GNU bug report logs - #78402
treesit after-change-functions

Previous Next

Full log

View this message in rfc822 format

From: Yuan Fu <casouri <at> gmail.com>
To: Juri Linkov <juri <at> linkov.net>
Cc: Stefan Monnier <monnier <at> iro.umontreal.ca>, 78402 <at> debbugs.gnu.org
Subject: bug#78402: treesit after-change-functions
Date: Tue, 8 Jul 2025 23:42:21 -0700

> On Jun 6, 2025, at 8:42 AM, Juri Linkov <juri <at> linkov.net> wrote:
> 
>>> 1. M-: (setq outline-minor-mode-cycle t)
>>> 2. M-: (setq outline-minor-mode-use-buttons t)
>>> 3. M-: (setq debug-on-error t)
>>> 4. C-x p f test/manual/indent/html-multi.html
>>> 5. M-x load-library RET treesit-x
>>> 6. M-x liquid-generic-ts-mode
>>> 7. M-x outline-minor-mode
>>> 8. move point to the first 's' in 'script' on line 19
>>> 9. type 'M-d' ('kill-word')
>> 
>> Aha, I reproduced it. It didn’t happen when I kill-word, but happened when
>> I undo.
> 
> This is what I observe too: when it doesn't fail on edit,
> then in any case it fails on undo.
> 
>> I’m still trying to understand why the node-outdated error can
>> happen. Each parser has a tick, which is incremented whenever it
>> re-parses. When we create a node from the parser, the node inherits the
>> current tick number of the parser. We only consider a node outdated when
>> its tick is less than its parser’s tick, meaning the parser re-parsed after
>> the node is created. For a parser to re-parse, it has to a) receive
>> a buffer edit or b) change it’s ranges; and then someone needs to request
>> a node from it.
>> 
>> That means in the following backtrace, the parser received a buffer
>> edit or changed its ranges after the `prev` node is created. And
>> that’s baffling to me.
> 
> Maybe calling this from after-change-functions affects the order of
> calling parser functions.

Ok, after much struggle, the bug should be fixed now. Quoting my commit message:

Right now in treesit-outline-search -> treesit-navigate-thing, a
freshly created tree-sitter node becomes outdated within the
function.  I'm not sure _exactly_ how it happend, but it might
look like this: we first get a node from, say, html parser, then
get another node from, say, liquid parser.  Creating the node
from liquid parser causes a reparse which updated the range of
the html parser, which rendered the html node outdated.

There're several problems with the current design, let's start
with the most obvious one: we add
treesit--font-lock-mark-ranges-to-fontify as a notifier of the
primar parser in treesit-major-mode-setup.  Now, if a ts major
mode inherits another major mode, treesit-major-mode-setup will
be called twice, once in the parent mode and once in the child
mode, and two parsers will have the notifier.  But
treesit--font-lock-mark-ranges-to-fontify is designed to run
only once.

I believe this bug, together with some mysterious async
execution order, led to the problems we saw in the bug report.
My solution is to just make everything synchronous.

So I added treesit-parser-changed-regions, and modified
treesit--font-lock-mark-ranges-to-fontify to use it.  Now we
don't need to add the notifier to the primary parser anymore.

I also applied the tree-sitter-outline change we discussed in
the bug report.  (Change to treesit-outline-search, and remove
treesit--after-change.)

Yuan

Previous Next

GNU bug tracking system
Copyright (C) 1999 Darren O. Benham, 1997,2003 nCipher Corporation Ltd, 1994-97 Ian Jackson.