improve performance of tree sitter query captures (for text object motions in particular) (#4707)

* add tree sitter match limit to avoid slowdowns for larger files

Affects all tree sitter queries and should speedup both
syntax highlighting and text object queries.
This has been shown to fix significant slowdowns with textobjects
for rust files as small as 3k loc.

* Apply suggestions from code review

Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>

Co-authored-by: Blaž Hrastnik <blaz@mxxn.io>
This commit is contained in:
Pascal Kuthe 2022-11-11 02:47:12 +01:00 committed by GitHub
parent e17ad2722a
commit bb5a122cde
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23

View file

@ -354,6 +354,25 @@ impl<'a> CapturedNode<'a> {
}
}
/// The number of matches a TS cursor can at once to avoid performance problems for medium to large files.
/// Set with `set_match_limit`.
/// Using such a limit means that we lose valid captures in, so there is fundamentally a tradeoff here.
///
///
/// Old tree sitter versions used a limit of 32 by default until this limit was removed in version `0.19.5` (must now be set manually).
/// However, this causes performance issues for medium to large files.
/// In helix, this problem caused treesitter motions to take multiple seconds to complete in medium-sized rust files (3k loc).
/// Neovim also encountered this problem and reintroduced this limit after it was removed upstream
/// (see <https://github.com/neovim/neovim/issues/14897> and <https://github.com/neovim/neovim/pull/14915>).
/// The number used here is fundamentally a tradeoff between breaking some obscure edge cases and performance.
///
///
/// A value of 64 was chosen because neovim uses that value.
/// Neovim chose this value somewhat arbitrarily (<https://github.com/neovim/neovim/pull/18397>) adjusting it whenever issues occur in practice.
/// However this value has been in use for a long time and due to the large userbase of neovim it is probably a good choice.
/// If this limit causes problems for a grammar in the future, it could be increased.
const TREE_SITTER_MATCH_LIMIT: u32 = 64;
impl TextObjectQuery {
/// Run the query on the given node and return sub nodes which match given
/// capture ("function.inside", "class.around", etc).
@ -394,6 +413,8 @@ impl TextObjectQuery {
.iter()
.find_map(|cap| self.query.capture_index_for_name(cap))?;
cursor.set_match_limit(TREE_SITTER_MATCH_LIMIT);
let nodes = cursor
.captures(&self.query, node, RopeProvider(slice))
.filter_map(move |(mat, _)| {
@ -843,6 +864,7 @@ impl Syntax {
let mut cursor = ts_parser.cursors.pop().unwrap_or_else(QueryCursor::new);
// TODO: might need to set cursor range
cursor.set_byte_range(0..usize::MAX);
cursor.set_match_limit(TREE_SITTER_MATCH_LIMIT);
let source_slice = source.slice(..);
@ -1032,6 +1054,7 @@ impl Syntax {
// if reusing cursors & no range this resets to whole range
cursor_ref.set_byte_range(range.clone().unwrap_or(0..usize::MAX));
cursor_ref.set_match_limit(TREE_SITTER_MATCH_LIMIT);
let mut captures = cursor_ref
.captures(