[Someone]

I am not sure I grok the code, but I think something along the following lines would work:

  - a 'recovery' list of (object pointer, old refcount) pairs
  - whenever you allocate something, add (object pointer, 0) to the list
  - when you update a reference count, check whether the object is in the list.
    If it is not, add it to the list, thus remembering the original refcount
  - whenever the code decreases reference count to zero, add the object to a
   'to be deleted' list (this could possibly be a linked list chained through
    the 'refcount' fields; the old refcount will be in the recovery list)
  - when the parser reaches toplevel without longjmp, clear the first list,
    and delete any objects in the 'to be deleted' list.
  - when a longjmp occurs, walk the list and reset reference counts. For
    zero 'old refcounts', delete the object

Elegant? Not really, but with proper macros/functions, it should not be much less elegant than reference counting on its own.

[silentbicycle]

That's pretty close to Deutsch-Bobrow deferred referenced counting. If you decide to implement it, you may want to read their paper about the technique.