On 7/5/09, Paul L
Previously you had lastOrNil taking m [a] as input, presumably generated by mapM. So mapM is actually building an entire list before it returns the argument for you to call lastOrNil. This is where you had unexpected memory behavior.
Now you are "fusing" lastOrNil and mapM together, and instead of building a list, you traverse it and perform monadic action along the way. This can happen in a constant memory if the original pure list is generated lazily.
I think the real problem you had was a mis-understanding of mapM, and there was nothing wrong with your previous lastOrNil function. mapM will only return a list after all monadic actions are performed, and in doing so, it inevitably has to build the entire list along the way.
-- Regards, Paul Liu
Yale Haskell Group http://www.haskell.org/yale
Hi, Paul, thanks for the comments. You're quite right that I am fusing the two functions together, but I think I wasn't mis-understanding mapM... I knew I was generating the entire list, and aside from the slight inefficiency of generating it only to tear it down an instant later, that would have been no problem. But I was expecting all of the memory associated with the list to be reclaimed after I had processed it, and that was what was not happening as far as I could tell. (This isn't one monolithic list, by the way; it's the small bodies of a couple of small scheme functions that get evaluated over and over. So the setup and teardown happens a lot.) I don't have very good intuition yet about what should get garbage-collected and what should get kept in such situations, and in fact I'm kind of in the same boat again: the test case now runs much better, but it still leaks memory, and I am again stumped as to why. Could I see something useful by examining ghc core? I haven't looked at that yet, no idea what to look for... Uwe