Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, uint32_t cur_match) {

    const unsigned wmask = W_MASK(s);

    unsigned int strstart = s->strstart;

    const unsigned char *window = s->window;

    const Pos *prev = s->prev;

#ifdef LONGEST_MATCH_ROLL

    const Pos *head = s->head;

#endif

    const unsigned char *scan;

    const unsigned char *mbase_start = window;

    const unsigned char *mbase_end;

    uint32_t limit;

#ifdef LONGEST_MATCH_ROLL

    uint32_t limit_base;

#else

    int32_t early_exit;

#endif

    uint32_t chain_length = s->max_chain_length;

    uint32_t nice_match = (uint32_t)s->nice_match;

    uint32_t best_len, offset;

    uint32_t lookahead = s->lookahead;

    uint32_t match_offset = 0;

    uint64_t scan_start;

    uint64_t scan_end;

    /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */

    Assert(STD_MAX_MATCH == 258, "Code too clever");

    best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1;

    if (UNLIKELY(best_len >= lookahead))

        return lookahead;

    /* Calculate read offset which should only extend an extra byte to find the

     * next best match length. When best_len is shorter than the read width, we

     * diff the mismatched bytes instead.

     */

    offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0;

    scan = window + strstart;

    scan_start = zng_memread_8(scan);

    scan_end = zng_memread_8(scan+offset);

    mbase_end = (mbase_start+offset);

    /* Do not waste too much time if we already have a good match */

    if (UNLIKELY(best_len >= s->good_match))

        chain_length >>= 2;

    /* Stop when cur_match becomes <= limit. To simplify the code,

     * we prevent matches with the string of window index 0

     */

    limit = strstart > MAX_DIST(s) ? (strstart - MAX_DIST(s)) : 0;

#ifdef LONGEST_MATCH_ROLL

    limit_base = limit;

    if (best_len >= STD_MIN_MATCH) {

        /* We're continuing search (lazy evaluation). */

        uint32_t hash;

        uint32_t pos;

        /* Find a most distant chain starting from scan with index=1 (index=0 corresponds

         * to cur_match). We cannot use s->prev[strstart+1,...] immediately, because

         * these strings are not yet inserted into the hash table.

         */

        // use update_hash_roll for deflate_slow

        hash = update_hash_roll(0, scan[1]);

        hash = update_hash_roll(hash, scan[2]);

        for (uint32_t i = 3; i <= best_len; i++) {

            // use update_hash_roll for deflate_slow

            hash = update_hash_roll(hash, scan[i]);

            /* If we're starting with best_len >= 3, we can use offset search. */

            pos = head[hash];

            if (UNLIKELY(pos < cur_match)) {

                match_offset = i - 2;

                cur_match = pos;

            }

        }

        /* Update offset-dependent variables */

        limit = limit_base+match_offset;

        if (UNLIKELY(cur_match <= limit))

            return best_len;

        mbase_start -= match_offset;

        mbase_end -= match_offset;

    }

#else

    early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL;

#endif

    Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead");

    for (;;) {

        if (UNLIKELY(cur_match >= strstart))

            break;

        /* Skip to next match if the match length cannot increase or if the match length is

         * less than 2. Note that the checks below for insufficient lookahead only occur

         * occasionally for performance reasons.

         * Therefore uninitialized memory will be accessed and conditional jumps will be made

         * that depend on those values. However the length of the match is limited to the

         * lookahead, so the output of deflate is not affected by the uninitialized values.

         */

        uint32_t len;

        if (best_len < sizeof(uint64_t)) {

            uint64_t cand_start = zng_memread_8(mbase_start + cur_match);

            if (scan_start != cand_start) {

                /* Peel the first candidate out of the loop. A full 8-byte match falls straight

                 * through to compare256, and single-candidate chains (barely-compressible data)

                 * run with no loop overhead. */

                uint64_t first_mask = zng_first_bytes_mask64(best_len + 1);

                uint64_t diff = scan_start ^ cand_start;

                /* A candidate beats best_len only when its first best_len+1 bytes match, i.e.

                 * those bytes of the XOR are zero. The masked test rejects without running ctz. */

                if (UNLIKELY((diff & first_mask) == 0)) {

                    len = zng_first_diff_byte64(diff);

                    goto short_match_accept;

                }

                if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit)

                    return best_len;

                cand_start = zng_memread_8(mbase_start + cur_match);

                if (scan_start != cand_start) {

                    /* Walk the remaining candidates with the chain advance kept inline. */

                    for (;;) {

                        diff = scan_start ^ cand_start;

                        if (UNLIKELY((diff & first_mask) == 0)) {

                            len = zng_first_diff_byte64(diff);

                            goto short_match_accept;

                        }

                        if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit)

                            return best_len;

                        cand_start = zng_memread_8(mbase_start + cur_match);

                        if (scan_start == cand_start)

                            break;

                    }

                }

            }

            /* All 8 bytes match, fallthrough to compare256 for the tail. */

        } else {

            /* Pre-filter the candidate on the start and end sentinels before compare256 using

             * simple 8-byte comparison since best_len >= 8. */

            for (;;) {

                /* First check the end of the candidate at best_len+1 due to the higher

                 * likelihood of a mismatch. */

                if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 &&

                    zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0)

                    break;

                GOTO_NEXT_CHAIN;

            }

        }

        len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2;

        Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan");

        if (len > best_len)

short_match_accept:

        {

            uint32_t match_start = cur_match - match_offset;

            s->match_start = match_start;

            /* Do not look for better matches if the current match reaches

             * or exceeds the end of the input.

             */

            if (UNLIKELY(len >= lookahead))

                return lookahead;

            if (UNLIKELY(len >= nice_match))

                return len;

            best_len = len;

            offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0;

            scan_end = zng_memread_8(scan+offset);

#ifdef LONGEST_MATCH_ROLL

            /* Look for a better string offset */

            if (UNLIKELY(len > STD_MIN_MATCH && match_start + len < strstart)) {

                const unsigned char *scan_endstr;

                uint32_t hash;

                uint32_t pos, next_pos;

                /* Go back to offset 0 */

                cur_match -= match_offset;

                match_offset = 0;

                next_pos = cur_match;

                for (uint32_t i = 0; i <= len - STD_MIN_MATCH; i++) {

                    pos = prev[(cur_match + i) & wmask];

                    if (UNLIKELY(pos < next_pos)) {

                        /* Hash chain is more distant, use it */

                        if (UNLIKELY(pos <= limit_base + i))

                            return best_len;

                        next_pos = pos;

                        match_offset = i;

                    }

                }

                /* Switch cur_match to next_pos chain */

                cur_match = next_pos;

                /* Try hash head at len-(STD_MIN_MATCH-1) position to see if we could get

                 * a better cur_match at the end of string. Using (STD_MIN_MATCH-1) lets

                 * us include one more byte into hash - the byte which will be checked

                 * in main loop now, and which allows to grow match by 1.

                 */

                scan_endstr = scan + len - (STD_MIN_MATCH-1);

                hash = update_hash_roll(0, scan_endstr[0]);

                hash = update_hash_roll(hash, scan_endstr[1]);

                hash = update_hash_roll(hash, scan_endstr[2]);

                pos = head[hash];

                if (UNLIKELY(pos < cur_match)) {

                    match_offset = len - (STD_MIN_MATCH-1);

                    if (pos <= limit_base + match_offset)

                        return best_len;

                    cur_match = pos;

                }

                /* Update offset-dependent variables */

                limit = limit_base+match_offset;

                mbase_start = window-match_offset;

                mbase_end = (mbase_start+offset);

                continue;

            }

#endif

            mbase_end = (mbase_start+offset);

        }

#ifndef LONGEST_MATCH_ROLL

        else if (UNLIKELY(early_exit)) {

            /* The probability of finding a match later if we here is pretty low, so for

             * performance it's best to outright stop here for the lower compression levels

             */

            break;

        }

#endif

        GOTO_NEXT_CHAIN;

    }

    return best_len;

}

Z_INTERNAL uint32_t LONGEST_MATCH(deflate_state *const s, uint32_t cur_match) {

    const unsigned wmask = W_MASK(s);

    unsigned int strstart = s->strstart;

    const unsigned char *window = s->window;

    const Pos *prev = s->prev;

#ifdef LONGEST_MATCH_ROLL

    const Pos *head = s->head;

#endif

    const unsigned char *scan;

    const unsigned char *mbase_start = window;

    const unsigned char *mbase_end;

    uint32_t limit;

#ifdef LONGEST_MATCH_ROLL

    uint32_t limit_base;

#else

    int32_t early_exit;

#endif

    uint32_t chain_length = s->max_chain_length;

    uint32_t nice_match = (uint32_t)s->nice_match;

    uint32_t best_len, offset;

    uint32_t lookahead = s->lookahead;

    uint32_t match_offset = 0;

    uint64_t scan_start;

    uint64_t scan_end;

    /* The code is optimized for STD_MAX_MATCH-2 multiple of 16. */

    Assert(STD_MAX_MATCH == 258, "Code too clever");

    best_len = s->prev_length ? s->prev_length : STD_MIN_MATCH-1;

    if (UNLIKELY(best_len >= lookahead))

        return lookahead;

    /* Calculate read offset which should only extend an extra byte to find the

     * next best match length. When best_len is shorter than the read width, we

     * diff the mismatched bytes instead.

     */

    offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0;

    scan = window + strstart;

    scan_start = zng_memread_8(scan);

    scan_end = zng_memread_8(scan+offset);

    mbase_end = (mbase_start+offset);

    /* Do not waste too much time if we already have a good match */

    if (UNLIKELY(best_len >= s->good_match))

        chain_length >>= 2;

    /* Stop when cur_match becomes <= limit. To simplify the code,

     * we prevent matches with the string of window index 0

     */

    limit = strstart > MAX_DIST(s) ? (strstart - MAX_DIST(s)) : 0;

#ifdef LONGEST_MATCH_ROLL

    limit_base = limit;

    if (best_len >= STD_MIN_MATCH) {

        /* We're continuing search (lazy evaluation). */

        uint32_t hash;

        uint32_t pos;

        /* Find a most distant chain starting from scan with index=1 (index=0 corresponds

         * to cur_match). We cannot use s->prev[strstart+1,...] immediately, because

         * these strings are not yet inserted into the hash table.

         */

        // use update_hash_roll for deflate_slow

        hash = update_hash_roll(0, scan[1]);

        hash = update_hash_roll(hash, scan[2]);

        for (uint32_t i = 3; i <= best_len; i++) {

            // use update_hash_roll for deflate_slow

            hash = update_hash_roll(hash, scan[i]);

            /* If we're starting with best_len >= 3, we can use offset search. */

            pos = head[hash];

            if (UNLIKELY(pos < cur_match)) {

                match_offset = i - 2;

                cur_match = pos;

            }

        }

        /* Update offset-dependent variables */

        limit = limit_base+match_offset;

        if (UNLIKELY(cur_match <= limit))

            return best_len;

        mbase_start -= match_offset;

        mbase_end -= match_offset;

    }

#else

    early_exit = s->level < EARLY_EXIT_TRIGGER_LEVEL;

#endif

    Assert((unsigned long)strstart <= s->window_size - MIN_LOOKAHEAD, "need lookahead");

    for (;;) {

        if (UNLIKELY(cur_match >= strstart))

            break;

        /* Skip to next match if the match length cannot increase or if the match length is

         * less than 2. Note that the checks below for insufficient lookahead only occur

         * occasionally for performance reasons.

         * Therefore uninitialized memory will be accessed and conditional jumps will be made

         * that depend on those values. However the length of the match is limited to the

         * lookahead, so the output of deflate is not affected by the uninitialized values.

         */

        uint32_t len;

        if (best_len < sizeof(uint64_t)) {

            uint64_t cand_start = zng_memread_8(mbase_start + cur_match);

            if (scan_start != cand_start) {

                /* Peel the first candidate out of the loop. A full 8-byte match falls straight

                 * through to compare256, and single-candidate chains (barely-compressible data)

                 * run with no loop overhead. */

                uint64_t first_mask = zng_first_bytes_mask64(best_len + 1);

                uint64_t diff = scan_start ^ cand_start;

                /* A candidate beats best_len only when its first best_len+1 bytes match, i.e.

                 * those bytes of the XOR are zero. The masked test rejects without running ctz. */

                if (UNLIKELY((diff & first_mask) == 0)) {

                    len = zng_first_diff_byte64(diff);

                    goto short_match_accept;

                }

                if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit)

                    return best_len;

                cand_start = zng_memread_8(mbase_start + cur_match);

                if (scan_start != cand_start) {

                    /* Walk the remaining candidates with the chain advance kept inline. */

                    for (;;) {

                        diff = scan_start ^ cand_start;

                        if (UNLIKELY((diff & first_mask) == 0)) {

                            len = zng_first_diff_byte64(diff);

                            goto short_match_accept;

                        }

                        if (--chain_length == 0 || (cur_match = prev[cur_match & wmask]) <= limit)

                            return best_len;

                        cand_start = zng_memread_8(mbase_start + cur_match);

                        if (scan_start == cand_start)

                            break;

                    }

                }

            }

            /* All 8 bytes match, fallthrough to compare256 for the tail. */

        } else {

            /* Pre-filter the candidate on the start and end sentinels before compare256 using

             * simple 8-byte comparison since best_len >= 8. */

            for (;;) {

                /* First check the end of the candidate at best_len+1 due to the higher

                 * likelihood of a mismatch. */

                if (zng_memcmp_8(mbase_end+cur_match, &scan_end) == 0 &&

                    zng_memcmp_8(mbase_start+cur_match, &scan_start) == 0)

                    break;

                GOTO_NEXT_CHAIN;

            }

        }

        len = COMPARE256(scan+2, mbase_start+cur_match+2) + 2;

        Assert(scan+len <= window+(unsigned)(s->window_size-1), "wild scan");

        if (len > best_len)

short_match_accept:

        {

            uint32_t match_start = cur_match - match_offset;

            s->match_start = match_start;

            /* Do not look for better matches if the current match reaches

             * or exceeds the end of the input.

             */

            if (UNLIKELY(len >= lookahead))

                return lookahead;

            if (UNLIKELY(len >= nice_match))

                return len;

            best_len = len;

            offset = best_len >= sizeof(uint64_t) ? best_len - 7 : 0;

            scan_end = zng_memread_8(scan+offset);

#ifdef LONGEST_MATCH_ROLL

            /* Look for a better string offset */

            if (UNLIKELY(len > STD_MIN_MATCH && match_start + len < strstart)) {

                const unsigned char *scan_endstr;

                uint32_t hash;

                uint32_t pos, next_pos;

                /* Go back to offset 0 */

                cur_match -= match_offset;

                match_offset = 0;

                next_pos = cur_match;

                for (uint32_t i = 0; i <= len - STD_MIN_MATCH; i++) {

                    pos = prev[(cur_match + i) & wmask];

                    if (UNLIKELY(pos < next_pos)) {

                        /* Hash chain is more distant, use it */

                        if (UNLIKELY(pos <= limit_base + i))

                            return best_len;

                        next_pos = pos;

                        match_offset = i;

                    }

                }

                /* Switch cur_match to next_pos chain */

                cur_match = next_pos;

                /* Try hash head at len-(STD_MIN_MATCH-1) position to see if we could get

                 * a better cur_match at the end of string. Using (STD_MIN_MATCH-1) lets

                 * us include one more byte into hash - the byte which will be checked

                 * in main loop now, and which allows to grow match by 1.

                 */

                scan_endstr = scan + len - (STD_MIN_MATCH-1);

                hash = update_hash_roll(0, scan_endstr[0]);

                hash = update_hash_roll(hash, scan_endstr[1]);

                hash = update_hash_roll(hash, scan_endstr[2]);

                pos = head[hash];

                if (UNLIKELY(pos < cur_match)) {

                    match_offset = len - (STD_MIN_MATCH-1);

                    if (pos <= limit_base + match_offset)

                        return best_len;

                    cur_match = pos;

                }

                /* Update offset-dependent variables */

                limit = limit_base+match_offset;

                mbase_start = window-match_offset;

                mbase_end = (mbase_start+offset);

                continue;

            }

#endif

            mbase_end = (mbase_start+offset);

        }

#ifndef LONGEST_MATCH_ROLL

        else if (UNLIKELY(early_exit)) {

            /* The probability of finding a match later if we here is pretty low, so for

             * performance it's best to outright stop here for the lower compression levels

             */

            break;

        }

#endif

        GOTO_NEXT_CHAIN;

    }

    return best_len;

}