For a compacting GC, you'd have to hunt down a complete set of references from running programs written in unknown languages, with strange errors if the set is

For a compacting GC, you'd have to hunt down a complete set of references from running programs written in unknown languages, with strange errors if the set is underapproximated even once. I'd rather have reference counting.