New type of ($) operator in GHC 8.0 is problematic

[Yuras Shumovich]

On Mon, 2016-02-15 at 12:35 +0100, Herbert Valerio Riedel wrote:
> On 2016-02-15 at 12:00:23 +0100, Yuras Shumovich wrote:
> 
> [...]
> 
> > > - It is possible to have unlifted types about even without
> > > -XMagicHash. -XMagicHash is simply a lexer extension, nothing
> > > more.
> > > By convention, we use the # suffix with unlifted things, but
> > > there's
> > > no requirement here. Having -XMagicHash thus imply a flag about
> > > the
> > > type system is bizarre.
> > 
> > OK, I always forget about that. But is not it a bug already?
> > Usually we
> > don't allow code that uses GHC-specific extensions to compile
> > without a
> > language pragma. Why we don't have such pragma for levity
> > polymorphism?
> 
> There are extensions which are only needed at the definition
> site. Take {-# LANGUAGE PolyKinds #-} for instance; this is enabled
> inside the Data.Proxy module, which defines the following type
> 
>   {-# LANGUAGE PolyKinds #-}
>   module Data.Proxy where
> 
>   data Proxy t = Proxy
> 
> Now when you query via GHCi 7.10, you get the following output
> 
>   > import Data.Proxy
>   > :i Proxy 
>   type role Proxy phantom
>   data Proxy (t :: k) = Proxy
>     	-- Defined in ‘Data.Proxy’
>   instance forall (k :: BOX) (s :: k). Bounded (Proxy s) -- Defined
> in ‘Data.Proxy’
>   instance forall (k :: BOX) (s :: k). Enum (Proxy s) -- Defined in
> ‘Data.Proxy’
>   instance forall (k :: BOX) (s :: k). Eq (Proxy s) -- Defined in
> ‘Data.Proxy’
>   instance Monad Proxy -- Defined in ‘Data.Proxy’
>   instance Functor Proxy -- Defined in ‘Data.Proxy’
>   instance forall (k :: BOX) (s :: k). Ord (Proxy s) -- Defined in
> ‘Data.Proxy’
>   instance forall (k :: BOX) (s :: k). Read (Proxy s) -- Defined in
> ‘Data.Proxy’
>   instance forall (k :: BOX) (s :: k). Show (Proxy s) -- Defined in
> ‘Data.Proxy’
>   instance Applicative Proxy -- Defined in ‘Data.Proxy’
>   instance Foldable Proxy -- Defined in ‘Data.Foldable’
>   instance Traversable Proxy -- Defined in ‘Data.Traversable’
>   instance forall (k :: BOX) (s :: k). Monoid (Proxy s) -- Defined in
> ‘Data.Proxy’
> 
> even though you never enabled any extensions beyond what Haskell2010
> provides.
> 
> Do you consider this a bug as well?

Yes, IMO it is a bug. Though people didn't complain so far, so lets say
it is a minor design flow. Probably there are more important bugs to
fix.

Ideally language extensions should not leak to Haskell2010. E.g. making
lens using TemplateHaskell doens't leak to use side because I can
define lens by hands and preserve the API. But if something can't be
expressed in Haskell2010, then it should require extension to be
enabled both of definition and use sides.

In case of ($) people complain, and everybody seem to agree that levity
polymorphism leaking to Haskell2010 is bad. Fixing the leakage IMO is
the right way, while hiding the issue behind -fshow-rutime-rep is a
hack and a lie.

Probably the right way is harder in terms of development efforts (I
have no idea). In that case it probably makes sense to choose easier
way and introduce a hack. Life consists of compromises.