Lua (and to a somewhat lesser extent Luau) are small in terms of the learned surface of the value language, not necessarily in terms of lines of code. That being said, any runtime use of the language needn't depend on Analysis, which is the biggest compilation unit by far.

Probably also worth mentioning that Analysis currently contains two full type system implementations because we've spent the better part of the past three years building a new type system to address a lot of the fundamental limitations and architectural issues that we'd run into after years of working on the original one. The new one is not without issues still, but is definitely very usable, and gets better and better every week. At some point in the future, we will clip the old one altogether, and that'll shave off some 30,000 lines of code or something.

▲

pizlonator 3 days ago | parent [-]

Is the primary goal of the type system performance, or dev productivity, or something else?

If performance, are you comparing against a baseline that does the classic Self-style optimizations (like what JS engines do)?

FWIW I don't consider LuaJIT to be an interesting baseline because of what I've heard about it being overtuned to specific benchmark that the author(s) care about. (Too much of an attitude where if your code doesn't run well then it's your fault.)

▲

parenwielder 3 days ago | parent [-]

I would say that the primary interest in building the type system is in supporting a better developer experience, both in terms of productivity (better autocomplete, go-to definition, etc.) and in terms of correctness (identifying bugs earlier, which is why we're actively concerned with having a _sound_ type system). There's a pretty limited surface of areas where we accept unsoundness today (outside of casting) and they're all connected to limitations of the type system that we've been working to resolve.

Longer-term, there's definitely some interest in how we could leverage types to support more optimized code, but it's been a notoriously difficult problem for gradually-typed languages in general, see [Is sound gradual typing dead?](https://dl.acm.org/doi/pdf/10.1145/2837614.2837630)

	▲	pizlonator 3 days ago \| parent [-]
		Gotcha. For some reason I was under the impression that getting perf from gradual typing had been a goal of y'all's. > primary interest in building the type system is in supporting a better developer experience Yeah this is the right reason to do it :-) > there's definitely some interest in how we could leverage types to support more optimized code, but it's been a notoriously difficult problem for gradually-typed languages in general I know. I still get folks asking why TypeScript types can't be used to make JS fast and it's quite exhausting to explain this. Hard to imagine gradual types beating what you get from the combo of PICs, speculative JITing, and static analysis. (I worked on JavaScriptCore's optimizer, hence why I'm curious about where y'all are coming from on this)