After an edit, LanguageTree:_edit() updates the current trees. When the
LanguageTree manages explicit regions, _edit() also refreshes _regions
from tree:included_ranges(true), so those regions have the edited byte
offsets.
A later injection pass may call set_included_regions() with a different
number of child regions. That path discards the old trees and emits
changedtree callbacks for them. invalidate(true) does the same when a
buffer is reloaded. Before this change, both discard paths called
tree:included_ranges(true) for every old tree, even if _edit() had just
collected those exact ranges.
That duplicate range extraction is expensive with many injection trees.
Realistic shapes include generated C files with many macro bodies parsed
by the C preproc_arg injection, Markdown documents with many fenced blocks
of the same language, and template files with many embedded-language
islands. The stock highlighter registers recursive changedtree callbacks,
so this is on the normal highlighting edit path.
Track whether _regions currently came from tree:included_ranges(true)
with _regions_from_tree_ranges. _do_changedtree_callbacks() reuses
_regions only in that state; otherwise it falls back to calling
tree:included_ranges(true). Clear the marker when regions are replaced by
injection ranges, when a tree is reparsed, or when trees are discarded.
This avoids keeping a second copy of the ranges while preserving callback
precision: changedtree still receives tree:included_ranges(true) for the
old tree, not the broader managed region.
Benchmark on 100k C macro injections, one-line edit, recursive
changedtree callback:
- HEAD median: edited parse 84.2 ms, child region replacement 58.7 ms
- This change: edited parse 34.6 ms, child region replacement 8.5 ms
That is about 2.4x faster for the edit parse and 6.9x faster for child
region replacement in this workload.
Add a regression test that replacing injection regions still fires
changedtree and still reports the old tree's exact included ranges.
AI-assisted: Codex
33ea63011c