The following code:

module Main where

import Control.Monad.ST.Lazy
import Data.STRef.Lazy
import Data.Array.ST
import Int
import Debug.Trace

data Refs s = Refs
    { memory    :: STArray s Int8 Int8
    , pc        :: STRef s Int8
    }

main :: IO ()
main = do
    print $ runST m
  where
    m = do
      m <- newArray_ (0,30)
      p <- newSTRef 0
      let r = Refs m p
      writeArray m 0 0x4
      v <- readSTRef p
      modifySTRef p (+1)
      trace ("v: " ++ show v) $ return ()
      op <- readArray m v
      case trace ("v: " ++ show v) $ op of
          0x4 -> modifySTRef p (+100) -- should run this
          n   -> error ("should never match this: " ++ show n)

generates this output:

> ./st                                              
v: 0
v: 1
st: MArray: undefined array element

That is, the two instances of trace (show v) are printing different results, for the same v!

Looking at the Core, the problem seems to be that a call to readMutVar# has been inlined and duplicated. The readMutVar# primop doesn't have the has_side_effects flag set, and setting it (in primops.txt.pp) fixes it, but I'm not completely sure if that's the right fix. If it is, then there are lots of other similar primops that also need that flag set.

Bug report originally from a haskell-cafe post: http://www.haskell.org/pipermail/haskell-cafe/2009-May/061010.html