bug#60953: The :match predicate with large regexp in tree-sitter font-lock seems inefficient

[Dmitry Gutov <dgutov@yandex.ru>]

On 26/01/2023 20:24, Eli Zaretskii wrote:
>> Date: Thu, 26 Jan 2023 19:15:51 +0200
>> Cc:60953@debbugs.gnu.org
>> From: Dmitry Gutov<dgutov@yandex.ru>
>>
>> On 26/01/2023 10:10, Eli Zaretskii wrote:
>>> Perhaps Dmitry could present comparison of profiles from perf which
>>> would allow us to understand the reason(s)?
>> I believe I did that in the second message in this thread:
>> https://debbugs.gnu.org/cgi/bugreport.cgi?bug=60953#8
>>
>> To quote the specific profiles, it's
>>
>>     15.30%  emacs         libtree-sitter.so.0.0       [.]
>> ts_tree_cursor_current_status
>>     14.92%  emacs         emacs                       [.] process_mark_stack
>>      9.75%  emacs         libtree-sitter.so.0.0       [.]
>> ts_tree_cursor_goto_next_sibling
>>      8.90%  emacs         libtree-sitter.so.0.0       [.]
>> ts_tree_cursor_goto_first_child
>>      3.87%  emacs         libtree-sitter.so.0.0       [.] ts_node_start_point
>>
>> for :pred vs.
>>
>>     23.72%  emacs         emacs                    [.] process_mark_stack
>>     12.33%  emacs         libtree-sitter.so.0.0    [.]
>> ts_tree_cursor_current_status
>>      7.96%  emacs         libtree-sitter.so.0.0    [.]
>> ts_tree_cursor_goto_next_sibling
>>      7.38%  emacs         libtree-sitter.so.0.0    [.]
>> ts_tree_cursor_goto_first_child
>>      3.37%  emacs         libtree-sitter.so.0.0    [.] ts_node_start_point
>>
>> for :match.
>>
>> And to continue the quote:
>>
>>     Here's a significant jump in GC time which is almost the same as the
>>     difference in runtime. And all of it is spent marking?
>>
>>     I suppose if the problem is allocation of a large string (many times
>>     over), the GC could be spending a lot of time scanning through the
>>     memory. Could this be avoided by passing some substitute handle to TS,
>>     instead of the full string? E.g. some kind of reference to it in the
>>     regexp cache.
> If you are saying that GC is responsible, then running the benchmark
> with gc-cons-threshold set to most-positive-fixnum should produce a
> more interesting profile and perhaps a more interesting comparison.

That really helps:

(benchmark-run 1000 (progn (font-lock-mode -1) (font-lock-mode 1) (let 
(treesit--font-lock-fast-mode) (font-lock-ensure))))

=> (16.078430587 251 5.784299419999996)

(let ((gc-cons-threshold most-positive-fixnum)) (benchmark-run 1000 
(progn (font-lock-mode -1) (font-lock-mode 1) (let 
(treesit--font-lock-fast-mode) (font-lock-ensure)))))

=> (10.369389725 0 0.0)

Do you want a perf profile for the latter? It might not be very useful.

> (But I thought you concluded that GC alone cannot explain the
> difference in performance?)

I'm inclined to think the difference is related to copying of the regexp 
string, but whether the time is spent in actually copying it, or 
scanning its copies for garbage later, it was harder to say. Seems like 
it's the latter, though.