[carry_bit]

The state-of-the-art in refcounting[1] greatly improves the barrier situation over a naïve implementation: no read barrier, and the write barrier only uses atomics when mutating an (apparently) unmodified field in an old generation object.

[1] https://dl.acm.org/doi/pdf/10.1145/3519939.3523440

[pron]

Certainly improvements in refcounting can bring its performance closer to that of tracing (and there are constant improvements in tracing, too), but one of the main reasons some languages choose refcounting is due to simplicitly of implementation, and these improvements bring the complexity of tracing and refcounting closer to each other. Currently, tracing leads in performance, but if that changes we'll see a shift in the algorithm for languages that depend heavily on GC performance.

BTW, our GC team investigated the implementation in that paper, and it is still significantly behind that of tracing in too many relevant workloads to be considered for adoption in production.

[pkolaczk]

And, additionally, reference counting on a single thread doesn’t need barriers at all.