Non-deterministic results from `noThunksInValues`

[sjakobi]

I'm getting non-deterministic results for a regression test in unordered-containers.

You can find the code at https://github.com/sjakobi/unordered-containers/tree/nothunks-issue. Run the test with

$ cabal run tests -- -p issue383

With GHC 9.2.2 I'm getting failures and successes in roughly equal proportion.

If I remove the source of randomness in the test with

--- a/tests/Regressions.hs
+++ b/tests/Regressions.hs
@@ -198,7 +198,7 @@ instance Functor MyIdentity where
 
 issue383 :: Assertion
 issue383 = do
-  i :: Int <- randomIO
+  let i :: Int = 42
   let f Nothing = MyIdentity (Just (show i))
       f Just{}  = MyIdentity (error "Nah")
   let (MyIdentity m) = HMS.alterF f () mempty

… I get consistent successes with GHC 9.2.2 and consistent failures with GHC 9.0.2.

The expected test result is a success, independent of the value of i.

[sjakobi]

I have much reduced the test on the mentioned branch, until only this is left:

issue383 :: Assertion
issue383 = do
  i :: Int <- randomIO
  let !v = show i
  mThunkInfo <- noThunks mempty v
  assert $ isNothing mThunkInfo

What I had initially overlooked is that a String forced to WHNF may still contain thunks. However I'm still surprised that the result is non-deterministic. Any ideas?

[lehins]

Isn't order of evaluation of pure values in Haskell non-deterministic by design?

[sjakobi]

Isn't order of evaluation of pure values in Haskell non-deterministic by design?

Surely GHC can be surprisingly eager to evaluate values in some cases, when demand analysis figures out that the value will be forced anyway.

But in this case I don't see much reason for it. The strict let-binding should force the string to the outer (:) constructor and no further before noThunks is applied.

[TeofilC]

I think this is explained by the show instance for Int.
It looks like for a negative number forcing the result to whnf gives you '-': _ whereas for a non-negative number you get the whole number.

Eg, the following ghci session:

> showWhnf :: Int -> IO String; showWhnf = evaluate . show
> x1 <- showWhnf 1
> :sprint x1
x1 = "1"
> x2 <- showWhnf $ -1
> :sprint x2
x2 = '-' : _
>

Cf. https://hackage.haskell.org/package/base-4.17.0.0/docs/src/GHC.Show.html#itos
You can clearly see there that the tail of the result is not forced in the negative case

[sjakobi]

Ha, that explains it! :) Thank you, @TeofilC!