bug#60691: 29.0.60; Slow tree-sitter font-lock in ruby-ts-mode

[Dmitry Gutov <dgutov@yandex.ru>]

On 10/01/2023 10:10, Juri Linkov wrote:
>>> After more rules were added recently to ruby-ts--font-lock-settings,
>>> font-lock became slow even on very small files.  Some measurements:
>>
>> If you saw a particular commit that made things slower, did you try
>> reverting it? What was the performance after?
> 
> No particular commit, just adding more rules degrades performance
> gradually.

But I don't think I added that many rules recently. No more than a 
quarter anyway.

>>> M-: (benchmark-run 1000 (progn (font-lock-mode -1) (font-lock-mode 1) (font-lock-ensure)))
>>> M-x ruby-mode
>>> (1.3564674989999999 0 0.0)
>>> M-x ruby-ts-mode
>>> (8.349582391999999 2 6.489918534000001)
>>
>> I have tried this scenario (which, to be frank, is pretty artificial, given
>> that fontification is usually performed in chunks, not over the whole
>> buffer).
>>
>> Perhaps the results depend on a particular file. The ones I have tried
>> (ruby.rb and ruby-after-operator-indent.rb) show only 2x difference (or
>> less). The difference was in favor of ruby-mode, but given the difference
>> in approaches I wouldn't be surprised if ruby-ts-mode incurs a fixed
>> overhead somewhere.
> 
> On test/lisp/progmodes/ruby-mode-resources/ruby.rb I see these numbers:
> 
> ruby-mode
> (8.701560543000001 95 1.045961102)
> 
> ruby-ts-mode
> (34.653148898000005 1464 16.904981779)

Interesting. It's 12s vs 36s for me, as I've retested now.

>>> This is not a problem when files are visited infrequently, but
>>> becomes a problem for diff-syntax fontification that wants to
>>> highlight simultaneously many files from git logs.
>>> So a temporary measure would be not to enable ruby-ts-mode
>>> in internal buffers:
>>
>> Is it common to try to highlight 1000 or even 100 files in one diff?
> 
> 100 is rare, but tens is pretty common, so this problem affects
> only this specific case.

So it's a 0,8-3s delay in those cases? That's not ideal.

>>> (add-hook 'find-file-hook
>>>             (lambda ()
>>>               (when (and (eq major-mode 'ruby-mode)
>>>                          ;; Only when not internal as from diff-syntax
>>>                          (not (string-prefix-p " " (buffer-name))))
>>>                 (ruby-ts-mode))))
>>
>> Have you tried similar tests with other -ts- modes? Ones with complex
>> font-lock rules in particular.
> 
> I tried with c-ts-mode, and it's very fast.

Just how fast is it? The number of font-lock features is has is 
comparable (though a little smaller).

I've tried the same benchmark for it in admin/alloc-colors.c, and it 
comes out to

   (3.2004193190000003 30 0.9609690980000067)

Which seems comparable.

Not sure how to directly test the modes against each other, but if I 
enable ruby-ts-mode in the same file, the benchmark comes to 1s.

Or if I enable c-ts-mode in ruby.rb -- 16s.

>> I've tried commenting out different rules in ruby-ts--font-lock-settings,
>> but none of them seem to have particularly outsides impact. Performance
>> seems, roughly, inversely proportional to the number of separate
>> "features".
> 
> Indeed, this is what I see - no particular rule, only their number
> affects performance.
> 
>> And if all ts modes turn out to have this problem, perhaps the place to
>> improve this is inside some common code.
> 
> I noticed that while most library files are small, e.g.
> libtree-sitter-c.so is 401,528 bytes,
> libtree-sitter-ruby.so is 2,130,616 bytes
> that means that it has more complex logic
> that might explain its performance.

ruby is indeed one of the larger ones. Among the ones I have here 
compiled, it's exceeded only by cpp. 2.29 MB vs 2.12 MB.

But testing admin/alloc-colors.c with c++-ts-mode vs c-ts-mode gives 
very similar performance, so it's unlikely that the complexity of the 
grammar is directly responsible.

> In this case, when nothing could be done to improve performance,
> please close this request.

Perhaps Yuan has some further ideas. There are some strong oddities here:

- Some time into debugging and repeating the benchmark again and again, 
I get the "Pure Lisp storage overflowed" message. Just once per Emacs 
session. It doesn't seem to change much, so it might be unimportant.

- The profiler output looks like this:

   18050  75%                    - font-lock-fontify-syntactically-region
   15686  65%                     - treesit-font-lock-fontify-region
    3738  15% 
treesit--children-covering-range-recurse
     188   0%                        treesit-fontify-with-override

- When running the benchmark for the first time in a buffer (such as 
ruby.rb), the variable treesit--font-lock-fast-mode is usually changed 
to t. In one Emacs session, after I changed it to nil and re-ran the 
benchmark, the variable stayed nil, and the benchmark ran much faster 
(like 10s vs 36s).

In the next session, after I restarted Emacs, that didn't happen: it 
always stayed at t, even if I reset it to nil between runs. But if I 
comment out the block in treesit-font-lock-fontify-region that uses it

     ;; (when treesit--font-lock-fast-mode
     ;;   (setq nodes (treesit--children-covering-range-recurse
     ;;                (car nodes) start end (* 4 jit-lock-chunk-size))))

and evaluate the defun, the benchmark runs much faster again: 11s.

(But then I brought it all back, and re-ran the tests, and the variable 
stayed nil that time around; to sum up: the way it's turned on is unstable.)

Should treesit--font-lock-fast-mode be locally bound inside that 
function, so that it's reset between chunks? Or maybe the condition for 
its enabling should be tweaked? E.g. I don't think there are any 
particularly large or deep nodes in ruby.rb's parse tree. It's a very 
shallow file.