I believe there is a third way (And no, I'm not talking about Tony:).

This consists of using reference counting and mark'n'sweep. The reference counting deals with the majority of GC, ensuring timely destruction and avoiding the "pregnent pauses" that are another, more insidious signature of M&S. The mark&sweep is only used to break the bonds of circular references. Either invoked automatically at "resources low" time, or by the programmer when he knows he has just finished with data that may include circular refs.

That leaves the problem of reference counting code being distributed throughout the codebase. That can be addressed by using real references (Handles).

If references aren't "addresses with a flag and some magic", but are instances of a Reference class that overloads the "take a ref" operator and assignment, such that whenever a reference to an object is taken, or a reference is assigned, a Reference object method is invoked to do it, then all the reference counting code can live in one place.

Further, by storing the actual addresses behind the Reference handles in a single place (say an array that is class data in the Reference class, and using the index into that array as the Handle), when the time comes that a Mark'n'Sweep is required, the amount of memory that must be scanned to locate live/dead objects is confined to things pointed to by that array, and is therefore very fast.

The penalty is that every data access through a reference has an extra level of indirection. That may sound like a high overhead, but in the scheme of things where most references are to (fat) objects rather than individual intrisic types, the overhead can become lost in the greater levels of indirection they already entail. Moore's law means that the costs going forward are ever reducing and only become significant if compared to a non-indirected implementation which will always be faster--until the GC runs.

There are other benefits of using a "references are handles not addresses" implementation to do with reentrancy which can really come into their own once threading, continuations and co-routines come into play.

Overall, I think that the benefits of "one more level of indirection" could more than outweigh costs, but so far there is no existing implementation that I can find upon which to base substance to this speculation.