Suppose a Haskell library designer decides to use UndecidableInstances for some reason. The library compiles fine. Now suppose some program uses the library (like defines some instances of its type classes), but doesn't use the extension. Can it happen that the compilation fails (doesn't terminate)?
If such a scenario can happen, I'd be happy to see an example. For example, as mtl uses UndecidableInstances a lot, is it possible to write a program that depends on mtl (or any other standard library that uses the extension), doesn't use UndecidableInstances itself, but fails to compile because of undecidability?
Great question!
In general this is certainly possible. Consider this module:
{-# LANGUAGE MultiParamTypeClasses, FunctionalDependencies, UndecidableInstances #-}
module M where
class C a b | a -> b where
f :: a -> b
instance C a b => C [a] [b]
where f = map f
It compiles by itself just fine. However, if you import this module and define
g x = x + f [x]
you'll get
Context reduction stack overflow; size = 201
Use -fcontext-stack=N to increase stack size to N
C [b] b
In the second argument of `(+)', namely `f [x]'
In the expression: x + f [x]
In an equation for `g': g x = x + f [x]
Regarding the mtl instances, I don't see how something like this is possible, but I also don't have a proof that it's not.
Related
I am using data-reify and graphviz to transform an eDSL into a nice graphical representation, for introspection purposes.
As simple, contrived example, consider:
{-# LANGUAGE GADTs #-}
data Expr a where
Constant :: a -> Expr a
Map :: (other -> a) -> Expr a -> Expr a
Apply :: Expr (other -> a) -> Expr a -> Expr a
instance Functor Expr where
fmap fun val = Map fun val
instance Applicative Expr where
fun_expr <*> data_expr = Apply fun_expr data_expr
pure val = Constant val
-- And then some functions to optimize an Expr AST, evaluate Exprs, etc.
To make introspection nicer, I would like to print the values which are stored inside certain AST nodes of the DSL datatype.
However, in general any a might be stored in Constant, even those that do not implement Show. This is not necessarily a problem since we can constrain the instance of Expr like so:
instance Show a => Show (Expr a) where
...
This is not what I want however: I would still like to be able to print Expr even if a is not Show-able, by printing some placeholder value (such as just its type and a message that it is unprintable) instead.
So we want to do one thing if we have an a implementing Show, and another if a particular a does not.
Furthermore, the DSL also has the constructors Map and Apply which are even more problematic. The constructor is existential in other, and thus we cannot assume anything about other, a or (other -> a). Adding constraints to the type of other to the Map resp. Apply constructors would break the implementation of Functor resp. Applicative which forwards to them.
But here also I'd like to print for the functions:
a unique reference. This is always possible (even though it is not pretty as it requires unsafePerformIO) using System.Mem.StableName.
Its type, if possible (one technique is to use show (typeOf fun), but it requires that fun is Typeable).
Again we reach the issue where we want to do one thing if we have an f implementing Typeable and another if f does not.
How to do this?
Extra disclaimer: The goal here is not to create 'correct' Show instances for types that do not support it. There is no aspiration to be able to Read them later, or that print a != print b implies a != b.
The goal is to print any datastructure in a 'nice for human introspection' way.
The part I am stuck at, is that I want to use one implementation if extra constraints are holding for a resp. (other -> a), but a 'default' one if these do not exist.
Maybe type classes with FlexibleInstances, or maybe type families are needed here? I have not been able to figure it out (and maybe I am on the wrong track all together).
Not all problems have solutions. Not all constraint systems have a satisfying assignment.
So... relax the constraints. Store the data you need to make a sensible introspective function in your data structure, and use functions with type signatures like show, fmap, pure, and (<*>), but not exactly equal to them. If you need IO, use IO in your type signature. In short: free yourself from the expectation that your exceptional needs fit into the standard library.
To deal with things where you may either have an instance or not, store data saying whether you have an instance or not:
data InstanceOrNot c where
Instance :: c => InstanceOrNot c
Not :: InstanceOrNot c
(Perhaps a Constraint-kinded Either-alike, rather than Maybe-alike, would be more appropriate. I suspect as you start coding this you will discover what's needed.) Demand that clients that call notFmap and friends supply these as appropriate.
In the comments, I propose parameterizing your type by the constraints you demand, and giving a Functor instance for the no-constraints version. Here's a short example showing how that might look:
{-# LANGUAGE GADTs #-}
{-# LANGUAGE DataKinds #-}
{-# LANGUAGE TypeOperators #-}
{-# LANGUAGE TypeFamilies #-}
{-# LANGUAGE FlexibleInstances #-}
import Data.Kind
type family All cs a :: Constraint where
All '[] a = ()
All (c:cs) a = (c a, All cs a)
data Lol cs a where
Leaf :: a -> Lol cs a
Fmap :: All cs b => (a -> b) -> Lol cs a -> Lol cs b
instance Functor (Lol '[]) where
fmap f (Leaf a) = Leaf (f a)
fmap f (Fmap g garg) = Fmap (f . g) garg
Great timing! Well-typed recently released a library which allows you to recover runtime information. They specifically have an example of showing arbitrary values. It's on github at https://github.com/well-typed/recover-rtti.
It turns out that this is a problem which has been recognized by multiple people in the past, known as the 'Constrained Monad Problem'. There is an elegant solution, explained in detail in the paper The Constrained-Monad Problem by Neil Sculthorpe and Jan Bracker and George Giorgidze and Andy Gill.
A brief summary of the technique: Monads (and other typeclasses) have a 'normal form'. We can 'lift' primitives (which are constrained any way we wish) into this 'normal form' construction, itself an existential datatype, and then use any of the operations available for the typeclass we have lifted into. These operations themselves are not constrained, and thus we can use all of Haskell's normal typeclass functions.
Finally, to turn this back into the concrete type (which again has all the constraints we are interested in) we 'lower' it, which is an operation that takes for each of the typeclass' operations a function which it will apply at the appropriate time.
This way, constraints from the outside (which are part of the functions supplied to the lowering) and constraints from the inside (which are part of the primitives we lifted) are able to be matched, and finally we end up with one big happy constrained datatype for which we have been able to use any of the normal Functor/Monoid/Monad/etc. operations.
Interestingly, while the intermediate operations are not constrained, to my knowledge it is impossible to write something which 'breaks' them as this would break the categorical laws that the typeclass under consideration should adhere to.
This is available in the constrained-normal Hackage package to use in your own code.
The example I struggled with, could be implemented as follows:
{-# LANGUAGE GADTs #-}
{-# LANGUAGE DataKinds #-}
{-# LANGUAGE FlexibleInstances #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE MultiParamTypeClasses #-}
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE StandaloneDeriving #-}
{-# LANGUAGE ConstraintKinds #-}
{-# LANGUAGE KindSignatures #-}
{-# LANGUAGE UndecidableInstances #-}
module Example where
import Data.Dynamic
import Data.Kind
import Data.Typeable
import Control.Monad.ConstrainedNormal
-- | Required to have a simple constraint which we can use as argument to `Expr` / `Expr'`.
-- | This is definitely the part of the example with the roughest edges: I have yet to figure out
-- | how to make Haskell happy with constraints
class (Show a, Typeable a) => Introspectable a where {}
instance (Show a, Typeable a) => Introspectable a where {}
data Expr' (c :: * -> Constraint) a where
C :: a -> Expr' c a
-- M :: (a -> b) -> Expr' a -> Expr' b --^ NOTE: This one is actually never used as ConstrainedNormal will use the 'free' implementation based on A + C.
A :: c a => Expr' c (a -> b) -> Expr' c a -> Expr' c b
instance Introspectable a => Show (Expr' Introspectable a) where
show e = case e of
C x -> "(C " ++ show x ++ ")"
-- M f x = "(M " ++ show val ++ ")"
A fx x -> "(A " ++ show (typeOf fx) ++ " " ++ show x ++ ")"
-- | In user-facing code you'd not want to expose the guts of this construction
-- So let's introduce a 'wrapper type' which is what a user would normally interact with.
type Expr c a = NAF c (Expr' c) a
liftExpr :: c a => Expr' c a -> Expr c a
liftExpr expr = liftNAF expr
lowerExpr :: c a => Expr c a -> Expr' c a
lowerExpr lifted_expr = lowerNAF C A lifted_expr
constant :: Introspectable a => a -> Expr c a
constant val = pure val -- liftExpr (C val)
You could now for instance write
ghci> val = constant 10 :: Expr Introspectable Int
(C 10)
ghci> (+2) <$> val
(C 12)
ghci> (+) <$> constant 10 <*> constant 32 :: Expr Introspectable Int
And by using Data.Constraint.Trivial (part of the trivial-constrained library, although it is also possible to write your own 'empty constrained') one could instead write e.g.
ghci> val = constant 10 :: Expr Unconstrained Int
which will work just as before, but now val cannot be printed.
The one thing I have not yet figured out, is how to properly work with subsets of constraints (i.e. if I have a function that only requires Show, make it work with something that is Introspectable). Currently everything has to work with the 'big' set of constraints.
Another minor drawback is of course that you'll have to annotate the constraint type (e.g. if you do not want constraints, write Unconstrained manually), as GHC will otherwise complain that c0 is not known.
We've reached the goal of having a type which can be optionally be constrained to be printable, with all machinery that does not need printing to work also on all instances of the family of types including those that are not printable, and the types can be used as Monoids, Functors, Applicatives, etc just as you like.
I think it is a beautiful approach, and want to commend Neil Sculthorpe et al. for their work on the paper and the constrained-normal library that makes this possible. It's very cool!
Consider the following code:
{-# LANGUAGE FlexibleInstances #-}
{-# LANGUAGE UndecidableInstances #-}
class X a
class Y a
instance Y Bool
instance (Y a) => X a
instance {-# OVERLAPPING #-} X Int
f :: (X a) => a -> a
f x = x
These LANGUAGE pragma's are needed to write the above instances.
Now, say we want to write a function g:
g :: (Y a) => a -> a
g = f
Without IncoherentInstances or adding {-# INCOHERENT #-} to one of the instances, this doesn't typecheck.
But when we do add this, and ask ghci
ghci> :t f
f :: Y a => a -> a
Suddenly the type of 'f' changed?
With this small example, programs still typecheck when I give f an Int (indicating that the above would be merely a 'visual bug', but in a bigger example the same does not typecheck, giving me an error like:
Could not deduce (Y a) arising from a use of 'f
(...)
from the context: (..., X a, ...)
This also happens when we say
h = f
and try to call h with an Int
:type f does not report the type of the defined entity f. It reports the type of the expression f. GHC tries really hard to stamp polymorphism out of expressions. In particular, using f in an expression triggers simplification of the X a constraint (as does using any definition with a constraint). Without IncoherentInstances, GHC refrains from using instance Y a => X a, because there is another instance that overlaps it, so GHC needs to wait to see which one it should use. This ensures coherence; the only X Int instance that is ever used is the explicitly "specialized" one. With IncoherentInstances, you say that you don't care about coherence, so GHC goes ahead and simplifies X a to Y a using the polymorphic instance whenever f appears in an expression. The weird behavior you see where sometimes GHC is OK with using X Int and sometimes complains that there is no Y Int is a result of GHC making different internal decisions about when it wants to simplify constraints (you did ask for incoherence!). The command for seeing the type of a definition is :type +v. :type +v f should show the type of f "as declared". Hopefully, you can also see that IncoherentInstances is a bad idea. Don't use it.
Consider this code:
{-# language FlexibleInstances, UndecidableInstances #-}
module Y where
class C m where
x :: m
instance {-# overlappable #-} Monoid m => C m where
x = mempty
instance C Int where
x = 53
What is the type of x?
λ :type x
x :: C m => m
So far — so good. Now remove the Int instance. What is the type of x?
λ :type x
x :: Monoid m => m
Surprise!
Why is this happening?
This behaviour is explained in the following blog post:
Opaque constraint synonyms
In short: GHC is smart enough to see that you have only one instance of the C typeclass and decided that it's the only possible instance, so every time it sees C m constraint, it replaces it with Monoid m because they are equivalent.
N.B. As #chi further explains in a comment:
When GHC finds a constraint C t, it tries to solve it. If if finds a matching instance (...) => C t where ..., the constraint is replaced with the context (...). This is repeated as much as possible. The final constraint appears in the type (or triggers a "unsolved" type error). This process is justified since there can only be at most one matching instance. Overlapping instances change this, and prevent this context reduction when multiple instances (in scope!) match, roughly. It is a fragile extension, to be used with some care.
The following snippet makes GHC (checked with 8.6.2 & 8.4.4) stuck during compilation:
{-# LANGUAGE TypeFamilies #-}
{-# LANGUAGE FlexibleContexts #-}
{-# LANGUAGE UndecidableInstances #-}
{-# LANGUAGE UndecidableSuperClasses #-}
import GHC.Exts (Constraint)
data T = T
type family F t (c :: * -> Constraint) :: Constraint
type instance F T c = c T
class F t C => C t where
t :: C T => t
t = undefined
I think that it gets stuck because for t GHC tries to find C T, which leads to F T C which expands via type family F back to C T, which is what it was looking for (infinite loop).
I suppose that theoretically GHC could tell that it reached its quest for C T from itself and that anything that depends on itself can work fine recursively, or am I misunderstanding something?
Side note: in the real example where I stumbled upon this behaviour I was able to achieve what I wanted without the compiler being stuck by replacing UndecidableSuperClasses with Data.Constraint.Dict instead.
UndecidableSuperClasses does not make instance resolution lazy. The compiler will still expand superclass constraints as far as possible. I believe that the fields in instance dictionaries that point to the superclass dictionaries are strict, and GHC actually pins them down at compile time. This is in contrast to UndecidableInstances, which allows instance constraints to be treated (a bit) lazily.
deriving instance Show (f (Fix f)) => Show (Fix f)
will work just fine. When resolving an instance for, e.g., Show (Fix Maybe)), GHC will see that it needs Show (Maybe (Fix Maybe)). It then sees it needs Show (Fix Maybe) (which it's currently resolving) and accept that thanks to UndecidableInstances.
All UndecidableSuperClases does is disable the checks that guarantee that expansion won't loop. See the bit near the beginning of Ed Kmett's talk where he describes the process "reaching a fixed point".
Consider a working example (ripped from Data.Constraint.Forall):
type family Skolem (p :: k -> Constraint) :: k
class p (Skolem p) => Forall (p :: k -> Constraint)
GHC only accepts this with UndecidableSuperclasses. Why? Because it doesn't know anything about what that constraint might mean. As far as it knows, it could be the case that p (Skolem p) will reduce to Forall p. And that could actually happen!
class Forall P => P x
-- This definition loops the type checker
foo :: P x => x
foo = undefined
This is modification of a question a just ask some minutes ago - but for this situation I got problems...
I have the following problem: I defined a type class and want to declare tuples of types of this class to be instances as well. But I don't know how to get GHC to accept this declaration. Here a very simple example:
class Test x a where
elm :: a
And know for tuples I want to do something like
instance (Test x a, Test x b) => Test x (a,b) where
elm = (elm, elm)
Moreover, I am using
{-# LANGUAGE MultiParamTypeClasses #-}
{-# LANGUAGE FlexibleInstances #-}
The addition type x causes the trouble here...
How can this been done? Thanks in advance for any suggestions!
This has nothing whatsoever to do with tuples. The problem is simply that x occurs in the class head, but not the signature of elm – so there's no way to determine what to use as the x type, when elm turns up somewhere in your code.
Indeed, if elm is the only method of Test, then there's no need to have that x parameter at all – simply remove it, and your instance will be fine:
class Test a where
elm :: a
instance (Test a, Test b) => Test (a,b) where
elm = (elm, elm)
More likely, you actually have some other method that does use x, like
class Test x a where
elm :: a
beech :: x -> a
In this case, it might make sense to factor out elm into a simpler superclass:
class PreTest a where
elm :: a
class (PreTest a) => Test x a where
beech :: x -> a
Alternatively, you can obtain the x information in some other way, not through the method's signature. This can be done with a fundep:
{-# LANGUAGE FunctionalDependencies #-}
class Test x a | a->x where
elm :: a
This indicates that there can only be exactly one Test _ a instance for any type a, hence the compiler can unambiguously infer x from it.
But I heavily suspect that you don't really want that – such a class is much less flexible than a full two-parameter class. Modern Haskell tends to favour the largely equivalent associated type synonyms over fundep MultiParamTypeClasses, and mostly uses MTPCs when x and a are really two independent parameters.