> Sure, but I don't think SML/NJ actually *detects* that it's still
> relevant.  (The GC could, I guess.)  This means that the retained C
> frame may still keep data alive which is really dead.  This might
> create a space leak.

Roland did this work for Olin Shivers' group at MIT; I don't know if
it has made it into the main distribution.  But that's correct --- you
need the GC to tell you when KS2 is garbage, and thus C2 can be
destroyed.