That's fair; Python, Swift, and most Lisps all use or have previously used reference-counting. But the quoted sentence isn't wrong, since it said no "garbage collection pauses" rather than "garbage collection."

Ok, I concede the point, “garbage collection” technically includes reference counting. However the practical point remains - reference counting doesn’t come with the same performance penalties as ‘normal’ garbage collection. It has essentially the same performance characteristics of manual memory management because that’s essentially what it’s doing.

It's not as simple as that.

Roc does static reference counting too, otherwise it wouldn't be able to do opportunistic in place mutation. It can do static reference counting up to a known compile time bound, whereas rust can only count to one. Both of them can do runtime reference counting, but it's implicit in roc and explicit with Rc and Arc in rust.

For example, consider the pseudocode
{
h = "Hello, "
hw = h + "world."
hm = h + "Mum!"
}

In real life, this could be something less swervable.

Roc counts, at compile time, 1,2,3,0, drop. No problem.

Depending on how you declare these variables (with what additional keywords, symbols, string types and concepts), rust counts, at compile time, 1,release,1,2! No no no stop broken! Bad programmer! This was in this case an unnecessary premature optimisation. That's what I mean by rust counts references, but only counts up to 1.

The borrow checker is a static reference counter with an arbitrary number of immutable references that you must declare explicitly and a maximum of one mutable reference that you declare explicitly with mut or let under different circumstances. Arc and Rc are runtime reference counters that you declare explicitly. This is essentially all tracked in the type system.

Roc does the static reference counting and if the total doesn't rise above rust's maximum of 1, uses in place mutation (as opposed to the default immutability). If it is bounded it can use static (compile time) reference counting so that when, for example, all four local references fall out of scope, the memory is dropped. If the number is unbounded (eg parameter passing recursion that can't be tail-cpseudocode ilarly removed), runtime reference counting is used. This is all essentially tracked in the runtime system, but calls to clone are automated in roc. A beginner absolutely can write a memory hog in roc, but the same beginner is likely to overuse clone in rust and write a similar memory hog.

runtime check. Which in turn results in a runtime performance.

If you're calling drop on a mutable string that's been extended repeatedly, you're recursively dropping all kinds of mess all over the heap. Checking for zero beforehand has an insignificant impact. Those cache misses you had because rust pays less attention to "where" than it does to "whether", they cost you a lot more than the reference count check. In the real world, in practice, under profiling of real code, the cache misses and the branch misses are more expensive than the reference counting.

You sound a little bit like a C programmer who claims his code is fast because his arrays don't do bounds checking. That's not why C is fast. Similarly rust isn't fast because it never does runtime reference counting. It does sometimes, but that code isn't pathologically slow.

Also, rust isn't just fast because of the borrow checker, primarily it's memory safe because of the borrow checker.

If it's any consolation, afaik, most of the roc platforms are written in rust. Also afaik only application specific code is written in roc. There are no memory management primitives in roc code unless a platform author exposes them in their api/interface, and I don't think anyone is working on implementing C on top of roc.

Roc has runtime overhead to do garbage collection, it says so right on their own page.

I was sceptical about your assertion because the language authors made a design decision not do do garbage collection. So I did a google search for garbage on roc-lang.org to try and find evidence of your claim. It doesn't say it does garbage collection. It does say overhead, but you're talking about it like it's a big slow thing that takes up time and makes thread pauses, but it's a small thing like array bounds checking. You do believe in array bounds checking, don't you?

So no, that's not what it says and you're using the phrase garbage collection to mean a much wider class of things than is merited. Garbage collection involves searching the heap for data which has fallen out of scope and freeing that memory up. It's slow and it necessitates pausing the main thread, causing unpredictably long delays. Roc does not do this.

Here's what the website actually says on the topic.

https://www.roc-lang.org/fast

Roc is a memory-safe language with automatic memory management. Automatic memory management has some unavoidable runtime overhead, and memory safety based on static analysis rules out certain performance optimizations—which is why unsafe Rust can outperform safe Rust. This gives Roc a lower performance ceiling than languages which support memory unsafety and manual memory management, such as C, C++, Zig, and Rust.

Just in case you missed it, that was unsafe rust that lacks the overheads. If you're advocating for using unsafe to gain a tiny performance benefit, you may as well be writing C, or zig, which at least has some tools to cope with all that stuff.

https://www.roc-lang.org/fast

When benchmarking compiled Roc programs, the goal is to have them normally outperform the fastest mainstream garbage-collected languages (for example, Go, C#, Java, and JavaScript)

Just in case you missed it, roc is not in the list of garbage collected languages.

https://www.roc-lang.org/platforms

The bigger benefit is tailoring memory management itself based on the domain. For example, nea is a work-in-progress Web server which performs arena allocation on each request handler. In Roc terms, this means the host's implementation of malloc can allocate into the current handler's arena, and free can be a no-op. Instead, the arena can be reset when the response has been sent.

In this design, heap allocations in a Web server running on nea are about as cheap as stack allocations, and deallocations are essentially free. This is much better for the server's throughput, latency, and predictability than (for example) having to pay for periodic garbage collection!

Summary: roc doesn't have the performance disadvantages of garbage collected languages because it's not a garbage collected language.