[Haskell-cafe] Best way to implement "scoped exceptions"?

28 Dec 2022

      I'd like to implement "scoped exceptions"[1], that is, a combinator of
type

    withScopedException ::
        ((forall a. e -> IO a) -> IO r) ->
	IO (Either e r)

so I can use exceptions in a convenient way, without having to
set up fragile "action at a distance" based on the type of the thing
I'm throwing and catching, for example:

  scopedExceptionExample :: IO (Either String (Either Int Void))
  scopedExceptionExample = do
    withScopedException $ \throw1 ->
      withScopedException $ \throw2 ->
        if (1 :: Int) < 0
          then throw1 "Hello"
          else throw2 1234

  -- ghci> scopedExceptionExample
  -- Right (Left 1234)

Here's a hacky way of doing it, based on tagging each exception with a
unique value, and then filtering when handling:

  data MyException where
    MyException :: e -> Data.Unique.Unique -> MyException

  instance Show MyException where
    show _ = "<MyException>"

  instance Exception MyException

  withScopedException ::
    ((forall a. e -> IO a) -> IO r) -> IO (Either e r)
  wiathScopedException f = do
    fresh <- Data.Unique.newUnique
    flip tryJust (f (\e -> throwIO (MyException e fresh))) $ \case
      MyException e tag ->
        -- unsafeCoerce is very unpleasant
        if tag == fresh then Just (unsafeCoerce e) else Nothing

This is the approach taken by the effectful library[2].  But is there
a better way?  Can I persuade GHC's RTS to work like this directly?

Tom

[1] I don't know if these already have a name

[2] https://hackage.haskell.org/package/effectful-core-2.2.1.0/docs/src/Effectfu...

[Haskell-cafe] Best way to implement "scoped exceptions"?

Tom Ellis