At Tue, 17 May 2011 02:18:55 +1000, Bernie Pope wrote:
http://augustss.blogspot.com/2011/04/ ugly-memoization-heres-problem-that-i.html
He says that "There's no guarantee about readIORef and writeIORef when doing multi-threading.". But I was wondering if that was true, and if it were, what the consequences would be. If you read his reply to my question on the blog, then I believe that he was saying that sequential consistency was not guaranteed.
While I don't know how IORefs work and I'd love to understand this better, I can't imagine any IORef implementation in which memoIO (in the blog post) would give the wrong answer on x86. It might, of course, cause f x to be evaluated multiple times on the same x. However, on other CPUs (e.g., the DEC alpha), there could maybe, maybe be issues. Though I'm not sure, since to avoid crashes, the alpha implementation of IORef would need to include a memory barrier. The question is whether there is an architecture in which IORef avoids crashes AND memoIO can give you the wrong answer. Also, if Simon's original post means that IORef operations all contain barrier instructions, it could be that memoIO is simply correct and the blog post is simply wrong about needing MVars. David