X-Git-Url: https://wimlib.net/git/?p=wimlib;a=blobdiff_plain;f=src%2Flzx_compress.c;h=ec90ba1640cc0a11f18b1b848e5dfee28e915356;hp=1506b8842ca567533e1fafea263edaaac89fe8ff;hb=8d91746c5772779e0e0593112c0143284452ee6e;hpb=dadccd777e56b473128d87e485fcf8565b8cb93b diff --git a/src/lzx_compress.c b/src/lzx_compress.c index 1506b884..ec90ba16 100644 --- a/src/lzx_compress.c +++ b/src/lzx_compress.c @@ -105,7 +105,7 @@ /* * LZX_BIT_COST is a scaling factor that represents the cost to output one bit. - * THis makes it possible to consider fractional bit costs. + * This makes it possible to consider fractional bit costs. * * Note: this is only useful as a statistical trick for when the true costs are * unknown. In reality, each token in LZX requires a whole number of bits to @@ -114,10 +114,9 @@ #define LZX_BIT_COST 16 /* - * Consideration of aligned offset costs is disabled for now, due to - * insufficient benefit gained from the time spent. + * Should the compressor take into account the costs of aligned offset symbols? */ -#define LZX_CONSIDER_ALIGNED_COSTS 0 +#define LZX_CONSIDER_ALIGNED_COSTS 1 /* * LZX_MAX_FAST_LEVEL is the maximum compression level at which we use the @@ -126,32 +125,46 @@ #define LZX_MAX_FAST_LEVEL 34 /* - * LZX_HASH2_ORDER is the log base 2 of the number of entries in the hash table - * for finding length 2 matches. This can be as high as 16 (in which case the - * hash function is trivial), but using a smaller hash table actually speeds up - * compression due to reduced cache pressure. + * BT_MATCHFINDER_HASH2_ORDER is the log base 2 of the number of entries in the + * hash table for finding length 2 matches. This could be as high as 16, but + * using a smaller hash table speeds up compression due to reduced cache + * pressure. */ -#define LZX_HASH2_ORDER 12 -#define LZX_HASH2_LENGTH (1UL << LZX_HASH2_ORDER) - -#include "wimlib/lzx_common.h" +#define BT_MATCHFINDER_HASH2_ORDER 12 /* - * The maximum allowed window order for the matchfinder. + * These are the compressor-side limits on the codeword lengths for each Huffman + * code. To make outputting bits slightly faster, some of these limits are + * lower than the limits defined by the LZX format. This does not significantly + * affect the compression ratio, at least for the block sizes we use. */ -#define MATCHFINDER_MAX_WINDOW_ORDER LZX_MAX_WINDOW_ORDER +#define MAIN_CODEWORD_LIMIT 12 /* 64-bit: can buffer 4 main symbols */ +#define LENGTH_CODEWORD_LIMIT 12 +#define ALIGNED_CODEWORD_LIMIT 7 +#define PRE_CODEWORD_LIMIT 7 -#include - -#include "wimlib/bt_matchfinder.h" #include "wimlib/compress_common.h" #include "wimlib/compressor_ops.h" #include "wimlib/error.h" -#include "wimlib/hc_matchfinder.h" #include "wimlib/lz_extend.h" +#include "wimlib/lzx_common.h" #include "wimlib/unaligned.h" #include "wimlib/util.h" +/* Matchfinders with 16-bit positions */ +#define mf_pos_t u16 +#define MF_SUFFIX _16 +#include "wimlib/bt_matchfinder.h" +#include "wimlib/hc_matchfinder.h" + +/* Matchfinders with 32-bit positions */ +#undef mf_pos_t +#undef MF_SUFFIX +#define mf_pos_t u32 +#define MF_SUFFIX _32 +#include "wimlib/bt_matchfinder.h" +#include "wimlib/hc_matchfinder.h" + struct lzx_output_bitstream; /* Codewords for the LZX Huffman codes. */ @@ -164,8 +177,8 @@ struct lzx_codewords { /* Codeword lengths (in bits) for the LZX Huffman codes. * A zero length means the corresponding codeword has zero frequency. */ struct lzx_lens { - u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS]; - u8 len[LZX_LENCODE_NUM_SYMBOLS]; + u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS + 1]; + u8 len[LZX_LENCODE_NUM_SYMBOLS + 1]; u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; @@ -201,14 +214,28 @@ struct lzx_freqs { u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; -/* Intermediate LZX match/literal format */ -struct lzx_item { - - /* Bits 0 - 9: Main symbol - * Bits 10 - 17: Length symbol - * Bits 18 - 22: Number of extra offset bits - * Bits 23+ : Extra offset bits */ - u64 data; +/* + * Represents a run of literals followed by a match or end-of-block. This + * struct is needed to temporarily store items chosen by the parser, since items + * cannot be written until all items for the block have been chosen and the + * block's Huffman codes have been computed. + */ +struct lzx_sequence { + + /* The number of literals in the run. This may be 0. The literals are + * not stored explicitly in this structure; instead, they are read + * directly from the uncompressed data. */ + u16 litrunlen; + + /* If the next field doesn't indicate end-of-block, then this is the + * match length minus LZX_MIN_MATCH_LEN. */ + u16 adjusted_length; + + /* If bit 31 is clear, then this field contains the match header in bits + * 0-8 and the match offset minus LZX_OFFSET_ADJUSTMENT in bits 9-30. + * Otherwise, this sequence's literal run was the last literal run in + * the block, so there is no match that follows it. */ + u32 adjusted_offset_and_match_hdr; }; /* @@ -235,7 +262,7 @@ struct lzx_optimum_node { * This variable is divided into two bitfields. * * Literals: - * Low bits are 1, high bits are the literal. + * Low bits are 0, high bits are the literal. * * Explicit offset matches: * Low bits are the match length, high bits are the offset plus 2. @@ -381,36 +408,29 @@ struct lzx_compressor { struct lzx_codes codes[2]; unsigned codes_index; - /* - * The match/literal sequence the algorithm chose for the current block. - * - * Notes on how large this array actually needs to be: - * - * - In lzx_compress_near_optimal(), the maximum block size is - * 'LZX_DIV_BLOCK_SIZE + LZX_MAX_MATCH_LEN - 1' bytes. This occurs if - * a match of the maximum length is found on the last byte. Although - * it is impossible for this particular case to actually result in a - * parse of all literals, we reserve this many spaces anyway. - * - * - The worst case for lzx_compress_lazy() is a block of almost all - * literals that ends with a series of matches of increasing scores, - * causing a sequence of literals to be chosen before the last match - * is finally chosen. The number of items actually chosen in this - * scenario is limited by the number of distinct match scores that - * exist for matches shorter than 'nice_match_length'. Having - * 'LZX_MAX_MATCH_LEN - 1' extra spaces is plenty for now. - */ - struct lzx_item chosen_items[LZX_DIV_BLOCK_SIZE + LZX_MAX_MATCH_LEN - 1]; + /* The matches and literals that the parser has chosen for the current + * block. The required length of this array is limited by the maximum + * number of matches that can ever be chosen for a single block, plus + * one for the special entry at the end. */ + struct lzx_sequence chosen_sequences[ + DIV_ROUND_UP(LZX_DIV_BLOCK_SIZE, LZX_MIN_MATCH_LEN) + 1]; + + /* Tables for mapping adjusted offsets to offset slots */ - /* Table mapping match offset => offset slot for small offsets */ -#define LZX_NUM_FAST_OFFSETS 32768 - u8 offset_slot_fast[LZX_NUM_FAST_OFFSETS]; + /* offset slots [0, 29] */ + u8 offset_slot_tab_1[32768]; + + /* offset slots [30, 49] */ + u8 offset_slot_tab_2[128]; union { /* Data for greedy or lazy parsing */ struct { /* Hash chains matchfinder (MUST BE LAST!!!) */ - struct hc_matchfinder hc_mf; + union { + struct hc_matchfinder_16 hc_mf_16; + struct hc_matchfinder_32 hc_mf_32; + }; }; /* Data for near-optimal parsing */ @@ -465,16 +485,43 @@ struct lzx_compressor { LZX_MAX_MATCHES_PER_POS + LZX_MAX_MATCH_LEN - 1]; - /* Hash table for finding length 2 matches */ - pos_t hash2_tab[LZX_HASH2_LENGTH] - _aligned_attribute(MATCHFINDER_ALIGNMENT); - /* Binary trees matchfinder (MUST BE LAST!!!) */ - struct bt_matchfinder bt_mf; + union { + struct bt_matchfinder_16 bt_mf_16; + struct bt_matchfinder_32 bt_mf_32; + }; }; }; }; +/* + * Will a matchfinder using 16-bit positions be sufficient for compressing + * buffers of up to the specified size? The limit could be 65536 bytes, but we + * also want to optimize out the use of offset_slot_tab_2 in the 16-bit case. + * This requires that the limit be no more than the length of offset_slot_tab_1 + * (currently 32768). + */ +static inline bool +lzx_is_16_bit(size_t max_bufsize) +{ + STATIC_ASSERT(ARRAY_LEN(((struct lzx_compressor *)0)->offset_slot_tab_1) == 32768); + return max_bufsize <= 32768; +} + +/* + * The following macros call either the 16-bit or the 32-bit version of a + * matchfinder function based on the value of 'is_16_bit', which will be known + * at compilation time. + */ + +#define CALL_HC_MF(is_16_bit, c, funcname, ...) \ + ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->hc_mf_16, ##__VA_ARGS__) : \ + CONCAT(funcname, _32)(&(c)->hc_mf_32, ##__VA_ARGS__)); + +#define CALL_BT_MF(is_16_bit, c, funcname, ...) \ + ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->bt_mf_16, ##__VA_ARGS__) : \ + CONCAT(funcname, _32)(&(c)->bt_mf_32, ##__VA_ARGS__)); + /* * Structure to keep track of the current state of sending bits to the * compressed output buffer. @@ -484,7 +531,7 @@ struct lzx_compressor { struct lzx_output_bitstream { /* Bits that haven't yet been written to the output buffer. */ - u32 bitbuf; + machine_word_t bitbuf; /* Number of bits currently held in @bitbuf. */ u32 bitcount; @@ -501,6 +548,10 @@ struct lzx_output_bitstream { u8 *end; }; +/* Can the specified number of bits always be added to 'bitbuf' after any + * pending 16-bit coding units have been flushed? */ +#define CAN_BUFFER(n) ((n) <= (8 * sizeof(machine_word_t)) - 15) + /* * Initialize the output bitstream. * @@ -521,68 +572,38 @@ lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size) os->end = os->start + (size & ~1); } -/* - * Write some bits to the output bitstream. - * - * The bits are given by the low-order @num_bits bits of @bits. Higher-order - * bits in @bits cannot be set. At most 17 bits can be written at once. - * - * @max_num_bits is a compile-time constant that specifies the maximum number of - * bits that can ever be written at the call site. It is used to optimize away - * the conditional code for writing a second 16-bit coding unit when writing - * fewer than 17 bits. - * - * If the output buffer space is exhausted, then the bits will be ignored, and - * lzx_flush_output() will return 0 when it gets called. - */ +/* Add some bits to the bitbuffer variable of the output bitstream. The caller + * must make sure there is enough room. */ static inline void -lzx_write_varbits(struct lzx_output_bitstream *os, - const u32 bits, const unsigned num_bits, - const unsigned max_num_bits) +lzx_add_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits) { - /* This code is optimized for LZX, which never needs to write more than - * 17 bits at once. */ - LZX_ASSERT(num_bits <= 17); - LZX_ASSERT(num_bits <= max_num_bits); - LZX_ASSERT(os->bitcount <= 15); - - /* Add the bits to the bit buffer variable. @bitcount will be at most - * 15, so there will be just enough space for the maximum possible - * @num_bits of 17. */ - os->bitcount += num_bits; os->bitbuf = (os->bitbuf << num_bits) | bits; + os->bitcount += num_bits; +} - /* Check whether any coding units need to be written. */ - if (os->bitcount >= 16) { - - os->bitcount -= 16; - - /* Write a coding unit, unless it would overflow the buffer. */ - if (os->next != os->end) { - put_unaligned_u16_le(os->bitbuf >> os->bitcount, os->next); - os->next += 2; - } - - /* If writing 17 bits, a second coding unit might need to be - * written. But because 'max_num_bits' is a compile-time - * constant, the compiler will optimize away this code at most - * call sites. */ - if (max_num_bits == 17 && os->bitcount == 16) { - if (os->next != os->end) { - put_unaligned_u16_le(os->bitbuf, os->next); - os->next += 2; - } - os->bitcount = 0; - } - } +/* Flush bits from the bitbuffer variable to the output buffer. 'max_num_bits' + * specifies the maximum number of bits that may have been added since the last + * flush. */ +static inline void +lzx_flush_bits(struct lzx_output_bitstream *os, unsigned max_num_bits) +{ + if (os->end - os->next < 6) + return; + put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 16), os->next + 0); + if (max_num_bits > 16) + put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 32), os->next + 2); + if (max_num_bits > 32) + put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 48), os->next + 4); + os->next += (os->bitcount >> 4) << 1; + os->bitcount &= 15; } -/* Use when @num_bits is a compile-time constant. Otherwise use - * lzx_write_varbits(). */ +/* Add at most 16 bits to the bitbuffer and flush it. */ static inline void lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits) { - lzx_write_varbits(os, bits, num_bits, num_bits); + lzx_add_bits(os, bits, num_bits); + lzx_flush_bits(os, 16); } /* @@ -592,7 +613,7 @@ lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits) static u32 lzx_flush_output(struct lzx_output_bitstream *os) { - if (os->next == os->end) + if (os->end - os->next < 6) return 0; if (os->bitcount != 0) { @@ -613,20 +634,27 @@ lzx_make_huffman_codes(struct lzx_compressor *c) const struct lzx_freqs *freqs = &c->freqs; struct lzx_codes *codes = &c->codes[c->codes_index]; + STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 && + MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN); + STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 9 && + LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN); + STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS && + ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN); + make_canonical_huffman_code(c->num_main_syms, - LZX_MAX_MAIN_CODEWORD_LEN, + MAIN_CODEWORD_LIMIT, freqs->main, codes->lens.main, codes->codewords.main); make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS, - LZX_MAX_LEN_CODEWORD_LEN, + LENGTH_CODEWORD_LIMIT, freqs->len, codes->lens.len, codes->codewords.len); make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS, - LZX_MAX_ALIGNED_CODEWORD_LEN, + ALIGNED_CODEWORD_LIMIT, freqs->aligned, codes->lens.aligned, codes->codewords.aligned); @@ -642,7 +670,6 @@ lzx_reset_symbol_frequencies(struct lzx_compressor *c) static unsigned lzx_compute_precode_items(const u8 lens[restrict], const u8 prev_lens[restrict], - const unsigned num_lens, u32 precode_freqs[restrict], unsigned precode_items[restrict]) { @@ -655,16 +682,17 @@ lzx_compute_precode_items(const u8 lens[restrict], itemptr = precode_items; run_start = 0; - do { - /* Find the next run of codeword lengths. */ + + while (!((len = lens[run_start]) & 0x80)) { /* len = the length being repeated */ - len = lens[run_start]; + + /* Find the next run of codeword lengths. */ run_end = run_start + 1; /* Fast case for a single length. */ - if (likely(run_end == num_lens || len != lens[run_end])) { + if (likely(len != lens[run_end])) { delta = prev_lens[run_start] - len; if (delta < 0) delta += 17; @@ -677,7 +705,7 @@ lzx_compute_precode_items(const u8 lens[restrict], /* Extend the run. */ do { run_end++; - } while (run_end != num_lens && len == lens[run_end]); + } while (len == lens[run_end]); if (len == 0) { /* Run of zeroes. */ @@ -723,7 +751,7 @@ lzx_compute_precode_items(const u8 lens[restrict], *itemptr++ = delta; run_start++; } - } while (run_start != num_lens); + } return itemptr - precode_items; } @@ -771,6 +799,8 @@ lzx_write_compressed_code(struct lzx_output_bitstream *os, unsigned precode_item; unsigned precode_sym; unsigned i; + u8 saved = lens[num_lens]; + *(u8 *)(lens + num_lens) = 0x80; for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) precode_freqs[i] = 0; @@ -779,13 +809,14 @@ lzx_write_compressed_code(struct lzx_output_bitstream *os, * the codeword lengths in the larger code will be output. */ num_precode_items = lzx_compute_precode_items(lens, prev_lens, - num_lens, precode_freqs, precode_items); /* Build the precode. */ + STATIC_ASSERT(PRE_CODEWORD_LIMIT >= 5 && + PRE_CODEWORD_LIMIT <= LZX_MAX_PRE_CODEWORD_LEN); make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, - LZX_MAX_PRE_CODEWORD_LEN, + PRE_CODEWORD_LIMIT, precode_freqs, precode_lens, precode_codewords); @@ -797,78 +828,25 @@ lzx_write_compressed_code(struct lzx_output_bitstream *os, for (i = 0; i < num_precode_items; i++) { precode_item = precode_items[i]; precode_sym = precode_item & 0x1F; - lzx_write_varbits(os, precode_codewords[precode_sym], - precode_lens[precode_sym], - LZX_MAX_PRE_CODEWORD_LEN); + lzx_add_bits(os, precode_codewords[precode_sym], + precode_lens[precode_sym]); if (precode_sym >= 17) { if (precode_sym == 17) { - lzx_write_bits(os, precode_item >> 5, 4); + lzx_add_bits(os, precode_item >> 5, 4); } else if (precode_sym == 18) { - lzx_write_bits(os, precode_item >> 5, 5); + lzx_add_bits(os, precode_item >> 5, 5); } else { - lzx_write_bits(os, (precode_item >> 5) & 1, 1); + lzx_add_bits(os, (precode_item >> 5) & 1, 1); precode_sym = precode_item >> 6; - lzx_write_varbits(os, precode_codewords[precode_sym], - precode_lens[precode_sym], - LZX_MAX_PRE_CODEWORD_LEN); + lzx_add_bits(os, precode_codewords[precode_sym], + precode_lens[precode_sym]); } } - } -} - -/* Output a match or literal. */ -static inline void -lzx_write_item(struct lzx_output_bitstream *os, struct lzx_item item, - unsigned ones_if_aligned, const struct lzx_codes *codes) -{ - u64 data = item.data; - unsigned main_symbol; - unsigned len_symbol; - unsigned num_extra_bits; - u32 extra_bits; - - main_symbol = data & 0x3FF; - - lzx_write_varbits(os, codes->codewords.main[main_symbol], - codes->lens.main[main_symbol], - LZX_MAX_MAIN_CODEWORD_LEN); - - if (main_symbol < LZX_NUM_CHARS) /* Literal? */ - return; - - len_symbol = (data >> 10) & 0xFF; - - if (len_symbol != LZX_LENCODE_NUM_SYMBOLS) { - lzx_write_varbits(os, codes->codewords.len[len_symbol], - codes->lens.len[len_symbol], - LZX_MAX_LEN_CODEWORD_LEN); + STATIC_ASSERT(CAN_BUFFER(2 * PRE_CODEWORD_LIMIT + 1)); + lzx_flush_bits(os, 2 * PRE_CODEWORD_LIMIT + 1); } - num_extra_bits = (data >> 18) & 0x1F; - if (num_extra_bits == 0) /* Small offset or repeat offset match? */ - return; - - extra_bits = data >> 23; - - if ((num_extra_bits & ones_if_aligned) >= LZX_NUM_ALIGNED_OFFSET_BITS) { - - /* Aligned offset blocks: The low 3 bits of the extra offset - * bits are Huffman-encoded using the aligned offset code. The - * remaining bits are output literally. */ - - lzx_write_varbits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS, - num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS, - 17 - LZX_NUM_ALIGNED_OFFSET_BITS); - - lzx_write_varbits(os, - codes->codewords.aligned[extra_bits & LZX_ALIGNED_OFFSET_BITMASK], - codes->lens.aligned[extra_bits & LZX_ALIGNED_OFFSET_BITMASK], - LZX_MAX_ALIGNED_CODEWORD_LEN); - } else { - /* Verbatim blocks, or fewer than 3 extra bits: All extra - * offset bits are output literally. */ - lzx_write_varbits(os, extra_bits, num_extra_bits, 17); - } + *(u8 *)(lens + num_lens) = saved; } /* @@ -881,39 +859,169 @@ lzx_write_item(struct lzx_output_bitstream *os, struct lzx_item item, * @block_type * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or * LZX_BLOCKTYPE_VERBATIM). - * @items - * The array of matches/literals to output. - * @num_items - * Number of matches/literals to output (length of @items). + * @block_data + * The uncompressed data of the block. + * @sequences + * The matches and literals to output, given as a series of sequences. * @codes * The main, length, and aligned offset Huffman codes for the current * LZX compressed block. */ static void -lzx_write_items(struct lzx_output_bitstream *os, int block_type, - const struct lzx_item items[], u32 num_items, - const struct lzx_codes *codes) +lzx_write_sequences(struct lzx_output_bitstream *os, int block_type, + const u8 *block_data, const struct lzx_sequence sequences[], + const struct lzx_codes *codes) { - unsigned ones_if_aligned = 0U - (block_type == LZX_BLOCKTYPE_ALIGNED); + const struct lzx_sequence *seq = sequences; + u32 ones_if_aligned = 0 - (block_type == LZX_BLOCKTYPE_ALIGNED); + + for (;;) { + /* Output the next sequence. */ + + unsigned litrunlen = seq->litrunlen; + unsigned match_hdr; + unsigned main_symbol; + unsigned adjusted_length; + u32 adjusted_offset; + unsigned offset_slot; + unsigned num_extra_bits; + u32 extra_bits; + + /* Output the literal run of the sequence. */ + + if (litrunlen) { /* Is the literal run nonempty? */ + + /* Verify optimization is enabled on 64-bit */ + STATIC_ASSERT(sizeof(machine_word_t) < 8 || + CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT)); + + if (CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT)) { + + /* 64-bit: write 4 literals at a time. */ + while (litrunlen >= 4) { + unsigned lit0 = block_data[0]; + unsigned lit1 = block_data[1]; + unsigned lit2 = block_data[2]; + unsigned lit3 = block_data[3]; + lzx_add_bits(os, codes->codewords.main[lit0], codes->lens.main[lit0]); + lzx_add_bits(os, codes->codewords.main[lit1], codes->lens.main[lit1]); + lzx_add_bits(os, codes->codewords.main[lit2], codes->lens.main[lit2]); + lzx_add_bits(os, codes->codewords.main[lit3], codes->lens.main[lit3]); + lzx_flush_bits(os, 4 * MAIN_CODEWORD_LIMIT); + block_data += 4; + litrunlen -= 4; + } + if (litrunlen--) { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]); + if (litrunlen--) { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]); + if (litrunlen--) { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]); + lzx_flush_bits(os, 3 * MAIN_CODEWORD_LIMIT); + } else { + lzx_flush_bits(os, 2 * MAIN_CODEWORD_LIMIT); + } + } else { + lzx_flush_bits(os, 1 * MAIN_CODEWORD_LIMIT); + } + } + } else { + /* 32-bit: write 1 literal at a time. */ + do { + unsigned lit = *block_data++; + lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]); + lzx_flush_bits(os, MAIN_CODEWORD_LIMIT); + } while (--litrunlen); + } + } + + /* Was this the last literal run? */ + if (seq->adjusted_offset_and_match_hdr & 0x80000000) + return; + + /* Nope; output the match. */ + + match_hdr = seq->adjusted_offset_and_match_hdr & 0x1FF; + main_symbol = LZX_NUM_CHARS + match_hdr; + adjusted_length = seq->adjusted_length; + + block_data += adjusted_length + LZX_MIN_MATCH_LEN; + + offset_slot = match_hdr / LZX_NUM_LEN_HEADERS; + adjusted_offset = seq->adjusted_offset_and_match_hdr >> 9; - for (u32 i = 0; i < num_items; i++) - lzx_write_item(os, items[i], ones_if_aligned, codes); + num_extra_bits = lzx_extra_offset_bits[offset_slot]; + extra_bits = adjusted_offset - lzx_offset_slot_base[offset_slot]; + + #define MAX_MATCH_BITS (MAIN_CODEWORD_LIMIT + LENGTH_CODEWORD_LIMIT + \ + 14 + ALIGNED_CODEWORD_LIMIT) + + /* Verify optimization is enabled on 64-bit */ + STATIC_ASSERT(sizeof(machine_word_t) < 8 || CAN_BUFFER(MAX_MATCH_BITS)); + + /* Output the main symbol for the match. */ + + lzx_add_bits(os, codes->codewords.main[main_symbol], + codes->lens.main[main_symbol]); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, MAIN_CODEWORD_LIMIT); + + /* If needed, output the length symbol for the match. */ + + if (adjusted_length >= LZX_NUM_PRIMARY_LENS) { + lzx_add_bits(os, codes->codewords.len[adjusted_length - LZX_NUM_PRIMARY_LENS], + codes->lens.len[adjusted_length - LZX_NUM_PRIMARY_LENS]); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, LENGTH_CODEWORD_LIMIT); + } + + /* Output the extra offset bits for the match. In aligned + * offset blocks, the lowest 3 bits of the adjusted offset are + * Huffman-encoded using the aligned offset code, provided that + * there are at least extra 3 offset bits required. All other + * extra offset bits are output verbatim. */ + + if ((adjusted_offset & ones_if_aligned) >= 16) { + + lzx_add_bits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS, + num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, 14); + + lzx_add_bits(os, codes->codewords.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK], + codes->lens.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT); + } else { + STATIC_ASSERT(CAN_BUFFER(17)); + + lzx_add_bits(os, extra_bits, num_extra_bits); + if (!CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, 17); + } + + if (CAN_BUFFER(MAX_MATCH_BITS)) + lzx_flush_bits(os, MAX_MATCH_BITS); + + /* Advance to the next sequence. */ + seq++; + } } static void -lzx_write_compressed_block(int block_type, +lzx_write_compressed_block(const u8 *block_begin, + int block_type, u32 block_size, unsigned window_order, unsigned num_main_syms, - const struct lzx_item chosen_items[], - u32 num_chosen_items, + const struct lzx_sequence sequences[], const struct lzx_codes * codes, const struct lzx_lens * prev_lens, struct lzx_output_bitstream * os) { - LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED || - block_type == LZX_BLOCKTYPE_VERBATIM); - /* The first three bits indicate the type of block and are one of the * LZX_BLOCKTYPE_* constants. */ lzx_write_bits(os, block_type, 3); @@ -966,7 +1074,7 @@ lzx_write_compressed_block(int block_type, LZX_LENCODE_NUM_SYMBOLS); /* Output the compressed matches and literals. */ - lzx_write_items(os, block_type, chosen_items, num_chosen_items, codes); + lzx_write_sequences(os, block_type, block_begin, sequences, codes); } /* Given the frequencies of symbols in an LZX-compressed block and the @@ -996,6 +1104,19 @@ lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs, return LZX_BLOCKTYPE_VERBATIM; } +/* + * Return the offset slot for the specified adjusted match offset, using the + * compressor's acceleration tables to speed up the mapping. + */ +static inline unsigned +lzx_comp_get_offset_slot(struct lzx_compressor *c, u32 adjusted_offset, + bool is_16_bit) +{ + if (is_16_bit || adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1)) + return c->offset_slot_tab_1[adjusted_offset]; + return c->offset_slot_tab_2[adjusted_offset >> 14]; +} + /* * Finish an LZX block: * @@ -1006,7 +1127,7 @@ lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs, */ static void lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os, - u32 block_size, u32 num_chosen_items) + const u8 *block_begin, u32 block_size, u32 seq_idx) { int block_type; @@ -1014,181 +1135,245 @@ lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os, block_type = lzx_choose_verbatim_or_aligned(&c->freqs, &c->codes[c->codes_index]); - lzx_write_compressed_block(block_type, + lzx_write_compressed_block(block_begin, + block_type, block_size, c->window_order, c->num_main_syms, - c->chosen_items, - num_chosen_items, + &c->chosen_sequences[seq_idx], &c->codes[c->codes_index], &c->codes[c->codes_index ^ 1].lens, os); c->codes_index ^= 1; } -/* Return the offset slot for the specified offset, which must be - * less than LZX_NUM_FAST_OFFSETS. */ -static inline unsigned -lzx_get_offset_slot_fast(struct lzx_compressor *c, u32 offset) +/* Tally the Huffman symbol for a literal and increment the literal run length. + */ +static inline void +lzx_record_literal(struct lzx_compressor *c, unsigned literal, u32 *litrunlen_p) { - LZX_ASSERT(offset < LZX_NUM_FAST_OFFSETS); - return c->offset_slot_fast[offset]; + c->freqs.main[literal]++; + ++*litrunlen_p; } -/* Tally, and optionally record, the specified literal byte. */ +/* Tally the Huffman symbol for a match, save the match data and the length of + * the preceding literal run in the next lzx_sequence, and update the recent + * offsets queue. */ static inline void -lzx_declare_literal(struct lzx_compressor *c, unsigned literal, - struct lzx_item **next_chosen_item) +lzx_record_match(struct lzx_compressor *c, unsigned length, u32 offset_data, + u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], bool is_16_bit, + u32 *litrunlen_p, struct lzx_sequence **next_seq_p) { - unsigned main_symbol = lzx_main_symbol_for_literal(literal); + u32 litrunlen = *litrunlen_p; + struct lzx_sequence *next_seq = *next_seq_p; + unsigned offset_slot; + unsigned v; - c->freqs.main[main_symbol]++; + v = length - LZX_MIN_MATCH_LEN; - if (next_chosen_item) { - *(*next_chosen_item)++ = (struct lzx_item) { - .data = main_symbol, - }; + /* Save the literal run length and adjusted length. */ + next_seq->litrunlen = litrunlen; + next_seq->adjusted_length = v; + + /* Compute the length header and tally the length symbol if needed */ + if (v >= LZX_NUM_PRIMARY_LENS) { + c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++; + v = LZX_NUM_PRIMARY_LENS; } -} -/* Tally, and optionally record, the specified repeat offset match. */ -static inline void -lzx_declare_repeat_offset_match(struct lzx_compressor *c, - unsigned len, unsigned rep_index, - struct lzx_item **next_chosen_item) -{ - unsigned len_header; - unsigned len_symbol; - unsigned main_symbol; + /* Compute the offset slot */ + offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit); + + /* Compute the match header. */ + v += offset_slot * LZX_NUM_LEN_HEADERS; - if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) { - len_header = len - LZX_MIN_MATCH_LEN; - len_symbol = LZX_LENCODE_NUM_SYMBOLS; + /* Save the adjusted offset and match header. */ + next_seq->adjusted_offset_and_match_hdr = (offset_data << 9) | v; + + /* Tally the main symbol. */ + c->freqs.main[LZX_NUM_CHARS + v]++; + + /* Update the recent offsets queue. */ + if (offset_data < LZX_NUM_RECENT_OFFSETS) { + /* Repeat offset match */ + swap(recent_offsets[0], recent_offsets[offset_data]); } else { - len_header = LZX_NUM_PRIMARY_LENS; - len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS; - c->freqs.len[len_symbol]++; + /* Explicit offset match */ + + /* Tally the aligned offset symbol if needed */ + if (offset_data >= 16) + c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++; + + recent_offsets[2] = recent_offsets[1]; + recent_offsets[1] = recent_offsets[0]; + recent_offsets[0] = offset_data - LZX_OFFSET_ADJUSTMENT; } - main_symbol = lzx_main_symbol_for_match(rep_index, len_header); + /* Reset the literal run length and advance to the next sequence. */ + *next_seq_p = next_seq + 1; + *litrunlen_p = 0; +} - c->freqs.main[main_symbol]++; +/* Finish the last lzx_sequence. The last lzx_sequence is just a literal run; + * there is no match. This literal run may be empty. */ +static inline void +lzx_finish_sequence(struct lzx_sequence *last_seq, u32 litrunlen) +{ + last_seq->litrunlen = litrunlen; - if (next_chosen_item) { - *(*next_chosen_item)++ = (struct lzx_item) { - .data = (u64)main_symbol | ((u64)len_symbol << 10), - }; - } + /* Special value to mark last sequence */ + last_seq->adjusted_offset_and_match_hdr = 0x80000000; } -/* Tally, and optionally record, the specified explicit offset match. */ +/* + * Given the minimum-cost path computed through the item graph for the current + * block, walk the path and count how many of each symbol in each Huffman-coded + * alphabet would be required to output the items (matches and literals) along + * the path. + * + * Note that the path will be walked backwards (from the end of the block to the + * beginning of the block), but this doesn't matter because this function only + * computes frequencies. + */ static inline void -lzx_declare_explicit_offset_match(struct lzx_compressor *c, unsigned len, u32 offset, - struct lzx_item **next_chosen_item) +lzx_tally_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit) { - unsigned len_header; - unsigned len_symbol; - unsigned main_symbol; - unsigned offset_slot; - unsigned num_extra_bits; - u32 extra_bits; + u32 node_idx = block_size; + for (;;) { + u32 len; + u32 offset_data; + unsigned v; + unsigned offset_slot; - if (len - LZX_MIN_MATCH_LEN < LZX_NUM_PRIMARY_LENS) { - len_header = len - LZX_MIN_MATCH_LEN; - len_symbol = LZX_LENCODE_NUM_SYMBOLS; - } else { - len_header = LZX_NUM_PRIMARY_LENS; - len_symbol = len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS; - c->freqs.len[len_symbol]++; - } + /* Tally literals until either a match or the beginning of the + * block is reached. */ + for (;;) { + u32 item = c->optimum_nodes[node_idx].item; - offset_slot = (offset < LZX_NUM_FAST_OFFSETS) ? - lzx_get_offset_slot_fast(c, offset) : - lzx_get_offset_slot(offset); + len = item & OPTIMUM_LEN_MASK; + offset_data = item >> OPTIMUM_OFFSET_SHIFT; - main_symbol = lzx_main_symbol_for_match(offset_slot, len_header); + if (len != 0) /* Not a literal? */ + break; - c->freqs.main[main_symbol]++; + /* Tally the main symbol for the literal. */ + c->freqs.main[offset_data]++; - num_extra_bits = lzx_extra_offset_bits[offset_slot]; + if (--node_idx == 0) /* Beginning of block was reached? */ + return; + } - if (num_extra_bits >= LZX_NUM_ALIGNED_OFFSET_BITS) - c->freqs.aligned[(offset + LZX_OFFSET_ADJUSTMENT) & - LZX_ALIGNED_OFFSET_BITMASK]++; + node_idx -= len; - if (next_chosen_item) { + /* Tally a match. */ - extra_bits = (offset + LZX_OFFSET_ADJUSTMENT) - - lzx_offset_slot_base[offset_slot]; + /* Tally the aligned offset symbol if needed. */ + if (offset_data >= 16) + c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++; - BUILD_BUG_ON(LZX_MAINCODE_MAX_NUM_SYMBOLS > (1 << 10)); - BUILD_BUG_ON(LZX_LENCODE_NUM_SYMBOLS > (1 << 8)); - *(*next_chosen_item)++ = (struct lzx_item) { - .data = (u64)main_symbol | - ((u64)len_symbol << 10) | - ((u64)num_extra_bits << 18) | - ((u64)extra_bits << 23), - }; - } -} + /* Tally the length symbol if needed. */ + v = len - LZX_MIN_MATCH_LEN;; + if (v >= LZX_NUM_PRIMARY_LENS) { + c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++; + v = LZX_NUM_PRIMARY_LENS; + } + /* Tally the main symbol. */ + offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit); + v += offset_slot * LZX_NUM_LEN_HEADERS; + c->freqs.main[LZX_NUM_CHARS + v]++; -/* Tally, and optionally record, the specified match or literal. */ -static inline void -lzx_declare_item(struct lzx_compressor *c, u32 item, - struct lzx_item **next_chosen_item) -{ - u32 len = item & OPTIMUM_LEN_MASK; - u32 offset_data = item >> OPTIMUM_OFFSET_SHIFT; - - if (len == 1) - lzx_declare_literal(c, offset_data, next_chosen_item); - else if (offset_data < LZX_NUM_RECENT_OFFSETS) - lzx_declare_repeat_offset_match(c, len, offset_data, - next_chosen_item); - else - lzx_declare_explicit_offset_match(c, len, - offset_data - LZX_OFFSET_ADJUSTMENT, - next_chosen_item); + if (node_idx == 0) /* Beginning of block was reached? */ + return; + } } -static inline void -lzx_record_item_list(struct lzx_compressor *c, - struct lzx_optimum_node *cur_node, - struct lzx_item **next_chosen_item) +/* + * Like lzx_tally_item_list(), but this function also generates the list of + * lzx_sequences for the minimum-cost path and writes it to c->chosen_sequences, + * ready to be output to the bitstream after the Huffman codes are computed. + * The lzx_sequences will be written to decreasing memory addresses as the path + * is walked backwards, which means they will end up in the expected + * first-to-last order. The return value is the index in c->chosen_sequences at + * which the lzx_sequences begin. + */ +static inline u32 +lzx_record_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit) { - struct lzx_optimum_node *end_node; - u32 saved_item; - u32 item; + u32 node_idx = block_size; + u32 seq_idx = ARRAY_LEN(c->chosen_sequences) - 1; + u32 lit_start_node; - /* The list is currently in reverse order (last item to first item). - * Reverse it. */ - end_node = cur_node; - saved_item = cur_node->item; - do { - item = saved_item; - cur_node -= item & OPTIMUM_LEN_MASK; - saved_item = cur_node->item; - cur_node->item = item; - } while (cur_node != c->optimum_nodes); - - /* Walk the list of items from beginning to end, tallying and recording - * each item. */ - do { - lzx_declare_item(c, cur_node->item, next_chosen_item); - cur_node += (cur_node->item) & OPTIMUM_LEN_MASK; - } while (cur_node != end_node); -} + /* Special value to mark last sequence */ + c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = 0x80000000; -static inline void -lzx_tally_item_list(struct lzx_compressor *c, struct lzx_optimum_node *cur_node) -{ - /* Since we're just tallying the items, we don't need to reverse the - * list. Processing the items in reverse order is fine. */ - do { - lzx_declare_item(c, cur_node->item, NULL); - cur_node -= (cur_node->item & OPTIMUM_LEN_MASK); - } while (cur_node != c->optimum_nodes); + lit_start_node = node_idx; + for (;;) { + u32 len; + u32 offset_data; + unsigned v; + unsigned offset_slot; + + /* Record literals until either a match or the beginning of the + * block is reached. */ + for (;;) { + u32 item = c->optimum_nodes[node_idx].item; + + len = item & OPTIMUM_LEN_MASK; + offset_data = item >> OPTIMUM_OFFSET_SHIFT; + + if (len != 0) /* Not a literal? */ + break; + + /* Tally the main symbol for the literal. */ + c->freqs.main[offset_data]++; + + if (--node_idx == 0) /* Beginning of block was reached? */ + goto out; + } + + /* Save the literal run length for the next sequence (the + * "previous sequence" when walking backwards). */ + c->chosen_sequences[seq_idx--].litrunlen = lit_start_node - node_idx; + node_idx -= len; + lit_start_node = node_idx; + + /* Record a match. */ + + /* Tally the aligned offset symbol if needed. */ + if (offset_data >= 16) + c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++; + + /* Save the adjusted length. */ + v = len - LZX_MIN_MATCH_LEN; + c->chosen_sequences[seq_idx].adjusted_length = v; + + /* Tally the length symbol if needed. */ + if (v >= LZX_NUM_PRIMARY_LENS) { + c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++; + v = LZX_NUM_PRIMARY_LENS; + } + + /* Tally the main symbol. */ + offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit); + v += offset_slot * LZX_NUM_LEN_HEADERS; + c->freqs.main[LZX_NUM_CHARS + v]++; + + /* Save the adjusted offset and match header. */ + c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = + (offset_data << 9) | v; + + if (node_idx == 0) /* Beginning of block was reached? */ + goto out; + } + +out: + /* Save the literal run length for the first sequence. */ + c->chosen_sequences[seq_idx].litrunlen = lit_start_node - node_idx; + + /* Return the index in c->chosen_sequences at which the lzx_sequences + * begin. */ + return seq_idx; } /* @@ -1222,11 +1407,12 @@ lzx_tally_item_list(struct lzx_compressor *c, struct lzx_optimum_node *cur_node) * later. The algorithm does not solve this problem; it only considers the * lowest cost to reach each individual position. */ -static struct lzx_lru_queue +static inline struct lzx_lru_queue lzx_find_min_cost_path(struct lzx_compressor * const restrict c, const u8 * const restrict block_begin, const u32 block_size, - const struct lzx_lru_queue initial_queue) + const struct lzx_lru_queue initial_queue, + bool is_16_bit) { struct lzx_optimum_node *cur_node = c->optimum_nodes; struct lzx_optimum_node * const end_node = &c->optimum_nodes[block_size]; @@ -1242,7 +1428,7 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, * it is no longer needed. */ struct lzx_lru_queue queues[512]; - BUILD_BUG_ON(ARRAY_LEN(queues) < LZX_MAX_MATCH_LEN + 1); + STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1); #define QUEUE(in) (queues[(uintptr_t)(in) % ARRAY_LEN(queues)]) /* Initially, the cost to reach each node is "infinity". */ @@ -1294,7 +1480,7 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, matchptr = in_next - lzx_lru_queue_R0(QUEUE(in_next)); if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next)) goto R0_done; - BUILD_BUG_ON(LZX_MIN_MATCH_LEN != 2); + STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2); do { u32 cost = cur_node->cost + c->costs.match_cost[0][ @@ -1372,19 +1558,20 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, do { u32 offset = cache_ptr->offset; u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT; - unsigned offset_slot = (offset < LZX_NUM_FAST_OFFSETS) ? - lzx_get_offset_slot_fast(c, offset) : - lzx_get_offset_slot(offset); + unsigned offset_slot = lzx_comp_get_offset_slot(c, offset_data, + is_16_bit); + u32 base_cost = cur_node->cost; + + #if LZX_CONSIDER_ALIGNED_COSTS + if (offset_data >= 16) + base_cost += c->costs.aligned[offset_data & + LZX_ALIGNED_OFFSET_BITMASK]; + #endif + do { - u32 cost = cur_node->cost + + u32 cost = base_cost + c->costs.match_cost[offset_slot][ next_len - LZX_MIN_MATCH_LEN]; - #if LZX_CONSIDER_ALIGNED_COSTS - if (lzx_extra_offset_bits[offset_slot] >= - LZX_NUM_ALIGNED_OFFSET_BITS) - cost += c->costs.aligned[offset_data & - LZX_ALIGNED_OFFSET_BITMASK]; - #endif if (cost < (cur_node + next_len)->cost) { (cur_node + next_len)->cost = cost; (cur_node + next_len)->item = @@ -1402,8 +1589,7 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, * of coding the literal is integrated into the queue update * code below. */ literal = *in_next++; - cost = cur_node->cost + - c->costs.main[lzx_main_symbol_for_literal(literal)]; + cost = cur_node->cost + c->costs.main[literal]; /* Advance to the next position. */ cur_node++; @@ -1415,7 +1601,7 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, if (cost <= cur_node->cost) { /* Literal: queue remains unchanged. */ cur_node->cost = cost; - cur_node->item = (literal << OPTIMUM_OFFSET_SHIFT) | 1; + cur_node->item = (u32)literal << OPTIMUM_OFFSET_SHIFT; QUEUE(in_next) = QUEUE(in_next - 1); } else { /* Match: queue update is needed. */ @@ -1443,17 +1629,17 @@ lzx_find_min_cost_path(struct lzx_compressor * const restrict c, static void lzx_compute_match_costs(struct lzx_compressor *c) { - unsigned num_offset_slots = lzx_get_num_offset_slots(c->window_order); + unsigned num_offset_slots = (c->num_main_syms - LZX_NUM_CHARS) / LZX_NUM_LEN_HEADERS; struct lzx_costs *costs = &c->costs; for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) { u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST; - unsigned main_symbol = lzx_main_symbol_for_match(offset_slot, 0); + unsigned main_symbol = LZX_NUM_CHARS + (offset_slot * LZX_NUM_LEN_HEADERS); unsigned i; #if LZX_CONSIDER_ALIGNED_COSTS - if (lzx_extra_offset_bits[offset_slot] >= LZX_NUM_ALIGNED_OFFSET_BITS) + if (offset_slot >= 8) extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST; #endif @@ -1479,7 +1665,7 @@ lzx_set_default_costs(struct lzx_compressor *c, const u8 *block, u32 block_size) unsigned num_used_bytes; /* The costs below are hard coded to use a scaling factor of 16. */ - BUILD_BUG_ON(LZX_BIT_COST != 16); + STATIC_ASSERT(LZX_BIT_COST == 16); /* * Heuristics: @@ -1527,29 +1713,37 @@ lzx_update_costs(struct lzx_compressor *c) unsigned i; const struct lzx_lens *lens = &c->codes[c->codes_index].lens; - for (i = 0; i < c->num_main_syms; i++) - c->costs.main[i] = (lens->main[i] ? lens->main[i] : 15) * LZX_BIT_COST; + for (i = 0; i < c->num_main_syms; i++) { + c->costs.main[i] = (lens->main[i] ? lens->main[i] : + MAIN_CODEWORD_LIMIT) * LZX_BIT_COST; + } - for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) - c->costs.len[i] = (lens->len[i] ? lens->len[i] : 15) * LZX_BIT_COST; + for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) { + c->costs.len[i] = (lens->len[i] ? lens->len[i] : + LENGTH_CODEWORD_LIMIT) * LZX_BIT_COST; + } #if LZX_CONSIDER_ALIGNED_COSTS - for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) - c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : 7) * LZX_BIT_COST; + for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) { + c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : + ALIGNED_CODEWORD_LIMIT) * LZX_BIT_COST; + } #endif lzx_compute_match_costs(c); } -static struct lzx_lru_queue -lzx_optimize_and_write_block(struct lzx_compressor *c, - struct lzx_output_bitstream *os, - const u8 *block_begin, const u32 block_size, - const struct lzx_lru_queue initial_queue) +static inline struct lzx_lru_queue +lzx_optimize_and_write_block(struct lzx_compressor * const restrict c, + struct lzx_output_bitstream * const restrict os, + const u8 * const restrict block_begin, + const u32 block_size, + const struct lzx_lru_queue initial_queue, + bool is_16_bit) { unsigned num_passes_remaining = c->num_optim_passes; - struct lzx_item *next_chosen_item; struct lzx_lru_queue new_queue; + u32 seq_idx; /* The first optimization pass uses a default cost model. Each * additional optimization pass uses a cost model derived from the @@ -1559,18 +1753,17 @@ lzx_optimize_and_write_block(struct lzx_compressor *c, lzx_reset_symbol_frequencies(c); do { new_queue = lzx_find_min_cost_path(c, block_begin, block_size, - initial_queue); + initial_queue, is_16_bit); if (num_passes_remaining > 1) { - lzx_tally_item_list(c, c->optimum_nodes + block_size); + lzx_tally_item_list(c, block_size, is_16_bit); lzx_make_huffman_codes(c); lzx_update_costs(c); lzx_reset_symbol_frequencies(c); } } while (--num_passes_remaining); - next_chosen_item = c->chosen_items; - lzx_record_item_list(c, c->optimum_nodes + block_size, &next_chosen_item); - lzx_finish_block(c, os, block_size, next_chosen_item - c->chosen_items); + seq_idx = lzx_record_item_list(c, block_size, is_16_bit); + lzx_finish_block(c, os, block_begin, block_size, seq_idx); return new_queue; } @@ -1587,21 +1780,20 @@ lzx_optimize_and_write_block(struct lzx_compressor *c, * time, but rather to produce a compression ratio significantly better than a * simpler "greedy" or "lazy" parse while still being relatively fast. */ -static void +static inline void lzx_compress_near_optimal(struct lzx_compressor *c, - struct lzx_output_bitstream *os) + struct lzx_output_bitstream *os, + bool is_16_bit) { const u8 * const in_begin = c->in_buffer; const u8 * in_next = in_begin; const u8 * const in_end = in_begin + c->in_nbytes; - unsigned max_len = LZX_MAX_MATCH_LEN; - unsigned nice_len = min(c->nice_match_length, max_len); - u32 next_hash; + u32 max_len = LZX_MAX_MATCH_LEN; + u32 nice_len = min(c->nice_match_length, max_len); + u32 next_hashes[2] = {}; struct lzx_lru_queue queue; - bt_matchfinder_init(&c->bt_mf); - matchfinder_init(c->hash2_tab, LZX_HASH2_LENGTH); - next_hash = bt_matchfinder_hash_3_bytes(in_next); + CALL_BT_MF(is_16_bit, c, bt_matchfinder_init); lzx_lru_queue_init(&queue); do { @@ -1614,24 +1806,14 @@ lzx_compress_near_optimal(struct lzx_compressor *c, struct lz_match *cache_ptr = c->match_cache; do { struct lz_match *lz_matchptr; - u32 hash2; - pos_t cur_match; - unsigned best_len; + u32 best_len; /* If approaching the end of the input buffer, adjust * 'max_len' and 'nice_len' accordingly. */ if (unlikely(max_len > in_end - in_next)) { max_len = in_end - in_next; nice_len = min(max_len, nice_len); - - /* This extra check is needed to ensure that - * reading the next 3 bytes when looking for a - * length 2 match is valid. In addition, we - * cannot allow ourselves to find a length 2 - * match of the very last two bytes with the - * very first two bytes, since such a match has - * an offset too large to be represented. */ - if (unlikely(max_len < 3)) { + if (unlikely(max_len < BT_MATCHFINDER_REQUIRED_NBYTES)) { in_next++; cache_ptr->length = 0; cache_ptr++; @@ -1639,33 +1821,16 @@ lzx_compress_near_optimal(struct lzx_compressor *c, } } - lz_matchptr = cache_ptr + 1; - - /* Check for a length 2 match. */ - hash2 = lz_hash_2_bytes(in_next, LZX_HASH2_ORDER); - cur_match = c->hash2_tab[hash2]; - c->hash2_tab[hash2] = in_next - in_begin; - if (matchfinder_node_valid(cur_match) && - (LZX_HASH2_ORDER == 16 || - load_u16_unaligned(&in_begin[cur_match]) == - load_u16_unaligned(in_next))) - { - lz_matchptr->length = 2; - lz_matchptr->offset = in_next - &in_begin[cur_match]; - lz_matchptr++; - } - - /* Check for matches of length >= 3. */ - lz_matchptr = bt_matchfinder_get_matches(&c->bt_mf, - in_begin, - in_next, - 3, - max_len, - nice_len, - c->max_search_depth, - &next_hash, - &best_len, - lz_matchptr); + /* Check for matches. */ + lz_matchptr = CALL_BT_MF(is_16_bit, c, bt_matchfinder_get_matches, + in_begin, + in_next - in_begin, + max_len, + nice_len, + c->max_search_depth, + next_hashes, + &best_len, + cache_ptr + 1); in_next++; cache_ptr->length = lz_matchptr - (cache_ptr + 1); cache_ptr = lz_matchptr; @@ -1688,22 +1853,20 @@ lzx_compress_near_optimal(struct lzx_compressor *c, if (unlikely(max_len > in_end - in_next)) { max_len = in_end - in_next; nice_len = min(max_len, nice_len); - if (unlikely(max_len < 3)) { + if (unlikely(max_len < BT_MATCHFINDER_REQUIRED_NBYTES)) { in_next++; cache_ptr->length = 0; cache_ptr++; continue; } } - c->hash2_tab[lz_hash_2_bytes(in_next, LZX_HASH2_ORDER)] = - in_next - in_begin; - bt_matchfinder_skip_position(&c->bt_mf, - in_begin, - in_next, - in_end, - nice_len, - c->max_search_depth, - &next_hash); + CALL_BT_MF(is_16_bit, c, bt_matchfinder_skip_position, + in_begin, + in_next - in_begin, + max_len, + nice_len, + c->max_search_depth, + next_hashes); in_next++; cache_ptr->length = 0; cache_ptr++; @@ -1717,10 +1880,24 @@ lzx_compress_near_optimal(struct lzx_compressor *c, queue = lzx_optimize_and_write_block(c, os, in_block_begin, in_next - in_block_begin, - queue); + queue, is_16_bit); } while (in_next != in_end); } +static void +lzx_compress_near_optimal_16(struct lzx_compressor *c, + struct lzx_output_bitstream *os) +{ + lzx_compress_near_optimal(c, os, true); +} + +static void +lzx_compress_near_optimal_32(struct lzx_compressor *c, + struct lzx_output_bitstream *os) +{ + lzx_compress_near_optimal(c, os, false); +} + /* * Given a pointer to the current byte sequence and the current list of recent * match offsets, find the longest repeat offset match. @@ -1733,11 +1910,10 @@ lzx_compress_near_optimal(struct lzx_compressor *c, static unsigned lzx_find_longest_repeat_offset_match(const u8 * const in_next, const u32 bytes_remaining, - struct lzx_lru_queue queue, + const u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], unsigned *rep_max_idx_ret) { - BUILD_BUG_ON(LZX_NUM_RECENT_OFFSETS != 3); - LZX_ASSERT(bytes_remaining >= 2); + STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3); const unsigned max_len = min(bytes_remaining, LZX_MAX_MATCH_LEN); const u16 next_2_bytes = load_u16_unaligned(in_next); @@ -1746,14 +1922,14 @@ lzx_find_longest_repeat_offset_match(const u8 * const in_next, unsigned rep_max_idx; unsigned rep_len; - matchptr = in_next - lzx_lru_queue_pop(&queue); + matchptr = in_next - recent_offsets[0]; if (load_u16_unaligned(matchptr) == next_2_bytes) rep_max_len = lz_extend(in_next, matchptr, 2, max_len); else rep_max_len = 0; rep_max_idx = 0; - matchptr = in_next - lzx_lru_queue_pop(&queue); + matchptr = in_next - recent_offsets[1]; if (load_u16_unaligned(matchptr) == next_2_bytes) { rep_len = lz_extend(in_next, matchptr, 2, max_len); if (rep_len > rep_max_len) { @@ -1762,7 +1938,7 @@ lzx_find_longest_repeat_offset_match(const u8 * const in_next, } } - matchptr = in_next - lzx_lru_queue_pop(&queue); + matchptr = in_next - recent_offsets[2]; if (load_u16_unaligned(matchptr) == next_2_bytes) { rep_len = lz_extend(in_next, matchptr, 2, max_len); if (rep_len > rep_max_len) { @@ -1797,18 +1973,20 @@ lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx) } /* This is the "lazy" LZX compressor. */ -static void -lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) +static inline void +lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os, + bool is_16_bit) { const u8 * const in_begin = c->in_buffer; const u8 * in_next = in_begin; const u8 * const in_end = in_begin + c->in_nbytes; unsigned max_len = LZX_MAX_MATCH_LEN; unsigned nice_len = min(c->nice_match_length, max_len); - struct lzx_lru_queue queue; + STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3); + u32 recent_offsets[3] = {1, 1, 1}; + u32 next_hashes[2] = {}; - hc_matchfinder_init(&c->hc_mf); - lzx_lru_queue_init(&queue); + CALL_HC_MF(is_16_bit, c, hc_matchfinder_init); do { /* Starting a new block */ @@ -1816,7 +1994,7 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) const u8 * const in_block_begin = in_next; const u8 * const in_block_end = in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next); - struct lzx_item *next_chosen_item = c->chosen_items; + struct lzx_sequence *next_seq = c->chosen_sequences; unsigned cur_len; u32 cur_offset; u32 cur_offset_data; @@ -1829,6 +2007,7 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) unsigned rep_max_idx; unsigned rep_score; unsigned skip_len; + u32 litrunlen = 0; lzx_reset_symbol_frequencies(c); @@ -1840,29 +2019,29 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) /* Find the longest match at the current position. */ - cur_len = hc_matchfinder_longest_match(&c->hc_mf, - in_begin, - in_next, - 2, - max_len, - nice_len, - c->max_search_depth, - &cur_offset); + cur_len = CALL_HC_MF(is_16_bit, c, hc_matchfinder_longest_match, + in_begin, + in_next - in_begin, + 2, + max_len, + nice_len, + c->max_search_depth, + next_hashes, + &cur_offset); if (cur_len < 3 || (cur_len == 3 && cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT && - cur_offset != lzx_lru_queue_R0(queue) && - cur_offset != lzx_lru_queue_R1(queue) && - cur_offset != lzx_lru_queue_R2(queue))) + cur_offset != recent_offsets[0] && + cur_offset != recent_offsets[1] && + cur_offset != recent_offsets[2])) { /* There was no match found, or the only match found * was a distant length 3 match. Output a literal. */ - lzx_declare_literal(c, *in_next++, - &next_chosen_item); + lzx_record_literal(c, *in_next++, &litrunlen); continue; } - if (cur_offset == lzx_lru_queue_R0(queue)) { + if (cur_offset == recent_offsets[0]) { in_next++; cur_offset_data = 0; skip_len = cur_len - 1; @@ -1875,7 +2054,7 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) /* Consider a repeat offset match */ rep_max_len = lzx_find_longest_repeat_offset_match(in_next, in_end - in_next, - queue, + recent_offsets, &rep_max_idx); in_next++; @@ -1906,14 +2085,15 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) nice_len = min(max_len, nice_len); } - next_len = hc_matchfinder_longest_match(&c->hc_mf, - in_begin, - in_next, - cur_len - 2, - max_len, - nice_len, - c->max_search_depth / 2, - &next_offset); + next_len = CALL_HC_MF(is_16_bit, c, hc_matchfinder_longest_match, + in_begin, + in_next - in_begin, + cur_len - 2, + max_len, + nice_len, + c->max_search_depth / 2, + next_hashes, + &next_offset); if (next_len <= cur_len - 2) { in_next++; @@ -1926,7 +2106,7 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) rep_max_len = lzx_find_longest_repeat_offset_match(in_next, in_end - in_next, - queue, + recent_offsets, &rep_max_idx); in_next++; @@ -1938,8 +2118,8 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) if (rep_score > cur_score) { /* The next match is better, and it's a * repeat offset match. */ - lzx_declare_literal(c, *(in_next - 2), - &next_chosen_item); + lzx_record_literal(c, *(in_next - 2), + &litrunlen); cur_len = rep_max_len; cur_offset_data = rep_max_idx; skip_len = cur_len - 1; @@ -1949,8 +2129,8 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) if (next_score > cur_score) { /* The next match is better, and it's an * explicit offset match. */ - lzx_declare_literal(c, *(in_next - 2), - &next_chosen_item); + lzx_record_literal(c, *(in_next - 2), + &litrunlen); cur_len = next_len; cur_offset_data = next_offset_data; cur_score = next_score; @@ -1962,42 +2142,59 @@ lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os) skip_len = cur_len - 2; choose_cur_match: - if (cur_offset_data < LZX_NUM_RECENT_OFFSETS) { - lzx_declare_repeat_offset_match(c, cur_len, - cur_offset_data, - &next_chosen_item); - queue = lzx_lru_queue_swap(queue, cur_offset_data); - } else { - lzx_declare_explicit_offset_match(c, cur_len, - cur_offset_data - LZX_OFFSET_ADJUSTMENT, - &next_chosen_item); - queue = lzx_lru_queue_push(queue, cur_offset_data - LZX_OFFSET_ADJUSTMENT); - } - - hc_matchfinder_skip_positions(&c->hc_mf, - in_begin, - in_next, - in_end, - skip_len); - in_next += skip_len; + lzx_record_match(c, cur_len, cur_offset_data, + recent_offsets, is_16_bit, + &litrunlen, &next_seq); + in_next = CALL_HC_MF(is_16_bit, c, hc_matchfinder_skip_positions, + in_begin, + in_next - in_begin, + in_end - in_begin, + skip_len, + next_hashes); } while (in_next < in_block_end); - lzx_finish_block(c, os, in_next - in_block_begin, - next_chosen_item - c->chosen_items); + lzx_finish_sequence(next_seq, litrunlen); + + lzx_finish_block(c, os, in_block_begin, in_next - in_block_begin, 0); + } while (in_next != in_end); } static void -lzx_init_offset_slot_fast(struct lzx_compressor *c) +lzx_compress_lazy_16(struct lzx_compressor *c, struct lzx_output_bitstream *os) { - u8 slot = 0; + lzx_compress_lazy(c, os, true); +} - for (u32 offset = 0; offset < LZX_NUM_FAST_OFFSETS; offset++) { +static void +lzx_compress_lazy_32(struct lzx_compressor *c, struct lzx_output_bitstream *os) +{ + lzx_compress_lazy(c, os, false); +} - while (offset + LZX_OFFSET_ADJUSTMENT >= lzx_offset_slot_base[slot + 1]) +/* Generate the acceleration tables for offset slots. */ +static void +lzx_init_offset_slot_tabs(struct lzx_compressor *c) +{ + u32 adjusted_offset = 0; + unsigned slot = 0; + + /* slots [0, 29] */ + for (; adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1); + adjusted_offset++) + { + if (adjusted_offset >= lzx_offset_slot_base[slot + 1]) slot++; + c->offset_slot_tab_1[adjusted_offset] = slot; + } - c->offset_slot_fast[offset] = slot; + /* slots [30, 49] */ + for (; adjusted_offset < LZX_MAX_WINDOW_SIZE; + adjusted_offset += (u32)1 << 14) + { + if (adjusted_offset >= lzx_offset_slot_base[slot + 1]) + slot++; + c->offset_slot_tab_2[adjusted_offset >> 14] = slot; } } @@ -2005,11 +2202,19 @@ static size_t lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level) { if (compression_level <= LZX_MAX_FAST_LEVEL) { - return offsetof(struct lzx_compressor, hc_mf) + - hc_matchfinder_size(max_bufsize); + if (lzx_is_16_bit(max_bufsize)) + return offsetof(struct lzx_compressor, hc_mf_16) + + hc_matchfinder_size_16(max_bufsize); + else + return offsetof(struct lzx_compressor, hc_mf_32) + + hc_matchfinder_size_32(max_bufsize); } else { - return offsetof(struct lzx_compressor, bt_mf) + - bt_matchfinder_size(max_bufsize); + if (lzx_is_16_bit(max_bufsize)) + return offsetof(struct lzx_compressor, bt_mf_16) + + bt_matchfinder_size_16(max_bufsize); + else + return offsetof(struct lzx_compressor, bt_mf_32) + + bt_matchfinder_size_32(max_bufsize); } } @@ -2039,9 +2244,7 @@ lzx_create_compressor(size_t max_bufsize, unsigned compression_level, if (window_order == 0) return WIMLIB_ERR_INVALID_PARAM; - c = ALIGNED_MALLOC(lzx_get_compressor_size(max_bufsize, - compression_level), - MATCHFINDER_ALIGNMENT); + c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level)); if (!c) goto oom0; @@ -2060,9 +2263,12 @@ lzx_create_compressor(size_t max_bufsize, unsigned compression_level, /* Fast compression: Use lazy parsing. */ - c->impl = lzx_compress_lazy; - c->max_search_depth = (36 * compression_level) / 20; - c->nice_match_length = (72 * compression_level) / 20; + if (lzx_is_16_bit(max_bufsize)) + c->impl = lzx_compress_lazy_16; + else + c->impl = lzx_compress_lazy_32; + c->max_search_depth = (60 * compression_level) / 20; + c->nice_match_length = (80 * compression_level) / 20; /* lzx_compress_lazy() needs max_search_depth >= 2 because it * halves the max_search_depth when attempting a lazy match, and @@ -2073,12 +2279,15 @@ lzx_create_compressor(size_t max_bufsize, unsigned compression_level, /* Normal / high compression: Use near-optimal parsing. */ - c->impl = lzx_compress_near_optimal; + if (lzx_is_16_bit(max_bufsize)) + c->impl = lzx_compress_near_optimal_16; + else + c->impl = lzx_compress_near_optimal_32; /* Scale nice_match_length and max_search_depth with the * compression level. */ c->max_search_depth = (24 * compression_level) / 50; - c->nice_match_length = (32 * compression_level) / 50; + c->nice_match_length = (48 * compression_level) / 50; /* Set a number of optimization passes appropriate for the * compression level. */ @@ -2111,19 +2320,19 @@ lzx_create_compressor(size_t max_bufsize, unsigned compression_level, if (c->nice_match_length > LZX_MAX_MATCH_LEN) c->nice_match_length = LZX_MAX_MATCH_LEN; - lzx_init_offset_slot_fast(c); + lzx_init_offset_slot_tabs(c); *c_ret = c; return 0; oom1: - ALIGNED_FREE(c); + FREE(c); oom0: return WIMLIB_ERR_NOMEM; } static size_t -lzx_compress(const void *in, size_t in_nbytes, - void *out, size_t out_nbytes_avail, void *_c) +lzx_compress(const void *restrict in, size_t in_nbytes, + void *restrict out, size_t out_nbytes_avail, void *restrict _c) { struct lzx_compressor *c = _c; struct lzx_output_bitstream os; @@ -2139,7 +2348,7 @@ lzx_compress(const void *in, size_t in_nbytes, else memcpy(c->in_buffer, in, in_nbytes); c->in_nbytes = in_nbytes; - lzx_do_e8_preprocessing(c->in_buffer, in_nbytes); + lzx_preprocess(c->in_buffer, in_nbytes); /* Initially, the previous Huffman codeword lengths are all zeroes. */ c->codes_index = 0; @@ -2154,7 +2363,7 @@ lzx_compress(const void *in, size_t in_nbytes, /* Flush the output bitstream and return the compressed size or 0. */ result = lzx_flush_output(&os); if (!result && c->destructive) - lzx_undo_e8_preprocessing(c->in_buffer, c->in_nbytes); + lzx_postprocess(c->in_buffer, c->in_nbytes); return result; } @@ -2165,7 +2374,7 @@ lzx_free_compressor(void *_c) if (!c->destructive) FREE(c->in_buffer); - ALIGNED_FREE(c); + FREE(c); } const struct compressor_ops lzx_compressor_ops = {