What's wrong with GHC Haskell's current constraint system?

Question

I've heard that there are some problems with Haskell's "broken" constraint system, as of GHC 7.6 and below. What's "wrong" with it? Is there a comparable existing system that overcomes those flaws?

For example, both edwardk and tekmo have run into trouble (e.g. this comment from tekmo).

Answer 1

Ok, I had several discussions with other people before posting here because I wanted to get this right. They all showed me that all the problems I described boil down to the lack of polymorphic constraints.

The simplest example of this problem is the MonadPlus class, defined as:

class MonadPlus m where     mzero :: m a     mplus :: m a -> m a -> m a 

... with the following laws:

mzero `mplus` m = m  m `mplus` mzero = m  (m1 `mplus` m2) `mplus` m3 = m1 `mplus` (m2 `mplus` m3) 

Notice that these are the Monoid laws, where the Monoid class is given by:

class Monoid a where     mempty :: a     mappend :: a -> a -> a  mempty `mplus` a = a  a `mplus` mempty = a  (a1 `mplus` a2) `mplus` a3 = a1 `mplus` (a2 `mplus` a3) 

So why do we even have the MonadPlus class? The reason is because Haskell forbids us from writing constraints of the form:

(forall a . Monoid (m a)) => ... 

So Haskell programmers must work around this flaw of the type system by defining a separate class to handle this particular polymorphic case.

However, this isn't always a viable solution. For example, in my own work on the pipes library, I frequently encountered the need to pose constraints of the form:

(forall a' a b' b . Monad (p a a' b' b m)) => ... 

Unlike the MonadPlus solution, I cannot afford to switch the Monad type class to a different type class to get around the polymorphic constraint problem because then users of my library would lose do notation, which is a high price to pay.

This also comes up when composing transformers, both monad transformers and the proxy transformers I include in my library. We'd like to write something like:

data Compose t1 t2 m r = C (t1 (t2 m) r)  instance (MonadTrans t1, MonadTrans t2) => MonadTrans (Compose t1 t2) where     lift = C . lift . lift 

This first attempt doesn't work because lift does not constrain its result to be a Monad. We'd actually need:

class (forall m . Monad m => Monad (t m)) => MonadTrans t where     lift :: (Monad m) => m r -> t m r 

... but Haskell's constraint system does not permit that.

This problem will grow more and more pronounced as Haskell users move on to type constructors of higher kinds. You will typically have a type class of the form:

class SomeClass someHigherKindedTypeConstructor where     ... 

... but you will want to constrain some lower-kinded derived type constructor:

class (SomeConstraint (someHigherKindedTypeConstructor a b c))     => SomeClass someHigherKindedTypeConstructor where     ... 

However, without polymorphic constraints, that constraint is not legal. I've been the one complaining about this problem the most recently because my pipes library uses types of very high kinds, so I run into this problem constantly.

There are workarounds using data types that several people have proposed to me, but I haven't (yet) had the time to evaluate them to understand which extensions they require or which one solves my problem correctly. Somebody more familiar with this issue could perhaps provide a separate answer detailing the solution to this and why it works.

Answer 2

[a follow-up to Gabriel Gonzalez answer]

The right notation for constraints and quantifications in Haskell is the following:

<functions-definition> ::= <functions> :: <quantified-type-expression>  <quantified-type-expression> ::= forall <type-variables-with-kinds> . (<constraints>) => <type-expression>  <type-expression> ::= <type-expression> -> <quantified-type-expression>                     | ...  ... 

Kinds can be omitted, as well as foralls for rank-1 types:

<simply-quantified-type-expression> ::= (<constraints-that-uses-rank-1-type-variables>) => <type-expression> 

For example:

{-# LANGUAGE Rank2Types #-}  msum :: forall m a. Monoid (m a) => [m a] -> m a msum = mconcat  mfilter :: forall m a. (Monad m, Monoid (m a)) => (a -> Bool) -> m a -> m a mfilter p ma = do { a <- ma; if p a then return a else mempty }  guard :: forall m. (Monad m, Monoid (m ())) => Bool -> m () guard True = return () guard False = mempty 

or without Rank2Types (since we only have rank-1 types here), and using CPP (j4f):

{-# LANGUAGE CPP #-}  #define MonadPlus(m, a) (Monad m, Monoid (m a))  msum :: MonadPlus(m, a) => [m a] -> m a msum = mconcat  mfilter :: MonadPlus(m, a) => (a -> Bool) -> m a -> m a mfilter p ma = do { a <- ma; if p a then return a else mempty }  guard :: MonadPlus(m, ()) => Bool -> m () guard True = return () guard False = mempty 

The "problem" is that we can't write

class (Monad m, Monoid (m a)) => MonadPlus m where   ... 

or

class forall m a. (Monad m, Monoid (m a)) => MonadPlus m where   ... 

That is, forall m a. (Monad m, Monoid (m a)) can be used as a standalone constraint, but can't be aliased with a new one-parametric typeclass for *->* types.

This is because the typeclass defintion mechanism works like this:

class (constraints[a, b, c, d, e, ...]) => ClassName (a b c) (d e) ... 

i.e. the rhs side introduce type variables, not the lhs or forall at the lhs.

Instead, we need to write 2-parametric typeclass:

{-# LANGUAGE MultiParamTypeClasses, FlexibleContexts, FlexibleInstances #-}  class (Monad m, Monoid (m a)) => MonadPlus m a where   mzero :: m a   mzero = mempty   mplus :: m a -> m a -> m a   mplus = mappend  instance MonadPlus [] a instance Monoid a => MonadPlus Maybe a  msum :: MonadPlus m a => [m a] -> m a msum = mconcat  mfilter :: MonadPlus m a => (a -> Bool) -> m a -> m a mfilter p ma = do { a <- ma; if p a then return a else mzero }  guard :: MonadPlus m () => Bool -> m () guard True = return () guard False = mzero 

Cons: we need to specify second parameter every time we use MonadPlus.

Question: how

instance Monoid a => MonadPlus Maybe a 

can be written if MonadPlus is one-parametric typeclass? MonadPlus Maybe from base:

instance MonadPlus Maybe where    mzero = Nothing    Nothing `mplus` ys  = ys    xs      `mplus` _ys = xs 

works not like Monoid Maybe:

instance Monoid a => Monoid (Maybe a) where   mempty = Nothing   Nothing `mappend` m = m   m `mappend` Nothing = m   Just m1 `mappend` Just m2 = Just (m1 `mappend` m2) -- < here 

:

(Just [1,2] `mplus` Just [3,4]) `mplus` Just [5,6] => Just [1,2] (Just [1,2] `mappend` Just [3,4]) `mappend` Just [5,6] => Just [1,2,3,4,5,6] 

---

Analogically, forall m a b n c d e. (Foo (m a b), Bar (n c d) e) gives rise for (7 - 2 * 2)-parametric typeclass if we want * types, (7 - 2 * 1)-parametric typeclass for * -> * types, and (7 - 2 * 0) for * -> * -> * types.