I don't know if my CPS conversion is really the result of an optimization or
not, but I agree one could make the argument that it does.
I would have thought that there would be some movement required when you
realize that a call is never going to return. Any way, it certainly has
never come up in anything I have done. It was just the result of some
thinking about if there are any cases where a stack-based implementation is
costing us. It was partly inspired by some comment some where about some
SML/NJ code which required the author to add these non-terminating loops just
so that SML/NJ would realize that the continuation was dead. I think that
putting it on the todo list at low priority is probably the right place for
it. Also I don't think I would like any overhead at non-GC time.
(text/plain)