[ezdiy]

> A fully concurrent GC is impossible

That's a common misconception, of course mutability gc barriers can be made atomic. But it comes at significant synchronization cost, plus using full shared world like in C does not seem like a good design decision anyway in high level language like crystal.

Which is why I'd be more in favour of refcounting, and let the user make the choice - a simple stop-world gc mark&sweep is alright for tasks which can afford the higher memory usage and pauses (one gains good throughput), or rc - good for low latency, low memory usage (and low throughput and high cache pollution).

Regarding multi core threading, crystal has next to none. All modern gc design decisions depend on how exactly multicore threading will be eventually implemented. hence why fixing gc is not a priority, but rc could be readily useful.