->match_window_end); + /* Main code (part 1): Literal symbols */ + for (i = 0; i < LZX_NUM_CHARS; i++) + costs->main[i] = 8; - 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; - } + /* Main code (part 2): Match header symbols */ + for (; i < num_main_syms; i++) + costs->main[i] = 10; + + /* Length code */ + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) + costs->len[i] = 8; + + /* Aligned offset code */ + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) + costs->aligned[i] = 3; +} + +/* Return the cost, in bits, to output a literal byte using the specified cost + * model. */ +static inline u32 +lzx_literal_cost(unsigned literal, const struct lzx_costs * costs) +{ + return costs->main[literal]; +} + +/* Return the cost, in bits, to output a match of the specified length and + * offset slot using the specified cost model. Does not take into account + * extra offset bits. */ +static inline u32 +lzx_match_cost_raw(unsigned len, unsigned offset_slot, + const struct lzx_costs *costs) +{ + u32 cost; + unsigned len_header; + unsigned main_symbol; + + if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) { + len_header = len - LZX_MIN_MATCH_LEN; + cost = 0; } else { - num_matches = 0; + len_header = LZX_NUM_PRIMARY_LENS; + + /* Account for length symbol. */ + cost = costs->len[len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS]; } - /* 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; + /* Account for main symbol. */ + main_symbol = LZX_NUM_CHARS + ((offset_slot << 3) | len_header); + cost += costs->main[main_symbol]; - if (limit >= LZX_MIN_MATCH_LEN) { + return cost; +} - unsigned i = num_matches - 1; - do { - if (matches[i].len >= limit) { - matches[i].len = limit; +/* Equivalent to lzx_match_cost_raw(), but assumes the length is small enough + * that it doesn't require a length symbol. */ +static inline u32 +lzx_match_cost_raw_smalllen(unsigned len, unsigned offset_slot, + const struct lzx_costs *costs) +{ + LZX_ASSERT(len < LZX_MIN_MATCH_LEN + LZX_NUM_PRIMARY_LENS); + return costs->main[LZX_NUM_CHARS + + ((offset_slot << 3) | (len - LZX_MIN_MATCH_LEN))]; +} - /* 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; +/* + * Consider coding the match at repeat offset index @rep_idx. Consider each + * length from the minimum (2) to the full match length (@rep_len). + */ +static inline void +lzx_consider_repeat_offset_match(struct lzx_compressor *c, + struct lzx_mc_pos_data *cur_optimum_ptr, + unsigned rep_len, unsigned rep_idx) +{ + u32 base_cost = cur_optimum_ptr->cost; + u32 cost; + unsigned len; + +#if 1 /* Optimized version */ + + if (rep_len < LZX_MIN_MATCH_LEN + LZX_NUM_PRIMARY_LENS) { + /* All lengths being considered are small. */ + len = 2; + do { + cost = base_cost + + lzx_match_cost_raw_smalllen(len, rep_idx, &c->costs); + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->mc_item_data = + (rep_idx << MC_OFFSET_SHIFT) | len; + (cur_optimum_ptr + len)->cost = cost; + } + } while (++len <= rep_len); + } else { + /* Some lengths being considered are small, and some are big. + * Start with the optimized loop for small lengths, then switch + * to the optimized loop for big lengths. */ + len = 2; + do { + cost = base_cost + + lzx_match_cost_raw_smalllen(len, rep_idx, &c->costs); + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->mc_item_data = + (rep_idx << MC_OFFSET_SHIFT) | len; + (cur_optimum_ptr + len)->cost = cost; + } + } while (++len < LZX_MIN_MATCH_LEN + LZX_NUM_PRIMARY_LENS); + + /* The main symbol is now fixed. */ + base_cost += c->costs.main[LZX_NUM_CHARS + + ((rep_idx << 3) | LZX_NUM_PRIMARY_LENS)]; + do { + cost = base_cost + + c->costs.len[len - LZX_MIN_MATCH_LEN - + LZX_NUM_PRIMARY_LENS]; + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->mc_item_data = + (rep_idx << MC_OFFSET_SHIFT) | len; + (cur_optimum_ptr + len)->cost = cost; + } + } while (++len <= rep_len); } -#if 0 - fprintf(stderr, "Pos %u/%u: %u matches\n", - 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 +#else /* Unoptimized version */ -#ifdef ENABLE_LZX_DEBUG - 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(matches[i].offset > 0); - LZX_ASSERT(matches[i].offset <= 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); + len = 2; + do { + cost = base_cost + + lzx_match_cost_raw(len, rep_idx, &c->costs); + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->mc_item_data = + (rep_idx << MC_OFFSET_SHIFT) | len; + (cur_optimum_ptr + len)->cost = cost; } - } + } while (++len <= rep_len); #endif - ctx->match_window_pos++; - ctx->cache_ptr = matches + num_matches; - *matches_ret = matches; - return num_matches; } -static void -lzx_skip_bytes(struct lzx_compressor *ctx, unsigned n) +/* + * Consider coding each match in @matches as an explicit offset match. + * + * @matches must be sorted by strictly increasing length and strictly + * increasing offset. This is guaranteed by the match-finder. + * + * We consider each length from the minimum (2) to the longest + * (matches[num_matches - 1].len). For each length, we consider only + * the smallest offset for which that length is available. Although + * this is not guaranteed to be optimal due to the possibility of a + * larger offset costing less than a smaller offset to code, this is a + * very useful heuristic. + */ +static inline void +lzx_consider_explicit_offset_matches(struct lzx_compressor *c, + struct lzx_mc_pos_data *cur_optimum_ptr, + const struct lz_match matches[], + unsigned num_matches) { - struct lz_match *cache_ptr; + LZX_ASSERT(num_matches > 0); - LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos); + unsigned i; + unsigned len; + unsigned offset_slot; + u32 position_cost; + u32 cost; + u32 offset_data; - 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; + +#if 1 /* Optimized version */ + + if (matches[num_matches - 1].offset < LZX_NUM_FAST_OFFSETS) { + + /* + * Offset is small; the offset slot can be looked up directly in + * c->offset_slot_fast. + * + * Additional optimizations: + * + * - Since the offset is small, it falls in the exponential part + * of the offset slot bases and the number of extra offset + * bits can be calculated directly as (offset_slot >> 1) - 1. + * + * - Just consider the number of extra offset bits; don't + * account for the aligned offset code. Usually this has + * almost no effect on the compression ratio. + * + * - Start out in a loop optimized for small lengths. When the + * length becomes high enough that a length symbol will be + * needed, jump into a loop optimized for big lengths. + */ + + LZX_ASSERT(offset_slot <= 37); /* for extra bits formula */ + + len = 2; + i = 0; + do { + offset_slot = c->offset_slot_fast[matches[i].offset]; + position_cost = cur_optimum_ptr->cost + + ((offset_slot >> 1) - 1); + offset_data = matches[i].offset + LZX_OFFSET_OFFSET; + do { + if (len >= LZX_MIN_MATCH_LEN + LZX_NUM_PRIMARY_LENS) + goto biglen; + cost = position_cost + + lzx_match_cost_raw_smalllen(len, offset_slot, + &c->costs); + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->cost = cost; + (cur_optimum_ptr + len)->mc_item_data = + (offset_data << MC_OFFSET_SHIFT) | len; + } + } while (++len <= matches[i].len); + } while (++i != num_matches); + + return; + + do { + offset_slot = c->offset_slot_fast[matches[i].offset]; + biglen: + position_cost = cur_optimum_ptr->cost + + ((offset_slot >> 1) - 1) + + c->costs.main[LZX_NUM_CHARS + + ((offset_slot << 3) | + LZX_NUM_PRIMARY_LENS)]; + offset_data = matches[i].offset + LZX_OFFSET_OFFSET; + do { + cost = position_cost + + c->costs.len[len - LZX_MIN_MATCH_LEN - + LZX_NUM_PRIMARY_LENS]; + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->cost = cost; + (cur_optimum_ptr + len)->mc_item_data = + (offset_data << MC_OFFSET_SHIFT) | len; + } + } while (++len <= matches[i].len); + } while (++i != num_matches); } else { - lz_bt_skip_positions(&ctx->mf, n); - while (n-- && cache_ptr <= ctx->cache_limit) { - cache_ptr->len = 0; - cache_ptr += 1; - } + len = 2; + i = 0; + do { + offset_data = matches[i].offset + LZX_OFFSET_OFFSET; + offset_slot = lzx_get_offset_slot_raw(offset_data); + position_cost = cur_optimum_ptr->cost + + lzx_extra_offset_bits[offset_slot]; + do { + cost = position_cost + + lzx_match_cost_raw(len, offset_slot, &c->costs); + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->cost = cost; + (cur_optimum_ptr + len)->mc_item_data = + (offset_data << MC_OFFSET_SHIFT) | len; + } + } while (++len <= matches[i].len); + } while (++i != num_matches); } - ctx->cache_ptr = cache_ptr; + +#else /* Unoptimized version */ + + unsigned num_extra_bits; + + len = 2; + i = 0; + do { + offset_data = matches[i].offset + LZX_OFFSET_OFFSET; + position_cost = cur_optimum_ptr->cost; + offset_slot = lzx_get_offset_slot_raw(offset_data); + num_extra_bits = lzx_extra_offset_bits[offset_slot]; + if (num_extra_bits >= 3) { + position_cost += num_extra_bits - 3; + position_cost += c->costs.aligned[offset_data & 7]; + } else { + position_cost += num_extra_bits; + } + do { + cost = position_cost + + lzx_match_cost_raw(len, offset_slot, &c->costs); + if (cost < (cur_optimum_ptr + len)->cost) { + (cur_optimum_ptr + len)->cost = cost; + (cur_optimum_ptr + len)->mc_item_data = + (offset_data << MC_OFFSET_SHIFT) | len; + } + } while (++len <= matches[i].len); + } while (++i != num_matches); +#endif } /* - * Reverse the linked list of near-optimal matches so that they can be returned - * in forwards order. - * - * Returns the first match in the list. + * Search for repeat offset matches with the current position. */ -static struct lz_match -lzx_match_chooser_reverse_list(struct lzx_compressor *ctx, unsigned cur_pos) +static inline unsigned +lzx_repsearch(const u8 * const strptr, const u32 bytes_remaining, + const struct lzx_lru_queue *queue, unsigned *rep_max_idx_ret) { - unsigned prev_link, saved_prev_link; - unsigned prev_match_offset, saved_prev_match_offset; + BUILD_BUG_ON(LZX_NUM_RECENT_OFFSETS != 3); + return lz_repsearch3(strptr, min(bytes_remaining, LZX_MAX_MATCH_LEN), + queue->R, rep_max_idx_ret); +} + +/* + * The main near-optimal parsing routine. + * + * Briefly, the algorithm does an approximate minimum-cost path search to find a + * "near-optimal" sequence of matches and literals to output, based on the + * current cost model. The algorithm steps forward, position by position (byte + * by byte), and updates the minimum cost path to reach each later position that + * can be reached using a match or literal from the current position. This is + * essentially Dijkstra's algorithm in disguise: the graph nodes are positions, + * the graph edges are possible matches/literals to code, and the cost of each + * edge is the estimated number of bits that will be required to output the + * corresponding match or literal. But one difference is that we actually + * compute the lowest-cost path in pieces, where each piece is terminated when + * there are no choices to be made. + * + * This function will run this algorithm on the portion of the window from + * &c->cur_window[c->match_window_pos] to &c->cur_window[c->match_window_end]. + * + * On entry, c->queue must be the current state of the match offset LRU queue, + * and c->costs must be the current cost model to use for Huffman symbols. + * + * On exit, c->queue will be the state that the LRU queue would be in if the + * chosen items were to be coded. + * + * If next_chosen_item != NULL, then all items chosen will be recorded (saved in + * the chosen_items array). Otherwise, all items chosen will only be tallied + * (symbol frequencies tallied in c->freqs). + */ +static void +lzx_optim_pass(struct lzx_compressor *c, struct lzx_item **next_chosen_item) +{ + const u8 *block_end; + struct lzx_lru_queue *begin_queue; + const u8 *window_ptr; + struct lzx_mc_pos_data *cur_optimum_ptr; + struct lzx_mc_pos_data *end_optimum_ptr; + const struct lz_match *matches; + unsigned num_matches; + unsigned longest_len; + unsigned rep_max_len; + unsigned rep_max_idx; + unsigned literal; + unsigned len; + u32 cost; + u32 offset_data; - ctx->optimum_end_idx = cur_pos; + block_end = &c->cur_window[c->match_window_end]; + begin_queue = &c->queue; +begin: + /* Start building a new list of items, which will correspond to the next + * piece of the overall minimum-cost path. + * + * *begin_queue is the current state of the match offset LRU queue. */ - saved_prev_link = ctx->optimum[cur_pos].prev.link; - saved_prev_match_offset = ctx->optimum[cur_pos].prev.match_offset; + window_ptr = &c->cur_window[c->match_window_pos]; - do { - prev_link = saved_prev_link; - prev_match_offset = saved_prev_match_offset; + if (window_ptr == block_end) { + c->queue = *begin_queue; + return; + } + + cur_optimum_ptr = c->optimum; + cur_optimum_ptr->cost = 0; + cur_optimum_ptr->queue = *begin_queue; + + end_optimum_ptr = cur_optimum_ptr; + + /* The following loop runs once for each per byte in the window, except + * in a couple shortcut cases. */ + for (;;) { + + /* Find explicit offset matches with the current position. */ + num_matches = lzx_get_matches(c, &matches); + + if (num_matches) { + /* + * Find the longest repeat offset match with the current + * position. + * + * Heuristics: + * + * - Only search for repeat offset matches if the + * match-finder already found at least one match. + * + * - Only consider the longest repeat offset match. It + * seems to be rare for the optimal parse to include a + * repeat offset match that doesn't have the longest + * length (allowing for the possibility that not all + * of that length is actually used). + */ + rep_max_len = lzx_repsearch(window_ptr, + block_end - window_ptr, + &cur_optimum_ptr->queue, + &rep_max_idx); + + if (rep_max_len) { + /* If there's a very long repeat offset match, + * choose it immediately. */ + if (rep_max_len >= c->params.nice_match_length) { + + swap(cur_optimum_ptr->queue.R[0], + cur_optimum_ptr->queue.R[rep_max_idx]); + begin_queue = &cur_optimum_ptr->queue; + + cur_optimum_ptr += rep_max_len; + cur_optimum_ptr->mc_item_data = + (rep_max_idx << MC_OFFSET_SHIFT) | + rep_max_len; + + lzx_skip_bytes(c, rep_max_len - 1); + break; + } + + /* If reaching any positions for the first time, + * initialize their costs to "infinity". */ + while (end_optimum_ptr < cur_optimum_ptr + rep_max_len) + (++end_optimum_ptr)->cost = MC_INFINITE_COST; + + /* Consider coding a repeat offset match. */ + lzx_consider_repeat_offset_match(c, + cur_optimum_ptr, + rep_max_len, + rep_max_idx); + } - saved_prev_link = ctx->optimum[prev_link].prev.link; - saved_prev_match_offset = ctx->optimum[prev_link].prev.match_offset; + longest_len = matches[num_matches - 1].len; - ctx->optimum[prev_link].next.link = cur_pos; - ctx->optimum[prev_link].next.match_offset = prev_match_offset; + /* If there's a very long explicit offset match, choose + * it immediately. */ + if (longest_len >= c->params.nice_match_length) { - cur_pos = prev_link; - } while (cur_pos != 0); + cur_optimum_ptr->queue.R[2] = + cur_optimum_ptr->queue.R[1]; + cur_optimum_ptr->queue.R[1] = + cur_optimum_ptr->queue.R[0]; + cur_optimum_ptr->queue.R[0] = + matches[num_matches - 1].offset; + begin_queue = &cur_optimum_ptr->queue; - ctx->optimum_cur_idx = ctx->optimum[0].next.link; + offset_data = matches[num_matches - 1].offset + + LZX_OFFSET_OFFSET; + cur_optimum_ptr += longest_len; + cur_optimum_ptr->mc_item_data = + (offset_data << MC_OFFSET_SHIFT) | + longest_len; - return (struct lz_match) - { .len = ctx->optimum_cur_idx, - .offset = ctx->optimum[0].next.match_offset, - }; -} + lzx_skip_bytes(c, longest_len - 1); + break; + } -/* - * 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) -{ - 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; - } + /* If reaching any positions for the first time, + * initialize their costs to "infinity". */ + while (end_optimum_ptr < cur_optimum_ptr + longest_len) + (++end_optimum_ptr)->cost = MC_INFINITE_COST; - /* 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; + /* Consider coding an explicit offset match. */ + lzx_consider_explicit_offset_matches(c, cur_optimum_ptr, + matches, num_matches); + } else { + /* No matches found. The only choice at this position + * is to code a literal. */ + + if (end_optimum_ptr == cur_optimum_ptr) { + #if 1 + /* Optimization for single literals. */ + if (likely(cur_optimum_ptr == c->optimum)) { + lzx_declare_literal(c, *window_ptr++, + next_chosen_item); + if (window_ptr == block_end) { + c->queue = cur_optimum_ptr->queue; + return; + } + continue; + } + #endif + (++end_optimum_ptr)->cost = MC_INFINITE_COST; } } - } - /* 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, - }; - } + /* Consider coding a literal. - /* Search other matches. */ - num_matches = lzx_get_matches(ctx, &matches); + * To avoid an extra unpredictable brench, actually checking the + * preferability of coding a literal is integrated into the + * queue update code below. */ + literal = *window_ptr++; + cost = cur_optimum_ptr->cost + lzx_literal_cost(literal, &c->costs); - /* 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; - } + /* Advance to the next position. */ + cur_optimum_ptr++; - /* 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; + /* The lowest-cost path to the current position is now known. + * Finalize the recent offsets queue that results from taking + * this lowest-cost path. */ - /* 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]; + if (cost < cur_optimum_ptr->cost) { + /* Literal: queue remains unchanged. */ + cur_optimum_ptr->cost = cost; + cur_optimum_ptr->mc_item_data = (literal << MC_OFFSET_SHIFT) | 1; + cur_optimum_ptr->queue = (cur_optimum_ptr - 1)->queue; } else { - position_cost += num_extra_bits; + /* Match: queue update is needed. */ + len = cur_optimum_ptr->mc_item_data & MC_LEN_MASK; + offset_data = cur_optimum_ptr->mc_item_data >> MC_OFFSET_SHIFT; + if (offset_data >= LZX_NUM_RECENT_OFFSETS) { + /* Explicit offset match: offset is inserted at front */ + cur_optimum_ptr->queue.R[0] = offset_data - LZX_OFFSET_OFFSET; + cur_optimum_ptr->queue.R[1] = (cur_optimum_ptr - len)->queue.R[0]; + cur_optimum_ptr->queue.R[2] = (cur_optimum_ptr - len)->queue.R[1]; + } else { + /* Repeat offset match: offset is swapped to front */ + cur_optimum_ptr->queue = (cur_optimum_ptr - len)->queue; + swap(cur_optimum_ptr->queue.R[0], + cur_optimum_ptr->queue.R[offset_data]); + } } - 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; + /* + * This loop will terminate when either of the following + * conditions is true: + * + * (1) cur_optimum_ptr == end_optimum_ptr + * + * There are no paths that extend beyond the current + * position. In this case, any path to a later position + * must pass through the current position, so we can go + * ahead and choose the list of items that led to this + * position. + * + * (2) cur_optimum_ptr == &c->optimum[LZX_OPTIM_ARRAY_LENGTH] + * + * This bounds the number of times the algorithm can step + * forward before it is guaranteed to start choosing items. + * This limits the memory usage. But + * LZX_OPTIM_ARRAY_LENGTH is high enough that on most + * inputs this limit is never reached. + * + * Note: no check for end-of-block is needed because + * end-of-block will trigger condition (1). + */ + if (cur_optimum_ptr == end_optimum_ptr || + cur_optimum_ptr == &c->optimum[LZX_OPTIM_ARRAY_LENGTH]) + { + begin_queue = &cur_optimum_ptr->queue; + break; } } - /* 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); + /* Choose the current list of items that constitute the minimum-cost + * path to the current position. */ + lzx_declare_item_list(c, cur_optimum_ptr, next_chosen_item); + goto begin; +} - /* 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; +/* Fast heuristic scoring for lazy parsing: how "good" is this match? */ +static inline unsigned +lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset) +{ + unsigned score = len; - /* Skip over the remaining bytes of the long match. */ - lzx_skip_bytes(ctx, longest_rep_len); + if (adjusted_offset < 2048) + score++; - /* Return first match in the list. */ - return match; - } + if (adjusted_offset < 1024) + score++; - /* Search other matches. */ - num_matches = lzx_get_matches(ctx, &matches); + return score; +} - /* 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); +static inline unsigned +lzx_repeat_offset_match_score(unsigned len, unsigned slot) +{ + return len + 3; +} - /* 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; +/* Lazy parsing */ +static u32 +lzx_choose_lazy_items_for_block(struct lzx_compressor *c, + u32 block_start_pos, u32 block_size) +{ + const u8 *window_ptr; + const u8 *block_end; + struct lz_mf *mf; + struct lz_match *matches; + unsigned num_matches; + unsigned cur_len; + u32 cur_offset_data; + unsigned cur_score; + unsigned rep_max_len; + unsigned rep_max_idx; + unsigned rep_score; + unsigned prev_len; + unsigned prev_score; + u32 prev_offset_data; + unsigned skip_len; + struct lzx_item *next_chosen_item; + + window_ptr = &c->cur_window[block_start_pos]; + block_end = window_ptr + block_size; + matches = c->cached_matches; + mf = c->mf; + next_chosen_item = c->chosen_items; + + prev_len = 0; + prev_offset_data = 0; + prev_score = 0; + + while (window_ptr != block_end) { + + /* Find explicit offset matches with the current position. */ + num_matches = lz_mf_get_matches(mf, matches); + window_ptr++; + + if (num_matches == 0 || + (matches[num_matches - 1].len == 3 && + matches[num_matches - 1].offset >= 8192 - LZX_OFFSET_OFFSET && + matches[num_matches - 1].offset != c->queue.R[0] && + matches[num_matches - 1].offset != c->queue.R[1] && + matches[num_matches - 1].offset != c->queue.R[2])) + { + /* No match found, or the only match found was a distant + * length 3 match. Output the previous match if there + * is one; otherwise output a literal. */ - /* Skip over the remaining bytes of the long match. */ - lzx_skip_bytes(ctx, longest_len - 1); + no_match_found: - /* Return first match in the list. */ - return match; + if (prev_len) { + skip_len = prev_len - 2; + goto output_prev_match; + } else { + lzx_declare_literal(c, *(window_ptr - 1), + &next_chosen_item); + continue; } - } 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; + /* Find the longest repeat offset match with the current + * position. */ + if (likely(block_end - (window_ptr - 1) >= 2)) { + rep_max_len = lzx_repsearch((window_ptr - 1), + block_end - (window_ptr - 1), + &c->queue, &rep_max_idx); + } else { + rep_max_len = 0; } - /* 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; - } + cur_len = matches[num_matches - 1].len; + cur_offset_data = matches[num_matches - 1].offset + LZX_OFFSET_OFFSET; + cur_score = lzx_explicit_offset_match_score(cur_len, cur_offset_data); - 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); + /* Select the better of the explicit and repeat offset matches. */ + if (rep_max_len >= 3 && + (rep_score = lzx_repeat_offset_match_score(rep_max_len, + rep_max_idx)) >= cur_score) + { + cur_len = rep_max_len; + cur_offset_data = rep_max_idx; + cur_score = rep_score; } - 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; - } + if (unlikely(cur_len > block_end - (window_ptr - 1))) { + /* Nearing end of block. */ + cur_len = block_end - (window_ptr - 1); + if (cur_len < 3) + goto no_match_found; } - } -} -/* Set default symbol costs for the LZX Huffman codes. */ -static void -lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms) -{ - unsigned i; + if (prev_len == 0 || cur_score > prev_score) { + /* No previous match, or the current match is better + * than the previous match. + * + * If there's a previous match, then output a literal in + * its place. + * + * In both cases, if the current match is very long, + * then output it immediately. Otherwise, attempt a + * lazy match by waiting to see if there's a better + * match at the next position. */ - /* Main code (part 1): Literal symbols */ - for (i = 0; i < LZX_NUM_CHARS; i++) - costs->main[i] = 8; + if (prev_len) + lzx_declare_literal(c, *(window_ptr - 2), &next_chosen_item); - /* Main code (part 2): Match header symbols */ - for (; i < num_main_syms; i++) - costs->main[i] = 10; + prev_len = cur_len; + prev_offset_data = cur_offset_data; + prev_score = cur_score; - /* Length code */ - for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) - costs->len[i] = 8; + if (prev_len >= c->params.nice_match_length) { + skip_len = prev_len - 1; + goto output_prev_match; + } + continue; + } - /* Aligned offset code */ - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) - costs->aligned[i] = 3; + /* Current match is not better than the previous match, so + * output the previous match. */ + + skip_len = prev_len - 2; + + output_prev_match: + if (prev_offset_data < LZX_NUM_RECENT_OFFSETS) { + lzx_declare_repeat_offset_match(c, prev_len, + prev_offset_data, + &next_chosen_item); + swap(c->queue.R[0], c->queue.R[prev_offset_data]); + } else { + lzx_declare_explicit_offset_match(c, prev_len, + prev_offset_data - LZX_OFFSET_OFFSET, + &next_chosen_item); + c->queue.R[2] = c->queue.R[1]; + c->queue.R[1] = c->queue.R[0]; + c->queue.R[0] = prev_offset_data - LZX_OFFSET_OFFSET; + } + lz_mf_skip_positions(mf, skip_len); + window_ptr += skip_len; + prev_len = 0; + } + + return next_chosen_item - c->chosen_items; } /* Given the frequencies of symbols in an LZX-compressed block and the @@ -1829,548 +1891,442 @@ static int lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs, const struct lzx_codes * codes) { - unsigned aligned_cost = 0; - unsigned verbatim_cost = 0; + u32 aligned_cost = 0; + u32 verbatim_cost = 0; - /* Verbatim blocks have a constant 3 bits per position footer. Aligned - * offset blocks have an aligned offset symbol per position footer, plus - * an extra 24 bits per block to output the lengths necessary to - * reconstruct the aligned offset code itself. */ + /* A verbatim block requires 3 bits in each place that an aligned symbol + * would be used in an aligned offset block. */ for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { verbatim_cost += 3 * freqs->aligned[i]; aligned_cost += codes->lens.aligned[i] * freqs->aligned[i]; } + + /* Account for output of the aligned offset code. */ aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS; + if (aligned_cost < verbatim_cost) return LZX_BLOCKTYPE_ALIGNED; else return LZX_BLOCKTYPE_VERBATIM; } -/* Find a near-optimal sequence of matches/literals with which to output the - * specified LZX block, then set the block's type to that which has the minimum - * cost to output (either verbatim or aligned). */ -static void -lzx_optimize_block(struct lzx_compressor *ctx, struct lzx_block_spec *spec, - unsigned num_passes) +/* Near-optimal parsing */ +static u32 +lzx_choose_near_optimal_items_for_block(struct lzx_compressor *c, + u32 block_start_pos, u32 block_size) { - 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; - - LZX_ASSERT(num_passes >= 1); - LZX_ASSERT(lz_bt_get_position(&ctx->mf) == spec->window_pos); + u32 num_passes_remaining = c->params.num_optim_passes; + struct lzx_lru_queue orig_queue; + struct lzx_item *next_chosen_item; + struct lzx_item **next_chosen_item_ptr; + + /* Choose appropriate match-finder wrapper functions. */ + if (num_passes_remaining > 1) { + if (block_size == c->cur_window_size) + c->get_matches_func = lzx_get_matches_fillcache_singleblock; + else + c->get_matches_func = lzx_get_matches_fillcache_multiblock; + c->skip_bytes_func = lzx_skip_bytes_fillcache; + } else { + if (block_size == c->cur_window_size) + c->get_matches_func = lzx_get_matches_nocache_singleblock; + else + c->get_matches_func = lzx_get_matches_nocache_multiblock; + c->skip_bytes_func = lzx_skip_bytes_nocache; + } - ctx->match_window_end = spec->window_pos + spec->block_size; - ctx->matches_cached = false; + /* No matches will extend beyond the end of the block. */ + c->match_window_end = block_start_pos + block_size; - /* 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. + /* The first optimization pass will use a default cost model. Each + * additional optimization pass will use a cost model 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. */ - - 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; - - 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_tally_literal(*window_ptr, &freqs); - window_ptr += 1; - } - } - 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; - } + * matches and literals in intermediate form on the final pass. For + * earlier passes, tallying symbol frequencies is sufficient. */ + lzx_set_default_costs(&c->costs, c->num_main_syms); - 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_item_ptr = NULL; + orig_queue = c->queue; + do { + /* Reset the match-finder wrapper. */ + c->match_window_pos = block_start_pos; + c->cache_ptr = c->cached_matches; + + if (num_passes_remaining == 1) { + /* Last pass: actually generate the items. */ + next_chosen_item = c->chosen_items; + next_chosen_item_ptr = &next_chosen_item; } - *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) -{ - /* Set up a default cost model. */ - lzx_set_default_costs(&ctx->costs, ctx->num_main_syms); + /* Choose the items. */ + lzx_optim_pass(c, next_chosen_item_ptr); - /* 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); - } + if (num_passes_remaining > 1) { + /* This isn't the last pass. */ - /* 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_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); - } + /* Make the Huffman codes from the symbol frequencies. */ + lzx_make_huffman_codes(&c->freqs, &c->codes[c->codes_index], + c->num_main_syms); + + /* Update symbol costs. */ + lzx_set_costs(c, &c->codes[c->codes_index].lens); + + /* Reset symbol frequencies. */ + memset(&c->freqs, 0, sizeof(c->freqs)); + + /* Reset the match offset LRU queue to what it was at + * the beginning of the block. */ + c->queue = orig_queue; + + /* Choose appopriate match-finder wrapper functions. */ + if (c->cache_ptr <= c->cache_limit) { + c->get_matches_func = lzx_get_matches_usecache_nocheck; + c->skip_bytes_func = lzx_skip_bytes_usecache_nocheck; + } else { + c->get_matches_func = lzx_get_matches_usecache; + c->skip_bytes_func = lzx_skip_bytes_usecache; + } + } + } while (--num_passes_remaining); + + /* Return the number of items chosen. */ + return next_chosen_item - c->chosen_items; } /* - * This is the fast version of lzx_prepare_blocks(). This version "quickly" - * prepares a single compressed block containing the entire input. See the - * description of the "Fast algorithm" at the beginning of this file for more - * information. + * Choose the matches/literals with which to output the block of data beginning + * at '&c->cur_window[block_start_pos]' and extending for 'block_size' bytes. * - * Input --- the preprocessed data: + * The frequences of the Huffman symbols in the block will be tallied in + * 'c->freqs'. * - * ctx->window[] - * ctx->window_size - * - * Output --- the block specification and the corresponding match/literal data: + * 'c->queue' must specify the state of the queue at the beginning of this block. + * This function will update it to the state of the queue at the end of this + * block. * - * ctx->block_specs[] - * ctx->num_blocks - * ctx->chosen_items[] + * Returns the number of matches/literals that were chosen and written to + * 'c->chosen_items' in the 'struct lzx_item' intermediate representation. */ -static void -lzx_prepare_block_fast(struct lzx_compressor * ctx) +static u32 +lzx_choose_items_for_block(struct lzx_compressor *c, + u32 block_start_pos, u32 block_size) { - struct lzx_record_ctx record_ctx; - struct lzx_block_spec *spec; - - /* Parameters to hash chain LZ match finder - * (lazy with 1 match lookahead) */ - static const struct lz_params lzx_lz_params = { - /* Although LZX_MIN_MATCH_LEN == 2, length 2 matches typically - * aren't worth choosing when using greedy or lazy parsing. */ - .min_match = 3, - .max_match = LZX_MAX_MATCH_LEN, - .max_offset = LZX_MAX_WINDOW_SIZE, - .good_match = LZX_MAX_MATCH_LEN, - .nice_match = LZX_MAX_MATCH_LEN, - .max_chain_len = LZX_MAX_MATCH_LEN, - .max_lazy_match = LZX_MAX_MATCH_LEN, - .too_far = 4096, - }; - - /* 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_items; - - /* Determine series of matches/literals to output. */ - lz_analyze_block(ctx->window, - ctx->window_size, - lzx_record_match, - lzx_record_literal, - &record_ctx, - &lzx_lz_params, - ctx->prev_tab); - - /* Set up block specification. */ - spec = &ctx->block_specs[0]; - spec->block_type = LZX_BLOCKTYPE_ALIGNED; - spec->window_pos = 0; - spec->block_size = ctx->window_size; - 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; + return (*c->params.choose_items_for_block)(c, block_start_pos, block_size); } -static size_t -lzx_compress(const void *uncompressed_data, size_t uncompressed_size, - void *compressed_data, size_t compressed_size_avail, void *_ctx) +/* Initialize c->offset_slot_fast. */ +static void +lzx_init_offset_slot_fast(struct lzx_compressor *c) { - struct lzx_compressor *ctx = _ctx; - struct output_bitstream ostream; - size_t compressed_size; + u8 slot = 0; - if (uncompressed_size < 100) { - LZX_DEBUG("Too small to bother compressing."); - return 0; - } + for (u32 offset = 0; offset < LZX_NUM_FAST_OFFSETS; offset++) { - if (uncompressed_size > ctx->max_window_size) { - LZX_DEBUG("Can't compress %zu bytes using window of %u bytes!", - uncompressed_size, ctx->max_window_size); - return 0; + while (offset + LZX_OFFSET_OFFSET >= lzx_offset_slot_base[slot + 1]) + slot++; + + c->offset_slot_fast[offset] = slot; } +} - LZX_DEBUG("Attempting to compress %zu bytes...", - uncompressed_size); +/* Set internal compression parameters for the specified compression level and + * maximum window size. */ +static void +lzx_build_params(unsigned int compression_level, u32 max_window_size, + struct lzx_compressor_params *lzx_params) +{ + if (compression_level < 25) { - /* The input data must be preprocessed. To avoid changing the original - * input, copy it to a temporary buffer. */ - memcpy(ctx->window, uncompressed_data, uncompressed_size); - ctx->window_size = uncompressed_size; + /* Fast compression: Use lazy parsing. */ - /* This line is unnecessary; it just avoids inconsequential accesses of - * uninitialized memory that would show up in memory-checking tools such - * as valgrind. */ - memset(&ctx->window[ctx->window_size], 0, 12); + lzx_params->choose_items_for_block = lzx_choose_lazy_items_for_block; + lzx_params->num_optim_passes = 1; - LZX_DEBUG("Preprocessing data..."); + /* When lazy parsing, the hash chain match-finding algorithm is + * fastest unless the window is too large. + * + * TODO: something like hash arrays would actually be better + * than binary trees on large windows. */ + if (max_window_size <= 262144) + lzx_params->mf_algo = LZ_MF_HASH_CHAINS; + else + lzx_params->mf_algo = LZ_MF_BINARY_TREES; - /* Before doing any actual compression, do the call instruction (0xe8 - * byte) translation on the uncompressed data. */ - lzx_do_e8_preprocessing(ctx->window, ctx->window_size); + /* When lazy parsing, don't bother with length 2 matches. */ + lzx_params->min_match_length = 3; - LZX_DEBUG("Preparing blocks..."); + /* Scale nice_match_length and max_search_depth with the + * compression level. */ + lzx_params->nice_match_length = 25 + compression_level * 2; + lzx_params->max_search_depth = 25 + compression_level; + } else { - /* Prepare the compressed data. */ - if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_FAST) - lzx_prepare_block_fast(ctx); - else - lzx_prepare_blocks(ctx); + /* Normal / high compression: Use near-optimal parsing. */ - LZX_DEBUG("Writing compressed blocks..."); + lzx_params->choose_items_for_block = lzx_choose_near_optimal_items_for_block; - /* Generate the compressed data. */ - init_output_bitstream(&ostream, compressed_data, compressed_size_avail); - lzx_write_all_blocks(ctx, &ostream); + /* Set a number of optimization passes appropriate for the + * compression level. */ - LZX_DEBUG("Flushing bitstream..."); - compressed_size = flush_output_bitstream(&ostream); - if (compressed_size == (u32)~0UL) { - LZX_DEBUG("Data did not compress to %zu bytes or less!", - compressed_size_avail); - return 0; - } + lzx_params->num_optim_passes = 1; - LZX_DEBUG("Done: compressed %zu => %zu bytes.", - uncompressed_size, compressed_size); - - /* Verify that we really get the same thing back when decompressing. - * Although this could be disabled by default in all cases, it only - * takes around 2-3% of the running time of the slow algorithm to do the - * verification. */ - if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_SLOW - #if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION) - || 1 - #endif - ) - { - struct wimlib_decompressor *decompressor; - - if (0 == wimlib_create_decompressor(WIMLIB_COMPRESSION_TYPE_LZX, - ctx->max_window_size, - NULL, - &decompressor)) - { - int ret; - ret = wimlib_decompress(compressed_data, - compressed_size, - ctx->window, - uncompressed_size, - decompressor); - wimlib_free_decompressor(decompressor); - - if (ret) { - ERROR("Failed to decompress data we " - "compressed using LZX algorithm"); - wimlib_assert(0); - return 0; - } - if (memcmp(uncompressed_data, ctx->window, uncompressed_size)) { - ERROR("Data we compressed using LZX algorithm " - "didn't decompress to original"); - wimlib_assert(0); - return 0; - } - } else { - WARNING("Failed to create decompressor for " - "data verification!"); + if (compression_level >= 40) + lzx_params->num_optim_passes++; + + /* Use more optimization passes for higher compression levels. + * But the more passes there are, the less they help --- so + * don't add them linearly. */ + if (compression_level >= 70) { + lzx_params->num_optim_passes++; + if (compression_level >= 100) + lzx_params->num_optim_passes++; + if (compression_level >= 150) + lzx_params->num_optim_passes++; + if (compression_level >= 200) + lzx_params->num_optim_passes++; + if (compression_level >= 300) + lzx_params->num_optim_passes++; } + + /* When doing near-optimal parsing, the hash chain match-finding + * algorithm is good if the window size is small and we're only + * doing one optimization pass. Otherwise, the binary tree + * algorithm is the way to go. */ + if (max_window_size <= 32768 && lzx_params->num_optim_passes == 1) + lzx_params->mf_algo = LZ_MF_HASH_CHAINS; + else + lzx_params->mf_algo = LZ_MF_BINARY_TREES; + + /* When doing near-optimal parsing, allow length 2 matches if + * the compression level is sufficiently high. */ + if (compression_level >= 45) + lzx_params->min_match_length = 2; + else + lzx_params->min_match_length = 3; + + /* Scale nice_match_length and max_search_depth with the + * compression level. */ + lzx_params->nice_match_length = min(((u64)compression_level * 32) / 50, + LZX_MAX_MATCH_LEN); + lzx_params->max_search_depth = min(((u64)compression_level * 50) / 50, + LZX_MAX_MATCH_LEN); } - return compressed_size; } +/* Given the internal compression parameters and maximum window size, build the + * Lempel-Ziv match-finder parameters. */ static void -lzx_free_compressor(void *_ctx) +lzx_build_mf_params(const struct lzx_compressor_params *lzx_params, + u32 max_window_size, struct lz_mf_params *mf_params) { - struct lzx_compressor *ctx = _ctx; - - if (ctx) { - FREE(ctx->chosen_items); - FREE(ctx->cached_matches); - FREE(ctx->optimum); - lz_bt_destroy(&ctx->mf); - FREE(ctx->block_specs); - FREE(ctx->prev_tab); - FREE(ctx->window); - FREE(ctx); - } + memset(mf_params, 0, sizeof(*mf_params)); + + mf_params->algorithm = lzx_params->mf_algo; + mf_params->max_window_size = max_window_size; + mf_params->min_match_len = lzx_params->min_match_length; + mf_params->max_match_len = LZX_MAX_MATCH_LEN; + mf_params->max_search_depth = lzx_params->max_search_depth; + mf_params->nice_match_len = lzx_params->nice_match_length; } -static const struct wimlib_lzx_compressor_params lzx_fast_default = { - .hdr = { - .size = sizeof(struct wimlib_lzx_compressor_params), - }, - .algorithm = WIMLIB_LZX_ALGORITHM_FAST, - .use_defaults = 0, - .alg_params = { - .fast = { - }, - }, -}; -static const struct wimlib_lzx_compressor_params lzx_slow_default = { - .hdr = { - .size = sizeof(struct wimlib_lzx_compressor_params), - }, - .algorithm = WIMLIB_LZX_ALGORITHM_SLOW, - .use_defaults = 0, - .alg_params = { - .slow = { - .use_len2_matches = 1, - .nice_match_length = 32, - .num_optim_passes = 2, - .max_search_depth = 50, - .main_nostat_cost = 15, - .len_nostat_cost = 15, - .aligned_nostat_cost = 7, - }, - }, -}; +static void +lzx_free_compressor(void *_c); -static const struct wimlib_lzx_compressor_params * -lzx_get_params(const struct wimlib_compressor_params_header *_params) +static u64 +lzx_get_needed_memory(size_t max_block_size, unsigned int compression_level) { - const struct wimlib_lzx_compressor_params *params = - (const struct wimlib_lzx_compressor_params*)_params; + struct lzx_compressor_params params; + u64 size = 0; + unsigned window_order; + u32 max_window_size; - if (params == NULL) { - LZX_DEBUG("Using default algorithm and parameters."); - params = &lzx_slow_default; - } else { - if (params->use_defaults) { - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) - params = &lzx_slow_default; - else - params = &lzx_fast_default; - } - } - return params; + window_order = lzx_get_window_order(max_block_size); + if (window_order == 0) + return 0; + max_window_size = max_block_size; + + lzx_build_params(compression_level, max_window_size, ¶ms); + + size += sizeof(struct lzx_compressor); + + /* cur_window */ + size += max_window_size; + + /* mf */ + size += lz_mf_get_needed_memory(params.mf_algo, max_window_size); + + /* cached_matches */ + if (params.num_optim_passes > 1) + size += LZX_CACHE_LEN * sizeof(struct lz_match); + else + size += LZX_MAX_MATCHES_PER_POS * sizeof(struct lz_match); + return size; } static int -lzx_create_compressor(size_t window_size, - const struct wimlib_compressor_params_header *_params, - void **ctx_ret) +lzx_create_compressor(size_t max_block_size, unsigned int compression_level, + void **c_ret) { - const struct wimlib_lzx_compressor_params *params = lzx_get_params(_params); - struct lzx_compressor *ctx; - - LZX_DEBUG("Allocating LZX context..."); + struct lzx_compressor *c; + struct lzx_compressor_params params; + struct lz_mf_params mf_params; + unsigned window_order; + u32 max_window_size; - if (!lzx_window_size_valid(window_size)) + window_order = lzx_get_window_order(max_block_size); + if (window_order == 0) return WIMLIB_ERR_INVALID_PARAM; + max_window_size = max_block_size; - ctx = CALLOC(1, sizeof(struct lzx_compressor)); - if (ctx == NULL) - goto oom; + lzx_build_params(compression_level, max_window_size, ¶ms); + lzx_build_mf_params(¶ms, max_window_size, &mf_params); + if (!lz_mf_params_valid(&mf_params)) + return WIMLIB_ERR_INVALID_PARAM; - ctx->num_main_syms = lzx_get_num_main_syms(window_size); - ctx->max_window_size = window_size; - ctx->window = MALLOC(window_size + 12); - if (ctx->window == NULL) + c = CALLOC(1, sizeof(struct lzx_compressor)); + if (!c) goto oom; - if (params->algorithm == WIMLIB_LZX_ALGORITHM_FAST) { - ctx->prev_tab = MALLOC(window_size * sizeof(ctx->prev_tab[0])); - if (ctx->prev_tab == NULL) - goto oom; - } + c->params = params; + c->num_main_syms = lzx_get_num_main_syms(window_order); + c->window_order = window_order; - size_t block_specs_length = DIV_ROUND_UP(window_size, LZX_DIV_BLOCK_SIZE); - ctx->block_specs = MALLOC(block_specs_length * sizeof(ctx->block_specs[0])); - if (ctx->block_specs == NULL) + /* The window is allocated as 16-byte aligned to speed up memcpy() and + * enable lzx_e8_filter() optimization on x86_64. */ + c->cur_window = ALIGNED_MALLOC(max_window_size, 16); + if (!c->cur_window) goto oom; - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - unsigned min_match_len = LZX_MIN_MATCH_LEN; - if (!params->alg_params.slow.use_len2_matches) - min_match_len = max(min_match_len, 3); - - 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; - } + c->mf = lz_mf_alloc(&mf_params); + if (!c->mf) + goto oom; - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - 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) + if (params.num_optim_passes > 1) { + c->cached_matches = MALLOC(LZX_CACHE_LEN * + sizeof(struct lz_match)); + if (!c->cached_matches) goto oom; - } - - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - ctx->cached_matches = MALLOC(LZX_CACHE_SIZE); - if (ctx->cached_matches == NULL) + c->cache_limit = c->cached_matches + LZX_CACHE_LEN - + (LZX_MAX_MATCHES_PER_POS + 1); + } else { + c->cached_matches = MALLOC(LZX_MAX_MATCHES_PER_POS * + sizeof(struct lz_match)); + if (!c->cached_matches) goto oom; - ctx->cache_limit = ctx->cached_matches + - LZX_CACHE_LEN - (LZX_MAX_MATCHES_PER_POS + 1); } - 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)); - memset(&ctx->zero_codes, 0, sizeof(ctx->zero_codes)); - - LZX_DEBUG("Successfully allocated new LZX context."); + lzx_init_offset_slot_fast(c); - *ctx_ret = ctx; + *c_ret = c; return 0; oom: - lzx_free_compressor(ctx); + lzx_free_compressor(c); return WIMLIB_ERR_NOMEM; } -static u64 -lzx_get_needed_memory(size_t max_block_size, - const struct wimlib_compressor_params_header *_params) +static size_t +lzx_compress(const void *uncompressed_data, size_t uncompressed_size, + void *compressed_data, size_t compressed_size_avail, void *_c) { - const struct wimlib_lzx_compressor_params *params = lzx_get_params(_params); + struct lzx_compressor *c = _c; + struct lzx_output_bitstream os; + u32 num_chosen_items; + const struct lzx_lens *prev_lens; + u32 block_start_pos; + u32 block_size; + int block_type; - u64 size = 0; + /* Don't bother compressing very small inputs. */ + if (uncompressed_size < 100) + return 0; - size += sizeof(struct lzx_compressor); + /* The input data must be preprocessed. To avoid changing the original + * input data, copy it to a temporary buffer. */ + memcpy(c->cur_window, uncompressed_data, uncompressed_size); + c->cur_window_size = uncompressed_size; - size += max_block_size + 12; + /* Preprocess the data. */ + lzx_do_e8_preprocessing(c->cur_window, c->cur_window_size); - size += DIV_ROUND_UP(max_block_size, LZX_DIV_BLOCK_SIZE) * - sizeof(((struct lzx_compressor*)0)->block_specs[0]); + /* Load the window into the match-finder. */ + lz_mf_load_window(c->mf, c->cur_window, c->cur_window_size); - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - 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]); - } - return size; -} + /* Initialize the match offset LRU queue. */ + lzx_lru_queue_init(&c->queue); -static bool -lzx_params_valid(const struct wimlib_compressor_params_header *_params) -{ - const struct wimlib_lzx_compressor_params *params = - (const struct wimlib_lzx_compressor_params*)_params; + /* Initialize the output bitstream. */ + lzx_init_output(&os, compressed_data, compressed_size_avail); - if (params->hdr.size != sizeof(struct wimlib_lzx_compressor_params)) { - LZX_DEBUG("Invalid parameter structure size!"); - return false; - } + /* Compress the data block by block. + * + * 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. */ + block_start_pos = 0; + c->codes_index = 0; + prev_lens = &c->zero_lens; + do { + /* Compute the block size. */ + block_size = min(LZX_DIV_BLOCK_SIZE, + uncompressed_size - block_start_pos); - if (params->algorithm != WIMLIB_LZX_ALGORITHM_SLOW && - params->algorithm != WIMLIB_LZX_ALGORITHM_FAST) - { - LZX_DEBUG("Invalid algorithm."); - return false; - } + /* Reset symbol frequencies. */ + memset(&c->freqs, 0, sizeof(c->freqs)); - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW && - !params->use_defaults) - { - if (params->alg_params.slow.num_optim_passes < 1) - { - LZX_DEBUG("Invalid number of optimization passes!"); - return false; - } + /* Prepare the matches/literals for the block. */ + num_chosen_items = lzx_choose_items_for_block(c, + block_start_pos, + block_size); - if (params->alg_params.slow.main_nostat_cost < 1 || - params->alg_params.slow.main_nostat_cost > 16) - { - LZX_DEBUG("Invalid main_nostat_cost!"); - return false; - } + /* Make the Huffman codes from the symbol frequencies. */ + lzx_make_huffman_codes(&c->freqs, &c->codes[c->codes_index], + c->num_main_syms); - if (params->alg_params.slow.len_nostat_cost < 1 || - params->alg_params.slow.len_nostat_cost > 16) - { - LZX_DEBUG("Invalid len_nostat_cost!"); - return false; - } + /* Choose the best block type. + * + * Note: we currently don't consider uncompressed blocks. */ + block_type = lzx_choose_verbatim_or_aligned(&c->freqs, + &c->codes[c->codes_index]); + + /* Write the compressed block to the output buffer. */ + lzx_write_compressed_block(block_type, + block_size, + c->window_order, + c->num_main_syms, + c->chosen_items, + num_chosen_items, + &c->codes[c->codes_index], + prev_lens, + &os); + + /* The current codeword lengths become the previous lengths. */ + prev_lens = &c->codes[c->codes_index].lens; + c->codes_index ^= 1; + + block_start_pos += block_size; + + } while (block_start_pos != uncompressed_size); + + return lzx_flush_output(&os); +} - if (params->alg_params.slow.aligned_nostat_cost < 1 || - params->alg_params.slow.aligned_nostat_cost > 8) - { - LZX_DEBUG("Invalid aligned_nostat_cost!"); - return false; - } +static void +lzx_free_compressor(void *_c) +{ + struct lzx_compressor *c = _c; + + if (c) { + ALIGNED_FREE(c->cur_window); + lz_mf_free(c->mf); + FREE(c->cached_matches); + FREE(c); } - return true; } const struct compressor_ops lzx_compressor_ops = { - .params_valid = lzx_params_valid, .get_needed_memory = lzx_get_needed_memory, .create_compressor = lzx_create_compressor, .compress = lzx_compress,