Turns out the previous commit message and comment is not entirely
correct: the old behavior is in fact wrong, not just "correct but has
problems".
Here is why the old code is wrong:
|visible range| -> markup for visible range
updated range -> markup for updated range
-------------
First we have some text
|aaaaaa|
Now we insert something at the beginning, because we clip
new_end_offset to visible_end, out of eight b's inserted, only the
first six are known to tree-sitter.
|
bbbbbbbbaaaa|aa start: 0, old_end: 0, new_end: 6
------
In treesit_sync_visible_region, we sync up visible region, but the two
missing b's are not in the updated range.
|
bbbbbbbbaaaaaa| start: 12, old_end: 12, new_end: 14
--
The old behavior not only is wrong, but also doesn't make much sense.
* src/treesit.c (treesit_record_change): Update comment.
max (visible_beg, old_end_byte))
- visible_beg);
/* We don't clip new_end_offset under visible_end, because
- inserting in narrowed region always extends the visible
- region. If we clip new_end_offset here, and re-add the
- clipped "tail" in treesit_sync_visible_region later,
- while it is technically equivalent, tree-sitter's
- incremental parsing algorithm doesn't seem to like it
- (bug#61369). */
+ otherwise we would miss updating the clipped part. Plus,
+ when inserting in narrowed region, the narrowed region
+ will grow to accommodate the new text, so this is the
+ correct behavior. (Bug#61369). */
ptrdiff_t new_end_offset = (max (visible_beg, new_end_byte)
- visible_beg);
eassert (start_offset <= old_end_offset);