I'm curious, how does GHC determine that the CAF is no longer required if it is referenced by code (somehow)? If code was also some kind of heap allocated data structure I guess this would be possible, but I thought this wasn't so with GHC.

GHC actually tracks references to top-level entities from code, so that the GC can trace the transitive closure of live code and hence find all the reachable CAFs. It's a real pain, but worth it.

But in any case, I'm not sure this really helps me. I don't really mind if the unreduced form of the CAF is garbage collected or not (it's only going to be a few words of memory). The effect I'm trying to get is to ensure that the (partially) reduced form of the CAF only lives as long as any (non code) heap object which references it. Does that make sense? (probably not:-)

If I understand correctly, that's the behaviour you'll get with GHC, and (I think) nhc98. Cheers, Simon