state hack causes unneeded value to be evaluated

[int-e]

The following code prints "Main: Why?" when compiled with ghc -O1. With -fno-state-hack it works fine.

data X = X ()

{-# NOINLINE newX #-}
newX :: () -> IO X
newX n = do
    let {-# NOINLINE value #-}
        value = n
    return (X value)

main = do
    x <- newX (error "Why?")
    case x of
        X _ -> return ()

Both pragmas are needed to exhibit the bug.

(In my actual code, value is an unsafePerformIO -- this was the motivation for adding the pragmas.)

[simonmar]

Good bug. I don't think this is caused by -fno-state-hack. The code coming out of the simplifier is fine; this is newX:

a_rus =
  \ (n_afy :: ()) (eta_ssU :: GHC.Prim.State# GHC.Prim.RealWorld) ->
    (# eta_ssU,
       Main.X
         (let {
            value_sub [NEVER Just S] :: ()
            [Str: DmdType]
            value_sub = n_afy } in
          value_sub) #)

The prep phase lifts out that let, but doesn't reset its strictness annotation, which changes the meaning of the program:

a_rus =
  \ (n_suA :: ()) (eta_suy :: GHC.Prim.State# GHC.Prim.RealWorld) ->
    let {
      sat_suE :: Main.X
      []
      sat_suE =
        case n_suA of value_suC [NEVER Just S] { __DEFAULT ->
        Main.X value_suC
        } } in
    (# eta_suy, sat_suE #)

[simonpj]

Fixed by

Wed Nov 26 14:34:48 GMT 2008  simonpj@microsoft.com
  * Fix Trac #2756: CorePrep strictness bug

Pls merge.

Simon

[igloo]

Merged