[beagle3]

> Reference counting leaks cycles unless accompanied with a tracing GC (at which point reference counting makes little sense).

Python proves otherwise. Reference counting gives you deterministic memory use and finalization except when a cycle is involved. The tracing GC helps for those cases (and libraries) that do introduce cycles.

If each one of your objects is in a cycle, then -- yes, reference counting makes no sense. If only 1% of your objects are in a cycle, it makes 99% sense.

[tomp]

Python is a trivial special case as it has no concurrency. In a single-threaded language, that indeed makes sense. Multi-threaded RC is extremely tricky to implement efficiently, so unless you can prove there's no cycles (or you don't care it it leaks memory), it makes little sense.

[beagle3]

Oh, so now it's about "efficiently". Goal posts have moved.

No, it's not hard at all to implement efficiently as long as objects don't cross a thread boundary (and e.g. Nim's older GC used to enforce that condition, an old version of K tracked it and switched to "lock; xadd" to count references when something did cross a thread boundary IIRC, which made it inefficient only for those objects that crossed the boundary which usually weren't many.

It's way simpler than multithreaded mark&sweep, for example. Regardless - it makes a lot of sense. It might not make a lot of sense to you, but it does in general in most contexts.