Changes in ParallelRegex.php

1c00c02121477c81b95fe750b94acdf109cba20a - fix(parser): validate inline formatting closers with a single memoized scan

Thu, 09 Jul 2026 12:09:52 +0000

fix(parser): validate inline formatting closers with a single memoized scanThe inline formatting modes only open a span when a valid closer existsahead. That check was a lookahead built on CONTENT_UNTIL_PARA, testedcharacter by character up to the next paragraph break and re-evaluatedfrom scratch for every opener candidate — openers times paragraphlength. With pcre.jit=0 a crafted 32KB page took 16s and an ordinary34KB page with long paragraphs 37s; with the JIT on (the PHP default)the per-character lookahead exhausted the JIT stack, the match silentlyfailed, and the formatting — or everything after it — rendered as plaintext.The check also decided the wrong thing. It scanned raw text, so a closerlookalike inside content the lexer consumes atomically — a nowiki or %%span, a backtick code span, a link, a URL — counted as a real closereven though the mode's exit pattern can never fire there. And it ignoredthe enclosing span: an inner delimiter whose only closer lay past thecloser of the mode it sits in was entered anyway, so a stray delimiterpaired with one in a following sibling span and dragged the boundaryalong — the `*` in ''glob/*.conf'' joined the `*` of the next ''...''span and corrupted the paragraph; the same held for //, ** and __ insidemonospace, and an emphasis opened inside ((...)) ran past the footnote's)) and the enclosing bold's **.Each formatting mode now declares its closer throughLexer::addCloserPattern(), mirroring addExitPattern(), and the lexeranswers "does a valid closer exist ahead" with one anchored possessivescan per range instead of a lookahead per opener:- The scan runs left to right from the opener, hopping over opaque spans derived from already-registered patterns — a plain or special match is consumed in one step, an entry into a verbatim mode (nowiki, the backtick code spans) extends to that mode's first exit — so a closer lookalike inside consumed content is never mistaken for a closer. Each hop finds the earliest of boundary, closer, or opaque span in a single leftmost search, keeping the check linear.- An opener is rejected when the nearest enclosing mode that has a closer of its own would close before the opener's own closer, so a delimiter that can never close within its span stays literal. That ancestor is found by walking the mode stack past modes that declare no closer (plugins, footnotes); the nearest guarded ancestor suffices, as it was itself validated against its own when it opened.- Both verdicts are memoized and reset per parse() run: a proven closer validates every earlier candidate, and a proven closer-free range rejects every later candidate before the next boundary. With the lexer consuming each opened span, the whole parse is linear in document size.Closer patterns match the closing delimiter itself with flanking contextin lookarounds — the convention exit patterns already follow — so closerpositions compare exactly across modes and a closer directly after aninner opener is seen. AbstractFormatting derives the closer from the exitpattern and registers it with the paragraph break as the boundary,preserving the rule that formatting never spans paragraphs; a mode withother needs can pass a different boundary or none.Footnote declares its )) as a closer rather than guarding its (( entrywith a (?=.*)) lookahead, so the footnote becomes a boundary the scansees and formatting inside it no longer pairs across the )); its closertakes no paragraph boundary, as footnotes are block-level. GfmEmphasisgains a closer pattern so single * emphasis is validated the same way,while its entry lookahead still enforces CommonMark nearest-delimiterpairing. GfmEmphasis and GfmStrong span bodies cannot contain theirdelimiter, so their in-pattern lookaheads stay linear on their own; theGFM backtick span bodies get deterministic alternatives with possessivequantifiers, removing their per-character backtracking.CONTENT_UNTIL_PARA is removed: any entry pattern built with it recreatesthe quadratic scan. ParallelRegex gains escapePattern() so embeddedcloser fragments follow the lexer's bare-parenthesis convention, reportsPREG_JIT_STACKLIMIT_ERROR so a future JIT exhaustion surfaces instead ofsilently truncating, and no longer rewrites its registered patterns inplace while compiling the compound regex.The adversarial 32KB page drops to 0.1s, the 37-second benign page to0.1s, and a 128KB variant stays under 0.6s. List of files: /dokuwiki/inc/Parsing/Lexer/ParallelRegex.php

864d6c6daac408fa4fc25801adcaf0f744288b30 - fix Lexer so exit-pattern lookbehinds see chars consumed by prior tokens

Tue, 21 Apr 2026 15:28:18 +0000

fix Lexer so exit-pattern lookbehinds see chars consumed by prior tokensLexer::reduce used to hand PCRE a shrinking tail of the subject — eachmatched token was chopped off the front of $raw and the next preg_matchran on what remained. Once a token was consumed, the bytes before thecursor were gone, and any lookbehind assertion in a subsequent patternsilently failed.The bug was latent for DokuWiki's entire history because literal exitpatterns like `\*\*`, ``, or `%%` don't care what's behind them.It surfaced with c3755410a ("require non-whitespace adjacency forinline formatting delimiters"), which added `(?<=[^\s])` to Strong,Emphasis, Underline, Monospace, Subscript, Superscript and Deleted atonce. After that commit, `**[[link]]**` stopped closing — the `]` thatwould satisfy the lookbehind had just been consumed by the link match,so Strong stayed open until end-of-section and swallowed everythingafter it (list items, headings, the lot).Fix:* Lexer::parse and Lexer::reduce track a byte offset into $raw instead of mutating $raw. $initialLength and the shrinking-length arithmetic for absolute match positions are replaced by straight offset increments; the no-progress guard and the trailing-unmatched dispatch both shift to the same cursor.* ParallelRegex::split takes an optional $offset and passes it to preg_match together with PREG_OFFSET_CAPTURE. PCRE scans from the offset forward but still sees the whole subject, so lookbehinds work across already-consumed tokens. The secondary preg_split call used to carve out pre/post is no longer needed — PREG_OFFSET_CAPTURE gives the match start for free, saving one regex operation per reduce() step.Regression tests at all three layers:* ParallelRegexTest — offset plumbing and pre/match accounting.* LexerTest::testIndexLookbehindAcrossConsumedToken — exit-pattern lookbehind targeting the `/>` of a self-closing `` that was consumed as a SPECIAL token on the previous step. Fails under the old Lexer.* FormattingTest — `**[[link]]**` and `**foo//bar//**` round-trip with correct open/close instructions through the full pipeline.Also updates ListsTest::testUnorderedListStrong, whose expectationsdocumented the pre-fix buggy behaviour ("formatting able to spreadacross list items"). With the fix, bold correctly stays within asingle list item; the expected call sequence and the comment areupdated to match. List of files: /dokuwiki/inc/Parsing/Lexer/ParallelRegex.php