"How does the approach in mitten compare to Automatic Reference Counting in Objective-C (and I think Swift too)? From my experience, ARC can still add a surprising amount of memory management overhead to a program and needs a lot of hand-holding to keep that overhead down to an acceptable level (e.g. low single-digit percentage of overall execution time in programs that talk to Obj-C APIs a lot). I would be surprised if a "traditional GC" can do any worse in that regard (maybe reference counting smears the overhead over a wider area, e.g. no obvious spikes, but instead "death by a thousand cuts")."
The reference counts have to be incremented when a new reference is made and decremented when one is deleted; freeing the memory when the count goes to zero. (This activity is cache- and atomicity-unfriendly (in the presence of threads).) A sufficiently smart compiler can omit many if not most of the count activity, but this kind of static analysis promises to remove all of it.
Further, reference counting has difficulty with circular references as the counts never go to zero. This should also be able to handle that.
Both this and reference counting are likely victims of the "death by a thousand cuts" you mention, as well as "drop the last pointer to a large structure and wait for a long time while the pieces are deleted"---the reference counting equivalent of a stop-the-world-and-trace garbage collection.
Luckily, they also present amore cache-friendly variation on reference counting that halves that overhead.
They modified the Swift compiler, so I think there’s a decent chance we’ll see this added to Swift.
Swift also, I think, is somewhat designed around the inefficiency of reference counting by promoting the use of structs for small objects (structs in Swift are value objects and not reference-counted in the language).
The reference counts have to be incremented when a new reference is made and decremented when one is deleted; freeing the memory when the count goes to zero. (This activity is cache- and atomicity-unfriendly (in the presence of threads).) A sufficiently smart compiler can omit many if not most of the count activity, but this kind of static analysis promises to remove all of it.
Further, reference counting has difficulty with circular references as the counts never go to zero. This should also be able to handle that.
Both this and reference counting are likely victims of the "death by a thousand cuts" you mention, as well as "drop the last pointer to a large structure and wait for a long time while the pieces are deleted"---the reference counting equivalent of a stop-the-world-and-trace garbage collection.