X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=src%2Flzx_compress.c;h=f0c3c5268c82c5e26347b9c42e5c008e811886c8;hp=6032f1e26f55621bc40a2c622efd6875d6106145;hb=48c386f471f50c2e8d07f9766be2410d401297c1;hpb=40a690416a3951361ec77d33a723dd4497fb7585 diff --git a/src/lzx_compress.c b/src/lzx_compress.c index 6032f1e2..f0c3c526 100644 --- a/src/lzx_compress.c +++ b/src/lzx_compress.c @@ -1,11 +1,11 @@ /* * lzx_compress.c * - * A compressor for the LZX compression format, as used in WIM files. + * A compressor for the LZX compression format, as used in WIM archives. */ /* - * Copyright (C) 2012, 2013, 2014 Eric Biggers + * Copyright (C) 2012-2016 Eric Biggers * * This file is free software; you can redistribute it and/or modify it under * the terms of the GNU Lesser General Public License as published by the Free @@ -26,20 +26,19 @@ * This file contains a compressor for the LZX ("Lempel-Ziv eXtended") * compression format, as used in the WIM (Windows IMaging) file format. * - * 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. + * Two different LZX-compatible 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. * * 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. + * slightly different, and sliding window support might be required. * - * 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 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. @@ -54,379 +53,637 @@ * ("verbatim" and "aligned"). * * - 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. + * useful, but generally only if the compressor is smart about choosing them. * * - 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. */ +/******************************************************************************/ +/* General parameters */ +/*----------------------------------------------------------------------------*/ + +/* + * The compressor uses the faster algorithm at levels <= MAX_FAST_LEVEL. It + * uses the slower algorithm at levels > MAX_FAST_LEVEL. + */ +#define MAX_FAST_LEVEL 34 + +/* + * The compressor-side limits on the codeword lengths (in bits) for each Huffman + * code. To make outputting bits slightly faster, some of these limits are + * lower than the limits defined by the LZX format. This does not significantly + * affect the compression ratio. + */ +#define MAIN_CODEWORD_LIMIT 16 +#define LENGTH_CODEWORD_LIMIT 12 +#define ALIGNED_CODEWORD_LIMIT 7 +#define PRE_CODEWORD_LIMIT 7 + + +/******************************************************************************/ +/* Block splitting parameters */ +/*----------------------------------------------------------------------------*/ + +/* + * The compressor always outputs blocks of at least this size in bytes, except + * for the last block which may need to be smaller. + */ +#define MIN_BLOCK_SIZE 6500 + +/* + * The compressor attempts to end a block when it reaches this size in bytes. + * The final size might be slightly larger due to matches extending beyond the + * end of the block. Specifically: + * + * - The near-optimal compressor may choose a match of up to LZX_MAX_MATCH_LEN + * bytes starting at position 'SOFT_MAX_BLOCK_SIZE - 1'. + * + * - The lazy compressor may choose a sequence of literals starting at position + * 'SOFT_MAX_BLOCK_SIZE - 1' when it sees a sequence of increasingly better + * matches. The final match may be up to LZX_MAX_MATCH_LEN bytes. The + * length of the literal sequence is approximately limited by the "nice match + * length" parameter. + */ +#define SOFT_MAX_BLOCK_SIZE 100000 + +/* + * The number of observed items (matches and literals) that represents + * sufficient data for the compressor to decide whether the current block should + * be ended or not. + */ +#define NUM_OBSERVATIONS_PER_BLOCK_CHECK 400 + + +/******************************************************************************/ +/* Parameters for slower algorithm */ +/*----------------------------------------------------------------------------*/ + +/* + * The log base 2 of the number of entries in the hash table for finding length + * 2 matches. This could be as high as 16, but using a smaller hash table + * speeds up compression due to reduced cache pressure. + */ +#define BT_MATCHFINDER_HASH2_ORDER 12 + +/* + * The number of lz_match structures in the match cache, excluding the extra + * "overflow" entries. This value should be high enough so that nearly the + * time, all matches found in a given block can fit in the match cache. + * However, fallback behavior (immediately terminating the block) on cache + * overflow is still required. + */ +#define CACHE_LENGTH (SOFT_MAX_BLOCK_SIZE * 5) + +/* + * An upper bound on the number of matches that can ever be saved in the match + * cache for a single position. Since each match we save for a single position + * has a distinct length, we can use the number of possible match lengths in LZX + * as this bound. This bound is guaranteed to be valid in all cases, although + * if 'nice_match_length < LZX_MAX_MATCH_LEN', then it will never actually be + * reached. + */ +#define MAX_MATCHES_PER_POS LZX_NUM_LENS + +/* + * A scaling factor that makes it possible to consider fractional bit costs. A + * single bit has a cost of BIT_COST. + * + * Note: this is only useful as a statistical trick for when the true costs are + * unknown. Ultimately, each token in LZX requires a whole number of bits to + * output. + */ +#define BIT_COST 64 + +/* + * Should the compressor take into account the costs of aligned offset symbols + * instead of assuming that all are equally likely? + */ +#define CONSIDER_ALIGNED_COSTS 1 + +/* + * Should the "minimum" cost path search algorithm consider "gap" matches, where + * a normal match is followed by a literal, then by a match with the same + * offset? This is one specific, somewhat common situation in which the true + * minimum cost path is often different from the path found by looking only one + * edge ahead. + */ +#define CONSIDER_GAP_MATCHES 1 + +/******************************************************************************/ +/* Includes */ +/*----------------------------------------------------------------------------*/ + #ifdef HAVE_CONFIG_H # include "config.h" #endif #include "wimlib/compress_common.h" #include "wimlib/compressor_ops.h" -#include "wimlib/endianness.h" #include "wimlib/error.h" -#include "wimlib/lz_mf.h" -#include "wimlib/lz_repsearch.h" +#include "wimlib/lz_extend.h" #include "wimlib/lzx_common.h" +#include "wimlib/unaligned.h" #include "wimlib/util.h" -#include -#include - -#define LZX_OPTIM_ARRAY_LENGTH 4096 - -#define LZX_DIV_BLOCK_SIZE 32768 +/* Note: BT_MATCHFINDER_HASH2_ORDER must be defined before including + * bt_matchfinder.h. */ -#define LZX_CACHE_PER_POS 8 +/* Matchfinders with 16-bit positions */ +#define mf_pos_t u16 +#define MF_SUFFIX _16 +#include "wimlib/bt_matchfinder.h" +#include "wimlib/hc_matchfinder.h" -#define LZX_MAX_MATCHES_PER_POS (LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1) +/* Matchfinders with 32-bit positions */ +#undef mf_pos_t +#undef MF_SUFFIX +#define mf_pos_t u32 +#define MF_SUFFIX _32 +#include "wimlib/bt_matchfinder.h" +#include "wimlib/hc_matchfinder.h" -#define LZX_CACHE_LEN (LZX_DIV_BLOCK_SIZE * (LZX_CACHE_PER_POS + 1)) +/******************************************************************************/ +/* Compressor structure */ +/*----------------------------------------------------------------------------*/ -struct lzx_compressor; - -/* Codewords for the LZX Huffman codes. */ +/* Codewords for the 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 Huffman codes. - * A zero length means the corresponding codeword has zero frequency. */ +/* + * Codeword lengths, in bits, for the Huffman codes. + * + * A codeword length of 0 means the corresponding codeword has zero frequency. + * + * The main and length codes each have one extra entry for use as a sentinel. + * See lzx_write_compressed_code(). + */ struct lzx_lens { - u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; - u8 len[LZX_LENCODE_NUM_SYMBOLS]; - u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; -}; - -/* 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 main[LZX_MAINCODE_MAX_NUM_SYMBOLS + 1]; + u8 len[LZX_LENCODE_NUM_SYMBOLS + 1]; u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* Codewords and lengths for the LZX Huffman codes. */ +/* Codewords and lengths for the Huffman codes */ struct lzx_codes { struct lzx_codewords codewords; struct lzx_lens lens; }; -/* Symbol frequency counters for the LZX Huffman codes. */ +/* Symbol frequency counters for the Huffman-encoded alphabets */ struct lzx_freqs { u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; u32 len[LZX_LENCODE_NUM_SYMBOLS]; u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* Intermediate LZX match/literal format */ -struct lzx_item { - - /* 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; +/* Block split statistics. See the "Block splitting algorithm" section later in + * this file for details. */ +#define NUM_LITERAL_OBSERVATION_TYPES 8 +#define NUM_MATCH_OBSERVATION_TYPES 2 +#define NUM_OBSERVATION_TYPES (NUM_LITERAL_OBSERVATION_TYPES + \ + NUM_MATCH_OBSERVATION_TYPES) +struct lzx_block_split_stats { + u32 new_observations[NUM_OBSERVATION_TYPES]; + u32 observations[NUM_OBSERVATION_TYPES]; + u32 num_new_observations; + u32 num_observations; }; -/* 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; +/* + * Represents a run of literals followed by a match or end-of-block. This + * structure is needed to temporarily store items chosen by the compressor, + * since items cannot be written until all items for the block have been chosen + * and the block's Huffman codes have been computed. + */ +struct lzx_sequence { + + /* The number of literals in the run. This may be 0. The literals are + * not stored explicitly in this structure; instead, they are read + * directly from the uncompressed data. */ + u16 litrunlen; + + /* If the next field doesn't indicate end-of-block, then this is the + * match length minus LZX_MIN_MATCH_LEN. */ + u16 adjusted_length; + + /* If bit 31 is clear, then this field contains the match header in bits + * 0-8, and either the match offset plus LZX_OFFSET_ADJUSTMENT or a + * recent offset code in bits 9-30. Otherwise (if bit 31 is set), this + * sequence's literal run was the last literal run in the block, so + * there is no match that follows it. */ + u32 adjusted_offset_and_match_hdr; }; /* - * Match chooser position data: + * This structure represents a byte position in the input buffer and a node in + * the graph of possible match/literal choices. * - * 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. + * Logically, each incoming edge to this node is labeled with a literal or a + * match that can be taken to reach this position from an earlier position; and + * each outgoing edge from this node is labeled with a literal or a match that + * can be taken to advance from this position to a later position. */ -struct lzx_mc_pos_data { +struct lzx_optimum_node { /* 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 or literal that was taken to reach this position. This can - * change as progressively lower cost paths are found to reach this - * position. + /* + * The best arrival to this node, i.e. the match or literal that was + * used to arrive to this position at the given 'cost'. This can change + * as progressively lower cost paths are found to reach this position. * - * This variable is divided into two bitfields. + * For non-gap matches, this variable is divided into two bitfields + * whose meanings depend on the item type: * * Literals: - * Low bits are 1, high bits are the literal. + * Low bits are 0, high bits are the literal. * * Explicit offset matches: - * Low bits are the match length, high bits are the offset plus 2. + * Low bits are the match length, high bits are the offset plus + * LZX_OFFSET_ADJUSTMENT. * * Repeat offset matches: * Low bits are the match length, high bits are the queue index. + * + * For gap matches, identified by OPTIMUM_GAP_MATCH set, special + * behavior applies --- see the code. */ - u32 mc_item_data; -#define MC_OFFSET_SHIFT 9 -#define MC_LEN_MASK ((1 << MC_OFFSET_SHIFT) - 1) + u32 item; +#define OPTIMUM_OFFSET_SHIFT 9 +#define OPTIMUM_LEN_MASK ((1 << OPTIMUM_OFFSET_SHIFT) - 1) +#if CONSIDER_GAP_MATCHES +# define OPTIMUM_GAP_MATCH 0x80000000 +#endif - /* 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 */ +} _aligned_attribute(8); + +/* The cost model for near-optimal parsing */ +struct lzx_costs { + + /* + * 'match_cost[offset_slot][len - LZX_MIN_MATCH_LEN]' is the cost of a + * length 'len' match which has an offset belonging to 'offset_slot'. + * The cost includes the main symbol, the length symbol if required, and + * the extra offset bits if any, excluding any entropy-coded bits + * (aligned offset bits). It does *not* include the cost of the aligned + * offset symbol which may be required. + */ + u16 match_cost[LZX_MAX_OFFSET_SLOTS][LZX_NUM_LENS]; + + /* Cost of each symbol in the main code */ + u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; + + /* Cost of each symbol in the length code */ + u32 len[LZX_LENCODE_NUM_SYMBOLS]; + +#if CONSIDER_ALIGNED_COSTS + /* Cost of each symbol in the aligned offset code */ + u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; +#endif +}; + +struct lzx_output_bitstream; + +/* The main LZX compressor structure */ struct lzx_compressor { - /* Internal compression parameters */ - struct lzx_compressor_params params; + /* The buffer for preprocessed input data, if not using destructive + * compression */ + void *in_buffer; - /* The preprocessed buffer of data being compressed */ - u8 *cur_window; + /* If true, then the compressor need not preserve the input buffer if it + * compresses the data successfully */ + bool destructive; - /* 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; + /* Pointer to the compress() implementation chosen at allocation time */ + void (*impl)(struct lzx_compressor *, const u8 *, size_t, + struct lzx_output_bitstream *); - /* 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: 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().) */ + /* The log base 2 of the window size for match offset encoding purposes. + * This will be >= LZX_MIN_WINDOW_ORDER and <= LZX_MAX_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. */ + /* The number of symbols in the main alphabet. This depends on the + * window order, since the window order determines the maximum possible + * match offset. */ unsigned num_main_syms; - /* Lempel-Ziv match-finder */ - struct lz_mf *mf; + /* The "nice" match length: if a match of this length is found, then it + * is chosen immediately without further consideration. */ + unsigned nice_match_length; - /* 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; - - /* 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; - - /* Frequency counters for the current block. */ + /* The maximum search depth: at most this many potential matches are + * considered at each position. */ + unsigned max_search_depth; + + /* The number of optimization passes per block */ + unsigned num_optim_passes; + + /* The symbol frequency counters for the current block */ struct lzx_freqs freqs; - /* The Huffman codes for the current and previous blocks. */ - struct lzx_codes codes[2]; + /* Block split statistics for the current block */ + struct lzx_block_split_stats split_stats; - /* 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; + /* The Huffman codes for the current and previous blocks. The one with + * index 'codes_index' is for the current block, and the other one is + * for the previous block. */ + struct lzx_codes codes[2]; + unsigned codes_index; + + /* The matches and literals that the compressor has chosen for the + * current block. The required length of this array is limited by the + * maximum number of matches that can ever be chosen for a single block, + * plus one for the special entry at the end. */ + struct lzx_sequence chosen_sequences[ + DIV_ROUND_UP(SOFT_MAX_BLOCK_SIZE, LZX_MIN_MATCH_LEN) + 1]; + + /* Tables for mapping adjusted offsets to offset slots */ + u8 offset_slot_tab_1[32768]; /* offset slots [0, 29] */ + u8 offset_slot_tab_2[128]; /* offset slots [30, 49] */ + + union { + /* Data for lzx_compress_lazy() */ + struct { + /* Hash chains matchfinder (MUST BE LAST!!!) */ + union { + struct hc_matchfinder_16 hc_mf_16; + struct hc_matchfinder_32 hc_mf_32; + }; + }; - /* Dummy lengths that are always 0. */ - struct lzx_lens zero_lens; + /* Data for lzx_compress_near_optimal() */ + struct { + /* + * Array of nodes, one per position, for running the + * minimum-cost path algorithm. + * + * This array must be large enough to accommodate the + * worst-case number of nodes, which occurs if the + * compressor finds a match of length LZX_MAX_MATCH_LEN + * at position 'SOFT_MAX_BLOCK_SIZE - 1', producing a + * block of size 'SOFT_MAX_BLOCK_SIZE - 1 + + * LZX_MAX_MATCH_LEN'. Add one for the end-of-block + * node. + */ + struct lzx_optimum_node optimum_nodes[ + SOFT_MAX_BLOCK_SIZE - 1 + + LZX_MAX_MATCH_LEN + 1]; - /* Matches/literals that were chosen for the current block. */ - struct lzx_item chosen_items[LZX_DIV_BLOCK_SIZE]; + /* The cost model for the current optimization pass */ + 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]; + /* + * Cached matches for the current block. This array + * contains the matches that were found at each position + * in the block. Specifically, for each position, there + * is a special 'struct lz_match' whose 'length' field + * contains the number of matches that were found at + * that position; this is followed by the matches + * themselves, if any, sorted by strictly increasing + * length. + * + * Note: in rare cases, there will be a very high number + * of matches in the block and this array will overflow. + * If this happens, we force the end of the current + * block. CACHE_LENGTH is the length at which we + * actually check for overflow. The extra slots beyond + * this are enough to absorb the worst case overflow, + * which occurs if starting at &match_cache[CACHE_LENGTH + * - 1], we write the match count header, then write + * MAX_MATCHES_PER_POS matches, then skip searching for + * matches at 'LZX_MAX_MATCH_LEN - 1' positions and + * write the match count header for each. + */ + struct lz_match match_cache[CACHE_LENGTH + + MAX_MATCHES_PER_POS + + LZX_MAX_MATCH_LEN - 1]; + + /* Binary trees matchfinder (MUST BE LAST!!!) */ + union { + struct bt_matchfinder_16 bt_mf_16; + struct bt_matchfinder_32 bt_mf_32; + }; + }; + }; }; +/******************************************************************************/ +/* Matchfinder utilities */ +/*----------------------------------------------------------------------------*/ + +/* + * Will a matchfinder using 16-bit positions be sufficient for compressing + * buffers of up to the specified size? The limit could be 65536 bytes, but we + * also want to optimize out the use of offset_slot_tab_2 in the 16-bit case. + * This requires that the limit be no more than the length of offset_slot_tab_1 + * (currently 32768). + */ +static inline bool +lzx_is_16_bit(size_t max_bufsize) +{ + STATIC_ASSERT(ARRAY_LEN(((struct lzx_compressor *)0)->offset_slot_tab_1) == 32768); + return max_bufsize <= 32768; +} + +/* + * Return the offset slot for the specified adjusted match offset. + */ +static inline unsigned +lzx_get_offset_slot(struct lzx_compressor *c, u32 adjusted_offset, + bool is_16_bit) +{ + if (is_16_bit || adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1)) + return c->offset_slot_tab_1[adjusted_offset]; + return c->offset_slot_tab_2[adjusted_offset >> 14]; +} + +/* + * The following macros call either the 16-bit or the 32-bit version of a + * matchfinder function based on the value of 'is_16_bit', which will be known + * at compilation time. + */ + +#define CALL_HC_MF(is_16_bit, c, funcname, ...) \ + ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->hc_mf_16, ##__VA_ARGS__) : \ + CONCAT(funcname, _32)(&(c)->hc_mf_32, ##__VA_ARGS__)); + +#define CALL_BT_MF(is_16_bit, c, funcname, ...) \ + ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->bt_mf_16, ##__VA_ARGS__) : \ + CONCAT(funcname, _32)(&(c)->bt_mf_32, ##__VA_ARGS__)); + +/******************************************************************************/ +/* Output bitstream */ +/*----------------------------------------------------------------------------*/ + +/* + * The LZX bitstream is encoded as a sequence of little endian 16-bit coding + * units. Bits are ordered from most significant to least significant within + * each coding unit. + */ + /* * 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 { - /* Bits that haven't yet been written to the output buffer. */ - u32 bitbuf; + /* Bits that haven't yet been written to the output buffer */ + machine_word_t bitbuf; - /* Number of bits currently held in @bitbuf. */ - u32 bitcount; + /* Number of bits currently held in @bitbuf */ + machine_word_t bitcount; - /* Pointer to the start of the output buffer. */ - le16 *start; + /* Pointer to the start of the output buffer */ + u8 *start; /* Pointer to the position in the output buffer at which the next coding - * unit should be written. */ - le16 *next; + * unit should be written */ + u8 *next; - /* Pointer past the end of the output buffer. */ - le16 *end; + /* Pointer to just past the end of the output buffer, rounded down by + * one byte if needed to make 'end - start' a multiple of 2 */ + u8 *end; }; -/* - * Initialize the output bitstream. - * - * @os - * The output bitstream structure to initialize. - * @buffer - * The buffer being written to. - * @size - * Size of @buffer, in bytes. - */ +/* Can the specified number of bits always be added to 'bitbuf' after all + * pending 16-bit coding units have been flushed? */ +#define CAN_BUFFER(n) ((n) <= WORDBITS - 15) + +/* Initialize the output bitstream to write to the specified buffer. */ static void -lzx_init_output(struct lzx_output_bitstream *os, void *buffer, u32 size) +lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size) { os->bitbuf = 0; os->bitcount = 0; os->start = buffer; - os->next = os->start; - os->end = os->start + size / sizeof(le16); + os->next = buffer; + os->end = (u8 *)buffer + (size & ~1); } /* - * 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. + * Add some bits to the bitbuffer variable of the output bitstream. The caller + * must make sure there is enough room. */ 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) +lzx_add_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits) { - /* 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; + os->bitcount += num_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) - put_unaligned_u16_le(os->bitbuf >> os->bitcount, os->next++); +/* + * Flush bits from the bitbuffer variable to the output buffer. 'max_num_bits' + * specifies the maximum number of bits that may have been added since the last + * flush. + */ +static inline void +lzx_flush_bits(struct lzx_output_bitstream *os, unsigned max_num_bits) +{ + /* Masking the number of bits to shift is only needed to avoid undefined + * behavior; we don't actually care about the results of bad shifts. On + * x86, the explicit masking generates no extra code. */ + const u32 shift_mask = WORDBITS - 1; - /* 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) - put_unaligned_u16_le(os->bitbuf, os->next++); - os->bitcount = 0; - } - } + if (os->end - os->next < 6) + return; + put_unaligned_le16(os->bitbuf >> ((os->bitcount - 16) & + shift_mask), os->next + 0); + if (max_num_bits > 16) + put_unaligned_le16(os->bitbuf >> ((os->bitcount - 32) & + shift_mask), os->next + 2); + if (max_num_bits > 32) + put_unaligned_le16(os->bitbuf >> ((os->bitcount - 48) & + shift_mask), os->next + 4); + os->next += (os->bitcount >> 4) << 1; + os->bitcount &= 15; } -/* Use when @num_bits is a compile-time constant. Otherwise use - * lzx_write_varbits(). */ +/* Add at most 16 bits to the bitbuffer and flush it. */ static inline void -lzx_write_bits(struct lzx_output_bitstream *os, - const u32 bits, const unsigned int num_bits) +lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits) { - lzx_write_varbits(os, bits, num_bits, num_bits); + lzx_add_bits(os, bits, num_bits); + lzx_flush_bits(os, 16); } /* * 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 +static size_t lzx_flush_output(struct lzx_output_bitstream *os) { - if (os->next == os->end) + if (os->end - os->next < 6) return 0; - if (os->bitcount != 0) - put_unaligned_u16_le(os->bitbuf << (16 - os->bitcount), os->next++); + if (os->bitcount != 0) { + put_unaligned_le16(os->bitbuf << (16 - os->bitcount), os->next); + os->next += 2; + } - return (const u8 *)os->next - (const u8 *)os->start; + return os->next - os->start; } -/* Build the main, length, and aligned offset Huffman codes used in LZX. - * - * 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. */ +/******************************************************************************/ +/* Preparing Huffman codes */ +/*----------------------------------------------------------------------------*/ + +/* + * Build the Huffman codes. 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, - unsigned num_main_syms) +lzx_build_huffman_codes(struct lzx_compressor *c) { - make_canonical_huffman_code(num_main_syms, - LZX_MAX_MAIN_CODEWORD_LEN, + const struct lzx_freqs *freqs = &c->freqs; + struct lzx_codes *codes = &c->codes[c->codes_index]; + + STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 && + MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN); + make_canonical_huffman_code(c->num_main_syms, + MAIN_CODEWORD_LIMIT, freqs->main, codes->lens.main, codes->codewords.main); + STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 8 && + LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN); make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS, - LZX_MAX_LEN_CODEWORD_LEN, + LENGTH_CODEWORD_LIMIT, freqs->len, codes->lens.len, codes->codewords.len); + STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS && + ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN); make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS, - LZX_MAX_ALIGNED_CODEWORD_LEN, + ALIGNED_CODEWORD_LIMIT, freqs->aligned, codes->lens.aligned, codes->codewords.aligned); } +/* Reset the symbol frequencies for the current block. */ +static void +lzx_reset_symbol_frequencies(struct lzx_compressor *c) +{ + memset(&c->freqs, 0, sizeof(c->freqs)); +} + 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]) { @@ -439,16 +696,17 @@ lzx_compute_precode_items(const u8 lens[restrict], itemptr = precode_items; run_start = 0; - do { - /* Find the next run of codeword lengths. */ + + while (!((len = lens[run_start]) & 0x80)) { /* len = the length being repeated */ - len = lens[run_start]; + + /* Find the next run of codeword lengths. */ run_end = run_start + 1; /* Fast case for a single length. */ - if (likely(run_end == num_lens || len != lens[run_end])) { + if (likely(len != lens[run_end])) { delta = prev_lens[run_start] - len; if (delta < 0) delta += 17; @@ -461,14 +719,14 @@ lzx_compute_precode_items(const u8 lens[restrict], /* Extend the run. */ do { run_end++; - } while (run_end != num_lens && len == lens[run_end]); + } while (len == lens[run_end]); if (len == 0) { /* Run of zeroes. */ /* 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); + extra_bits = min((run_end - run_start) - 20, 0x1F); precode_freqs[18]++; *itemptr++ = 18 | (extra_bits << 5); run_start += 20 + extra_bits; @@ -476,7 +734,7 @@ lzx_compute_precode_items(const u8 lens[restrict], /* 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); + extra_bits = min((run_end - run_start) - 4, 0xF); precode_freqs[17]++; *itemptr++ = 17 | (extra_bits << 5); run_start += 4 + extra_bits; @@ -507,11 +765,15 @@ lzx_compute_precode_items(const u8 lens[restrict], *itemptr++ = delta; run_start++; } - } while (run_start != num_lens); + } return itemptr - precode_items; } +/******************************************************************************/ +/* Outputting compressed data */ +/*----------------------------------------------------------------------------*/ + /* * Output a Huffman code in the compressed form used in LZX. * @@ -555,6 +817,8 @@ lzx_write_compressed_code(struct lzx_output_bitstream *os, unsigned precode_item; unsigned precode_sym; unsigned i; + u8 saved = lens[num_lens]; + *(u8 *)(lens + num_lens) = 0x80; for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) precode_freqs[i] = 0; @@ -563,13 +827,13 @@ lzx_write_compressed_code(struct lzx_output_bitstream *os, * 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); /* Build the precode. */ - make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, - LZX_MAX_PRE_CODEWORD_LEN, + STATIC_ASSERT(PRE_CODEWORD_LIMIT >= 5 && + PRE_CODEWORD_LIMIT <= LZX_MAX_PRE_CODEWORD_LEN); + make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, PRE_CODEWORD_LIMIT, precode_freqs, precode_lens, precode_codewords); @@ -581,76 +845,25 @@ lzx_write_compressed_code(struct lzx_output_bitstream *os, 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); + lzx_add_bits(os, precode_codewords[precode_sym], + precode_lens[precode_sym]); if (precode_sym >= 17) { if (precode_sym == 17) { - lzx_write_bits(os, precode_item >> 5, 4); + lzx_add_bits(os, precode_item >> 5, 4); } else if (precode_sym == 18) { - lzx_write_bits(os, precode_item >> 5, 5); + lzx_add_bits(os, precode_item >> 5, 5); } else { - lzx_write_bits(os, (precode_item >> 5) & 1, 1); + lzx_add_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); + lzx_add_bits(os, precode_codewords[precode_sym], + precode_lens[precode_sym]); } } - } -} - -/* 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); + STATIC_ASSERT(CAN_BUFFER(2 * PRE_CODEWORD_LIMIT + 1)); + lzx_flush_bits(os, 2 * PRE_CODEWORD_LIMIT + 1); } - 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); - } + *(u8 *)(lens + num_lens) = saved; } /* @@ -663,59 +876,193 @@ lzx_write_item(struct lzx_output_bitstream *os, struct lzx_item item, * @block_type * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or * LZX_BLOCKTYPE_VERBATIM). - * @items - * The array of matches/literals to output. - * @num_items - * Number of matches/literals to output (length of @items). + * @block_data + * The uncompressed data of the block. + * @sequences + * The matches and literals to output, given as a series of sequences. * @codes - * The main, length, and aligned offset Huffman codes for the current - * LZX compressed block. + * The main, length, and aligned offset Huffman codes for the block. */ static void -lzx_write_items(struct lzx_output_bitstream *os, int block_type, - const struct lzx_item items[], u32 num_items, - const struct lzx_codes *codes) +lzx_write_sequences(struct lzx_output_bitstream *os, int block_type, + const u8 *block_data, const struct lzx_sequence sequences[], + const struct lzx_codes *codes) { - unsigned ones_if_aligned = 0U - (block_type == LZX_BLOCKTYPE_ALIGNED); + const struct lzx_sequence *seq = sequences; + u32 ones_if_aligned = 0 - (block_type == LZX_BLOCKTYPE_ALIGNED); + + for (;;) { + /* Output the next sequence. */ + + unsigned litrunlen = seq->litrunlen; + unsigned match_hdr; + unsigned main_symbol; + unsigned adjusted_length; + u32 adjusted_offset; + unsigned offset_slot; + unsigned num_extra_bits; + u32 extra_bits; + + /* Output the literal run of the sequence. */ + + if (litrunlen) { /* Is the literal run nonempty? */ + + /* Verify optimization is enabled on 64-bit */ + STATIC_ASSERT(WORDBITS < 64 || + CAN_BUFFER(3 * MAIN_CODEWORD_LIMIT)); + + if (CAN_BUFFER(3 * MAIN_CODEWORD_LIMIT)) { + + /* 64-bit: write 3 literals at a time. */ + while (litrunlen >= 3) { + unsigned lit0 = block_data[0]; + unsigned lit1 = block_data[1]; + unsigned lit2 = block_data[2]; + lzx_add_bits(os, codes->codewords.main[lit0], + codes->lens.main[lit0]); + lzx_add_bits(os, codes->codewords.main[lit1], + codes->lens.main[lit1]); + lzx_add_bits(os, codes->codewords.main[lit2], + codes->lens.main[lit2]); + lzx_flush_bits(os, 3 * MAIN_CODEWORD_LIMIT); + block_data += 3; + litrunlen -= 3; + } + if (litrunlen--) { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], + codes->lens.main[lit]); + if (litrunlen--) { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], + codes->lens.main[lit]); + lzx_flush_bits(os, 2 * MAIN_CODEWORD_LIMIT); + } else { + lzx_flush_bits(os, 1 * MAIN_CODEWORD_LIMIT); + } + } + } else { + /* 32-bit: write 1 literal at a time. */ + do { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], + codes->lens.main[lit]); + lzx_flush_bits(os, MAIN_CODEWORD_LIMIT); + } while (--litrunlen); + } + } + + /* Was this the last literal run? */ + if (seq->adjusted_offset_and_match_hdr & 0x80000000) + return; + + /* Nope; output the match. */ + + match_hdr = seq->adjusted_offset_and_match_hdr & 0x1FF; + main_symbol = LZX_NUM_CHARS + match_hdr; + adjusted_length = seq->adjusted_length; + + block_data += adjusted_length + LZX_MIN_MATCH_LEN; + + offset_slot = match_hdr / LZX_NUM_LEN_HEADERS; + adjusted_offset = seq->adjusted_offset_and_match_hdr >> 9; + + num_extra_bits = lzx_extra_offset_bits[offset_slot]; + extra_bits = adjusted_offset - lzx_offset_slot_base[offset_slot]; + + #define MAX_MATCH_BITS (MAIN_CODEWORD_LIMIT + LENGTH_CODEWORD_LIMIT + \ + 14 + ALIGNED_CODEWORD_LIMIT) + + /* Verify optimization is enabled on 64-bit */ + STATIC_ASSERT(WORDBITS < 64 || CAN_BUFFER(MAX_MATCH_BITS)); + + /* Output the main symbol for the match. */ + + lzx_add_bits(os, codes->codewords.main[main_symbol], + codes->lens.main[main_symbol]); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, MAIN_CODEWORD_LIMIT); + + /* If needed, output the length symbol for the match. */ + + if (adjusted_length >= LZX_NUM_PRIMARY_LENS) { + lzx_add_bits(os, codes->codewords.len[adjusted_length - + LZX_NUM_PRIMARY_LENS], + codes->lens.len[adjusted_length - + LZX_NUM_PRIMARY_LENS]); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, LENGTH_CODEWORD_LIMIT); + } + + /* Output the extra offset bits for the match. In aligned + * offset blocks, the lowest 3 bits of the adjusted offset are + * Huffman-encoded using the aligned offset code, provided that + * there are at least extra 3 offset bits required. All other + * extra offset bits are output verbatim. */ + + if ((adjusted_offset & ones_if_aligned) >= 16) { + + lzx_add_bits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS, + num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, 14); + + lzx_add_bits(os, codes->codewords.aligned[adjusted_offset & + LZX_ALIGNED_OFFSET_BITMASK], + codes->lens.aligned[adjusted_offset & + LZX_ALIGNED_OFFSET_BITMASK]); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT); + } else { + STATIC_ASSERT(CAN_BUFFER(17)); - for (u32 i = 0; i < num_items; i++) - lzx_write_item(os, items[i], ones_if_aligned, codes); + lzx_add_bits(os, extra_bits, num_extra_bits); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, 17); + } + + if (CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, MAX_MATCH_BITS); + + /* Advance to the next sequence. */ + seq++; + } } -/* Write an LZX aligned offset or verbatim block to the output bitstream. */ static void -lzx_write_compressed_block(int block_type, +lzx_write_compressed_block(const u8 *block_begin, + int block_type, u32 block_size, unsigned window_order, unsigned num_main_syms, - struct lzx_item * chosen_items, - u32 num_chosen_items, + const struct lzx_sequence sequences[], const struct lzx_codes * codes, const struct lzx_lens * prev_lens, struct lzx_output_bitstream * os) { - LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED || - block_type == LZX_BLOCKTYPE_VERBATIM); - /* The first three bits indicate the type of block and are one of the * LZX_BLOCKTYPE_* constants. */ lzx_write_bits(os, block_type, 3); - /* Output the block size. + /* + * Output the block size. * - * The original LZX format seemed to always encode the block size in 3 - * bytes. However, the implementation in WIMGAPI, as used in WIM files, - * uses the first bit to indicate whether the block is the default size - * (32768) or a different size given explicitly by the next 16 bits. + * The original LZX format encoded the block size in 24 bits. However, + * the LZX format used in WIM archives uses 1 bit to specify whether the + * block has the default size of 32768 bytes, then optionally 16 bits to + * specify a non-default size. This works fine for Microsoft's WIM + * software (WIMGAPI), which never compresses more than 32768 bytes at a + * time with LZX. However, as an extension, our LZX compressor supports + * compressing up to 2097152 bytes, with a corresponding increase in + * window size. It is possible for blocks in these larger buffers to + * exceed 65535 bytes; such blocks cannot have their size represented in + * 16 bits. * - * By default, this compressor uses a window size of 32768 and therefore - * follows the WIMGAPI behavior. However, this compressor also supports - * window sizes greater than 32768 bytes, which do not appear to be - * supported by WIMGAPI. In such cases, we retain the default size bit - * to mean a size of 32768 bytes but output non-default block size in 24 - * bits rather than 16. The compatibility of this behavior is unknown - * because WIMs created with chunk size greater than 32768 can seemingly - * only be opened by wimlib anyway. */ + * The chosen solution was to use 24 bits for the block size when + * possibly required --- specifically, when the compressor has been + * allocated to be capable of compressing more than 32768 bytes at once + * (which also causes the number of main symbols to be increased). + */ if (block_size == LZX_DEFAULT_BLOCK_SIZE) { lzx_write_bits(os, 1, 1); } else { @@ -749,1579 +1096,1866 @@ lzx_write_compressed_block(int block_type, LZX_LENCODE_NUM_SYMBOLS); /* Output the compressed matches and literals. */ - lzx_write_items(os, block_type, chosen_items, num_chosen_items, codes); + lzx_write_sequences(os, block_type, block_begin, sequences, codes); } -/* 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) +/* + * Given the frequencies of symbols in an LZX-compressed block and the + * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or + * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively, + * will take fewer bits to output. + */ +static int +lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs, + const struct lzx_codes * codes) { - if (c->match_window_end < c->cur_window_size && num_matches != 0) { - u32 limit = c->match_window_end - c->match_window_pos; + u32 verbatim_cost = 0; + u32 aligned_cost = 0; - if (limit >= LZX_MIN_MATCH_LEN) { + /* A verbatim block requires 3 bits in each place that an aligned offset + * symbol would be used in an aligned offset block. */ + for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i]; + aligned_cost += codes->lens.aligned[i] * freqs->aligned[i]; + } - unsigned i = num_matches - 1; - do { - if (matches[i].len >= limit) { - matches[i].len = limit; + /* Account for the cost of sending the codeword lengths of the aligned + * offset code. */ + aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * + LZX_ALIGNEDCODE_NUM_SYMBOLS; - /* 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; + if (aligned_cost < verbatim_cost) + return LZX_BLOCKTYPE_ALIGNED; + else + return LZX_BLOCKTYPE_VERBATIM; } -static unsigned -lzx_get_matches_fillcache_singleblock(struct lzx_compressor *c, - const struct lz_match **matches_ret) +/* + * Flush an LZX block: + * + * 1. Build the Huffman codes. + * 2. Decide whether to output the block as VERBATIM or ALIGNED. + * 3. Write the block. + * 4. Swap the indices of the current and previous Huffman codes. + * + * Note: we never output UNCOMPRESSED blocks. This probably should be + * implemented sometime, but it doesn't make much difference. + */ +static void +lzx_flush_block(struct lzx_compressor *c, struct lzx_output_bitstream *os, + const u8 *block_begin, u32 block_size, u32 seq_idx) { - 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; -} + int block_type; -static unsigned -lzx_get_matches_fillcache_multiblock(struct lzx_compressor *c, - const struct lz_match **matches_ret) -{ - 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 { - num_matches = 0; - } - c->match_window_pos++; - *matches_ret = matches; - return num_matches; + lzx_build_huffman_codes(c); + + block_type = lzx_choose_verbatim_or_aligned(&c->freqs, + &c->codes[c->codes_index]); + lzx_write_compressed_block(block_begin, + block_type, + block_size, + c->window_order, + c->num_main_syms, + &c->chosen_sequences[seq_idx], + &c->codes[c->codes_index], + &c->codes[c->codes_index ^ 1].lens, + os); + c->codes_index ^= 1; } -static unsigned -lzx_get_matches_usecache(struct lzx_compressor *c, - const struct lz_match **matches_ret) +/******************************************************************************/ +/* Block splitting algorithm */ +/*----------------------------------------------------------------------------*/ + +/* + * The problem of block splitting is to decide when it is worthwhile to start a + * new block with new entropy codes. There is a theoretically optimal solution: + * recursively consider every possible block split, considering the exact cost + * of each block, and choose the minimum cost approach. But this is far too + * slow. Instead, as an approximation, we can count symbols and after every N + * symbols, compare the expected distribution of symbols based on the previous + * data with the actual distribution. If they differ "by enough", then start a + * new block. + * + * As an optimization and heuristic, we don't distinguish between every symbol + * but rather we combine many symbols into a single "observation type". For + * literals we only look at the high bits and low bits, and for matches we only + * look at whether the match is long or not. The assumption is that for typical + * "real" data, places that are good block boundaries will tend to be noticable + * based only on changes in these aggregate frequencies, without looking for + * subtle differences in individual symbols. For example, a change from ASCII + * bytes to non-ASCII bytes, or from few matches (generally less compressible) + * to many matches (generally more compressible), would be easily noticed based + * on the aggregates. + * + * For determining whether the frequency distributions are "different enough" to + * start a new block, the simply heuristic of splitting when the sum of absolute + * differences exceeds a constant seems to be good enough. + * + * Finally, for an approximation, it is not strictly necessary that the exact + * symbols being used are considered. With "near-optimal parsing", for example, + * the actual symbols that will be used are unknown until after the block + * boundary is chosen and the block has been optimized. Since the final choices + * cannot be used, we can use preliminary "greedy" choices instead. + */ + +/* Initialize the block split statistics when starting a new block. */ +static void +lzx_init_block_split_stats(struct lzx_block_split_stats *stats) { - struct lz_match *cache_ptr; - struct lz_match *matches; - unsigned num_matches; - - 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; + memset(stats, 0, sizeof(*stats)); } -static unsigned -lzx_get_matches_usecache_nocheck(struct lzx_compressor *c, - const struct lz_match **matches_ret) +/* Literal observation. Heuristic: use the top 2 bits and low 1 bits of the + * literal, for 8 possible literal observation types. */ +static inline void +lzx_observe_literal(struct lzx_block_split_stats *stats, u8 lit) { - struct lz_match *cache_ptr; - struct lz_match *matches; - unsigned num_matches; - - 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; + stats->new_observations[((lit >> 5) & 0x6) | (lit & 1)]++; + stats->num_new_observations++; } -static unsigned -lzx_get_matches_nocache_singleblock(struct lzx_compressor *c, - const struct lz_match **matches_ret) +/* Match observation. Heuristic: use one observation type for "short match" and + * one observation type for "long match". */ +static inline void +lzx_observe_match(struct lzx_block_split_stats *stats, unsigned length) { - 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; + stats->new_observations[NUM_LITERAL_OBSERVATION_TYPES + (length >= 5)]++; + stats->num_new_observations++; } -static unsigned -lzx_get_matches_nocache_multiblock(struct lzx_compressor *c, - const struct lz_match **matches_ret) +static bool +lzx_should_end_block(struct lzx_block_split_stats *stats) { - 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; + if (stats->num_observations > 0) { + + /* Note: to avoid slow divisions, we do not divide by + * 'num_observations', but rather do all math with the numbers + * multiplied by 'num_observations'. */ + u32 total_delta = 0; + for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) { + u32 expected = stats->observations[i] * + stats->num_new_observations; + u32 actual = stats->new_observations[i] * + stats->num_observations; + u32 delta = (actual > expected) ? actual - expected : + expected - actual; + total_delta += delta; + } + + /* Ready to end the block? */ + if (total_delta >= + stats->num_new_observations * 7 / 8 * stats->num_observations) + return true; + } + + for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) { + stats->num_observations += stats->new_observations[i]; + stats->observations[i] += stats->new_observations[i]; + stats->new_observations[i] = 0; + } + stats->num_new_observations = 0; + return false; } +/******************************************************************************/ +/* Slower ("near-optimal") compression algorithm */ +/*----------------------------------------------------------------------------*/ + /* - * Find matches at the next position in the window. + * Least-recently-used queue for match offsets. * - * 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. + * This is represented as a 64-bit integer for efficiency. There are three + * offsets of 21 bits each. Bit 64 is garbage. */ -static inline unsigned -lzx_get_matches(struct lzx_compressor *c, const struct lz_match **matches_ret) +struct lzx_lru_queue { + u64 R; +} _aligned_attribute(8); + +#define LZX_QUEUE_OFFSET_SHIFT 21 +#define LZX_QUEUE_OFFSET_MASK (((u64)1 << LZX_QUEUE_OFFSET_SHIFT) - 1) + +#define LZX_QUEUE_R0_SHIFT (0 * LZX_QUEUE_OFFSET_SHIFT) +#define LZX_QUEUE_R1_SHIFT (1 * LZX_QUEUE_OFFSET_SHIFT) +#define LZX_QUEUE_R2_SHIFT (2 * LZX_QUEUE_OFFSET_SHIFT) + +#define LZX_QUEUE_R0_MASK (LZX_QUEUE_OFFSET_MASK << LZX_QUEUE_R0_SHIFT) +#define LZX_QUEUE_R1_MASK (LZX_QUEUE_OFFSET_MASK << LZX_QUEUE_R1_SHIFT) +#define LZX_QUEUE_R2_MASK (LZX_QUEUE_OFFSET_MASK << LZX_QUEUE_R2_SHIFT) + +#define LZX_QUEUE_INITIALIZER { \ + ((u64)1 << LZX_QUEUE_R0_SHIFT) | \ + ((u64)1 << LZX_QUEUE_R1_SHIFT) | \ + ((u64)1 << LZX_QUEUE_R2_SHIFT) } + +static inline u64 +lzx_lru_queue_R0(struct lzx_lru_queue queue) { - return (*c->get_matches_func)(c, matches_ret); + return (queue.R >> LZX_QUEUE_R0_SHIFT) & LZX_QUEUE_OFFSET_MASK; } -static void -lzx_skip_bytes_fillcache(struct lzx_compressor *c, unsigned n) +static inline u64 +lzx_lru_queue_R1(struct lzx_lru_queue queue) { - struct lz_match *cache_ptr; - - 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; + return (queue.R >> LZX_QUEUE_R1_SHIFT) & LZX_QUEUE_OFFSET_MASK; } -static void -lzx_skip_bytes_usecache(struct lzx_compressor *c, unsigned n) +static inline u64 +lzx_lru_queue_R2(struct lzx_lru_queue queue) { - struct lz_match *cache_ptr; - - 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; + return (queue.R >> LZX_QUEUE_R2_SHIFT) & LZX_QUEUE_OFFSET_MASK; } -static void -lzx_skip_bytes_usecache_nocheck(struct lzx_compressor *c, unsigned n) +/* Push a match offset onto the front (most recently used) end of the queue. */ +static inline struct lzx_lru_queue +lzx_lru_queue_push(struct lzx_lru_queue queue, u32 offset) { - struct lz_match *cache_ptr; + return (struct lzx_lru_queue) { + .R = (queue.R << LZX_QUEUE_OFFSET_SHIFT) | offset, + }; +} - cache_ptr = c->cache_ptr; - c->match_window_pos += n; - do { - cache_ptr += 1 + cache_ptr->len; - } while (--n); - c->cache_ptr = cache_ptr; +/* Swap a match offset to the front of the queue. */ +static inline struct lzx_lru_queue +lzx_lru_queue_swap(struct lzx_lru_queue queue, unsigned idx) +{ + unsigned shift = idx * 21; + const u64 mask = LZX_QUEUE_R0_MASK; + const u64 mask_high = mask << shift; + + return (struct lzx_lru_queue) { + (queue.R & ~(mask | mask_high)) | + ((queue.R & mask_high) >> shift) | + ((queue.R & mask) << shift) + }; } -static void -lzx_skip_bytes_nocache(struct lzx_compressor *c, unsigned n) +static inline u32 +lzx_walk_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit, + bool record) { - c->match_window_pos += n; - lz_mf_skip_positions(c->mf, n); + u32 node_idx = block_size; + u32 seq_idx = ARRAY_LEN(c->chosen_sequences) - 1; + u32 lit_start_node; + + if (record) { + /* Special value to mark last sequence */ + c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = 0x80000000; + lit_start_node = node_idx; + } + + for (;;) { + u32 item; + u32 len; + u32 adjusted_offset; + unsigned v; + unsigned offset_slot; + + /* Tally literals until either a match or the beginning of the + * block is reached. Note: the item in the node at the + * beginning of the block has all bits set, causing this loop to + * end when it is reached. */ + for (;;) { + item = c->optimum_nodes[node_idx].item; + if (item & OPTIMUM_LEN_MASK) + break; + c->freqs.main[item >> OPTIMUM_OFFSET_SHIFT]++; + node_idx--; + } + + #if CONSIDER_GAP_MATCHES + if (item & OPTIMUM_GAP_MATCH) { + + if (node_idx == 0) + break; + + /* Record the literal run length for the next sequence + * (the "previous sequence" when walking backwards). */ + len = item & OPTIMUM_LEN_MASK; + if (record) { + c->chosen_sequences[seq_idx--].litrunlen = + lit_start_node - node_idx; + lit_start_node = node_idx - len; + } + + /* Tally the rep0 match after the gap. */ + v = len - LZX_MIN_MATCH_LEN; + if (record) + c->chosen_sequences[seq_idx].adjusted_length = v; + if (v >= LZX_NUM_PRIMARY_LENS) { + c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++; + v = LZX_NUM_PRIMARY_LENS; + } + c->freqs.main[LZX_NUM_CHARS + v]++; + if (record) + c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = v; + + /* Tally the literal in the gap. */ + c->freqs.main[(u8)(item >> OPTIMUM_OFFSET_SHIFT)]++; + + /* Fall through and tally the match before the gap. + * (It was temporarily saved in the 'cost' field of the + * previous node, which was free to reuse.) */ + item = c->optimum_nodes[--node_idx].cost; + node_idx -= len; + } + #else /* CONSIDER_GAP_MATCHES */ + if (node_idx == 0) + break; + #endif /* !CONSIDER_GAP_MATCHES */ + + len = item & OPTIMUM_LEN_MASK; + adjusted_offset = item >> OPTIMUM_OFFSET_SHIFT; + + /* Record the literal run length for the next sequence (the + * "previous sequence" when walking backwards). */ + if (record) { + c->chosen_sequences[seq_idx--].litrunlen = + lit_start_node - node_idx; + node_idx -= len; + lit_start_node = node_idx; + } else { + node_idx -= len; + } + + /* Record a match. */ + + /* Tally the aligned offset symbol if needed. */ + if (adjusted_offset >= 16) + c->freqs.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]++; + + /* Record the adjusted length. */ + v = len - LZX_MIN_MATCH_LEN; + if (record) + c->chosen_sequences[seq_idx].adjusted_length = v; + + /* Tally the length symbol if needed. */ + if (v >= LZX_NUM_PRIMARY_LENS) { + c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++; + v = LZX_NUM_PRIMARY_LENS; + } + + /* Tally the main symbol. */ + offset_slot = lzx_get_offset_slot(c, adjusted_offset, is_16_bit); + v += offset_slot * LZX_NUM_LEN_HEADERS; + c->freqs.main[LZX_NUM_CHARS + v]++; + + /* Record the adjusted offset and match header. */ + if (record) { + c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = + (adjusted_offset << 9) | v; + } + } + + /* Record the literal run length for the first sequence. */ + if (record) + c->chosen_sequences[seq_idx].litrunlen = lit_start_node - node_idx; + + /* Return the index in chosen_sequences at which the sequences begin. */ + return seq_idx; } /* - * Skip the specified number of positions in the window (don't search for - * matches at them). + * Given the minimum-cost path computed through the item graph for the current + * block, walk the path and count how many of each symbol in each Huffman-coded + * alphabet would be required to output the items (matches and literals) along + * the path. * - * This uses a wrapper function around the underlying match-finder. + * Note that the path will be walked backwards (from the end of the block to the + * beginning of the block), but this doesn't matter because this function only + * computes frequencies. */ static inline void -lzx_skip_bytes(struct lzx_compressor *c, unsigned n) +lzx_tally_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit) { - return (*c->skip_bytes_func)(c, n); + lzx_walk_item_list(c, block_size, is_16_bit, false); } -/* 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) +/* + * Like lzx_tally_item_list(), but this function also generates the list of + * lzx_sequences for the minimum-cost path and writes it to c->chosen_sequences, + * ready to be output to the bitstream after the Huffman codes are computed. + * The lzx_sequences will be written to decreasing memory addresses as the path + * is walked backwards, which means they will end up in the expected + * first-to-last order. The return value is the index in c->chosen_sequences at + * which the lzx_sequences begin. + */ +static inline u32 +lzx_record_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit) { - unsigned main_symbol = literal; - - c->freqs.main[main_symbol]++; - - if (next_chosen_item) { - *(*next_chosen_item)++ = (struct lzx_item) { - .data = main_symbol, - }; - } + return lzx_walk_item_list(c, block_size, is_16_bit, true); } -/* 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) +/* + * Find an inexpensive path through the graph of possible match/literal choices + * for the current block. The nodes of the graph are + * c->optimum_nodes[0...block_size]. They correspond directly to the bytes in + * the current block, plus one extra node for end-of-block. The edges of the + * graph are matches and literals. The goal is to find the minimum cost path + * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]', given the cost + * model 'c->costs'. + * + * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and + * proceeding forwards one node at a time. At each node, a selection of matches + * (len >= 2), as well as the literal byte (len = 1), is considered. An item of + * length 'len' provides a new path to reach the node 'len' bytes later. If + * such a path is the lowest cost found so far to reach that later node, then + * that later node is updated with the new cost and the "arrival" which provided + * that cost. + * + * Note that although this algorithm is based on minimum cost path search, due + * to various simplifying assumptions the result is not guaranteed to be the + * true minimum cost, or "optimal", path over the graph of all valid LZX + * representations of this block. + * + * Also, note that because of the presence of the recent offsets queue (which is + * a type of adaptive state), the algorithm cannot work backwards and compute + * "cost to end" instead of "cost to beginning". Furthermore, the way the + * algorithm handles 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. The algorithm does not solve this problem in general; it only looks + * one step ahead, with the exception of special consideration for "gap + * matches". + */ +static inline struct lzx_lru_queue +lzx_find_min_cost_path(struct lzx_compressor * const restrict c, + const u8 * const restrict block_begin, + const u32 block_size, + const struct lzx_lru_queue initial_queue, + bool is_16_bit) { - unsigned len_header; - unsigned main_symbol; - unsigned len_symbol; + struct lzx_optimum_node *cur_node = c->optimum_nodes; + struct lzx_optimum_node * const end_node = cur_node + block_size; + struct lz_match *cache_ptr = c->match_cache; + const u8 *in_next = block_begin; + const u8 * const block_end = block_begin + block_size; + + /* + * Instead of storing the match offset LRU queues in the + * 'lzx_optimum_node' structures, we save memory (and cache lines) by + * storing them in a smaller array. This works because the algorithm + * only requires a limited history of the adaptive state. Once a given + * state is more than LZX_MAX_MATCH_LEN bytes behind the current node + * (more if gap match consideration is enabled; we just round up to 512 + * so it's a power of 2), it is no longer needed. + * + * The QUEUE() macro finds the queue for the given node. This macro has + * been optimized by taking advantage of 'struct lzx_lru_queue' and + * 'struct lzx_optimum_node' both being 8 bytes in size and alignment. + */ + struct lzx_lru_queue queues[512]; + STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1); + STATIC_ASSERT(sizeof(c->optimum_nodes[0]) == sizeof(queues[0])); +#define QUEUE(node) \ + (*(struct lzx_lru_queue *)((char *)queues + \ + ((uintptr_t)(node) % (ARRAY_LEN(queues) * sizeof(queues[0]))))) + /*(queues[(uintptr_t)(node) / sizeof(*(node)) % ARRAY_LEN(queues)])*/ + +#if CONSIDER_GAP_MATCHES + u32 matches_before_gap[ARRAY_LEN(queues)]; +#define MATCH_BEFORE_GAP(node) \ + (matches_before_gap[(uintptr_t)(node) / sizeof(*(node)) % \ + ARRAY_LEN(matches_before_gap)]) +#endif - 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]++; - } + /* + * Initially, the cost to reach each node is "infinity". + * + * The first node actually should have cost 0, but "infinity" + * (0xFFFFFFFF) works just as well because it immediately overflows. + * + * The following statement also intentionally sets the 'item' of the + * first node, which would otherwise have no meaning, to 0xFFFFFFFF for + * use as a sentinel. See lzx_walk_item_list(). + */ + memset(c->optimum_nodes, 0xFF, + (block_size + 1) * sizeof(c->optimum_nodes[0])); - main_symbol = LZX_NUM_CHARS + ((rep_index << 3) | len_header); + /* Initialize the recent offsets queue for the first node. */ + QUEUE(cur_node) = initial_queue; - c->freqs.main[main_symbol]++; + do { /* For each node in the block in position order... */ - if (next_chosen_item) { - *(*next_chosen_item)++ = (struct lzx_item) { - .data = (u64)main_symbol | ((u64)len_symbol << 10), - }; - } -} + unsigned num_matches; + unsigned literal; + u32 cost; -/* 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; + /* + * A selection of matches for the block was already saved in + * memory so that we don't have to run the uncompressed data + * through the matchfinder on every optimization pass. However, + * we still search for repeat offset matches during each + * optimization pass because we cannot predict the state of the + * recent offsets queue. But as a heuristic, we don't bother + * searching for repeat offset matches if the general-purpose + * matchfinder failed to find any matches. + * + * Note that a match of length n at some offset implies there is + * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at + * that same offset. In other words, we don't necessarily need + * to use the full length of a match. The key heuristic that + * saves a significicant amount of time is that for each + * distinct length, we only consider the smallest offset for + * which that length is available. This heuristic also applies + * to repeat offsets, which we order specially: R0 < R1 < R2 < + * any explicit offset. Of course, this heuristic may be + * produce suboptimal results because offset slots in LZX are + * subject to entropy encoding, but in practice this is a useful + * heuristic. + */ - 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]++; - } + num_matches = cache_ptr->length; + cache_ptr++; - offset_slot = lzx_get_offset_slot_raw(offset + LZX_OFFSET_OFFSET); + if (num_matches) { + struct lz_match *end_matches = cache_ptr + num_matches; + unsigned next_len = LZX_MIN_MATCH_LEN; + unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN); + const u8 *matchptr; + + /* Consider rep0 matches. */ + matchptr = in_next - lzx_lru_queue_R0(QUEUE(cur_node)); + if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next)) + goto rep0_done; + STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2); + do { + u32 cost = cur_node->cost + + c->costs.match_cost[0][ + next_len - LZX_MIN_MATCH_LEN]; + if (cost <= (cur_node + next_len)->cost) { + (cur_node + next_len)->cost = cost; + (cur_node + next_len)->item = + (0 << OPTIMUM_OFFSET_SHIFT) | next_len; + } + if (unlikely(++next_len > max_len)) { + cache_ptr = end_matches; + goto done_matches; + } + } while (in_next[next_len - 1] == matchptr[next_len - 1]); + + rep0_done: + + /* Consider rep1 matches. */ + matchptr = in_next - lzx_lru_queue_R1(QUEUE(cur_node)); + if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next)) + goto rep1_done; + if (matchptr[next_len - 1] != in_next[next_len - 1]) + goto rep1_done; + for (unsigned len = 2; len < next_len - 1; len++) + if (matchptr[len] != in_next[len]) + goto rep1_done; + do { + u32 cost = cur_node->cost + + c->costs.match_cost[1][ + next_len - LZX_MIN_MATCH_LEN]; + if (cost <= (cur_node + next_len)->cost) { + (cur_node + next_len)->cost = cost; + (cur_node + next_len)->item = + (1 << OPTIMUM_OFFSET_SHIFT) | next_len; + } + if (unlikely(++next_len > max_len)) { + cache_ptr = end_matches; + goto done_matches; + } + } while (in_next[next_len - 1] == matchptr[next_len - 1]); + + rep1_done: + + /* Consider rep2 matches. */ + matchptr = in_next - lzx_lru_queue_R2(QUEUE(cur_node)); + if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next)) + goto rep2_done; + if (matchptr[next_len - 1] != in_next[next_len - 1]) + goto rep2_done; + for (unsigned len = 2; len < next_len - 1; len++) + if (matchptr[len] != in_next[len]) + goto rep2_done; + do { + u32 cost = cur_node->cost + + c->costs.match_cost[2][ + next_len - LZX_MIN_MATCH_LEN]; + if (cost <= (cur_node + next_len)->cost) { + (cur_node + next_len)->cost = cost; + (cur_node + next_len)->item = + (2 << OPTIMUM_OFFSET_SHIFT) | next_len; + } + if (unlikely(++next_len > max_len)) { + cache_ptr = end_matches; + goto done_matches; + } + } while (in_next[next_len - 1] == matchptr[next_len - 1]); + + rep2_done: + + while (next_len > cache_ptr->length) + if (++cache_ptr == end_matches) + goto done_matches; + + /* Consider explicit offset matches. */ + for (;;) { + u32 offset = cache_ptr->offset; + u32 adjusted_offset = offset + LZX_OFFSET_ADJUSTMENT; + unsigned offset_slot = lzx_get_offset_slot(c, adjusted_offset, is_16_bit); + u32 base_cost = cur_node->cost; + u32 cost; + + #if CONSIDER_ALIGNED_COSTS + if (offset >= 16 - LZX_OFFSET_ADJUSTMENT) + base_cost += c->costs.aligned[adjusted_offset & + LZX_ALIGNED_OFFSET_BITMASK]; + #endif + do { + cost = base_cost + + c->costs.match_cost[offset_slot][ + next_len - LZX_MIN_MATCH_LEN]; + if (cost < (cur_node + next_len)->cost) { + (cur_node + next_len)->cost = cost; + (cur_node + next_len)->item = + (adjusted_offset << OPTIMUM_OFFSET_SHIFT) | next_len; + } + } while (++next_len <= cache_ptr->length); + + if (++cache_ptr == end_matches) { + #if CONSIDER_GAP_MATCHES + /* Also consider the longest explicit + * offset match as a "gap match": match + * + lit + rep0. */ + s32 remaining = (block_end - in_next) - (s32)next_len; + if (likely(remaining >= 2)) { + const u8 *strptr = in_next + next_len; + const u8 *matchptr = strptr - offset; + if (load_u16_unaligned(strptr) == load_u16_unaligned(matchptr)) { + STATIC_ASSERT(ARRAY_LEN(queues) - LZX_MAX_MATCH_LEN - 2 >= 250); + STATIC_ASSERT(ARRAY_LEN(queues) == ARRAY_LEN(matches_before_gap)); + unsigned limit = min(remaining, + min(ARRAY_LEN(queues) - LZX_MAX_MATCH_LEN - 2, + LZX_MAX_MATCH_LEN)); + unsigned rep0_len = lz_extend(strptr, matchptr, 2, limit); + u8 lit = strptr[-1]; + cost += c->costs.main[lit] + + c->costs.match_cost[0][rep0_len - LZX_MIN_MATCH_LEN]; + unsigned total_len = next_len + rep0_len; + if (cost < (cur_node + total_len)->cost) { + (cur_node + total_len)->cost = cost; + (cur_node + total_len)->item = + OPTIMUM_GAP_MATCH | + ((u32)lit << OPTIMUM_OFFSET_SHIFT) | + rep0_len; + MATCH_BEFORE_GAP(cur_node + total_len) = + (adjusted_offset << OPTIMUM_OFFSET_SHIFT) | + (next_len - 1); + } + } + } + #endif /* CONSIDER_GAP_MATCHES */ + break; + } + } + } - main_symbol = LZX_NUM_CHARS + ((offset_slot << 3) | len_header); + done_matches: - c->freqs.main[main_symbol]++; + /* Consider coding a literal. - if (offset_slot >= 8) - c->freqs.aligned[(offset + LZX_OFFSET_OFFSET) & 7]++; + * To avoid an extra branch, actually checking the preferability + * of coding the literal is integrated into the queue update + * code below. */ + literal = *in_next++; + cost = cur_node->cost + c->costs.main[literal]; - if (next_chosen_item) { + /* Advance to the next position. */ + cur_node++; - num_extra_bits = lzx_extra_offset_bits[offset_slot]; + /* The lowest-cost path to the current position is now known. + * Finalize the recent offsets queue that results from taking + * this lowest-cost path. */ - extra_bits = (offset + LZX_OFFSET_OFFSET) - - lzx_offset_slot_base[offset_slot]; + if (cost <= cur_node->cost) { + /* Literal: queue remains unchanged. */ + cur_node->cost = cost; + cur_node->item = (u32)literal << OPTIMUM_OFFSET_SHIFT; + QUEUE(cur_node) = QUEUE(cur_node - 1); + } else { + /* Match: queue update is needed. */ + unsigned len = cur_node->item & OPTIMUM_LEN_MASK; + #if CONSIDER_GAP_MATCHES + s32 adjusted_offset = (s32)cur_node->item >> OPTIMUM_OFFSET_SHIFT; + STATIC_ASSERT(OPTIMUM_GAP_MATCH == 0x80000000); /* assuming sign extension */ + #else + u32 adjusted_offset = cur_node->item >> OPTIMUM_OFFSET_SHIFT; + #endif + + if (adjusted_offset >= LZX_NUM_RECENT_OFFSETS) { + /* Explicit offset match: insert offset at front. */ + QUEUE(cur_node) = + lzx_lru_queue_push(QUEUE(cur_node - len), + adjusted_offset - LZX_OFFSET_ADJUSTMENT); + } + #if CONSIDER_GAP_MATCHES + else if (adjusted_offset < 0) { + /* "Gap match": Explicit offset match, then a + * literal, then rep0 match. Save the explicit + * offset match information in the cost field of + * the previous node, which isn't needed + * anymore. Then insert the offset at the front + * of the queue. */ + u32 match_before_gap = MATCH_BEFORE_GAP(cur_node); + (cur_node - 1)->cost = match_before_gap; + QUEUE(cur_node) = + lzx_lru_queue_push(QUEUE(cur_node - len - 1 - + (match_before_gap & OPTIMUM_LEN_MASK)), + (match_before_gap >> OPTIMUM_OFFSET_SHIFT) - + LZX_OFFSET_ADJUSTMENT); + } + #endif + else { + /* Repeat offset match: swap offset to front. */ + QUEUE(cur_node) = + lzx_lru_queue_swap(QUEUE(cur_node - len), + adjusted_offset); + } + } + } while (cur_node != end_node); - *(*next_chosen_item)++ = (struct lzx_item) { - .data = (u64)main_symbol | - ((u64)len_symbol << 10) | - ((u64)num_extra_bits << 18) | - ((u64)extra_bits << 23), - }; - } + /* Return the recent offsets queue at the end of the path. */ + return QUEUE(cur_node); } -/* 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) +/* + * Given the costs for the main and length codewords (c->costs.main and + * c->costs.len), initialize the match cost array (c->costs.match_cost) which + * directly provides the cost of every possible (length, offset slot) pair. + */ +static void +lzx_compute_match_costs(struct lzx_compressor *c) { - 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); -} + unsigned num_offset_slots = (c->num_main_syms - LZX_NUM_CHARS) / + LZX_NUM_LEN_HEADERS; + struct lzx_costs *costs = &c->costs; + unsigned main_symbol = LZX_NUM_CHARS; + + for (unsigned offset_slot = 0; offset_slot < num_offset_slots; + offset_slot++) + { + u32 extra_cost = lzx_extra_offset_bits[offset_slot] * BIT_COST; + unsigned i; + + #if CONSIDER_ALIGNED_COSTS + if (offset_slot >= 8) + extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * BIT_COST; + #endif + + for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++) { + costs->match_cost[offset_slot][i] = + costs->main[main_symbol++] + extra_cost; + } -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; + extra_cost += costs->main[main_symbol++]; - /* 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); + for (; i < LZX_NUM_LENS; i++) { + costs->match_cost[offset_slot][i] = + costs->len[i - LZX_NUM_PRIMARY_LENS] + + extra_cost; + } + } } -static inline void -lzx_tally_item_list(struct lzx_compressor *c, struct lzx_mc_pos_data *cur_optimum_ptr) +/* + * Fast approximation for log2f(x). This is not as accurate as the standard C + * version. It does not need to be perfectly accurate because it is only used + * for estimating symbol costs, which is very approximate anyway. + */ +static float +log2f_fast(float x) { - /* 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); + union { + float f; + s32 i; + } u = { .f = x }; + + /* Extract the exponent and subtract 127 to remove the bias. This gives + * the integer part of the result. */ + float res = ((u.i >> 23) & 0xFF) - 127; + + /* Set the exponent to 0 (plus bias of 127). This transforms the number + * to the range [1, 2) while retaining the same mantissa. */ + u.i = (u.i & ~(0xFF << 23)) | (127 << 23); + + /* + * Approximate the log2 of the transformed number using a degree 2 + * interpolating polynomial for log2(x) over the interval [1, 2). Then + * add this to the extracted exponent to produce the final approximation + * of log2(x). + * + * The coefficients of the interpolating polynomial used here were found + * using the script tools/log2_interpolation.r. + */ + return res - 1.653124006f + u.f * (1.9941812f - u.f * 0.3347490189f); + } -/* 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) +/* + * Return the estimated cost of a symbol which has been estimated to have the + * given probability. + */ +static u32 +lzx_cost_for_probability(float prob) { - if (next_chosen_item) - lzx_record_item_list(c, cur_optimum_ptr, next_chosen_item); - else - lzx_tally_item_list(c, cur_optimum_ptr); + /* + * The basic formula is: + * + * entropy = -log2(probability) + * + * Use this to get the cost in fractional bits. Then multiply by our + * scaling factor of BIT_COST and truncate to a u32. + * + * In addition, the minimum cost is BIT_COST (one bit) because the + * entropy coding method will be Huffman codes. + */ + u32 cost = -log2f_fast(prob) * BIT_COST; + return max(cost, BIT_COST); } -/* Set the cost model @c->costs from the Huffman codeword lengths specified in - * @lens. +/* + * Mapping: number of used literals => heuristic probability of a literal times + * 6870. Generated by running this R command: * - * The cost model and codeword lengths are almost the same thing, but the - * Huffman codewords with length 0 correspond to symbols with zero frequency - * that still need to be assigned actual costs. The specific values assigned - * are arbitrary, but they should be fairly high (near the maximum codeword - * 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 *c, const struct lzx_lens * lens) -{ - unsigned i; - - /* Main code */ - 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++) - c->costs.len[i] = lens->len[i] ? lens->len[i] : 15; + * cat(paste(round(6870*2^-((304+(0:256))/64)), collapse=", ")) + */ +static const u8 literal_scaled_probs[257] = { + 255, 253, 250, 247, 244, 242, 239, 237, 234, 232, 229, 227, 224, 222, + 219, 217, 215, 212, 210, 208, 206, 203, 201, 199, 197, 195, 193, 191, + 189, 186, 184, 182, 181, 179, 177, 175, 173, 171, 169, 167, 166, 164, + 162, 160, 159, 157, 155, 153, 152, 150, 149, 147, 145, 144, 142, 141, + 139, 138, 136, 135, 133, 132, 130, 129, 128, 126, 125, 124, 122, 121, + 120, 118, 117, 116, 115, 113, 112, 111, 110, 109, 107, 106, 105, 104, + 103, 102, 101, 100, 98, 97, 96, 95, 94, 93, 92, 91, 90, 89, 88, 87, 86, + 86, 85, 84, 83, 82, 81, 80, 79, 78, 78, 77, 76, 75, 74, 73, 73, 72, 71, + 70, 70, 69, 68, 67, 67, 66, 65, 65, 64, 63, 62, 62, 61, 60, 60, 59, 59, + 58, 57, 57, 56, 55, 55, 54, 54, 53, 53, 52, 51, 51, 50, 50, 49, 49, 48, + 48, 47, 47, 46, 46, 45, 45, 44, 44, 43, 43, 42, 42, 41, 41, 40, 40, 40, + 39, 39, 38, 38, 38, 37, 37, 36, 36, 36, 35, 35, 34, 34, 34, 33, 33, 33, + 32, 32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 27, 27, 27, + 27, 26, 26, 26, 25, 25, 25, 25, 24, 24, 24, 24, 23, 23, 23, 23, 22, 22, + 22, 22, 21, 21, 21, 21, 20, 20, 20, 20, 20, 19, 19, 19, 19, 19, 18, 18, + 18, 18, 18, 17, 17, 17, 17, 17, 16, 16, 16, 16 +}; - /* Aligned offset code */ - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) - c->costs.aligned[i] = lens->aligned[i] ? lens->aligned[i] : 7; -} +/* + * Mapping: length symbol => default cost of that symbol. This is derived from + * sample data but has been slightly edited to add more bias towards the + * shortest lengths, which are the most common. + */ +static const u16 lzx_default_len_costs[LZX_LENCODE_NUM_SYMBOLS] = { + 300, 310, 320, 330, 360, 396, 399, 416, 451, 448, 463, 466, 505, 492, + 503, 514, 547, 531, 566, 561, 589, 563, 592, 586, 623, 602, 639, 627, + 659, 643, 657, 650, 685, 662, 661, 672, 685, 686, 696, 680, 657, 682, + 666, 699, 674, 699, 679, 709, 688, 712, 692, 714, 694, 716, 698, 712, + 706, 727, 714, 727, 713, 723, 712, 718, 719, 719, 720, 735, 725, 735, + 728, 740, 727, 739, 727, 742, 716, 733, 733, 740, 738, 746, 737, 747, + 738, 745, 736, 748, 742, 749, 745, 749, 743, 748, 741, 752, 745, 752, + 747, 750, 747, 752, 748, 753, 750, 752, 753, 753, 749, 744, 752, 755, + 753, 756, 745, 748, 746, 745, 723, 757, 755, 758, 755, 758, 752, 757, + 754, 757, 755, 759, 755, 758, 753, 755, 755, 758, 757, 761, 755, 750, + 758, 759, 759, 760, 758, 751, 757, 757, 759, 759, 758, 759, 758, 761, + 750, 761, 758, 760, 759, 761, 758, 761, 760, 752, 759, 760, 759, 759, + 757, 762, 760, 761, 761, 748, 761, 760, 762, 763, 752, 762, 762, 763, + 762, 762, 763, 763, 762, 763, 762, 763, 762, 763, 763, 764, 763, 762, + 763, 762, 762, 762, 764, 764, 763, 764, 763, 763, 763, 762, 763, 763, + 762, 764, 764, 763, 762, 763, 763, 763, 763, 762, 764, 763, 762, 764, + 764, 763, 763, 765, 764, 764, 762, 763, 764, 765, 763, 764, 763, 764, + 762, 764, 764, 754, 763, 764, 763, 763, 762, 763, 584, +}; -/* Set default LZX Huffman symbol costs to bootstrap the iterative optimization - * algorithm. */ +/* Set default costs to bootstrap the iterative optimization algorithm. */ static void -lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms) +lzx_set_default_costs(struct lzx_compressor *c) { unsigned i; + u32 num_literals = 0; + u32 num_used_literals = 0; + float inv_num_matches = 1.0f / c->freqs.main[LZX_NUM_CHARS]; + float inv_num_items; + float prob_match = 1.0f; + u32 match_cost; + float base_literal_prob; + + /* Some numbers here have been hardcoded to assume a bit cost of 64. */ + STATIC_ASSERT(BIT_COST == 64); + + /* Estimate the number of literals that will used. 'num_literals' is + * the total number, whereas 'num_used_literals' is the number of + * distinct symbols. */ + for (i = 0; i < LZX_NUM_CHARS; i++) { + num_literals += c->freqs.main[i]; + num_used_literals += (c->freqs.main[i] != 0); + } - /* Main code (part 1): Literal symbols */ - for (i = 0; i < LZX_NUM_CHARS; i++) - costs->main[i] = 8; - - /* Main code (part 2): Match header symbols */ - for (; i < num_main_syms; i++) - costs->main[i] = 10; + /* Note: all match headers were tallied as symbol 'LZX_NUM_CHARS'. We + * don't attempt to estimate which ones will be used. */ + + inv_num_items = 1.0f / (num_literals + c->freqs.main[LZX_NUM_CHARS]); + base_literal_prob = literal_scaled_probs[num_used_literals] * + (1.0f / 6870.0f); + + /* Literal costs. We use two different methods to compute the + * probability of each literal and mix together their results. */ + for (i = 0; i < LZX_NUM_CHARS; i++) { + u32 freq = c->freqs.main[i]; + if (freq != 0) { + float prob = 0.5f * ((freq * inv_num_items) + + base_literal_prob); + c->costs.main[i] = lzx_cost_for_probability(prob); + prob_match -= prob; + } else { + c->costs.main[i] = 11 * BIT_COST; + } + } - /* Length code */ + /* Match header costs. We just assume that all match headers are + * equally probable, but we do take into account the relative cost of a + * match header vs. a literal depending on how common matches are + * expected to be vs. literals. */ + prob_match = max(prob_match, 0.15f); + match_cost = lzx_cost_for_probability(prob_match / (c->num_main_syms - + LZX_NUM_CHARS)); + for (; i < c->num_main_syms; i++) + c->costs.main[i] = match_cost; + + /* Length symbol costs. These are just set to fixed values which + * reflect the fact the smallest lengths are typically the most common, + * and therefore are typically the cheapest. */ 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]; + c->costs.len[i] = lzx_default_len_costs[i]; + +#if CONSIDER_ALIGNED_COSTS + /* Aligned offset symbol costs. These are derived from the estimated + * probability of each aligned offset symbol. */ + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + /* We intentionally tallied the frequencies in the wrong slots, + * not accounting for LZX_OFFSET_ADJUSTMENT, since doing the + * fixup here is faster: a constant 8 subtractions here vs. one + * addition for every match. */ + unsigned j = (i - LZX_OFFSET_ADJUSTMENT) & LZX_ALIGNED_OFFSET_BITMASK; + if (c->freqs.aligned[j] != 0) { + float prob = c->freqs.aligned[j] * inv_num_matches; + c->costs.aligned[i] = lzx_cost_for_probability(prob); + } else { + c->costs.aligned[i] = + (2 * LZX_NUM_ALIGNED_OFFSET_BITS) * BIT_COST; + } + } +#endif } -/* 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) +/* Update the current cost model to reflect the computed Huffman codes. */ +static void +lzx_set_costs_from_codes(struct lzx_compressor *c) { - 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 { - len_header = LZX_NUM_PRIMARY_LENS; + unsigned i; + const struct lzx_lens *lens = &c->codes[c->codes_index].lens; - /* Account for length symbol. */ - cost = costs->len[len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS]; + for (i = 0; i < c->num_main_syms; i++) { + c->costs.main[i] = (lens->main[i] ? lens->main[i] : + MAIN_CODEWORD_LIMIT) * BIT_COST; } - /* Account for main symbol. */ - main_symbol = LZX_NUM_CHARS + ((offset_slot << 3) | len_header); - cost += costs->main[main_symbol]; - - return cost; -} + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) { + c->costs.len[i] = (lens->len[i] ? lens->len[i] : + LENGTH_CODEWORD_LIMIT) * BIT_COST; + } -/* 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))]; +#if CONSIDER_ALIGNED_COSTS + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : + ALIGNED_CODEWORD_LIMIT) * BIT_COST; + } +#endif } /* - * Consider coding the match at repeat offset index @rep_idx. Consider each - * length from the minimum (2) to the full match length (@rep_len). + * Choose a "near-optimal" literal/match sequence to use for the current block, + * then flush the block. Because the cost of each Huffman symbol is unknown + * until the Huffman codes have been built and the Huffman codes themselves + * depend on the symbol frequencies, this uses an iterative optimization + * algorithm to approximate an optimal solution. The first optimization pass + * for the block uses default costs; additional passes use costs derived from + * the Huffman codes computed in the previous pass. */ -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) +static inline struct lzx_lru_queue +lzx_optimize_and_flush_block(struct lzx_compressor * const restrict c, + struct lzx_output_bitstream * const restrict os, + const u8 * const restrict block_begin, + const u32 block_size, + const struct lzx_lru_queue initial_queue, + bool is_16_bit) { - u32 base_cost = cur_optimum_ptr->cost; - u32 cost; - unsigned len; + unsigned num_passes_remaining = c->num_optim_passes; + struct lzx_lru_queue new_queue; + u32 seq_idx; -#if 1 /* Optimized version */ + lzx_set_default_costs(c); - 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); + for (;;) { + lzx_compute_match_costs(c); + new_queue = lzx_find_min_cost_path(c, block_begin, block_size, + initial_queue, is_16_bit); - /* 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 (--num_passes_remaining == 0) + break; -#else /* Unoptimized version */ + /* At least one optimization pass remains. Update the costs. */ + lzx_reset_symbol_frequencies(c); + lzx_tally_item_list(c, block_size, is_16_bit); + lzx_build_huffman_codes(c); + lzx_set_costs_from_codes(c); + } - 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 + /* Done optimizing. Generate the sequence list and flush the block. */ + lzx_reset_symbol_frequencies(c); + seq_idx = lzx_record_item_list(c, block_size, is_16_bit); + lzx_flush_block(c, os, block_begin, block_size, seq_idx); + return new_queue; } /* - * Consider coding each match in @matches as an explicit offset match. + * This is the "near-optimal" LZX compressor. * - * @matches must be sorted by strictly increasing length and strictly - * increasing offset. This is guaranteed by the match-finder. + * For each block, it performs a relatively thorough graph search to find an + * inexpensive (in terms of compressed size) way to output the block. * - * 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. + * Note: there are actually many things this algorithm leaves on the table in + * terms of compression ratio. So although it may be "near-optimal", it is + * certainly not "optimal". The goal is not to produce the optimal compression + * ratio, which for LZX is probably impossible within any practical amount of + * time, but rather to produce a compression ratio significantly better than a + * simpler "greedy" or "lazy" parse while still being relatively fast. */ 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) +lzx_compress_near_optimal(struct lzx_compressor * restrict c, + const u8 * const restrict in_begin, size_t in_nbytes, + struct lzx_output_bitstream * restrict os, + bool is_16_bit) { - LZX_ASSERT(num_matches > 0); - - unsigned i; - unsigned len; - unsigned offset_slot; - u32 position_cost; - u32 cost; - u32 offset_data; + const u8 * in_next = in_begin; + const u8 * const in_end = in_begin + in_nbytes; + u32 max_len = LZX_MAX_MATCH_LEN; + u32 nice_len = min(c->nice_match_length, max_len); + u32 next_hashes[2] = {0, 0}; + struct lzx_lru_queue queue = LZX_QUEUE_INITIALIZER; + /* Initialize the matchfinder. */ + CALL_BT_MF(is_16_bit, c, bt_matchfinder_init); -#if 1 /* Optimized version */ - - if (matches[num_matches - 1].offset < LZX_NUM_FAST_OFFSETS) { + do { + /* Starting a new block */ + + const u8 * const in_block_begin = in_next; + const u8 * const in_max_block_end = + in_next + min(SOFT_MAX_BLOCK_SIZE, in_end - in_next); + struct lz_match *cache_ptr = c->match_cache; + const u8 *next_search_pos = in_next; + const u8 *next_observation = in_next; + const u8 *next_pause_point = + min(in_next + min(MIN_BLOCK_SIZE, + in_max_block_end - in_next), + in_max_block_end - min(LZX_MAX_MATCH_LEN - 1, + in_max_block_end - in_next)); + + lzx_init_block_split_stats(&c->split_stats); + lzx_reset_symbol_frequencies(c); + + if (in_next >= next_pause_point) + goto pause; /* - * 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. + * Run the input buffer through the matchfinder, caching the + * matches, until we decide to end the block. * - * - 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. + * For a tighter matchfinding loop, we compute a "pause point", + * which is the next position at which we may need to check + * whether to end the block or to decrease max_len. We then + * only do these extra checks upon reaching the pause point. */ - - LZX_ASSERT(offset_slot <= 37); /* for extra bits formula */ - - len = 2; - i = 0; + resume_matchfinding: 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; + if (in_next >= next_search_pos) { + /* Search for matches at this position. */ + struct lz_match *lz_matchptr; + u32 best_len; + + lz_matchptr = CALL_BT_MF(is_16_bit, c, + bt_matchfinder_get_matches, + in_begin, + in_next - in_begin, + max_len, + nice_len, + c->max_search_depth, + next_hashes, + &best_len, + cache_ptr + 1); + cache_ptr->length = lz_matchptr - (cache_ptr + 1); + cache_ptr = lz_matchptr; + + /* Accumulate literal/match statistics for block + * splitting and for generating the initial cost + * model. */ + if (in_next >= next_observation) { + best_len = cache_ptr[-1].length; + if (best_len >= 3) { + /* Match (len >= 3) */ + + /* + * Note: for performance reasons this has + * been simplified significantly: + * + * - We wait until later to account for + * LZX_OFFSET_ADJUSTMENT. + * - We don't account for repeat offsets. + * - We don't account for different match headers. + */ + c->freqs.aligned[cache_ptr[-1].offset & + LZX_ALIGNED_OFFSET_BITMASK]++; + c->freqs.main[LZX_NUM_CHARS]++; + + lzx_observe_match(&c->split_stats, best_len); + next_observation = in_next + best_len; + } else { + /* Literal */ + c->freqs.main[*in_next]++; + lzx_observe_literal(&c->split_stats, *in_next); + next_observation = in_next + 1; + } } - } 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 { - 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; + /* + * If there was a very long match found, then + * don't cache any matches for the bytes covered + * by that match. This avoids degenerate + * behavior when compressing highly redundant + * data, where the number of matches can be very + * large. + * + * This heuristic doesn't actually hurt the + * compression ratio *too* much. If there's a + * long match, then the data must be highly + * compressible, so it doesn't matter as much + * what we do. + */ + if (best_len >= nice_len) + next_search_pos = in_next + best_len; + } else { + /* Don't search for matches at this position. */ + CALL_BT_MF(is_16_bit, c, + bt_matchfinder_skip_position, + in_begin, + in_next - in_begin, + nice_len, + c->max_search_depth, + next_hashes); + cache_ptr->length = 0; + cache_ptr++; + } + } while (++in_next < next_pause_point && + likely(cache_ptr < &c->match_cache[CACHE_LENGTH])); + + pause: + + /* Adjust max_len and nice_len if we're nearing the end of the + * input buffer. In addition, if we are so close to the end of + * the input buffer that there cannot be any more matches, then + * just advance through the last few positions and record no + * matches. */ + if (unlikely(max_len > in_end - in_next)) { + max_len = in_end - in_next; + nice_len = min(max_len, nice_len); + if (max_len < BT_MATCHFINDER_REQUIRED_NBYTES) { + while (in_next != in_end) { + cache_ptr->length = 0; + cache_ptr++; + in_next++; } - } while (++len <= matches[i].len); - } while (++i != num_matches); - } + } + } -#else /* Unoptimized version */ + /* End the block if the match cache may overflow. */ + if (unlikely(cache_ptr >= &c->match_cache[CACHE_LENGTH])) + goto end_block; + + /* End the block if the soft maximum size has been reached. */ + if (in_next >= in_max_block_end) + goto end_block; + + /* End the block if the block splitting algorithm thinks this is + * a good place to do so. */ + if (c->split_stats.num_new_observations >= + NUM_OBSERVATIONS_PER_BLOCK_CHECK && + in_max_block_end - in_next >= MIN_BLOCK_SIZE && + lzx_should_end_block(&c->split_stats)) + goto end_block; + + /* It's not time to end the block yet. Compute the next pause + * point and resume matchfinding. */ + next_pause_point = + min(in_next + min(NUM_OBSERVATIONS_PER_BLOCK_CHECK * 2 - + c->split_stats.num_new_observations, + in_max_block_end - in_next), + in_max_block_end - min(LZX_MAX_MATCH_LEN - 1, + in_max_block_end - in_next)); + goto resume_matchfinding; + + end_block: + /* We've decided on a block boundary and cached matches. Now + * choose a match/literal sequence and flush the block. */ + queue = lzx_optimize_and_flush_block(c, os, in_block_begin, + in_next - in_block_begin, + queue, is_16_bit); + } while (in_next != in_end); +} - unsigned num_extra_bits; +static void +lzx_compress_near_optimal_16(struct lzx_compressor *c, const u8 *in, + size_t in_nbytes, struct lzx_output_bitstream *os) +{ + lzx_compress_near_optimal(c, in, in_nbytes, os, true); +} - 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 +static void +lzx_compress_near_optimal_32(struct lzx_compressor *c, const u8 *in, + size_t in_nbytes, struct lzx_output_bitstream *os) +{ + lzx_compress_near_optimal(c, in, in_nbytes, os, false); } +/******************************************************************************/ +/* Faster ("lazy") compression algorithm */ +/*----------------------------------------------------------------------------*/ + /* - * Search for repeat offset matches with the current position. + * Called when the compressor chooses to use a literal. This tallies the + * Huffman symbol for the literal, increments the current literal run length, + * and "observes" the literal for the block split statistics. */ -static inline unsigned -lzx_repsearch(const u8 * const strptr, const u32 bytes_remaining, - const struct lzx_lru_queue *queue, unsigned *rep_max_idx_ret) +static inline void +lzx_choose_literal(struct lzx_compressor *c, unsigned literal, u32 *litrunlen_p) { - 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); + lzx_observe_literal(&c->split_stats, literal); + c->freqs.main[literal]++; + ++*litrunlen_p; } /* - * 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). + * Called when the compressor chooses to use a match. This tallies the Huffman + * symbol(s) for a match, saves the match data and the length of the preceding + * literal run, updates the recent offsets queue, and "observes" the match for + * the block split statistics. */ -static void -lzx_optim_pass(struct lzx_compressor *c, struct lzx_item **next_chosen_item) +static inline void +lzx_choose_match(struct lzx_compressor *c, unsigned length, u32 adjusted_offset, + u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], bool is_16_bit, + u32 *litrunlen_p, struct lzx_sequence **next_seq_p) { - 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; - - 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. */ - - window_ptr = &c->cur_window[c->match_window_pos]; - - 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); + u32 litrunlen = *litrunlen_p; + struct lzx_sequence *next_seq = *next_seq_p; + unsigned offset_slot; + unsigned v; - 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; - } + lzx_observe_match(&c->split_stats, length); - /* 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; + v = length - LZX_MIN_MATCH_LEN; - /* Consider coding a repeat offset match. */ - lzx_consider_repeat_offset_match(c, - cur_optimum_ptr, - rep_max_len, - rep_max_idx); - } + /* Save the literal run length and adjusted length. */ + next_seq->litrunlen = litrunlen; + next_seq->adjusted_length = v; - longest_len = matches[num_matches - 1].len; + /* Compute the length header, then tally the length symbol if needed. */ + if (v >= LZX_NUM_PRIMARY_LENS) { + c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++; + v = LZX_NUM_PRIMARY_LENS; + } - /* If there's a very long explicit offset match, choose - * it immediately. */ - if (longest_len >= c->params.nice_match_length) { + /* Compute the offset slot. */ + offset_slot = lzx_get_offset_slot(c, adjusted_offset, is_16_bit); - 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; + /* Compute the match header. */ + v += offset_slot * LZX_NUM_LEN_HEADERS; - 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; + /* Save the adjusted offset and match header. */ + next_seq->adjusted_offset_and_match_hdr = (adjusted_offset << 9) | v; - lzx_skip_bytes(c, longest_len - 1); - break; - } + /* Tally the main symbol. */ + c->freqs.main[LZX_NUM_CHARS + v]++; - /* 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; + /* Update the recent offsets queue. */ + if (adjusted_offset < LZX_NUM_RECENT_OFFSETS) { + /* Repeat offset match. */ + swap(recent_offsets[0], recent_offsets[adjusted_offset]); + } else { + /* Explicit offset match. */ - /* 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; - } - } + /* Tally the aligned offset symbol if needed. */ + if (adjusted_offset >= 16) + c->freqs.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]++; - /* Consider coding a literal. + recent_offsets[2] = recent_offsets[1]; + recent_offsets[1] = recent_offsets[0]; + recent_offsets[0] = adjusted_offset - LZX_OFFSET_ADJUSTMENT; + } - * 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); + /* Reset the literal run length and advance to the next sequence. */ + *next_seq_p = next_seq + 1; + *litrunlen_p = 0; +} - /* Advance to the next position. */ - cur_optimum_ptr++; +/* + * Called when the compressor ends a block. This finshes the last lzx_sequence, + * which is just a literal run with no following match. This literal run might + * be empty. + */ +static inline void +lzx_finish_sequence(struct lzx_sequence *last_seq, u32 litrunlen) +{ + last_seq->litrunlen = litrunlen; - /* The lowest-cost path to the current position is now known. - * Finalize the recent offsets queue that results from taking - * this lowest-cost path. */ + /* Special value to mark last sequence */ + last_seq->adjusted_offset_and_match_hdr = 0x80000000; +} - 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 { - /* 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]); - } +/* + * Find the longest repeat offset match with the current position. If a match + * is found, return its length and set *best_rep_idx_ret to the index of its + * offset in @recent_offsets. Otherwise, return 0. + * + * Don't bother with length 2 matches; consider matches of length >= 3 only. + * Also assume that max_len >= 3. + */ +static unsigned +lzx_find_longest_repeat_offset_match(const u8 * const in_next, + const u32 recent_offsets[], + const unsigned max_len, + unsigned *best_rep_idx_ret) +{ + STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3); /* loop is unrolled */ + + const u32 seq3 = load_u24_unaligned(in_next); + const u8 *matchptr; + unsigned best_rep_len = 0; + unsigned best_rep_idx = 0; + unsigned rep_len; + + /* Check for rep0 match (most recent offset) */ + matchptr = in_next - recent_offsets[0]; + if (load_u24_unaligned(matchptr) == seq3) + best_rep_len = lz_extend(in_next, matchptr, 3, max_len); + + /* Check for rep1 match (second most recent offset) */ + matchptr = in_next - recent_offsets[1]; + if (load_u24_unaligned(matchptr) == seq3) { + rep_len = lz_extend(in_next, matchptr, 3, max_len); + if (rep_len > best_rep_len) { + best_rep_len = rep_len; + best_rep_idx = 1; } + } - /* - * 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; + /* Check for rep2 match (third most recent offset) */ + matchptr = in_next - recent_offsets[2]; + if (load_u24_unaligned(matchptr) == seq3) { + rep_len = lz_extend(in_next, matchptr, 3, max_len); + if (rep_len > best_rep_len) { + best_rep_len = rep_len; + best_rep_idx = 2; } } - /* 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; + *best_rep_idx_ret = best_rep_idx; + return best_rep_len; } -/* Fast heuristic scoring for lazy parsing: how "good" is this match? */ +/* + * Fast heuristic scoring for lazy parsing: how "good" is this match? + * This is mainly determined by the length: longer matches are better. + * However, we also give a bonus to close (small offset) matches and to repeat + * offset matches, since those require fewer bits to encode. + */ + static inline unsigned lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset) { unsigned score = len; - if (adjusted_offset < 2048) + if (adjusted_offset < 4096) score++; - - if (adjusted_offset < 1024) + if (adjusted_offset < 256) score++; return score; } static inline unsigned -lzx_repeat_offset_match_score(unsigned len, unsigned slot) +lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx) { - return len + 3; + return rep_len + 3; } -/* Lazy parsing */ -static u32 -lzx_choose_lazy_items_for_block(struct lzx_compressor *c, - u32 block_start_pos, u32 block_size) +/* + * This is the "lazy" LZX compressor. The basic idea is that before it chooses + * a match, it checks to see if there's a longer match at the next position. If + * yes, it chooses a literal and continues to the next position. If no, it + * chooses the match. + * + * Some additional heuristics are used as well. Repeat offset matches are + * considered favorably and sometimes are chosen immediately. In addition, long + * matches (at least "nice_len" bytes) are chosen immediately as well. Finally, + * when we decide whether a match is "better" than another, we take the offset + * into consideration as well as the length. + */ +static inline void +lzx_compress_lazy(struct lzx_compressor * restrict c, + const u8 * const restrict in_begin, size_t in_nbytes, + struct lzx_output_bitstream * restrict os, bool is_16_bit) { - 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. */ - - no_match_found: - - if (prev_len) { - skip_len = prev_len - 2; - goto output_prev_match; - } else { - lzx_declare_literal(c, *(window_ptr - 1), - &next_chosen_item); - continue; - } - } - - /* 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; - } - - 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); - - /* 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 (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; - } - - 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. */ + const u8 * in_next = in_begin; + const u8 * const in_end = in_begin + in_nbytes; + unsigned max_len = LZX_MAX_MATCH_LEN; + unsigned nice_len = min(c->nice_match_length, max_len); + STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3); + u32 recent_offsets[LZX_NUM_RECENT_OFFSETS] = {1, 1, 1}; + u32 next_hashes[2] = {0, 0}; - if (prev_len) - lzx_declare_literal(c, *(window_ptr - 2), &next_chosen_item); + /* Initialize the matchfinder. */ + CALL_HC_MF(is_16_bit, c, hc_matchfinder_init); - prev_len = cur_len; - prev_offset_data = cur_offset_data; - prev_score = cur_score; + do { + /* Starting a new block */ + + const u8 * const in_block_begin = in_next; + const u8 * const in_max_block_end = + in_next + min(SOFT_MAX_BLOCK_SIZE, in_end - in_next); + struct lzx_sequence *next_seq = c->chosen_sequences; + u32 litrunlen = 0; + unsigned cur_len; + u32 cur_offset; + u32 cur_adjusted_offset; + unsigned cur_score; + unsigned next_len; + u32 next_offset; + u32 next_adjusted_offset; + unsigned next_score; + unsigned best_rep_len; + unsigned best_rep_idx; + unsigned rep_score; + unsigned skip_len; + + lzx_reset_symbol_frequencies(c); + lzx_init_block_split_stats(&c->split_stats); - if (prev_len >= c->params.nice_match_length) { - skip_len = prev_len - 1; - goto output_prev_match; + do { + /* Adjust max_len and nice_len if we're nearing the end + * of the input buffer. */ + if (unlikely(max_len > in_end - in_next)) { + max_len = in_end - in_next; + nice_len = min(max_len, nice_len); } - continue; - } - - /* 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 - * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or - * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively, - * will take fewer bits to output. */ -static int -lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs, - const struct lzx_codes * codes) -{ - u32 aligned_cost = 0; - u32 verbatim_cost = 0; - - /* 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; -} - -/* Near-optimal parsing */ -static u32 -lzx_choose_near_optimal_items_for_block(struct lzx_compressor *c, - u32 block_start_pos, u32 block_size) -{ - 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; - } - - /* No matches will extend beyond the end of the block. */ - c->match_window_end = block_start_pos + block_size; - - /* 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. For - * earlier passes, tallying symbol frequencies is sufficient. */ - lzx_set_default_costs(&c->costs, c->num_main_syms); - - 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; - } + /* Find the longest match (subject to the + * max_search_depth cutoff parameter) with the current + * position. Don't bother with length 2 matches; only + * look for matches of length >= 3. */ + cur_len = CALL_HC_MF(is_16_bit, c, + hc_matchfinder_longest_match, + in_begin, + in_next - in_begin, + 2, + max_len, + nice_len, + c->max_search_depth, + next_hashes, + &cur_offset); + + /* If there was no match found, or the only match found + * was a distant short match, then choose a literal. */ + if (cur_len < 3 || + (cur_len == 3 && + cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT && + cur_offset != recent_offsets[0] && + cur_offset != recent_offsets[1] && + cur_offset != recent_offsets[2])) + { + lzx_choose_literal(c, *in_next, &litrunlen); + in_next++; + continue; + } - /* Choose the items. */ - lzx_optim_pass(c, next_chosen_item_ptr); + /* Heuristic: if this match has the most recent offset, + * then go ahead and choose it as a rep0 match. */ + if (cur_offset == recent_offsets[0]) { + in_next++; + skip_len = cur_len - 1; + cur_adjusted_offset = 0; + goto choose_cur_match; + } - if (num_passes_remaining > 1) { - /* This isn't the last pass. */ + /* Compute the longest match's score as an explicit + * offset match. */ + cur_adjusted_offset = cur_offset + LZX_OFFSET_ADJUSTMENT; + cur_score = lzx_explicit_offset_match_score(cur_len, cur_adjusted_offset); + + /* Find the longest repeat offset match at this + * position. If we find one and it's "better" than the + * explicit offset match we found, then go ahead and + * choose the repeat offset match immediately. */ + best_rep_len = lzx_find_longest_repeat_offset_match(in_next, + recent_offsets, + max_len, + &best_rep_idx); + in_next++; + + if (best_rep_len != 0 && + (rep_score = lzx_repeat_offset_match_score(best_rep_len, + best_rep_idx)) >= cur_score) + { + cur_len = best_rep_len; + cur_adjusted_offset = best_rep_idx; + skip_len = best_rep_len - 1; + goto choose_cur_match; + } - /* Make the Huffman codes from the symbol frequencies. */ - lzx_make_huffman_codes(&c->freqs, &c->codes[c->codes_index], - c->num_main_syms); + have_cur_match: + /* + * We have a match at the current position. If the + * match is very long, then choose it immediately. + * Otherwise, see if there's a better match at the next + * position. + */ - /* Update symbol costs. */ - lzx_set_costs(c, &c->codes[c->codes_index].lens); + if (cur_len >= nice_len) { + skip_len = cur_len - 1; + goto choose_cur_match; + } - /* Reset symbol frequencies. */ - memset(&c->freqs, 0, sizeof(c->freqs)); + if (unlikely(max_len > in_end - in_next)) { + max_len = in_end - in_next; + nice_len = min(max_len, nice_len); + } - /* Reset the match offset LRU queue to what it was at - * the beginning of the block. */ - c->queue = orig_queue; + next_len = CALL_HC_MF(is_16_bit, c, + hc_matchfinder_longest_match, + in_begin, + in_next - in_begin, + cur_len - 2, + max_len, + nice_len, + c->max_search_depth / 2, + next_hashes, + &next_offset); + + if (next_len <= cur_len - 2) { + /* No potentially better match was found. */ + in_next++; + skip_len = cur_len - 2; + goto choose_cur_match; + } - /* 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; + next_adjusted_offset = next_offset + LZX_OFFSET_ADJUSTMENT; + next_score = lzx_explicit_offset_match_score(next_len, next_adjusted_offset); + + best_rep_len = lzx_find_longest_repeat_offset_match(in_next, + recent_offsets, + max_len, + &best_rep_idx); + in_next++; + + if (best_rep_len != 0 && + (rep_score = lzx_repeat_offset_match_score(best_rep_len, + best_rep_idx)) >= next_score) + { + + if (rep_score > cur_score) { + /* The next match is better, and it's a + * repeat offset match. */ + lzx_choose_literal(c, *(in_next - 2), + &litrunlen); + cur_len = best_rep_len; + cur_adjusted_offset = best_rep_idx; + skip_len = cur_len - 1; + goto choose_cur_match; + } } else { - c->get_matches_func = lzx_get_matches_usecache; - c->skip_bytes_func = lzx_skip_bytes_usecache; + if (next_score > cur_score) { + /* The next match is better, and it's an + * explicit offset match. */ + lzx_choose_literal(c, *(in_next - 2), + &litrunlen); + cur_len = next_len; + cur_adjusted_offset = next_adjusted_offset; + cur_score = next_score; + goto have_cur_match; + } } - } - } while (--num_passes_remaining); - /* Return the number of items chosen. */ - return next_chosen_item - c->chosen_items; + /* The original match was better; choose it. */ + skip_len = cur_len - 2; + + choose_cur_match: + /* Choose a match and have the matchfinder skip over its + * remaining bytes. */ + lzx_choose_match(c, cur_len, cur_adjusted_offset, + recent_offsets, is_16_bit, + &litrunlen, &next_seq); + in_next = CALL_HC_MF(is_16_bit, c, + hc_matchfinder_skip_positions, + in_begin, + in_next - in_begin, + in_end - in_begin, + skip_len, + next_hashes); + + /* Keep going until it's time to end the block. */ + } while (in_next < in_max_block_end && + !(c->split_stats.num_new_observations >= + NUM_OBSERVATIONS_PER_BLOCK_CHECK && + in_next - in_block_begin >= MIN_BLOCK_SIZE && + in_end - in_next >= MIN_BLOCK_SIZE && + lzx_should_end_block(&c->split_stats))); + + /* Flush the block. */ + lzx_finish_sequence(next_seq, litrunlen); + lzx_flush_block(c, os, in_block_begin, in_next - in_block_begin, 0); + + /* Keep going until we've reached the end of the input buffer. */ + } while (in_next != in_end); } -/* - * 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. - * - * The frequences of the Huffman symbols in the block will be tallied in - * 'c->freqs'. - * - * '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. - * - * Returns the number of matches/literals that were chosen and written to - * 'c->chosen_items' in the 'struct lzx_item' intermediate representation. - */ -static u32 -lzx_choose_items_for_block(struct lzx_compressor *c, - u32 block_start_pos, u32 block_size) +static void +lzx_compress_lazy_16(struct lzx_compressor *c, const u8 *in, size_t in_nbytes, + struct lzx_output_bitstream *os) { - return (*c->params.choose_items_for_block)(c, block_start_pos, block_size); + lzx_compress_lazy(c, in, in_nbytes, os, true); } -/* Initialize c->offset_slot_fast. */ static void -lzx_init_offset_slot_fast(struct lzx_compressor *c) +lzx_compress_lazy_32(struct lzx_compressor *c, const u8 *in, size_t in_nbytes, + struct lzx_output_bitstream *os) { - u8 slot = 0; + lzx_compress_lazy(c, in, in_nbytes, os, false); +} - for (u32 offset = 0; offset < LZX_NUM_FAST_OFFSETS; offset++) { +/******************************************************************************/ +/* Compressor operations */ +/*----------------------------------------------------------------------------*/ - while (offset + LZX_OFFSET_OFFSET >= lzx_offset_slot_base[slot + 1]) +/* + * Generate tables for mapping match offsets (actually, "adjusted" match + * offsets) to offset slots. + */ +static void +lzx_init_offset_slot_tabs(struct lzx_compressor *c) +{ + u32 adjusted_offset = 0; + unsigned slot = 0; + + /* slots [0, 29] */ + for (; adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1); + adjusted_offset++) + { + if (adjusted_offset >= lzx_offset_slot_base[slot + 1]) slot++; + c->offset_slot_tab_1[adjusted_offset] = slot; + } - c->offset_slot_fast[offset] = slot; + /* slots [30, 49] */ + for (; adjusted_offset < LZX_MAX_WINDOW_SIZE; + adjusted_offset += (u32)1 << 14) + { + if (adjusted_offset >= lzx_offset_slot_base[slot + 1]) + slot++; + c->offset_slot_tab_2[adjusted_offset >> 14] = slot; } } -/* 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) +static size_t +lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level) { - if (compression_level < 25) { - - /* Fast compression: Use lazy parsing. */ - - lzx_params->choose_items_for_block = lzx_choose_lazy_items_for_block; - lzx_params->num_optim_passes = 1; - - /* 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; + if (compression_level <= MAX_FAST_LEVEL) { + if (lzx_is_16_bit(max_bufsize)) + return offsetof(struct lzx_compressor, hc_mf_16) + + hc_matchfinder_size_16(max_bufsize); else - lzx_params->mf_algo = LZ_MF_BINARY_TREES; - - /* When lazy parsing, don't bother with length 2 matches. */ - lzx_params->min_match_length = 3; - - /* 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; + return offsetof(struct lzx_compressor, hc_mf_32) + + hc_matchfinder_size_32(max_bufsize); } else { - - /* Normal / high compression: Use near-optimal parsing. */ - - lzx_params->choose_items_for_block = lzx_choose_near_optimal_items_for_block; - - /* Set a number of optimization passes appropriate for the - * compression level. */ - - lzx_params->num_optim_passes = 1; - - 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; + if (lzx_is_16_bit(max_bufsize)) + return offsetof(struct lzx_compressor, bt_mf_16) + + bt_matchfinder_size_16(max_bufsize); 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 offsetof(struct lzx_compressor, bt_mf_32) + + bt_matchfinder_size_32(max_bufsize); } } -/* Given the internal compression parameters and maximum window size, build the - * Lempel-Ziv match-finder parameters. */ -static void -lzx_build_mf_params(const struct lzx_compressor_params *lzx_params, - u32 max_window_size, struct lz_mf_params *mf_params) -{ - 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 void -lzx_free_compressor(void *_c); - +/* Compute the amount of memory needed to allocate an LZX compressor. */ static u64 -lzx_get_needed_memory(size_t max_block_size, unsigned int compression_level) +lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level, + bool destructive) { - struct lzx_compressor_params params; u64 size = 0; - unsigned window_order; - u32 max_window_size; - window_order = lzx_get_window_order(max_block_size); - if (window_order == 0) + if (max_bufsize > LZX_MAX_WINDOW_SIZE) 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); + size += lzx_get_compressor_size(max_bufsize, compression_level); + if (!destructive) + size += max_bufsize; /* account for in_buffer */ return size; } +/* Allocate an LZX compressor. */ static int -lzx_create_compressor(size_t max_block_size, unsigned int compression_level, - void **c_ret) +lzx_create_compressor(size_t max_bufsize, unsigned compression_level, + bool destructive, void **c_ret) { - struct lzx_compressor *c; - struct lzx_compressor_params params; - struct lz_mf_params mf_params; unsigned window_order; - u32 max_window_size; + struct lzx_compressor *c; - window_order = lzx_get_window_order(max_block_size); + /* Validate the maximum buffer size and get the window order from it. */ + window_order = lzx_get_window_order(max_bufsize); if (window_order == 0) return WIMLIB_ERR_INVALID_PARAM; - max_window_size = max_block_size; - - 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; - c = CALLOC(1, sizeof(struct lzx_compressor)); + /* Allocate the compressor. */ + c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level)); if (!c) - goto oom; + goto oom0; - c->params = params; - c->num_main_syms = lzx_get_num_main_syms(window_order); c->window_order = window_order; + c->num_main_syms = lzx_get_num_main_syms(window_order); + c->destructive = destructive; + + /* Allocate the buffer for preprocessed data if needed. */ + if (!c->destructive) { + c->in_buffer = MALLOC(max_bufsize); + if (!c->in_buffer) + goto oom1; + } + + if (compression_level <= MAX_FAST_LEVEL) { + + /* Fast compression: Use lazy parsing. */ + if (lzx_is_16_bit(max_bufsize)) + c->impl = lzx_compress_lazy_16; + else + c->impl = lzx_compress_lazy_32; + + /* Scale max_search_depth and nice_match_length with the + * compression level. */ + c->max_search_depth = (60 * compression_level) / 20; + c->nice_match_length = (80 * compression_level) / 20; - /* 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; - - c->mf = lz_mf_alloc(&mf_params); - if (!c->mf) - goto oom; - - if (params.num_optim_passes > 1) { - c->cached_matches = MALLOC(LZX_CACHE_LEN * - sizeof(struct lz_match)); - if (!c->cached_matches) - goto oom; - c->cache_limit = c->cached_matches + LZX_CACHE_LEN - - (LZX_MAX_MATCHES_PER_POS + 1); + /* lzx_compress_lazy() needs max_search_depth >= 2 because it + * halves the max_search_depth when attempting a lazy match, and + * max_search_depth must be at least 1. */ + c->max_search_depth = max(c->max_search_depth, 2); } else { - c->cached_matches = MALLOC(LZX_MAX_MATCHES_PER_POS * - sizeof(struct lz_match)); - if (!c->cached_matches) - goto oom; + + /* Normal / high compression: Use near-optimal parsing. */ + if (lzx_is_16_bit(max_bufsize)) + c->impl = lzx_compress_near_optimal_16; + else + c->impl = lzx_compress_near_optimal_32; + + /* Scale max_search_depth and nice_match_length with the + * compression level. */ + c->max_search_depth = (24 * compression_level) / 50; + c->nice_match_length = (48 * compression_level) / 50; + + /* Also scale num_optim_passes with the compression level. But + * the more passes there are, the less they help --- so don't + * add them linearly. */ + c->num_optim_passes = 1; + c->num_optim_passes += (compression_level >= 45); + c->num_optim_passes += (compression_level >= 70); + c->num_optim_passes += (compression_level >= 100); + c->num_optim_passes += (compression_level >= 150); + c->num_optim_passes += (compression_level >= 200); + c->num_optim_passes += (compression_level >= 300); + + /* max_search_depth must be at least 1. */ + c->max_search_depth = max(c->max_search_depth, 1); } - lzx_init_offset_slot_fast(c); + /* Prepare the offset => offset slot mapping. */ + lzx_init_offset_slot_tabs(c); *c_ret = c; return 0; -oom: - lzx_free_compressor(c); +oom1: + FREE(c); +oom0: return WIMLIB_ERR_NOMEM; } +/* Compress a buffer of data. */ static size_t -lzx_compress(const void *uncompressed_data, size_t uncompressed_size, - void *compressed_data, size_t compressed_size_avail, void *_c) +lzx_compress(const void *restrict in, size_t in_nbytes, + void *restrict out, size_t out_nbytes_avail, void *restrict _c) { 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; + size_t result; - /* Don't bother compressing very small inputs. */ - if (uncompressed_size < 100) + /* Don't bother trying to compress very small inputs. */ + if (in_nbytes < 64) return 0; - /* 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; - - /* Preprocess the data. */ - lzx_do_e8_preprocessing(c->cur_window, c->cur_window_size); - - /* Load the window into the match-finder. */ - lz_mf_load_window(c->mf, c->cur_window, c->cur_window_size); - - /* Initialize the match offset LRU queue. */ - lzx_lru_queue_init(&c->queue); + /* If the compressor is in "destructive" mode, then we can directly + * preprocess the input data. Otherwise, we need to copy it into an + * internal buffer first. */ + if (!c->destructive) { + memcpy(c->in_buffer, in, in_nbytes); + in = c->in_buffer; + } - /* Initialize the output bitstream. */ - lzx_init_output(&os, compressed_data, compressed_size_avail); + /* Preprocess the input data. */ + lzx_preprocess((void *)in, in_nbytes); - /* 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; + /* Initially, the previous Huffman codeword lengths are all zeroes. */ 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); + memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens)); - /* Reset symbol frequencies. */ - memset(&c->freqs, 0, sizeof(c->freqs)); + /* Initialize the output bitstream. */ + lzx_init_output(&os, out, out_nbytes_avail); - /* Prepare the matches/literals for the block. */ - num_chosen_items = lzx_choose_items_for_block(c, - block_start_pos, - block_size); + /* Call the compression level-specific compress() function. */ + (*c->impl)(c, in, in_nbytes, &os); - /* Make the Huffman codes from the symbol frequencies. */ - lzx_make_huffman_codes(&c->freqs, &c->codes[c->codes_index], - c->num_main_syms); + /* Flush the output bitstream. */ + result = lzx_flush_output(&os); - /* 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 the data did not compress to less than its original size and we + * preprocessed the original buffer, then postprocess it to restore it + * to its original state. */ + if (result == 0 && c->destructive) + lzx_postprocess((void *)in, in_nbytes); + + /* Return the number of compressed bytes, or 0 if the input did not + * compress to less than its original size. */ + return result; } +/* Free an LZX compressor. */ 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); - } + if (!c->destructive) + FREE(c->in_buffer); + FREE(c); } const struct compressor_ops lzx_compressor_ops = {