[flohofwoe]

But with ARC the compiler also does compile-time tracking to figure out when object references are shared and unshared, and based on that analysis inserts retain/release calls at the right points in the program (or more importantly: avoids the refcount overhead completely when it is not needed - e.g. when ownership is moved, not shared).

If mitten can do complete compile-time analysis also for all sorts of shared references and thus can avoid refcounting completely at all times then this would indeed be a nice improvement.

[_bxg1]

I didn't know that some reference-counting languages optimize away some cases of runtime counting by attempting to track ownership, though it makes sense. But I think when people say "reference counting" they mean the naiive approach. Even Rust has reference-counting structures, if you need them, and they actually expand what you're allowed to do with those values because you're partially stepping outside of the ownership system (at the cost of runtime performance).

[Ono-Sendai]

It's pretty easy to optimise away some reference counting operations. For example if you allocate something on the heap that is not returned from the function, nor passed to any other function, or captured in a closure, then you know it will be dead at the end of the function, so you don't need to emit reference counting operations for it.

[_bxg1]

Yeah. I guess Rust's secret sauce is just handling the long tail of much harder cases

[Ono-Sendai]

Probably yeah. One thing is that it helps to have a statically typed language to do these optimisations, since you can used type-based analysis for them. So dynamically typed languages like Lisp or Python are going to have a harder time doing them.