Proposal: merge either into transformers

[Ross Paterson]

On Wed, Apr 30, 2014 at 06:00:50PM +0900, Kei Hibino wrote:
> I discovered mplus of ExceptT doesn't call mappend to accumulate
> error states which is different from origitnal EitherT like below.
> I suppose this EitherT semantics is more useful than fixed adoption
> of last error state.
> (For example, Last Monoid is pre-defined in base)
> 
> ExceptT
> >instance (Monad m, Monoid e) => MonadPlus (ExceptT e m) where
> >    mzero = ExceptT $ return (Left mempty)
> >    ExceptT m `mplus` ExceptT n = ExceptT $ do
> >        a <- m
> >        case a of
> >            Left _ -> n                  -- throw left error away
> >            Right x -> return (Right x)
> 
> EitherT
> >instance (Monad m, Monoid e) => Alternative (EitherT e m) where
> >  EitherT m <|> EitherT n = EitherT $ m >>= \a -> case a of
> >    Left l -> liftM (\b -> case b of
> >      Left l' -> Left (mappend l l')     -- mappend error states
> >      Right r -> Right r) n
> >    Right r -> return (Right r)
> >
> >  empty = EitherT $ return (Left mempty)

Yes, the ExceptT instance is similar to the ErrorT one.  The accumulating
one in EitherT is more flexible (and consistent with using mempty in
the mzero case), but it's also more expensive than the tail call in the
ErrorT version.

What do others think?