Fix Match Whole Word skipping overlapping matches

[varshitha (dobbydobap)]

Fixes #291591. Match Whole Word is applied as a post-filter on each regex match; when a candidate fails the boundary check, the search resumed at the end of that candidate, skipping any valid whole-word match that starts inside it (e.g. the trailing aa-aa in xaa-aa-aa, which grep -w finds). Now it advances one code point on a failed check so overlapping matches are considered. Added a regression test.

[Copilot]

Pull request overview

Fixes whole-word search skipping overlapping candidates after a boundary-check failure.

Changes:

• Advances rejected matches by one Unicode code point.
• Adds regression coverage for issue #291591.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/vs/editor/common/model/textModelSearch.ts	Adjusts search progression after rejected matches.
src/vs/editor/test/common/model/textModelSearch.test.ts	Adds overlapping-match regression test.

[Copilot]

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Balanced

[varshitha (dobbydobap)]

right on both counts. pieceTreeBase.ts had its own copy of that loop and it's the one model.findMatches ends up in. rather than patch the same loop twice I pulled it out into findSimpleMatchesInLine in textModelSearch.ts, and both TextModelSearch._findMatchesInLine and PieceTreeBase._findMatchesInLine call it now — that duplication is what let the first commit fix only half of it.

one thing worth flagging on the test: findMatchesLineByLine returns early through findMatchesInNode when the range starts and ends in the same node, and that path goes through Searcher, which was already fixed. so a plain createTextModel(text) passes with or without this change. the test builds the model from a chunked PieceTreeTextBufferBuilder so the search crosses a node boundary and actually reaches the line scanner — it fails without this commit and passes with it.

[Dmitriy Vasyura (dmitrivMS)]

AI Review: Retrying from every code point after a rejected whole-word candidate makes the new regex path quadratic. For example, whole-word /a+/ over x plus 40k as reruns the regex 40k times and took about 542 ms locally instead of one scan; the literal helper has the same repeated-suffix behavior. Because find runs synchronously, long/minified lines can freeze the renderer. Please advance to the next position that can satisfy the left-boundary rules (while honoring separators, CR/LF, and code points) and add long-line coverage for both paths.

[varshitha (dobbydobap)]

you're right, the one code point retry makes the regex path quadratic, and the literal helper has the same repeated-suffix problem. reworking it to advance to the next position that can actually satisfy the left boundary rule instead of stepping one code point at a time on your x + 40k as case no such position exists, so it scans once instead of 40k times, and the overlap fix still holds since a whole word match can only start right after a boundary. adding long line coverage for both paths.

[Dmitriy Vasyura (dmitrivMS)]

AI Review: Searcher accepts any RegExp, and SearchData does not require g/y. For a non-global /abc/ on xabc, exec() ignores the assigned lastIndex and repeats index 1; clearing the previous-match state here defeats the duplicate-range guard, so next() loops forever. Current factories use global regexes, but an existing piece-tree test already constructs non-global SearchData. Please preserve duplicate detection for regexes that cannot advance via lastIndex, or explicitly normalize/reject them, with regression coverage.

[varshitha (dobbydobap)]

that's a real hang. clearing the prev-match state was only there to get past the end-of-line guard at the top of the loop, but it takes the duplicate-range check out with it, so a regex that can't advance via lastIndex never terminates. switching to a one-shot flag that skips just the end-of-line check and leaves _prevMatchStartIndex/_prevMatchLength intact. adding a non-global regression test pieceTreeTextBuffer.test.ts already constructs a non-global SearchData with separators, it just happens to run on an empty buffer today so it never spins.