This is a problem mentioned in the Haskell on SMP paper in section 3.5. If a thunk contains an unsafePerformIO or was created by unsafeInterleaveIO, evaluating it multiple times is harmful. The code below triggers this problem.

import Data.List
import Data.Char
import Control.Parallel
import System.IO

{-# NOINLINE fool #-}
-- 'fool' just makes sure CSE doesn't meddle with the code below
fool :: [a] -> [a]
fool (x:xs) = xs

main = do
    hSetBuffering stdin NoBuffering
    hs <- getContents
    let -- create copies of the input
        ts = map (:hs) ['a'..'z']
        -- sum the characters of each copy
        qs = map (foldl' (+) 0 . map ord . fool) ts
        -- in parallel
        rs = foldr (\x y -> x `par` y `par` (x:y)) [] qs
    -- compare the results and print 'True' if they are all equal.
    -- This should never print 'False'
    print $ all (uncurry (==)) $ zip rs (tail rs)

Don Steward kindly tried it on an SMP machine with the following results:

<dons> $ ghc -O A.hs -threaded
<dons> /tmp/ghc5400_0/ghc5400_0.hc:275:0:
<dons>      warning: implicit declaration of function 'newSpark'
<dons> -- good sign I got the right rts
<dons> $ yes abqszzzq | head -c 100000 | ./a.out +RTS -N1
<dons> True
<dons> $ yes abqszzzq | head -c 100000 | ./a.out +RTS -N2
<dons> False
<dons> $ uname -msr
<dons> Linux 2.6.15.3-general i686
<dons> with SMP kernel
<dons> CPU7: Intel(R) Xeon(TM) CPU 2.80GHz stepping 08
<dons> Total of 8 processors activated (44807.44 BogoMIPS).

To solve this problem, we need something like the justOnce suggested in the paper. It should be used in unsafePerformIO and unsafeInterleaveIO. Non-locking versions could be provided as unlockedUnsafePerformIO and unlockedInterleaveIO. This would allow existing libraries to work on SMP without or with just minor modifications, as far as I can see. The assumption that a thunk is evaluated at most once is fairly widespread.