On Sat, 2009-03-21 at 11:15 -0700, Conal Elliott wrote:
I'm suggesting that we have well-defined denotations for the pure types in Haskell, and that the various Haskell implementations be expected to implement those denotations.
I'm fine with IO continuing to be the (non-denotational) "sin bin" until we find more appealing denotationally-founded replacements.
I didn't answer your question as stated because I don't know what you include in "behaviour" for a functional program. I have operational associations with that word.
You're right of course, once we have one machine-dependent type then all of them are "infected" even simple things like Bool. The question is what should we do about it, if anything? A certain amount of machine dependent behaviour would seem to be useful. Even machine-dependent Int it's not in the H98 standard, implementations would want to add it as an efficiency extension and then we're back to the same place because one machine-dependent type infects all types. It doesn't even need machine dependent compilers. We've got plenty of libraries that have different "pure" semantics on different platforms. For example System.FilePath has a bunch of pure functions for manipulating file paths. We could construct a similar dodgy :: Bool example using functions from System.FilePath. As, Ganesh said, I'm not sure what we can actually do, unless you want to explain the denotation of everything with an extra (MachineInfo ->) context. Duncan