mirror of
https://github.com/neovim/neovim.git
synced 2026-11-06 05:16:52 +00:00
jdrouhard
e8e914e41d
Problem: Semantic token request/response handling has several races: - active_request and active_range_request were shared, mutable tables reused for every request. Responses compared a mutable request_id that had already been cleared on cancellation or completion, so a stale, cancelled, or superseded response could still be applied to live state. - tokens_to_ranges mutated current_result.highlights in place and the range merge inserted incrementally while the decoder yielded. An overlapping range result could duplicate tokens or leave stale ones that the server no longer reported. - mark_dirty cleared the document version but kept result_id, so a following request could be a delta built on tokens that a range request had already mutated. Solution: Make request/response handling identity-based and non-mutating: - represent each in-flight request as an object captured by its response handler and compare identity instead of the mutable request_id. - drop responses when there is no matching active request (cancelled, replaced, or already completed). - make tokens_to_ranges a pure decoder that returns a new list, and merge range results separately with merge_highlight_ranges, which keeps tokens outside the requested range and substitutes the decoded tokens over the hole. - re-check that the request is still current after decoding, since decoding may yield, via is_current(). - clear result_id in mark_dirty so the next request is a full request rather than a delta over range-mutated tokens. - clear the active request field before cancelling it. Add regression tests covering overlapping range results, cancelled and superseded responses, and the mark_dirty full/delta selection. The tests defer the fake server's response callback to reproduce the races. AI-assisted
