Compression updates

[wimlib] / src / lzms-compress.c

/*
 * lzms-compress.c
 *
 * A compressor that produces output compatible with the LZMS compression format.
 */

/*
 * Copyright (C) 2013, 2014 Eric Biggers
 *
 * This file is part of wimlib, a library for working with WIM files.
 *
 * wimlib is free software; you can redistribute it and/or modify it under the
 * terms of the GNU General Public License as published by the Free
 * Software Foundation; either version 3 of the License, or (at your option)
 * any later version.
 *
 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
 * A PARTICULAR PURPOSE. See the GNU General Public License for more
 * details.
 *
 * You should have received a copy of the GNU General Public License
 * along with wimlib; if not, see http://www.gnu.org/licenses/.
 */

#ifdef HAVE_CONFIG_H
#  include "config.h"
#endif

#include "wimlib/compress_common.h"
#include "wimlib/compressor_ops.h"
#include "wimlib/endianness.h"
#include "wimlib/error.h"
#include "wimlib/lz_mf.h"
#include "wimlib/lz_repsearch.h"
#include "wimlib/lzms.h"
#include "wimlib/util.h"

#include <string.h>
#include <limits.h>
#include <pthread.h>

/* Stucture used for writing raw bits as a series of 16-bit little endian coding
 * units.  This starts at the *end* of the compressed data buffer and proceeds
 * backwards.  */
struct lzms_output_bitstream {

        /* Bits that haven't yet been written to the output buffer.  */
        u64 bitbuf;

        /* Number of bits currently held in @bitbuf.  */
        unsigned bitcount;

        /* Pointer to one past the next position in the compressed data buffer
         * at which to output a 16-bit coding unit.  */
        le16 *next;

        /* Pointer to the beginning of the output buffer.  (The "end" when
         * writing backwards!)  */
        le16 *begin;
};

/* Stucture used for range encoding (raw version).  This starts at the
 * *beginning* of the compressed data buffer and proceeds forward.  */
struct lzms_range_encoder_raw {

        /* A 33-bit variable that holds the low boundary of the current range.
         * The 33rd bit is needed to catch carries.  */
        u64 low;

        /* Size of the current range.  */
        u32 range;

        /* Next 16-bit coding unit to output.  */
        u16 cache;

        /* Number of 16-bit coding units whose output has been delayed due to
         * possible carrying.  The first such coding unit is @cache; all
         * subsequent such coding units are 0xffff.  */
        u32 cache_size;

        /* Pointer to the beginning of the output buffer.  */
        le16 *begin;

        /* Pointer to the position in the output buffer at which the next coding
         * unit must be written.  */
        le16 *next;

        /* Pointer just past the end of the output buffer.  */
        le16 *end;
};

/* Structure used for range encoding.  This wraps around `struct
 * lzms_range_encoder_raw' to use and maintain probability entries.  */
struct lzms_range_encoder {

        /* Pointer to the raw range encoder, which has no persistent knowledge
         * of probabilities.  Multiple lzms_range_encoder's share the same
         * lzms_range_encoder_raw.  */
        struct lzms_range_encoder_raw *rc;

        /* Bits recently encoded by this range encoder.  This is used as an
         * index into @prob_entries.  */
        u32 state;

        /* Bitmask for @state to prevent its value from exceeding the number of
         * probability entries.  */
        u32 mask;

        /* Probability entries being used for this range encoder.  */
        struct lzms_probability_entry prob_entries[LZMS_MAX_NUM_STATES];
};

/* Structure used for Huffman encoding.  */
struct lzms_huffman_encoder {

        /* Bitstream to write Huffman-encoded symbols and verbatim bits to.
         * Multiple lzms_huffman_encoder's share the same lzms_output_bitstream.
         */
        struct lzms_output_bitstream *os;

        /* Number of symbols that have been written using this code far.  Reset
         * to 0 whenever the code is rebuilt.  */
        u32 num_syms_written;

        /* When @num_syms_written reaches this number, the Huffman code must be
         * rebuilt.  */
        u32 rebuild_freq;

        /* Number of symbols in the represented Huffman code.  */
        unsigned num_syms;

        /* Running totals of symbol frequencies.  These are diluted slightly
         * whenever the code is rebuilt.  */
        u32 sym_freqs[LZMS_MAX_NUM_SYMS];

        /* The length, in bits, of each symbol in the Huffman code.  */
        u8 lens[LZMS_MAX_NUM_SYMS];

        /* The codeword of each symbol in the Huffman code.  */
        u32 codewords[LZMS_MAX_NUM_SYMS];
};

/* Internal compression parameters  */
struct lzms_compressor_params {
        u32 min_match_length;
        u32 nice_match_length;
        u32 max_search_depth;
        u32 optim_array_length;
};

/* State of the LZMS compressor  */
struct lzms_compressor {

        /* Internal compression parameters  */
        struct lzms_compressor_params params;

        /* Data currently being compressed  */
        u8 *cur_window;
        u32 cur_window_size;

        /* Lempel-Ziv match-finder  */
        struct lz_mf *mf;

        /* Temporary space to store found matches  */
        struct lz_match *matches;

        /* Per-position data for near-optimal parsing  */
        struct lzms_mc_pos_data *optimum;
        struct lzms_mc_pos_data *optimum_end;

        /* Raw range encoder which outputs to the beginning of the compressed
         * data buffer, proceeding forwards  */
        struct lzms_range_encoder_raw rc;

        /* Bitstream which outputs to the end of the compressed data buffer,
         * proceeding backwards  */
        struct lzms_output_bitstream os;

        /* Range encoders  */
        struct lzms_range_encoder main_range_encoder;
        struct lzms_range_encoder match_range_encoder;
        struct lzms_range_encoder lz_match_range_encoder;
        struct lzms_range_encoder lz_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
        struct lzms_range_encoder delta_match_range_encoder;
        struct lzms_range_encoder delta_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];

        /* Huffman encoders  */
        struct lzms_huffman_encoder literal_encoder;
        struct lzms_huffman_encoder lz_offset_encoder;
        struct lzms_huffman_encoder length_encoder;
        struct lzms_huffman_encoder delta_power_encoder;
        struct lzms_huffman_encoder delta_offset_encoder;

        /* Used for preprocessing  */
        s32 last_target_usages[65536];

#define LZMS_NUM_FAST_LENGTHS 256
        /* Table: length => length slot for small lengths  */
        u8 length_slot_fast[LZMS_NUM_FAST_LENGTHS];

        /* Table: length => current cost for small match lengths  */
        u32 length_cost_fast[LZMS_NUM_FAST_LENGTHS];

#define LZMS_NUM_FAST_OFFSETS 32768
        /* Table: offset => offset slot for small offsets  */
        u8 offset_slot_fast[LZMS_NUM_FAST_OFFSETS];
};

/*
 * Match chooser position data:
 *
 * An array of these structures is used during the near-optimal match-choosing
 * algorithm.  They correspond to consecutive positions in the window and are
 * used to keep track of the cost to reach each position, and the match/literal
 * choices that need to be chosen to reach that position.
 */
struct lzms_mc_pos_data {

        /* The cost, in bits, of the lowest-cost path that has been found to
         * reach this position.  This can change as progressively lower cost
         * paths are found to reach this position.  */
        u32 cost;
#define MC_INFINITE_COST UINT32_MAX

        /* The match or literal that was taken to reach this position.  This can
         * change as progressively lower cost paths are found to reach this
         * position.
         *
         * This variable is divided into two bitfields.
         *
         * Literals:
         *      Low bits are 1, high bits are the literal.
         *
         * Explicit offset matches:
         *      Low bits are the match length, high bits are the offset plus 2.
         *
         * Repeat offset matches:
         *      Low bits are the match length, high bits are the queue index.
         */
        u64 mc_item_data;
#define MC_OFFSET_SHIFT 32
#define MC_LEN_MASK (((u64)1 << MC_OFFSET_SHIFT) - 1)

        /* The LZMS adaptive state that exists at this position.  This is filled
         * in lazily, only after the minimum-cost path to this position is
         * found.
         *
         * Note: the way we handle this adaptive state in the "minimum-cost"
         * parse is actually only an approximation.  It's possible for the
         * globally optimal, minimum cost path to contain a prefix, ending at a
         * position, where that path prefix is *not* the minimum cost path to
         * that position.  This can happen if such a path prefix results in a
         * different adaptive state which results in lower costs later.  We do
         * not solve this problem; we only consider the lowest cost to reach
         * each position, which seems to be an acceptable approximation.
         *
         * Note: this adaptive state also does not include the probability
         * entries or current Huffman codewords.  Those aren't maintained
         * per-position and are only updated occassionally.  */
        struct lzms_adaptive_state {
                struct lzms_lz_lru_queues lru;
                u8 main_state;
                u8 match_state;
                u8 lz_match_state;
                u8 lz_repeat_match_state[LZMS_NUM_RECENT_OFFSETS - 1];
        } state;
};

static void
lzms_init_fast_slots(struct lzms_compressor *c)
{
        /* Create table mapping small lengths to length slots.  */
        for (unsigned slot = 0, i = 0; i < LZMS_NUM_FAST_LENGTHS; i++) {
                while (i >= lzms_length_slot_base[slot + 1])
                        slot++;
                c->length_slot_fast[i] = slot;
        }

        /* Create table mapping small offsets to offset slots.  */
        for (unsigned slot = 0, i = 0; i < LZMS_NUM_FAST_OFFSETS; i++) {
                while (i >= lzms_offset_slot_base[slot + 1])
                        slot++;
                c->offset_slot_fast[i] = slot;
        }
}

static inline unsigned
lzms_get_length_slot_fast(const struct lzms_compressor *c, u32 length)
{
        if (likely(length < LZMS_NUM_FAST_LENGTHS))
                return c->length_slot_fast[length];
        else
                return lzms_get_length_slot(length);
}

static inline unsigned
lzms_get_offset_slot_fast(const struct lzms_compressor *c, u32 offset)
{
        if (offset < LZMS_NUM_FAST_OFFSETS)
                return c->offset_slot_fast[offset];
        else
                return lzms_get_offset_slot(offset);
}

/* Initialize the output bitstream @os to write backwards to the specified
 * compressed data buffer @out that is @out_limit 16-bit integers long.  */
static void
lzms_output_bitstream_init(struct lzms_output_bitstream *os,
                           le16 *out, size_t out_limit)
{
        os->bitbuf = 0;
        os->bitcount = 0;
        os->next = out + out_limit;
        os->begin = out;
}

/*
 * Write some bits, contained in the low @num_bits bits of @bits (ordered from
 * high-order to low-order), to the output bitstream @os.
 *
 * @max_num_bits is a compile-time constant that specifies the maximum number of
 * bits that can ever be written at this call site.
 */
static inline void
lzms_output_bitstream_put_varbits(struct lzms_output_bitstream *os,
                                  u32 bits, unsigned num_bits,
                                  unsigned max_num_bits)
{
        LZMS_ASSERT(num_bits <= 48);

        /* Add the bits to the bit buffer variable.  */
        os->bitcount += num_bits;
        os->bitbuf = (os->bitbuf << num_bits) | bits;

        /* Check whether any coding units need to be written.  */
        while (os->bitcount >= 16) {

                os->bitcount -= 16;

                /* Write a coding unit, unless it would underflow the buffer. */
                if (os->next != os->begin)
                        *--os->next = cpu_to_le16(os->bitbuf >> os->bitcount);

                /* Optimization for call sites that never write more than 16
                 * bits at once.  */
                if (max_num_bits <= 16)
                        break;
        }
}

/* Use when @num_bits is a compile-time constant.  Otherwise use
 * lzms_output_bitstream_put_bits().  */
static inline void
lzms_output_bitstream_put_bits(struct lzms_output_bitstream *os,
                               u32 bits, unsigned num_bits)
{
        lzms_output_bitstream_put_varbits(os, bits, num_bits, num_bits);
}

/* Flush the output bitstream, ensuring that all bits written to it have been
 * written to memory.  Returns %true if all bits have been output successfully,
 * or %false if an overrun occurred.  */
static bool
lzms_output_bitstream_flush(struct lzms_output_bitstream *os)
{
        if (os->next == os->begin)
                return false;

        if (os->bitcount != 0)
                *--os->next = cpu_to_le16(os->bitbuf << (16 - os->bitcount));

        return true;
}

/* Initialize the range encoder @rc to write forwards to the specified
 * compressed data buffer @out that is @out_limit 16-bit integers long.  */
static void
lzms_range_encoder_raw_init(struct lzms_range_encoder_raw *rc,
                            le16 *out, size_t out_limit)
{
        rc->low = 0;
        rc->range = 0xffffffff;
        rc->cache = 0;
        rc->cache_size = 1;
        rc->begin = out;
        rc->next = out - 1;
        rc->end = out + out_limit;
}

/*
 * Attempt to flush bits from the range encoder.
 *
 * Note: this is based on the public domain code for LZMA written by Igor
 * Pavlov.  The only differences in this function are that in LZMS the bits must
 * be output in 16-bit coding units instead of 8-bit coding units, and that in
 * LZMS the first coding unit is not ignored by the decompressor, so the encoder
 * cannot output a dummy value to that position.
 *
 * The basic idea is that we're writing bits from @rc->low to the output.
 * However, due to carrying, the writing of coding units with value 0xffff, as
 * well as one prior coding unit, must be delayed until it is determined whether
 * a carry is needed.
 */
static void
lzms_range_encoder_raw_shift_low(struct lzms_range_encoder_raw *rc)
{
        if ((u32)(rc->low) < 0xffff0000 ||
            (u32)(rc->low >> 32) != 0)
        {
                /* Carry not needed (rc->low < 0xffff0000), or carry occurred
                 * ((rc->low >> 32) != 0, a.k.a. the carry bit is 1).  */
                do {
                        if (likely(rc->next >= rc->begin)) {
                                if (rc->next != rc->end)
                                        *rc->next++ = cpu_to_le16(rc->cache +
                                                                  (u16)(rc->low >> 32));
                        } else {
                                rc->next++;
                        }
                        rc->cache = 0xffff;
                } while (--rc->cache_size != 0);

                rc->cache = (rc->low >> 16) & 0xffff;
        }
        ++rc->cache_size;
        rc->low = (rc->low & 0xffff) << 16;
}

static void
lzms_range_encoder_raw_normalize(struct lzms_range_encoder_raw *rc)
{
        if (rc->range <= 0xffff) {
                rc->range <<= 16;
                lzms_range_encoder_raw_shift_low(rc);
        }
}

static bool
lzms_range_encoder_raw_flush(struct lzms_range_encoder_raw *rc)
{
        for (unsigned i = 0; i < 4; i++)
                lzms_range_encoder_raw_shift_low(rc);
        return rc->next != rc->end;
}

/* Encode the next bit using the range encoder (raw version).
 *
 * @prob is the chance out of LZMS_PROBABILITY_MAX that the next bit is 0.  */
static inline void
lzms_range_encoder_raw_encode_bit(struct lzms_range_encoder_raw *rc,
                                  int bit, u32 prob)
{
        lzms_range_encoder_raw_normalize(rc);

        u32 bound = (rc->range >> LZMS_PROBABILITY_BITS) * prob;
        if (bit == 0) {
                rc->range = bound;
        } else {
                rc->low += bound;
                rc->range -= bound;
        }
}

/* Encode a bit using the specified range encoder. This wraps around
 * lzms_range_encoder_raw_encode_bit() to handle using and updating the
 * appropriate state and probability entry.  */
static void
lzms_range_encode_bit(struct lzms_range_encoder *enc, int bit)
{
        struct lzms_probability_entry *prob_entry;
        u32 prob;

        /* Load the probability entry corresponding to the current state.  */
        prob_entry = &enc->prob_entries[enc->state];

        /* Update the state based on the next bit.  */
        enc->state = ((enc->state << 1) | bit) & enc->mask;

        /* Get the probability that the bit is 0.  */
        prob = lzms_get_probability(prob_entry);

        /* Update the probability entry.  */
        lzms_update_probability_entry(prob_entry, bit);

        /* Encode the bit.  */
        lzms_range_encoder_raw_encode_bit(enc->rc, bit, prob);
}

/* Called when an adaptive Huffman code needs to be rebuilt.  */
static void
lzms_rebuild_huffman_code(struct lzms_huffman_encoder *enc)
{
        make_canonical_huffman_code(enc->num_syms,
                                    LZMS_MAX_CODEWORD_LEN,
                                    enc->sym_freqs,
                                    enc->lens,
                                    enc->codewords);

        /* Dilute the frequencies.  */
        for (unsigned i = 0; i < enc->num_syms; i++) {
                enc->sym_freqs[i] >>= 1;
                enc->sym_freqs[i] += 1;
        }
        enc->num_syms_written = 0;
}

/* Encode a symbol using the specified Huffman encoder.  */
static inline void
lzms_huffman_encode_symbol(struct lzms_huffman_encoder *enc, unsigned sym)
{
        lzms_output_bitstream_put_varbits(enc->os,
                                          enc->codewords[sym],
                                          enc->lens[sym],
                                          LZMS_MAX_CODEWORD_LEN);
        ++enc->sym_freqs[sym];
        if (++enc->num_syms_written == enc->rebuild_freq)
                lzms_rebuild_huffman_code(enc);
}

static void
lzms_update_fast_length_costs(struct lzms_compressor *c);

/* Encode a match length.  */
static void
lzms_encode_length(struct lzms_compressor *c, u32 length)
{
        unsigned slot;
        unsigned num_extra_bits;
        u32 extra_bits;

        slot = lzms_get_length_slot_fast(c, length);

        extra_bits = length - lzms_length_slot_base[slot];
        num_extra_bits = lzms_extra_length_bits[slot];

        lzms_huffman_encode_symbol(&c->length_encoder, slot);
        if (c->length_encoder.num_syms_written == 0)
                lzms_update_fast_length_costs(c);

        lzms_output_bitstream_put_varbits(c->length_encoder.os,
                                          extra_bits, num_extra_bits, 30);
}

/* Encode an LZ match offset.  */
static void
lzms_encode_lz_offset(struct lzms_compressor *c, u32 offset)
{
        unsigned slot;
        unsigned num_extra_bits;
        u32 extra_bits;

        slot = lzms_get_offset_slot_fast(c, offset);

        extra_bits = offset - lzms_offset_slot_base[slot];
        num_extra_bits = lzms_extra_offset_bits[slot];

        lzms_huffman_encode_symbol(&c->lz_offset_encoder, slot);
        lzms_output_bitstream_put_varbits(c->lz_offset_encoder.os,
                                          extra_bits, num_extra_bits, 30);
}

/* Encode a literal byte.  */
static void
lzms_encode_literal(struct lzms_compressor *c, unsigned literal)
{
        /* Main bit: 0 = a literal, not a match.  */
        lzms_range_encode_bit(&c->main_range_encoder, 0);

        /* Encode the literal using the current literal Huffman code.  */
        lzms_huffman_encode_symbol(&c->literal_encoder, literal);
}

/* Encode an LZ repeat offset match.  */
static void
lzms_encode_lz_repeat_offset_match(struct lzms_compressor *c,
                                   u32 length, unsigned rep_index)
{
        unsigned i;

        /* Main bit: 1 = a match, not a literal.  */
        lzms_range_encode_bit(&c->main_range_encoder, 1);

        /* Match bit: 0 = an LZ match, not a delta match.  */
        lzms_range_encode_bit(&c->match_range_encoder, 0);

        /* LZ match bit: 1 = repeat offset, not an explicit offset.  */
        lzms_range_encode_bit(&c->lz_match_range_encoder, 1);

        /* Encode the repeat offset index.  A 1 bit is encoded for each index
         * passed up.  This sequence of 1 bits is terminated by a 0 bit, or
         * automatically when (LZMS_NUM_RECENT_OFFSETS - 1) 1 bits have been
         * encoded.  */
        for (i = 0; i < rep_index; i++)
                lzms_range_encode_bit(&c->lz_repeat_match_range_encoders[i], 1);

        if (i < LZMS_NUM_RECENT_OFFSETS - 1)
                lzms_range_encode_bit(&c->lz_repeat_match_range_encoders[i], 0);

        /* Encode the match length.  */
        lzms_encode_length(c, length);
}

/* Encode an LZ explicit offset match.  */
static void
lzms_encode_lz_explicit_offset_match(struct lzms_compressor *c,
                                     u32 length, u32 offset)
{
        /* Main bit: 1 = a match, not a literal.  */
        lzms_range_encode_bit(&c->main_range_encoder, 1);

        /* Match bit: 0 = an LZ match, not a delta match.  */
        lzms_range_encode_bit(&c->match_range_encoder, 0);

        /* LZ match bit: 0 = explicit offset, not a repeat offset.  */
        lzms_range_encode_bit(&c->lz_match_range_encoder, 0);

        /* Encode the match offset.  */
        lzms_encode_lz_offset(c, offset);

        /* Encode the match length.  */
        lzms_encode_length(c, length);
}

static void
lzms_encode_item(struct lzms_compressor *c, u64 mc_item_data)
{
        u32 len = mc_item_data & MC_LEN_MASK;
        u32 offset_data = mc_item_data >> MC_OFFSET_SHIFT;

        if (len == 1)
                lzms_encode_literal(c, offset_data);
        else if (offset_data < LZMS_NUM_RECENT_OFFSETS)
                lzms_encode_lz_repeat_offset_match(c, len, offset_data);
        else
                lzms_encode_lz_explicit_offset_match(c, len, offset_data - LZMS_OFFSET_OFFSET);
}

/* Encode a list of matches and literals chosen by the parsing algorithm.  */
static void
lzms_encode_item_list(struct lzms_compressor *c,
                      struct lzms_mc_pos_data *cur_optimum_ptr)
{
        struct lzms_mc_pos_data *end_optimum_ptr;
        u64 saved_item;
        u64 item;

        /* The list is currently in reverse order (last item to first item).
         * Reverse it.  */
        end_optimum_ptr = cur_optimum_ptr;
        saved_item = cur_optimum_ptr->mc_item_data;
        do {
                item = saved_item;
                cur_optimum_ptr -= item & MC_LEN_MASK;
                saved_item = cur_optimum_ptr->mc_item_data;
                cur_optimum_ptr->mc_item_data = item;
        } while (cur_optimum_ptr != c->optimum);

        /* Walk the list of items from beginning to end, encoding each item.  */
        do {
                lzms_encode_item(c, cur_optimum_ptr->mc_item_data);
                cur_optimum_ptr += (cur_optimum_ptr->mc_item_data) & MC_LEN_MASK;
        } while (cur_optimum_ptr != end_optimum_ptr);
}

/* Each bit costs 1 << LZMS_COST_SHIFT units.  */
#define LZMS_COST_SHIFT 6

/*#define LZMS_RC_COSTS_USE_FLOATING_POINT*/

static u32
lzms_rc_costs[LZMS_PROBABILITY_MAX + 1];

#ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
#  include <math.h>
#endif

static void
lzms_do_init_rc_costs(void)
{
        /* Fill in a table that maps range coding probabilities needed to code a
         * bit X (0 or 1) to the number of bits (scaled by a constant factor, to
         * handle fractional costs) needed to code that bit X.
         *
         * Consider the range of the range decoder.  To eliminate exactly half
         * the range (logical probability of 0.5), we need exactly 1 bit.  For
         * lower probabilities we need more bits and for higher probabilities we
         * need fewer bits.  In general, a logical probability of N will
         * eliminate the proportion 1 - N of the range; this information takes
         * log2(1 / N) bits to encode.
         *
         * The below loop is simply calculating this number of bits for each
         * possible probability allowed by the LZMS compression format, but
         * without using real numbers.  To handle fractional probabilities, each
         * cost is multiplied by (1 << LZMS_COST_SHIFT).  These techniques are
         * based on those used by LZMA.
         *
         * Note that in LZMS, a probability x really means x / 64, and 0 / 64 is
         * really interpreted as 1 / 64 and 64 / 64 is really interpreted as
         * 63 / 64.
         */
        for (u32 i = 0; i <= LZMS_PROBABILITY_MAX; i++) {
                u32 prob = i;

                if (prob == 0)
                        prob = 1;
                else if (prob == LZMS_PROBABILITY_MAX)
                        prob = LZMS_PROBABILITY_MAX - 1;

        #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
                lzms_rc_costs[i] = log2((double)LZMS_PROBABILITY_MAX / prob) *
                                        (1 << LZMS_COST_SHIFT);
        #else
                u32 w = prob;
                u32 bit_count = 0;
                for (u32 j = 0; j < LZMS_COST_SHIFT; j++) {
                        w *= w;
                        bit_count <<= 1;
                        while (w >= (1U << 16)) {
                                w >>= 1;
                                ++bit_count;
                        }
                }
                lzms_rc_costs[i] = (LZMS_PROBABILITY_BITS << LZMS_COST_SHIFT) -
                                   (15 + bit_count);
        #endif
        }
}

static void
lzms_init_rc_costs(void)
{
        static pthread_once_t once = PTHREAD_ONCE_INIT;

        pthread_once(&once, lzms_do_init_rc_costs);
}

/* Return the cost to range-encode the specified bit from the specified state.*/
static inline u32
lzms_rc_bit_cost(const struct lzms_range_encoder *enc, u8 cur_state, int bit)
{
        u32 prob_zero;
        u32 prob_correct;

        prob_zero = enc->prob_entries[cur_state].num_recent_zero_bits;

        if (bit == 0)
                prob_correct = prob_zero;
        else
                prob_correct = LZMS_PROBABILITY_MAX - prob_zero;

        return lzms_rc_costs[prob_correct];
}

/* Return the cost to Huffman-encode the specified symbol.  */
static inline u32
lzms_huffman_symbol_cost(const struct lzms_huffman_encoder *enc, unsigned sym)
{
        return (u32)enc->lens[sym] << LZMS_COST_SHIFT;
}

/* Return the cost to encode the specified literal byte.  */
static inline u32
lzms_literal_cost(const struct lzms_compressor *c, unsigned literal,
                  const struct lzms_adaptive_state *state)
{
        return lzms_rc_bit_cost(&c->main_range_encoder, state->main_state, 0) +
               lzms_huffman_symbol_cost(&c->literal_encoder, literal);
}

/* Update the table that directly provides the costs for small lengths.  */
static void
lzms_update_fast_length_costs(struct lzms_compressor *c)
{
        u32 len;
        int slot = -1;
        u32 cost = 0;

        for (len = 1; len < LZMS_NUM_FAST_LENGTHS; len++) {

                while (len >= lzms_length_slot_base[slot + 1]) {
                        slot++;
                        cost = (u32)(c->length_encoder.lens[slot] +
                                     lzms_extra_length_bits[slot]) << LZMS_COST_SHIFT;
                }

                c->length_cost_fast[len] = cost;
        }
}

/* Return the cost to encode the specified match length, which must be less than
 * LZMS_NUM_FAST_LENGTHS.  */
static inline u32
lzms_fast_length_cost(const struct lzms_compressor *c, u32 length)
{
        LZMS_ASSERT(length < LZMS_NUM_FAST_LENGTHS);
        return c->length_cost_fast[length];
}

/* Return the cost to encode the specified LZ match offset.  */
static inline u32
lzms_lz_offset_cost(const struct lzms_compressor *c, u32 offset)
{
        unsigned slot = lzms_get_offset_slot_fast(c, offset);

        return (u32)(c->lz_offset_encoder.lens[slot] +
                     lzms_extra_offset_bits[slot]) << LZMS_COST_SHIFT;
}

/*
 * Consider coding the match at repeat offset index @rep_idx.  Consider each
 * length from the minimum (2) to the full match length (@rep_len).
 */
static inline void
lzms_consider_lz_repeat_offset_match(const struct lzms_compressor *c,
                                     struct lzms_mc_pos_data *cur_optimum_ptr,
                                     u32 rep_len, unsigned rep_idx)
{
        u32 len;
        u32 base_cost;
        u32 cost;
        unsigned i;

        base_cost = cur_optimum_ptr->cost;

        base_cost += lzms_rc_bit_cost(&c->main_range_encoder,
                                      cur_optimum_ptr->state.main_state, 1);

        base_cost += lzms_rc_bit_cost(&c->match_range_encoder,
                                      cur_optimum_ptr->state.match_state, 0);

        base_cost += lzms_rc_bit_cost(&c->lz_match_range_encoder,
                                      cur_optimum_ptr->state.lz_match_state, 1);

        for (i = 0; i < rep_idx; i++)
                base_cost += lzms_rc_bit_cost(&c->lz_repeat_match_range_encoders[i],
                                              cur_optimum_ptr->state.lz_repeat_match_state[i], 1);

        if (i < LZMS_NUM_RECENT_OFFSETS - 1)
                base_cost += lzms_rc_bit_cost(&c->lz_repeat_match_range_encoders[i],
                                              cur_optimum_ptr->state.lz_repeat_match_state[i], 0);

        len = 2;
        do {
                cost = base_cost + lzms_fast_length_cost(c, len);
                if (cost < (cur_optimum_ptr + len)->cost) {
                        (cur_optimum_ptr + len)->mc_item_data =
                                ((u64)rep_idx << MC_OFFSET_SHIFT) | len;
                        (cur_optimum_ptr + len)->cost = cost;
                }
        } while (++len <= rep_len);
}

/*
 * Consider coding each match in @matches as an explicit offset match.
 *
 * @matches must be sorted by strictly increasing length and strictly increasing
 * offset.  This is guaranteed by the match-finder.
 *
 * We consider each length from the minimum (2) to the longest
 * (matches[num_matches - 1].len).  For each length, we consider only the
 * smallest offset for which that length is available.  Although this is not
 * guaranteed to be optimal due to the possibility of a larger offset costing
 * less than a smaller offset to code, this is a very useful heuristic.
 */
static inline void
lzms_consider_lz_explicit_offset_matches(const struct lzms_compressor *c,
                                         struct lzms_mc_pos_data *cur_optimum_ptr,
                                         const struct lz_match matches[],
                                         u32 num_matches)
{
        u32 len;
        u32 i;
        u32 base_cost;
        u32 position_cost;
        u32 cost;

        base_cost = cur_optimum_ptr->cost;

        base_cost += lzms_rc_bit_cost(&c->main_range_encoder,
                                      cur_optimum_ptr->state.main_state, 1);

        base_cost += lzms_rc_bit_cost(&c->match_range_encoder,
                                      cur_optimum_ptr->state.match_state, 0);

        base_cost += lzms_rc_bit_cost(&c->lz_match_range_encoder,
                                      cur_optimum_ptr->state.lz_match_state, 0);
        len = 2;
        i = 0;
        do {
                position_cost = base_cost + lzms_lz_offset_cost(c,
                                                                matches[i].offset);
                do {
                        cost = position_cost + lzms_fast_length_cost(c, len);
                        if (cost < (cur_optimum_ptr + len)->cost) {
                                (cur_optimum_ptr + len)->mc_item_data =
                                        ((u64)(matches[i].offset + LZMS_OFFSET_OFFSET)
                                                << MC_OFFSET_SHIFT) | len;
                                (cur_optimum_ptr + len)->cost = cost;
                        }
                } while (++len <= matches[i].len);
        } while (++i != num_matches);
}

static void
lzms_init_adaptive_state(struct lzms_adaptive_state *state)
{
        unsigned i;

        lzms_init_lz_lru_queues(&state->lru);
        state->main_state = 0;
        state->match_state = 0;
        state->lz_match_state = 0;
        for (i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
                state->lz_repeat_match_state[i] = 0;
}

static inline void
lzms_update_main_state(struct lzms_adaptive_state *state, int is_match)
{
        state->main_state = ((state->main_state << 1) | is_match) % LZMS_NUM_MAIN_STATES;
}

static inline void
lzms_update_match_state(struct lzms_adaptive_state *state, int is_delta)
{
        state->match_state = ((state->match_state << 1) | is_delta) % LZMS_NUM_MATCH_STATES;
}

static inline void
lzms_update_lz_match_state(struct lzms_adaptive_state *state, int is_repeat_offset)
{
        state->lz_match_state = ((state->lz_match_state << 1) | is_repeat_offset) % LZMS_NUM_LZ_MATCH_STATES;
}

static inline void
lzms_update_lz_repeat_match_state(struct lzms_adaptive_state *state, int rep_idx)
{
        int i;

        for (i = 0; i < rep_idx; i++)
                state->lz_repeat_match_state[i] =
                        ((state->lz_repeat_match_state[i] << 1) | 1) %
                                LZMS_NUM_LZ_REPEAT_MATCH_STATES;

        if (i < LZMS_NUM_RECENT_OFFSETS - 1)
                state->lz_repeat_match_state[i] =
                        ((state->lz_repeat_match_state[i] << 1) | 0) %
                                LZMS_NUM_LZ_REPEAT_MATCH_STATES;
}

/*
 * The main near-optimal parsing routine.
 *
 * Briefly, the algorithm does an approximate minimum-cost path search to find a
 * "near-optimal" sequence of matches and literals to output, based on the
 * current cost model.  The algorithm steps forward, position by position (byte
 * by byte), and updates the minimum cost path to reach each later position that
 * can be reached using a match or literal from the current position.  This is
 * essentially Dijkstra's algorithm in disguise: the graph nodes are positions,
 * the graph edges are possible matches/literals to code, and the cost of each
 * edge is the estimated number of bits that will be required to output the
 * corresponding match or literal.  But one difference is that we actually
 * compute the lowest-cost path in pieces, where each piece is terminated when
 * there are no choices to be made.
 *
 * Notes:
 *
 * - This does not output any delta matches.
 *
 * - The costs of literals and matches are estimated using the range encoder
 *   states and the semi-adaptive Huffman codes.  Except for range encoding
 *   states, costs are assumed to be constant throughout a single run of the
 *   parsing algorithm, which can parse up to @optim_array_length bytes of data.
 *   This introduces a source of inaccuracy because the probabilities and
 *   Huffman codes can change over this part of the data.
 */
static void
lzms_near_optimal_parse(struct lzms_compressor *c)
{
        const u8 *window_ptr;
        const u8 *window_end;
        struct lzms_mc_pos_data *cur_optimum_ptr;
        struct lzms_mc_pos_data *end_optimum_ptr;
        u32 num_matches;
        u32 longest_len;
        u32 rep_max_len;
        unsigned rep_max_idx;
        unsigned literal;
        unsigned i;
        u32 cost;
        u32 len;
        u32 offset_data;

        window_ptr = c->cur_window;
        window_end = window_ptr + c->cur_window_size;

        lzms_init_adaptive_state(&c->optimum[0].state);

begin:
        /* Start building a new list of items, which will correspond to the next
         * piece of the overall minimum-cost path.  */

        cur_optimum_ptr = c->optimum;
        cur_optimum_ptr->cost = 0;
        end_optimum_ptr = cur_optimum_ptr;

        /* States should currently be consistent with the encoders.  */
        LZMS_ASSERT(cur_optimum_ptr->state.main_state == c->main_range_encoder.state);
        LZMS_ASSERT(cur_optimum_ptr->state.match_state == c->match_range_encoder.state);
        LZMS_ASSERT(cur_optimum_ptr->state.lz_match_state == c->lz_match_range_encoder.state);
        for (i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
                LZMS_ASSERT(cur_optimum_ptr->state.lz_repeat_match_state[i] ==
                            c->lz_repeat_match_range_encoders[i].state);

        if (window_ptr == window_end)
                return;

        /* The following loop runs once for each per byte in the window, except
         * in a couple shortcut cases.  */
        for (;;) {

                /* Find explicit offset matches with the current position.  */
                num_matches = lz_mf_get_matches(c->mf, c->matches);

                if (num_matches) {
                        /*
                         * Find the longest repeat offset match with the current
                         * position.
                         *
                         * Heuristics:
                         *
                         * - Only search for repeat offset matches if the
                         *   match-finder already found at least one match.
                         *
                         * - Only consider the longest repeat offset match.  It
                         *   seems to be rare for the optimal parse to include a
                         *   repeat offset match that doesn't have the longest
                         *   length (allowing for the possibility that not all
                         *   of that length is actually used).
                         */
                        if (likely(window_ptr - c->cur_window >= LZMS_MAX_INIT_RECENT_OFFSET)) {
                                BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
                                rep_max_len = lz_repsearch3(window_ptr,
                                                            window_end - window_ptr,
                                                            cur_optimum_ptr->state.lru.recent_offsets,
                                                            &rep_max_idx);
                        } else {
                                rep_max_len = 0;
                        }

                        if (rep_max_len) {
                                /* If there's a very long repeat offset match,
                                 * choose it immediately.  */
                                if (rep_max_len >= c->params.nice_match_length) {

                                        lz_mf_skip_positions(c->mf, rep_max_len - 1);
                                        window_ptr += rep_max_len;

                                        if (cur_optimum_ptr != c->optimum)
                                                lzms_encode_item_list(c, cur_optimum_ptr);

                                        lzms_encode_lz_repeat_offset_match(c, rep_max_len,
                                                                           rep_max_idx);

                                        c->optimum[0].state = cur_optimum_ptr->state;

                                        lzms_update_main_state(&c->optimum[0].state, 1);
                                        lzms_update_match_state(&c->optimum[0].state, 0);
                                        lzms_update_lz_match_state(&c->optimum[0].state, 1);
                                        lzms_update_lz_repeat_match_state(&c->optimum[0].state,
                                                                          rep_max_idx);

                                        c->optimum[0].state.lru.upcoming_offset =
                                                c->optimum[0].state.lru.recent_offsets[rep_max_idx];

                                        for (i = rep_max_idx; i < LZMS_NUM_RECENT_OFFSETS; i++)
                                                c->optimum[0].state.lru.recent_offsets[i] =
                                                        c->optimum[0].state.lru.recent_offsets[i + 1];

                                        lzms_update_lz_lru_queue(&c->optimum[0].state.lru);
                                        goto begin;
                                }

                                /* If reaching any positions for the first time,
                                 * initialize their costs to "infinity".  */
                                while (end_optimum_ptr < cur_optimum_ptr + rep_max_len)
                                        (++end_optimum_ptr)->cost = MC_INFINITE_COST;

                                /* Consider coding a repeat offset match.  */
                                lzms_consider_lz_repeat_offset_match(c, cur_optimum_ptr,
                                                                     rep_max_len, rep_max_idx);
                        }

                        longest_len = c->matches[num_matches - 1].len;

                        /* If there's a very long explicit offset match, choose
                         * it immediately.  */
                        if (longest_len >= c->params.nice_match_length) {

                                lz_mf_skip_positions(c->mf, longest_len - 1);
                                window_ptr += longest_len;

                                if (cur_optimum_ptr != c->optimum)
                                        lzms_encode_item_list(c, cur_optimum_ptr);

                                lzms_encode_lz_explicit_offset_match(c, longest_len,
                                                                     c->matches[num_matches - 1].offset);

                                c->optimum[0].state = cur_optimum_ptr->state;

                                lzms_update_main_state(&c->optimum[0].state, 1);
                                lzms_update_match_state(&c->optimum[0].state, 0);
                                lzms_update_lz_match_state(&c->optimum[0].state, 0);

                                c->optimum[0].state.lru.upcoming_offset =
                                        c->matches[num_matches - 1].offset;

                                lzms_update_lz_lru_queue(&c->optimum[0].state.lru);
                                goto begin;
                        }

                        /* If reaching any positions for the first time,
                         * initialize their costs to "infinity".  */
                        while (end_optimum_ptr < cur_optimum_ptr + longest_len)
                                (++end_optimum_ptr)->cost = MC_INFINITE_COST;

                        /* Consider coding an explicit offset match.  */
                        lzms_consider_lz_explicit_offset_matches(c, cur_optimum_ptr,
                                                                 c->matches, num_matches);
                } else {
                        /* No matches found.  The only choice at this position
                         * is to code a literal.  */

                        if (end_optimum_ptr == cur_optimum_ptr)
                                (++end_optimum_ptr)->cost = MC_INFINITE_COST;
                }

                /* Consider coding a literal.

                 * To avoid an extra unpredictable brench, actually checking the
                 * preferability of coding a literal is integrated into the
                 * adaptive state update code below.  */
                literal = *window_ptr++;
                cost = cur_optimum_ptr->cost +
                       lzms_literal_cost(c, literal, &cur_optimum_ptr->state);

                /* Advance to the next position.  */
                cur_optimum_ptr++;

                /* The lowest-cost path to the current position is now known.
                 * Finalize the adaptive state that results from taking this
                 * lowest-cost path.  */

                if (cost < cur_optimum_ptr->cost) {
                        /* Literal  */
                        cur_optimum_ptr->cost = cost;
                        cur_optimum_ptr->mc_item_data = ((u64)literal << MC_OFFSET_SHIFT) | 1;

                        cur_optimum_ptr->state = (cur_optimum_ptr - 1)->state;

                        lzms_update_main_state(&cur_optimum_ptr->state, 0);

                        cur_optimum_ptr->state.lru.upcoming_offset = 0;
                } else {
                        /* LZ match  */
                        len = cur_optimum_ptr->mc_item_data & MC_LEN_MASK;
                        offset_data = cur_optimum_ptr->mc_item_data >> MC_OFFSET_SHIFT;

                        cur_optimum_ptr->state = (cur_optimum_ptr - len)->state;

                        lzms_update_main_state(&cur_optimum_ptr->state, 1);
                        lzms_update_match_state(&cur_optimum_ptr->state, 0);

                        if (offset_data >= LZMS_NUM_RECENT_OFFSETS) {

                                /* Explicit offset LZ match  */

                                lzms_update_lz_match_state(&cur_optimum_ptr->state, 0);

                                cur_optimum_ptr->state.lru.upcoming_offset =
                                        offset_data - LZMS_OFFSET_OFFSET;
                        } else {
                                /* Repeat offset LZ match  */

                                lzms_update_lz_match_state(&cur_optimum_ptr->state, 1);
                                lzms_update_lz_repeat_match_state(&cur_optimum_ptr->state,
                                                                  offset_data);

                                cur_optimum_ptr->state.lru.upcoming_offset =
                                        cur_optimum_ptr->state.lru.recent_offsets[offset_data];

                                for (i = offset_data; i < LZMS_NUM_RECENT_OFFSETS; i++)
                                        cur_optimum_ptr->state.lru.recent_offsets[i] =
                                                cur_optimum_ptr->state.lru.recent_offsets[i + 1];
                        }
                }

                lzms_update_lz_lru_queue(&cur_optimum_ptr->state.lru);

                /*
                 * This loop will terminate when either of the following
                 * conditions is true:
                 *
                 * (1) cur_optimum_ptr == end_optimum_ptr
                 *
                 *      There are no paths that extend beyond the current
                 *      position.  In this case, any path to a later position
                 *      must pass through the current position, so we can go
                 *      ahead and choose the list of items that led to this
                 *      position.
                 *
                 * (2) cur_optimum_ptr == c->optimum_end
                 *
                 *      This bounds the number of times the algorithm can step
                 *      forward before it is guaranteed to start choosing items.
                 *      This limits the memory usage.  It also guarantees that
                 *      the parser will not go too long without updating the
                 *      probability tables.
                 *
                 * Note: no check for end-of-block is needed because
                 * end-of-block will trigger condition (1).
                 */
                if (cur_optimum_ptr == end_optimum_ptr ||
                    cur_optimum_ptr == c->optimum_end)
                {
                        c->optimum[0].state = cur_optimum_ptr->state;
                        break;
                }
        }

        /* Output the current list of items that constitute the minimum-cost
         * path to the current position.  */
        lzms_encode_item_list(c, cur_optimum_ptr);
        goto begin;
}

static void
lzms_init_range_encoder(struct lzms_range_encoder *enc,
                        struct lzms_range_encoder_raw *rc, u32 num_states)
{
        enc->rc = rc;
        enc->state = 0;
        LZMS_ASSERT(is_power_of_2(num_states));
        enc->mask = num_states - 1;
        for (u32 i = 0; i < num_states; i++) {
                enc->prob_entries[i].num_recent_zero_bits = LZMS_INITIAL_PROBABILITY;
                enc->prob_entries[i].recent_bits = LZMS_INITIAL_RECENT_BITS;
        }
}

static void
lzms_init_huffman_encoder(struct lzms_huffman_encoder *enc,
                          struct lzms_output_bitstream *os,
                          unsigned num_syms,
                          unsigned rebuild_freq)
{
        enc->os = os;
        enc->num_syms_written = 0;
        enc->rebuild_freq = rebuild_freq;
        enc->num_syms = num_syms;
        for (unsigned i = 0; i < num_syms; i++)
                enc->sym_freqs[i] = 1;

        make_canonical_huffman_code(enc->num_syms,
                                    LZMS_MAX_CODEWORD_LEN,
                                    enc->sym_freqs,
                                    enc->lens,
                                    enc->codewords);
}

/* Prepare the LZMS compressor for compressing a block of data.  */
static void
lzms_prepare_compressor(struct lzms_compressor *c, const u8 *udata, u32 ulen,
                        le16 *cdata, u32 clen16)
{
        unsigned num_offset_slots;

        /* Copy the uncompressed data into the @c->cur_window buffer.  */
        memcpy(c->cur_window, udata, ulen);
        c->cur_window_size = ulen;

        /* Initialize the raw range encoder (writing forwards).  */
        lzms_range_encoder_raw_init(&c->rc, cdata, clen16);

        /* Initialize the output bitstream for Huffman symbols and verbatim bits
         * (writing backwards).  */
        lzms_output_bitstream_init(&c->os, cdata, clen16);

        /* Calculate the number of offset slots required.  */
        num_offset_slots = lzms_get_offset_slot(ulen - 1) + 1;

        /* Initialize a Huffman encoder for each alphabet.  */
        lzms_init_huffman_encoder(&c->literal_encoder, &c->os,
                                  LZMS_NUM_LITERAL_SYMS,
                                  LZMS_LITERAL_CODE_REBUILD_FREQ);

        lzms_init_huffman_encoder(&c->lz_offset_encoder, &c->os,
                                  num_offset_slots,
                                  LZMS_LZ_OFFSET_CODE_REBUILD_FREQ);

        lzms_init_huffman_encoder(&c->length_encoder, &c->os,
                                  LZMS_NUM_LEN_SYMS,
                                  LZMS_LENGTH_CODE_REBUILD_FREQ);

        lzms_init_huffman_encoder(&c->delta_offset_encoder, &c->os,
                                  num_offset_slots,
                                  LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ);

        lzms_init_huffman_encoder(&c->delta_power_encoder, &c->os,
                                  LZMS_NUM_DELTA_POWER_SYMS,
                                  LZMS_DELTA_POWER_CODE_REBUILD_FREQ);

        /* Initialize range encoders, all of which wrap around the same
         * lzms_range_encoder_raw.  */
        lzms_init_range_encoder(&c->main_range_encoder,
                                &c->rc, LZMS_NUM_MAIN_STATES);

        lzms_init_range_encoder(&c->match_range_encoder,
                                &c->rc, LZMS_NUM_MATCH_STATES);

        lzms_init_range_encoder(&c->lz_match_range_encoder,
                                &c->rc, LZMS_NUM_LZ_MATCH_STATES);

        for (unsigned i = 0; i < ARRAY_LEN(c->lz_repeat_match_range_encoders); i++)
                lzms_init_range_encoder(&c->lz_repeat_match_range_encoders[i],
                                        &c->rc, LZMS_NUM_LZ_REPEAT_MATCH_STATES);

        lzms_init_range_encoder(&c->delta_match_range_encoder,
                                &c->rc, LZMS_NUM_DELTA_MATCH_STATES);

        for (unsigned i = 0; i < ARRAY_LEN(c->delta_repeat_match_range_encoders); i++)
                lzms_init_range_encoder(&c->delta_repeat_match_range_encoders[i],
                                        &c->rc, LZMS_NUM_DELTA_REPEAT_MATCH_STATES);

        /* Set initial length costs for lengths < LZMS_NUM_FAST_LENGTHS.  */
        lzms_update_fast_length_costs(c);
}

/* Flush the output streams, prepare the final compressed data, and return its
 * size in bytes.
 *
 * A return value of 0 indicates that the data could not be compressed to fit in
 * the available space.  */
static size_t
lzms_finalize(struct lzms_compressor *c, u8 *cdata, size_t csize_avail)
{
        size_t num_forwards_bytes;
        size_t num_backwards_bytes;

        /* Flush both the forwards and backwards streams, and make sure they
         * didn't cross each other and start overwriting each other's data.  */
        if (!lzms_output_bitstream_flush(&c->os))
                return 0;

        if (!lzms_range_encoder_raw_flush(&c->rc))
                return 0;

        if (c->rc.next > c->os.next)
                return 0;

        /* Now the compressed buffer contains the data output by the forwards
         * bitstream, then empty space, then data output by the backwards
         * bitstream.  Move the data output by the backwards bitstream to be
         * adjacent to the data output by the forward bitstream, and calculate
         * the compressed size that this results in.  */
        num_forwards_bytes = (u8*)c->rc.next - (u8*)cdata;
        num_backwards_bytes = ((u8*)cdata + csize_avail) - (u8*)c->os.next;

        memmove(cdata + num_forwards_bytes, c->os.next, num_backwards_bytes);

        return num_forwards_bytes + num_backwards_bytes;
}

/* Set internal compression parameters for the specified compression level and
 * maximum window size.  */
static void
lzms_build_params(unsigned int compression_level,
                  struct lzms_compressor_params *lzms_params)
{
        /* Allow length 2 matches if the compression level is sufficiently high.
         */
        if (compression_level >= 45)
                lzms_params->min_match_length = 2;
        else
                lzms_params->min_match_length = 3;

        /* Scale nice_match_length and max_search_depth with the compression
         * level.  But to allow an optimization on length cost calculations,
         * don't allow nice_match_length to exceed LZMS_NUM_FAST_LENGTH.  */
        lzms_params->nice_match_length = ((u64)compression_level * 32) / 50;
        if (lzms_params->nice_match_length < lzms_params->min_match_length)
                lzms_params->nice_match_length = lzms_params->min_match_length;
        if (lzms_params->nice_match_length > LZMS_NUM_FAST_LENGTHS)
                lzms_params->nice_match_length = LZMS_NUM_FAST_LENGTHS;
        lzms_params->max_search_depth = compression_level;

        lzms_params->optim_array_length = 1024;
}

/* Given the internal compression parameters and maximum window size, build the
 * Lempel-Ziv match-finder parameters.  */
static void
lzms_build_mf_params(const struct lzms_compressor_params *lzms_params,
                     u32 max_window_size, struct lz_mf_params *mf_params)
{
        memset(mf_params, 0, sizeof(*mf_params));

        /* Choose an appropriate match-finding algorithm.  */
        if (max_window_size <= 2097152)
                mf_params->algorithm = LZ_MF_BINARY_TREES;
        else if (max_window_size <= 33554432)
                mf_params->algorithm = LZ_MF_LCP_INTERVAL_TREE;
        else
                mf_params->algorithm = LZ_MF_LINKED_SUFFIX_ARRAY;

        mf_params->max_window_size = max_window_size;
        mf_params->min_match_len = lzms_params->min_match_length;
        mf_params->max_search_depth = lzms_params->max_search_depth;
        mf_params->nice_match_len = lzms_params->nice_match_length;
}

static void
lzms_free_compressor(void *_c);

static u64
lzms_get_needed_memory(size_t max_block_size, unsigned int compression_level)
{
        struct lzms_compressor_params params;
        struct lz_mf_params mf_params;
        u64 size = 0;

        if (max_block_size >= INT32_MAX)
                return 0;

        lzms_build_params(compression_level, &params);
        lzms_build_mf_params(&params, max_block_size, &mf_params);

        size += sizeof(struct lzms_compressor);

        /* cur_window */
        size += max_block_size;

        /* mf */
        size += lz_mf_get_needed_memory(mf_params.algorithm, max_block_size);

        /* matches */
        size += min(params.max_search_depth, params.nice_match_length) *
                sizeof(struct lz_match);

        /* optimum */
        size += (params.optim_array_length + params.nice_match_length) *
                sizeof(struct lzms_mc_pos_data);

        return size;
}

static int
lzms_create_compressor(size_t max_block_size, unsigned int compression_level,
                       void **ctx_ret)
{
        struct lzms_compressor *c;
        struct lzms_compressor_params params;
        struct lz_mf_params mf_params;

        if (max_block_size >= INT32_MAX)
                return WIMLIB_ERR_INVALID_PARAM;

        lzms_build_params(compression_level, &params);
        lzms_build_mf_params(&params, max_block_size, &mf_params);
        if (!lz_mf_params_valid(&mf_params))
                return WIMLIB_ERR_INVALID_PARAM;

        c = CALLOC(1, sizeof(struct lzms_compressor));
        if (!c)
                goto oom;

        c->params = params;

        c->cur_window = MALLOC(max_block_size);
        if (!c->cur_window)
                goto oom;

        c->mf = lz_mf_alloc(&mf_params);
        if (!c->mf)
                goto oom;

        c->matches = MALLOC(min(params.max_search_depth,
                                params.nice_match_length) *
                            sizeof(struct lz_match));
        if (!c->matches)
                goto oom;

        c->optimum = MALLOC((params.optim_array_length +
                             params.nice_match_length) *
                            sizeof(struct lzms_mc_pos_data));
        if (!c->optimum)
                goto oom;
        c->optimum_end = &c->optimum[params.optim_array_length];

        lzms_init_slots();

        lzms_init_rc_costs();

        lzms_init_fast_slots(c);

        *ctx_ret = c;
        return 0;

oom:
        lzms_free_compressor(c);
        return WIMLIB_ERR_NOMEM;
}

static size_t
lzms_compress(const void *uncompressed_data, size_t uncompressed_size,
              void *compressed_data, size_t compressed_size_avail, void *_c)
{
        struct lzms_compressor *c = _c;

        /* Don't bother compressing extremely small inputs.  */
        if (uncompressed_size < 4)
                return 0;

        /* Cap the available compressed size to a 32-bit integer and round it
         * down to the nearest multiple of 2.  */
        if (compressed_size_avail > UINT32_MAX)
                compressed_size_avail = UINT32_MAX;
        if (compressed_size_avail & 1)
                compressed_size_avail--;

        /* Initialize the compressor structures.  */
        lzms_prepare_compressor(c, uncompressed_data, uncompressed_size,
                                compressed_data, compressed_size_avail / 2);

        /* Preprocess the uncompressed data.  */
        lzms_x86_filter(c->cur_window, c->cur_window_size,
                        c->last_target_usages, false);

        /* Load the window into the match-finder.  */
        lz_mf_load_window(c->mf, c->cur_window, c->cur_window_size);

        /* Compute and encode a literal/match sequence that decompresses to the
         * preprocessed data.  */
        lzms_near_optimal_parse(c);

        /* Return the compressed data size or 0.  */
        return lzms_finalize(c, compressed_data, compressed_size_avail);
}

static void
lzms_free_compressor(void *_c)
{
        struct lzms_compressor *c = _c;

        if (c) {
                FREE(c->cur_window);
                lz_mf_free(c->mf);
                FREE(c->matches);
                FREE(c->optimum);
                FREE(c);
        }
}

const struct compressor_ops lzms_compressor_ops = {
        .get_needed_memory  = lzms_get_needed_memory,
        .create_compressor  = lzms_create_compressor,
        .compress           = lzms_compress,
        .free_compressor    = lzms_free_compressor,
};