While GC requires tuning for the best performance, (pure) reference counting imposes a different cost: because it can't collect circular structures, the programmer has to break cycles herself. This clutters the application code. So you can't simply take a program written to run on a GC platform and move it to a reference counting platform unmodified.
Alternatively, you could use both: do most of your reclamation with reference counting, but throw in an occasional GC to collect circular structures. CPython, the most commonly used Python implementation, does this. I doubt it is the best thing for performance, but CPython isn't very fast anyway.
Comments
There's no free lunch.
While GC requires tuning for the best performance, (pure) reference counting imposes a different cost: because it can't collect circular structures, the programmer has to break cycles herself. This clutters the application code. So you can't simply take a program written to run on a GC platform and move it to a reference counting platform unmodified.
Alternatively, you could use both: do most of your reclamation with reference counting, but throw in an occasional GC to collect circular structures. CPython, the most commonly used Python implementation, does this. I doubt it is the best thing for performance, but CPython isn't very fast anyway.