From bba6158f9e7f07d3f3d67aa300b14140253a0e36 Mon Sep 17 00:00:00 2001 From: Eric Biggers Date: Sun, 3 Jul 2016 14:13:28 -0500 Subject: [PATCH] Use subtables instead of binary trees for decoding long Huffman codewords --- include/wimlib/compiler.h | 4 + include/wimlib/decompress_common.h | 210 +++++++++--- src/decompress_common.c | 514 ++++++++++++----------------- src/lzms_decompress.c | 67 ++-- src/lzx_decompress.c | 26 +- src/xpress_decompress.c | 8 +- 6 files changed, 402 insertions(+), 427 deletions(-) diff --git a/include/wimlib/compiler.h b/include/wimlib/compiler.h index a43bb769..3832ef68 100644 --- a/include/wimlib/compiler.h +++ b/include/wimlib/compiler.h @@ -192,6 +192,10 @@ # define STATIC_ASSERT(expr) ((void)sizeof(char[1 - 2 * !(expr)])) #endif +/* STATIC_ASSERT_ZERO() - verify the truth of an expression at compilation time + * and also produce a result of 0 to be used in constant expressions */ +#define STATIC_ASSERT_ZERO(expr) ((int)sizeof(char[0 - 2 * !(expr)])) + #define CONCAT_IMPL(s1, s2) s1##s2 /* CONCAT() - concatenate two tokens at preprocessing time. */ diff --git a/include/wimlib/decompress_common.h b/include/wimlib/decompress_common.h index c3016475..20790464 100644 --- a/include/wimlib/decompress_common.h +++ b/include/wimlib/decompress_common.h @@ -198,68 +198,165 @@ bitstream_align(struct input_bitstream *is) * about dealing with the unaligned case. */ #define DECODE_TABLE_ALIGNMENT 16 -/* Maximum supported symbol count for make_huffman_decode_table(). - * - * Reason: In direct mapping entries, we store the symbol in 11 bits. */ -#define DECODE_TABLE_MAX_SYMBOLS 2048 +#define DECODE_TABLE_SYMBOL_SHIFT 4 +#define DECODE_TABLE_LENGTH_MASK DECODE_TABLE_MAX_LENGTH -/* Maximum supported table bits for make_huffman_decode_table(). - * - * Reason: In internal binary tree nodes, offsets are encoded in 14 bits. - * But the real limit is 13, because we allocate entries past the end of - * the direct lookup part of the table for binary tree nodes. (Note: if - * needed this limit could be removed by encoding the offsets relative to - * &decode_table[1 << table_bits].) */ -#define DECODE_TABLE_MAX_TABLE_BITS 13 - -/* Maximum supported codeword length for make_huffman_decode_table(). +#define DECODE_TABLE_MAX_NUM_SYMS ((1 << (16 - DECODE_TABLE_SYMBOL_SHIFT)) - 1) +#define DECODE_TABLE_MAX_LENGTH ((1 << DECODE_TABLE_SYMBOL_SHIFT) - 1) + +/* + * Reads and returns the next Huffman-encoded symbol from a bitstream. * - * Reason: In direct mapping entries, we encode the codeword length in 5 - * bits, and the top 2 bits can't both be set because that has special - * meaning. */ -#define DECODE_TABLE_MAX_CODEWORD_LEN 23 - -/* Reads and returns the next Huffman-encoded symbol from a bitstream. If the - * input data is exhausted, the Huffman symbol is decoded as if the missing bits - * are all zeroes. + * If the input data is exhausted, the Huffman symbol is decoded as if the + * missing bits are all zeroes. * * XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in - * lzms_decompress.c. */ + * lzms_decompress.c. + */ static inline unsigned -read_huffsym(struct input_bitstream *istream, const u16 decode_table[], +read_huffsym(struct input_bitstream *is, const u16 decode_table[], unsigned table_bits, unsigned max_codeword_len) { unsigned entry; - unsigned key_bits; - - bitstream_ensure_bits(istream, max_codeword_len); - - /* Index the decode table by the next table_bits bits of the input. */ - key_bits = bitstream_peek_bits(istream, table_bits); - entry = decode_table[key_bits]; - if (likely(entry < 0xC000)) { - /* Fast case: The decode table directly provided the - * symbol and codeword length. The low 11 bits are the - * symbol, and the high 5 bits are the codeword length. */ - bitstream_remove_bits(istream, entry >> 11); - return entry & 0x7FF; - } else { - /* Slow case: The codeword for the symbol is longer than - * table_bits, so the symbol does not have an entry - * directly in the first (1 << table_bits) entries of the - * decode table. Traverse the appropriate binary tree - * bit-by-bit to decode the symbol. */ - bitstream_remove_bits(istream, table_bits); - do { - key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1); - } while ((entry = decode_table[key_bits]) >= 0xC000); - return entry; + unsigned sym; + unsigned len; + + /* If the bitbuffer contains fewer bits than might be required by a + * single codeword, then refill it. */ + bitstream_ensure_bits(is, max_codeword_len); + + /* Index the root table by the next 'table_bits' bits of input. */ + entry = decode_table[bitstream_peek_bits(is, table_bits)]; + + /* Extract the symbol and length from the entry. */ + sym = entry >> DECODE_TABLE_SYMBOL_SHIFT; + len = entry & DECODE_TABLE_LENGTH_MASK; + + /* If the root table is indexed by the full 'max_codeword_len' bits, + * then there cannot be any subtables. This will be known at compile + * time. Otherwise, we must check whether the decoded symbol is really + * a subtable pointer. If so, we must discard the bits with which the + * root table was indexed, then index the subtable by the next 'len' + * bits of input to get the real entry. */ + if (max_codeword_len > table_bits && + entry >= (1U << (table_bits + DECODE_TABLE_SYMBOL_SHIFT))) + { + /* Subtable required */ + bitstream_remove_bits(is, table_bits); + entry = decode_table[sym + bitstream_peek_bits(is, len)]; + sym = entry >> DECODE_TABLE_SYMBOL_SHIFT;; + len = entry & DECODE_TABLE_LENGTH_MASK; } + + /* Discard the bits (or the remaining bits, if a subtable was required) + * of the codeword. */ + bitstream_remove_bits(is, len); + + /* Return the decoded symbol. */ + return sym; } +/* + * The ENOUGH() macro returns the maximum number of decode table entries, + * including all subtable entries, that may be required for decoding a given + * Huffman code. This depends on three parameters: + * + * num_syms: the maximum number of symbols in the code + * table_bits: the number of bits with which the root table will be indexed + * max_codeword_len: the maximum allowed codeword length + * + * Given these parameters, the utility program 'enough' from zlib, when run as + * './enough num_syms table_bits max_codeword_len', will compute the maximum + * number of entries required. This has already been done for the combinations + * we need (or may need) and incorporated into the macro below so that the + * mapping can be done at compilation time. If an unknown combination is used, + * then a compilation error will result. To fix this, use 'enough' to find the + * missing value and add it below. + */ +#define ENOUGH(num_syms, table_bits, max_codeword_len) ( \ + ((num_syms) == 8 && (table_bits) == 7 && (max_codeword_len) == 15) ? 128 : \ + ((num_syms) == 8 && (table_bits) == 5 && (max_codeword_len) == 7) ? 36 : \ + ((num_syms) == 8 && (table_bits) == 6 && (max_codeword_len) == 7) ? 66 : \ + ((num_syms) == 8 && (table_bits) == 7 && (max_codeword_len) == 7) ? 128 : \ + ((num_syms) == 20 && (table_bits) == 5 && (max_codeword_len) == 15) ? 1062 : \ + ((num_syms) == 20 && (table_bits) == 6 && (max_codeword_len) == 15) ? 582 : \ + ((num_syms) == 20 && (table_bits) == 7 && (max_codeword_len) == 15) ? 390 : \ + ((num_syms) == 54 && (table_bits) == 9 && (max_codeword_len) == 15) ? 618 : \ + ((num_syms) == 54 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1098 : \ + ((num_syms) == 249 && (table_bits) == 9 && (max_codeword_len) == 16) ? 878 : \ + ((num_syms) == 249 && (table_bits) == 10 && (max_codeword_len) == 16) ? 1326 : \ + ((num_syms) == 249 && (table_bits) == 11 && (max_codeword_len) == 16) ? 2318 : \ + ((num_syms) == 256 && (table_bits) == 9 && (max_codeword_len) == 15) ? 822 : \ + ((num_syms) == 256 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1302 : \ + ((num_syms) == 256 && (table_bits) == 11 && (max_codeword_len) == 15) ? 2310 : \ + ((num_syms) == 512 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1558 : \ + ((num_syms) == 512 && (table_bits) == 11 && (max_codeword_len) == 15) ? 2566 : \ + ((num_syms) == 512 && (table_bits) == 12 && (max_codeword_len) == 15) ? 4606 : \ + ((num_syms) == 656 && (table_bits) == 10 && (max_codeword_len) == 16) ? 1734 : \ + ((num_syms) == 656 && (table_bits) == 11 && (max_codeword_len) == 16) ? 2726 : \ + ((num_syms) == 656 && (table_bits) == 12 && (max_codeword_len) == 16) ? 4758 : \ + ((num_syms) == 799 && (table_bits) == 9 && (max_codeword_len) == 15) ? 1366 : \ + ((num_syms) == 799 && (table_bits) == 10 && (max_codeword_len) == 15) ? 1846 : \ + ((num_syms) == 799 && (table_bits) == 11 && (max_codeword_len) == 15) ? 2854 : \ + -1) + +/* Wrapper around ENOUGH() that does additional compile-time validation. */ +#define DECODE_TABLE_LENGTH(num_syms, table_bits, max_codeword_len) ( \ + \ + /* Every possible symbol value must fit into the symbol portion \ + * of a decode table entry. */ \ + STATIC_ASSERT_ZERO((num_syms) <= DECODE_TABLE_MAX_NUM_SYMS) + \ + \ + /* There cannot be more symbols than possible codewords. */ \ + STATIC_ASSERT_ZERO((num_syms) <= 1U << (max_codeword_len)) + \ + \ + /* It doesn't make sense to use a table_bits more than the \ + * maximum codeword length. */ \ + STATIC_ASSERT_ZERO((max_codeword_len) >= (table_bits)) + \ + \ + /* The maximum length in the root table must fit into the \ + * length portion of a decode table entry. */ \ + STATIC_ASSERT_ZERO((table_bits) <= DECODE_TABLE_MAX_LENGTH) + \ + \ + /* The maximum length in a subtable must fit into the length + * portion of a decode table entry. */ \ + STATIC_ASSERT_ZERO((max_codeword_len) - (table_bits) <= \ + DECODE_TABLE_MAX_LENGTH) + \ + \ + /* The needed 'enough' value must have been defined. */ \ + STATIC_ASSERT_ZERO(ENOUGH((num_syms), (table_bits), \ + (max_codeword_len)) >= 0) + \ + \ + /* The maximum subtable index must fit in the field which would \ + * normally hold a symbol value. */ \ + STATIC_ASSERT_ZERO(ENOUGH((num_syms), (table_bits), \ + (max_codeword_len)) <= \ + DECODE_TABLE_MAX_NUM_SYMS) + \ + \ + /* The minimum subtable index must be greater than the greatest \ + * possible symbol value. */ \ + STATIC_ASSERT_ZERO((1U << table_bits) >= num_syms) + \ + \ + ENOUGH(num_syms, table_bits, max_codeword_len) \ +) + +/* + * Declare the decode table for a Huffman code, given several compile-time + * constants that describe that code (see ENOUGH() for details). + * + * Decode tables must be aligned to a DECODE_TABLE_ALIGNMENT-boundary. This + * implies that if a decode table is nested a dynamically allocated structure, + * then the outer structure must be allocated on a DECODE_TABLE_ALIGNMENT-byte + * boundary as well. + */ +#define DECODE_TABLE(name, num_syms, table_bits, max_codeword_len) \ + u16 name[DECODE_TABLE_LENGTH((num_syms), (table_bits), \ + (max_codeword_len))] \ + _aligned_attribute(DECODE_TABLE_ALIGNMENT) + extern int make_huffman_decode_table(u16 decode_table[], unsigned num_syms, - unsigned num_bits, const u8 lens[], + unsigned table_bits, const u8 lens[], unsigned max_codeword_len); static inline void @@ -269,19 +366,22 @@ copy_word_unaligned(const void *src, void *dst) } static inline machine_word_t -repeat_byte(u8 b) +repeat_u16(u16 b) { - machine_word_t v; + machine_word_t v = b; STATIC_ASSERT(WORDBITS == 32 || WORDBITS == 64); - - v = b; - v |= v << 8; v |= v << 16; v |= v << ((WORDBITS == 64) ? 32 : 0); return v; } +static inline machine_word_t +repeat_u8(u8 b) +{ + return repeat_u16(((u16)b << 8) | b); +} + /* * Copy an LZ77 match at (dst - offset) to dst. * @@ -340,7 +440,7 @@ lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length) * encoding of the previous byte. This case is common * if the data contains many repeated bytes. */ - machine_word_t v = repeat_byte(*(dst - 1)); + machine_word_t v = repeat_u8(*(dst - 1)); do { store_word_unaligned(v, dst); src += WORDBYTES; diff --git a/src/decompress_common.c b/src/decompress_common.c index 973f467c..a1a1e6a6 100644 --- a/src/decompress_common.c +++ b/src/decompress_common.c @@ -5,7 +5,7 @@ * * The following copying information applies to this specific source code file: * - * Written in 2012-2015 by Eric Biggers + * Written in 2012-2016 by Eric Biggers * * To the extent possible under law, the author(s) have dedicated all copyright * and related and neighboring rights to this software to the public domain @@ -24,212 +24,152 @@ # include "config.h" #endif -#include "wimlib/decompress_common.h" - #include -#define USE_WORD_FILL - -#ifdef __GNUC__ -# ifdef __SSE2__ -# undef USE_WORD_FILL -# define USE_SSE2_FILL -# include -# endif +#ifdef __SSE2__ +# include #endif -/* Construct a direct mapping entry in the lookup table. */ -#define MAKE_DIRECT_ENTRY(symbol, length) ((symbol) | ((length) << 11)) +#include "wimlib/decompress_common.h" + +#define MAKE_ENTRY(sym, len) (((sym) << DECODE_TABLE_SYMBOL_SHIFT) | (len)) /* * make_huffman_decode_table() - * - * Build a decoding table for a canonical prefix code, or "Huffman code". - * - * This takes as input the length of the codeword for each symbol in the - * alphabet and produces as output a table that can be used for fast - * decoding of prefix-encoded symbols using read_huffsym(). + * Build a decoding table for a canonical prefix code, or "Huffman code". This + * takes as input the length of the codeword for each symbol in the code and + * produces as output a table for fast symbol decoding with read_huffsym(). * - * Strictly speaking, a canonical prefix code might not be a Huffman - * code. But this algorithm will work either way; and in fact, since - * Huffman codes are defined in terms of symbol frequencies, there is no - * way for the decompressor to know whether the code is a true Huffman - * code or not until all symbols have been decoded. + * Because the code is assumed to be "canonical", it can be reconstructed + * directly from the codeword lengths. A prefix code is canonical if and only + * if a longer codeword never lexicographically precedes a shorter codeword, and + * the lexicographic ordering of codewords of the same length is the same as the + * lexicographic ordering of the corresponding symbols. Consequently, we can + * sort the symbols primarily by codeword length and secondarily by symbol + * value, then reconstruct the code by generating codewords lexicographically in + * that order. * - * Because the prefix code is assumed to be "canonical", it can be - * reconstructed directly from the codeword lengths. A prefix code is - * canonical if and only if a longer codeword never lexicographically - * precedes a shorter codeword, and the lexicographic ordering of - * codewords of the same length is the same as the lexicographic ordering - * of the corresponding symbols. Consequently, we can sort the symbols - * primarily by codeword length and secondarily by symbol value, then - * reconstruct the prefix code by generating codewords lexicographically - * in that order. + * This function does not, however, generate the code explicitly. Instead, it + * directly builds a table for decoding symbols using the code. The basic idea + * is this: given the next 'max_codeword_len' bits in the input, we can look up + * the decoded symbol by indexing a table containing 2**max_codeword_len + * entries. A codeword with length 'max_codeword_len' will have exactly one + * entry in this table, whereas a codeword shorter than 'max_codeword_len' will + * have multiple entries in this table. Precisely, a codeword of length n will + * be represented by 2**(max_codeword_len - n) entries in this table. The + * 0-based index of each such entry will contain the corresponding codeword as a + * prefix when zero-padded on the left to 'max_codeword_len' binary digits. * - * This function does not, however, generate the prefix code explicitly. - * Instead, it directly builds a table for decoding symbols using the - * code. The basic idea is this: given the next 'max_codeword_len' bits - * in the input, we can look up the decoded symbol by indexing a table - * containing 2**max_codeword_len entries. A codeword with length - * 'max_codeword_len' will have exactly one entry in this table, whereas - * a codeword shorter than 'max_codeword_len' will have multiple entries - * in this table. Precisely, a codeword of length n will be represented - * by 2**(max_codeword_len - n) entries in this table. The 0-based index - * of each such entry will contain the corresponding codeword as a prefix - * when zero-padded on the left to 'max_codeword_len' binary digits. + * That's the basic idea, but we implement two optimizations: * - * That's the basic idea, but we implement two optimizations regarding - * the format of the decode table itself: + * - Often the maximum codeword length is too long for it to be efficient to + * build the full decoding table whenever a new code is used. Instead, we can + * build the table using only 2**table_bits entries, where 'table_bits <= + * max_codeword_len'. Then, a lookup of 'table_bits' bits will produce either + * a codeword directly (for codewords not longer than 'table_bits') or the + * index of a subtable which must be indexed with additional bits of input to + * decode the full codeword (for codewords longer than 'table_bits'). * - * - For many compression formats, the maximum codeword length is too - * long for it to be efficient to build the full decoding table - * whenever a new prefix code is used. Instead, we can build the table - * using only 2**table_bits entries, where 'table_bits' is some number - * less than or equal to 'max_codeword_len'. Then, only codewords of - * length 'table_bits' and shorter can be directly looked up. For - * longer codewords, the direct lookup instead produces the root of a - * binary tree. Using this tree, the decoder can do traditional - * bit-by-bit decoding of the remainder of the codeword. Child nodes - * are allocated in extra entries at the end of the table; leaf nodes - * contain symbols. Note that the long-codeword case is, in general, - * not performance critical, since in Huffman codes the most frequently - * used symbols are assigned the shortest codeword lengths. - * - * - When we decode a symbol using a direct lookup of the table, we still - * need to know its length so that the bitstream can be advanced by the - * appropriate number of bits. The simple solution is to simply retain - * the 'lens' array and use the decoded symbol as an index into it. - * However, this requires two separate array accesses in the fast path. - * The optimization is to store the length directly in the decode - * table. We use the bottom 11 bits for the symbol and the top 5 bits - * for the length. In addition, to combine this optimization with the - * previous one, we introduce a special case where the top 2 bits of - * the length are both set if the entry is actually the root of a - * binary tree. + * - When we decode a symbol, we still need to know its codeword length so that + * the bitstream can be advanced by the appropriate number of bits. The + * obvious solution is to simply retain the 'lens' array and use the decoded + * symbol as an index into it. However, this requires two separate array + * accesses in the fast path. The optimization is to store the length + * directly in the decode table, along with the symbol. * * @decode_table: - * The array in which to create the decoding table. This must be - * 16-byte aligned and must have a length of at least - * ((2**table_bits) + 2 * num_syms) entries. This is permitted to - * alias @lens, since all information from @lens is consumed before -* anything is written to @decode_table. + * The array in which to build the decode table. This must have been + * declared by the DECODE_TABLE() macro. This may alias @lens, since all + * @lens are consumed before the decode table is written to. * * @num_syms: - * The number of symbols in the alphabet; also, the length of the - * 'lens' array. Must be less than or equal to - * DECODE_TABLE_MAX_SYMBOLS. + * The number of symbols in the alphabet. * * @table_bits: - * The order of the decode table size, as explained above. Must be - * less than or equal to DECODE_TABLE_MAX_TABLE_BITS. + * The log base 2 of the number of entries in the root table. * * @lens: - * An array of length @num_syms, indexable by symbol, that gives the - * length of the codeword, in bits, for that symbol. The length can - * be 0, which means that the symbol does not have a codeword - * assigned. This is permitted to alias @decode_table, since all - * information from @lens is consumed before anything is written to - * @decode_table. + * An array of length @num_syms, indexable by symbol, that gives the length + * of the codeword, in bits, for that symbol. The length can be 0, which + * means that the symbol does not have a codeword assigned. In addition, + * @lens may alias @decode_table, as noted above. * * @max_codeword_len: - * The longest codeword length allowed in the compression format. - * All entries in 'lens' must be less than or equal to this value. - * This must be less than or equal to DECODE_TABLE_MAX_CODEWORD_LEN. + * The maximum codeword length permitted for this code. All entries in + * 'lens' must be less than or equal to this value. * - * Returns 0 on success, or -1 if the lengths do not form a valid prefix - * code. + * Returns 0 on success, or -1 if the lengths do not form a valid prefix code. */ int -make_huffman_decode_table(u16 decode_table[const], - const unsigned num_syms, - const unsigned table_bits, - const u8 lens[const], - const unsigned max_codeword_len) +make_huffman_decode_table(u16 decode_table[], unsigned num_syms, + unsigned table_bits, const u8 lens[], + unsigned max_codeword_len) { - const unsigned table_num_entries = 1 << table_bits; - unsigned len_counts[max_codeword_len + 1]; + u16 offsets[max_codeword_len + 1]; + u16 len_counts[max_codeword_len + 1]; u16 sorted_syms[num_syms]; - int left; - void *decode_table_ptr; + s32 remainder = 1; + void *entry_ptr = decode_table; + unsigned codeword_len = 1; unsigned sym_idx; - unsigned codeword_len; - unsigned stores_per_loop; - unsigned decode_table_pos; - -#ifdef USE_WORD_FILL - const unsigned entries_per_word = WORDBYTES / sizeof(decode_table[0]); -#endif + unsigned codeword; + unsigned subtable_pos; + unsigned subtable_bits; + unsigned subtable_prefix; -#ifdef USE_SSE2_FILL - const unsigned entries_per_xmm = sizeof(__m128i) / sizeof(decode_table[0]); -#endif - - /* Count how many symbols have each possible codeword length. - * Note that a length of 0 indicates the corresponding symbol is not - * used in the code and therefore does not have a codeword. */ + /* Count how many codewords have each length, including 0. */ for (unsigned len = 0; len <= max_codeword_len; len++) len_counts[len] = 0; for (unsigned sym = 0; sym < num_syms; sym++) len_counts[lens[sym]]++; - /* We can assume all lengths are <= max_codeword_len, but we - * cannot assume they form a valid prefix code. A codeword of - * length n should require a proportion of the codespace equaling - * (1/2)^n. The code is valid if and only if the codespace is - * exactly filled by the lengths, by this measure. */ - left = 1; + /* It is already guaranteed that all lengths are <= max_codeword_len, + * but it cannot be assumed they form a complete prefix code. A + * codeword of length n should require a proportion of the codespace + * equaling (1/2)^n. The code is complete if and only if, by this + * measure, the codespace is exactly filled by the lengths. */ for (unsigned len = 1; len <= max_codeword_len; len++) { - left <<= 1; - left -= len_counts[len]; - if (unlikely(left < 0)) { - /* The lengths overflow the codespace; that is, the code - * is over-subscribed. */ + remainder = (remainder << 1) - len_counts[len]; + /* Do the lengths overflow the codespace? */ + if (unlikely(remainder < 0)) return -1; - } } - if (unlikely(left != 0)) { + if (remainder != 0) { /* The lengths do not fill the codespace; that is, they form an - * incomplete set. */ - if (left == (1 << max_codeword_len)) { - /* The code is completely empty. This is arguably - * invalid, but in fact it is valid in LZX and XPRESS, - * so we must allow it. By definition, no symbols can - * be decoded with an empty code. Consequently, we - * technically don't even need to fill in the decode - * table. However, to avoid accessing uninitialized - * memory if the algorithm nevertheless attempts to - * decode symbols using such a code, we zero out the - * decode table. */ - memset(decode_table, 0, - table_num_entries * sizeof(decode_table[0])); - return 0; - } - return -1; + * incomplete code. This is permitted only if the code is empty + * (contains no symbols). */ + + if (unlikely(remainder != (s32)1 << max_codeword_len)) + return -1; + + /* The code is empty. When processing a well-formed stream, the + * decode table need not be initialized in this case. However, + * we cannot assume the stream is well-formed, so we must + * initialize the decode table anyway. Setting all entries to 0 + * makes this table always produce symbol '0' without consuming + * any bits, which is good enough. */ + memset(decode_table, 0, + (1U << table_bits) * sizeof(decode_table[0])); + return 0; } - /* Sort the symbols primarily by length and secondarily by symbol order. - */ - { - unsigned offsets[max_codeword_len + 1]; + /* Sort the symbols primarily by increasing codeword length and + * secondarily by increasing symbol value. */ - /* Initialize 'offsets' so that offsets[len] for 1 <= len <= - * max_codeword_len is the number of codewords shorter than - * 'len' bits. */ - offsets[1] = 0; - for (unsigned len = 1; len < max_codeword_len; len++) - offsets[len + 1] = offsets[len] + len_counts[len]; + /* Initialize 'offsets' so that 'offsets[len]' is the number of + * codewords shorter than 'len' bits, including length 0. */ + offsets[0] = 0; + for (unsigned len = 0; len < max_codeword_len; len++) + offsets[len + 1] = offsets[len] + len_counts[len]; - /* Use the 'offsets' array to sort the symbols. - * Note that we do not include symbols that are not used in the - * code. Consequently, fewer than 'num_syms' entries in - * 'sorted_syms' may be filled. */ - for (unsigned sym = 0; sym < num_syms; sym++) - if (lens[sym] != 0) - sorted_syms[offsets[lens[sym]]++] = sym; - } + /* Use the 'offsets' array to sort the symbols. */ + for (unsigned sym = 0; sym < num_syms; sym++) + sorted_syms[offsets[lens[sym]]++] = sym; - /* Fill entries for codewords with length <= table_bits + /* + * Fill entries for codewords with length <= table_bits * --- that is, those short enough for a direct mapping. * * The table will start with entries for the shortest codeword(s), which @@ -237,46 +177,38 @@ make_huffman_decode_table(u16 decode_table[const], * codeword will decrease. As an optimization, we may begin filling * entries with SSE2 vector accesses (8 entries/store), then change to * 'machine_word_t' accesses (2 or 4 entries/store), then change to - * 16-bit accesses (1 entry/store). */ - decode_table_ptr = decode_table; - sym_idx = 0; - codeword_len = 1; -#ifdef USE_SSE2_FILL - /* Fill the entries one 128-bit vector at a time. - * This is 8 entries per store. */ - stores_per_loop = (1 << (table_bits - codeword_len)) / entries_per_xmm; - for (; stores_per_loop != 0; codeword_len++, stores_per_loop >>= 1) { + * 16-bit accesses (1 entry/store). + */ + sym_idx = offsets[0]; +#ifdef __SSE2__ + /* Fill entries one 128-bit vector (8 entries) at a time. */ + for (unsigned stores_per_loop = (1U << (table_bits - codeword_len)) / + (sizeof(__m128i) / sizeof(decode_table[0])); + stores_per_loop != 0; codeword_len++, stores_per_loop >>= 1) + { unsigned end_sym_idx = sym_idx + len_counts[codeword_len]; for (; sym_idx < end_sym_idx; sym_idx++) { - /* Note: unlike in the machine_word_t version below, the - * __m128i type already has __attribute__((may_alias)), - * so using it to access the decode table, which is an - * array of unsigned shorts, will not violate strict + /* Note: unlike in the "word" version below, the __m128i + * type already has __attribute__((may_alias)), so using + * it to access an array of u16 will not violate strict * aliasing. */ - u16 entry; - __m128i v; - __m128i *p; - unsigned n; - - entry = MAKE_DIRECT_ENTRY(sorted_syms[sym_idx], codeword_len); - - v = _mm_set1_epi16(entry); - p = (__m128i*)decode_table_ptr; - n = stores_per_loop; + __m128i v = _mm_set1_epi16( + MAKE_ENTRY(sorted_syms[sym_idx], codeword_len)); + unsigned n = stores_per_loop; do { - *p++ = v; + *(__m128i *)entry_ptr = v; + entry_ptr += sizeof(v); } while (--n); - decode_table_ptr = p; } } -#endif /* USE_SSE2_FILL */ +#endif /* __SSE2__ */ -#ifdef USE_WORD_FILL - /* Fill the entries one machine word at a time. - * On 32-bit systems this is 2 entries per store, while on 64-bit - * systems this is 4 entries per store. */ - stores_per_loop = (1 << (table_bits - codeword_len)) / entries_per_word; - for (; stores_per_loop != 0; codeword_len++, stores_per_loop >>= 1) { +#ifdef __GNUC__ + /* Fill entries one word (2 or 4 entries) at a time. */ + for (unsigned stores_per_loop = (1U << (table_bits - codeword_len)) / + (WORDBYTES / sizeof(decode_table[0])); + stores_per_loop != 0; codeword_len++, stores_per_loop >>= 1) + { unsigned end_sym_idx = sym_idx + len_counts[codeword_len]; for (; sym_idx < end_sym_idx; sym_idx++) { @@ -285,140 +217,100 @@ make_huffman_decode_table(u16 decode_table[const], * the code with -fno-strict-aliasing to guarantee * correctness. To work around this problem, use the * gcc 'may_alias' extension. */ - typedef machine_word_t _may_alias_attribute aliased_word_t; - - machine_word_t v; - aliased_word_t *p; - unsigned n; - - STATIC_ASSERT(WORDBITS == 32 || WORDBITS == 64); - - v = MAKE_DIRECT_ENTRY(sorted_syms[sym_idx], codeword_len); - v |= v << 16; - v |= v << (WORDBITS == 64 ? 32 : 0); - - p = (aliased_word_t *)decode_table_ptr; - n = stores_per_loop; + typedef machine_word_t + __attribute__((may_alias)) aliased_word_t; + aliased_word_t v = repeat_u16( + MAKE_ENTRY(sorted_syms[sym_idx], codeword_len)); + unsigned n = stores_per_loop; do { - *p++ = v; + *(aliased_word_t *)entry_ptr = v; + entry_ptr += sizeof(v); } while (--n); - decode_table_ptr = p; } } -#endif /* USE_WORD_FILL */ +#endif /* __GNUC__ */ - /* Fill the entries one 16-bit integer at a time. */ - stores_per_loop = (1 << (table_bits - codeword_len)); - for (; stores_per_loop != 0; codeword_len++, stores_per_loop >>= 1) { + /* Fill entries one at a time. */ + for (unsigned stores_per_loop = (1U << (table_bits - codeword_len)); + stores_per_loop != 0; codeword_len++, stores_per_loop >>= 1) + { unsigned end_sym_idx = sym_idx + len_counts[codeword_len]; for (; sym_idx < end_sym_idx; sym_idx++) { - u16 entry; - u16 *p; - unsigned n; - - entry = MAKE_DIRECT_ENTRY(sorted_syms[sym_idx], codeword_len); - - p = (u16*)decode_table_ptr; - n = stores_per_loop; - + u16 v = MAKE_ENTRY(sorted_syms[sym_idx], codeword_len); + unsigned n = stores_per_loop; do { - *p++ = entry; + *(u16 *)entry_ptr = v; + entry_ptr += sizeof(v); } while (--n); - - decode_table_ptr = p; } } - /* If we've filled in the entire table, we are done. Otherwise, - * there are codewords longer than table_bits for which we must - * generate binary trees. */ - - decode_table_pos = (u16*)decode_table_ptr - decode_table; - if (decode_table_pos != table_num_entries) { - unsigned j; - unsigned next_free_tree_slot; - unsigned cur_codeword; - - /* First, zero out the remaining entries. This is - * necessary so that these entries appear as - * "unallocated" in the next part. Each of these entries - * will eventually be filled with the representation of - * the root node of a binary tree. */ - j = decode_table_pos; - do { - decode_table[j] = 0; - } while (++j != table_num_entries); - - /* We allocate child nodes starting at the end of the - * direct lookup table. Note that there should be - * 2*num_syms extra entries for this purpose, although - * fewer than this may actually be needed. */ - next_free_tree_slot = table_num_entries; - - /* Iterate through each codeword with length greater than - * 'table_bits', primarily in order of codeword length - * and secondarily in order of symbol. */ - for (cur_codeword = decode_table_pos << 1; - codeword_len <= max_codeword_len; - codeword_len++, cur_codeword <<= 1) - { - unsigned end_sym_idx = sym_idx + len_counts[codeword_len]; - for (; sym_idx < end_sym_idx; sym_idx++, cur_codeword++) - { - /* 'sym' is the symbol represented by the - * codeword. */ - unsigned sym = sorted_syms[sym_idx]; - - unsigned extra_bits = codeword_len - table_bits; + /* If all symbols were processed, then no subtables are required. */ + if (sym_idx == num_syms) + return 0; + + /* At least one subtable is required. Process the remaining symbols. */ + codeword = ((u16 *)entry_ptr - decode_table) << 1; + subtable_pos = 1U << table_bits; + subtable_bits = table_bits; + subtable_prefix = -1; + do { + while (len_counts[codeword_len] == 0) { + codeword_len++; + codeword <<= 1; + } - unsigned node_idx = cur_codeword >> extra_bits; + unsigned prefix = codeword >> (codeword_len - table_bits); + + /* Start a new subtable if the first 'table_bits' bits of the + * codeword don't match the prefix for the previous subtable, or + * if this will be the first subtable. */ + if (prefix != subtable_prefix) { + + subtable_prefix = prefix; + + /* Calculate the subtable length. If the codeword + * length exceeds 'table_bits' by n, the subtable needs + * at least 2**n entries. But it may need more; if + * there are fewer than 2**n codewords of length + * 'table_bits + n' remaining, then n will need to be + * incremented to bring in longer codewords until the + * subtable can be filled completely. Note that it + * always will, eventually, be possible to fill the + * subtable, since it was previously verified that the + * code is complete. */ + subtable_bits = codeword_len - table_bits; + remainder = (s32)1 << subtable_bits; + for (;;) { + remainder -= len_counts[table_bits + + subtable_bits]; + if (remainder <= 0) + break; + subtable_bits++; + remainder <<= 1; + } - /* Go through each bit of the current codeword - * beyond the prefix of length @table_bits and - * walk the appropriate binary tree, allocating - * any slots that have not yet been allocated. - * - * Note that the 'pointer' entry to the binary - * tree, which is stored in the direct lookup - * portion of the table, is represented - * identically to other internal (non-leaf) - * nodes of the binary tree; it can be thought - * of as simply the root of the tree. The - * representation of these internal nodes is - * simply the index of the left child combined - * with the special bits 0xC000 to distinguish - * the entry from direct mapping and leaf node - * entries. */ - do { + /* Create the entry that points from the root table to + * the subtable. This entry contains the index of the + * start of the subtable and the number of bits with + * which the subtable is indexed (the log base 2 of the + * number of entries it contains). */ + decode_table[subtable_prefix] = + MAKE_ENTRY(subtable_pos, subtable_bits); + } - /* At least one bit remains in the - * codeword, but the current node is an - * unallocated leaf. Change it to an - * internal node. */ - if (decode_table[node_idx] == 0) { - decode_table[node_idx] = - next_free_tree_slot | 0xC000; - decode_table[next_free_tree_slot++] = 0; - decode_table[next_free_tree_slot++] = 0; - } + u16 entry = MAKE_ENTRY(sorted_syms[sym_idx], + codeword_len - table_bits); + unsigned n = 1U << (subtable_bits - (codeword_len - + table_bits)); + do { + decode_table[subtable_pos++] = entry; + } while (--n); - /* Go to the left child if the next bit - * in the codeword is 0; otherwise go to - * the right child. */ - node_idx = decode_table[node_idx] & 0x3FFF; - --extra_bits; - node_idx += (cur_codeword >> extra_bits) & 1; - } while (extra_bits != 0); + len_counts[codeword_len]--; + codeword++; + } while (++sym_idx < num_syms); - /* We've traversed the tree using the entire - * codeword, and we're now at the entry where - * the actual symbol will be stored. This is - * distinguished from internal nodes by not - * having its high two bits set. */ - decode_table[node_idx] = sym; - } - } - } return 0; } diff --git a/src/lzms_decompress.c b/src/lzms_decompress.c index 1ea0ac6d..ac1ddeeb 100644 --- a/src/lzms_decompress.c +++ b/src/lzms_decompress.c @@ -5,7 +5,7 @@ */ /* - * Copyright (C) 2013, 2014, 2015 Eric Biggers + * Copyright (C) 2013-2016 Eric Biggers * * This file is free software; you can redistribute it and/or modify it under * the terms of the GNU Lesser General Public License as published by the Free @@ -257,10 +257,10 @@ /* The TABLEBITS values can be changed; they only affect decoding speed. */ #define LZMS_LITERAL_TABLEBITS 10 -#define LZMS_LENGTH_TABLEBITS 10 -#define LZMS_LZ_OFFSET_TABLEBITS 10 -#define LZMS_DELTA_OFFSET_TABLEBITS 10 -#define LZMS_DELTA_POWER_TABLEBITS 8 +#define LZMS_LENGTH_TABLEBITS 9 +#define LZMS_LZ_OFFSET_TABLEBITS 11 +#define LZMS_DELTA_OFFSET_TABLEBITS 11 +#define LZMS_DELTA_POWER_TABLEBITS 7 struct lzms_range_decoder { @@ -323,33 +323,28 @@ struct lzms_decompressor { struct lzms_probabilites probs; - u16 literal_decode_table[(1 << LZMS_LITERAL_TABLEBITS) + - (2 * LZMS_NUM_LITERAL_SYMS)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(literal_decode_table, LZMS_NUM_LITERAL_SYMS, + LZMS_LITERAL_TABLEBITS, LZMS_MAX_CODEWORD_LENGTH); u32 literal_freqs[LZMS_NUM_LITERAL_SYMS]; struct lzms_huffman_rebuild_info literal_rebuild_info; - u16 lz_offset_decode_table[(1 << LZMS_LZ_OFFSET_TABLEBITS) + - ( 2 * LZMS_MAX_NUM_OFFSET_SYMS)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(lz_offset_decode_table, LZMS_MAX_NUM_OFFSET_SYMS, + LZMS_LZ_OFFSET_TABLEBITS, LZMS_MAX_CODEWORD_LENGTH); u32 lz_offset_freqs[LZMS_MAX_NUM_OFFSET_SYMS]; struct lzms_huffman_rebuild_info lz_offset_rebuild_info; - u16 length_decode_table[(1 << LZMS_LENGTH_TABLEBITS) + - (2 * LZMS_NUM_LENGTH_SYMS)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(length_decode_table, LZMS_NUM_LENGTH_SYMS, + LZMS_LENGTH_TABLEBITS, LZMS_MAX_CODEWORD_LENGTH); u32 length_freqs[LZMS_NUM_LENGTH_SYMS]; struct lzms_huffman_rebuild_info length_rebuild_info; - u16 delta_offset_decode_table[(1 << LZMS_DELTA_OFFSET_TABLEBITS) + - (2 * LZMS_MAX_NUM_OFFSET_SYMS)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(delta_offset_decode_table, LZMS_MAX_NUM_OFFSET_SYMS, + LZMS_DELTA_OFFSET_TABLEBITS, LZMS_MAX_CODEWORD_LENGTH); u32 delta_offset_freqs[LZMS_MAX_NUM_OFFSET_SYMS]; struct lzms_huffman_rebuild_info delta_offset_rebuild_info; - u16 delta_power_decode_table[(1 << LZMS_DELTA_POWER_TABLEBITS) + - (2 * LZMS_NUM_DELTA_POWER_SYMS)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(delta_power_decode_table, LZMS_NUM_DELTA_POWER_SYMS, + LZMS_DELTA_POWER_TABLEBITS, LZMS_MAX_CODEWORD_LENGTH); u32 delta_power_freqs[LZMS_NUM_DELTA_POWER_SYMS]; struct lzms_huffman_rebuild_info delta_power_rebuild_info; @@ -599,34 +594,26 @@ lzms_decode_huffman_symbol(struct lzms_input_bitstream *is, u16 decode_table[], unsigned table_bits, u32 freqs[], struct lzms_huffman_rebuild_info *rebuild_info) { - unsigned key_bits; unsigned entry; unsigned sym; + unsigned len; lzms_ensure_bits(is, LZMS_MAX_CODEWORD_LENGTH); - /* Index the decode table by the next table_bits bits of the input. */ - key_bits = lzms_peek_bits(is, table_bits); - entry = decode_table[key_bits]; - if (likely(entry < 0xC000)) { - /* Fast case: The decode table directly provided the symbol and - * codeword length. The low 11 bits are the symbol, and the - * high 5 bits are the codeword length. */ - lzms_remove_bits(is, entry >> 11); - sym = entry & 0x7FF; - } else { - /* Slow case: The codeword for the symbol is longer than - * table_bits, so the symbol does not have an entry directly in - * the first (1 << table_bits) entries of the decode table. - * Traverse the appropriate binary tree bit-by-bit in order to - * decode the symbol. */ + entry = decode_table[lzms_peek_bits(is, table_bits)]; + sym = entry >> DECODE_TABLE_SYMBOL_SHIFT; + len = entry & DECODE_TABLE_LENGTH_MASK; + + if (entry >= (1U << (table_bits + DECODE_TABLE_SYMBOL_SHIFT))) { + /* Subtable required */ lzms_remove_bits(is, table_bits); - do { - key_bits = (entry & 0x3FFF) + lzms_pop_bits(is, 1); - } while ((entry = decode_table[key_bits]) >= 0xC000); - sym = entry; + entry = decode_table[sym + lzms_peek_bits(is, len)]; + sym = entry >> DECODE_TABLE_SYMBOL_SHIFT;; + len = entry & DECODE_TABLE_LENGTH_MASK; } + lzms_remove_bits(is, len); + freqs[sym]++; if (--rebuild_info->num_syms_until_rebuild == 0) lzms_rebuild_huffman_code(rebuild_info); diff --git a/src/lzx_decompress.c b/src/lzx_decompress.c index c2a3f6a9..33d42268 100644 --- a/src/lzx_decompress.c +++ b/src/lzx_decompress.c @@ -64,7 +64,7 @@ /* These values are chosen for fast decompression. */ #define LZX_MAINCODE_TABLEBITS 11 -#define LZX_LENCODE_TABLEBITS 10 +#define LZX_LENCODE_TABLEBITS 9 #define LZX_PRECODE_TABLEBITS 6 #define LZX_ALIGNEDCODE_TABLEBITS 7 @@ -73,31 +73,23 @@ /* Reusable heap-allocated memory for LZX decompression */ struct lzx_decompressor { - /* Huffman decoding tables, and arrays that map symbols to codeword - * lengths */ - - u16 maincode_decode_table[(1 << LZX_MAINCODE_TABLEBITS) + - (LZX_MAINCODE_MAX_NUM_SYMBOLS * 2)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(maincode_decode_table, LZX_MAINCODE_MAX_NUM_SYMBOLS, + LZX_MAINCODE_TABLEBITS, LZX_MAX_MAIN_CODEWORD_LEN); u8 maincode_lens[LZX_MAINCODE_MAX_NUM_SYMBOLS + LZX_READ_LENS_MAX_OVERRUN]; - - u16 lencode_decode_table[(1 << LZX_LENCODE_TABLEBITS) + - (LZX_LENCODE_NUM_SYMBOLS * 2)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(lencode_decode_table, LZX_LENCODE_NUM_SYMBOLS, + LZX_LENCODE_TABLEBITS, LZX_MAX_LEN_CODEWORD_LEN); u8 lencode_lens[LZX_LENCODE_NUM_SYMBOLS + LZX_READ_LENS_MAX_OVERRUN]; union { - u16 alignedcode_decode_table[(1 << LZX_ALIGNEDCODE_TABLEBITS) + - (LZX_ALIGNEDCODE_NUM_SYMBOLS * 2)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(alignedcode_decode_table, LZX_ALIGNEDCODE_NUM_SYMBOLS, + LZX_ALIGNEDCODE_TABLEBITS, LZX_MAX_ALIGNED_CODEWORD_LEN); u8 alignedcode_lens[LZX_ALIGNEDCODE_NUM_SYMBOLS]; }; union { - u16 precode_decode_table[(1 << LZX_PRECODE_TABLEBITS) + - (LZX_PRECODE_NUM_SYMBOLS * 2)] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(precode_decode_table, LZX_PRECODE_NUM_SYMBOLS, + LZX_PRECODE_TABLEBITS, LZX_MAX_PRE_CODEWORD_LEN); u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS]; }; diff --git a/src/xpress_decompress.c b/src/xpress_decompress.c index ec5b1831..9fd5ac5e 100644 --- a/src/xpress_decompress.c +++ b/src/xpress_decompress.c @@ -6,7 +6,7 @@ /* * - * Copyright (C) 2012, 2013, 2015 Eric Biggers + * Copyright (C) 2012-2016 Eric Biggers * * This file is free software; you can redistribute it and/or modify it under * the terms of the GNU Lesser General Public License as published by the Free @@ -73,7 +73,7 @@ #include "wimlib/xpress_constants.h" /* This value is chosen for fast decompression. */ -#define XPRESS_TABLEBITS 12 +#define XPRESS_TABLEBITS 11 static int xpress_decompress(const void *restrict compressed_data, size_t compressed_size, @@ -85,8 +85,8 @@ xpress_decompress(const void *restrict compressed_data, size_t compressed_size, u8 *out_next = out_begin; u8 * const out_end = out_begin + uncompressed_size; union { - u16 decode_table[(1 << XPRESS_TABLEBITS) + 2 * XPRESS_NUM_SYMBOLS] - _aligned_attribute(DECODE_TABLE_ALIGNMENT); + DECODE_TABLE(decode_table, XPRESS_NUM_SYMBOLS, + XPRESS_TABLEBITS, XPRESS_MAX_CODEWORD_LEN); u8 lens[XPRESS_NUM_SYMBOLS]; } u; struct input_bitstream is; -- 2.43.0