With NOINLINE you should be safe, but you never mentioned that originally. :)
On Thu, Apr 16, 2009 at 10:29 AM, Peter Verswyvelen <[email protected]> wrote: > Well, the documentation says: > Use {-# NOINLINE foo #-} as a pragma on any function foo that > calls unsafePerformIO. If the call is inlined, the I/O may be performed more > than once. > So you claim this does not prevent GHC to inline it anyway? That feels like > a bug then, both in the documentation and NOINLINE > On Thu, Apr 16, 2009 at 10:18 AM, Lennart Augustsson > <[email protected]> wrote: >> >> There's no guarantee about unsafePerformIO not being inlined, that's >> just how ghc treats it. >> >> 2009/4/16 Patai Gergely <[email protected]>: >> >> On the other hand, breaking referential transparency in the >> >> external interface is a very bad idea, in my opinion. Actually, >> >> this means that the library user would have to turn certain >> >> compiler optimizations off to get the intended behavior. >> > However, in practice you can compile Elerea with -O2 without ill >> > effects. In fact, that's what happens if you install it with cabal. >> > >> >> Just have a look at the Haddock docs of unsafePerformIO. >> > Yes, I did that too, and came up with the following checklist: >> > >> > - the order of side effects doesn't matter much, since the resulting >> > networks are equivalent if we don't rely on the automatic delay feature >> > (applicative optimisations can be different, but still with the same net >> > effect) >> > - unsafePerformIO is apparently never inlined, i.e. each instance is >> > executed once, so sharing works as desired >> > - let-floating is no problem, because all instances of unsafePerformIO >> > rely on surrounding function arguments >> > - CSE is no problem either, it even helps if it's performed (and it is >> > with optimisations turned on), since it results in smaller equivalent >> > networks >> > >> > I think we can expect it to be fairly well-behaving, because the 'side >> > effect' of Elerea primitives is basically the same as that of pure >> > values in general: upon evaluation a value is created in the memory and >> > we get a reference to it. We only have an extra constraint for the >> > compiler: never duplicate these values. Merging identical ones is okay, >> > and in fact desirable. The following code demonstrates this if you >> > compile it with and without optimisations: >> > >> > import Control.Applicative >> > import Control.Monad >> > import FRP.Elerea >> > import System.IO.Unsafe >> > >> > cint a b = unsafePerformIO (putStrLn "!") `seq` >> > transfer 0 (\dt x x0 -> x0+x*dt) b >> > >> > mysig = (latcher 0 (b >@ 0.3) (const (cint a b) <$> cint a b)) + >> > (cint a b) + (cint a b) + a >> > where a = pure 4 >> > b = stateful 0 (+) >> > >> > main = replicateM 10 (superstep mysig 0.1) >>= print >> > >> > I'd like to see an example where optimisation does make a difference, >> > because I'm still unsure about the consequences of 'unsafeness'. >> > >> > Gergely >> > >> > -- >> > http://www.fastmail.fm - Or how I learned to stop worrying and >> > love email again >> > >> > _______________________________________________ >> > Haskell-Cafe mailing list >> > [email protected] >> > http://www.haskell.org/mailman/listinfo/haskell-cafe >> > >> _______________________________________________ >> Haskell-Cafe mailing list >> [email protected] >> http://www.haskell.org/mailman/listinfo/haskell-cafe > > _______________________________________________ Haskell-Cafe mailing list [email protected] http://www.haskell.org/mailman/listinfo/haskell-cafe
