For context Sublime Text 3 takes 2 seconds with the same grammar and same file on the same computer due to using a better custom regex engine written specifically for highlighting.
Given what alexdima mentioned in a different comment about spending most of the time in the regex engine, I'm not sure that my engine would be substantially faster under exactly identical conditions since I'm also bottlenecked by Oniguruma.
However, maybe after I port my engine to https://github.com/google/fancy-regex I'll be substantially faster. And if I do it is likely they could also benefit from a fancy-regex port.
Just eyeballing the cited numbers, they take 3939ms to handle a 1.18MB input on "a somewhat powerful desktop machine". Assuming that that means a chip running at 2GHz, we're talking about over 6300 cycles per byte!
That's quite frankly ridiculous. An improvement by at least one order of magnitude should be possible. Where's the ambition?
(Yes, there's always a trade-off with these things. But I feel someone has to point this out when the OP is explicitly about getting kudos for performance work.)
I've looked in the past for optimization opportunities in the C land (mostly through better caching), which yielded quite nice results [1][2]. I would love if you'd want to take a look too.
At this point, in tokenization, 90% of the time is spent in C, matching regular expressions in oniguruma. More precisely, regular expressions are executed 3,933,859 times to tokenize checker.ts -- the 1.18MB file. That is with some very good caching in node-oniguruma and it just speaks to the inefficiency of the TM grammars regex based design, more than anything else.
It is definitely possible to write faster tokenizers, especially when writing them by hand (even in JS), see for example the Monaco Editor[3] where we use the TypeScript compiler's lexer as a tokenizer.
At least in this case, inefficiencies are not caused by our runtime.
[1] https://github.com/atom/node-oniguruma/pull/40
It's non-trivial because TextMate grammar seem like they're just a little bit too general to be convenient. So there's definitely a trade-off. But if I wanted to really get as fast as possible, I would try to see if I can get there.
I could get way better performance by rewriting all those grammars using compiled parsers in Rust (like Xi has as an option https://github.com/google/xi-editor/blob/master/rust/lang/Ca...) but it would take an absurd amount of effort.
The speed of highlighting tmLanguage files is limited mostly by the regex library. My code spends 50% of its time in Oniguruma. You can improve that a bit by using a fancy regex engine, which is how Sublime is faster than my engine, but this evening I'm going to try to port my library to a faster engine based on Rust's regex crate. But that library (https://github.com/google/fancy-regex) is brand new and almost untested. It's the only open source library that is faster than Oniguruma while supporting all the right features. In the future VSCode may be able to gain some speed by switching to it, but they have much more overhead over the regex library to eat away first than I do.
We had to implement getuid [0] as an executable in C, which just looks up /etc/passwd by name and returns a number.
By using specialised trie-like data structures and, in the end, also custom memory alignment instead of malloc, we squeezed the whole algorithm runtime down to an average of 45 cycles or so.
And, do you have an issue for the regex port? Would love to see the benchmark.
I have yet to do the regex engine port, I'm doing that later this evening. I'm porting it from Oniguruma to https://github.com/google/fancy-regex which accelerates common types of regexes using the awesome Rust regex crate which is super fast.
This evening I'm even trying to port it to a pure Rust regex engine that should eliminate non-Rust code and make it substantially faster.
It also implements the sublime-syntax format which is a superset of tmlanguage that allows even nicer highlighting.
https://github.com/trishume/syntect