I'm talking about languages that already use a GC as the primary heap management mechanism. Any kind of GC algorithm, whether tracing or refcounting, frees developers from thinking about lifetimes, but languages that primarily rely on GC and care about performance almost always pick tracing.
The only languages that choose a refcounting GC are either those that care little about performance (Python) or those that don't rely heavily on GC (Rust).
> Both largely for reasons of predictability, lower latency, lower memory use
Refcounting offers worse latency, worse predictability [1] (compared to modern low-latency tracing GCs, like ZGC), and worse throughput; it does offer a significantly lower memory footprint. The reason Rust chose refcounting is primarily because it doesn't rely on the GC heavily and so it might as well use a crude GC (and enjoy the lower effort of implementing it as well as the lower footprint), as a crude refcounting GC is usually better than a crude tracing GC.
Furthermore, an sophisticated tracing GC is not a great fit for languages like Rust or C++ because it requires a more elaborate code generation and some "magical" mutator thread behaviour, when low-level languages like the generated code to be as close to the source code as possible. This isn't so much predictability of performance, but predictability of the generated code and the behaviour of mutators (in terms of lowering the surprise of "why is my thread doing _that_ now?" when observed at a low level). For example, with ZGC, a mutator thread may find itself moving an object when accessing it; this is surprising at a low level -- which may not be a good fit for a low-level language -- even though it ends up offering better performance predictability overall. On the other hand, a mutator thread doing work freeing memory with a refcounting GC when it's dropping a reference to it may offer worse latency and worse predictability overall, but it's not as surprising at a lower level in the sense that it's easy to understand why it's doing that.
As for Swift, I think it picked a refcounting algorithm because it cares about memory footprint (most Swift applications run in memory-constrained environments) and because some unique language features allow to reduce the cost of refcounting sufficiently for a simple refcounting GC to be acceptable.
[1]: Again, not in principle (as both refcounting and tracing can converge to a similar behaviour) but as practiced by most refcounting GCs in industry use.
Not true. The creators of Swift chose RC exactly for performance reasons: low memory footprint AND low latency.
//edit: ah, you actually admitted it at the end. So you’re contradicting yourself.
If someday there's some advancement in refcounting that makes it better (or even as good but with significantly lower footprint), you'll see all languages that currenly use tracing switch quickly. For the time being, though, tracing is winning.
These days, tracing collectors offer better throughput, latency and predictability than refcounting, while refcounting offers better footprint. That's the main tradeoff, and I mentioned some secondary ones (re ease of implementation and low-level "surprises") before.
I disagree.
IMO the reason is simplicity, most languages don’t want to burden programmers with lifetimes and/or soft references.
Rust and Swift, both high-performance languages targeting specific niches, chose differently. Both largely for reasons of predictability, lower latency, lower memory use (which you admitted above) as well as lower memory trashing.