/*
 * MD4C: Markdown parser for C
 * (http://github.com/mity/md4c)
 *
 * Copyright (c) 2016-2020 Martin Mitas
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 */

#include "md4c.h"

#include <limits.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>


/*****************************
 ***  Miscellaneous Stuff  ***
 *****************************/

#if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 199409L
    /* C89/90 or old compilers in general may not understand "inline". */
    #if defined __GNUC__
        #define inline __inline__
    #elif defined _MSC_VER
        #define inline __inline
    #else
        #define inline
    #endif
#endif

/* Make the UTF-8 support the default. */
#if !defined MD4C_USE_ASCII && !defined MD4C_USE_UTF8 && !defined MD4C_USE_UTF16
    #define MD4C_USE_UTF8
#endif

/* Magic for making wide literals with MD4C_USE_UTF16. */
#ifdef _T
    #undef _T
#endif
#if defined MD4C_USE_UTF16
    #define _T(x)           L##x
#else
    #define _T(x)           x
#endif

/* Misc. macros. */
#define SIZEOF_ARRAY(a)     (sizeof(a) / sizeof(a[0]))

#define STRINGIZE_(x)       #x
#define STRINGIZE(x)        STRINGIZE_(x)

#ifndef TRUE
    #define TRUE            1
    #define FALSE           0
#endif

#define MD_LOG(msg)                                                     \
    do {                                                                \
        if(ctx->parser.debug_log != NULL)                               \
            ctx->parser.debug_log((msg), ctx->userdata);                \
    } while(0)

#ifdef DEBUG
    #define MD_ASSERT(cond)                                             \
            do {                                                        \
                if(!(cond)) {                                           \
                    MD_LOG(__FILE__ ":" STRINGIZE(__LINE__) ": "        \
                           "Assertion '" STRINGIZE(cond) "' failed.");  \
                    exit(1);                                            \
                }                                                       \
            } while(0)

    #define MD_UNREACHABLE()        MD_ASSERT(1 == 0)
#else
    #if defined __GNUC__ && !defined __TINYC__
        #define MD_ASSERT(cond)     do { if(!(cond)) __builtin_unreachable(); } while(0)
        #define MD_UNREACHABLE()    do { __builtin_unreachable(); } while(0)
    #elif defined _MSC_VER  &&  _MSC_VER > 120
        #define MD_ASSERT(cond)     do { __assume(cond); } while(0)
        #define MD_UNREACHABLE()    do { __assume(0); } while(0)
    #else
        #define MD_ASSERT(cond)     do {} while(0)
        #define MD_UNREACHABLE()    do {} while(0)
    #endif
#endif

/* For falling through case labels in switch statements. */
#if defined __clang__ && __clang_major__ >= 12
    #define MD_FALLTHROUGH()        __attribute__((fallthrough))
#elif defined __GNUC__ && __GNUC__ >= 7
    #define MD_FALLTHROUGH()        __attribute__((fallthrough))
#else
    #define MD_FALLTHROUGH()        ((void)0)
#endif

/* Suppress "unused parameter" warnings. */
#define MD_UNUSED(x)                ((void)x)


/************************
 ***  Internal Types  ***
 ************************/

/* These are omnipresent so lets save some typing. */
#define CHAR    MD_CHAR
#define SZ      MD_SIZE
#define OFF     MD_OFFSET

typedef struct MD_MARK_tag MD_MARK;
typedef struct MD_BLOCK_tag MD_BLOCK;
typedef struct MD_CONTAINER_tag MD_CONTAINER;
typedef struct MD_REF_DEF_tag MD_REF_DEF;


/* During analyzes of inline marks, we need to manage some "mark chains",
 * of (yet unresolved) openers. This structure holds start/end of the chain.
 * The chain internals are then realized through MD_MARK::prev and ::next.
 */
typedef struct MD_MARKCHAIN_tag MD_MARKCHAIN;
struct MD_MARKCHAIN_tag {
    int head;   /* Index of first mark in the chain, or -1 if empty. */
    int tail;   /* Index of last mark in the chain, or -1 if empty. */
};

/* Context propagated through all the parsing. */
typedef struct MD_CTX_tag MD_CTX;
struct MD_CTX_tag {
    /* Immutable stuff (parameters of md_parse()). */
    const CHAR* text;
    SZ size;
    MD_PARSER parser;
    void* userdata;

    /* When this is true, it allows some optimizations. */
    int doc_ends_with_newline;

    /* Helper temporary growing buffer. */
    CHAR* buffer;
    unsigned alloc_buffer;

    /* Reference definitions. */
    MD_REF_DEF* ref_defs;
    int n_ref_defs;
    int alloc_ref_defs;
    void** ref_def_hashtable;
    int ref_def_hashtable_size;

    /* Stack of inline/span markers.
     * This is only used for parsing a single block contents but by storing it
     * here we may reuse the stack for subsequent blocks; i.e. we have fewer
     * (re)allocations. */
    MD_MARK* marks;
    int n_marks;
    int alloc_marks;

#if defined MD4C_USE_UTF16
    char mark_char_map[128];
#else
    char mark_char_map[256];
#endif

    /* For resolving of inline spans. */
    MD_MARKCHAIN mark_chains[13];
#define PTR_CHAIN                               (ctx->mark_chains[0])
#define TABLECELLBOUNDARIES                     (ctx->mark_chains[1])
#define ASTERISK_OPENERS_extraword_mod3_0       (ctx->mark_chains[2])
#define ASTERISK_OPENERS_extraword_mod3_1       (ctx->mark_chains[3])
#define ASTERISK_OPENERS_extraword_mod3_2       (ctx->mark_chains[4])
#define ASTERISK_OPENERS_intraword_mod3_0       (ctx->mark_chains[5])
#define ASTERISK_OPENERS_intraword_mod3_1       (ctx->mark_chains[6])
#define ASTERISK_OPENERS_intraword_mod3_2       (ctx->mark_chains[7])
#define UNDERSCORE_OPENERS                      (ctx->mark_chains[8])
#define TILDE_OPENERS_1                         (ctx->mark_chains[9])
#define TILDE_OPENERS_2                         (ctx->mark_chains[10])
#define BRACKET_OPENERS                         (ctx->mark_chains[11])
#define DOLLAR_OPENERS                          (ctx->mark_chains[12])
#define OPENERS_CHAIN_FIRST                     1
#define OPENERS_CHAIN_LAST                      12

    int n_table_cell_boundaries;

    /* For resolving links. */
    int unresolved_link_head;
    int unresolved_link_tail;

    /* For resolving raw HTML. */
    OFF html_comment_horizon;
    OFF html_proc_instr_horizon;
    OFF html_decl_horizon;
    OFF html_cdata_horizon;

    /* For block analysis.
     * Notes:
     *   -- It holds MD_BLOCK as well as MD_LINE structures. After each
     *      MD_BLOCK, its (multiple) MD_LINE(s) follow.
     *   -- For MD_BLOCK_HTML and MD_BLOCK_CODE, MD_VERBATIMLINE(s) are used
     *      instead of MD_LINE(s).
     */
    void* block_bytes;
    MD_BLOCK* current_block;
    int n_block_bytes;
    int alloc_block_bytes;

    /* For container block analysis. */
    MD_CONTAINER* containers;
    int n_containers;
    int alloc_containers;

    /* Minimal indentation to call the block "indented code block". */
    unsigned code_indent_offset;

    /* Contextual info for line analysis. */
    SZ code_fence_length;   /* For checking closing fence length. */
    int html_block_type;    /* For checking closing raw HTML condition. */
    int last_line_has_list_loosening_effect;
    int last_list_item_starts_with_two_blank_lines;
};

enum MD_LINETYPE_tag {
    MD_LINE_BLANK,
    MD_LINE_HR,
    MD_LINE_ATXHEADER,
    MD_LINE_SETEXTHEADER,
    MD_LINE_SETEXTUNDERLINE,
    MD_LINE_INDENTEDCODE,
    MD_LINE_FENCEDCODE,
    MD_LINE_HTML,
    MD_LINE_TEXT,
    MD_LINE_TABLE,
    MD_LINE_TABLEUNDERLINE
};
typedef enum MD_LINETYPE_tag MD_LINETYPE;

typedef struct MD_LINE_ANALYSIS_tag MD_LINE_ANALYSIS;
struct MD_LINE_ANALYSIS_tag {
    MD_LINETYPE type    : 16;
    unsigned data       : 16;
    OFF beg;
    OFF end;
    unsigned indent;        /* Indentation level. */
};

typedef struct MD_LINE_tag MD_LINE;
struct MD_LINE_tag {
    OFF beg;
    OFF end;
};

typedef struct MD_VERBATIMLINE_tag MD_VERBATIMLINE;
struct MD_VERBATIMLINE_tag {
    OFF beg;
    OFF end;
    OFF indent;
};


/*****************
 ***  Helpers  ***
 *****************/

/* Character accessors. */
#define CH(off)                 (ctx->text[(off)])
#define STR(off)                (ctx->text + (off))

/* Character classification.
 * Note we assume ASCII compatibility of code points < 128 here. */
#define ISIN_(ch, ch_min, ch_max)       ((ch_min) <= (unsigned)(ch) && (unsigned)(ch) <= (ch_max))
#define ISANYOF_(ch, palette)           ((ch) != _T('\0')  &&  md_strchr((palette), (ch)) != NULL)
#define ISANYOF2_(ch, ch1, ch2)         ((ch) == (ch1) || (ch) == (ch2))
#define ISANYOF3_(ch, ch1, ch2, ch3)    ((ch) == (ch1) || (ch) == (ch2) || (ch) == (ch3))
#define ISASCII_(ch)                    ((unsigned)(ch) <= 127)
#define ISBLANK_(ch)                    (ISANYOF2_((ch), _T(' '), _T('\t')))
#define ISNEWLINE_(ch)                  (ISANYOF2_((ch), _T('\r'), _T('\n')))
#define ISWHITESPACE_(ch)               (ISBLANK_(ch) || ISANYOF2_((ch), _T('\v'), _T('\f')))
#define ISCNTRL_(ch)                    ((unsigned)(ch) <= 31 || (unsigned)(ch) == 127)
#define ISPUNCT_(ch)                    (ISIN_(ch, 33, 47) || ISIN_(ch, 58, 64) || ISIN_(ch, 91, 96) || ISIN_(ch, 123, 126))
#define ISUPPER_(ch)                    (ISIN_(ch, _T('A'), _T('Z')))
#define ISLOWER_(ch)                    (ISIN_(ch, _T('a'), _T('z')))
#define ISALPHA_(ch)                    (ISUPPER_(ch) || ISLOWER_(ch))
#define ISDIGIT_(ch)                    (ISIN_(ch, _T('0'), _T('9')))
#define ISXDIGIT_(ch)                   (ISDIGIT_(ch) || ISIN_(ch, _T('A'), _T('F')) || ISIN_(ch, _T('a'), _T('f')))
#define ISALNUM_(ch)                    (ISALPHA_(ch) || ISDIGIT_(ch))

#define ISANYOF(off, palette)           ISANYOF_(CH(off), (palette))
#define ISANYOF2(off, ch1, ch2)         ISANYOF2_(CH(off), (ch1), (ch2))
#define ISANYOF3(off, ch1, ch2, ch3)    ISANYOF3_(CH(off), (ch1), (ch2), (ch3))
#define ISASCII(off)                    ISASCII_(CH(off))
#define ISBLANK(off)                    ISBLANK_(CH(off))
#define ISNEWLINE(off)                  ISNEWLINE_(CH(off))
#define ISWHITESPACE(off)               ISWHITESPACE_(CH(off))
#define ISCNTRL(off)                    ISCNTRL_(CH(off))
#define ISPUNCT(off)                    ISPUNCT_(CH(off))
#define ISUPPER(off)                    ISUPPER_(CH(off))
#define ISLOWER(off)                    ISLOWER_(CH(off))
#define ISALPHA(off)                    ISALPHA_(CH(off))
#define ISDIGIT(off)                    ISDIGIT_(CH(off))
#define ISXDIGIT(off)                   ISXDIGIT_(CH(off))
#define ISALNUM(off)                    ISALNUM_(CH(off))


#if defined MD4C_USE_UTF16
    #define md_strchr wcschr
#else
    #define md_strchr strchr
#endif


/* Case insensitive check of string equality. */
static inline int
md_ascii_case_eq(const CHAR* s1, const CHAR* s2, SZ n)
{
    OFF i;
    for(i = 0; i < n; i++) {
        CHAR ch1 = s1[i];
        CHAR ch2 = s2[i];

        if(ISLOWER_(ch1))
            ch1 += ('A'-'a');
        if(ISLOWER_(ch2))
            ch2 += ('A'-'a');
        if(ch1 != ch2)
            return FALSE;
    }
    return TRUE;
}

static inline int
md_ascii_eq(const CHAR* s1, const CHAR* s2, SZ n)
{
    return memcmp(s1, s2, n * sizeof(CHAR)) == 0;
}

static int
md_text_with_null_replacement(MD_CTX* ctx, MD_TEXTTYPE type, const CHAR* str, SZ size)
{
    OFF off = 0;
    int ret = 0;

    while(1) {
        while(off < size  &&  str[off] != _T('\0'))
            off++;

        if(off > 0) {
            ret = ctx->parser.text(type, str, off, ctx->userdata);
            if(ret != 0)
                return ret;

            str += off;
            size -= off;
            off = 0;
        }

        if(off >= size)
            return 0;

        ret = ctx->parser.text(MD_TEXT_NULLCHAR, _T(""), 1, ctx->userdata);
        if(ret != 0)
            return ret;
        off++;
    }
}


#define MD_CHECK(func)                                                      \
    do {                                                                    \
        ret = (func);                                                       \
        if(ret < 0)                                                         \
            goto abort;                                                     \
    } while(0)


#define MD_TEMP_BUFFER(sz)                                                  \
    do {                                                                    \
        if(sz > ctx->alloc_buffer) {                                        \
            CHAR* new_buffer;                                               \
            SZ new_size = ((sz) + (sz) / 2 + 128) & ~127;                   \
                                                                            \
            new_buffer = realloc(ctx->buffer, new_size);                    \
            if(new_buffer == NULL) {                                        \
                MD_LOG("realloc() failed.");                                \
                ret = -1;                                                   \
                goto abort;                                                 \
            }                                                               \
                                                                            \
            ctx->buffer = new_buffer;                                       \
            ctx->alloc_buffer = new_size;                                   \
        }                                                                   \
    } while(0)


#define MD_ENTER_BLOCK(type, arg)                                           \
    do {                                                                    \
        ret = ctx->parser.enter_block((type), (arg), ctx->userdata);        \
        if(ret != 0) {                                                      \
            MD_LOG("Aborted from enter_block() callback.");                 \
            goto abort;                                                     \
        }                                                                   \
    } while(0)

#define MD_LEAVE_BLOCK(type, arg)                                           \
    do {                                                                    \
        ret = ctx->parser.leave_block((type), (arg), ctx->userdata);        \
        if(ret != 0) {                                                      \
            MD_LOG("Aborted from leave_block() callback.");                 \
            goto abort;                                                     \
        }                                                                   \
    } while(0)

#define MD_ENTER_SPAN(type, arg)                                            \
    do {                                                                    \
        ret = ctx->parser.enter_span((type), (arg), ctx->userdata);         \
        if(ret != 0) {                                                      \
            MD_LOG("Aborted from enter_span() callback.");                  \
            goto abort;                                                     \
        }                                                                   \
    } while(0)

#define MD_LEAVE_SPAN(type, arg)                                            \
    do {                                                                    \
        ret = ctx->parser.leave_span((type), (arg), ctx->userdata);         \
        if(ret != 0) {                                                      \
            MD_LOG("Aborted from leave_span() callback.");                  \
            goto abort;                                                     \
        }                                                                   \
    } while(0)

#define MD_TEXT(type, str, size)                                            \
    do {                                                                    \
        if(size > 0) {                                                      \
            ret = ctx->parser.text((type), (str), (size), ctx->userdata);   \
            if(ret != 0) {                                                  \
                MD_LOG("Aborted from text() callback.");                    \
                goto abort;                                                 \
            }                                                               \
        }                                                                   \
    } while(0)

#define MD_TEXT_INSECURE(type, str, size)                                   \
    do {                                                                    \
        if(size > 0) {                                                      \
            ret = md_text_with_null_replacement(ctx, type, str, size);      \
            if(ret != 0) {                                                  \
                MD_LOG("Aborted from text() callback.");                    \
                goto abort;                                                 \
            }                                                               \
        }                                                                   \
    } while(0)


/* If the offset falls into a gap between line, we return the following
 * line. */
static const MD_LINE*
md_lookup_line(OFF off, const MD_LINE* lines, int n_lines)
{
    int lo, hi;
    int pivot;
    const MD_LINE* line;

    lo = 0;
    hi = n_lines - 1;
    while(lo <= hi) {
        pivot = (lo + hi) / 2;
        line = &lines[pivot];

        if(off < line->beg) {
            hi = pivot - 1;
            if(hi < 0  ||  lines[hi].end <= off)
                return line;
        } else if(off > line->end) {
            lo = pivot + 1;
        } else {
            return line;
        }
    }

    return NULL;
}


/*************************
 ***  Unicode Support  ***
 *************************/

typedef struct MD_UNICODE_FOLD_INFO_tag MD_UNICODE_FOLD_INFO;
struct MD_UNICODE_FOLD_INFO_tag {
    unsigned codepoints[3];
    unsigned n_codepoints;
};


#if defined MD4C_USE_UTF16 || defined MD4C_USE_UTF8
    /* Binary search over sorted "map" of codepoints. Consecutive sequences
     * of codepoints may be encoded in the map by just using the
     * (MIN_CODEPOINT | 0x40000000) and (MAX_CODEPOINT | 0x80000000).
     *
     * Returns index of the found record in the map (in the case of ranges,
     * the minimal value is used); or -1 on failure. */
    static int
    md_unicode_bsearch__(unsigned codepoint, const unsigned* map, size_t map_size)
    {
        int beg, end;
        int pivot_beg, pivot_end;

        beg = 0;
        end = (int) map_size-1;
        while(beg <= end) {
            /* Pivot may be a range, not just a single value. */
            pivot_beg = pivot_end = (beg + end) / 2;
            if(map[pivot_end] & 0x40000000)
                pivot_end++;
            if(map[pivot_beg] & 0x80000000)
                pivot_beg--;

            if(codepoint < (map[pivot_beg] & 0x00ffffff))
                end = pivot_beg - 1;
            else if(codepoint > (map[pivot_end] & 0x00ffffff))
                beg = pivot_end + 1;
            else
                return pivot_beg;
        }

        return -1;
    }

    static int
    md_is_unicode_whitespace__(unsigned codepoint)
    {
#define R(cp_min, cp_max)   ((cp_min) | 0x40000000), ((cp_max) | 0x80000000)
#define S(cp)               (cp)
        /* Unicode "Zs" category.
         * (generated by scripts/build_whitespace_map.py) */
        static const unsigned WHITESPACE_MAP[] = {
            S(0x0020), S(0x00a0), S(0x1680), R(0x2000,0x200a), S(0x202f), S(0x205f), S(0x3000)
        };
#undef R
#undef S

        /* The ASCII ones are the most frequently used ones, also CommonMark
         * specification requests few more in this range. */
        if(codepoint <= 0x7f)
            return ISWHITESPACE_(codepoint);

        return (md_unicode_bsearch__(codepoint, WHITESPACE_MAP, SIZEOF_ARRAY(WHITESPACE_MAP)) >= 0);
    }

    static int
    md_is_unicode_punct__(unsigned codepoint)
    {
#define R(cp_min, cp_max)   ((cp_min) | 0x40000000), ((cp_max) | 0x80000000)
#define S(cp)               (cp)
        /* Unicode "Pc", "Pd", "Pe", "Pf", "Pi", "Po", "Ps" categories.
         * (generated by scripts/build_punct_map.py) */
        static const unsigned PUNCT_MAP[] = {
            R(0x0021,0x0023), R(0x0025,0x002a), R(0x002c,0x002f), R(0x003a,0x003b), R(0x003f,0x0040),
            R(0x005b,0x005d), S(0x005f), S(0x007b), S(0x007d), S(0x00a1), S(0x00a7), S(0x00ab), R(0x00b6,0x00b7),
            S(0x00bb), S(0x00bf), S(0x037e), S(0x0387), R(0x055a,0x055f), R(0x0589,0x058a), S(0x05be), S(0x05c0),
            S(0x05c3), S(0x05c6), R(0x05f3,0x05f4), R(0x0609,0x060a), R(0x060c,0x060d), S(0x061b), R(0x061e,0x061f),
            R(0x066a,0x066d), S(0x06d4), R(0x0700,0x070d), R(0x07f7,0x07f9), R(0x0830,0x083e), S(0x085e),
            R(0x0964,0x0965), S(0x0970), S(0x09fd), S(0x0a76), S(0x0af0), S(0x0c77), S(0x0c84), S(0x0df4), S(0x0e4f),
            R(0x0e5a,0x0e5b), R(0x0f04,0x0f12), S(0x0f14), R(0x0f3a,0x0f3d), S(0x0f85), R(0x0fd0,0x0fd4),
            R(0x0fd9,0x0fda), R(0x104a,0x104f), S(0x10fb), R(0x1360,0x1368), S(0x1400), S(0x166e), R(0x169b,0x169c),
            R(0x16eb,0x16ed), R(0x1735,0x1736), R(0x17d4,0x17d6), R(0x17d8,0x17da), R(0x1800,0x180a),
            R(0x1944,0x1945), R(0x1a1e,0x1a1f), R(0x1aa0,0x1aa6), R(0x1aa8,0x1aad), R(0x1b5a,0x1b60),
            R(0x1bfc,0x1bff), R(0x1c3b,0x1c3f), R(0x1c7e,0x1c7f), R(0x1cc0,0x1cc7), S(0x1cd3), R(0x2010,0x2027),
            R(0x2030,0x2043), R(0x2045,0x2051), R(0x2053,0x205e), R(0x207d,0x207e), R(0x208d,0x208e),
            R(0x2308,0x230b), R(0x2329,0x232a), R(0x2768,0x2775), R(0x27c5,0x27c6), R(0x27e6,0x27ef),
            R(0x2983,0x2998), R(0x29d8,0x29db), R(0x29fc,0x29fd), R(0x2cf9,0x2cfc), R(0x2cfe,0x2cff), S(0x2d70),
            R(0x2e00,0x2e2e), R(0x2e30,0x2e4f), S(0x2e52), R(0x3001,0x3003), R(0x3008,0x3011), R(0x3014,0x301f),
            S(0x3030), S(0x303d), S(0x30a0), S(0x30fb), R(0xa4fe,0xa4ff), R(0xa60d,0xa60f), S(0xa673), S(0xa67e),
            R(0xa6f2,0xa6f7), R(0xa874,0xa877), R(0xa8ce,0xa8cf), R(0xa8f8,0xa8fa), S(0xa8fc), R(0xa92e,0xa92f),
            S(0xa95f), R(0xa9c1,0xa9cd), R(0xa9de,0xa9df), R(0xaa5c,0xaa5f), R(0xaade,0xaadf), R(0xaaf0,0xaaf1),
            S(0xabeb), R(0xfd3e,0xfd3f), R(0xfe10,0xfe19), R(0xfe30,0xfe52), R(0xfe54,0xfe61), S(0xfe63), S(0xfe68),
            R(0xfe6a,0xfe6b), R(0xff01,0xff03), R(0xff05,0xff0a), R(0xff0c,0xff0f), R(0xff1a,0xff1b),
            R(0xff1f,0xff20), R(0xff3b,0xff3d), S(0xff3f), S(0xff5b), S(0xff5d), R(0xff5f,0xff65), R(0x10100,0x10102),
            S(0x1039f), S(0x103d0), S(0x1056f), S(0x10857), S(0x1091f), S(0x1093f), R(0x10a50,0x10a58), S(0x10a7f),
            R(0x10af0,0x10af6), R(0x10b39,0x10b3f), R(0x10b99,0x10b9c), S(0x10ead), R(0x10f55,0x10f59),
            R(0x11047,0x1104d), R(0x110bb,0x110bc), R(0x110be,0x110c1), R(0x11140,0x11143), R(0x11174,0x11175),
            R(0x111c5,0x111c8), S(0x111cd), S(0x111db), R(0x111dd,0x111df), R(0x11238,0x1123d), S(0x112a9),
            R(0x1144b,0x1144f), R(0x1145a,0x1145b), S(0x1145d), S(0x114c6), R(0x115c1,0x115d7), R(0x11641,0x11643),
            R(0x11660,0x1166c), R(0x1173c,0x1173e), S(0x1183b), R(0x11944,0x11946), S(0x119e2), R(0x11a3f,0x11a46),
            R(0x11a9a,0x11a9c), R(0x11a9e,0x11aa2), R(0x11c41,0x11c45), R(0x11c70,0x11c71), R(0x11ef7,0x11ef8),
            S(0x11fff), R(0x12470,0x12474), R(0x16a6e,0x16a6f), S(0x16af5), R(0x16b37,0x16b3b), S(0x16b44),
            R(0x16e97,0x16e9a), S(0x16fe2), S(0x1bc9f), R(0x1da87,0x1da8b), R(0x1e95e,0x1e95f)
        };
#undef R
#undef S

        /* The ASCII ones are the most frequently used ones, also CommonMark
         * specification requests few more in this range. */
        if(codepoint <= 0x7f)
            return ISPUNCT_(codepoint);

        return (md_unicode_bsearch__(codepoint, PUNCT_MAP, SIZEOF_ARRAY(PUNCT_MAP)) >= 0);
    }

    static void
    md_get_unicode_fold_info(unsigned codepoint, MD_UNICODE_FOLD_INFO* info)
    {
#define R(cp_min, cp_max)   ((cp_min) | 0x40000000), ((cp_max) | 0x80000000)
#define S(cp)               (cp)
        /* Unicode "Pc", "Pd", "Pe", "Pf", "Pi", "Po", "Ps" categories.
         * (generated by scripts/build_folding_map.py) */
        static const unsigned FOLD_MAP_1[] = {
            R(0x0041,0x005a), S(0x00b5), R(0x00c0,0x00d6), R(0x00d8,0x00de), R(0x0100,0x012e), R(0x0132,0x0136),
            R(0x0139,0x0147), R(0x014a,0x0176), S(0x0178), R(0x0179,0x017d), S(0x017f), S(0x0181), S(0x0182),
            S(0x0184), S(0x0186), S(0x0187), S(0x0189), S(0x018a), S(0x018b), S(0x018e), S(0x018f), S(0x0190),
            S(0x0191), S(0x0193), S(0x0194), S(0x0196), S(0x0197), S(0x0198), S(0x019c), S(0x019d), S(0x019f),
            R(0x01a0,0x01a4), S(0x01a6), S(0x01a7), S(0x01a9), S(0x01ac), S(0x01ae), S(0x01af), S(0x01b1), S(0x01b2),
            S(0x01b3), S(0x01b5), S(0x01b7), S(0x01b8), S(0x01bc), S(0x01c4), S(0x01c5), S(0x01c7), S(0x01c8),
            S(0x01ca), R(0x01cb,0x01db), R(0x01de,0x01ee), S(0x01f1), S(0x01f2), S(0x01f4), S(0x01f6), S(0x01f7),
            R(0x01f8,0x021e), S(0x0220), R(0x0222,0x0232), S(0x023a), S(0x023b), S(0x023d), S(0x023e), S(0x0241),
            S(0x0243), S(0x0244), S(0x0245), R(0x0246,0x024e), S(0x0345), S(0x0370), S(0x0372), S(0x0376), S(0x037f),
            S(0x0386), R(0x0388,0x038a), S(0x038c), S(0x038e), S(0x038f), R(0x0391,0x03a1), R(0x03a3,0x03ab),
            S(0x03c2), S(0x03cf), S(0x03d0), S(0x03d1), S(0x03d5), S(0x03d6), R(0x03d8,0x03ee), S(0x03f0), S(0x03f1),
            S(0x03f4), S(0x03f5), S(0x03f7), S(0x03f9), S(0x03fa), R(0x03fd,0x03ff), R(0x0400,0x040f),
            R(0x0410,0x042f), R(0x0460,0x0480), R(0x048a,0x04be), S(0x04c0), R(0x04c1,0x04cd), R(0x04d0,0x052e),
            R(0x0531,0x0556), R(0x10a0,0x10c5), S(0x10c7), S(0x10cd), R(0x13f8,0x13fd), S(0x1c80), S(0x1c81),
            S(0x1c82), S(0x1c83), S(0x1c84), S(0x1c85), S(0x1c86), S(0x1c87), S(0x1c88), R(0x1c90,0x1cba),
            R(0x1cbd,0x1cbf), R(0x1e00,0x1e94), S(0x1e9b), R(0x1ea0,0x1efe), R(0x1f08,0x1f0f), R(0x1f18,0x1f1d),
            R(0x1f28,0x1f2f), R(0x1f38,0x1f3f), R(0x1f48,0x1f4d), S(0x1f59), S(0x1f5b), S(0x1f5d), S(0x1f5f),
            R(0x1f68,0x1f6f), S(0x1fb8), S(0x1fb9), S(0x1fba), S(0x1fbb), S(0x1fbe), R(0x1fc8,0x1fcb), S(0x1fd8),
            S(0x1fd9), S(0x1fda), S(0x1fdb), S(0x1fe8), S(0x1fe9), S(0x1fea), S(0x1feb), S(0x1fec), S(0x1ff8),
            S(0x1ff9), S(0x1ffa), S(0x1ffb), S(0x2126), S(0x212a), S(0x212b), S(0x2132), R(0x2160,0x216f), S(0x2183),
            R(0x24b6,0x24cf), R(0x2c00,0x2c2e), S(0x2c60), S(0x2c62), S(0x2c63), S(0x2c64), R(0x2c67,0x2c6b),
            S(0x2c6d), S(0x2c6e), S(0x2c6f), S(0x2c70), S(0x2c72), S(0x2c75), S(0x2c7e), S(0x2c7f), R(0x2c80,0x2ce2),
            S(0x2ceb), S(0x2ced), S(0x2cf2), R(0xa640,0xa66c), R(0xa680,0xa69a), R(0xa722,0xa72e), R(0xa732,0xa76e),
            S(0xa779), S(0xa77b), S(0xa77d), R(0xa77e,0xa786), S(0xa78b), S(0xa78d), S(0xa790), S(0xa792),
            R(0xa796,0xa7a8), S(0xa7aa), S(0xa7ab), S(0xa7ac), S(0xa7ad), S(0xa7ae), S(0xa7b0), S(0xa7b1), S(0xa7b2),
            S(0xa7b3), R(0xa7b4,0xa7be), S(0xa7c2), S(0xa7c4), S(0xa7c5), S(0xa7c6), S(0xa7c7), S(0xa7c9), S(0xa7f5),
            R(0xab70,0xabbf), R(0xff21,0xff3a), R(0x10400,0x10427), R(0x104b0,0x104d3), R(0x10c80,0x10cb2),
            R(0x118a0,0x118bf), R(0x16e40,0x16e5f), R(0x1e900,0x1e921)
        };
        static const unsigned FOLD_MAP_1_DATA[] = {
            0x0061, 0x007a, 0x03bc, 0x00e0, 0x00f6, 0x00f8, 0x00fe, 0x0101, 0x012f, 0x0133, 0x0137, 0x013a, 0x0148,
            0x014b, 0x0177, 0x00ff, 0x017a, 0x017e, 0x0073, 0x0253, 0x0183, 0x0185, 0x0254, 0x0188, 0x0256, 0x0257,
            0x018c, 0x01dd, 0x0259, 0x025b, 0x0192, 0x0260, 0x0263, 0x0269, 0x0268, 0x0199, 0x026f, 0x0272, 0x0275,
            0x01a1, 0x01a5, 0x0280, 0x01a8, 0x0283, 0x01ad, 0x0288, 0x01b0, 0x028a, 0x028b, 0x01b4, 0x01b6, 0x0292,
            0x01b9, 0x01bd, 0x01c6, 0x01c6, 0x01c9, 0x01c9, 0x01cc, 0x01cc, 0x01dc, 0x01df, 0x01ef, 0x01f3, 0x01f3,
            0x01f5, 0x0195, 0x01bf, 0x01f9, 0x021f, 0x019e, 0x0223, 0x0233, 0x2c65, 0x023c, 0x019a, 0x2c66, 0x0242,
            0x0180, 0x0289, 0x028c, 0x0247, 0x024f, 0x03b9, 0x0371, 0x0373, 0x0377, 0x03f3, 0x03ac, 0x03ad, 0x03af,
            0x03cc, 0x03cd, 0x03ce, 0x03b1, 0x03c1, 0x03c3, 0x03cb, 0x03c3, 0x03d7, 0x03b2, 0x03b8, 0x03c6, 0x03c0,
            0x03d9, 0x03ef, 0x03ba, 0x03c1, 0x03b8, 0x03b5, 0x03f8, 0x03f2, 0x03fb, 0x037b, 0x037d, 0x0450, 0x045f,
            0x0430, 0x044f, 0x0461, 0x0481, 0x048b, 0x04bf, 0x04cf, 0x04c2, 0x04ce, 0x04d1, 0x052f, 0x0561, 0x0586,
            0x2d00, 0x2d25, 0x2d27, 0x2d2d, 0x13f0, 0x13f5, 0x0432, 0x0434, 0x043e, 0x0441, 0x0442, 0x0442, 0x044a,
            0x0463, 0xa64b, 0x10d0, 0x10fa, 0x10fd, 0x10ff, 0x1e01, 0x1e95, 0x1e61, 0x1ea1, 0x1eff, 0x1f00, 0x1f07,
            0x1f10, 0x1f15, 0x1f20, 0x1f27, 0x1f30, 0x1f37, 0x1f40, 0x1f45, 0x1f51, 0x1f53, 0x1f55, 0x1f57, 0x1f60,
            0x1f67, 0x1fb0, 0x1fb1, 0x1f70, 0x1f71, 0x03b9, 0x1f72, 0x1f75, 0x1fd0, 0x1fd1, 0x1f76, 0x1f77, 0x1fe0,
            0x1fe1, 0x1f7a, 0x1f7b, 0x1fe5, 0x1f78, 0x1f79, 0x1f7c, 0x1f7d, 0x03c9, 0x006b, 0x00e5, 0x214e, 0x2170,
            0x217f, 0x2184, 0x24d0, 0x24e9, 0x2c30, 0x2c5e, 0x2c61, 0x026b, 0x1d7d, 0x027d, 0x2c68, 0x2c6c, 0x0251,
            0x0271, 0x0250, 0x0252, 0x2c73, 0x2c76, 0x023f, 0x0240, 0x2c81, 0x2ce3, 0x2cec, 0x2cee, 0x2cf3, 0xa641,
            0xa66d, 0xa681, 0xa69b, 0xa723, 0xa72f, 0xa733, 0xa76f, 0xa77a, 0xa77c, 0x1d79, 0xa77f, 0xa787, 0xa78c,
            0x0265, 0xa791, 0xa793, 0xa797, 0xa7a9, 0x0266, 0x025c, 0x0261, 0x026c, 0x026a, 0x029e, 0x0287, 0x029d,
            0xab53, 0xa7b5, 0xa7bf, 0xa7c3, 0xa794, 0x0282, 0x1d8e, 0xa7c8, 0xa7ca, 0xa7f6, 0x13a0, 0x13ef, 0xff41,
            0xff5a, 0x10428, 0x1044f, 0x104d8, 0x104fb, 0x10cc0, 0x10cf2, 0x118c0, 0x118df, 0x16e60, 0x16e7f, 0x1e922,
            0x1e943
        };
        static const unsigned FOLD_MAP_2[] = {
            S(0x00df), S(0x0130), S(0x0149), S(0x01f0), S(0x0587), S(0x1e96), S(0x1e97), S(0x1e98), S(0x1e99),
            S(0x1e9a), S(0x1e9e), S(0x1f50), R(0x1f80,0x1f87), R(0x1f88,0x1f8f), R(0x1f90,0x1f97), R(0x1f98,0x1f9f),
            R(0x1fa0,0x1fa7), R(0x1fa8,0x1faf), S(0x1fb2), S(0x1fb3), S(0x1fb4), S(0x1fb6), S(0x1fbc), S(0x1fc2),
            S(0x1fc3), S(0x1fc4), S(0x1fc6), S(0x1fcc), S(0x1fd6), S(0x1fe4), S(0x1fe6), S(0x1ff2), S(0x1ff3),
            S(0x1ff4), S(0x1ff6), S(0x1ffc), S(0xfb00), S(0xfb01), S(0xfb02), S(0xfb05), S(0xfb06), S(0xfb13),
            S(0xfb14), S(0xfb15), S(0xfb16), S(0xfb17)
        };
        static const unsigned FOLD_MAP_2_DATA[] = {
            0x0073,0x0073, 0x0069,0x0307, 0x02bc,0x006e, 0x006a,0x030c, 0x0565,0x0582, 0x0068,0x0331, 0x0074,0x0308,
            0x0077,0x030a, 0x0079,0x030a, 0x0061,0x02be, 0x0073,0x0073, 0x03c5,0x0313, 0x1f00,0x03b9, 0x1f07,0x03b9,
            0x1f00,0x03b9, 0x1f07,0x03b9, 0x1f20,0x03b9, 0x1f27,0x03b9, 0x1f20,0x03b9, 0x1f27,0x03b9, 0x1f60,0x03b9,
            0x1f67,0x03b9, 0x1f60,0x03b9, 0x1f67,0x03b9, 0x1f70,0x03b9, 0x03b1,0x03b9, 0x03ac,0x03b9, 0x03b1,0x0342,
            0x03b1,0x03b9, 0x1f74,0x03b9, 0x03b7,0x03b9, 0x03ae,0x03b9, 0x03b7,0x0342, 0x03b7,0x03b9, 0x03b9,0x0342,
            0x03c1,0x0313, 0x03c5,0x0342, 0x1f7c,0x03b9, 0x03c9,0x03b9, 0x03ce,0x03b9, 0x03c9,0x0342, 0x03c9,0x03b9,
            0x0066,0x0066, 0x0066,0x0069, 0x0066,0x006c, 0x0073,0x0074, 0x0073,0x0074, 0x0574,0x0576, 0x0574,0x0565,
            0x0574,0x056b, 0x057e,0x0576, 0x0574,0x056d
        };
        static const unsigned FOLD_MAP_3[] = {
            S(0x0390), S(0x03b0), S(0x1f52), S(0x1f54), S(0x1f56), S(0x1fb7), S(0x1fc7), S(0x1fd2), S(0x1fd3),
            S(0x1fd7), S(0x1fe2), S(0x1fe3), S(0x1fe7), S(0x1ff7), S(0xfb03), S(0xfb04)
        };
        static const unsigned FOLD_MAP_3_DATA[] = {
            0x03b9,0x0308,0x0301, 0x03c5,0x0308,0x0301, 0x03c5,0x0313,0x0300, 0x03c5,0x0313,0x0301,
            0x03c5,0x0313,0x0342, 0x03b1,0x0342,0x03b9, 0x03b7,0x0342,0x03b9, 0x03b9,0x0308,0x0300,
            0x03b9,0x0308,0x0301, 0x03b9,0x0308,0x0342, 0x03c5,0x0308,0x0300, 0x03c5,0x0308,0x0301,
            0x03c5,0x0308,0x0342, 0x03c9,0x0342,0x03b9, 0x0066,0x0066,0x0069, 0x0066,0x0066,0x006c
        };
#undef R
#undef S
        static const struct {
            const unsigned* map;
            const unsigned* data;
            size_t map_size;
            unsigned n_codepoints;
        } FOLD_MAP_LIST[] = {
            { FOLD_MAP_1, FOLD_MAP_1_DATA, SIZEOF_ARRAY(FOLD_MAP_1), 1 },
            { FOLD_MAP_2, FOLD_MAP_2_DATA, SIZEOF_ARRAY(FOLD_MAP_2), 2 },
            { FOLD_MAP_3, FOLD_MAP_3_DATA, SIZEOF_ARRAY(FOLD_MAP_3), 3 }
        };

        int i;

        /* Fast path for ASCII characters. */
        if(codepoint <= 0x7f) {
            info->codepoints[0] = codepoint;
            if(ISUPPER_(codepoint))
                info->codepoints[0] += 'a' - 'A';
            info->n_codepoints = 1;
            return;
        }

        /* Try to locate the codepoint in any of the maps. */
        for(i = 0; i < (int) SIZEOF_ARRAY(FOLD_MAP_LIST); i++) {
            int index;

            index = md_unicode_bsearch__(codepoint, FOLD_MAP_LIST[i].map, FOLD_MAP_LIST[i].map_size);
            if(index >= 0) {
                /* Found the mapping. */
                unsigned n_codepoints = FOLD_MAP_LIST[i].n_codepoints;
                const unsigned* map = FOLD_MAP_LIST[i].map;
                const unsigned* codepoints = FOLD_MAP_LIST[i].data + (index * n_codepoints);

                memcpy(info->codepoints, codepoints, sizeof(unsigned) * n_codepoints);
                info->n_codepoints = n_codepoints;

                if(FOLD_MAP_LIST[i].map[index] != codepoint) {
                    /* The found mapping maps whole range of codepoints,
                     * i.e. we have to offset info->codepoints[0] accordingly. */
                    if((map[index] & 0x00ffffff)+1 == codepoints[0]) {
                        /* Alternating type of the range. */
                        info->codepoints[0] = codepoint + ((codepoint & 0x1) == (map[index] & 0x1) ? 1 : 0);
                    } else {
                        /* Range to range kind of mapping. */
                        info->codepoints[0] += (codepoint - (map[index] & 0x00ffffff));
                    }
                }

                return;
            }
        }

        /* No mapping found. Map the codepoint to itself. */
        info->codepoints[0] = codepoint;
        info->n_codepoints = 1;
    }
#endif


#if defined MD4C_USE_UTF16
    #define IS_UTF16_SURROGATE_HI(word)     (((WORD)(word) & 0xfc00) == 0xd800)
    #define IS_UTF16_SURROGATE_LO(word)     (((WORD)(word) & 0xfc00) == 0xdc00)
    #define UTF16_DECODE_SURROGATE(hi, lo)  (0x10000 + ((((unsigned)(hi) & 0x3ff) << 10) | (((unsigned)(lo) & 0x3ff) << 0)))

    static unsigned
    md_decode_utf16le__(const CHAR* str, SZ str_size, SZ* p_size)
    {
        if(IS_UTF16_SURROGATE_HI(str[0])) {
            if(1 < str_size && IS_UTF16_SURROGATE_LO(str[1])) {
                if(p_size != NULL)
                    *p_size = 2;
                return UTF16_DECODE_SURROGATE(str[0], str[1]);
            }
        }

        if(p_size != NULL)
            *p_size = 1;
        return str[0];
    }

    static unsigned
    md_decode_utf16le_before__(MD_CTX* ctx, OFF off)
    {
        if(off > 2 && IS_UTF16_SURROGATE_HI(CH(off-2)) && IS_UTF16_SURROGATE_LO(CH(off-1)))
            return UTF16_DECODE_SURROGATE(CH(off-2), CH(off-1));

        return CH(off);
    }

    /* No whitespace uses surrogates, so no decoding needed here. */
    #define ISUNICODEWHITESPACE_(codepoint) md_is_unicode_whitespace__(codepoint)
    #define ISUNICODEWHITESPACE(off)        md_is_unicode_whitespace__(CH(off))
    #define ISUNICODEWHITESPACEBEFORE(off)  md_is_unicode_whitespace__(CH((off)-1))

    #define ISUNICODEPUNCT(off)             md_is_unicode_punct__(md_decode_utf16le__(STR(off), ctx->size - (off), NULL))
    #define ISUNICODEPUNCTBEFORE(off)       md_is_unicode_punct__(md_decode_utf16le_before__(ctx, off))

    static inline int
    md_decode_unicode(const CHAR* str, OFF off, SZ str_size, SZ* p_char_size)
    {
        return md_decode_utf16le__(str+off, str_size-off, p_char_size);
    }
#elif defined MD4C_USE_UTF8
    #define IS_UTF8_LEAD1(byte)     ((unsigned char)(byte) <= 0x7f)
    #define IS_UTF8_LEAD2(byte)     (((unsigned char)(byte) & 0xe0) == 0xc0)
    #define IS_UTF8_LEAD3(byte)     (((unsigned char)(byte) & 0xf0) == 0xe0)
    #define IS_UTF8_LEAD4(byte)     (((unsigned char)(byte) & 0xf8) == 0xf0)
    #define IS_UTF8_TAIL(byte)      (((unsigned char)(byte) & 0xc0) == 0x80)

    static unsigned
    md_decode_utf8__(const CHAR* str, SZ str_size, SZ* p_size)
    {
        if(!IS_UTF8_LEAD1(str[0])) {
            if(IS_UTF8_LEAD2(str[0])) {
                if(1 < str_size && IS_UTF8_TAIL(str[1])) {
                    if(p_size != NULL)
                        *p_size = 2;

                    return (((unsigned int)str[0] & 0x1f) << 6) |
                           (((unsigned int)str[1] & 0x3f) << 0);
                }
            } else if(IS_UTF8_LEAD3(str[0])) {
                if(2 < str_size && IS_UTF8_TAIL(str[1]) && IS_UTF8_TAIL(str[2])) {
                    if(p_size != NULL)
                        *p_size = 3;

                    return (((unsigned int)str[0] & 0x0f) << 12) |
                           (((unsigned int)str[1] & 0x3f) << 6) |
                           (((unsigned int)str[2] & 0x3f) << 0);
                }
            } else if(IS_UTF8_LEAD4(str[0])) {
                if(3 < str_size && IS_UTF8_TAIL(str[1]) && IS_UTF8_TAIL(str[2]) && IS_UTF8_TAIL(str[3])) {
                    if(p_size != NULL)
                        *p_size = 4;

                    return (((unsigned int)str[0] & 0x07) << 18) |
                           (((unsigned int)str[1] & 0x3f) << 12) |
                           (((unsigned int)str[2] & 0x3f) << 6) |
                           (((unsigned int)str[3] & 0x3f) << 0);
                }
            }
        }

        if(p_size != NULL)
            *p_size = 1;
        return (unsigned) str[0];
    }

    static unsigned
    md_decode_utf8_before__(MD_CTX* ctx, OFF off)
    {
        if(!IS_UTF8_LEAD1(CH(off-1))) {
            if(off > 1 && IS_UTF8_LEAD2(CH(off-2)) && IS_UTF8_TAIL(CH(off-1)))
                return (((unsigned int)CH(off-2) & 0x1f) << 6) |
                       (((unsigned int)CH(off-1) & 0x3f) << 0);

            if(off > 2 && IS_UTF8_LEAD3(CH(off-3)) && IS_UTF8_TAIL(CH(off-2)) && IS_UTF8_TAIL(CH(off-1)))
                return (((unsigned int)CH(off-3) & 0x0f) << 12) |
                       (((unsigned int)CH(off-2) & 0x3f) << 6) |
                       (((unsigned int)CH(off-1) & 0x3f) << 0);

            if(off > 3 && IS_UTF8_LEAD4(CH(off-4)) && IS_UTF8_TAIL(CH(off-3)) && IS_UTF8_TAIL(CH(off-2)) && IS_UTF8_TAIL(CH(off-1)))
                return (((unsigned int)CH(off-4) & 0x07) << 18) |
                       (((unsigned int)CH(off-3) & 0x3f) << 12) |
                       (((unsigned int)CH(off-2) & 0x3f) << 6) |
                       (((unsigned int)CH(off-1) & 0x3f) << 0);
        }

        return (unsigned) CH(off-1);
    }

    #define ISUNICODEWHITESPACE_(codepoint) md_is_unicode_whitespace__(codepoint)
    #define ISUNICODEWHITESPACE(off)        md_is_unicode_whitespace__(md_decode_utf8__(STR(off), ctx->size - (off), NULL))
    #define ISUNICODEWHITESPACEBEFORE(off)  md_is_unicode_whitespace__(md_decode_utf8_before__(ctx, off))

    #define ISUNICODEPUNCT(off)             md_is_unicode_punct__(md_decode_utf8__(STR(off), ctx->size - (off), NULL))
    #define ISUNICODEPUNCTBEFORE(off)       md_is_unicode_punct__(md_decode_utf8_before__(ctx, off))

    static inline unsigned
    md_decode_unicode(const CHAR* str, OFF off, SZ str_size, SZ* p_char_size)
    {
        return md_decode_utf8__(str+off, str_size-off, p_char_size);
    }
#else
    #define ISUNICODEWHITESPACE_(codepoint) ISWHITESPACE_(codepoint)
    #define ISUNICODEWHITESPACE(off)        ISWHITESPACE(off)
    #define ISUNICODEWHITESPACEBEFORE(off)  ISWHITESPACE((off)-1)

    #define ISUNICODEPUNCT(off)             ISPUNCT(off)
    #define ISUNICODEPUNCTBEFORE(off)       ISPUNCT((off)-1)

    static inline void
    md_get_unicode_fold_info(unsigned codepoint, MD_UNICODE_FOLD_INFO* info)
    {
        info->codepoints[0] = codepoint;
        if(ISUPPER_(codepoint))
            info->codepoints[0] += 'a' - 'A';
        info->n_codepoints = 1;
    }

    static inline unsigned
    md_decode_unicode(const CHAR* str, OFF off, SZ str_size, SZ* p_size)
    {
        *p_size = 1;
        return (unsigned) str[off];
    }
#endif


/*************************************
 ***  Helper string manipulations  ***
 *************************************/

/* Fill buffer with copy of the string between 'beg' and 'end' but replace any
 * line breaks with given replacement character.
 *
 * NOTE: Caller is responsible to make sure the buffer is large enough.
 * (Given the output is always shorter then input, (end - beg) is good idea
 * what the caller should allocate.)
 */
static void
md_merge_lines(MD_CTX* ctx, OFF beg, OFF end, const MD_LINE* lines, int n_lines,
               CHAR line_break_replacement_char, CHAR* buffer, SZ* p_size)
{
    CHAR* ptr = buffer;
    int line_index = 0;
    OFF off = beg;

    MD_UNUSED(n_lines);

    while(1) {
        const MD_LINE* line = &lines[line_index];
        OFF line_end = line->end;
        if(end < line_end)
            line_end = end;

        while(off < line_end) {
            *ptr = CH(off);
            ptr++;
            off++;
        }

        if(off >= end) {
            *p_size = (MD_SIZE)(ptr - buffer);
            return;
        }

        *ptr = line_break_replacement_char;
        ptr++;

        line_index++;
        off = lines[line_index].beg;
    }
}

/* Wrapper of md_merge_lines() which allocates new buffer for the output string.
 */
static int
md_merge_lines_alloc(MD_CTX* ctx, OFF beg, OFF end, const MD_LINE* lines, int n_lines,
                    CHAR line_break_replacement_char, CHAR** p_str, SZ* p_size)
{
    CHAR* buffer;

    buffer = (CHAR*) malloc(sizeof(CHAR) * (end - beg));
    if(buffer == NULL) {
        MD_LOG("malloc() failed.");
        return -1;
    }

    md_merge_lines(ctx, beg, end, lines, n_lines,
                line_break_replacement_char, buffer, p_size);

    *p_str = buffer;
    return 0;
}

static OFF
md_skip_unicode_whitespace(const CHAR* label, OFF off, SZ size)
{
    SZ char_size;
    unsigned codepoint;

    while(off < size) {
        codepoint = md_decode_unicode(label, off, size, &char_size);
        if(!ISUNICODEWHITESPACE_(codepoint)  &&  !ISNEWLINE_(label[off]))
            break;
        off += char_size;
    }

    return off;
}


/******************************
 ***  Recognizing raw HTML  ***
 ******************************/

/* md_is_html_tag() may be called when processing inlines (inline raw HTML)
 * or when breaking document to blocks (checking for start of HTML block type 7).
 *
 * When breaking document to blocks, we do not yet know line boundaries, but
 * in that case the whole tag has to live on a single line. We distinguish this
 * by n_lines == 0.
 */
static int
md_is_html_tag(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg, OFF max_end, OFF* p_end)
{
    int attr_state;
    OFF off = beg;
    OFF line_end = (n_lines > 0) ? lines[0].end : ctx->size;
    int i = 0;

    MD_ASSERT(CH(beg) == _T('<'));

    if(off + 1 >= line_end)
        return FALSE;
    off++;

    /* For parsing attributes, we need a little state automaton below.
     * State -1: no attributes are allowed.
     * State 0: attribute could follow after some whitespace.
     * State 1: after a whitespace (attribute name may follow).
     * State 2: after attribute name ('=' MAY follow).
     * State 3: after '=' (value specification MUST follow).
     * State 41: in middle of unquoted attribute value.
     * State 42: in middle of single-quoted attribute value.
     * State 43: in middle of double-quoted attribute value.
     */
    attr_state = 0;

    if(CH(off) == _T('/')) {
        /* Closer tag "</ ... >". No attributes may be present. */
        attr_state = -1;
        off++;
    }

    /* Tag name */
    if(off >= line_end  ||  !ISALPHA(off))
        return FALSE;
    off++;
    while(off < line_end  &&  (ISALNUM(off)  ||  CH(off) == _T('-')))
        off++;

    /* (Optional) attributes (if not closer), (optional) '/' (if not closer)
     * and final '>'. */
    while(1) {
        while(off < line_end  &&  !ISNEWLINE(off)) {
            if(attr_state > 40) {
                if(attr_state == 41 && (ISBLANK(off) || ISANYOF(off, _T("\"'=<>`")))) {
                    attr_state = 0;
                    off--;  /* Put the char back for re-inspection in the new state. */
                } else if(attr_state == 42 && CH(off) == _T('\'')) {
                    attr_state = 0;
                } else if(attr_state == 43 && CH(off) == _T('"')) {
                    attr_state = 0;
                }
                off++;
            } else if(ISWHITESPACE(off)) {
                if(attr_state == 0)
                    attr_state = 1;
                off++;
            } else if(attr_state <= 2 && CH(off) == _T('>')) {
                /* End. */
                goto done;
            } else if(attr_state <= 2 && CH(off) == _T('/') && off+1 < line_end && CH(off+1) == _T('>')) {
                /* End with digraph '/>' */
                off++;
                goto done;
            } else if((attr_state == 1 || attr_state == 2) && (ISALPHA(off) || CH(off) == _T('_') || CH(off) == _T(':'))) {
                off++;
                /* Attribute name */
                while(off < line_end && (ISALNUM(off) || ISANYOF(off, _T("_.:-"))))
                    off++;
                attr_state = 2;
            } else if(attr_state == 2 && CH(off) == _T('=')) {
                /* Attribute assignment sign */
                off++;
                attr_state = 3;
            } else if(attr_state == 3) {
                /* Expecting start of attribute value. */
                if(CH(off) == _T('"'))
                    attr_state = 43;
                else if(CH(off) == _T('\''))
                    attr_state = 42;
                else if(!ISANYOF(off, _T("\"'=<>`"))  &&  !ISNEWLINE(off))
                    attr_state = 41;
                else
                    return FALSE;
                off++;
            } else {
                /* Anything unexpected. */
                return FALSE;
            }
        }

        /* We have to be on a single line. See definition of start condition
         * of HTML block, type 7. */
        if(n_lines == 0)
            return FALSE;

        i++;
        if(i >= n_lines)
            return FALSE;

        off = lines[i].beg;
        line_end = lines[i].end;

        if(attr_state == 0  ||  attr_state == 41)
            attr_state = 1;

        if(off >= max_end)
            return FALSE;
    }

done:
    if(off >= max_end)
        return FALSE;

    *p_end = off+1;
    return TRUE;
}

static int
md_scan_for_html_closer(MD_CTX* ctx, const MD_CHAR* str, MD_SIZE len,
                        const MD_LINE* lines, int n_lines,
                        OFF beg, OFF max_end, OFF* p_end,
                        OFF* p_scan_horizon)
{
    OFF off = beg;
    int i = 0;

    if(off < *p_scan_horizon  &&  *p_scan_horizon >= max_end - len) {
        /* We have already scanned the range up to the max_end so we know
         * there is nothing to see. */
        return FALSE;
    }

    while(TRUE) {
        while(off + len <= lines[i].end  &&  off + len <= max_end) {
            if(md_ascii_eq(STR(off), str, len)) {
                /* Success. */
                *p_end = off + len;
                return TRUE;
            }
            off++;
        }

        i++;
        if(off >= max_end  ||  i >= n_lines) {
            /* Failure. */
            *p_scan_horizon = off;
            return FALSE;
        }

        off = lines[i].beg;
    }
}

static int
md_is_html_comment(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg;

    MD_ASSERT(CH(beg) == _T('<'));

    if(off + 4 >= lines[0].end)
        return FALSE;
    if(CH(off+1) != _T('!')  ||  CH(off+2) != _T('-')  ||  CH(off+3) != _T('-'))
        return FALSE;
    off += 4;

    /* ">" and "->" must not follow the opening. */
    if(off < lines[0].end  &&  CH(off) == _T('>'))
        return FALSE;
    if(off+1 < lines[0].end  &&  CH(off) == _T('-')  &&  CH(off+1) == _T('>'))
        return FALSE;

    /* HTML comment must not contain "--", so we scan just for "--" instead
     * of "-->" and verify manually that '>' follows. */
    if(md_scan_for_html_closer(ctx, _T("--"), 2,
                lines, n_lines, off, max_end, p_end, &ctx->html_comment_horizon))
    {
        if(*p_end < max_end  &&  CH(*p_end) == _T('>')) {
            *p_end = *p_end + 1;
            return TRUE;
        }
    }

    return FALSE;
}

static int
md_is_html_processing_instruction(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg;

    if(off + 2 >= lines[0].end)
        return FALSE;
    if(CH(off+1) != _T('?'))
        return FALSE;
    off += 2;

    return md_scan_for_html_closer(ctx, _T("?>"), 2,
                lines, n_lines, off, max_end, p_end, &ctx->html_proc_instr_horizon);
}

static int
md_is_html_declaration(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg;

    if(off + 2 >= lines[0].end)
        return FALSE;
    if(CH(off+1) != _T('!'))
        return FALSE;
    off += 2;

    /* Declaration name. */
    if(off >= lines[0].end  ||  !ISALPHA(off))
        return FALSE;
    off++;
    while(off < lines[0].end  &&  ISALPHA(off))
        off++;
    if(off < lines[0].end  &&  !ISWHITESPACE(off))
        return FALSE;

    return md_scan_for_html_closer(ctx, _T(">"), 1,
                lines, n_lines, off, max_end, p_end, &ctx->html_decl_horizon);
}

static int
md_is_html_cdata(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg, OFF max_end, OFF* p_end)
{
    static const CHAR open_str[] = _T("<![CDATA[");
    static const SZ open_size = SIZEOF_ARRAY(open_str) - 1;

    OFF off = beg;

    if(off + open_size >= lines[0].end)
        return FALSE;
    if(memcmp(STR(off), open_str, open_size) != 0)
        return FALSE;
    off += open_size;

    if(lines[n_lines-1].end < max_end)
        max_end = lines[n_lines-1].end - 2;

    return md_scan_for_html_closer(ctx, _T("]]>"), 3,
                lines, n_lines, off, max_end, p_end, &ctx->html_cdata_horizon);
}

static int
md_is_html_any(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg, OFF max_end, OFF* p_end)
{
    MD_ASSERT(CH(beg) == _T('<'));
    return (md_is_html_tag(ctx, lines, n_lines, beg, max_end, p_end)  ||
            md_is_html_comment(ctx, lines, n_lines, beg, max_end, p_end)  ||
            md_is_html_processing_instruction(ctx, lines, n_lines, beg, max_end, p_end)  ||
            md_is_html_declaration(ctx, lines, n_lines, beg, max_end, p_end)  ||
            md_is_html_cdata(ctx, lines, n_lines, beg, max_end, p_end));
}


/****************************
 ***  Recognizing Entity  ***
 ****************************/

static int
md_is_hex_entity_contents(MD_CTX* ctx, const CHAR* text, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg;
    MD_UNUSED(ctx);

    while(off < max_end  &&  ISXDIGIT_(text[off])  &&  off - beg <= 8)
        off++;

    if(1 <= off - beg  &&  off - beg <= 6) {
        *p_end = off;
        return TRUE;
    } else {
        return FALSE;
    }
}

static int
md_is_dec_entity_contents(MD_CTX* ctx, const CHAR* text, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg;
    MD_UNUSED(ctx);

    while(off < max_end  &&  ISDIGIT_(text[off])  &&  off - beg <= 8)
        off++;

    if(1 <= off - beg  &&  off - beg <= 7) {
        *p_end = off;
        return TRUE;
    } else {
        return FALSE;
    }
}

static int
md_is_named_entity_contents(MD_CTX* ctx, const CHAR* text, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg;
    MD_UNUSED(ctx);

    if(off < max_end  &&  ISALPHA_(text[off]))
        off++;
    else
        return FALSE;

    while(off < max_end  &&  ISALNUM_(text[off])  &&  off - beg <= 48)
        off++;

    if(2 <= off - beg  &&  off - beg <= 48) {
        *p_end = off;
        return TRUE;
    } else {
        return FALSE;
    }
}

static int
md_is_entity_str(MD_CTX* ctx, const CHAR* text, OFF beg, OFF max_end, OFF* p_end)
{
    int is_contents;
    OFF off = beg;

    MD_ASSERT(text[off] == _T('&'));
    off++;

    if(off+2 < max_end  &&  text[off] == _T('#')  &&  (text[off+1] == _T('x') || text[off+1] == _T('X')))
        is_contents = md_is_hex_entity_contents(ctx, text, off+2, max_end, &off);
    else if(off+1 < max_end  &&  text[off] == _T('#'))
        is_contents = md_is_dec_entity_contents(ctx, text, off+1, max_end, &off);
    else
        is_contents = md_is_named_entity_contents(ctx, text, off, max_end, &off);

    if(is_contents  &&  off < max_end  &&  text[off] == _T(';')) {
        *p_end = off+1;
        return TRUE;
    } else {
        return FALSE;
    }
}

static inline int
md_is_entity(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end)
{
    return md_is_entity_str(ctx, ctx->text, beg, max_end, p_end);
}


/******************************
 ***  Attribute Management  ***
 ******************************/

typedef struct MD_ATTRIBUTE_BUILD_tag MD_ATTRIBUTE_BUILD;
struct MD_ATTRIBUTE_BUILD_tag {
    CHAR* text;
    MD_TEXTTYPE* substr_types;
    OFF* substr_offsets;
    int substr_count;
    int substr_alloc;
    MD_TEXTTYPE trivial_types[1];
    OFF trivial_offsets[2];
};


#define MD_BUILD_ATTR_NO_ESCAPES    0x0001

static int
md_build_attr_append_substr(MD_CTX* ctx, MD_ATTRIBUTE_BUILD* build,
                            MD_TEXTTYPE type, OFF off)
{
    if(build->substr_count >= build->substr_alloc) {
        MD_TEXTTYPE* new_substr_types;
        OFF* new_substr_offsets;

        build->substr_alloc = (build->substr_alloc > 0
                ? build->substr_alloc + build->substr_alloc / 2
                : 8);
        new_substr_types = (MD_TEXTTYPE*) realloc(build->substr_types,
                                    build->substr_alloc * sizeof(MD_TEXTTYPE));
        if(new_substr_types == NULL) {
            MD_LOG("realloc() failed.");
            return -1;
        }
        /* Note +1 to reserve space for final offset (== raw_size). */
        new_substr_offsets = (OFF*) realloc(build->substr_offsets,
                                    (build->substr_alloc+1) * sizeof(OFF));
        if(new_substr_offsets == NULL) {
            MD_LOG("realloc() failed.");
            free(new_substr_types);
            return -1;
        }

        build->substr_types = new_substr_types;
        build->substr_offsets = new_substr_offsets;
    }

    build->substr_types[build->substr_count] = type;
    build->substr_offsets[build->substr_count] = off;
    build->substr_count++;
    return 0;
}

static void
md_free_attribute(MD_CTX* ctx, MD_ATTRIBUTE_BUILD* build)
{
    MD_UNUSED(ctx);

    if(build->substr_alloc > 0) {
        free(build->text);
        free(build->substr_types);
        free(build->substr_offsets);
    }
}

static int
md_build_attribute(MD_CTX* ctx, const CHAR* raw_text, SZ raw_size,
                   unsigned flags, MD_ATTRIBUTE* attr, MD_ATTRIBUTE_BUILD* build)
{
    OFF raw_off, off;
    int is_trivial;
    int ret = 0;

    memset(build, 0, sizeof(MD_ATTRIBUTE_BUILD));

    /* If there is no backslash and no ampersand, build trivial attribute
     * without any malloc(). */
    is_trivial = TRUE;
    for(raw_off = 0; raw_off < raw_size; raw_off++) {
        if(ISANYOF3_(raw_text[raw_off], _T('\\'), _T('&'), _T('\0'))) {
            is_trivial = FALSE;
            break;
        }
    }

    if(is_trivial) {
        build->text = (CHAR*) (raw_size ? raw_text : NULL);
        build->substr_types = build->trivial_types;
        build->substr_offsets = build->trivial_offsets;
        build->substr_count = 1;
        build->substr_alloc = 0;
        build->trivial_types[0] = MD_TEXT_NORMAL;
        build->trivial_offsets[0] = 0;
        build->trivial_offsets[1] = raw_size;
        off = raw_size;
    } else {
        build->text = (CHAR*) malloc(raw_size * sizeof(CHAR));
        if(build->text == NULL) {
            MD_LOG("malloc() failed.");
            goto abort;
        }

        raw_off = 0;
        off = 0;

        while(raw_off < raw_size) {
            if(raw_text[raw_off] == _T('\0')) {
                MD_CHECK(md_build_attr_append_substr(ctx, build, MD_TEXT_NULLCHAR, off));
                memcpy(build->text + off, raw_text + raw_off, 1);
                off++;
                raw_off++;
                continue;
            }

            if(raw_text[raw_off] == _T('&')) {
                OFF ent_end;

                if(md_is_entity_str(ctx, raw_text, raw_off, raw_size, &ent_end)) {
                    MD_CHECK(md_build_attr_append_substr(ctx, build, MD_TEXT_ENTITY, off));
                    memcpy(build->text + off, raw_text + raw_off, ent_end - raw_off);
                    off += ent_end - raw_off;
                    raw_off = ent_end;
                    continue;
                }
            }

            if(build->substr_count == 0  ||  build->substr_types[build->substr_count-1] != MD_TEXT_NORMAL)
                MD_CHECK(md_build_attr_append_substr(ctx, build, MD_TEXT_NORMAL, off));

            if(!(flags & MD_BUILD_ATTR_NO_ESCAPES)  &&
               raw_text[raw_off] == _T('\\')  &&  raw_off+1 < raw_size  &&
               (ISPUNCT_(raw_text[raw_off+1]) || ISNEWLINE_(raw_text[raw_off+1])))
                raw_off++;

            build->text[off++] = raw_text[raw_off++];
        }
        build->substr_offsets[build->substr_count] = off;
    }

    attr->text = build->text;
    attr->size = off;
    attr->substr_offsets = build->substr_offsets;
    attr->substr_types = build->substr_types;
    return 0;

abort:
    md_free_attribute(ctx, build);
    return -1;
}


/*********************************************
 ***  Dictionary of Reference Definitions  ***
 *********************************************/

#define MD_FNV1A_BASE       2166136261U
#define MD_FNV1A_PRIME      16777619U

static inline unsigned
md_fnv1a(unsigned base, const void* data, size_t n)
{
    const unsigned char* buf = (const unsigned char*) data;
    unsigned hash = base;
    size_t i;

    for(i = 0; i < n; i++) {
        hash ^= buf[i];
        hash *= MD_FNV1A_PRIME;
    }

    return hash;
}


struct MD_REF_DEF_tag {
    CHAR* label;
    CHAR* title;
    unsigned hash;
    SZ label_size;
    SZ title_size;
    OFF dest_beg;
    OFF dest_end;
    unsigned char label_needs_free : 1;
    unsigned char title_needs_free : 1;
};

/* Label equivalence is quite complicated with regards to whitespace and case
 * folding. This complicates computing a hash of it as well as direct comparison
 * of two labels. */

static unsigned
md_link_label_hash(const CHAR* label, SZ size)
{
    unsigned hash = MD_FNV1A_BASE;
    OFF off;
    unsigned codepoint;
    int is_whitespace = FALSE;

    off = md_skip_unicode_whitespace(label, 0, size);
    while(off < size) {
        SZ char_size;

        codepoint = md_decode_unicode(label, off, size, &char_size);
        is_whitespace = ISUNICODEWHITESPACE_(codepoint) || ISNEWLINE_(label[off]);

        if(is_whitespace) {
            codepoint = ' ';
            hash = md_fnv1a(hash, &codepoint, sizeof(unsigned));
            off = md_skip_unicode_whitespace(label, off, size);
        } else {
            MD_UNICODE_FOLD_INFO fold_info;

            md_get_unicode_fold_info(codepoint, &fold_info);
            hash = md_fnv1a(hash, fold_info.codepoints, fold_info.n_codepoints * sizeof(unsigned));
            off += char_size;
        }
    }

    return hash;
}

static OFF
md_link_label_cmp_load_fold_info(const CHAR* label, OFF off, SZ size,
                                 MD_UNICODE_FOLD_INFO* fold_info)
{
    unsigned codepoint;
    SZ char_size;

    if(off >= size) {
        /* Treat end of a link label as a whitespace. */
        goto whitespace;
    }

    codepoint = md_decode_unicode(label, off, size, &char_size);
    off += char_size;
    if(ISUNICODEWHITESPACE_(codepoint)) {
        /* Treat all whitespace as equivalent */
        goto whitespace;
    }

    /* Get real folding info. */
    md_get_unicode_fold_info(codepoint, fold_info);
    return off;

whitespace:
    fold_info->codepoints[0] = _T(' ');
    fold_info->n_codepoints = 1;
    return md_skip_unicode_whitespace(label, off, size);
}

static int
md_link_label_cmp(const CHAR* a_label, SZ a_size, const CHAR* b_label, SZ b_size)
{
    OFF a_off;
    OFF b_off;
    MD_UNICODE_FOLD_INFO a_fi = { { 0 }, 0 };
    MD_UNICODE_FOLD_INFO b_fi = { { 0 }, 0 };
    OFF a_fi_off = 0;
    OFF b_fi_off = 0;
    int cmp;

    a_off = md_skip_unicode_whitespace(a_label, 0, a_size);
    b_off = md_skip_unicode_whitespace(b_label, 0, b_size);
    while(a_off < a_size || a_fi_off < a_fi.n_codepoints ||
          b_off < b_size || b_fi_off < b_fi.n_codepoints)
    {
        /* If needed, load fold info for next char. */
        if(a_fi_off >= a_fi.n_codepoints) {
            a_fi_off = 0;
            a_off = md_link_label_cmp_load_fold_info(a_label, a_off, a_size, &a_fi);
        }
        if(b_fi_off >= b_fi.n_codepoints) {
            b_fi_off = 0;
            b_off = md_link_label_cmp_load_fold_info(b_label, b_off, b_size, &b_fi);
        }

        cmp = b_fi.codepoints[b_fi_off] - a_fi.codepoints[a_fi_off];
        if(cmp != 0)
            return cmp;

        a_fi_off++;
        b_fi_off++;
    }

    return 0;
}

typedef struct MD_REF_DEF_LIST_tag MD_REF_DEF_LIST;
struct MD_REF_DEF_LIST_tag {
    int n_ref_defs;
    int alloc_ref_defs;
    MD_REF_DEF* ref_defs[];  /* Valid items always  point into ctx->ref_defs[] */
};

static int
md_ref_def_cmp(const void* a, const void* b)
{
    const MD_REF_DEF* a_ref = *(const MD_REF_DEF**)a;
    const MD_REF_DEF* b_ref = *(const MD_REF_DEF**)b;

    if(a_ref->hash < b_ref->hash)
        return -1;
    else if(a_ref->hash > b_ref->hash)
        return +1;
    else
        return md_link_label_cmp(a_ref->label, a_ref->label_size, b_ref->label, b_ref->label_size);
}

static int
md_ref_def_cmp_for_sort(const void* a, const void* b)
{
    int cmp;

    cmp = md_ref_def_cmp(a, b);

    /* Ensure stability of the sorting. */
    if(cmp == 0) {
        const MD_REF_DEF* a_ref = *(const MD_REF_DEF**)a;
        const MD_REF_DEF* b_ref = *(const MD_REF_DEF**)b;

        if(a_ref < b_ref)
            cmp = -1;
        else if(a_ref > b_ref)
            cmp = +1;
        else
            cmp = 0;
    }

    return cmp;
}

static int
md_build_ref_def_hashtable(MD_CTX* ctx)
{
    int i, j;

    if(ctx->n_ref_defs == 0)
        return 0;

    ctx->ref_def_hashtable_size = (ctx->n_ref_defs * 5) / 4;
    ctx->ref_def_hashtable = malloc(ctx->ref_def_hashtable_size * sizeof(void*));
    if(ctx->ref_def_hashtable == NULL) {
        MD_LOG("malloc() failed.");
        goto abort;
    }
    memset(ctx->ref_def_hashtable, 0, ctx->ref_def_hashtable_size * sizeof(void*));

    /* Each member of ctx->ref_def_hashtable[] can be:
     *  -- NULL,
     *  -- pointer to the MD_REF_DEF in ctx->ref_defs[], or
     *  -- pointer to a MD_REF_DEF_LIST, which holds multiple pointers to
     *     such MD_REF_DEFs.
     */
    for(i = 0; i < ctx->n_ref_defs; i++) {
        MD_REF_DEF* def = &ctx->ref_defs[i];
        void* bucket;
        MD_REF_DEF_LIST* list;

        def->hash = md_link_label_hash(def->label, def->label_size);
        bucket = ctx->ref_def_hashtable[def->hash % ctx->ref_def_hashtable_size];

        if(bucket == NULL) {
            /* The bucket is empty. Make it just point to the def. */
            ctx->ref_def_hashtable[def->hash % ctx->ref_def_hashtable_size] = def;
            continue;
        }

        if(ctx->ref_defs <= (MD_REF_DEF*) bucket  &&  (MD_REF_DEF*) bucket < ctx->ref_defs + ctx->n_ref_defs) {
            /* The bucket already contains one ref. def. Lets see whether it
             * is the same label (ref. def. duplicate) or different one
             * (hash conflict). */
            MD_REF_DEF* old_def = (MD_REF_DEF*) bucket;

            if(md_link_label_cmp(def->label, def->label_size, old_def->label, old_def->label_size) == 0) {
                /* Duplicate label: Ignore this ref. def. */
                continue;
            }

            /* Make the bucket complex, i.e. able to hold more ref. defs. */
            list = (MD_REF_DEF_LIST*) malloc(sizeof(MD_REF_DEF_LIST) + 2 * sizeof(MD_REF_DEF*));
            if(list == NULL) {
                MD_LOG("malloc() failed.");
                goto abort;
            }
            list->ref_defs[0] = old_def;
            list->ref_defs[1] = def;
            list->n_ref_defs = 2;
            list->alloc_ref_defs = 2;
            ctx->ref_def_hashtable[def->hash % ctx->ref_def_hashtable_size] = list;
            continue;
        }

        /* Append the def to the complex bucket list.
         *
         * Note in this case we ignore potential duplicates to avoid expensive
         * iterating over the complex bucket. Below, we revisit all the complex
         * buckets and handle it more cheaply after the complex bucket contents
         * is sorted. */
        list = (MD_REF_DEF_LIST*) bucket;
        if(list->n_ref_defs >= list->alloc_ref_defs) {
            int alloc_ref_defs = list->alloc_ref_defs + list->alloc_ref_defs / 2;
            MD_REF_DEF_LIST* list_tmp = (MD_REF_DEF_LIST*) realloc(list,
                        sizeof(MD_REF_DEF_LIST) + alloc_ref_defs * sizeof(MD_REF_DEF*));
            if(list_tmp == NULL) {
                MD_LOG("realloc() failed.");
                goto abort;
            }
            list = list_tmp;
            list->alloc_ref_defs = alloc_ref_defs;
            ctx->ref_def_hashtable[def->hash % ctx->ref_def_hashtable_size] = list;
        }

        list->ref_defs[list->n_ref_defs] = def;
        list->n_ref_defs++;
    }

    /* Sort the complex buckets so we can use bsearch() with them. */
    for(i = 0; i < ctx->ref_def_hashtable_size; i++) {
        void* bucket = ctx->ref_def_hashtable[i];
        MD_REF_DEF_LIST* list;

        if(bucket == NULL)
            continue;
        if(ctx->ref_defs <= (MD_REF_DEF*) bucket  &&  (MD_REF_DEF*) bucket < ctx->ref_defs + ctx->n_ref_defs)
            continue;

        list = (MD_REF_DEF_LIST*) bucket;
        qsort(list->ref_defs, list->n_ref_defs, sizeof(MD_REF_DEF*), md_ref_def_cmp_for_sort);

        /* Disable all duplicates in the complex bucket by forcing all such
         * records to point to the 1st such ref. def. I.e. no matter which
         * record is found during the lookup, it will always point to the right
         * ref. def. in ctx->ref_defs[]. */
        for(j = 1; j < list->n_ref_defs; j++) {
            if(md_ref_def_cmp(&list->ref_defs[j-1], &list->ref_defs[j]) == 0)
                list->ref_defs[j] = list->ref_defs[j-1];
        }
    }

    return 0;

abort:
    return -1;
}

static void
md_free_ref_def_hashtable(MD_CTX* ctx)
{
    if(ctx->ref_def_hashtable != NULL) {
        int i;

        for(i = 0; i < ctx->ref_def_hashtable_size; i++) {
            void* bucket = ctx->ref_def_hashtable[i];
            if(bucket == NULL)
                continue;
            if(ctx->ref_defs <= (MD_REF_DEF*) bucket  &&  (MD_REF_DEF*) bucket < ctx->ref_defs + ctx->n_ref_defs)
                continue;
            free(bucket);
        }

        free(ctx->ref_def_hashtable);
    }
}

static const MD_REF_DEF*
md_lookup_ref_def(MD_CTX* ctx, const CHAR* label, SZ label_size)
{
    unsigned hash;
    void* bucket;

    if(ctx->ref_def_hashtable_size == 0)
        return NULL;

    hash = md_link_label_hash(label, label_size);
    bucket = ctx->ref_def_hashtable[hash % ctx->ref_def_hashtable_size];

    if(bucket == NULL) {
        return NULL;
    } else if(ctx->ref_defs <= (MD_REF_DEF*) bucket  &&  (MD_REF_DEF*) bucket < ctx->ref_defs + ctx->n_ref_defs) {
        const MD_REF_DEF* def = (MD_REF_DEF*) bucket;

        if(md_link_label_cmp(def->label, def->label_size, label, label_size) == 0)
            return def;
        else
            return NULL;
    } else {
        MD_REF_DEF_LIST* list = (MD_REF_DEF_LIST*) bucket;
        MD_REF_DEF key_buf;
        const MD_REF_DEF* key = &key_buf;
        const MD_REF_DEF** ret;

        key_buf.label = (CHAR*) label;
        key_buf.label_size = label_size;
        key_buf.hash = md_link_label_hash(key_buf.label, key_buf.label_size);

        ret = (const MD_REF_DEF**) bsearch(&key, list->ref_defs,
                    list->n_ref_defs, sizeof(MD_REF_DEF*), md_ref_def_cmp);
        if(ret != NULL)
            return *ret;
        else
            return NULL;
    }
}


/***************************
 ***  Recognizing Links  ***
 ***************************/

/* Note this code is partially shared between processing inlines and blocks
 * as reference definitions and links share some helper parser functions.
 */

typedef struct MD_LINK_ATTR_tag MD_LINK_ATTR;
struct MD_LINK_ATTR_tag {
    OFF dest_beg;
    OFF dest_end;

    CHAR* title;
    SZ title_size;
    int title_needs_free;
};


static int
md_is_link_label(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg,
                 OFF* p_end, int* p_beg_line_index, int* p_end_line_index,
                 OFF* p_contents_beg, OFF* p_contents_end)
{
    OFF off = beg;
    OFF contents_beg = 0;
    OFF contents_end = 0;
    int line_index = 0;
    int len = 0;

    if(CH(off) != _T('['))
        return FALSE;
    off++;

    while(1) {
        OFF line_end = lines[line_index].end;

        while(off < line_end) {
            if(CH(off) == _T('\\')  &&  off+1 < ctx->size  &&  (ISPUNCT(off+1) || ISNEWLINE(off+1))) {
                if(contents_end == 0) {
                    contents_beg = off;
                    *p_beg_line_index = line_index;
                }
                contents_end = off + 2;
                off += 2;
            } else if(CH(off) == _T('[')) {
                return FALSE;
            } else if(CH(off) == _T(']')) {
                if(contents_beg < contents_end) {
                    /* Success. */
                    *p_contents_beg = contents_beg;
                    *p_contents_end = contents_end;
                    *p_end = off+1;
                    *p_end_line_index = line_index;
                    return TRUE;
                } else {
                    /* Link label must have some non-whitespace contents. */
                    return FALSE;
                }
            } else {
                unsigned codepoint;
                SZ char_size;

                codepoint = md_decode_unicode(ctx->text, off, ctx->size, &char_size);
                if(!ISUNICODEWHITESPACE_(codepoint)) {
                    if(contents_end == 0) {
                        contents_beg = off;
                        *p_beg_line_index = line_index;
                    }
                    contents_end = off + char_size;
                }

                off += char_size;
            }

            len++;
            if(len > 999)
                return FALSE;
        }

        line_index++;
        len++;
        if(line_index < n_lines)
            off = lines[line_index].beg;
        else
            break;
    }

    return FALSE;
}

static int
md_is_link_destination_A(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end,
                         OFF* p_contents_beg, OFF* p_contents_end)
{
    OFF off = beg;

    if(off >= max_end  ||  CH(off) != _T('<'))
        return FALSE;
    off++;

    while(off < max_end) {
        if(CH(off) == _T('\\')  &&  off+1 < max_end  &&  ISPUNCT(off+1)) {
            off += 2;
            continue;
        }

        if(ISNEWLINE(off)  ||  CH(off) == _T('<'))
            return FALSE;

        if(CH(off) == _T('>')) {
            /* Success. */
            *p_contents_beg = beg+1;
            *p_contents_end = off;
            *p_end = off+1;
            return TRUE;
        }

        off++;
    }

    return FALSE;
}

static int
md_is_link_destination_B(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end,
                         OFF* p_contents_beg, OFF* p_contents_end)
{
    OFF off = beg;
    int parenthesis_level = 0;

    while(off < max_end) {
        if(CH(off) == _T('\\')  &&  off+1 < max_end  &&  ISPUNCT(off+1)) {
            off += 2;
            continue;
        }

        if(ISWHITESPACE(off) || ISCNTRL(off))
            break;

        /* Link destination may include balanced pairs of unescaped '(' ')'.
         * Note we limit the maximal nesting level by 32 to protect us from
         * https://github.com/jgm/cmark/issues/214 */
        if(CH(off) == _T('(')) {
            parenthesis_level++;
            if(parenthesis_level > 32)
                return FALSE;
        } else if(CH(off) == _T(')')) {
            if(parenthesis_level == 0)
                break;
            parenthesis_level--;
        }

        off++;
    }

    if(parenthesis_level != 0  ||  off == beg)
        return FALSE;

    /* Success. */
    *p_contents_beg = beg;
    *p_contents_end = off;
    *p_end = off;
    return TRUE;
}

static inline int
md_is_link_destination(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end,
                       OFF* p_contents_beg, OFF* p_contents_end)
{
    if(CH(beg) == _T('<'))
        return md_is_link_destination_A(ctx, beg, max_end, p_end, p_contents_beg, p_contents_end);
    else
        return md_is_link_destination_B(ctx, beg, max_end, p_end, p_contents_beg, p_contents_end);
}

static int
md_is_link_title(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg,
                 OFF* p_end, int* p_beg_line_index, int* p_end_line_index,
                 OFF* p_contents_beg, OFF* p_contents_end)
{
    OFF off = beg;
    CHAR closer_char;
    int line_index = 0;

    /* White space with up to one line break. */
    while(off < lines[line_index].end  &&  ISWHITESPACE(off))
        off++;
    if(off >= lines[line_index].end) {
        line_index++;
        if(line_index >= n_lines)
            return FALSE;
        off = lines[line_index].beg;
    }
    if(off == beg)
        return FALSE;

    *p_beg_line_index = line_index;

    /* First char determines how to detect end of it. */
    switch(CH(off)) {
        case _T('"'):   closer_char = _T('"'); break;
        case _T('\''):  closer_char = _T('\''); break;
        case _T('('):   closer_char = _T(')'); break;
        default:        return FALSE;
    }
    off++;

    *p_contents_beg = off;

    while(line_index < n_lines) {
        OFF line_end = lines[line_index].end;

        while(off < line_end) {
            if(CH(off) == _T('\\')  &&  off+1 < ctx->size  &&  (ISPUNCT(off+1) || ISNEWLINE(off+1))) {
                off++;
            } else if(CH(off) == closer_char) {
                /* Success. */
                *p_contents_end = off;
                *p_end = off+1;
                *p_end_line_index = line_index;
                return TRUE;
            } else if(closer_char == _T(')')  &&  CH(off) == _T('(')) {
                /* ()-style title cannot contain (unescaped '(')) */
                return FALSE;
            }

            off++;
        }

        line_index++;
    }

    return FALSE;
}

/* Returns 0 if it is not a reference definition.
 *
 * Returns N > 0 if it is a reference definition. N then corresponds to the
 * number of lines forming it). In this case the definition is stored for
 * resolving any links referring to it.
 *
 * Returns -1 in case of an error (out of memory).
 */
static int
md_is_link_reference_definition(MD_CTX* ctx, const MD_LINE* lines, int n_lines)
{
    OFF label_contents_beg;
    OFF label_contents_end;
    int label_contents_line_index = -1;
    int label_is_multiline = FALSE;
    OFF dest_contents_beg;
    OFF dest_contents_end;
    OFF title_contents_beg;
    OFF title_contents_end;
    int title_contents_line_index;
    int title_is_multiline = FALSE;
    OFF off;
    int line_index = 0;
    int tmp_line_index;
    MD_REF_DEF* def = NULL;
    int ret = 0;

    /* Link label. */
    if(!md_is_link_label(ctx, lines, n_lines, lines[0].beg,
                &off, &label_contents_line_index, &line_index,
                &label_contents_beg, &label_contents_end))
        return FALSE;
    label_is_multiline = (label_contents_line_index != line_index);

    /* Colon. */
    if(off >= lines[line_index].end  ||  CH(off) != _T(':'))
        return FALSE;
    off++;

    /* Optional white space with up to one line break. */
    while(off < lines[line_index].end  &&  ISWHITESPACE(off))
        off++;
    if(off >= lines[line_index].end) {
        line_index++;
        if(line_index >= n_lines)
            return FALSE;
        off = lines[line_index].beg;
    }

    /* Link destination. */
    if(!md_is_link_destination(ctx, off, lines[line_index].end,
                &off, &dest_contents_beg, &dest_contents_end))
        return FALSE;

    /* (Optional) title. Note we interpret it as an title only if nothing
     * more follows on its last line. */
    if(md_is_link_title(ctx, lines + line_index, n_lines - line_index, off,
                &off, &title_contents_line_index, &tmp_line_index,
                &title_contents_beg, &title_contents_end)
        &&  off >= lines[line_index + tmp_line_index].end)
    {
        title_is_multiline = (tmp_line_index != title_contents_line_index);
        title_contents_line_index += line_index;
        line_index += tmp_line_index;
    } else {
        /* Not a title. */
        title_is_multiline = FALSE;
        title_contents_beg = off;
        title_contents_end = off;
        title_contents_line_index = 0;
    }

    /* Nothing more can follow on the last line. */
    if(off < lines[line_index].end)
        return FALSE;

    /* So, it _is_ a reference definition. Remember it. */
    if(ctx->n_ref_defs >= ctx->alloc_ref_defs) {
        MD_REF_DEF* new_defs;

        ctx->alloc_ref_defs = (ctx->alloc_ref_defs > 0
                ? ctx->alloc_ref_defs + ctx->alloc_ref_defs / 2
                : 16);
        new_defs = (MD_REF_DEF*) realloc(ctx->ref_defs, ctx->alloc_ref_defs * sizeof(MD_REF_DEF));
        if(new_defs == NULL) {
            MD_LOG("realloc() failed.");
            goto abort;
        }

        ctx->ref_defs = new_defs;
    }
    def = &ctx->ref_defs[ctx->n_ref_defs];
    memset(def, 0, sizeof(MD_REF_DEF));

    if(label_is_multiline) {
        MD_CHECK(md_merge_lines_alloc(ctx, label_contents_beg, label_contents_end,
                    lines + label_contents_line_index, n_lines - label_contents_line_index,
                    _T(' '), &def->label, &def->label_size));
        def->label_needs_free = TRUE;
    } else {
        def->label = (CHAR*) STR(label_contents_beg);
        def->label_size = label_contents_end - label_contents_beg;
    }

    if(title_is_multiline) {
        MD_CHECK(md_merge_lines_alloc(ctx, title_contents_beg, title_contents_end,
                    lines + title_contents_line_index, n_lines - title_contents_line_index,
                    _T('\n'), &def->title, &def->title_size));
        def->title_needs_free = TRUE;
    } else {
        def->title = (CHAR*) STR(title_contents_beg);
        def->title_size = title_contents_end - title_contents_beg;
    }

    def->dest_beg = dest_contents_beg;
    def->dest_end = dest_contents_end;

    /* Success. */
    ctx->n_ref_defs++;
    return line_index + 1;

abort:
    /* Failure. */
    if(def != NULL  &&  def->label_needs_free)
        free(def->label);
    if(def != NULL  &&  def->title_needs_free)
        free(def->title);
    return ret;
}

static int
md_is_link_reference(MD_CTX* ctx, const MD_LINE* lines, int n_lines,
                     OFF beg, OFF end, MD_LINK_ATTR* attr)
{
    const MD_REF_DEF* def;
    const MD_LINE* beg_line;
    int is_multiline;
    CHAR* label;
    SZ label_size;
    int ret;

    MD_ASSERT(CH(beg) == _T('[') || CH(beg) == _T('!'));
    MD_ASSERT(CH(end-1) == _T(']'));

    beg += (CH(beg) == _T('!') ? 2 : 1);
    end--;

    /* Find lines corresponding to the beg and end positions. */
    beg_line = md_lookup_line(beg, lines, n_lines);
    is_multiline = (end > beg_line->end);

    if(is_multiline) {
        MD_CHECK(md_merge_lines_alloc(ctx, beg, end, beg_line,
                 (int)(n_lines - (beg_line - lines)), _T(' '), &label, &label_size));
    } else {
        label = (CHAR*) STR(beg);
        label_size = end - beg;
    }

    def = md_lookup_ref_def(ctx, label, label_size);
    if(def != NULL) {
        attr->dest_beg = def->dest_beg;
        attr->dest_end = def->dest_end;
        attr->title = def->title;
        attr->title_size = def->title_size;
        attr->title_needs_free = FALSE;
    }

    if(is_multiline)
        free(label);

    ret = (def != NULL);

abort:
    return ret;
}

static int
md_is_inline_link_spec(MD_CTX* ctx, const MD_LINE* lines, int n_lines,
                       OFF beg, OFF* p_end, MD_LINK_ATTR* attr)
{
    int line_index = 0;
    int tmp_line_index;
    OFF title_contents_beg;
    OFF title_contents_end;
    int title_contents_line_index;
    int title_is_multiline;
    OFF off = beg;
    int ret = FALSE;

    while(off >= lines[line_index].end)
        line_index++;

    MD_ASSERT(CH(off) == _T('('));
    off++;

    /* Optional white space with up to one line break. */
    while(off < lines[line_index].end  &&  ISWHITESPACE(off))
        off++;
    if(off >= lines[line_index].end  &&  (off >= ctx->size  ||  ISNEWLINE(off))) {
        line_index++;
        if(line_index >= n_lines)
            return FALSE;
        off = lines[line_index].beg;
    }

    /* Link destination may be omitted, but only when not also having a title. */
    if(off < ctx->size  &&  CH(off) == _T(')')) {
        attr->dest_beg = off;
        attr->dest_end = off;
        attr->title = NULL;
        attr->title_size = 0;
        attr->title_needs_free = FALSE;
        off++;
        *p_end = off;
        return TRUE;
    }

    /* Link destination. */
    if(!md_is_link_destination(ctx, off, lines[line_index].end,
                        &off, &attr->dest_beg, &attr->dest_end))
        return FALSE;

    /* (Optional) title. */
    if(md_is_link_title(ctx, lines + line_index, n_lines - line_index, off,
                &off, &title_contents_line_index, &tmp_line_index,
                &title_contents_beg, &title_contents_end))
    {
        title_is_multiline = (tmp_line_index != title_contents_line_index);
        title_contents_line_index += line_index;
        line_index += tmp_line_index;
    } else {
        /* Not a title. */
        title_is_multiline = FALSE;
        title_contents_beg = off;
        title_contents_end = off;
        title_contents_line_index = 0;
    }

    /* Optional whitespace followed with final ')'. */
    while(off < lines[line_index].end  &&  ISWHITESPACE(off))
        off++;
    if (off >= lines[line_index].end  &&  (off >= ctx->size || ISNEWLINE(off))) {
        line_index++;
        if(line_index >= n_lines)
            return FALSE;
        off = lines[line_index].beg;
    }
    if(CH(off) != _T(')'))
        goto abort;
    off++;

    if(title_contents_beg >= title_contents_end) {
        attr->title = NULL;
        attr->title_size = 0;
        attr->title_needs_free = FALSE;
    } else if(!title_is_multiline) {
        attr->title = (CHAR*) STR(title_contents_beg);
        attr->title_size = title_contents_end - title_contents_beg;
        attr->title_needs_free = FALSE;
    } else {
        MD_CHECK(md_merge_lines_alloc(ctx, title_contents_beg, title_contents_end,
                    lines + title_contents_line_index, n_lines - title_contents_line_index,
                    _T('\n'), &attr->title, &attr->title_size));
        attr->title_needs_free = TRUE;
    }

    *p_end = off;
    ret = TRUE;

abort:
    return ret;
}

static void
md_free_ref_defs(MD_CTX* ctx)
{
    int i;

    for(i = 0; i < ctx->n_ref_defs; i++) {
        MD_REF_DEF* def = &ctx->ref_defs[i];

        if(def->label_needs_free)
            free(def->label);
        if(def->title_needs_free)
            free(def->title);
    }

    free(ctx->ref_defs);
}


/******************************************
 ***  Processing Inlines (a.k.a Spans)  ***
 ******************************************/

/* We process inlines in few phases:
 *
 * (1) We go through the block text and collect all significant characters
 *     which may start/end a span or some other significant position into
 *     ctx->marks[]. Core of this is what md_collect_marks() does.
 *
 *     We also do some very brief preliminary context-less analysis, whether
 *     it might be opener or closer (e.g. of an emphasis span).
 *
 *     This speeds the other steps as we do not need to re-iterate over all
 *     characters anymore.
 *
 * (2) We analyze each potential mark types, in order by their precedence.
 *
 *     In each md_analyze_XXX() function, we re-iterate list of the marks,
 *     skipping already resolved regions (in preceding precedences) and try to
 *     resolve them.
 *
 * (2.1) For trivial marks, which are single (e.g. HTML entity), we just mark
 *       them as resolved.
 *
 * (2.2) For range-type marks, we analyze whether the mark could be closer
 *       and, if yes, whether there is some preceding opener it could satisfy.
 *
 *       If not we check whether it could be really an opener and if yes, we
 *       remember it so subsequent closers may resolve it.
 *
 * (3) Finally, when all marks were analyzed, we render the block contents
 *     by calling MD_RENDERER::text() callback, interrupting by ::enter_span()
 *     or ::close_span() whenever we reach a resolved mark.
 */


/* The mark structure.
 *
 * '\\': Maybe escape sequence.
 * '\0': NULL char.
 *  '*': Maybe (strong) emphasis start/end.
 *  '_': Maybe (strong) emphasis start/end.
 *  '~': Maybe strikethrough start/end (needs MD_FLAG_STRIKETHROUGH).
 *  '`': Maybe code span start/end.
 *  '&': Maybe start of entity.
 *  ';': Maybe end of entity.
 *  '<': Maybe start of raw HTML or autolink.
 *  '>': Maybe end of raw HTML or autolink.
 *  '[': Maybe start of link label or link text.
 *  '!': Equivalent of '[' for image.
 *  ']': Maybe end of link label or link text.
 *  '@': Maybe permissive e-mail auto-link (needs MD_FLAG_PERMISSIVEEMAILAUTOLINKS).
 *  ':': Maybe permissive URL auto-link (needs MD_FLAG_PERMISSIVEURLAUTOLINKS).
 *  '.': Maybe permissive WWW auto-link (needs MD_FLAG_PERMISSIVEWWWAUTOLINKS).
 *  'D': Dummy mark, it reserves a space for splitting a previous mark
 *       (e.g. emphasis) or to make more space for storing some special data
 *       related to the preceding mark (e.g. link).
 *
 * Note that not all instances of these chars in the text imply creation of the
 * structure. Only those which have (or may have, after we see more context)
 * the special meaning.
 *
 * (Keep this struct as small as possible to fit as much of them into CPU
 * cache line.)
 */
struct MD_MARK_tag {
    OFF beg;
    OFF end;

    /* For unresolved openers, 'prev' and 'next' form the chain of open openers
     * of given type 'ch'.
     *
     * During resolving, we disconnect from the chain and point to the
     * corresponding counterpart so opener points to its closer and vice versa.
     */
    int prev;
    int next;
    CHAR ch;
    unsigned char flags;
};

/* Mark flags (these apply to ALL mark types). */
#define MD_MARK_POTENTIAL_OPENER            0x01  /* Maybe opener. */
#define MD_MARK_POTENTIAL_CLOSER            0x02  /* Maybe closer. */
#define MD_MARK_OPENER                      0x04  /* Definitely opener. */
#define MD_MARK_CLOSER                      0x08  /* Definitely closer. */
#define MD_MARK_RESOLVED                    0x10  /* Resolved in any definite way. */

/* Mark flags specific for various mark types (so they can share bits). */
#define MD_MARK_EMPH_INTRAWORD              0x20  /* Helper for the "rule of 3". */
#define MD_MARK_EMPH_MOD3_0                 0x40
#define MD_MARK_EMPH_MOD3_1                 0x80
#define MD_MARK_EMPH_MOD3_2                 (0x40 | 0x80)
#define MD_MARK_EMPH_MOD3_MASK              (0x40 | 0x80)
#define MD_MARK_AUTOLINK                    0x20  /* Distinguisher for '<', '>'. */
#define MD_MARK_VALIDPERMISSIVEAUTOLINK     0x20  /* For permissive autolinks. */
#define MD_MARK_HASNESTEDBRACKETS           0x20  /* For '[' to rule out invalid link labels early */

static MD_MARKCHAIN*
md_asterisk_chain(MD_CTX* ctx, unsigned flags)
{
    switch(flags & (MD_MARK_EMPH_INTRAWORD | MD_MARK_EMPH_MOD3_MASK)) {
        case MD_MARK_EMPH_INTRAWORD | MD_MARK_EMPH_MOD3_0:  return &ASTERISK_OPENERS_intraword_mod3_0;
        case MD_MARK_EMPH_INTRAWORD | MD_MARK_EMPH_MOD3_1:  return &ASTERISK_OPENERS_intraword_mod3_1;
        case MD_MARK_EMPH_INTRAWORD | MD_MARK_EMPH_MOD3_2:  return &ASTERISK_OPENERS_intraword_mod3_2;
        case MD_MARK_EMPH_MOD3_0:                           return &ASTERISK_OPENERS_extraword_mod3_0;
        case MD_MARK_EMPH_MOD3_1:                           return &ASTERISK_OPENERS_extraword_mod3_1;
        case MD_MARK_EMPH_MOD3_2:                           return &ASTERISK_OPENERS_extraword_mod3_2;
        default:                                            MD_UNREACHABLE();
    }
    return NULL;
}

static MD_MARKCHAIN*
md_mark_chain(MD_CTX* ctx, int mark_index)
{
    MD_MARK* mark = &ctx->marks[mark_index];

    switch(mark->ch) {
        case _T('*'):   return md_asterisk_chain(ctx, mark->flags);
        case _T('_'):   return &UNDERSCORE_OPENERS;
        case _T('~'):   return (mark->end - mark->beg == 1) ? &TILDE_OPENERS_1 : &TILDE_OPENERS_2;
        case _T('!'):   MD_FALLTHROUGH();
        case _T('['):   return &BRACKET_OPENERS;
        case _T('|'):   return &TABLECELLBOUNDARIES;
        default:        return NULL;
    }
}

static MD_MARK*
md_push_mark(MD_CTX* ctx)
{
    if(ctx->n_marks >= ctx->alloc_marks) {
        MD_MARK* new_marks;

        ctx->alloc_marks = (ctx->alloc_marks > 0
                ? ctx->alloc_marks + ctx->alloc_marks / 2
                : 64);
        new_marks = realloc(ctx->marks, ctx->alloc_marks * sizeof(MD_MARK));
        if(new_marks == NULL) {
            MD_LOG("realloc() failed.");
            return NULL;
        }

        ctx->marks = new_marks;
    }

    return &ctx->marks[ctx->n_marks++];
}

#define PUSH_MARK_()                                                    \
        do {                                                            \
            mark = md_push_mark(ctx);                                   \
            if(mark == NULL) {                                          \
                ret = -1;                                               \
                goto abort;                                             \
            }                                                           \
        } while(0)

#define PUSH_MARK(ch_, beg_, end_, flags_)                              \
        do {                                                            \
            PUSH_MARK_();                                               \
            mark->beg = (beg_);                                         \
            mark->end = (end_);                                         \
            mark->prev = -1;                                            \
            mark->next = -1;                                            \
            mark->ch = (char)(ch_);                                     \
            mark->flags = (flags_);                                     \
        } while(0)


static void
md_mark_chain_append(MD_CTX* ctx, MD_MARKCHAIN* chain, int mark_index)
{
    if(chain->tail >= 0)
        ctx->marks[chain->tail].next = mark_index;
    else
        chain->head = mark_index;

    ctx->marks[mark_index].prev = chain->tail;
    ctx->marks[mark_index].next = -1;
    chain->tail = mark_index;
}

/* Sometimes, we need to store a pointer into the mark. It is quite rare
 * so we do not bother to make MD_MARK use union, and it can only happen
 * for dummy marks. */
static inline void
md_mark_store_ptr(MD_CTX* ctx, int mark_index, void* ptr)
{
    MD_MARK* mark = &ctx->marks[mark_index];
    MD_ASSERT(mark->ch == 'D');

    /* Check only members beg and end are misused for this. */
    MD_ASSERT(sizeof(void*) <= 2 * sizeof(OFF));
    memcpy(mark, &ptr, sizeof(void*));
}

static inline void*
md_mark_get_ptr(MD_CTX* ctx, int mark_index)
{
    void* ptr;
    MD_MARK* mark = &ctx->marks[mark_index];
    MD_ASSERT(mark->ch == 'D');
    memcpy(&ptr, mark, sizeof(void*));
    return ptr;
}

static void
md_resolve_range(MD_CTX* ctx, MD_MARKCHAIN* chain, int opener_index, int closer_index)
{
    MD_MARK* opener = &ctx->marks[opener_index];
    MD_MARK* closer = &ctx->marks[closer_index];

    /* Remove opener from the list of openers. */
    if(chain != NULL) {
        if(opener->prev >= 0)
            ctx->marks[opener->prev].next = opener->next;
        else
            chain->head = opener->next;

        if(opener->next >= 0)
            ctx->marks[opener->next].prev = opener->prev;
        else
            chain->tail = opener->prev;
    }

    /* Interconnect opener and closer and mark both as resolved. */
    opener->next = closer_index;
    opener->flags |= MD_MARK_OPENER | MD_MARK_RESOLVED;
    closer->prev = opener_index;
    closer->flags |= MD_MARK_CLOSER | MD_MARK_RESOLVED;
}


#define MD_ROLLBACK_ALL         0
#define MD_ROLLBACK_CROSSING    1

/* In the range ctx->marks[opener_index] ... [closer_index], undo some or all
 * resolvings accordingly to these rules:
 *
 * (1) All openers BEFORE the range corresponding to any closer inside the
 *     range are un-resolved and they are re-added to their respective chains
 *     of unresolved openers. This ensures we can reuse the opener for closers
 *     AFTER the range.
 *
 * (2) If 'how' is MD_ROLLBACK_ALL, then ALL resolved marks inside the range
 *     are discarded.
 *
 * (3) If 'how' is MD_ROLLBACK_CROSSING, only closers with openers handled
 *     in (1) are discarded. I.e. pairs of openers and closers which are both
 *     inside the range are retained as well as any unpaired marks.
 */
static void
md_rollback(MD_CTX* ctx, int opener_index, int closer_index, int how)
{
    int i;
    int mark_index;

    /* Cut all unresolved openers at the mark index. */
    for(i = OPENERS_CHAIN_FIRST; i < OPENERS_CHAIN_LAST+1; i++) {
        MD_MARKCHAIN* chain = &ctx->mark_chains[i];

        while(chain->tail >= opener_index) {
            int same = chain->tail == opener_index;
            chain->tail = ctx->marks[chain->tail].prev;
            if (same) break;
        }

        if(chain->tail >= 0)
            ctx->marks[chain->tail].next = -1;
        else
            chain->head = -1;
    }

    /* Go backwards so that unresolved openers are re-added into their
     * respective chains, in the right order. */
    mark_index = closer_index - 1;
    while(mark_index > opener_index) {
        MD_MARK* mark = &ctx->marks[mark_index];
        int mark_flags = mark->flags;
        int discard_flag = (how == MD_ROLLBACK_ALL);

        if(mark->flags & MD_MARK_CLOSER) {
            int mark_opener_index = mark->prev;

            /* Undo opener BEFORE the range. */
            if(mark_opener_index < opener_index) {
                MD_MARK* mark_opener = &ctx->marks[mark_opener_index];
                MD_MARKCHAIN* chain;

                mark_opener->flags &= ~(MD_MARK_OPENER | MD_MARK_CLOSER | MD_MARK_RESOLVED);
                chain = md_mark_chain(ctx, opener_index);
                if(chain != NULL) {
                    md_mark_chain_append(ctx, chain, mark_opener_index);
                    discard_flag = 1;
                }
            }
        }

        /* And reset our flags. */
        if(discard_flag) {
            /* Make zero-length closer a dummy mark as that's how it was born */
            if((mark->flags & MD_MARK_CLOSER)  &&  mark->beg == mark->end)
                mark->ch = 'D';

            mark->flags &= ~(MD_MARK_OPENER | MD_MARK_CLOSER | MD_MARK_RESOLVED);
        }

        /* Jump as far as we can over unresolved or non-interesting marks. */
        switch(how) {
            case MD_ROLLBACK_CROSSING:
                if((mark_flags & MD_MARK_CLOSER)  &&  mark->prev > opener_index) {
                    /* If we are closer with opener INSIDE the range, there may
                     * not be any other crosser inside the subrange. */
                    mark_index = mark->prev;
                    break;
                }
                MD_FALLTHROUGH();
            default:
                mark_index--;
                break;
        }
    }
}

static void
md_build_mark_char_map(MD_CTX* ctx)
{
    memset(ctx->mark_char_map, 0, sizeof(ctx->mark_char_map));

    ctx->mark_char_map['\\'] = 1;
    ctx->mark_char_map['*'] = 1;
    ctx->mark_char_map['_'] = 1;
    ctx->mark_char_map['`'] = 1;
    ctx->mark_char_map['&'] = 1;
    ctx->mark_char_map[';'] = 1;
    ctx->mark_char_map['<'] = 1;
    ctx->mark_char_map['>'] = 1;
    ctx->mark_char_map['['] = 1;
    ctx->mark_char_map['!'] = 1;
    ctx->mark_char_map[']'] = 1;
    ctx->mark_char_map['\0'] = 1;

    if(ctx->parser.flags & MD_FLAG_STRIKETHROUGH)
        ctx->mark_char_map['~'] = 1;

    if(ctx->parser.flags & MD_FLAG_LATEXMATHSPANS)
        ctx->mark_char_map['$'] = 1;

    if(ctx->parser.flags & MD_FLAG_PERMISSIVEEMAILAUTOLINKS)
        ctx->mark_char_map['@'] = 1;

    if(ctx->parser.flags & MD_FLAG_PERMISSIVEURLAUTOLINKS)
        ctx->mark_char_map[':'] = 1;

    if(ctx->parser.flags & MD_FLAG_PERMISSIVEWWWAUTOLINKS)
        ctx->mark_char_map['.'] = 1;

    if((ctx->parser.flags & MD_FLAG_TABLES) || (ctx->parser.flags & MD_FLAG_WIKILINKS))
        ctx->mark_char_map['|'] = 1;

    if(ctx->parser.flags & MD_FLAG_COLLAPSEWHITESPACE) {
        int i;

        for(i = 0; i < (int) sizeof(ctx->mark_char_map); i++) {
            if(ISWHITESPACE_(i))
                ctx->mark_char_map[i] = 1;
        }
    }
}

/* We limit code span marks to lower than 32 backticks. This solves the
 * pathologic case of too many openers, each of different length: Their
 * resolving would be then O(n^2). */
#define CODESPAN_MARK_MAXLEN    32

static int
md_is_code_span(MD_CTX* ctx, const MD_LINE* lines, int n_lines, OFF beg,
                OFF* p_opener_beg, OFF* p_opener_end,
                OFF* p_closer_beg, OFF* p_closer_end,
                OFF last_potential_closers[CODESPAN_MARK_MAXLEN],
                int* p_reached_paragraph_end)
{
    OFF opener_beg = beg;
    OFF opener_end;
    OFF closer_beg;
    OFF closer_end;
    SZ mark_len;
    OFF line_end;
    int has_space_after_opener = FALSE;
    int has_eol_after_opener = FALSE;
    int has_space_before_closer = FALSE;
    int has_eol_before_closer = FALSE;
    int has_only_space = TRUE;
    int line_index = 0;

    line_end = lines[0].end;
    opener_end = opener_beg;
    while(opener_end < line_end  &&  CH(opener_end) == _T('`'))
        opener_end++;
    has_space_after_opener = (opener_end < line_end && CH(opener_end) == _T(' '));
    has_eol_after_opener = (opener_end == line_end);

    /* The caller needs to know end of the opening mark even if we fail. */
    *p_opener_end = opener_end;

    mark_len = opener_end - opener_beg;
    if(mark_len > CODESPAN_MARK_MAXLEN)
        return FALSE;

    /* Check whether we already know there is no closer of this length.
     * If so, re-scan does no sense. This fixes issue #59. */
    if(last_potential_closers[mark_len-1] >= lines[n_lines-1].end  ||
       (*p_reached_paragraph_end  &&  last_potential_closers[mark_len-1] < opener_end))
        return FALSE;

    closer_beg = opener_end;
    closer_end = opener_end;

    /* Find closer mark. */
    while(TRUE) {
        while(closer_beg < line_end  &&  CH(closer_beg) != _T('`')) {
            if(CH(closer_beg) != _T(' '))
                has_only_space = FALSE;
            closer_beg++;
        }
        closer_end = closer_beg;
        while(closer_end < line_end  &&  CH(closer_end) == _T('`'))
            closer_end++;

        if(closer_end - closer_beg == mark_len) {
            /* Success. */
            has_space_before_closer = (closer_beg > lines[line_index].beg && CH(closer_beg-1) == _T(' '));
            has_eol_before_closer = (closer_beg == lines[line_index].beg);
            break;
        }

        if(closer_end - closer_beg > 0) {
            /* We have found a back-tick which is not part of the closer. */
            has_only_space = FALSE;

            /* But if we eventually fail, remember it as a potential closer
             * of its own length for future attempts. This mitigates needs for
             * rescans. */
            if(closer_end - closer_beg < CODESPAN_MARK_MAXLEN) {
                if(closer_beg > last_potential_closers[closer_end - closer_beg - 1])
                    last_potential_closers[closer_end - closer_beg - 1] = closer_beg;
            }
        }

        if(closer_end >= line_end) {
            line_index++;
            if(line_index >= n_lines) {
                /* Reached end of the paragraph and still nothing. */
                *p_reached_paragraph_end = TRUE;
                return FALSE;
            }
            /* Try on the next line. */
            line_end = lines[line_index].end;
            closer_beg = lines[line_index].beg;
        } else {
            closer_beg = closer_end;
        }
    }

    /* If there is a space or a new line both after and before the opener
     * (and if the code span is not made of spaces only), consume one initial
     * and one trailing space as part of the marks. */
    if(!has_only_space  &&
       (has_space_after_opener || has_eol_after_opener)  &&
       (has_space_before_closer || has_eol_before_closer))
    {
        if(has_space_after_opener)
            opener_end++;
        else
            opener_end = lines[1].beg;

        if(has_space_before_closer)
            closer_beg--;
        else {
            closer_beg = lines[line_index-1].end;
            /* We need to eat the preceding "\r\n" but not any line trailing
             * spaces. */
            while(closer_beg < ctx->size  &&  ISBLANK(closer_beg))
                closer_beg++;
        }
    }

    *p_opener_beg = opener_beg;
    *p_opener_end = opener_end;
    *p_closer_beg = closer_beg;
    *p_closer_end = closer_end;
    return TRUE;
}

static int
md_is_autolink_uri(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg+1;

    MD_ASSERT(CH(beg) == _T('<'));

    /* Check for scheme. */
    if(off >= max_end  ||  !ISASCII(off))
        return FALSE;
    off++;
    while(1) {
        if(off >= max_end)
            return FALSE;
        if(off - beg > 32)
            return FALSE;
        if(CH(off) == _T(':')  &&  off - beg >= 3)
            break;
        if(!ISALNUM(off) && CH(off) != _T('+') && CH(off) != _T('-') && CH(off) != _T('.'))
            return FALSE;
        off++;
    }

    /* Check the path after the scheme. */
    while(off < max_end  &&  CH(off) != _T('>')) {
        if(ISWHITESPACE(off) || ISCNTRL(off) || CH(off) == _T('<'))
            return FALSE;
        off++;
    }

    if(off >= max_end)
        return FALSE;

    MD_ASSERT(CH(off) == _T('>'));
    *p_end = off+1;
    return TRUE;
}

static int
md_is_autolink_email(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end)
{
    OFF off = beg + 1;
    int label_len;

    MD_ASSERT(CH(beg) == _T('<'));

    /* The code should correspond to this regexp:
            /^[a-zA-Z0-9.!#$%&'*+\/=?^_`{|}~-]+
            @[a-zA-Z0-9](?:[a-zA-Z0-9-]{0,61}[a-zA-Z0-9])?
            (?:\.[a-zA-Z0-9](?:[a-zA-Z0-9-]{0,61}[a-zA-Z0-9])?)*$/
     */

    /* Username (before '@'). */
    while(off < max_end  &&  (ISALNUM(off) || ISANYOF(off, _T(".!#$%&'*+/=?^_`{|}~-"))))
        off++;
    if(off <= beg+1)
        return FALSE;

    /* '@' */
    if(off >= max_end  ||  CH(off) != _T('@'))
        return FALSE;
    off++;

    /* Labels delimited with '.'; each label is sequence of 1 - 63 alnum
     * characters or '-', but '-' is not allowed as first or last char. */
    label_len = 0;
    while(off < max_end) {
        if(ISALNUM(off))
            label_len++;
        else if(CH(off) == _T('-')  &&  label_len > 0)
            label_len++;
        else if(CH(off) == _T('.')  &&  label_len > 0  &&  CH(off-1) != _T('-'))
            label_len = 0;
        else
            break;

        if(label_len > 63)
            return FALSE;

        off++;
    }

    if(label_len <= 0  || off >= max_end  ||  CH(off) != _T('>') ||  CH(off-1) == _T('-'))
        return FALSE;

    *p_end = off+1;
    return TRUE;
}

static int
md_is_autolink(MD_CTX* ctx, OFF beg, OFF max_end, OFF* p_end, int* p_missing_mailto)
{
    if(md_is_autolink_uri(ctx, beg, max_end, p_end)) {
        *p_missing_mailto = FALSE;
        return TRUE;
    }

    if(md_is_autolink_email(ctx, beg, max_end, p_end)) {
        *p_missing_mailto = TRUE;
        return TRUE;
    }

    return FALSE;
}

static int
md_collect_marks(MD_CTX* ctx, const MD_LINE* lines, int n_lines, int table_mode)
{
    const MD_LINE* line_term = lines + n_lines;
    const MD_LINE* line;
    int ret = 0;
    MD_MARK* mark;
    OFF codespan_last_potential_closers[CODESPAN_MARK_MAXLEN] = { 0 };
    int codespan_scanned_till_paragraph_end = FALSE;

    for(line = lines; line < line_term; line++) {
        OFF off = line->beg;
        OFF line_end = line->end;

        while(TRUE) {
            CHAR ch;

#ifdef MD4C_USE_UTF16
    /* For UTF-16, mark_char_map[] covers only ASCII. */
    #define IS_MARK_CHAR(off)   ((CH(off) < SIZEOF_ARRAY(ctx->mark_char_map))  &&  \
                                (ctx->mark_char_map[(unsigned char) CH(off)]))
#else
    /* For 8-bit encodings, mark_char_map[] covers all 256 elements. */
    #define IS_MARK_CHAR(off)   (ctx->mark_char_map[(unsigned char) CH(off)])
#endif

            /* Optimization: Use some loop unrolling. */
            while(off + 3 < line_end  &&  !IS_MARK_CHAR(off+0)  &&  !IS_MARK_CHAR(off+1)
                                      &&  !IS_MARK_CHAR(off+2)  &&  !IS_MARK_CHAR(off+3))
                off += 4;
            while(off < line_end  &&  !IS_MARK_CHAR(off+0))
                off++;

            if(off >= line_end)
                break;

            ch = CH(off);

            /* A backslash escape.
             * It can go beyond line->end as it may involve escaped new
             * line to form a hard break. */
            if(ch == _T('\\')  &&  off+1 < ctx->size  &&  (ISPUNCT(off+1) || ISNEWLINE(off+1))) {
                /* Hard-break cannot be on the last line of the block. */
                if(!ISNEWLINE(off+1)  ||  line+1 < line_term)
                    PUSH_MARK(ch, off, off+2, MD_MARK_RESOLVED);
                off += 2;
                continue;
            }

            /* A potential (string) emphasis start/end. */
            if(ch == _T('*')  ||  ch == _T('_')) {
                OFF tmp = off+1;
                int left_level;     /* What precedes: 0 = whitespace; 1 = punctuation; 2 = other char. */
                int right_level;    /* What follows: 0 = whitespace; 1 = punctuation; 2 = other char. */

                while(tmp < line_end  &&  CH(tmp) == ch)
                    tmp++;

                if(off == line->beg  ||  ISUNICODEWHITESPACEBEFORE(off))
                    left_level = 0;
                else if(ISUNICODEPUNCTBEFORE(off))
                    left_level = 1;
                else
                    left_level = 2;

                if(tmp == line_end  ||  ISUNICODEWHITESPACE(tmp))
                    right_level = 0;
                else if(ISUNICODEPUNCT(tmp))
                    right_level = 1;
                else
                    right_level = 2;

                /* Intra-word underscore doesn't have special meaning. */
                if(ch == _T('_')  &&  left_level == 2  &&  right_level == 2) {
                    left_level = 0;
                    right_level = 0;
                }

                if(left_level != 0  ||  right_level != 0) {
                    unsigned flags = 0;

                    if(left_level > 0  &&  left_level >= right_level)
                        flags |= MD_MARK_POTENTIAL_CLOSER;
                    if(right_level > 0  &&  right_level >= left_level)
                        flags |= MD_MARK_POTENTIAL_OPENER;
                    if(left_level == 2  &&  right_level == 2)
                        flags |= MD_MARK_EMPH_INTRAWORD;

                    /* For "the rule of three" we need to remember the original
                     * size of the mark (modulo three), before we potentially
                     * split the mark when being later resolved partially by some
                     * shorter closer. */
                    switch((tmp - off) % 3) {
                        case 0: flags |= MD_MARK_EMPH_MOD3_0; break;
                        case 1: flags |= MD_MARK_EMPH_MOD3_1; break;
                        case 2: flags |= MD_MARK_EMPH_MOD3_2; break;
                    }

                    PUSH_MARK(ch, off, tmp, flags);

                    /* During resolving, multiple asterisks may have to be
                     * split into independent span start/ends. Consider e.g.
                     * "**foo* bar*". Therefore we push also some empty dummy
                     * marks to have enough space for that. */
                    off++;
                    while(off < tmp) {
                        PUSH_MARK('D', off, off, 0);
                        off++;
                    }
                    continue;
                }

                off = tmp;
                continue;
            }

            /* A potential code span start/end. */
            if(ch == _T('`')) {
                OFF opener_beg, opener_end;
                OFF closer_beg, closer_end;
                int is_code_span;

                is_code_span = md_is_code_span(ctx, line, line_term - line, off,
                                    &opener_beg, &opener_end, &closer_beg, &closer_end,
                                    codespan_last_potential_closers,
                                    &codespan_scanned_till_paragraph_end);
                if(is_code_span) {
                    PUSH_MARK(_T('`'), opener_beg, opener_end, MD_MARK_OPENER | MD_MARK_RESOLVED);
                    PUSH_MARK(_T('`'), closer_beg, closer_end, MD_MARK_CLOSER | MD_MARK_RESOLVED);
                    ctx->marks[ctx->n_marks-2].next = ctx->n_marks-1;
                    ctx->marks[ctx->n_marks-1].prev = ctx->n_marks-2;

                    off = closer_end;

                    /* Advance the current line accordingly. */
                    if(off > line_end) {
                        line = md_lookup_line(off, line, line_term - line);
                        if(NULL == line) exit(15);
                        line_end = line->end;
                    }
                    continue;
                }

                off = opener_end;
                continue;
            }

            /* A potential entity start. */
            if(ch == _T('&')) {
                PUSH_MARK(ch, off, off+1, MD_MARK_POTENTIAL_OPENER);
                off++;
                continue;
            }

            /* A potential entity end. */
            if(ch == _T(';')) {
                /* We surely cannot be entity unless the previous mark is '&'. */
                if(ctx->n_marks > 0  &&  ctx->marks[ctx->n_marks-1].ch == _T('&'))
                    PUSH_MARK(ch, off, off+1, MD_MARK_POTENTIAL_CLOSER);

                off++;
                continue;
            }

            /* A potential autolink or raw HTML start/end. */
            if(ch == _T('<')) {
                int is_autolink;
                OFF autolink_end;
                int missing_mailto;

                if(!(ctx->parser.flags & MD_FLAG_NOHTMLSPANS)) {
                    int is_html;
                    OFF html_end;

                    /* Given the nature of the raw HTML, we have to recognize
                     * it here. Doing so later in md_analyze_lt_gt() could
                     * open can of worms of quadratic complexity. */
                    is_html = md_is_html_any(ctx, line, line_term - line, off,
                                    lines[n_lines-1].end, &html_end);
                    if(is_html) {
                        PUSH_MARK(_T('<'), off, off, MD_MARK_OPENER | MD_MARK_RESOLVED);
                        PUSH_MARK(_T('>'), html_end, html_end, MD_MARK_CLOSER | MD_MARK_RESOLVED);
                        ctx->marks[ctx->n_marks-2].next = ctx->n_marks-1;
                        ctx->marks[ctx->n_marks-1].prev = ctx->n_marks-2;
                        off = html_end;

                        /* Advance the current line accordingly. */
                        if(off > line_end) {
                            line = md_lookup_line(off, line, line_term - line);
                            line_end = line->end;
                        }
                        continue;
                    }
                }

                is_autolink = md_is_autolink(ctx, off, lines[n_lines-1].end,
                                    &autolink_end, &missing_mailto);
                if(is_autolink) {
                    PUSH_MARK((missing_mailto ? _T('@') : _T('<')), off, off+1,
                                MD_MARK_OPENER | MD_MARK_RESOLVED | MD_MARK_AUTOLINK);
                    PUSH_MARK(_T('>'), autolink_end-1, autolink_end,
                                MD_MARK_CLOSER | MD_MARK_RESOLVED | MD_MARK_AUTOLINK);
                    ctx->marks[ctx->n_marks-2].next = ctx->n_marks-1;
                    ctx->marks[ctx->n_marks-1].prev = ctx->n_marks-2;
                    off = autolink_end;
                    continue;
                }

                off++;
                continue;
            }

            /* A potential link or its part. */
            if(ch == _T('[')  ||  (ch == _T('!') && off+1 < line_end && CH(off+1) == _T('['))) {
                OFF tmp = (ch == _T('[') ? off+1 : off+2);
                PUSH_MARK(ch, off, tmp, MD_MARK_POTENTIAL_OPENER);
                off = tmp;
                /* Two dummies to make enough place for data we need if it is
                 * a link. */
                PUSH_MARK('D', off, off, 0);
                PUSH_MARK('D', off, off, 0);
                continue;
            }
            if(ch == _T(']')) {
                PUSH_MARK(ch, off, off+1, MD_MARK_POTENTIAL_CLOSER);
                off++;
                continue;
            }

            /* A potential permissive e-mail autolink. */
            if(ch == _T('@')) {
                if(line->beg + 1 <= off  &&  ISALNUM(off-1)  &&
                    off + 3 < line->end  &&  ISALNUM(off+1))
                {
                    PUSH_MARK(ch, off, off+1, MD_MARK_POTENTIAL_OPENER);
                    /* Push a dummy as a reserve for a closer. */
                    PUSH_MARK('D', off, off, 0);
                }

                off++;
                continue;
            }

            /* A potential permissive URL autolink. */
            if(ch == _T(':')) {
                static struct {
                    const CHAR* scheme;
                    SZ scheme_size;
                    const CHAR* suffix;
                    SZ suffix_size;
                } scheme_map[] = {
                    /* In the order from the most frequently used, arguably. */
                    { _T("http"), 4,    _T("//"), 2 },
                    { _T("https"), 5,   _T("//"), 2 },
                    { _T("ftp"), 3,     _T("//"), 2 }
                };
                int scheme_index;

                for(scheme_index = 0; scheme_index < (int) SIZEOF_ARRAY(scheme_map); scheme_index++) {
                    const CHAR* scheme = scheme_map[scheme_index].scheme;
                    const SZ scheme_size = scheme_map[scheme_index].scheme_size;
                    const CHAR* suffix = scheme_map[scheme_index].suffix;
                    const SZ suffix_size = scheme_map[scheme_index].suffix_size;

                    if(line->beg + scheme_size <= off  &&  md_ascii_eq(STR(off-scheme_size), scheme, scheme_size)  &&
                        (line->beg + scheme_size == off || ISWHITESPACE(off-scheme_size-1) || ISANYOF(off-scheme_size-1, _T("*_~([")))  &&
                        off + 1 + suffix_size < line->end  &&  md_ascii_eq(STR(off+1), suffix, suffix_size))
                    {
                        PUSH_MARK(ch, off-scheme_size, off+1+suffix_size, MD_MARK_POTENTIAL_OPENER);
                        /* Push a dummy as a reserve for a closer. */
                        PUSH_MARK('D', off, off, 0);
                        off += 1 + suffix_size;
                        break;
                    }
                }

                off++;
                continue;
            }

            /* A potential permissive WWW autolink. */
            if(ch == _T('.')) {
                if(line->beg + 3 <= off  &&  md_ascii_eq(STR(off-3), _T("www"), 3)  &&
                    (line->beg + 3 == off || ISWHITESPACE(off-4) || ISANYOF(off-4, _T("*_~([")))  &&
                    off + 1 < line_end)
                {
                    PUSH_MARK(ch, off-3, off+1, MD_MARK_POTENTIAL_OPENER);
                    /* Push a dummy as a reserve for a closer. */
                    PUSH_MARK('D', off, off, 0);
                    off++;
                    continue;
                }

                off++;
                continue;
            }

            /* A potential table cell boundary or wiki link label delimiter. */
            if((table_mode || ctx->parser.flags & MD_FLAG_WIKILINKS) && ch == _T('|')) {
                PUSH_MARK(ch, off, off+1, 0);
                off++;
                continue;
            }

            /* A potential strikethrough start/end. */
            if(ch == _T('~')) {
                OFF tmp = off+1;

                while(tmp < line_end  &&  CH(tmp) == _T('~'))
                    tmp++;

                if(tmp - off < 3) {
                    unsigned flags = 0;

                    if(tmp < line_end  &&  !ISUNICODEWHITESPACE(tmp))
                        flags |= MD_MARK_POTENTIAL_OPENER;
                    if(off > line->beg  &&  !ISUNICODEWHITESPACEBEFORE(off))
                        flags |= MD_MARK_POTENTIAL_CLOSER;
                    if(flags != 0)
                        PUSH_MARK(ch, off, tmp, flags);
                }

                off = tmp;
                continue;
            }

            /* A potential equation start/end */
            if(ch == _T('$')) {
                /* We can have at most two consecutive $ signs,
                 * where two dollar signs signify a display equation. */
                OFF tmp = off+1;

                while(tmp < line_end && CH(tmp) == _T('$'))
                    tmp++;

                if (tmp - off <= 2)
                    PUSH_MARK(ch, off, tmp, MD_MARK_POTENTIAL_OPENER | MD_MARK_POTENTIAL_CLOSER);
                off = tmp;
                continue;
            }

            /* Turn non-trivial whitespace into single space. */
            if(ISWHITESPACE_(ch)) {
                OFF tmp = off+1;

                while(tmp < line_end  &&  ISWHITESPACE(tmp))
                    tmp++;

                if(tmp - off > 1  ||  ch != _T(' '))
                    PUSH_MARK(ch, off, tmp, MD_MARK_RESOLVED);

                off = tmp;
                continue;
            }

            /* NULL character. */
            if(ch == _T('\0')) {
                PUSH_MARK(ch, off, off+1, MD_MARK_RESOLVED);
                off++;
                continue;
            }

            off++;
        }
    }

    /* Add a dummy mark at the end of the mark vector to simplify
     * process_inlines(). */
    PUSH_MARK(127, ctx->size, ctx->size, MD_MARK_RESOLVED);

abort:
    return ret;
}

static void
md_analyze_bracket(MD_CTX* ctx, int mark_index)
{
    /* We cannot really resolve links here as for that we would need
     * more context. E.g. a following pair of brackets (reference link),
     * or enclosing pair of brackets (if the inner is the link, the outer
     * one cannot be.)
     *
     * Therefore we here only construct a list of '[' ']' pairs ordered by
     * position of the closer. This allows us to analyze what is or is not
     * link in the right order, from inside to outside in case of nested
     * brackets.
     *
     * The resolving itself is deferred to md_resolve_links().
     */

    MD_MARK* mark = &ctx->marks[mark_index];

    if(mark->flags & MD_MARK_POTENTIAL_OPENER) {
        if(BRACKET_OPENERS.head != -1)
            ctx->marks[BRACKET_OPENERS.tail].flags |= MD_MARK_HASNESTEDBRACKETS;

        md_mark_chain_append(ctx, &BRACKET_OPENERS, mark_index);
        return;
    }

    if(BRACKET_OPENERS.tail >= 0) {
        /* Pop the opener from the chain. */
        int opener_index = BRACKET_OPENERS.tail;
        MD_MARK* opener = &ctx->marks[opener_index];
        if(opener->prev >= 0)
            ctx->marks[opener->prev].next = -1;
        else
            BRACKET_OPENERS.head = -1;
        BRACKET_OPENERS.tail = opener->prev;

        /* Interconnect the opener and closer. */
        opener->next = mark_index;
        mark->prev = opener_index;

        /* Add the pair into chain of potential links for md_resolve_links().
         * Note we misuse opener->prev for this as opener->next points to its
         * closer. */
        if(ctx->unresolved_link_tail >= 0)
            ctx->marks[ctx->unresolved_link_tail].prev = opener_index;
        else
            ctx->unresolved_link_head = opener_index;
        ctx->unresolved_link_tail = opener_index;
        opener->prev = -1;
    }
}

/* Forward declaration. */
static void md_analyze_link_contents(MD_CTX* ctx, const MD_LINE* lines, int n_lines,
                                     int mark_beg, int mark_end);

static int
md_resolve_links(MD_CTX* ctx, const MD_LINE* lines, int n_lines)
{
    int opener_index = ctx->unresolved_link_head;
    OFF last_link_beg = 0;
    OFF last_link_end = 0;
    OFF last_img_beg = 0;
    OFF last_img_end = 0;

    while(opener_index >= 0) {
        MD_MARK* opener = &ctx->marks[opener_index];
        int closer_index = opener->next;
        MD_MARK* closer = &ctx->marks[closer_index];
        int next_index = opener->prev;
        MD_MARK* next_opener;
        MD_MARK* next_closer;
        MD_LINK_ATTR attr;
        int is_link = FALSE;

        if(next_index >= 0) {
            next_opener = &ctx->marks[next_index];
            next_closer = &ctx->marks[next_opener->next];
        } else {
            next_opener = NULL;
            next_closer = NULL;
        }

        /* If nested ("[ [ ] ]"), we need to make sure that:
         *   - The outer does not end inside of (...) belonging to the inner.
         *   - The outer cannot be link if the inner is link (i.e. not image).
         *
         * (Note we here analyze from inner to outer as the marks are ordered
         * by closer->beg.)
         */
        if((opener->beg < last_link_beg  &&  closer->end < last_link_end)  ||
           (opener->beg < last_img_beg  &&  closer->end < last_img_end)  ||
           (opener->beg < last_link_end  &&  opener->ch == '['))
        {
            opener_index = next_index;
            continue;
        }

        /* Recognize and resolve wiki links.
         * Wiki-links maybe '[[destination]]' or '[[destination|label]]'.
         */
        if ((ctx->parser.flags & MD_FLAG_WIKILINKS) &&
            (opener->end - opener->beg == 1) &&         /* not image */
            next_opener != NULL &&                      /* double '[' opener */
            next_opener->ch == '[' &&
            (next_opener->beg == opener->beg - 1) &&
            (next_opener->end - next_opener->beg == 1) &&
            next_closer != NULL &&                      /* double ']' closer */
            next_closer->ch == ']' &&
            (next_closer->beg == closer->beg + 1) &&
            (next_closer->end - next_closer->beg == 1))
        {
            MD_MARK* delim = NULL;
            int delim_index;
            OFF dest_beg, dest_end;

            is_link = TRUE;

            /* We don't allow destination to be longer than 100 characters.