[Haskell] generic catch in a MonadIO

[oleg at pobox.com]

The ability to use functions 'catch', 'bracket', 'catchDyn', etc. in
MonadIO other than IO itself has been a fairly frequently requested
feature:

http://www.haskell.org/pipermail/glasgow-haskell-users/2003-September/005660.html
http://haskell.org/pipermail/libraries/2003-February/000774.html

The reason it is not implemented is because these functions cannot be
defined for a general MonadIO. However, these functions can be easily
defined for a large and interesting subset of MonadIO. The following
code demonstrates that. It uses no extensions (other than those needed
for the Monad Transformer Library itself), patches no compilers, and
proposes no extensions. The generic catch has been useful in a
database library (Takusen), where many operations work in a monad
(ReaderT Session IO): IO with the environment containing the database
session data. Many other foreign libraries have a pattern of passing
around various handles, which are better hidden in a monad. Still, we
should be able to handle IO errors and user exceptions that arise in
these computations.

> {-# OPTIONS -fglasgow-exts #-}
>
> module CaughtMonadIO where
>
> import Data.Typeable
> import Data.Dynamic
> import Control.Monad.Trans
> import Control.Exception hiding (catch, catchDyn)
> import qualified Control.Exception (catch)
> import Control.Monad.Reader
> import Control.Monad.Writer
> import Control.Monad.State
> import Control.Monad.RWS
> import Control.Monad.Error
>
> ---------------------  Tests
>
> data MyException = MyException String deriving (Show, Typeable)
>
> testfn True = throwDyn (MyException "thrown")
> testfn False = return True
>
> testc m = catchDyn (m >>= return . show) (\ (MyException s) -> return s)
>
> test1 = do tf True >>= print; tf False >>= print
>  where
>   tf x = runReaderT (runWriterT (testc (do
> 					tell "begin"
> 					r <- ask
> 					testfn r))) x
>
> test2 = do tf True >>= print; tf False >>= print;
>  where
>   tf x = runReaderT (runErrorT (do
> 				r <- ask
> 				testfn r `catchDyn` 
> 				  (\ (MyException s) -> throwError s))) x



The implementation is quite trivial.

> class MonadIO m => CaughtMonadIO m where
>     gcatch :: m a -> (Exception -> m a) -> m a
>
> instance CaughtMonadIO IO where
>     gcatch = Control.Exception.catch

> instance (CaughtMonadIO m, Error e) => CaughtMonadIO (ErrorT e m) where
>     gcatch m f = mapErrorT (\m -> gcatch m (\e -> runErrorT $ f e)) m

The following is almost verbatim from `Control.Monad.Error'
Section "MonadError instances for other monad transformers"

> instance CaughtMonadIO m => CaughtMonadIO (ReaderT r m) where
>     gcatch m f = ReaderT $ 
>                  \r -> gcatch (runReaderT m r) (\e -> runReaderT (f e) r)

The following instances presume that an exception that occurs in
'm' discard the state accumulated since the beginning of 'm's execution.
If that is not desired -- don't use StateT. Rather, allocate
IORef and carry that _immutable_ value in a ReaderT. The accumulated
state will thus persist. One can always use IORefs within
any MonadIO.

> instance (Monoid w, CaughtMonadIO m) => CaughtMonadIO (WriterT w m) where
>         m `gcatch` h = WriterT $ runWriterT m
>                 `gcatch` \e -> runWriterT (h e)

> instance CaughtMonadIO m => CaughtMonadIO (StateT s m) where
>         m `gcatch` h = StateT $ \s -> runStateT m s
>                 `gcatch` \e -> runStateT (h e) s

> instance (Monoid w, CaughtMonadIO m) => CaughtMonadIO (RWST r w s m) where
>         m `gcatch` h = RWST $ \r s -> runRWST m r s
>                 `gcatch` \e -> runRWST (h e) r s


> catchDyn :: (Typeable e, CaughtMonadIO m) => m a -> (e -> m a) -> m a
> catchDyn m f = gcatch m (\e -> maybe (throw e) f ((dynExceptions e) 
>                                                   >>= fromDynamic))