@descent-vtt/spec-graph 0.4.0 → 0.5.0

package/dist/regex.js

@@ -0,0 +1,708 @@
+/**
+ * A regular-expression matcher that cannot backtrack.
+ *
+ * `~=` is the one predicate in the selector language that used to hand a
+ * user-supplied string to `RegExp` and wait. That is tolerable at a prompt,
+ * where a human can press Ctrl-C. It is not tolerable in `.spec-graph.json`,
+ * where the same string runs unattended in CI against every node in the corpus.
+ *
+ * The cost is neither theoretical nor exotic. `^([A-Za-z0-9_]+[ ]?)+$` is what
+ * somebody writes to say "the title is words separated by single spaces", and
+ * against a title that does not match, V8 tries one branch per way of cutting
+ * each word into pieces:
+ *
+ * ```text
+ *                                                          RegExp      here
+ * the quick brown fox jumps over the lazy dog!               0.9s      11us
+ * the quick brown fox jumps over the lazy dog and!           5.4s      12us
+ * the quick brown fox jumps over the lazy dog and keeps!     103s      13us
+ * ```
+ *
+ * Each word of k letters can be cut into pieces 2^(k-1) ways, and the engine
+ * tries the product of those, which is why adding one five-letter word costs
+ * another factor of sixteen. A sixty-character title does not finish this year.
+ *
+ * So a pattern here is compiled to a **non-deterministic finite automaton** and
+ * simulated over the subject one character at a time (Thompson, 1968), keeping
+ * a set of live states instead of a stack of choices. Each state is visited at
+ * most once per position, which makes the work exactly O(pattern x subject):
+ * the right-hand column above, growing with the subject and not with the
+ * pattern's ambiguity.
+ *
+ * What it costs is a dialect. Backreferences and lookaround are not regular and
+ * have no place in a linear automaton, so they are **rejected while compiling**,
+ * with a message and an offset, rather than accepted and then slow. That is the
+ * whole trade: a predicate that finishes, against a short list of constructs
+ * that a documentation selector never needed.
+ *
+ * Matching is case-insensitive and unanchored because that is what `~=` has
+ * always been - `new RegExp(source, 'i')` and `.test()`. There are no capture
+ * groups at all: nothing asks where the match was, only whether there was one,
+ * which is why this is a plain NFA simulation rather than a Pike VM.
+ *
+ * See ADR-0017.
+ */
+/* -------------------------------------------------------------------------- */
+/* Errors                                                                     */
+/* -------------------------------------------------------------------------- */
+/** A pattern that will not compile, and where in it the problem starts. */
+export class PatternError extends Error {
+    /** Offset into the pattern source. */
+    offset;
+    constructor(message, offset) {
+        super(message);
+        this.name = 'PatternError';
+        this.offset = offset;
+    }
+}
+/**
+ * Ceiling on the compiled program.
+ *
+ * Counted repetition is compiled by copying - `(ab){20}` is twenty copies -
+ * which is part of what keeps the simulation flat, and also what makes
+ * `(a{99}){99}` into ten thousand states. The bound has to live somewhere, and
+ * a number checked while compiling is the only place it can be enforced without
+ * measuring anything at match time.
+ *
+ * Far above any real selector: the widest pattern in this repository's own
+ * tests compiles to seventeen states.
+ */
+const MAX_PROGRAM = 4096;
+/** `.` - anything but a line terminator, which is the default `RegExp` reading. */
+const DOT = { negated: true, ranges: [[0x0a, 0x0a], [0x0d, 0x0d], [0x2028, 0x2029]] };
+const DIGIT = [[0x30, 0x39]];
+const WORD = [
+    [0x30, 0x39],
+    [0x41, 0x5a],
+    [0x5f, 0x5f],
+    [0x61, 0x7a],
+];
+/** `\s`, taken from the specification rather than guessed at. */
+const SPACE = [
+    [0x09, 0x0d],
+    [0x20, 0x20],
+    [0xa0, 0xa0],
+    [0x1680, 0x1680],
+    [0x2000, 0x200a],
+    [0x2028, 0x2029],
+    [0x202f, 0x202f],
+    [0x205f, 0x205f],
+    [0x3000, 0x3000],
+    [0xfeff, 0xfeff],
+];
+/**
+ * Folds a code unit the way a case-insensitive `RegExp` does.
+ *
+ * This is `Canonicalize` from the language specification, non-Unicode mode:
+ * upper-case the character, keep it unchanged if that produced anything other
+ * than one code unit, and keep it unchanged if upper-casing dragged a
+ * non-ASCII character into ASCII. The last clause is the one nobody expects and
+ * the reason U+017F (long s) does not match `s`.
+ */
+function canonical(code) {
+    const upper = String.fromCharCode(code).toUpperCase();
+    if (upper.length !== 1)
+        return code;
+    const folded = upper.charCodeAt(0);
+    if (code >= 128 && folded < 128)
+        return code;
+    return folded;
+}
+/**
+ * Whether a set accepts a code unit, case-insensitively.
+ *
+ * The specification canonicalises both sides: a character matches a set when
+ * *some* member of the set folds to the same thing it does. A single-character
+ * member can be asked that exactly. A range cannot without a fold table for
+ * every code unit, so a range is asked the question backwards - does the
+ * character, or either of its own case forms, fall inside it - which is exact
+ * for every character whose fold is its own case change, and therefore for
+ * every range anyone has written.
+ */
+function accepts(set, code) {
+    const folded = canonical(code);
+    let hit = false;
+    for (const [low, high] of set.ranges) {
+        if (low === high ? low === code || canonical(low) === folded : inRange(low, high, code)) {
+            hit = true;
+            break;
+        }
+    }
+    return set.negated ? !hit : hit;
+}
+/**
+ * Whether some member of a range folds to the same thing a code unit does.
+ *
+ * Asked backwards, by trying the character's own case forms against the raw
+ * range - and then checking that the form found there really does fold to the
+ * same place, which is the part that is easy to leave out. U+017F upper-cases
+ * to `S`, so without that check it matches `[A-Z]`, and it must not: the fold
+ * keeps it outside ASCII, which is the same clause `isWordChar` gets wrong if
+ * nobody is watching. The differential corpus caught both.
+ */
+function inRange(low, high, code) {
+    if (code >= low && code <= high)
+        return true;
+    const folded = canonical(code);
+    const char = String.fromCharCode(code);
+    for (const variant of [char.toLowerCase(), char.toUpperCase()]) {
+        if (variant.length !== 1)
+            continue;
+        const candidate = variant.charCodeAt(0);
+        if (candidate >= low && candidate <= high && canonical(candidate) === folded)
+            return true;
+    }
+    return false;
+}
+function within(ranges, code) {
+    for (const [low, high] of ranges) {
+        if (code >= low && code <= high)
+            return true;
+    }
+    return false;
+}
+/**
+ * Word characters for `\b`, which is exactly `\w` here.
+ *
+ * The specification does widen this set under case-insensitivity - to every
+ * character that folds into it, which is U+017F (long s) and U+212A (the
+ * Kelvin sign) - but only in Unicode mode, where `Canonicalize` has no clause
+ * keeping a non-ASCII character out of ASCII. This is the non-Unicode dialect,
+ * so both stay outside and the table for `\w` answers for `\b` as well.
+ *
+ * Written down because the first version of this function widened the set, and
+ * the differential test against `RegExp` is what said otherwise. Reading the
+ * clause was not enough; running it was. See ADR-0017.
+ */
+function isWordChar(code) {
+    return within(WORD, code);
+}
+/* -------------------------------------------------------------------------- */
+/* Parsing                                                                    */
+/* -------------------------------------------------------------------------- */
+class PatternParser {
+    source;
+    position = 0;
+    constructor(source) {
+        this.source = source;
+    }
+    parse() {
+        const node = this.parseAlternation();
+        // Every character is an atom except the two that close something, and
+        // `parseSequence` stops on both, so anything left over is an unopened `)`.
+        if (this.position < this.source.length)
+            throw new PatternError('unmatched ")"', this.position);
+        return node;
+    }
+    parseAlternation() {
+        const options = [this.parseSequence()];
+        while (this.peek() === '|') {
+            this.position += 1;
+            options.push(this.parseSequence());
+        }
+        return options.length === 1 ? options[0] : { kind: 'alternate', options };
+    }
+    parseSequence() {
+        const parts = [];
+        for (;;) {
+            const char = this.peek();
+            if (char === undefined || char === '|' || char === ')')
+                break;
+            parts.push(this.parseQuantified());
+        }
+        if (parts.length === 0)
+            return { kind: 'empty' };
+        return parts.length === 1 ? parts[0] : { kind: 'concat', parts };
+    }
+    parseQuantified() {
+        const start = this.position;
+        // A group makes its contents quantifiable: `RegExp` rejects `\b*` and
+        // accepts `(\b)*`, which is the same nothing said twice, and this parser
+        // treats a group as transparent - so whether one was written is the only
+        // thing left that distinguishes the two.
+        const grouped = this.peek() === '(';
+        const atom = this.parseAtom();
+        const bounds = this.parseQuantifier();
+        if (bounds === null)
+            return atom;
+        if (atom.kind === 'assert' && !grouped) {
+            throw new PatternError('nothing to repeat: a quantifier cannot follow an anchor', start);
+        }
+        return { kind: 'repeat', node: atom, min: bounds.min, max: bounds.max };
+    }
+    /** `*`, `+`, `?`, `{n}`, `{n,}`, `{n,m}`, and a trailing `?` that is ignored. */
+    parseQuantifier() {
+        const char = this.peek();
+        let bounds = null;
+        if (char === '*')
+            bounds = { min: 0, max: Infinity };
+        else if (char === '+')
+            bounds = { min: 1, max: Infinity };
+        else if (char === '?')
+            bounds = { min: 0, max: 1 };
+        if (bounds !== null)
+            this.position += 1;
+        else if (char === '{')
+            bounds = this.parseBraces();
+        if (bounds === null)
+            return null;
+        // Laziness decides *which* match is found, and nothing here asks: the
+        // question is whether one exists, and a lazy quantifier accepts the same
+        // language as a greedy one.
+        if (this.peek() === '?')
+            this.position += 1;
+        return bounds;
+    }
+    /**
+     * Reads `{n,m}`, or nothing.
+     *
+     * A `{` that does not open a counted repetition is a literal brace, which is
+     * what `RegExp` does with `a{` and what any pattern written against a
+     * document full of `{0}` placeholders depends on.
+     */
+    parseBraces() {
+        const match = /^\{(\d+)(?:(,)(\d*))?\}/.exec(this.source.slice(this.position));
+        if (!match)
+            return null;
+        const min = Number(match[1]);
+        const max = match[2] === undefined ? min : match[3] === '' ? Infinity : Number(match[3]);
+        if (min > max)
+            throw new PatternError('numbers out of order in a counted repetition', this.position);
+        this.position += match[0].length;
+        return { min, max };
+    }
+    parseAtom() {
+        const start = this.position;
+        const char = this.peek();
+        switch (char) {
+            case undefined:
+                throw new PatternError('pattern ends where an expression was expected', start);
+            case '(':
+                return this.parseGroup();
+            case '[':
+                return { kind: 'char', set: this.parseClass() };
+            case '.':
+                this.position += 1;
+                return { kind: 'char', set: DOT };
+            case '^':
+                this.position += 1;
+                return { kind: 'assert', assertion: 'start' };
+            case '$':
+                this.position += 1;
+                return { kind: 'assert', assertion: 'end' };
+            case '*':
+            case '+':
+                throw new PatternError(`nothing to repeat before "${char}"`, start);
+            case '\\':
+                return this.parseEscape();
+            default:
+                this.position += 1;
+                return { kind: 'char', set: literal(char.charCodeAt(0)) };
+        }
+    }
+    /**
+     * A group, which is never a capture.
+     *
+     * `(a)` and `(?:a)` compile to the same automaton because nothing downstream
+     * can ask what `(a)` caught. A named group is accepted and its name dropped,
+     * so a pattern copied out of somewhere else keeps working.
+     */
+    parseGroup() {
+        const start = this.position;
+        this.position += 1;
+        if (this.source.startsWith('?:', this.position)) {
+            this.position += 2;
+        }
+        else if (this.peek() === '?') {
+            const named = /^\?<([A-Za-z_$][\w$]*)>/.exec(this.source.slice(this.position));
+            if (named) {
+                this.position += named[0].length;
+            }
+            else if (/^\?(?:=|!|<=|<!)/.test(this.source.slice(this.position))) {
+                throw new PatternError('lookaround is not supported: an automaton that cannot backtrack cannot look ahead - use ^= $= or *= for a fixed prefix, suffix or substring', start);
+            }
+            else {
+                throw new PatternError('unsupported group: expected (?:...) or (?<name>...)', start);
+            }
+        }
+        const inner = this.parseAlternation();
+        if (this.peek() !== ')')
+            throw new PatternError('unmatched "("', start);
+        this.position += 1;
+        return inner;
+    }
+    /**
+     * Reads a `[...]` class.
+     *
+     * `[]` matches nothing and `[^]` matches anything, which reads like a mistake
+     * and is what every JavaScript engine does: the class ends at the first `]`,
+     * even when that `]` is the first character inside it.
+     */
+    parseClass() {
+        const start = this.position;
+        this.position += 1;
+        const negated = this.peek() === '^';
+        if (negated)
+            this.position += 1;
+        const ranges = [];
+        while (this.position < this.source.length && this.peek() !== ']') {
+            const low = this.parseClassAtom();
+            // A `-` between two single characters is a range. Anywhere else - at
+            // either end of the class, or beside a `\d`-style set, which has no
+            // endpoint - it is a literal hyphen, exactly as Annex B has it.
+            if (this.peek() !== '-' || this.source[this.position + 1] === ']' || this.position + 1 >= this.source.length) {
+                push(ranges, low);
+                continue;
+            }
+            const dash = this.position;
+            this.position += 1;
+            const high = this.parseClassAtom();
+            if (typeof low === 'number' && typeof high === 'number') {
+                if (low > high)
+                    throw new PatternError('characters out of order in a character class', dash);
+                ranges.push([low, high]);
+                continue;
+            }
+            push(ranges, low);
+            ranges.push([0x2d, 0x2d]);
+            push(ranges, high);
+        }
+        if (this.peek() !== ']')
+            throw new PatternError('unterminated character class', start);
+        this.position += 1;
+        return { negated, ranges };
+    }
+    /** One member of a class: a code unit, or a whole set for `\d` and its kind. */
+    parseClassAtom() {
+        if (this.peek() !== '\\') {
+            const code = this.peek().charCodeAt(0);
+            this.position += 1;
+            return code;
+        }
+        // `\b` is a backspace inside a class and a word boundary outside one, which
+        // is the only place the two readings of an escape diverge.
+        if (this.source[this.position + 1] === 'b') {
+            this.position += 2;
+            return 0x08;
+        }
+        if (this.source[this.position + 1] === 'B') {
+            throw new PatternError('"\\B" is an assertion, not a character, so it cannot be in a class', this.position);
+        }
+        const set = this.parseCharacterEscape();
+        if (!set.negated && set.ranges.length === 1) {
+            const [low, high] = set.ranges[0];
+            if (low === high)
+                return low;
+        }
+        return set.negated ? invert(set.ranges) : set.ranges;
+    }
+    /** An escape outside a class, where `\b` and `\B` are assertions. */
+    parseEscape() {
+        const next = this.source[this.position + 1];
+        if (next === 'b' || next === 'B') {
+            this.position += 2;
+            return { kind: 'assert', assertion: next === 'b' ? 'word-boundary' : 'not-word-boundary' };
+        }
+        return { kind: 'char', set: this.parseCharacterEscape() };
+    }
+    /** Every escape that stands for a character or a set of them. */
+    parseCharacterEscape() {
+        const start = this.position;
+        this.position += 1;
+        const char = this.peek();
+        if (char === undefined)
+            throw new PatternError('pattern ends with a backslash', start);
+        this.position += 1;
+        switch (char) {
+            case 'd':
+                return { negated: false, ranges: DIGIT };
+            case 'D':
+                return { negated: true, ranges: DIGIT };
+            case 'w':
+                return { negated: false, ranges: WORD };
+            case 'W':
+                return { negated: true, ranges: WORD };
+            case 's':
+                return { negated: false, ranges: SPACE };
+            case 'S':
+                return { negated: true, ranges: SPACE };
+            case 'n':
+                return literal(0x0a);
+            case 'r':
+                return literal(0x0d);
+            case 't':
+                return literal(0x09);
+            case 'f':
+                return literal(0x0c);
+            case 'v':
+                return literal(0x0b);
+            case '0':
+                if (/^\d/.test(this.source.slice(this.position))) {
+                    throw new PatternError('an octal escape is not supported - write \\xHH or \\uHHHH', start);
+                }
+                return literal(0);
+            case 'x':
+                return literal(this.readHex(2, start));
+            case 'u':
+                return literal(this.readHex(4, start));
+            case 'c': {
+                const letter = this.peek();
+                if (letter === undefined || !/[A-Za-z]/.test(letter)) {
+                    throw new PatternError('expected a letter after "\\c"', start);
+                }
+                this.position += 1;
+                return literal(letter.charCodeAt(0) % 32);
+            }
+            case 'k':
+                throw new PatternError('a named backreference is not supported: a backreference is not a regular expression, and this is an automaton', start);
+            case 'p':
+            case 'P':
+                throw new PatternError(`"\\${char}" needs a table of Unicode properties, which this matcher does not carry - write the characters out as a [class]`, start);
+            default:
+                if (/[1-9]/.test(char)) {
+                    throw new PatternError('a backreference is not supported: it is not a regular expression, and this is an automaton', start);
+                }
+                if (/[A-Za-z]/.test(char)) {
+                    // `RegExp` reads an unknown letter escape as the letter itself, so
+                    // `\A` quietly matches a capital A and `\z` a lower-case z. Saying so
+                    // is worth more than agreeing with a rule nobody wanted.
+                    throw new PatternError(`unknown escape "\\${char}"`, start);
+                }
+                return literal(char.charCodeAt(0));
+        }
+    }
+    readHex(digits, start) {
+        const raw = this.source.slice(this.position, this.position + digits);
+        if (raw.length !== digits || !/^[0-9a-fA-F]+$/.test(raw)) {
+            throw new PatternError(`expected ${digits} hex digits after "\\${this.source[start + 1]}"`, start);
+        }
+        this.position += digits;
+        return Number.parseInt(raw, 16);
+    }
+    peek() {
+        return this.source[this.position];
+    }
+}
+function literal(code) {
+    return { negated: false, ranges: [[code, code]] };
+}
+function push(ranges, atom) {
+    if (typeof atom === 'number')
+        ranges.push([atom, atom]);
+    else
+        ranges.push(...atom);
+}
+/**
+ * Complement of a range list, so `[\D]` can sit inside a class.
+ *
+ * A negated escape is a set with no endpoints, and a class is a union of
+ * ranges, so the only way to put one inside the other is to spell out what it
+ * leaves.
+ */
+function invert(ranges) {
+    const sorted = [...ranges].sort((a, b) => a[0] - b[0]);
+    const out = [];
+    let cursor = 0;
+    for (const [low, high] of sorted) {
+        if (low > cursor)
+            out.push([cursor, low - 1]);
+        cursor = Math.max(cursor, high + 1);
+    }
+    if (cursor <= 0xffff)
+        out.push([cursor, 0xffff]);
+    return out;
+}
+class Compiler {
+    code = [];
+    add(instruction) {
+        if (this.code.length >= MAX_PROGRAM) {
+            throw new PatternError(`pattern is too large: counted repetition is compiled by copying, and this needs more than ${MAX_PROGRAM} states`, 0);
+        }
+        this.code.push(instruction);
+        return instruction;
+    }
+    get here() {
+        return this.code.length;
+    }
+    emitMatch() {
+        this.add({ op: 'match' });
+    }
+    compile(node) {
+        switch (node.kind) {
+            case 'empty':
+                return;
+            case 'char':
+                this.add({ op: 'char', set: node.set });
+                return;
+            case 'assert':
+                this.add({ op: 'assert', assertion: node.assertion });
+                return;
+            case 'concat':
+                for (const part of node.parts)
+                    this.compile(part);
+                return;
+            case 'alternate': {
+                // One split per option but the last. Each split falls through to its
+                // own branch and points its second arm at the next split; each branch
+                // ends in a jump past all of them.
+                const ends = [];
+                for (let i = 0; i < node.options.length - 1; i += 1) {
+                    const split = this.add({ op: 'split', x: 0, y: 0 });
+                    split.x = this.here;
+                    this.compile(node.options[i]);
+                    ends.push(this.add({ op: 'jump', to: 0 }));
+                    split.y = this.here;
+                }
+                this.compile(node.options[node.options.length - 1]);
+                for (const end of ends)
+                    end.to = this.here;
+                return;
+            }
+            case 'repeat':
+                this.repeat(node.node, node.min, node.max);
+                return;
+        }
+    }
+    repeat(node, min, max) {
+        for (let i = 0; i < min; i += 1)
+            this.compile(node);
+        if (max === Infinity) {
+            const split = this.add({ op: 'split', x: 0, y: 0 });
+            const body = this.here;
+            split.x = body;
+            this.compile(node);
+            this.add({ op: 'jump', to: body - 1 });
+            split.y = this.here;
+            return;
+        }
+        // `{2,5}` is two copies and then three optional ones, each able to leave
+        // for the far end rather than for the next copy.
+        const exits = [];
+        for (let i = min; i < max; i += 1) {
+            const split = this.add({ op: 'split', x: 0, y: 0 });
+            split.x = this.here;
+            exits.push(split);
+            this.compile(node);
+        }
+        for (const split of exits)
+            split.y = this.here;
+    }
+}
+/**
+ * Compiles a pattern into a matcher.
+ *
+ * Throws {@link PatternError} on anything it cannot run in linear time, which
+ * is the point: the alternative to a message here is a build that stops.
+ */
+export function compilePattern(source) {
+    const tree = new PatternParser(source).parse();
+    const compiler = new Compiler();
+    compiler.compile(tree);
+    compiler.emitMatch();
+    const code = compiler.code;
+    // A pattern that begins by asserting the start of the subject cannot match
+    // anywhere else, so the simulation stops seeding new threads at every
+    // position. Nothing depends on this being noticed - it is the difference
+    // between one pass and two over a long body.
+    const anchored = startsAnchored(tree);
+    return {
+        size: code.length,
+        test: (subject) => run(code, subject, anchored),
+    };
+}
+/** True when every path through the pattern opens by asserting the start. */
+function startsAnchored(node) {
+    switch (node.kind) {
+        case 'assert':
+            return node.assertion === 'start';
+        case 'concat': {
+            const first = node.parts[0];
+            return first !== undefined && startsAnchored(first);
+        }
+        case 'alternate':
+            return node.options.every((option) => startsAnchored(option));
+        case 'repeat':
+            return node.min > 0 && startsAnchored(node.node);
+        default:
+            return false;
+    }
+}
+function run(code, subject, anchored) {
+    const length = subject.length;
+    // The position each state was last reached at. One visit per state per
+    // position is exactly what turns the exponent into a product.
+    const visited = new Int32Array(code.length).fill(-1);
+    let live = [];
+    for (let position = 0; position <= length; position += 1) {
+        if (position === 0 || !anchored) {
+            if (seedFrom(code, visited, live, 0, subject, position))
+                return true;
+        }
+        else if (live.length === 0) {
+            // An anchored pattern gets one chance, and every thread from it has died.
+            return false;
+        }
+        if (position === length)
+            return false;
+        const char = subject.charCodeAt(position);
+        const next = [];
+        for (const pc of live) {
+            const instruction = code[pc];
+            if (instruction.op !== 'char' || !accepts(instruction.set, char))
+                continue;
+            if (seedFrom(code, visited, next, pc + 1, subject, position + 1))
+                return true;
+        }
+        live = next;
+    }
+    return false;
+}
+/**
+ * Follows every zero-width transition out of one state.
+ *
+ * Returns true the moment a thread reaches the end of the pattern: the only
+ * question is whether a match exists, so there is no leftmost-longest to settle
+ * and no captures to keep, and the first thread to arrive is the answer.
+ */
+function seedFrom(code, visited, list, start, subject, position) {
+    const stack = [start];
+    while (stack.length > 0) {
+        const pc = stack.pop();
+        if (visited[pc] === position)
+            continue;
+        visited[pc] = position;
+        const instruction = code[pc];
+        switch (instruction.op) {
+            case 'match':
+                return true;
+            case 'jump':
+                stack.push(instruction.to);
+                break;
+            case 'split':
+                stack.push(instruction.y, instruction.x);
+                break;
+            case 'assert':
+                if (holds(instruction.assertion, subject, position))
+                    stack.push(pc + 1);
+                break;
+            case 'char':
+                list.push(pc);
+                break;
+        }
+    }
+    return false;
+}
+function holds(assertion, subject, position) {
+    switch (assertion) {
+        case 'start':
+            return position === 0;
+        case 'end':
+            return position === subject.length;
+        default: {
+            const before = position > 0 && isWordChar(subject.charCodeAt(position - 1));
+            const after = position < subject.length && isWordChar(subject.charCodeAt(position));
+            return (before !== after) === (assertion === 'word-boundary');
+        }
+    }
+}
+//# sourceMappingURL=regex.js.map