X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=src%2Flzx-compress.c;h=c72fd460c84d47a4612fde0214499acad1d17f5c;hp=ddfe1bbd1c3d0bb4774e102d98248eb5d462c65f;hb=69bf8e6b27c11c8dfb0e9794ea43c3b8af72ee38;hpb=2254a0fc3f1d7af1151ee83f3458f44339b5028b diff --git a/src/lzx-compress.c b/src/lzx-compress.c index ddfe1bbd..c72fd460 100644 --- a/src/lzx-compress.c +++ b/src/lzx-compress.c @@ -42,16 +42,18 @@ * and certain other details are quite similar, such as the method for storing * Huffman codes. However, some of the main differences are: * - * - LZX preprocesses the data before attempting to compress it. + * - LZX preprocesses the data to attempt to make x86 machine code slightly more + * compressible before attempting to compress it further. * - LZX uses a "main" alphabet which combines literals and matches, with the * match symbols containing a "length header" (giving all or part of the match - * length) and a "position footer" (giving, roughly speaking, the order of + * length) and a "position slot" (giving, roughly speaking, the order of * magnitude of the match offset). * - LZX does not have static Huffman blocks; however it does have two types of * dynamic Huffman blocks ("aligned offset" and "verbatim"). * - LZX has a minimum match length of 2 rather than 3. - * - In LZX, match offsets 0 through 2 actually represent entries in a LRU queue - * of match offsets. + * - In LZX, match offsets 0 through 2 actually represent entries in an LRU + * queue of match offsets. This is very useful for certain types of files, + * such as binary files that have repeating records. * * Algorithms * ========== @@ -66,48 +68,67 @@ * * The "slow" algorithm to generate LZX-compressed data is roughly as follows: * - * 1. Preprocess the input data to translate the targets of x86 call instructions - * to absolute offsets. - * - * 2. Determine the best known sequence of LZ77 matches ((offset, length) pairs) - * and literal bytes to divide the input into. Raw match-finding is done - * using a very clever binary tree search based on the "Bt3" algorithm from - * 7-Zip. Parsing, or match-choosing, is solved essentially as a - * minimum-cost path problem, but using a heuristic forward search based on - * the Deflate encoder from 7-Zip rather than a more intuitive backward - * search, the latter of which would naively require that all matches be - * found. This heuristic search, as well as other heuristics such as limits - * on the matches considered, considerably speed up this part of the - * algorithm, which is the main bottleneck. Finally, after matches and - * literals are chosen, the needed Huffman codes needed to output them are - * built. - * - * 3. Up to a certain number of iterations, use the resulting Huffman codes to - * refine a cost model and go back to Step #2 to determine an improved + * 1. Preprocess the input data to translate the targets of x86 call + * instructions to absolute offsets. + * + * 2. Build the suffix array and inverse suffix array for the input data. The + * suffix array contains the indices of all suffixes of the input data, + * sorted lexcographically by the corresponding suffixes. The "position" of + * a suffix is the index of that suffix in the original string, whereas the + * "rank" of a suffix is the index at which that suffix's position is found + * in the suffix array. + * + * 3. Build the longest common prefix array corresponding to the suffix array. + * + * 4. For each suffix, find the highest lower ranked suffix that has a lower + * position, the lowest higher ranked suffix that has a lower position, and + * the length of the common prefix shared between each. This information is + * later used to link suffix ranks into a doubly-linked list for searching + * the suffix array. + * + * 5. Set a default cost model for matches/literals. + * + * 6. Determine the lowest cost sequence of LZ77 matches ((offset, length) + * pairs) and literal bytes to divide the input into. Raw match-finding is + * done by searching the suffix array using a linked list to avoid + * considering any suffixes that start after the current position. Each run + * of the match-finder returns the approximate lowest-cost longest match as + * well as any shorter matches that have even lower approximate costs. Each + * such run also adds the suffix rank of the current position into the linked + * list being used to search the suffix array. Parsing, or match-choosing, + * is solved as a minimum-cost path problem using a forward "optimal parsing" + * algorithm based on the Deflate encoder from 7-Zip. This algorithm moves + * forward calculating the minimum cost to reach each byte until either a + * very long match is found or until a position is found at which no matches + * start or overlap. + * + * 7. Build the Huffman codes needed to output the matches/literals. + * + * 8. Up to a certain number of iterations, use the resulting Huffman codes to + * refine a cost model and go back to Step #6 to determine an improved * sequence of matches and literals. * - * 4. Up to a certain depth, try splitting the current block to see if the - * compression ratio can be improved. This may be the case if parts of the - * input differ greatly from each other and could benefit from different - * Huffman codes. + * 9. Output the resulting block using the match/literal sequences and the + * Huffman codes that were computed for the block. * - * 5. Output the resulting block(s) using the match/literal sequences and the - * Huffman codes that were computed for each block. + * Note: the algorithm does not yet attempt to split the input into multiple LZX + * blocks, instead using a series of blocks of LZX_DIV_BLOCK_SIZE bytes. * * Fast algorithm * -------------- * * The fast algorithm (and the only one available in wimlib v1.5.1 and earlier) * spends much less time on the main bottlenecks of the compression process --- - * that is the match finding, match choosing, and block splitting. Matches are - * found and chosen with hash chains using a greedy parse with one position of - * look-ahead. No block splitting is done; only compressing the full input into - * an aligned offset block is considered. + * that is, the match finding and match choosing. Matches are found and chosen + * with hash chains using a greedy parse with one position of look-ahead. No + * block splitting is done; only compressing the full input into an aligned + * offset block is considered. * * API * === * - * The old API (retained for backward compatibility) consists of just one function: + * The old API (retained for backward compatibility) consists of just one + * function: * * wimlib_lzx_compress() * @@ -118,34 +139,46 @@ * wimlib_lzx_alloc_context() * wimlib_lzx_compress2() * wimlib_lzx_free_context() + * wimlib_lzx_set_default_params() * * Both wimlib_lzx_compress() and wimlib_lzx_compress2() are designed to - * compress an in-memory buffer of up to 32768 bytes. There is no sliding - * window. This is suitable for the WIM format, which uses fixed-size chunks - * that are seemingly always 32768 bytes. If needed, the compressor potentially - * could be extended to support a larger and/or sliding window. + * compress an in-memory buffer of up to the window size, which can be any power + * of two between 2^15 and 2^21 inclusively. However, by default, the WIM + * format uses 2^15, and this is seemingly the only value that is compatible + * with WIMGAPI. In any case, the window is not a true "sliding window" since + * no data is ever "slid out" of the window. This is needed for the WIM format, + * which is designed such that chunks may be randomly accessed. * * Both wimlib_lzx_compress() and wimlib_lzx_compress2() return 0 if the data * could not be compressed to less than the size of the uncompressed data. * Again, this is suitable for the WIM format, which stores such data chunks * uncompressed. * - * The functions in this API are exported from the library, although this is - * only in case other programs happen to have uses for it other than WIM + * The functions in this LZX compression API are exported from the library, + * although with the possible exception of wimlib_lzx_set_default_params(), this + * is only in case other programs happen to have uses for it other than WIM * reading/writing as already handled through the rest of the library. * * Acknowledgments * =============== * - * Acknowledgments to several other open-source projects that made it possible - * to implement this code: + * Acknowledgments to several open-source projects and research papers that made + * it possible to implement this code: * - * - 7-Zip (author: Igor Pavlov), for the binary tree match-finding - * algorithm, the heuristic near-optimal forward match-choosing - * algorithm, and the block splitting algorithm. + * - divsufsort (author: Yuta Mori), for the suffix array construction code, + * located in a separate directory (divsufsort/). + * + * - "Linear-Time Longest-Common-Prefix Computation in Suffix Arrays and Its + * Applications" (Kasai et al. 2001), for the LCP array computation. + * + * - "LPF computation revisited" (Crochemore et al. 2009) for the prev and next + * array computations. + * + * - 7-Zip (author: Igor Pavlov) for the algorithm for forward optimal parsing + * (match-choosing). * * - zlib (author: Jean-loup Gailly and Mark Adler), for the hash table - * match-finding algorithm. + * match-finding algorithm (used in lz77.c). * * - lzx-compress (author: Matthew T. Russotto), on which some parts of this * code were originally based. @@ -157,46 +190,56 @@ #include "wimlib.h" #include "wimlib/compress.h" +#include "wimlib/endianness.h" #include "wimlib/error.h" #include "wimlib/lzx.h" #include "wimlib/util.h" +#include +#include +#include #ifdef ENABLE_LZX_DEBUG -# include +# include "wimlib/decompress.h" #endif -#include +#include "divsufsort/divsufsort.h" -/* Experimental parameters not exposed through the API */ -#define LZX_PARAM_OPTIM_ARRAY_SIZE 1024 -#define LZX_PARAM_ACCOUNT_FOR_LRU 1 -#define LZX_PARAM_DONT_SKIP_MATCHES 0 -#define LZX_PARAM_USE_EMPIRICAL_DEFAULT_COSTS 1 +typedef u32 block_cost_t; +#define INFINITE_BLOCK_COST ((block_cost_t)~0U) -/* Currently, this constant can't simply be changed because the code currently - * uses a static number of position slots. */ -#define LZX_MAX_WINDOW_SIZE 32768 +#define LZX_OPTIM_ARRAY_SIZE 4096 -/* This may be WIM-specific */ -#define LZX_DEFAULT_BLOCK_SIZE 32768 +#define LZX_DIV_BLOCK_SIZE 32768 -#define LZX_LZ_HASH_BITS 15 -#define LZX_LZ_HASH_SIZE (1 << LZX_LZ_HASH_BITS) -#define LZX_LZ_HASH_MASK (LZX_LZ_HASH_SIZE - 1) -#define LZX_LZ_HASH_SHIFT 5 +#define LZX_MAX_CACHE_PER_POS 10 /* Codewords for the LZX main, length, and aligned offset Huffman codes */ struct lzx_codewords { - u16 main[LZX_MAINTREE_NUM_SYMBOLS]; - u16 len[LZX_LENTREE_NUM_SYMBOLS]; - u16 aligned[LZX_ALIGNEDTREE_NUM_SYMBOLS]; + u16 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; + u16 len[LZX_LENCODE_NUM_SYMBOLS]; + u16 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* Lengths for the LZX main, length, and aligned offset Huffman codes */ +/* Codeword lengths (in bits) for the LZX main, length, and aligned offset + * Huffman codes. + * + * A 0 length means the codeword has zero frequency. + */ struct lzx_lens { - u8 main[LZX_MAINTREE_NUM_SYMBOLS]; - u8 len[LZX_LENTREE_NUM_SYMBOLS]; - u8 aligned[LZX_ALIGNEDTREE_NUM_SYMBOLS]; + u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; + u8 len[LZX_LENCODE_NUM_SYMBOLS]; + u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; +}; + +/* Costs for the LZX main, length, and aligned offset Huffman symbols. + * + * If a codeword has zero frequency, it must still be assigned some nonzero cost + * --- generally a high cost, since even if it gets used in the next iteration, + * it probably will not be used very times. */ +struct lzx_costs { + u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; + u8 len[LZX_LENCODE_NUM_SYMBOLS]; + u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; /* The LZX main, length, and aligned offset Huffman codes */ @@ -207,9 +250,9 @@ struct lzx_codes { /* Tables for tallying symbol frequencies in the three LZX alphabets */ struct lzx_freqs { - freq_t main[LZX_MAINTREE_NUM_SYMBOLS]; - freq_t len[LZX_LENTREE_NUM_SYMBOLS]; - freq_t aligned[LZX_ALIGNEDTREE_NUM_SYMBOLS]; + input_idx_t main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; + input_idx_t len[LZX_LENCODE_NUM_SYMBOLS]; + input_idx_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; /* LZX intermediate match/literal format */ @@ -223,50 +266,45 @@ struct lzx_match { * * 8-24 position footer. This is the offset of the real formatted * offset from the position base. This can be at most 17 bits - * (since lzx_extra_bits[LZX_NUM_POSITION_SLOTS - 1] is 17). + * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17). * * 0-7 length of match, minus 2. This can be at most - * (LZX_MAX_MATCH - 2) == 255, so it will fit in 8 bits. */ + * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */ u32 data; }; /* Raw LZ match/literal format: just a length and offset. * * The length is the number of bytes of the match, and the offset is the number - * of bytes back in the input the match is from the matched text. + * of bytes back in the input the match is from the current position. * - * If @len < LZX_MIN_MATCH, then it's really just a literal byte. */ + * If @len < LZX_MIN_MATCH_LEN, then it's really just a literal byte and @offset is + * meaningless. */ struct raw_match { u16 len; - u16 offset; + input_idx_t offset; }; -/* Specification for a LZX block */ +/* Specification for an LZX block. */ struct lzx_block_spec { - /* Set to 1 if this block has been split (in two --- we only considser - * binary splits). In such cases the rest of the fields are - * unimportant, since the relevant information is rather in the - * structures for the sub-blocks. */ - u8 is_split : 1; - /* One of the LZX_BLOCKTYPE_* constants indicating which type of this * block. */ - u8 block_type : 2; + int block_type; /* 0-based position in the window at which this block starts. */ - u16 window_pos; + input_idx_t window_pos; /* The number of bytes of uncompressed data this block represents. */ - u16 block_size; + input_idx_t block_size; /* The position in the 'chosen_matches' array in the `struct * lzx_compressor' at which the match/literal specifications for * this block begin. */ - unsigned chosen_matches_start_pos; + input_idx_t chosen_matches_start_pos; /* The number of match/literal specifications for this block. */ - u16 num_chosen_matches; + input_idx_t num_chosen_matches; /* Huffman codes for this block. */ struct lzx_codes codes; @@ -281,7 +319,7 @@ struct lzx_block_spec { struct lzx_optimal { /* The approximate minimum cost, in bits, to reach this position in the * window which has been found so far. */ - u32 cost; + block_cost_t cost; /* The union here is just for clarity, since the fields are used in two * slightly different ways. Initially, the @prev structure is filled in @@ -293,30 +331,58 @@ struct lzx_optimal { /* Position of the start of the match or literal that * was taken to get to this position in the approximate * minimum-cost parse. */ - u16 link; + input_idx_t link; - /* Offset, relative to its starting position, of the + /* Offset (as in an LZ (length, offset) pair) of the * match or literal that was taken to get to this * position in the approximate minimum-cost parse. */ - u16 match_offset; + input_idx_t match_offset; } prev; struct { /* Position at which the match or literal starting at * this position ends in the minimum-cost parse. */ - u16 link; + input_idx_t link; - /* Offset, relative to its starting position, of the + /* Offset (as in an LZ (length, offset) pair) of the * match or literal starting at this position in the * approximate minimum-cost parse. */ - u16 match_offset; + input_idx_t match_offset; } next; }; -#if LZX_PARAM_ACCOUNT_FOR_LRU + + /* The match offset LRU queue that will exist when the approximate + * minimum-cost path to reach this position is taken. */ struct lzx_lru_queue queue; -#endif }; -/* State of the LZX compressor */ +/* Suffix array link */ +struct salink { + /* Rank of highest ranked suffix that has rank lower than the suffix + * corresponding to this structure and either has a lower position + * (initially) or has a position lower than the highest position at + * which matches have been searched for so far, or -1 if there is no + * such suffix. */ + input_idx_t prev; + + /* Rank of lowest ranked suffix that has rank greater than the suffix + * corresponding to this structure and either has a lower position + * (intially) or has a position lower than the highest position at which + * matches have been searched for so far, or -1 if there is no such + * suffix. */ + input_idx_t next; + + /* Length of longest common prefix between the suffix corresponding to + * this structure and the suffix with rank @prev, or 0 if @prev is -1. + */ + input_idx_t lcpprev; + + /* Length of longest common prefix between the suffix corresponding to + * this structure and the suffix with rank @next, or 0 if @next is -1. + */ + input_idx_t lcpnext; +}; + +/* State of the LZX compressor. */ struct lzx_compressor { /* The parameters that were used to create the compressor. */ @@ -327,7 +393,7 @@ struct lzx_compressor { * 0xe8 byte preprocessing is done directly on the data here before * further compression. * - * Note that this compressor does *not* use a sliding window!!!! + * Note that this compressor does *not* use a real sliding window!!!! * It's not needed in the WIM format, since every chunk is compressed * independently. This is by design, to allow random access to the * chunks. @@ -335,82 +401,93 @@ struct lzx_compressor { * We reserve a few extra bytes to potentially allow reading off the end * of the array in the match-finding code for optimization purposes. */ - u8 window[LZX_MAX_WINDOW_SIZE + 12]; + u8 *window; /* Number of bytes of data to be compressed, which is the number of * bytes of data in @window that are actually valid. */ - unsigned window_size; + input_idx_t window_size; + + /* Allocated size of the @window. */ + input_idx_t max_window_size; + + /* Number of symbols in the main alphabet (depends on the + * @max_window_size since it determines the maximum allowed offset). */ + unsigned num_main_syms; /* The current match offset LRU queue. */ struct lzx_lru_queue queue; - /* Space for sequence of matches/literals that were chosen. - * - * Each LZX_MAX_WINDOW_SIZE-sized portion of this array is used for a - * different block splitting level. */ + /* Space for the sequences of matches/literals that were chosen for each + * block. */ struct lzx_match *chosen_matches; - /* Structures used during block splitting. - * - * This can be thought of as a binary tree. block_specs[(1) - 1] - * represents to the top-level block (root node), and block_specs[(i*2) - * - 1] and block_specs[(i*2+1) - 1] represent the sub-blocks (child - * nodes) resulting from a binary split of the block represented by - * block_spec[(i) - 1]. - */ + /* Information about the LZX blocks the preprocessed input was divided + * into. */ struct lzx_block_spec *block_specs; + /* Number of LZX blocks the input was divided into; a.k.a. the number of + * elements of @block_specs that are valid. */ + unsigned num_blocks; + /* This is simply filled in with zeroes and used to avoid special-casing * the output of the first compressed Huffman code, which conceptually * has a delta taken from a code with all symbols having zero-length * codewords. */ struct lzx_codes zero_codes; - /* Slow algorithm only: The current cost model. */ - struct lzx_lens costs; + /* The current cost model. */ + struct lzx_costs costs; - /* Slow algorithm only: Table that maps the hash codes for 3 character - * sequences to the most recent position that sequence (or a sequence - * sharing the same hash code) appeared in the window. */ - u16 *hash_tab; + /* Fast algorithm only: Array of hash table links. */ + input_idx_t *prev_tab; - /* Slow algorithm only: Table that maps 2-character sequences to the - * most recent position that sequence appeared in the window. */ - u16 *digram_tab; + /* Suffix array for window. + * This is a mapping from suffix rank to suffix position. */ + input_idx_t *SA; - /* Slow algorithm only: Table that contains the logical child pointers - * in the binary trees in the match-finding code. - * - * child_tab[i*2] and child_tab[i*2+1] are the left and right pointers, - * respectively, from the binary tree root corresponding to window - * position i. */ - u16 *child_tab; + /* Inverse suffix array for window. + * This is a mapping from suffix position to suffix rank. + * If 0 <= r < window_size, then ISA[SA[r]] == r. */ + input_idx_t *ISA; - /* Slow algorithm only: Matches that were already found and are saved in - * memory for subsequent queries (e.g. when block splitting). */ - struct raw_match *cached_matches; + /* Longest common prefix array corresponding to the suffix array SA. + * LCP[i] is the length of the longest common prefix between the + * suffixes with positions SA[i - 1] and SA[i]. LCP[0] is undefined. + */ + input_idx_t *LCP; - /* Slow algorithm only: Next position in 'cached_matches' to either - * return or fill in. */ + /* Suffix array links. + * + * During a linear scan of the input string to find matches, this array + * used to keep track of which rank suffixes in the suffix array appear + * before the current position. Instead of searching in the original + * suffix array, scans for matches at a given position traverse a linked + * list containing only suffixes that appear before that position. */ + struct salink *salink; + + /* Position in window of next match to return. */ + input_idx_t match_window_pos; + + /* The match-finder shall ensure the length of matches does not exceed + * this position in the input. */ + input_idx_t match_window_end; + + /* Matches found by the match-finder are cached in the following array + * to achieve a slight speedup when the same matches are needed on + * subsequent passes. This is suboptimal because different matches may + * be preferred with different cost models, but seems to be a worthwhile + * speedup. */ + struct raw_match *cached_matches; unsigned cached_matches_pos; - - /* Slow algorithm only: %true if reading from 'cached_matches'; %false - * if writing to 'cached_matches'. */ - bool matches_already_found; - - /* Slow algorithm only: Position in window of next match to return. */ - unsigned match_window_pos; - - /* Slow algorithm only: No matches returned shall reach past this - * position. */ - unsigned match_window_end; + bool matches_cached; /* Slow algorithm only: Temporary space used for match-choosing * algorithm. * - * The size of this array must be at least LZX_MAX_MATCH but otherwise - * is arbitrary. More space simply allows the match-choosing algorithm - * to find better matches (depending on the input, as always). */ + * The size of this array must be at least LZX_MAX_MATCH_LEN but + * otherwise is arbitrary. More space simply allows the match-choosing + * algorithm to potentially find better matches (depending on the input, + * as always). */ struct lzx_optimal *optimum; /* Slow algorithm only: Variables used by the match-choosing algorithm. @@ -423,87 +500,72 @@ struct lzx_compressor { u32 optimum_end_idx; }; -/* Returns the LZX position slot that corresponds to a given formatted offset. - * - * Logically, this returns the smallest i such that - * formatted_offset >= lzx_position_base[i]. - * - * The actual implementation below takes advantage of the regularity of the - * numbers in the lzx_position_base array to calculate the slot directly from - * the formatted offset without actually looking at the array. - */ +/* Returns the LZX position slot that corresponds to a given match offset, + * taking into account the recent offset queue and updating it if the offset is + * found in it. */ static unsigned -lzx_get_position_slot(unsigned formatted_offset) +lzx_get_position_slot(unsigned offset, struct lzx_lru_queue *queue) { -#if 0 - /* - * Slots 36-49 (formatted_offset >= 262144) can be found by - * (formatted_offset/131072) + 34 == (formatted_offset >> 17) + 34; - * however, this check for formatted_offset >= 262144 is commented out - * because WIM chunks cannot be that large. - */ - if (formatted_offset >= 262144) { - return (formatted_offset >> 17) + 34; - } else -#endif - { - /* Note: this part here only works if: - * - * 2 <= formatted_offset < 655360 - * - * It is < 655360 because the frequency of the position bases - * increases starting at the 655360 entry, and it is >= 2 - * because the below calculation fails if the most significant - * bit is lower than the 2's place. */ - LZX_ASSERT(2 <= formatted_offset && formatted_offset < 655360); - unsigned mssb_idx = bsr32(formatted_offset); - return (mssb_idx << 1) | - ((formatted_offset >> (mssb_idx - 1)) & 1); + unsigned position_slot; + + /* See if the offset was recently used. */ + for (unsigned i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) { + if (offset == queue->R[i]) { + /* Found it. */ + + /* Bring the repeat offset to the front of the + * queue. Note: this is, in fact, not a real + * LRU queue because repeat matches are simply + * swapped to the front. */ + swap(queue->R[0], queue->R[i]); + + /* The resulting position slot is simply the first index + * at which the offset was found in the queue. */ + return i; + } } -} -/* Compute the hash code for the next 3-character sequence in the window. */ -static unsigned -lzx_lz_compute_hash(const u8 *window) -{ - unsigned hash; - - hash = window[0]; - hash <<= LZX_LZ_HASH_SHIFT; - hash ^= window[1]; - hash <<= LZX_LZ_HASH_SHIFT; - hash ^= window[2]; - return hash & LZX_LZ_HASH_MASK; + /* The offset was not recently used; look up its real position slot. */ + position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET); + + /* Bring the new offset to the front of the queue. */ + for (unsigned i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--) + queue->R[i] = queue->R[i - 1]; + queue->R[0] = offset; + + return position_slot; } /* 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 lengths. */ + * 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) + struct lzx_codes *codes, + unsigned num_main_syms) { - make_canonical_huffman_code(LZX_MAINTREE_NUM_SYMBOLS, - LZX_MAX_CODEWORD_LEN, + make_canonical_huffman_code(num_main_syms, + LZX_MAX_MAIN_CODEWORD_LEN, freqs->main, codes->lens.main, codes->codewords.main); - make_canonical_huffman_code(LZX_LENTREE_NUM_SYMBOLS, - LZX_MAX_CODEWORD_LEN, + make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS, + LZX_MAX_LEN_CODEWORD_LEN, freqs->len, codes->lens.len, codes->codewords.len); - make_canonical_huffman_code(LZX_ALIGNEDTREE_NUM_SYMBOLS, 8, + make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS, + LZX_MAX_ALIGNED_CODEWORD_LEN, freqs->aligned, codes->lens.aligned, codes->codewords.aligned); } /* - * Output a LZX match. + * Output an LZX match. * * @out: The bitstream to write the match to. * @block_type: The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM) @@ -523,21 +585,20 @@ lzx_write_match(struct output_bitstream *out, int block_type, unsigned position_slot = (match.data >> 25) & 0x3f; /* 6 bits */ unsigned len_header; unsigned len_footer; - unsigned len_pos_header; unsigned main_symbol; unsigned num_extra_bits; unsigned verbatim_bits; unsigned aligned_bits; - /* If the match length is less than MIN_MATCH (= 2) + + /* If the match length is less than MIN_MATCH_LEN (= 2) + * NUM_PRIMARY_LENS (= 7), the length header contains - * the match length minus MIN_MATCH, and there is no + * the match length minus MIN_MATCH_LEN, and there is no * length footer. * * Otherwise, the length header contains * NUM_PRIMARY_LENS, and the length footer contains * the match length minus NUM_PRIMARY_LENS minus - * MIN_MATCH. */ + * MIN_MATCH_LEN. */ if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) { len_header = match_len_minus_2; /* No length footer-- mark it with a special @@ -548,15 +609,13 @@ lzx_write_match(struct output_bitstream *out, int block_type, len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS; } - /* Combine the position slot with the length header into - * a single symbol that will be encoded with the main - * tree. */ - len_pos_header = (position_slot << 3) | len_header; - - /* The actual main symbol is offset by LZX_NUM_CHARS because - * values under LZX_NUM_CHARS are used to indicate a literal - * byte rather than a match. */ - main_symbol = len_pos_header + LZX_NUM_CHARS; + /* Combine the position slot with the length header into a single symbol + * that will be encoded with the main code. + * + * The actual main symbol is offset by LZX_NUM_CHARS because values + * under LZX_NUM_CHARS are used to indicate a literal byte rather than a + * match. */ + main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS; /* Output main symbol. */ bitstream_put_bits(out, codes->codewords.main[main_symbol], @@ -573,7 +632,7 @@ lzx_write_match(struct output_bitstream *out, int block_type, /* For aligned offset blocks with at least 3 extra bits, output the * verbatim bits literally, then the aligned bits encoded using the - * aligned offset tree. Otherwise, only the verbatim bits need to be + * aligned offset code. Otherwise, only the verbatim bits need to be * output. */ if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) { @@ -595,23 +654,15 @@ lzx_write_match(struct output_bitstream *out, int block_type, static unsigned lzx_build_precode(const u8 lens[restrict], const u8 prev_lens[restrict], - unsigned num_syms, - freq_t precode_freqs[restrict LZX_PRETREE_NUM_SYMBOLS], + const unsigned num_syms, + input_idx_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS], u8 output_syms[restrict num_syms], - u8 precode_lens[restrict LZX_PRETREE_NUM_SYMBOLS], - u16 precode_codewords[restrict LZX_PRETREE_NUM_SYMBOLS], - unsigned * num_additional_bits_ret) + u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS], + u16 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS], + unsigned *num_additional_bits_ret) { - unsigned output_syms_idx; - unsigned cur_run_len; - unsigned i; - unsigned len_in_run; - unsigned additional_bits; - signed char delta; - unsigned num_additional_bits = 0; - memset(precode_freqs, 0, - LZX_PRETREE_NUM_SYMBOLS * sizeof(precode_freqs[0])); + LZX_PRECODE_NUM_SYMBOLS * sizeof(precode_freqs[0])); /* Since the code word lengths use a form of RLE encoding, the goal here * is to find each run of identical lengths when going through them in @@ -620,14 +671,14 @@ lzx_build_precode(const u8 lens[restrict], * literally. * * output_syms[] will be filled in with the length symbols that will be - * output, including RLE codes, not yet encoded using the pre-tree. + * output, including RLE codes, not yet encoded using the precode. * * cur_run_len keeps track of how many code word lengths are in the - * current run of identical lengths. - */ - output_syms_idx = 0; - cur_run_len = 1; - for (i = 1; i <= num_syms; i++) { + * current run of identical lengths. */ + unsigned output_syms_idx = 0; + unsigned cur_run_len = 1; + unsigned num_additional_bits = 0; + for (unsigned i = 1; i <= num_syms; i++) { if (i != num_syms && lens[i] == lens[i - 1]) { /* Still in a run--- keep going. */ @@ -640,7 +691,7 @@ lzx_build_precode(const u8 lens[restrict], /* The symbol that was repeated in the run--- not to be confused * with the length *of* the run (cur_run_len) */ - len_in_run = lens[i - 1]; + unsigned len_in_run = lens[i - 1]; if (len_in_run == 0) { /* A run of 0's. Encode it in as few length @@ -650,6 +701,7 @@ lzx_build_precode(const u8 lens[restrict], * where n is an uncompressed literal 5-bit integer that * follows the magic length. */ while (cur_run_len >= 20) { + unsigned additional_bits; additional_bits = min(cur_run_len - 20, 0x1f); num_additional_bits += 5; @@ -663,6 +715,8 @@ lzx_build_precode(const u8 lens[restrict], * where n is an uncompressed literal 4-bit integer that * follows the magic length. */ while (cur_run_len >= 4) { + unsigned additional_bits; + additional_bits = min(cur_run_len - 4, 0xf); num_additional_bits += 4; precode_freqs[17]++; @@ -683,9 +737,12 @@ lzx_build_precode(const u8 lens[restrict], * * The extra length symbol is encoded as a difference * from the length of the codeword for the first symbol - * in the run in the previous tree. + * in the run in the previous code. * */ while (cur_run_len >= 4) { + unsigned additional_bits; + signed char delta; + additional_bits = (cur_run_len > 4); num_additional_bits += 1; delta = (signed char)prev_lens[i - cur_run_len] - @@ -703,8 +760,10 @@ lzx_build_precode(const u8 lens[restrict], /* Any remaining lengths in the run are outputted without RLE, * as a difference from the length of that codeword in the - * previous tree. */ + * previous code. */ while (cur_run_len > 0) { + signed char delta; + delta = (signed char)prev_lens[i - cur_run_len] - (signed char)len_in_run; if (delta < 0) @@ -720,13 +779,12 @@ lzx_build_precode(const u8 lens[restrict], /* Build the precode from the frequencies of the length symbols. */ - make_canonical_huffman_code(LZX_PRETREE_NUM_SYMBOLS, - LZX_MAX_CODEWORD_LEN, + make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, + LZX_MAX_PRE_CODEWORD_LEN, precode_freqs, precode_lens, precode_codewords); - if (num_additional_bits_ret) - *num_additional_bits_ret = num_additional_bits; + *num_additional_bits_ret = num_additional_bits; return output_syms_idx; } @@ -738,7 +796,7 @@ lzx_build_precode(const u8 lens[restrict], * The Huffman code is represented in the output as a series of path lengths * from which the canonical Huffman code can be reconstructed. The path lengths * themselves are compressed using a separate Huffman code, the precode, which - * consists of LZX_PRETREE_NUM_SYMBOLS (= 20) symbols that cover all possible + * consists of LZX_PRECODE_NUM_SYMBOLS (= 20) symbols that cover all possible * code lengths, plus extra codes for repeated lengths. The path lengths of the * precode precede the path lengths of the larger code and are uncompressed, * consisting of 20 entries of 4 bits each. @@ -756,13 +814,14 @@ lzx_write_compressed_code(struct output_bitstream *out, const u8 prev_lens[restrict], unsigned num_syms) { - freq_t precode_freqs[LZX_PRETREE_NUM_SYMBOLS]; + input_idx_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS]; u8 output_syms[num_syms]; - u8 precode_lens[LZX_PRETREE_NUM_SYMBOLS]; - u16 precode_codewords[LZX_PRETREE_NUM_SYMBOLS]; + u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS]; + u16 precode_codewords[LZX_PRECODE_NUM_SYMBOLS]; unsigned i; unsigned num_output_syms; u8 precode_sym; + unsigned dummy; num_output_syms = lzx_build_precode(lens, prev_lens, @@ -771,12 +830,12 @@ lzx_write_compressed_code(struct output_bitstream *out, output_syms, precode_lens, precode_codewords, - NULL); + &dummy); /* Write the lengths of the precode codes to the output. */ - for (i = 0; i < LZX_PRETREE_NUM_SYMBOLS; i++) + for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) bitstream_put_bits(out, precode_lens[i], - LZX_PRETREE_ELEMENT_SIZE); + LZX_PRECODE_ELEMENT_SIZE); /* Write the length symbols, encoded with the precode, to the output. */ @@ -805,60 +864,21 @@ lzx_write_compressed_code(struct output_bitstream *out, } } -static unsigned -lzx_huffman_code_output_cost(const u8 lens[restrict], - const freq_t freqs[restrict], - unsigned num_syms) -{ - unsigned cost = 0; - - for (unsigned i = 0; i < num_syms; i++) - cost += (unsigned)lens[i] * (unsigned)freqs[i]; - - return cost; -} - -/* Return the number of bits required to output the lengths for the specified - * Huffman code in compressed format (encoded with a precode). */ -static unsigned -lzx_code_cost(const u8 lens[], const u8 prev_lens[], unsigned num_syms) -{ - u8 output_syms[num_syms]; - freq_t precode_freqs[LZX_PRETREE_NUM_SYMBOLS]; - u8 precode_lens[LZX_PRETREE_NUM_SYMBOLS]; - u16 precode_codewords[LZX_PRETREE_NUM_SYMBOLS]; - unsigned cost = 0; - unsigned num_additional_bits; - - /* Acount for the lengths of the precode itself. */ - cost += LZX_PRETREE_NUM_SYMBOLS * LZX_PRETREE_ELEMENT_SIZE; - - lzx_build_precode(lens, prev_lens, num_syms, - precode_freqs, output_syms, - precode_lens, precode_codewords, - &num_additional_bits); - - /* Account for all precode symbols output. */ - cost += lzx_huffman_code_output_cost(precode_lens, precode_freqs, - LZX_PRETREE_NUM_SYMBOLS); - - /* Account for additional bits. */ - cost += num_additional_bits; - - return cost; -} - /* - * Writes all compressed matches and literal bytes in a LZX block to the the + * Writes all compressed matches and literal bytes in an LZX block to the the * output bitstream. * - * @out: The output bitstream. - * @block_type: The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM) - * @match_tab[]: The array of matches that will be output. It has length - * of @num_compressed_literals. - * @num_compressed_literals: Number of compressed literals to be output. - * @codes: Pointer to a structure that contains the codewords for the - * main, length, and aligned offset Huffman codes. + * @ostream + * The output bitstream. + * @block_type + * The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM). + * @match_tab + * The array of matches/literals that will be output (length @match_count). + * @match_count + * Number of matches/literals to be output. + * @codes + * Pointer to a structure that contains the codewords for the main, length, + * and aligned offset Huffman codes. */ static void lzx_write_matches_and_literals(struct output_bitstream *ostream, @@ -885,48 +905,49 @@ lzx_write_matches_and_literals(struct output_bitstream *ostream, } } - static void -lzx_assert_codes_valid(const struct lzx_codes * codes) +lzx_assert_codes_valid(const struct lzx_codes * codes, unsigned num_main_syms) { #ifdef ENABLE_LZX_DEBUG unsigned i; - for (i = 0; i < LZX_MAINTREE_NUM_SYMBOLS; i++) - LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_CODEWORD_LEN); + for (i = 0; i < num_main_syms; i++) + LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_MAIN_CODEWORD_LEN); - for (i = 0; i < LZX_LENTREE_NUM_SYMBOLS; i++) - LZX_ASSERT(codes->lens.len[i] <= LZX_MAX_CODEWORD_LEN); + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) + LZX_ASSERT(codes->lens.len[i] <= LZX_MAX_LEN_CODEWORD_LEN); - for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++) - LZX_ASSERT(codes->lens.aligned[i] <= 8); + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) + LZX_ASSERT(codes->lens.aligned[i] <= LZX_MAX_ALIGNED_CODEWORD_LEN); const unsigned tablebits = 10; u16 decode_table[(1 << tablebits) + - (2 * max(LZX_MAINTREE_NUM_SYMBOLS, LZX_LENTREE_NUM_SYMBOLS))] + (2 * max(num_main_syms, LZX_LENCODE_NUM_SYMBOLS))] _aligned_attribute(DECODE_TABLE_ALIGNMENT); LZX_ASSERT(0 == make_huffman_decode_table(decode_table, - LZX_MAINTREE_NUM_SYMBOLS, - tablebits, + num_main_syms, + min(tablebits, LZX_MAINCODE_TABLEBITS), codes->lens.main, - LZX_MAX_CODEWORD_LEN)); + LZX_MAX_MAIN_CODEWORD_LEN)); LZX_ASSERT(0 == make_huffman_decode_table(decode_table, - LZX_LENTREE_NUM_SYMBOLS, - tablebits, + LZX_LENCODE_NUM_SYMBOLS, + min(tablebits, LZX_LENCODE_TABLEBITS), codes->lens.len, - LZX_MAX_CODEWORD_LEN)); + LZX_MAX_LEN_CODEWORD_LEN)); LZX_ASSERT(0 == make_huffman_decode_table(decode_table, - LZX_ALIGNEDTREE_NUM_SYMBOLS, - min(tablebits, 6), + LZX_ALIGNEDCODE_NUM_SYMBOLS, + min(tablebits, LZX_ALIGNEDCODE_TABLEBITS), codes->lens.aligned, - 8)); + LZX_MAX_ALIGNED_CODEWORD_LEN)); #endif /* ENABLE_LZX_DEBUG */ } -/* Write a LZX aligned offset or verbatim block to the output. */ +/* Write an LZX aligned offset or verbatim block to the output. */ static void lzx_write_compressed_block(int block_type, unsigned block_size, + unsigned max_window_size, + unsigned num_main_syms, struct lzx_match * chosen_matches, unsigned num_chosen_matches, const struct lzx_codes * codes, @@ -937,56 +958,70 @@ lzx_write_compressed_block(int block_type, LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED || block_type == LZX_BLOCKTYPE_VERBATIM); - LZX_ASSERT(block_size <= LZX_MAX_WINDOW_SIZE); - LZX_ASSERT(num_chosen_matches <= LZX_MAX_WINDOW_SIZE); - lzx_assert_codes_valid(codes); + lzx_assert_codes_valid(codes, num_main_syms); /* The first three bits indicate the type of block and are one of the * LZX_BLOCKTYPE_* constants. */ - bitstream_put_bits(ostream, block_type, LZX_BLOCKTYPE_NBITS); + bitstream_put_bits(ostream, block_type, 3); - /* The next bit indicates whether the block size is the default (32768), - * indicated by a 1 bit, or whether the block size is given by the next - * 16 bits, indicated by a 0 bit. */ + /* 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. + * + * 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. */ if (block_size == LZX_DEFAULT_BLOCK_SIZE) { bitstream_put_bits(ostream, 1, 1); } else { bitstream_put_bits(ostream, 0, 1); - bitstream_put_bits(ostream, block_size, LZX_BLOCKSIZE_NBITS); + + if (max_window_size >= 65536) + bitstream_put_bits(ostream, block_size >> 16, 8); + + bitstream_put_bits(ostream, block_size, 16); } /* Write out lengths of the main code. Note that the LZX specification * incorrectly states that the aligned offset code comes after the - * length code, but in fact it is the very first tree to be written + * length code, but in fact it is the very first code to be written * (before the main code). */ if (block_type == LZX_BLOCKTYPE_ALIGNED) - for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++) + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) bitstream_put_bits(ostream, codes->lens.aligned[i], - LZX_ALIGNEDTREE_ELEMENT_SIZE); + LZX_ALIGNEDCODE_ELEMENT_SIZE); LZX_DEBUG("Writing main code..."); - /* Write the pre-tree and lengths for the first LZX_NUM_CHARS symbols in + /* Write the precode and lengths for the first LZX_NUM_CHARS symbols in * the main code, which are the codewords for literal bytes. */ lzx_write_compressed_code(ostream, codes->lens.main, prev_codes->lens.main, LZX_NUM_CHARS); - /* Write the pre-tree and lengths for the rest of the main code, which + /* Write the precode and lengths for the rest of the main code, which * are the codewords for match headers. */ lzx_write_compressed_code(ostream, codes->lens.main + LZX_NUM_CHARS, prev_codes->lens.main + LZX_NUM_CHARS, - LZX_MAINTREE_NUM_SYMBOLS - LZX_NUM_CHARS); + num_main_syms - LZX_NUM_CHARS); LZX_DEBUG("Writing length code..."); - /* Write the pre-tree and lengths for the length code. */ + /* Write the precode and lengths for the length code. */ lzx_write_compressed_code(ostream, codes->lens.len, prev_codes->lens.len, - LZX_LENTREE_NUM_SYMBOLS); + LZX_LENCODE_NUM_SYMBOLS); LZX_DEBUG("Writing matches and literals..."); @@ -998,189 +1033,140 @@ lzx_write_compressed_block(int block_type, LZX_DEBUG("Done writing block."); } -/* Write the LZX block of index @block_number, or recurse to its children - * recursively if it is a split block. - * - * Return a pointer to the Huffman codes for the last block written. - */ -static struct lzx_codes * -lzx_write_block_recursive(struct lzx_compressor *ctx, - unsigned block_number, - struct lzx_codes * prev_codes, - struct output_bitstream *ostream) +/* Write out the LZX blocks that were computed. */ +static void +lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream) { - struct lzx_block_spec *spec = &ctx->block_specs[block_number - 1]; - if (spec->is_split) { - prev_codes = lzx_write_block_recursive(ctx, block_number * 2 + 0, - prev_codes, ostream); - prev_codes = lzx_write_block_recursive(ctx, block_number * 2 + 1, - prev_codes, ostream); - } else { - LZX_DEBUG("Writing block #%u (type=%d, size=%u, num_chosen_matches=%u)...", - block_number, spec->block_type, spec->block_size, + const struct lzx_codes *prev_codes = &ctx->zero_codes; + for (unsigned i = 0; i < ctx->num_blocks; i++) { + const struct lzx_block_spec *spec = &ctx->block_specs[i]; + + LZX_DEBUG("Writing block %u/%u (type=%d, size=%u, num_chosen_matches=%u)...", + i + 1, ctx->num_blocks, + spec->block_type, spec->block_size, spec->num_chosen_matches); + lzx_write_compressed_block(spec->block_type, spec->block_size, + ctx->max_window_size, + ctx->num_main_syms, &ctx->chosen_matches[spec->chosen_matches_start_pos], spec->num_chosen_matches, &spec->codes, prev_codes, ostream); + prev_codes = &spec->codes; } - return prev_codes; -} - -/* Write out the LZX blocks that were computed. */ -static void -lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream) -{ - lzx_write_block_recursive(ctx, 1, &ctx->zero_codes, ostream); } +/* Constructs an LZX match from a literal byte and updates the main code symbol + * frequencies. */ static u32 -lzx_record_literal(u8 literal, void *_freqs) +lzx_tally_literal(u8 lit, struct lzx_freqs *freqs) { - struct lzx_freqs *freqs = _freqs; - - freqs->main[literal]++; - - return (u32)literal; + freqs->main[lit]++; + return (u32)lit; } -/* Constructs a match from an offset and a length, and updates the LRU queue and - * the frequency of symbols in the main, length, and aligned offset alphabets. - * The return value is a 32-bit number that provides the match in an +/* Constructs an LZX match from an offset and a length, and updates the LRU + * queue and the frequency of symbols in the main, length, and aligned offset + * alphabets. The return value is a 32-bit number that provides the match in an * intermediate representation documented below. */ static u32 -lzx_record_match(unsigned match_offset, unsigned match_len, - void *_freqs, void *_queue) +lzx_tally_match(unsigned match_len, unsigned match_offset, + struct lzx_freqs *freqs, struct lzx_lru_queue *queue) { - struct lzx_freqs *freqs = _freqs; - struct lzx_lru_queue *queue = _queue; unsigned position_slot; - unsigned position_footer = 0; + unsigned position_footer; u32 len_header; - u32 len_pos_header; + unsigned main_symbol; unsigned len_footer; unsigned adjusted_match_len; - LZX_ASSERT(match_len >= LZX_MIN_MATCH && match_len <= LZX_MAX_MATCH); - - /* If possible, encode this offset as a repeated offset. */ - if (match_offset == queue->R0) { - position_slot = 0; - } else if (match_offset == queue->R1) { - swap(queue->R0, queue->R1); - position_slot = 1; - } else if (match_offset == queue->R2) { - swap(queue->R0, queue->R2); - position_slot = 2; - } else { - /* Not a repeated offset. */ - - /* offsets of 0, 1, and 2 are reserved for the repeated offset - * codes, so non-repeated offsets must be encoded as 3+. The - * minimum offset is 1, so encode the offsets offset by 2. */ - unsigned formatted_offset = match_offset + 2; - - queue->R2 = queue->R1; - queue->R1 = queue->R0; - queue->R0 = match_offset; - - /* The (now-formatted) offset will actually be encoded as a - * small position slot number that maps to a certain hard-coded - * offset (position base), followed by a number of extra bits--- - * the position footer--- that are added to the position base to - * get the original formatted offset. */ - - position_slot = lzx_get_position_slot(formatted_offset); - position_footer = formatted_offset & - ((1 << lzx_get_num_extra_bits(position_slot)) - 1); - } - - adjusted_match_len = match_len - LZX_MIN_MATCH; + LZX_ASSERT(match_len >= LZX_MIN_MATCH_LEN && match_len <= LZX_MAX_MATCH_LEN); + /* The match offset shall be encoded as a position slot (itself encoded + * as part of the main symbol) and a position footer. */ + position_slot = lzx_get_position_slot(match_offset, queue); + position_footer = (match_offset + LZX_OFFSET_OFFSET) & + ((1U << lzx_get_num_extra_bits(position_slot)) - 1); - /* The match length must be at least 2, so let the adjusted match length - * be the match length minus 2. - * - * If it is less than 7, the adjusted match length is encoded as a 3-bit - * number offset by 2. Otherwise, the 3-bit length header is all 1's - * and the actual adjusted length is given as a symbol encoded with the - * length tree, offset by 7. - */ + /* The match length shall be encoded as a length header (itself encoded + * as part of the main symbol) and an optional length footer. */ + adjusted_match_len = match_len - LZX_MIN_MATCH_LEN; if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) { + /* No length footer needed. */ len_header = adjusted_match_len; } else { + /* Length footer needed. It will be encoded using the length + * code. */ len_header = LZX_NUM_PRIMARY_LENS; len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS; freqs->len[len_footer]++; } - len_pos_header = (position_slot << 3) | len_header; - freqs->main[len_pos_header + LZX_NUM_CHARS]++; + /* Account for the main symbol. */ + main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS; + + freqs->main[main_symbol]++; - /* Equivalent to: - * if (lzx_extra_bits[position_slot] >= 3) */ + /* In an aligned offset block, 3 bits of the position footer are output + * as an aligned offset symbol. Account for this, although we may + * ultimately decide to output the block as verbatim. */ + + /* The following check is equivalent to: + * + * if (lzx_extra_bits[position_slot] >= 3) + * + * Note that this correctly excludes position slots that correspond to + * recent offsets. */ if (position_slot >= 8) freqs->aligned[position_footer & 7]++; /* Pack the position slot, position footer, and match length into an - * intermediate representation. - * - * bits description - * ---- ----------------------------------------------------------- - * - * 31 1 if a match, 0 if a literal. - * - * 30-25 position slot. This can be at most 50, so it will fit in 6 - * bits. - * - * 8-24 position footer. This is the offset of the real formatted - * offset from the position base. This can be at most 17 bits - * (since lzx_extra_bits[LZX_NUM_POSITION_SLOTS - 1] is 17). - * - * 0-7 length of match, offset by 2. This can be at most - * (LZX_MAX_MATCH - 2) == 255, so it will fit in 8 bits. */ + * intermediate representation. See `struct lzx_match' for details. + */ + LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64); + LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17); + LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256); + + LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1); + LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1); + LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1); return 0x80000000 | (position_slot << 25) | (position_footer << 8) | (adjusted_match_len); } -/* Set the cost model @ctx->costs from the Huffman codeword lengths specified in - * @lens. - * - * These are basically the same thing, except that Huffman codewords with length - * 0 corresponds to symbols with zero frequency. These 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. - */ +struct lzx_record_ctx { + struct lzx_freqs freqs; + struct lzx_lru_queue queue; + struct lzx_match *matches; +}; + static void -lzx_set_costs(struct lzx_compressor * ctx, const struct lzx_lens * lens) +lzx_record_match(unsigned len, unsigned offset, void *_ctx) { - unsigned i; + struct lzx_record_ctx *ctx = _ctx; - memcpy(&ctx->costs, lens, sizeof(struct lzx_lens)); - - for (i = 0; i < LZX_MAINTREE_NUM_SYMBOLS; i++) - if (ctx->costs.main[i] == 0) - ctx->costs.main[i] = ctx->params.slow.main_nostat_cost; + (ctx->matches++)->data = lzx_tally_match(len, offset, &ctx->freqs, &ctx->queue); +} - for (i = 0; i < LZX_LENTREE_NUM_SYMBOLS; i++) - if (ctx->costs.len[i] == 0) - ctx->costs.len[i] = ctx->params.slow.len_nostat_cost; +static void +lzx_record_literal(u8 lit, void *_ctx) +{ + struct lzx_record_ctx *ctx = _ctx; - for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++) - if (ctx->costs.aligned[i] == 0) - ctx->costs.aligned[i] = ctx->params.slow.aligned_nostat_cost; + (ctx->matches++)->data = lzx_tally_literal(lit, &ctx->freqs); } -static u32 -lzx_literal_cost(u8 c, const struct lzx_lens * costs) +/* Returns the cost, in bits, to output a literal byte using the specified cost + * model. */ +static unsigned +lzx_literal_cost(u8 c, const struct lzx_costs * costs) { return costs->main[c]; } @@ -1188,35 +1174,19 @@ lzx_literal_cost(u8 c, const struct lzx_lens * costs) /* Given a (length, offset) pair that could be turned into a valid LZX match as * well as costs for the codewords in the main, length, and aligned Huffman * codes, return the approximate number of bits it will take to represent this - * match in the compressed output. */ + * match in the compressed output. Take into account the match offset LRU + * queue and optionally update it. */ static unsigned -lzx_match_cost(unsigned length, unsigned offset, const struct lzx_lens *costs - -#if LZX_PARAM_ACCOUNT_FOR_LRU - , struct lzx_lru_queue *queue -#endif - ) +lzx_match_cost(unsigned length, unsigned offset, const struct lzx_costs *costs, + struct lzx_lru_queue *queue) { - unsigned position_slot, len_header, main_symbol; + unsigned position_slot; + unsigned len_header, main_symbol; unsigned cost = 0; - /* Calculate position slot and length header, then combine them into the - * main symbol. */ - -#if LZX_PARAM_ACCOUNT_FOR_LRU - if (offset == queue->R0) { - position_slot = 0; - } else if (offset == queue->R1) { - swap(queue->R0, queue->R1); - position_slot = 1; - } else if (offset == queue->R2) { - swap(queue->R0, queue->R2); - position_slot = 2; - } else -#endif - position_slot = lzx_get_position_slot(offset + 2); + position_slot = lzx_get_position_slot(offset, queue); - len_header = min(length - LZX_MIN_MATCH, LZX_NUM_PRIMARY_LENS); + len_header = min(length - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS); main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS; /* Account for main symbol. */ @@ -1226,385 +1196,375 @@ lzx_match_cost(unsigned length, unsigned offset, const struct lzx_lens *costs unsigned num_extra_bits = lzx_get_num_extra_bits(position_slot); if (num_extra_bits >= 3) { cost += num_extra_bits - 3; - cost += costs->aligned[(offset + LZX_MIN_MATCH) & 7]; + cost += costs->aligned[(offset + LZX_OFFSET_OFFSET) & 7]; } else { cost += num_extra_bits; } /* Account for extra length information. */ - if (length - LZX_MIN_MATCH >= LZX_NUM_PRIMARY_LENS) - cost += costs->len[length - LZX_MIN_MATCH - LZX_NUM_PRIMARY_LENS]; + if (len_header == LZX_NUM_PRIMARY_LENS) + cost += costs->len[length - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS]; return cost; + } -/* This procedure effectively creates a new binary tree corresponding to the - * current string at the same time that it searches the existing tree nodes for - * matches. */ -static unsigned -lzx_lz_get_matches(const u8 window[restrict], - const unsigned bytes_remaining, - const unsigned strstart, - const unsigned max_length, - u16 child_tab[restrict], - unsigned cur_match, - const unsigned prev_len, - struct raw_match * const matches) +/* Fast heuristic cost evaluation to use in the inner loop of the match-finder. + * Unlike lzx_match_cost() which does a true cost evaluation, this simply + * prioritize matches based on their offset. */ +static block_cost_t +lzx_match_cost_fast(unsigned offset, const struct lzx_lru_queue *queue) +{ + /* It seems well worth it to take the time to give priority to recently + * used offsets. */ + for (unsigned i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) + if (offset == queue->R[i]) + return i; + + BUILD_BUG_ON(LZX_MAX_WINDOW_SIZE >= (block_cost_t)~0U); + return offset; +} + +/* Set the cost model @ctx->costs from the Huffman codeword lengths specified in + * @lens. + * + * 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 * ctx, const struct lzx_lens * lens) { - u16 *new_tree_lt_ptr = &child_tab[strstart * 2]; - u16 *new_tree_gt_ptr = &child_tab[strstart * 2 + 1]; + unsigned i; + unsigned num_main_syms = ctx->num_main_syms; - u16 longest_lt_match_len = 0; - u16 longest_gt_match_len = 0; + /* Main code */ + for (i = 0; i < num_main_syms; i++) { + ctx->costs.main[i] = lens->main[i]; + if (ctx->costs.main[i] == 0) + ctx->costs.main[i] = ctx->params.alg_params.slow.main_nostat_cost; + } - /* Maximum number of nodes to walk down before stopping */ - unsigned depth = max_length; + /* Length code */ + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) { + ctx->costs.len[i] = lens->len[i]; + if (ctx->costs.len[i] == 0) + ctx->costs.len[i] = ctx->params.alg_params.slow.len_nostat_cost; + } - /* Length of longest match found so far */ - unsigned longest_match_len = prev_len; + /* Aligned offset code */ + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + ctx->costs.aligned[i] = lens->aligned[i]; + if (ctx->costs.aligned[i] == 0) + ctx->costs.aligned[i] = ctx->params.alg_params.slow.aligned_nostat_cost; + } +} - /* Maximum length of match to return */ - unsigned len_limit = min(bytes_remaining, max_length); +/* Advance the suffix array match-finder to the next position. */ +static void +lzx_lz_update_salink(input_idx_t i, + const input_idx_t SA[restrict], + const input_idx_t ISA[restrict], + struct salink link[restrict]) +{ + /* r = Rank of the suffix at the current position. */ + const input_idx_t r = ISA[i]; - /* Number of matches found so far */ - unsigned num_matches = 0; + /* next = rank of LOWEST ranked suffix that is ranked HIGHER than the + * current suffix AND has a LOWER position, or -1 if none exists. */ + const input_idx_t next = link[r].next; - for (;;) { + /* prev = rank of HIGHEST ranked suffix that is ranked LOWER than the + * current suffix AND has a LOWER position, or -1 if none exists. */ + const input_idx_t prev = link[r].prev; - /* Stop if too many nodes were traversed or if there is no next - * node */ - if (depth-- == 0 || cur_match == 0) { - *new_tree_gt_ptr = 0; - *new_tree_lt_ptr = 0; - return num_matches; - } + /* Link the suffix at the current position into the linked list that + * contains all suffixes in the suffix array that are appear at or + * before the current position, sorted by rank. + * + * Save the values of all fields we overwrite so that rollback is + * possible. */ + if (next != (input_idx_t)~0U) { - /* Load the pointers to the children of the binary tree node - * corresponding to the current match */ - u16 * const cur_match_ptrs = &child_tab[cur_match * 2]; - - /* Set up pointers to the current match and to the current - * string */ - const u8 * const matchptr = &window[cur_match]; - const u8 * const strptr = &window[strstart]; - - u16 len = min(longest_lt_match_len, - longest_gt_match_len); - - if (matchptr[len] == strptr[len]) { - while (++len != len_limit) - if (matchptr[len] != strptr[len]) - break; - - if (len > longest_match_len) { - longest_match_len = len; - matches[num_matches].len = len; - matches[num_matches].offset = strstart - cur_match; - num_matches++; - - if (len == len_limit) { - /* Length limit was reached. Link left pointer - * in the new tree with left subtree of current - * match tree, and link the right pointer in the - * new tree with the right subtree of the - * current match tree. This in effect deletes - * the node for the currrent match, which is - * desirable because the current match is the - * same as the current string up until the - * length limit, so in subsequent queries it - * will never be preferable to the current - * position. */ - *new_tree_lt_ptr = cur_match_ptrs[0]; - *new_tree_gt_ptr = cur_match_ptrs[1]; - return num_matches; - } - } - } + link[next].prev = r; + link[next].lcpprev = link[r].lcpnext; + } - if (matchptr[len] < strptr[len]) { - /* Case 1: The current match is lexicographically less - * than the current string. - * - * Since we are searching the binary tree structures, we - * need to walk down to the *right* subtree of the - * current match's node to get to a match that is - * lexicographically *greater* than the current match - * but still lexicographically *lesser* than the current - * string. - * - * At the same time, we link the entire binary tree - * corresponding to the current match into the - * appropriate place in the new binary tree being built - * for the current string. */ - *new_tree_lt_ptr = cur_match; - new_tree_lt_ptr = &cur_match_ptrs[1]; - cur_match = *new_tree_lt_ptr; - longest_lt_match_len = len; - } else { - /* Case 2: The current match is lexicographically - * greater than the current string. - * - * This is analogous to Case 1 above, but everything - * happens in the other direction. - */ - *new_tree_gt_ptr = cur_match; - new_tree_gt_ptr = &cur_match_ptrs[0]; - cur_match = *new_tree_gt_ptr; - longest_gt_match_len = len; - } + if (prev != (input_idx_t)~0U) { + + link[prev].next = r; + link[prev].lcpnext = link[r].lcpprev; } } -/* Equivalent to lzx_lz_get_matches(), but only updates the tree and doesn't - * return matches. See that function for details (including comments). */ -static void -lzx_lz_skip_matches(const u8 window[restrict], - const unsigned bytes_remaining, - const unsigned strstart, - const unsigned max_length, - u16 child_tab[restrict], - unsigned cur_match, - const unsigned prev_len) +/* + * Use the suffix array match-finder to retrieve a list of LZ matches at the + * current position. + * + * [in] @i Current position in the window. + * [in] @SA Suffix array for the window. + * [in] @ISA Inverse suffix array for the window. + * [inout] @link Suffix array links used internally by the match-finder. + * [out] @matches The (length, offset) pairs of the resulting matches will + * be written here, sorted in decreasing order by + * length. All returned lengths will be unique. + * [in] @queue Recently used match offsets, used when evaluating the + * cost of matches. + * [in] @min_match_len Minimum match length to return. + * [in] @max_matches_to_consider Maximum number of matches to consider at + * the position. + * [in] @max_matches_to_return Maximum number of matches to return. + * + * The return value is the number of matches found and written to @matches. + */ +static unsigned +lzx_lz_get_matches(const input_idx_t i, + const input_idx_t SA[const restrict], + const input_idx_t ISA[const restrict], + struct salink link[const restrict], + struct raw_match matches[const restrict], + const struct lzx_lru_queue * const restrict queue, + const unsigned min_match_len, + const u32 max_matches_to_consider, + const u32 max_matches_to_return) { - u16 *new_tree_lt_ptr = &child_tab[strstart * 2]; - u16 *new_tree_gt_ptr = &child_tab[strstart * 2 + 1]; + /* r = Rank of the suffix at the current position. */ + const input_idx_t r = ISA[i]; - u16 longest_lt_match_len = 0; - u16 longest_gt_match_len = 0; + /* Prepare for searching the current position. */ + lzx_lz_update_salink(i, SA, ISA, link); - unsigned depth = max_length; + /* L = rank of next suffix to the left; + * R = rank of next suffix to the right; + * lenL = length of match between current position and the suffix with rank L; + * lenR = length of match between current position and the suffix with rank R. + * + * This is left and right relative to the rank of the current suffix. + * Since the suffixes in the suffix array are sorted, the longest + * matches are immediately to the left and right (using the linked list + * to ignore all suffixes that occur later in the window). The match + * length decreases the farther left and right we go. We shall keep the + * length on both sides in sync in order to choose the lowest-cost match + * of each length. + */ + input_idx_t L = link[r].prev; + input_idx_t R = link[r].next; + input_idx_t lenL = link[r].lcpprev; + input_idx_t lenR = link[r].lcpnext; - unsigned longest_match_len = prev_len; + /* nmatches = number of matches found so far. */ + unsigned nmatches = 0; - unsigned len_limit = min(bytes_remaining, max_length); + /* best_cost = cost of lowest-cost match found so far. + * + * We keep track of this so that we can ignore shorter matches that do + * not have lower costs than a longer matches already found. + */ + block_cost_t best_cost = INFINITE_BLOCK_COST; - for (;;) { - if (depth-- == 0 || cur_match == 0) { - *new_tree_gt_ptr = 0; - *new_tree_lt_ptr = 0; - return; - } + /* count_remaining = maximum number of possible matches remaining to be + * considered. */ + u32 count_remaining = max_matches_to_consider; - u16 * const cur_match_ptrs = &child_tab[cur_match * 2]; + /* pending = match currently being considered for a specific length. */ + struct raw_match pending; + block_cost_t pending_cost; + + while (lenL >= min_match_len || lenR >= min_match_len) + { + pending.len = lenL; + pending_cost = INFINITE_BLOCK_COST; + block_cost_t cost; - const u8 * const matchptr = &window[cur_match]; - const u8 * const strptr = &window[strstart]; + /* Extend left. */ + if (lenL >= min_match_len && lenL >= lenR) { + for (;;) { - u16 len = min(longest_lt_match_len, - longest_gt_match_len); + if (--count_remaining == 0) + goto out_save_pending; - if (matchptr[len] == strptr[len]) { - while (++len != len_limit) - if (matchptr[len] != strptr[len]) - break; + input_idx_t offset = i - SA[L]; - if (len > longest_match_len) { - longest_match_len = len; + /* Save match if it has smaller cost. */ + cost = lzx_match_cost_fast(offset, queue); + if (cost < pending_cost) { + pending.offset = offset; + pending_cost = cost; + } - if (len == len_limit) { - *new_tree_lt_ptr = cur_match_ptrs[0]; - *new_tree_gt_ptr = cur_match_ptrs[1]; - return; + if (link[L].lcpprev < lenL) { + /* Match length decreased. */ + + lenL = link[L].lcpprev; + + /* Save the pending match unless the + * right side still may have matches of + * this length to be scanned, or if a + * previous (longer) match had lower + * cost. */ + if (pending.len > lenR) { + if (pending_cost < best_cost) { + best_cost = pending_cost; + matches[nmatches++] = pending; + if (nmatches == max_matches_to_return) + return nmatches; + } + pending.len = lenL; + pending_cost = INFINITE_BLOCK_COST; + } + if (lenL < min_match_len || lenL < lenR) + break; } + L = link[L].prev; } } - if (matchptr[len] < strptr[len]) { - *new_tree_lt_ptr = cur_match; - new_tree_lt_ptr = &cur_match_ptrs[1]; - cur_match = *new_tree_lt_ptr; - longest_lt_match_len = len; - } else { - *new_tree_gt_ptr = cur_match; - new_tree_gt_ptr = &cur_match_ptrs[0]; - cur_match = *new_tree_gt_ptr; - longest_gt_match_len = len; + pending.len = lenR; + + /* Extend right. */ + if (lenR >= min_match_len && lenR > lenL) { + for (;;) { + + if (--count_remaining == 0) + goto out_save_pending; + + input_idx_t offset = i - SA[R]; + + /* Save match if it has smaller cost. */ + cost = lzx_match_cost_fast(offset, queue); + if (cost < pending_cost) { + pending.offset = offset; + pending_cost = cost; + } + + if (link[R].lcpnext < lenR) { + /* Match length decreased. */ + + lenR = link[R].lcpnext; + + /* Save the pending match unless a + * previous (longer) match had lower + * cost. */ + if (pending_cost < best_cost) { + matches[nmatches++] = pending; + best_cost = pending_cost; + if (nmatches == max_matches_to_return) + return nmatches; + } + + if (lenR < min_match_len || lenR <= lenL) + break; + + pending.len = lenR; + pending_cost = INFINITE_BLOCK_COST; + } + R = link[R].next; + } } } + goto out; + +out_save_pending: + if (pending_cost != INFINITE_BLOCK_COST) + matches[nmatches++] = pending; + +out: + return nmatches; } -static unsigned -lzx_lz_get_matches_caching(struct lzx_compressor *ctx, - struct raw_match **matches_ret); /* Tell the match-finder to skip the specified number of bytes (@n) in the * input. */ static void lzx_lz_skip_bytes(struct lzx_compressor *ctx, unsigned n) { - -#if LZX_PARAM_DONT_SKIP_MATCHES - /* Option 1: Still cache the matches from the positions skipped. They - * will then be available in later passes. */ - struct raw_match *matches; - while (n--) - lzx_lz_get_matches_caching(ctx, &matches); -#else - /* Option 2: Simply mark the positions skipped as having no matches - * available. */ LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos); - if (ctx->matches_already_found) { + if (ctx->matches_cached) { + ctx->match_window_pos += n; while (n--) { - LZX_ASSERT(ctx->cached_matches[ctx->cached_matches_pos].offset == - ctx->match_window_pos); - ctx->cached_matches_pos += ctx->cached_matches[ctx->cached_matches_pos].len + 1; - ctx->match_window_pos++; + ctx->cached_matches_pos += + ctx->cached_matches[ctx->cached_matches_pos].len + 1; } } else { while (n--) { - if (ctx->params.slow.use_len2_matches && - ctx->match_window_end - ctx->match_window_pos >= 2) { - unsigned c1 = ctx->window[ctx->match_window_pos]; - unsigned c2 = ctx->window[ctx->match_window_pos + 1]; - unsigned digram = c1 | (c2 << 8); - ctx->digram_tab[digram] = ctx->match_window_pos; - } - if (ctx->match_window_end - ctx->match_window_pos >= 3) { - unsigned hash; - unsigned cur_match; - - hash = lzx_lz_compute_hash(&ctx->window[ctx->match_window_pos]); - - cur_match = ctx->hash_tab[hash]; - ctx->hash_tab[hash] = ctx->match_window_pos; - - lzx_lz_skip_matches(ctx->window, - ctx->match_window_end - ctx->match_window_pos, - ctx->match_window_pos, - ctx->params.slow.num_fast_bytes, - ctx->child_tab, - cur_match, 1); - } - ctx->cached_matches[ctx->cached_matches_pos].len = 0; - ctx->cached_matches[ctx->cached_matches_pos].offset = ctx->match_window_pos; - ctx->cached_matches_pos++; - ctx->match_window_pos++; + ctx->cached_matches[ctx->cached_matches_pos++].len = 0; + lzx_lz_update_salink(ctx->match_window_pos++, ctx->SA, + ctx->ISA, ctx->salink); } } -#endif /* !LZX_PARAM_DONT_SKIP_MATCHES */ } /* Retrieve a list of matches available at the next position in the input. * - * The return value is the number of matches found, and a pointer to them is - * written to @matches_ret. The matches will be sorted in order by length. - * - * This is essentially a wrapper around lzx_lz_get_matches() that caches its - * output the first time and also performs the needed hashing. - */ + * The matches are written to ctx->matches in decreasing order of length, and + * the return value is the number of matches found. */ static unsigned lzx_lz_get_matches_caching(struct lzx_compressor *ctx, + const struct lzx_lru_queue *queue, struct raw_match **matches_ret) { unsigned num_matches; struct raw_match *matches; - LZX_ASSERT(ctx->match_window_end >= ctx->match_window_pos); + LZX_ASSERT(ctx->match_window_pos <= ctx->match_window_end); matches = &ctx->cached_matches[ctx->cached_matches_pos + 1]; - if (ctx->matches_already_found) { - num_matches = ctx->cached_matches[ctx->cached_matches_pos].len; - LZX_ASSERT(ctx->cached_matches[ctx->cached_matches_pos].offset == ctx->match_window_pos); - - for (int i = (int)num_matches - 1; i >= 0; i--) { - if (ctx->match_window_pos + matches[i].len > ctx->match_window_end) - matches[i].len = ctx->match_window_end - ctx->match_window_pos; - else - break; - } + if (ctx->matches_cached) { + num_matches = matches[-1].len; } else { - unsigned prev_len = 1; - struct raw_match * matches_ret = &ctx->cached_matches[ctx->cached_matches_pos + 1]; - num_matches = 0; - - if (ctx->params.slow.use_len2_matches && - ctx->match_window_end - ctx->match_window_pos >= 3) { - unsigned c1 = ctx->window[ctx->match_window_pos]; - unsigned c2 = ctx->window[ctx->match_window_pos + 1]; - unsigned digram = c1 | (c2 << 8); - unsigned cur_match; - - cur_match = ctx->digram_tab[digram]; - ctx->digram_tab[digram] = ctx->match_window_pos; - if (cur_match != 0 && - ctx->window[cur_match + 2] != ctx->window[ctx->match_window_pos + 2]) - { - matches_ret->len = 2; - matches_ret->offset = ctx->match_window_pos - cur_match; - matches_ret++; - num_matches++; - prev_len = 2; - } - } - if (ctx->match_window_end - ctx->match_window_pos >= 3) { - unsigned hash; - unsigned cur_match; - - hash = lzx_lz_compute_hash(&ctx->window[ctx->match_window_pos]); - - cur_match = ctx->hash_tab[hash]; - ctx->hash_tab[hash] = ctx->match_window_pos; - num_matches += lzx_lz_get_matches(ctx->window, - ctx->match_window_end - ctx->match_window_pos, - ctx->match_window_pos, - ctx->params.slow.num_fast_bytes, - ctx->child_tab, - cur_match, - prev_len, - matches_ret); - } - - ctx->cached_matches[ctx->cached_matches_pos].len = num_matches; - ctx->cached_matches[ctx->cached_matches_pos].offset = ctx->match_window_pos; - - if (num_matches) { - struct raw_match *longest_match_ptr = - &ctx->cached_matches[ctx->cached_matches_pos + 1 + - num_matches - 1]; - u16 len = longest_match_ptr->len; - - /* If the longest match returned by the match-finder - * reached the number of fast bytes, extend it as much - * as possible. */ - if (len == ctx->params.slow.num_fast_bytes) { - const unsigned maxlen = - min(ctx->match_window_end - ctx->match_window_pos, - LZX_MAX_MATCH); - - const u8 * const matchptr = - &ctx->window[ctx->match_window_pos - longest_match_ptr->offset]; - - const u8 * const strptr = - &ctx->window[ctx->match_window_pos]; - - while (len < maxlen && matchptr[len] == strptr[len]) - len++; - } - longest_match_ptr->len = len; - } + unsigned min_match_len = LZX_MIN_MATCH_LEN; + if (!ctx->params.alg_params.slow.use_len2_matches) + min_match_len = max(min_match_len, 3); + const u32 max_search_depth = ctx->params.alg_params.slow.max_search_depth; + const u32 max_matches_per_pos = ctx->params.alg_params.slow.max_matches_per_pos; + + if (unlikely(max_search_depth == 0 || max_matches_per_pos == 0)) + num_matches = 0; + else + num_matches = lzx_lz_get_matches(ctx->match_window_pos, + ctx->SA, + ctx->ISA, + ctx->salink, + matches, + queue, + min_match_len, + max_search_depth, + max_matches_per_pos); + matches[-1].len = num_matches; } ctx->cached_matches_pos += num_matches + 1; *matches_ret = matches; + /* Cap the length of returned matches to the number of bytes remaining, + * if it is not the whole window. */ + if (ctx->match_window_end < ctx->window_size) { + unsigned maxlen = ctx->match_window_end - ctx->match_window_pos; + for (unsigned i = 0; i < num_matches; i++) + if (matches[i].len > maxlen) + matches[i].len = maxlen; + } #if 0 - printf("\n"); + fprintf(stderr, "Pos %u/%u: %u matches\n", + ctx->match_window_pos, ctx->match_window_end, num_matches); for (unsigned i = 0; i < num_matches; i++) - { - printf("Len %u Offset %u\n", matches[i].len, matches[i].offset); - } + fprintf(stderr, "\tLen %u Offset %u\n", matches[i].len, matches[i].offset); #endif +#ifdef ENABLE_LZX_DEBUG for (unsigned i = 0; i < num_matches; i++) { - LZX_ASSERT(matches[i].len <= LZX_MAX_MATCH); - if (matches[i].len >= LZX_MIN_MATCH) { - LZX_ASSERT(matches[i].offset <= ctx->match_window_pos); - LZX_ASSERT(matches[i].len <= ctx->match_window_end - ctx->match_window_pos); - LZX_ASSERT(!memcmp(&ctx->window[ctx->match_window_pos], - &ctx->window[ctx->match_window_pos - matches[i].offset], - matches[i].len)); - } + LZX_ASSERT(matches[i].len >= LZX_MIN_MATCH_LEN); + LZX_ASSERT(matches[i].len <= LZX_MAX_MATCH_LEN); + LZX_ASSERT(matches[i].len <= ctx->match_window_end - ctx->match_window_pos); + LZX_ASSERT(matches[i].offset > 0); + LZX_ASSERT(matches[i].offset <= ctx->match_window_pos); + LZX_ASSERT(!memcmp(&ctx->window[ctx->match_window_pos], + &ctx->window[ctx->match_window_pos - matches[i].offset], + matches[i].len)); } +#endif ctx->match_window_pos++; return num_matches; @@ -1652,21 +1612,22 @@ lzx_lz_reverse_near_optimal_match_list(struct lzx_compressor *ctx, /* * lzx_lz_get_near_optimal_match() - * - * Choose the "best" match or literal to use at the next position in the input. + * Choose the optimal match or literal to use at the next position in the input. * - * Unlike a "greedy" parser that always takes the longest match, or even a + * Unlike a greedy parser that always takes the longest match, or even a * parser with one match/literal look-ahead like zlib, the algorithm used here - * may look ahead many matches/literals to determine the best match/literal to + * may look ahead many matches/literals to determine the optimal match/literal to * output next. The motivation is that the compression ratio is improved if the * compressor can do things like use a shorter-than-possible match in order to * allow a longer match later, and also take into account the Huffman code cost - * model rather than simply assuming that longer is better. It is not a true - * "optimal" parser, however, since some shortcuts can be taken; for example, if - * a match is very long, the parser just chooses it immediately before too much - * time is wasting considering many different alternatives that are unlikely to - * be better. + * model rather than simply assuming that longer is better. + * + * Still, this is not truly an optimal parser because very long matches are + * taken immediately, and the raw match-finder takes some shortcuts. This is + * done to avoid considering many different alternatives that are unlikely to + * be significantly better. * - * This algorithm is based on that used in 7-Zip's deflate encoder. + * This algorithm is based on that used in 7-Zip's DEFLATE encoder. * * Each call to this function does one of two things: * @@ -1686,45 +1647,20 @@ lzx_lz_reverse_near_optimal_match_list(struct lzx_compressor *ctx, * ctx->optimum (internal state; leave uninitialized) * ctx->optimum_cur_idx (must set to 0 before first call) * ctx->optimum_end_idx (must set to 0 before first call) - * ctx->hash_tab (must set to 0 before first call) - * ctx->cached_matches (internal state; leave uninitialized) - * ctx->cached_matches_pos (initialize to 0 before first call; save and - * restore value if restarting parse from a - * certain position) - * ctx->match_window_pos (must initialize to position of next match to - * return; subsequent calls return subsequent - * matches) - * ctx->match_window_end (must initialize to limit of match-finding region; - * subsequent calls use the same limit) + * + * Plus any state used by the raw match-finder. * * The return value is a (length, offset) pair specifying the match or literal - * chosen. + * chosen. For literals, the length is less than LZX_MIN_MATCH_LEN and the + * offset is meaningless. */ static struct raw_match lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) { -#if 0 - /* Testing: literals only */ - ctx->match_window_pos++; - return (struct raw_match) { .len = 0 }; -#elif 0 - /* Testing: greedy parsing */ - struct raw_match *matches; - unsigned num_matches; - struct raw_match match = {.len = 0}; - - num_matches = lzx_lz_get_matches_caching(ctx, &matches); - if (num_matches) { - match = matches[num_matches - 1]; - lzx_lz_skip_bytes(ctx, match.len - 1); - } - return match; -#else unsigned num_possible_matches; struct raw_match *possible_matches; struct raw_match match; unsigned longest_match_len; - unsigned len, match_idx; if (ctx->optimum_cur_idx != ctx->optimum_end_idx) { /* Case 2: Return the next match/literal already found. */ @@ -1742,30 +1678,28 @@ lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) ctx->optimum_end_idx = 0; /* Get matches at this position. */ - num_possible_matches = lzx_lz_get_matches_caching(ctx, &possible_matches); + num_possible_matches = lzx_lz_get_matches_caching(ctx, &ctx->queue, &possible_matches); /* If no matches found, return literal. */ if (num_possible_matches == 0) return (struct raw_match){ .len = 0 }; - /* The matches that were found are sorted by length. Get the length of - * the longest one. */ - longest_match_len = possible_matches[num_possible_matches - 1].len; + /* The matches that were found are sorted in decreasing order by length. + * Get the length of the longest one. */ + longest_match_len = possible_matches[0].len; /* Greedy heuristic: if the longest match that was found is greater - * than LZX_PARAM_NUM_FAST_BYTES, return it immediately; don't both + * than the number of fast bytes, return it immediately; don't both * doing more work. */ - if (longest_match_len > ctx->params.slow.num_fast_bytes) { + if (longest_match_len > ctx->params.alg_params.slow.num_fast_bytes) { lzx_lz_skip_bytes(ctx, longest_match_len - 1); - return possible_matches[num_possible_matches - 1]; + return possible_matches[0]; } /* Calculate the cost to reach the next position by outputting a * literal. */ -#if LZX_PARAM_ACCOUNT_FOR_LRU ctx->optimum[0].queue = ctx->queue; ctx->optimum[1].queue = ctx->optimum[0].queue; -#endif ctx->optimum[1].cost = lzx_literal_cost(ctx->window[ctx->match_window_pos], &ctx->costs); ctx->optimum[1].prev.link = 0; @@ -1773,26 +1707,21 @@ lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) /* Calculate the cost to reach any position up to and including that * reached by the longest match, using the shortest (i.e. closest) match * that reaches each position. */ - match_idx = 0; - BUILD_BUG_ON(LZX_MIN_MATCH != 2); - for (len = LZX_MIN_MATCH; len <= longest_match_len; len++) { + BUILD_BUG_ON(LZX_MIN_MATCH_LEN != 2); + for (unsigned len = LZX_MIN_MATCH_LEN, match_idx = num_possible_matches - 1; + len <= longest_match_len; len++) { LZX_ASSERT(match_idx < num_possible_matches); - #if LZX_PARAM_ACCOUNT_FOR_LRU ctx->optimum[len].queue = ctx->optimum[0].queue; - #endif ctx->optimum[len].prev.link = 0; ctx->optimum[len].prev.match_offset = possible_matches[match_idx].offset; ctx->optimum[len].cost = lzx_match_cost(len, possible_matches[match_idx].offset, - &ctx->costs - #if LZX_PARAM_ACCOUNT_FOR_LRU - , &ctx->optimum[len].queue - #endif - ); + &ctx->costs, + &ctx->optimum[len].queue); if (len == possible_matches[match_idx].len) - match_idx++; + match_idx--; } unsigned cur_pos = 0; @@ -1801,26 +1730,25 @@ lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) * so far */ unsigned len_end = longest_match_len; - for (;;) { /* Advance to next position. */ cur_pos++; - if (cur_pos == len_end || cur_pos == LZX_PARAM_OPTIM_ARRAY_SIZE) + if (cur_pos == len_end || cur_pos == LZX_OPTIM_ARRAY_SIZE) return lzx_lz_reverse_near_optimal_match_list(ctx, cur_pos); /* retrieve the number of matches available at this position */ - num_possible_matches = lzx_lz_get_matches_caching(ctx, + num_possible_matches = lzx_lz_get_matches_caching(ctx, &ctx->optimum[cur_pos].queue, &possible_matches); unsigned new_len = 0; if (num_possible_matches != 0) { - new_len = possible_matches[num_possible_matches - 1].len; + new_len = possible_matches[0].len; /* Greedy heuristic: if we found a match greater than - * LZX_PARAM_NUM_FAST_BYTES, stop immediately. */ - if (new_len > ctx->params.slow.num_fast_bytes) { + * the number of fast bytes, stop immediately. */ + if (new_len > ctx->params.alg_params.slow.num_fast_bytes) { /* Build the list of matches to return and get * the first one. */ @@ -1828,7 +1756,7 @@ lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) /* Append the long match to the end of the list. */ ctx->optimum[cur_pos].next.match_offset = - possible_matches[num_possible_matches - 1].offset; + possible_matches[0].offset; ctx->optimum[cur_pos].next.link = cur_pos + new_len; ctx->optimum_end_idx = cur_pos + new_len; @@ -1841,16 +1769,14 @@ lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) } /* Consider proceeding with a literal byte. */ - u32 cur_cost = ctx->optimum[cur_pos].cost; - u32 cur_plus_literal_cost = cur_cost + + block_cost_t cur_cost = ctx->optimum[cur_pos].cost; + block_cost_t cur_plus_literal_cost = cur_cost + lzx_literal_cost(ctx->window[ctx->match_window_pos - 1], &ctx->costs); if (cur_plus_literal_cost < ctx->optimum[cur_pos + 1].cost) { ctx->optimum[cur_pos + 1].cost = cur_plus_literal_cost; ctx->optimum[cur_pos + 1].prev.link = cur_pos; - #if LZX_PARAM_ACCOUNT_FOR_LRU ctx->optimum[cur_pos + 1].queue = ctx->optimum[cur_pos].queue; - #endif } if (num_possible_matches == 0) @@ -1859,598 +1785,396 @@ lzx_lz_get_near_optimal_match(struct lzx_compressor * ctx) /* Consider proceeding with a match. */ while (len_end < cur_pos + new_len) - ctx->optimum[++len_end].cost = ~(u32)0; + ctx->optimum[++len_end].cost = INFINITE_BLOCK_COST; - match_idx = 0; - for (len = LZX_MIN_MATCH; len <= new_len; len++) { + for (unsigned len = LZX_MIN_MATCH_LEN, match_idx = num_possible_matches - 1; + len <= new_len; len++) { LZX_ASSERT(match_idx < num_possible_matches); - #if LZX_PARAM_ACCOUNT_FOR_LRU struct lzx_lru_queue q = ctx->optimum[cur_pos].queue; - #endif - u32 cost = cur_cost + lzx_match_cost(len, - possible_matches[match_idx].offset, - &ctx->costs - #if LZX_PARAM_ACCOUNT_FOR_LRU - , &q - #endif - ); + block_cost_t cost = cur_cost + lzx_match_cost(len, + possible_matches[match_idx].offset, + &ctx->costs, + &q); if (cost < ctx->optimum[cur_pos + len].cost) { ctx->optimum[cur_pos + len].cost = cost; ctx->optimum[cur_pos + len].prev.link = cur_pos; ctx->optimum[cur_pos + len].prev.match_offset = possible_matches[match_idx].offset; - #if LZX_PARAM_ACCOUNT_FOR_LRU ctx->optimum[cur_pos + len].queue = q; - #endif } if (len == possible_matches[match_idx].len) - match_idx++; + match_idx--; } } -#endif } -/* Account for extra bits in the main symbols. */ +/* + * Set default symbol costs. + */ static void -lzx_update_mainsym_match_costs(int block_type, - u8 main_lens[LZX_MAINTREE_NUM_SYMBOLS]) +lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms) { unsigned i; - LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED || - block_type == LZX_BLOCKTYPE_VERBATIM); + /* Literal symbols */ + for (i = 0; i < LZX_NUM_CHARS; i++) + costs->main[i] = 8; - for (i = LZX_NUM_CHARS; i < LZX_MAINTREE_NUM_SYMBOLS; i++) { - unsigned position_slot = (i >> 3) & 0x1f; + /* Match header symbols */ + for (; i < num_main_syms; i++) + costs->main[i] = 10; - /* If it's a verbatim block, add the number of extra bits - * corresponding to the position slot. - * - * If it's an aligned block and there would normally be at least - * 3 extra bits, count 3 less because they will be output as an - * aligned offset symbol instead. */ - unsigned num_extra_bits = lzx_get_num_extra_bits(position_slot); + /* Length symbols */ + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) + costs->len[i] = 8; - if (block_type == LZX_BLOCKTYPE_ALIGNED && num_extra_bits >= 3) - num_extra_bits -= 3; - main_lens[i] += num_extra_bits; - } + /* Aligned offset symbols */ + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) + costs->aligned[i] = 3; } -/* - * Compute the costs, in bits, to output a compressed block as aligned offset - * and verbatim. - * - * @block_size - * Number of bytes of uncompressed data this block represents. - * @codes - * Huffman codes that will be used to output the block. - * @prev_codes - * Huffman codes for the previous block, or all zeroes if this is the first - * block. - * @freqs - * Frequencies of Huffman symbol that will be output in the block. - * @aligned_cost_ret - * Cost of aligned block will be returned here. - * @verbatim_cost_ret - * Cost of verbatim block will be returned here. - */ -static void -lzx_compute_compressed_block_costs(unsigned block_size, - const struct lzx_codes *codes, - const struct lzx_codes *prev_codes, - const struct lzx_freqs *freqs, - unsigned * aligned_cost_ret, - unsigned * verbatim_cost_ret) +/* Given the frequencies of symbols in a 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) { - unsigned common_cost = 0; unsigned aligned_cost = 0; unsigned verbatim_cost = 0; - u8 updated_main_lens[LZX_MAINTREE_NUM_SYMBOLS]; - - /* Account for cost of block header. */ - common_cost += LZX_BLOCKTYPE_NBITS; - if (block_size == LZX_DEFAULT_BLOCK_SIZE) - common_cost += 1; + /* Verbatim blocks have a constant 3 bits per position footer. Aligned + * offset blocks have an aligned offset symbol per position footer, plus + * an extra 24 bits to output the lengths necessary to reconstruct the + * aligned offset code itself. */ + 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]; + } + aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS; + if (aligned_cost < verbatim_cost) + return LZX_BLOCKTYPE_ALIGNED; else - common_cost += LZX_BLOCKSIZE_NBITS; - - /* Account for cost of outputting aligned offset code. */ - aligned_cost += LZX_ALIGNEDTREE_NUM_SYMBOLS * LZX_ALIGNEDTREE_ELEMENT_SIZE; - - /* Account for cost of outputting main and length codes. */ - common_cost += lzx_code_cost(codes->lens.main, - prev_codes->lens.main, - LZX_NUM_CHARS); - common_cost += lzx_code_cost(codes->lens.main + LZX_NUM_CHARS, - prev_codes->lens.main + LZX_NUM_CHARS, - LZX_MAINTREE_NUM_SYMBOLS - LZX_NUM_CHARS); - common_cost += lzx_code_cost(codes->lens.len, - prev_codes->lens.len, - LZX_LENTREE_NUM_SYMBOLS); - - /* Account for cost to output main, length, and aligned symbols, taking - * into account extra position bits. */ - - memcpy(updated_main_lens, codes->lens.main, LZX_MAINTREE_NUM_SYMBOLS); - lzx_update_mainsym_match_costs(LZX_BLOCKTYPE_VERBATIM, updated_main_lens); - verbatim_cost += lzx_huffman_code_output_cost(updated_main_lens, - freqs->main, - LZX_MAINTREE_NUM_SYMBOLS); - memcpy(updated_main_lens, codes->lens.main, LZX_MAINTREE_NUM_SYMBOLS); - lzx_update_mainsym_match_costs(LZX_BLOCKTYPE_ALIGNED, updated_main_lens); - aligned_cost += lzx_huffman_code_output_cost(updated_main_lens, - freqs->main, - LZX_MAINTREE_NUM_SYMBOLS); - - common_cost += lzx_huffman_code_output_cost(codes->lens.len, - freqs->len, - LZX_LENTREE_NUM_SYMBOLS); - - aligned_cost += lzx_huffman_code_output_cost(codes->lens.aligned, - freqs->aligned, - LZX_ALIGNEDTREE_NUM_SYMBOLS); - - *aligned_cost_ret = aligned_cost + common_cost; - *verbatim_cost_ret = verbatim_cost + common_cost; + return LZX_BLOCKTYPE_VERBATIM; } -/* Prepare a (nonsplit) compressed block. */ -static unsigned -lzx_prepare_compressed_block(struct lzx_compressor *ctx, unsigned block_number, - struct lzx_codes *prev_codes) +/* Find a near-optimal sequence of matches/literals with which to output the + * specified LZX block, then set its type to that which has the minimum cost to + * output. */ +static void +lzx_optimize_block(struct lzx_compressor *ctx, struct lzx_block_spec *spec, + unsigned num_passes) { - struct lzx_block_spec *spec = &ctx->block_specs[block_number - 1]; - unsigned orig_cached_matches_pos = ctx->cached_matches_pos; - struct lzx_lru_queue orig_queue = ctx->queue; + const struct lzx_lru_queue orig_queue = ctx->queue; struct lzx_freqs freqs; - unsigned cost; - - /* Here's where the real work happens. The following loop runs one or - * more times, each time using a cost model based on the Huffman codes - * computed from the previous iteration (the first iteration uses a - * default model). Each iteration of the loop uses a heuristic - * algorithm to divide the block into near-optimal matches/literals from - * beginning to end. */ - LZX_ASSERT(ctx->params.slow.num_optim_passes >= 1); - spec->num_chosen_matches = 0; - for (unsigned pass = 0; pass < ctx->params.slow.num_optim_passes; pass++) - { - LZX_DEBUG("Block %u: Match-choosing pass %u of %u", - block_number, pass + 1, - ctx->params.slow.num_optim_passes); - - /* Reset frequency tables. */ - memset(&freqs, 0, sizeof(freqs)); - - /* Reset match offset LRU queue. */ - ctx->queue = orig_queue; - - /* Reset match-finding position. */ - ctx->cached_matches_pos = orig_cached_matches_pos; - ctx->match_window_pos = spec->window_pos; - ctx->match_window_end = spec->window_pos + spec->block_size; - /* Set cost model. */ - lzx_set_costs(ctx, &spec->codes.lens); + unsigned orig_window_pos = spec->window_pos; + unsigned orig_cached_pos = ctx->cached_matches_pos; - unsigned window_pos = spec->window_pos; - unsigned end = window_pos + spec->block_size; - - while (window_pos < end) { - struct raw_match match; - struct lzx_match lzx_match; + LZX_ASSERT(ctx->match_window_pos == spec->window_pos); - match = lzx_lz_get_near_optimal_match(ctx); + ctx->match_window_end = spec->window_pos + spec->block_size; + spec->chosen_matches_start_pos = spec->window_pos; - if (match.len >= LZX_MIN_MATCH) { + LZX_ASSERT(num_passes >= 1); - /* Best to output a match here. */ + /* The first optimal parsing pass is done using the cost model already + * set in ctx->costs. Each later pass is done using a cost model + * computed from the previous pass. */ + for (unsigned pass = 0; pass < num_passes; pass++) { - LZX_ASSERT(match.len <= LZX_MAX_MATCH); - LZX_ASSERT(!memcmp(&ctx->window[window_pos], - &ctx->window[window_pos - match.offset], - match.len)); + ctx->match_window_pos = orig_window_pos; + ctx->cached_matches_pos = orig_cached_pos; + ctx->queue = orig_queue; + spec->num_chosen_matches = 0; + memset(&freqs, 0, sizeof(freqs)); - /* Tally symbol frequencies. */ - lzx_match.data = lzx_record_match(match.offset, - match.len, - &freqs, - &ctx->queue); + for (unsigned i = spec->window_pos; i < spec->window_pos + spec->block_size; ) { + struct raw_match raw_match; + struct lzx_match lzx_match; - window_pos += match.len; + raw_match = lzx_lz_get_near_optimal_match(ctx); + if (raw_match.len >= LZX_MIN_MATCH_LEN) { + lzx_match.data = lzx_tally_match(raw_match.len, raw_match.offset, + &freqs, &ctx->queue); + i += raw_match.len; } else { - /* Best to output a literal here. */ - - /* Tally symbol frequencies. */ - lzx_match.data = lzx_record_literal(ctx->window[window_pos], - &freqs); - - window_pos += 1; - } - - /* If it's the last pass, save the match/literal in - * intermediate form. */ - if (pass == ctx->params.slow.num_optim_passes - 1) { - ctx->chosen_matches[spec->chosen_matches_start_pos + - spec->num_chosen_matches] = lzx_match; - - spec->num_chosen_matches++; + lzx_match.data = lzx_tally_literal(ctx->window[i], &freqs); + i += 1; } + ctx->chosen_matches[spec->chosen_matches_start_pos + + spec->num_chosen_matches++] = lzx_match; } - LZX_ASSERT(window_pos == end); - /* Build Huffman codes using the new frequencies. */ - lzx_make_huffman_codes(&freqs, &spec->codes); - - /* The first time we get here is when the full input has been - * processed, so the match-finding is done. */ - ctx->matches_already_found = true; + lzx_make_huffman_codes(&freqs, &spec->codes, + ctx->num_main_syms); + if (pass < num_passes - 1) + lzx_set_costs(ctx, &spec->codes.lens); + ctx->matches_cached = true; } + spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes); + ctx->matches_cached = false; +} - LZX_DEBUG("Block %u: saved %u matches/literals @ %u", - block_number, spec->num_chosen_matches, - spec->chosen_matches_start_pos); - - unsigned aligned_cost; - unsigned verbatim_cost; - - lzx_compute_compressed_block_costs(spec->block_size, - &spec->codes, - prev_codes, - &freqs, - &aligned_cost, - &verbatim_cost); - - /* Choose whether to make the block aligned offset or verbatim. */ - if (aligned_cost < verbatim_cost) { - spec->block_type = LZX_BLOCKTYPE_ALIGNED; - cost = aligned_cost; - LZX_DEBUG("Using aligned block (cost %u vs %u for verbatim)", - aligned_cost, verbatim_cost); - } else { - spec->block_type = LZX_BLOCKTYPE_VERBATIM; - cost = verbatim_cost; - LZX_DEBUG("Using verbatim block (cost %u vs %u for aligned)", - verbatim_cost, aligned_cost); - } +static void +lzx_optimize_blocks(struct lzx_compressor *ctx) +{ + lzx_lru_queue_init(&ctx->queue); + ctx->optimum_cur_idx = 0; + ctx->optimum_end_idx = 0; - LZX_DEBUG("Block %u is %u => %u bytes unsplit.", - block_number, spec->block_size, cost / 8); + const unsigned num_passes = ctx->params.alg_params.slow.num_optim_passes; - return cost; + for (unsigned i = 0; i < ctx->num_blocks; i++) + lzx_optimize_block(ctx, &ctx->block_specs[i], num_passes); } -/* - * lzx_prepare_block_recursive() - - * - * Given a (possibly nonproper) sub-sequence of the preprocessed input, compute - * the LZX block(s) that it should be output as. - * - * This function initially considers the case where the given sub-sequence of - * the preprocessed input be output as a single block. This block is calculated - * and its cost (number of bits required to output it) is computed. - * - * Then, if @max_split_level is greater than zero, a split into two evenly sized - * subblocks is considered. The block is recursively split in this way, - * potentially up to the depth specified by @max_split_level. The cost of the - * split block is compared to the cost of the single block, and the lower cost - * solution is used. - * - * For each compressed output block computed, the sequence of matches/literals - * and the corresponding Huffman codes for the block are produced and saved. - * - * The return value is the approximate number of bits the block (or all - * subblocks, in the case that the split block had lower cast), will take up - * when written to the compressed output. - */ -static unsigned -lzx_prepare_block_recursive(struct lzx_compressor * ctx, - unsigned block_number, - unsigned max_split_level, - struct lzx_codes **prev_codes_p) +/* Initialize the suffix array match-finder for the specified input. */ +static void +lzx_lz_init_matchfinder(const u8 T[const restrict], + const input_idx_t n, + input_idx_t SA[const restrict], + input_idx_t ISA[const restrict], + input_idx_t LCP[const restrict], + struct salink link[const restrict], + const unsigned max_match_len) { - struct lzx_block_spec *spec = &ctx->block_specs[block_number - 1]; - unsigned cost; - unsigned orig_cached_matches_pos; - struct lzx_lru_queue orig_queue, nonsplit_queue; - struct lzx_codes *prev_codes = *prev_codes_p; + /* Compute SA (Suffix Array). */ - LZX_DEBUG("Preparing block %u...", block_number); + { + /* ISA and link are used as temporary space. */ + BUILD_BUG_ON(LZX_MIN_WINDOW_SIZE * sizeof(ISA[0]) < 256 * sizeof(saidx_t)); + BUILD_BUG_ON(LZX_MIN_WINDOW_SIZE * 2 * sizeof(link[0]) < 256 * 256 * sizeof(saidx_t)); - /* Save positions of chosen and cached matches, and the match offset LRU - * queue, so that they can be restored if splitting is attempted. */ - orig_cached_matches_pos = ctx->cached_matches_pos; - orig_queue = ctx->queue; + if (sizeof(input_idx_t) == sizeof(saidx_t)) { + divsufsort(T, SA, n, (saidx_t*)ISA, (saidx_t*)link); + } else { + saidx_t sa[n]; + divsufsort(T, sa, n, (saidx_t*)ISA, (saidx_t*)link); + for (input_idx_t i = 0; i < n; i++) + SA[i] = sa[i]; + } + } - /* Consider outputting the input subsequence as a single block. */ - spec->is_split = 0; - cost = lzx_prepare_compressed_block(ctx, block_number, prev_codes); - nonsplit_queue = ctx->queue; +#ifdef ENABLE_LZX_DEBUG - *prev_codes_p = &spec->codes; + LZX_ASSERT(n > 0); - /* If the maximum split level is at least one, consider splitting the - * block in two. */ - if (max_split_level--) { + /* Verify suffix array. */ + { + bool found[n]; + ZERO_ARRAY(found); + for (input_idx_t r = 0; r < n; r++) { + input_idx_t i = SA[r]; + LZX_ASSERT(i < n); + LZX_ASSERT(!found[i]); + found[i] = true; + } + } - LZX_DEBUG("Calculating split of block %u...", block_number); + for (input_idx_t r = 0; r < n - 1; r++) { - struct lzx_block_spec *spec1, *spec2; - unsigned split_cost; + input_idx_t i1 = SA[r]; + input_idx_t i2 = SA[r + 1]; - ctx->cached_matches_pos = orig_cached_matches_pos; - ctx->queue = orig_queue; + input_idx_t n1 = n - i1; + input_idx_t n2 = n - i2; - /* Prepare and get the cost of the first sub-block. */ - spec1 = &ctx->block_specs[block_number * 2 - 1]; - spec1->codes.lens = spec->codes.lens; - spec1->window_pos = spec->window_pos; - spec1->block_size = spec->block_size / 2; - spec1->chosen_matches_start_pos = spec->chosen_matches_start_pos + - LZX_MAX_WINDOW_SIZE; - split_cost = lzx_prepare_block_recursive(ctx, - block_number * 2, - max_split_level, - &prev_codes); - - /* Prepare and get the cost of the second sub-block. */ - spec2 = spec1 + 1; - spec2->codes.lens = spec->codes.lens; - spec2->window_pos = spec->window_pos + spec1->block_size; - spec2->block_size = spec->block_size - spec1->block_size; - spec2->chosen_matches_start_pos = spec1->chosen_matches_start_pos + - spec1->block_size; - split_cost += lzx_prepare_block_recursive(ctx, - block_number * 2 + 1, - max_split_level, - &prev_codes); - - /* Compare the cost of the whole block with that of the split - * block. Choose the lower cost solution. */ - if (split_cost < cost) { - LZX_DEBUG("Splitting block %u is worth it " - "(%u => %u bytes).", - block_number, cost / 8, split_cost / 8); - spec->is_split = 1; - cost = split_cost; - *prev_codes_p = prev_codes; - } else { - LZX_DEBUG("Splitting block %u is NOT worth it " - "(%u => %u bytes).", - block_number, cost / 8, split_cost / 8); - ctx->queue = nonsplit_queue; - } + LZX_ASSERT(memcmp(&T[i1], &T[i2], min(n1, n2)) <= 0); } + LZX_DEBUG("Verified SA (len %u)", n); +#endif /* ENABLE_LZX_DEBUG */ - return cost; -} + /* Compute ISA (Inverse Suffix Array) */ + for (input_idx_t r = 0; r < n; r++) + ISA[SA[r]] = r; -/* Empirical averages */ -static const u8 lzx_default_mainsym_costs[LZX_MAINTREE_NUM_SYMBOLS] = { - 7, 9, 9, 10, 9, 10, 10, 10, 9, 10, 9, 10, 10, 9, 10, 10, 9, 10, 10, 11, - 10, 10, 10, 11, 10, 11, 11, 11, 10, 11, 11, 11, 8, 11, 9, 10, 9, 10, 11, - 11, 9, 9, 11, 10, 10, 9, 9, 9, 8, 8, 8, 8, 8, 9, 9, 9, 8, 8, 9, 9, 9, 9, - 10, 10, 10, 8, 9, 8, 8, 8, 8, 9, 9, 9, 10, 10, 8, 8, 9, 9, 8, 10, 9, 8, - 8, 9, 8, 9, 9, 10, 10, 10, 9, 10, 11, 9, 10, 8, 9, 8, 8, 8, 8, 9, 8, 8, - 9, 9, 8, 8, 8, 8, 8, 10, 8, 8, 7, 8, 9, 9, 9, 9, 10, 11, 10, 10, 11, 11, - 10, 11, 11, 10, 10, 11, 11, 11, 10, 10, 11, 10, 11, 10, 11, 11, 10, 11, - 11, 12, 11, 11, 11, 12, 11, 11, 11, 11, 11, 11, 11, 12, 10, 11, 11, 11, - 11, 11, 11, 12, 11, 11, 11, 11, 11, 12, 11, 11, 10, 11, 11, 11, 11, 11, - 11, 11, 10, 11, 11, 11, 11, 11, 11, 11, 10, 11, 11, 11, 11, 11, 11, 11, - 10, 11, 11, 11, 11, 11, 11, 11, 10, 11, 11, 11, 11, 12, 11, 11, 10, 11, - 11, 11, 11, 12, 11, 11, 10, 11, 11, 11, 10, 12, 11, 11, 10, 10, 11, 10, - 10, 11, 11, 11, 10, 11, 11, 11, 10, 11, 11, 11, 10, 11, 11, 11, 10, 11, - 10, 9, 8, 7, 10, 10, 11, 10, 11, 7, 9, 9, 11, 11, 11, 12, 11, 9, 10, 10, - 12, 12, 13, 13, 12, 11, 10, 12, 12, 14, 14, 14, 13, 12, 9, 12, 13, 14, - 14, 14, 14, 14, 9, 10, 13, 14, 14, 14, 14, 14, 9, 9, 11, 11, 13, 13, 13, - 14, 9, 9, 11, 12, 12, 13, 13, 13, 8, 8, 11, 11, 12, 12, 12, 11, 9, 9, - 10, 11, 12, 12, 12, 11, 8, 9, 10, 10, 11, 12, 11, 10, 9, 9, 10, 11, 11, - 12, 11, 10, 8, 9, 10, 10, 11, 11, 11, 9, 9, 9, 10, 11, 11, 11, 11, 9, 8, - 8, 10, 10, 11, 11, 11, 9, 9, 9, 10, 10, 11, 11, 11, 9, 9, 8, 9, 10, 11, - 11, 11, 9, 10, 9, 10, 11, 11, 11, 11, 9, 14, 9, 9, 10, 10, 11, 10, 9, - 14, 9, 10, 11, 11, 11, 11, 9, 14, 9, 10, 10, 11, 11, 11, 9, 14, 10, 10, - 11, 11, 12, 11, 10, 14, 10, 10, 10, 11, 11, 11, 10, 14, 11, 11, 11, 11, - 12, 12, 10, 14, 10, 11, 11, 11, 12, 11, 10, 14, 11, 11, 11, 12, 12, 12, - 11, 15, 11, 11, 11, 12, 12, 12, 11, 14, 12, 12, 12, 12, 13, 12, 11, 15, - 12, 12, 12, 13, 13, 13, 12, 15, 14, 13, 14, 14, 14, 14, 13, -}; + /* Compute LCP (longest common prefix) array. + * + * Algorithm adapted from Kasai et al. 2001: "Linear-Time + * Longest-Common-Prefix Computation in Suffix Arrays and Its + * Applications". */ + { + input_idx_t h = 0; + for (input_idx_t i = 0; i < n; i++) { + input_idx_t r = ISA[i]; + if (r > 0) { + input_idx_t j = SA[r - 1]; + + input_idx_t lim = min(n - i, n - j); + + while (h < lim && T[i + h] == T[j + h]) + h++; + LCP[r] = h; + if (h > 0) + h--; + } + } + } -/* Empirical averages */ -static const u8 lzx_default_lensym_costs[LZX_LENTREE_NUM_SYMBOLS] = { - 5, 5, 5, 5, 5, 6, 5, 5, 6, 7, 7, 7, 8, 8, 7, 8, 9, 9, 9, 9, 10, 9, 9, - 10, 9, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 12, 12, 12, 11, 12, 12, - 12, 12, 12, 12, 13, 12, 12, 12, 13, 12, 13, 13, 12, 12, 13, 12, 13, 13, - 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 14, 13, 14, 13, 14, 13, - 14, 13, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 13, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, - 14, 14, 14, 14, 14, 14, 14, 14, 14, 10, -}; +#ifdef ENABLE_LZX_DEBUG + /* Verify LCP array. */ + for (input_idx_t r = 0; r < n - 1; r++) { + LZX_ASSERT(ISA[SA[r]] == r); + LZX_ASSERT(ISA[SA[r + 1]] == r + 1); -/* - * Set default symbol costs. - */ -static void -lzx_set_default_costs(struct lzx_lens * lens) -{ - unsigned i; + input_idx_t i1 = SA[r]; + input_idx_t i2 = SA[r + 1]; + input_idx_t lcp = LCP[r + 1]; -#if LZX_PARAM_USE_EMPIRICAL_DEFAULT_COSTS - memcpy(&lens->main, lzx_default_mainsym_costs, LZX_MAINTREE_NUM_SYMBOLS); - memcpy(&lens->len, lzx_default_lensym_costs, LZX_LENTREE_NUM_SYMBOLS); + input_idx_t n1 = n - i1; + input_idx_t n2 = n - i2; -#else - /* Literal symbols */ - for (i = 0; i < LZX_NUM_CHARS; i++) - lens->main[i] = 8; + LZX_ASSERT(lcp <= min(n1, n2)); - /* Match header symbols */ - for (; i < LZX_MAINTREE_NUM_SYMBOLS; i++) - lens->main[i] = 10; + LZX_ASSERT(memcmp(&T[i1], &T[i2], lcp) == 0); + if (lcp < min(n1, n2)) + LZX_ASSERT(T[i1 + lcp] != T[i2 + lcp]); + } +#endif /* ENABLE_LZX_DEBUG */ - /* Length symbols */ - for (i = 0; i < LZX_LENTREE_NUM_SYMBOLS; i++) - lens->len[i] = 8; -#endif + /* Compute salink.next and salink.lcpnext. + * + * Algorithm adapted from Crochemore et al. 2009: + * "LPF computation revisited". + * + * Note: we cap lcpnext to the maximum match length so that the + * match-finder need not worry about it later. */ + link[n - 1].next = (input_idx_t)~0U; + link[n - 1].prev = (input_idx_t)~0U; + link[n - 1].lcpnext = 0; + link[n - 1].lcpprev = 0; + for (input_idx_t r = n - 2; r != (input_idx_t)~0U; r--) { + input_idx_t t = r + 1; + input_idx_t l = LCP[t]; + while (t != (input_idx_t)~0 && SA[t] > SA[r]) { + l = min(l, link[t].lcpnext); + t = link[t].next; + } + link[r].next = t; + link[r].lcpnext = min(l, max_match_len); + LZX_ASSERT(t == (input_idx_t)~0U || l <= n - SA[t]); + LZX_ASSERT(l <= n - SA[r]); + LZX_ASSERT(memcmp(&T[SA[r]], &T[SA[t]], l) == 0); + } - /* Aligned offset symbols */ - for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++) - lens->aligned[i] = 3; + /* Compute salink.prev and salink.lcpprev. + * + * Algorithm adapted from Crochemore et al. 2009: + * "LPF computation revisited". + * + * Note: we cap lcpprev to the maximum match length so that the + * match-finder need not worry about it later. */ + link[0].prev = (input_idx_t)~0; + link[0].next = (input_idx_t)~0; + link[0].lcpprev = 0; + link[0].lcpnext = 0; + for (input_idx_t r = 1; r < n; r++) { + input_idx_t t = r - 1; + input_idx_t l = LCP[r]; + while (t != (input_idx_t)~0 && SA[t] > SA[r]) { + l = min(l, link[t].lcpprev); + t = link[t].prev; + } + link[r].prev = t; + link[r].lcpprev = min(l, max_match_len); + LZX_ASSERT(t == (input_idx_t)~0 || l <= n - SA[t]); + LZX_ASSERT(l <= n - SA[r]); + LZX_ASSERT(memcmp(&T[SA[r]], &T[SA[t]], l) == 0); + } } -/* - * lzx_prepare_blocks() - - * - * Calculate the blocks to split the preprocessed data into. - * - * Input --- the preprocessed data: - * - * ctx->window[] - * ctx->window_size - * - * Working space: - * Match finding: - * ctx->hash_tab - * ctx->child_tab - * ctx->cached_matches - * ctx->cached_matches_pos - * ctx->matches_already_found - * - * Block cost modeling: - * ctx->costs - * ctx->block_specs (also an output) - * - * Match choosing: - * ctx->optimum - * ctx->optimum_cur_idx - * ctx->optimum_end_idx - * ctx->chosen_matches (also an output) - * - * Output --- the block specifications and the corresponding match/literal data: - * - * ctx->block_specs[] - * ctx->chosen_matches[] - * - * The return value is the approximate number of bits the compressed data will - * take up. - */ -static unsigned +/* Prepare the input window into one or more LZX blocks ready to be output. */ +static void lzx_prepare_blocks(struct lzx_compressor * ctx) { - /* This function merely does some initializations, then passes control - * to lzx_prepare_block_recursive(). */ - - /* 1. Initialize match-finding variables. */ - - /* Zero all entries in the hash table, indicating that no length-3 - * character sequences have been discovered in the input yet. */ - memset(ctx->hash_tab, 0, LZX_LZ_HASH_SIZE * 2 * sizeof(ctx->hash_tab[0])); - if (ctx->params.slow.use_len2_matches) - memset(ctx->digram_tab, 0, 256 * 256 * sizeof(ctx->digram_tab[0])); - /* Note: ctx->child_tab need not be initialized. */ - - /* No matches have been found and cached yet. */ + /* Initialize the match-finder. */ + lzx_lz_init_matchfinder(ctx->window, ctx->window_size, + ctx->SA, ctx->ISA, ctx->LCP, ctx->salink, + LZX_MAX_MATCH_LEN); ctx->cached_matches_pos = 0; - ctx->matches_already_found = false; - - /* 2. Initialize match-choosing variables. */ - ctx->optimum_cur_idx = 0; - ctx->optimum_end_idx = 0; - /* Note: ctx->optimum need not be initialized. */ - ctx->block_specs[0].chosen_matches_start_pos = 0; - - /* 3. Set block 1 (index 0) to represent the entire input data. */ - ctx->block_specs[0].block_size = ctx->window_size; - ctx->block_specs[0].window_pos = 0; + ctx->matches_cached = false; + ctx->match_window_pos = 0; - /* 4. Set up a default Huffman symbol cost model for block 1 (index 0). - * The model will be refined later. */ - lzx_set_default_costs(&ctx->block_specs[0].codes.lens); + /* Set up a default cost model. */ + lzx_set_default_costs(&ctx->costs, ctx->num_main_syms); - /* 5. Initialize the match offset LRU queue. */ - ctx->queue = (struct lzx_lru_queue){1, 1, 1}; + ctx->num_blocks = DIV_ROUND_UP(ctx->window_size, LZX_DIV_BLOCK_SIZE); + for (unsigned i = 0; i < ctx->num_blocks; i++) { + unsigned pos = LZX_DIV_BLOCK_SIZE * i; + ctx->block_specs[i].window_pos = pos; + ctx->block_specs[i].block_size = min(ctx->window_size - pos, LZX_DIV_BLOCK_SIZE); + } - /* 6. Pass control to recursive procedure. */ - struct lzx_codes * prev_codes = &ctx->zero_codes; - return lzx_prepare_block_recursive(ctx, 1, - ctx->params.slow.num_split_passes, - &prev_codes); + /* Determine sequence of matches/literals to output for each block. */ + lzx_optimize_blocks(ctx); } /* - * This is the fast version of lzx_prepare_blocks(), which "quickly" prepares a - * single compressed block containing the entire input. See the description of - * the "Fast algorithm" at the beginning of this file for more information. + * This is the fast version of lzx_prepare_blocks(). This version "quickly" + * prepares a single compressed block containing the entire input. See the + * description of the "Fast algorithm" at the beginning of this file for more + * information. * * Input --- the preprocessed data: * * ctx->window[] * ctx->window_size * - * Working space: - * ctx->queue - * - * Output --- the block specifications and the corresponding match/literal data: + * Output --- the block specification and the corresponding match/literal data: * * ctx->block_specs[] + * ctx->num_blocks * ctx->chosen_matches[] */ static void lzx_prepare_block_fast(struct lzx_compressor * ctx) { - unsigned num_matches; - struct lzx_freqs freqs; + struct lzx_record_ctx record_ctx; struct lzx_block_spec *spec; - /* Parameters to hash chain LZ match finder */ + /* Parameters to hash chain LZ match finder + * (lazy with 1 match lookahead) */ static const struct lz_params lzx_lz_params = { - /* LZX_MIN_MATCH == 2, but 2-character matches are rarely - * useful; the minimum match for compression is set to 3 - * instead. */ + /* Although LZX_MIN_MATCH_LEN == 2, length 2 matches typically + * aren't worth choosing when using greedy or lazy parsing. */ .min_match = 3, - .max_match = LZX_MAX_MATCH, - .good_match = LZX_MAX_MATCH, - .nice_match = LZX_MAX_MATCH, - .max_chain_len = LZX_MAX_MATCH, - .max_lazy_match = LZX_MAX_MATCH, + .max_match = LZX_MAX_MATCH_LEN, + .max_offset = LZX_MAX_WINDOW_SIZE, + .good_match = LZX_MAX_MATCH_LEN, + .nice_match = LZX_MAX_MATCH_LEN, + .max_chain_len = LZX_MAX_MATCH_LEN, + .max_lazy_match = LZX_MAX_MATCH_LEN, .too_far = 4096, }; /* Initialize symbol frequencies and match offset LRU queue. */ - memset(&freqs, 0, sizeof(struct lzx_freqs)); - ctx->queue = (struct lzx_lru_queue){ 1, 1, 1 }; + memset(&record_ctx.freqs, 0, sizeof(struct lzx_freqs)); + lzx_lru_queue_init(&record_ctx.queue); + record_ctx.matches = ctx->chosen_matches; /* Determine series of matches/literals to output. */ - num_matches = lz_analyze_block(ctx->window, - ctx->window_size, - (u32*)ctx->chosen_matches, - lzx_record_match, - lzx_record_literal, - &freqs, - &ctx->queue, - &freqs, - &lzx_lz_params); - + lz_analyze_block(ctx->window, + ctx->window_size, + lzx_record_match, + lzx_record_literal, + &record_ctx, + &lzx_lz_params, + ctx->prev_tab); /* Set up block specification. */ spec = &ctx->block_specs[0]; - spec->is_split = 0; spec->block_type = LZX_BLOCKTYPE_ALIGNED; spec->window_pos = 0; spec->block_size = ctx->window_size; - spec->num_chosen_matches = num_matches; + spec->num_chosen_matches = (record_ctx.matches - ctx->chosen_matches); spec->chosen_matches_start_pos = 0; - lzx_make_huffman_codes(&freqs, &spec->codes); + lzx_make_huffman_codes(&record_ctx.freqs, &spec->codes, + ctx->num_main_syms); + ctx->num_blocks = 1; } static void @@ -2496,20 +2220,19 @@ wimlib_lzx_compress2(const void * const restrict uncompressed_data, { struct lzx_compressor *ctx = (struct lzx_compressor*)lzx_ctx; struct output_bitstream ostream; - unsigned compressed_len; + input_idx_t compressed_len; if (uncompressed_len < 100) { LZX_DEBUG("Too small to bother compressing."); return 0; } - if (uncompressed_len > 32768) { - LZX_DEBUG("Only up to 32768 bytes of uncompressed data are supported."); + if (uncompressed_len > ctx->max_window_size) { + LZX_DEBUG("Can't compress %u bytes using window of %u bytes!", + uncompressed_len, ctx->max_window_size); return 0; } - wimlib_assert(lzx_ctx != NULL); - LZX_DEBUG("Attempting to compress %u bytes...", uncompressed_len); /* The input data must be preprocessed. To avoid changing the original @@ -2543,52 +2266,45 @@ wimlib_lzx_compress2(const void * const restrict uncompressed_data, lzx_write_all_blocks(ctx, &ostream); LZX_DEBUG("Flushing bitstream..."); - if (flush_output_bitstream(&ostream)) { - /* If the bitstream cannot be flushed, then the output space was - * exhausted. */ + compressed_len = flush_output_bitstream(&ostream); + if (compressed_len == ~(input_idx_t)0) { LZX_DEBUG("Data did not compress to less than original length!"); return 0; } - /* Compute the length of the compressed data. */ - compressed_len = ostream.bit_output - (u8*)compressed_data; - LZX_DEBUG("Done: compressed %u => %u bytes.", uncompressed_len, compressed_len); -#if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION) - /* Verify that we really get the same thing back when decompressing. */ + /* Verify that we really get the same thing back when decompressing. + * Although this could be disabled by default in all cases, it only + * takes around 2-3% of the running time of the slow algorithm to do the + * verification. */ + if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_SLOW + #if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION) + || 1 + #endif + ) { - u8 buf[uncompressed_len]; - int ret; - unsigned i; + /* The decompression buffer can be any temporary space that's no + * longer needed. */ + u8 *buf = (u8*)(ctx->SA ? ctx->SA : ctx->prev_tab); - ret = wimlib_lzx_decompress(compressed_data, compressed_len, - buf, uncompressed_len); - if (ret) { + if (wimlib_lzx_decompress2(compressed_data, compressed_len, + buf, uncompressed_len, ctx->max_window_size)) + { ERROR("Failed to decompress data we " "compressed using LZX algorithm"); wimlib_assert(0); return 0; } - bool bad = false; - const u8 * udata = uncompressed_data; - for (i = 0; i < uncompressed_len; i++) { - if (buf[i] != udata[i]) { - bad = true; - ERROR("Data we compressed using LZX algorithm " - "didn't decompress to original " - "(difference at idx %u: c %#02x, u %#02x)", - i, buf[i], udata[i]); - } - } - if (bad) { + if (memcmp(uncompressed_data, buf, uncompressed_len)) { + ERROR("Data we compressed using LZX algorithm " + "didn't decompress to original"); wimlib_assert(0); return 0; } } -#endif return compressed_len; } @@ -2599,48 +2315,121 @@ lzx_params_compatible(const struct wimlib_lzx_params *oldparams, return 0 == memcmp(oldparams, newparams, sizeof(struct wimlib_lzx_params)); } +static struct wimlib_lzx_params lzx_user_default_params; +static struct wimlib_lzx_params *lzx_user_default_params_ptr; + +static bool +lzx_params_valid(const struct wimlib_lzx_params *params) +{ + /* Validate parameters. */ + if (params->size_of_this != sizeof(struct wimlib_lzx_params)) { + LZX_DEBUG("Invalid parameter structure size!"); + return false; + } + + if (params->algorithm != WIMLIB_LZX_ALGORITHM_SLOW && + params->algorithm != WIMLIB_LZX_ALGORITHM_FAST) + { + LZX_DEBUG("Invalid algorithm."); + return false; + } + + if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { + if (params->alg_params.slow.num_optim_passes < 1) + { + LZX_DEBUG("Invalid number of optimization passes!"); + return false; + } + + if (params->alg_params.slow.main_nostat_cost < 1 || + params->alg_params.slow.main_nostat_cost > 16) + { + LZX_DEBUG("Invalid main_nostat_cost!"); + return false; + } + + if (params->alg_params.slow.len_nostat_cost < 1 || + params->alg_params.slow.len_nostat_cost > 16) + { + LZX_DEBUG("Invalid len_nostat_cost!"); + return false; + } + + if (params->alg_params.slow.aligned_nostat_cost < 1 || + params->alg_params.slow.aligned_nostat_cost > 8) + { + LZX_DEBUG("Invalid aligned_nostat_cost!"); + return false; + } + } + return true; +} + +/* API function documented in wimlib.h */ +WIMLIBAPI int +wimlib_lzx_set_default_params(const struct wimlib_lzx_params * params) +{ + if (params) { + if (!lzx_params_valid(params)) + return WIMLIB_ERR_INVALID_PARAM; + lzx_user_default_params = *params; + lzx_user_default_params_ptr = &lzx_user_default_params; + } else { + lzx_user_default_params_ptr = NULL; + } + return 0; +} + /* API function documented in wimlib.h */ WIMLIBAPI int -wimlib_lzx_alloc_context(const struct wimlib_lzx_params *params, +wimlib_lzx_alloc_context(u32 window_size, + const struct wimlib_lzx_params *params, struct wimlib_lzx_context **ctx_pp) { LZX_DEBUG("Allocating LZX context..."); + if (!lzx_window_size_valid(window_size)) + return WIMLIB_ERR_INVALID_PARAM; + struct lzx_compressor *ctx; static const struct wimlib_lzx_params fast_default = { .size_of_this = sizeof(struct wimlib_lzx_params), .algorithm = WIMLIB_LZX_ALGORITHM_FAST, .use_defaults = 0, - .fast = { + .alg_params = { + .fast = { + }, }, }; static const struct wimlib_lzx_params slow_default = { .size_of_this = sizeof(struct wimlib_lzx_params), .algorithm = WIMLIB_LZX_ALGORITHM_SLOW, .use_defaults = 0, - .slow = { - .use_len2_matches = 1, - .num_fast_bytes = 32, - .num_optim_passes = 3, - .num_split_passes = 3, - .main_nostat_cost = 15, - .len_nostat_cost = 15, - .aligned_nostat_cost = 7, + .alg_params = { + .slow = { + .use_len2_matches = 1, + .num_fast_bytes = 32, + .num_optim_passes = 2, + .max_search_depth = 50, + .max_matches_per_pos = 3, + .main_nostat_cost = 15, + .len_nostat_cost = 15, + .aligned_nostat_cost = 7, + }, }, }; - if (params == NULL) { + if (params) { + if (!lzx_params_valid(params)) + return WIMLIB_ERR_INVALID_PARAM; + } else { LZX_DEBUG("Using default algorithm and parameters."); - params = &slow_default; - } - - if (params->algorithm != WIMLIB_LZX_ALGORITHM_SLOW && - params->algorithm != WIMLIB_LZX_ALGORITHM_FAST) - { - LZX_DEBUG("Invalid algorithm."); - return WIMLIB_ERR_INVALID_PARAM; + if (lzx_user_default_params_ptr) + params = lzx_user_default_params_ptr; + else + params = &slow_default; } if (params->use_defaults) { @@ -2650,123 +2439,76 @@ wimlib_lzx_alloc_context(const struct wimlib_lzx_params *params, params = &fast_default; } - if (params->size_of_this != sizeof(struct wimlib_lzx_params)) { - LZX_DEBUG("Invalid parameter structure size!"); - return WIMLIB_ERR_INVALID_PARAM; - } - - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - if (params->slow.num_fast_bytes < 3 || - params->slow.num_fast_bytes > 257) - { - LZX_DEBUG("Invalid number of fast bytes!"); - return WIMLIB_ERR_INVALID_PARAM; - } - - if (params->slow.num_optim_passes < 1) - { - LZX_DEBUG("Invalid number of optimization passes!"); - return WIMLIB_ERR_INVALID_PARAM; - } - - if (params->slow.main_nostat_cost < 1 || - params->slow.main_nostat_cost > 16) - { - LZX_DEBUG("Invalid main_nostat_cost!"); - return WIMLIB_ERR_INVALID_PARAM; - } - - if (params->slow.len_nostat_cost < 1 || - params->slow.len_nostat_cost > 16) - { - LZX_DEBUG("Invalid len_nostat_cost!"); - return WIMLIB_ERR_INVALID_PARAM; - } + if (ctx_pp) { + ctx = *(struct lzx_compressor**)ctx_pp; - if (params->slow.aligned_nostat_cost < 1 || - params->slow.aligned_nostat_cost > 8) - { - LZX_DEBUG("Invalid aligned_nostat_cost!"); - return WIMLIB_ERR_INVALID_PARAM; - } - } - - if (ctx_pp == NULL) { + if (ctx && + lzx_params_compatible(&ctx->params, params) && + ctx->max_window_size == window_size) + return 0; + } else { LZX_DEBUG("Check parameters only."); return 0; } - ctx = *(struct lzx_compressor**)ctx_pp; - - if (ctx && lzx_params_compatible(&ctx->params, params)) - return 0; - LZX_DEBUG("Allocating memory."); - ctx = MALLOC(sizeof(struct lzx_compressor)); + ctx = CALLOC(1, sizeof(struct lzx_compressor)); if (ctx == NULL) goto err; - size_t block_specs_length; + ctx->num_main_syms = lzx_get_num_main_syms(window_size); + ctx->max_window_size = window_size; + ctx->window = MALLOC(window_size + 12); + if (ctx->window == NULL) + goto err; - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) - block_specs_length = ((1 << (params->slow.num_split_passes + 1)) - 1); - else - block_specs_length = 1; + if (params->algorithm == WIMLIB_LZX_ALGORITHM_FAST) { + ctx->prev_tab = MALLOC(window_size * sizeof(ctx->prev_tab[0])); + if (ctx->prev_tab == NULL) + goto err; + } + + size_t block_specs_length = DIV_ROUND_UP(window_size, LZX_DIV_BLOCK_SIZE); ctx->block_specs = MALLOC(block_specs_length * sizeof(ctx->block_specs[0])); if (ctx->block_specs == NULL) - goto err_free_ctx; + goto err; if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - ctx->hash_tab = MALLOC((LZX_LZ_HASH_SIZE + 2 * LZX_MAX_WINDOW_SIZE) * - sizeof(ctx->hash_tab[0])); - if (ctx->hash_tab == NULL) - goto err_free_block_specs; - ctx->child_tab = ctx->hash_tab + LZX_LZ_HASH_SIZE; - } else { - ctx->hash_tab = NULL; - ctx->child_tab = NULL; + ctx->SA = MALLOC(3U * window_size * sizeof(ctx->SA[0])); + if (ctx->SA == NULL) + goto err; + ctx->ISA = ctx->SA + window_size; + ctx->LCP = ctx->ISA + window_size; + + ctx->salink = MALLOC(window_size * sizeof(ctx->salink[0])); + if (ctx->salink == NULL) + goto err; } - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW && - params->slow.use_len2_matches) - { - ctx->digram_tab = MALLOC(256 * 256 * sizeof(ctx->digram_tab[0])); - if (ctx->digram_tab == NULL) - goto err_free_hash_tab; - } else { - ctx->digram_tab = NULL; + if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { + ctx->optimum = MALLOC((LZX_OPTIM_ARRAY_SIZE + LZX_MAX_MATCH_LEN) * + sizeof(ctx->optimum[0])); + if (ctx->optimum == NULL) + goto err; } if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - ctx->cached_matches = MALLOC(10 * LZX_MAX_WINDOW_SIZE * + u32 cache_per_pos; + + cache_per_pos = params->alg_params.slow.max_matches_per_pos; + if (cache_per_pos > LZX_MAX_CACHE_PER_POS) + cache_per_pos = LZX_MAX_CACHE_PER_POS; + + ctx->cached_matches = MALLOC(window_size * (cache_per_pos + 1) * sizeof(ctx->cached_matches[0])); if (ctx->cached_matches == NULL) - goto err_free_digram_tab; - } else { - ctx->cached_matches = NULL; + goto err; } - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) { - ctx->optimum = MALLOC((LZX_PARAM_OPTIM_ARRAY_SIZE + LZX_MAX_MATCH) * - sizeof(ctx->optimum[0])); - if (ctx->optimum == NULL) - goto err_free_cached_matches; - } else { - ctx->optimum = NULL; - } - - size_t chosen_matches_length; - if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) - chosen_matches_length = LZX_MAX_WINDOW_SIZE * - (params->slow.num_split_passes + 1); - else - chosen_matches_length = LZX_MAX_WINDOW_SIZE; - - ctx->chosen_matches = MALLOC(chosen_matches_length * - sizeof(ctx->chosen_matches[0])); + ctx->chosen_matches = MALLOC(window_size * sizeof(ctx->chosen_matches[0])); if (ctx->chosen_matches == NULL) - goto err_free_optimum; + goto err; memcpy(&ctx->params, params, sizeof(struct wimlib_lzx_params)); memset(&ctx->zero_codes, 0, sizeof(ctx->zero_codes)); @@ -2777,19 +2519,8 @@ wimlib_lzx_alloc_context(const struct wimlib_lzx_params *params, *ctx_pp = (struct wimlib_lzx_context*)ctx; return 0; -err_free_optimum: - FREE(ctx->optimum); -err_free_cached_matches: - FREE(ctx->cached_matches); -err_free_digram_tab: - FREE(ctx->digram_tab); -err_free_hash_tab: - FREE(ctx->hash_tab); -err_free_block_specs: - FREE(ctx->block_specs); -err_free_ctx: - FREE(ctx); err: + wimlib_lzx_free_context((struct wimlib_lzx_context*)ctx); LZX_DEBUG("Ran out of memory."); return WIMLIB_ERR_NOMEM; } @@ -2802,11 +2533,13 @@ wimlib_lzx_free_context(struct wimlib_lzx_context *_ctx) if (ctx) { FREE(ctx->chosen_matches); - FREE(ctx->optimum); FREE(ctx->cached_matches); - FREE(ctx->digram_tab); - FREE(ctx->hash_tab); + FREE(ctx->optimum); + FREE(ctx->salink); + FREE(ctx->SA); FREE(ctx->block_specs); + FREE(ctx->prev_tab); + FREE(ctx->window); FREE(ctx); } } @@ -2818,10 +2551,10 @@ wimlib_lzx_compress(const void * const restrict uncompressed_data, void * const restrict compressed_data) { int ret; - struct wimlib_lzx_context *ctx; + struct wimlib_lzx_context *ctx = NULL; unsigned compressed_len; - ret = wimlib_lzx_alloc_context(NULL, &ctx); + ret = wimlib_lzx_alloc_context(32768, NULL, &ctx); if (ret) { wimlib_assert(ret != WIMLIB_ERR_INVALID_PARAM); WARNING("Couldn't allocate LZX compression context: %"TS"",