Lifetime inference compiles to destination passing so for most cases a single stack allocation in a prior function can be used. For your example of mutually rec

Lifetime inference compiles to destination passing so for most cases a single stack allocation in a prior function can be used. For your example of mutually recursive functions allocating in a loop the destination would be a region that will grow dynamically like an arena allocator. Since refs are typed, ref elements of the same type will be allocated next to each other in memory. You don't touch on it, but there are some more difficult cases with lifetime inference as well. Namely branching the compiler must decide whether or not to extend a refs lifetime. This and a lack of granularity in container types are known problems with region inference (they lead to more memory than necessary being used by assuming the longest lifetimes), and are things I hope to tackle.