bug#66732: tree-sitter fontification doesn't update multi-line syntax reliably

[Yuan Fu <casouri@gmail.com>]

On 12/12/23 4:43 AM, Dmitry Gutov wrote:
> On 12/12/2023 09:50, Yuan Fu wrote:
>
>> I feel that font-lock and syntax-ppss have different problems and 
>> requires different solutions. For font-lock, we need to mark affected 
>> range unfontified; we just need to make sure we do that after 
>> jit-lock-fontify-now. For syntax-ppss, we need to force a reparse 
>> before entering syntax-ppss so that syntax text properties are 
>> up-to-date when syntax-ppss do its work.
>
> font-lock *could* have a different solution, but I don't see why any 
> solution that works for syntax-ppss properly, shouldn't work just as 
> well for font-lock. And if it's possible - why not do that, saving on 
> variations in code. We also save on having to handle two differently 
> callbacks from the parser.
>
> The proposed alternative is not without problems. For example, that 
> unfontifying a region inside a timer, you trigger a second 
> fontification rather than having the job finished in one go. Possibly 
> introducing a extra "blink" as well. That is a minor inconvenience by 
> itself, but ideally we'd avoid it.

I've come around it, right, extending the to-be-fontified region before 
jit-lock runs is also a solution (and is probably better than fontifying 
twice).

>>> Here's a draft solution based on *-extend-region-functions, attached.
>>>
>>> Alas, while it works fine in python-ts-mode (for both syntax and 
>>> font-lock), making it behave better than python-mode, in c-ts-mode 
>>> it doesn't quite have the same effect: when you backspace over the 
>>> closing "/", the highlighting is properly updated only after you 
>>> make the next edit (any edit), or select another window. I'm not 
>>> sure, though, if it's due to my own problems with Emacs's failure to 
>>> redisplay (reported elsewhere), so more testing is welcome.
>>>
>>> But that might also be related to the use of 
>>> c-ts-mode--emacs-set-ranges: printing a backtrace calls inside 
>>> treesit--font-lock-notifier shows that the last notification comes 
>>> also during font-lock but after treesit--font-lock-extend-region, 
>>> inside c-ts-mode--emacs-set-ranges. I don't quite understand this 
>>> design where the ranges are applied inside the font-lock code.
>> c-ts-mode--emacs-set-ranges is registered as a range rule, so many 
>> tree-sitter function calls it before doing anything to make sure 
>> range is up-to-date. treesit-font-lock-fontify-region calls 
>> treesit-update-ranges at the beginning of its body, and 
>> treesit-update-ranges calls c-ts-mode--emacs-set-ranges.
>
> That seems to mean that any feature accessing the parse tree should 
> call treesit-update-ranges first. Including 
> syntax-propertize-functions and *-extend-region-functions, which we 
> currently don't do.
>
> But which boundaries is it supposed to use? Should 
> treesit--syntax-extend-region call treesit-update-ranges, when wait 
> for the parser updates, then possibly call treesit-update-ranges again 
> on the extended boundaries, and so on, until the parser stops sending 
> notifications? Will it stop?

It'll stop. If the ranges don't change, no reparse will happen. And only 
buffer content change can cause range change. Reparse itself can't. So 
at most you'll see reparse -> update ranges -> reparse.

> Anyway, could you try my patch? Like I said, I'm not sure if the 
> insufficient fontification I'm observing in c-ts-mode is due to the 
> problem with the solution, or due to the other redisplay-related 
> problems on my system.

Yeah, it doesn't solve the problem in c-ts-mode regarding block comments.

We might need to run (progn (force-parse) (update-ranges) (force-parse)) 
before jit-lock-fontify-now and sytax-ppss.

Yuan