[Haskell-cafe] Proof that Haskell is RT

[David MacIver]

On Wed, Nov 12, 2008 at 8:35 PM, Lennart Augustsson
<lennart at augustsson.net> wrote:
> Actually, unsafeInterleaveIO is perfectly fine from a RT point of view.

Really? It seems easy to create things with it which when passed to
ostensibly pure functions yield different results depending on their
evaluation order:

module Main where

import System.IO.Unsafe
import Data.IORef

main = do w1 <- weirdTuple
          print w1
          w2 <- weirdTuple
          print $ swap w2

swap (x, y) = (y, x)

weirdTuple :: IO (Int, Int)
weirdTuple = do it <- newIORef 1
                x <- unsafeInterleaveIO $ readIORef it
                y <- unsafeInterleaveIO $ do writeIORef it 2 >> return 1
                return (x, y)

david at mel:~$ ./Unsafe
(1,1)
(1,2)

So show isn't acting in a referentially transparent way: If the second
part of the tuple were evaluated before the first part it would give a
different answer (as swapping demonstrates).