Nimrod borrowed Modula-3's idea of having both traced and untraced pointers (using ref and ptr as keywords, respectively). As long as you only use ptr references, no overhead for reference counting (or other forms of garbage collection) is generated.

Of course, in order to use untraced references, you also have to use manual memory management (though you can use the macro/metaprogramming system to reduce the pain somewhat).

For traced references, Nimrod uses deferred reference counting; i.e. reference counts are only updated when they are stored on the heap or in a global variable (similar to write barriers for generational/incremental garbage collectors). If a reference count reaches zero, the reference is stored in a zero count table. At intervals, a separate pass checks if any references in the zero count table are still referenced from the stack (and does cycle detection where needed).

Deferred reference counting avoids the biggest problem with naive reference counting, which is that just assigning a pointer to a local variable (to inspect the contents of an object, say), can trigger up to two reference count updates. Conversely, with deferred reference counting, a read-only traversal of a data structure will not perform any reference count changes at all. Similarly, purely temporary allocations will not require any reference count changes, either.

> That's going to take a fairly significant performance toll due to cache effects alone.

It's not so bad because in Nimrod they are non-atomically reference counted and deferred reference counting is used to avoid touching the RC for stack references, at the cost of some minor latency for stack scanning. (Disclaimer: I work on Rust and my Nimrod info is possibly wrong and/or out of date.)