Safer and more expressive type representations

This page is the main root page for a re-design of the Typeable class, to make type-indexed type representations.

The main ticket for tracking progress is #11011. There is currently an active implementation effort which is being documented at Typeable/BenGamari

Relevant links

Also relevant is the root page for distributed Haskell, which was a major driving force for the design described here.

The names of functions and type constructors is totally up for grabs.


Use Keyword = Typeable to ensure that a ticket ends up on these lists.

Open Tickets:

isInstance does not work on Typeable with base-4.8 anymore
[TypeApplications] Create Proxy-free alternatives of functions in base
Constraint vs *
No TypeRep for unboxed tuples
TemplateHaskell and Data.Typeable - tcIfaceGlobal (local): not found
Unboxed sums are not Typeable
Typeable regression in GHC HEAD

Closed Tickets:

Easily make Typeable keys pure, so that Typeable can be handled efficiently across communications
The type defaulting in GHCi with Typeable
Remove OldTypeable
(Try to) restructure `base` to allow more use of `AutoDeriveTypeable`
Export typeRepKinds in Data.Typeable
Make Typeable track kind information better
Add type-indexed type representations (`TypeRep a`)
Missing type representations
Kind of (->) type constructor is overly constrained
Allow unsaturated uses of unlifted types in Core
Typeable on RealWorld fails
GHC crashes when calling typeRep on a promoted tuple
Unboxed tuples have no type representations
Representation polymorphism validity check is too strict
Core lint failure with pattern synonym and levity polymorphism
Perplexing levity polymorphism issue


Consider Dynamic:

data Dynamic where
  Dyn :: TypeRep -> a -> Dynamic

We'd like to write dynApply:

dynApply :: Dynamic -> Dynamic -> Maybe Dynamic

which succeeds only when the application is type correct. But how? Let's try

dynApply (Dyn tf f) (Dyn tx x) = ???

We need some way to decompose tf (the type representation for f), ask if it is an arrow type, and if so extract the type representation for the argument and result type. Then we need to compare the argument type with xt's representation. If so, we are good to go.

But at the moment we don't have any type-safe way to decompose TypeRep. Indeed dynApply is defined like this right now:

dynApply (Dyn tf f) (Dyn tx x)
  = case funResultTy tf tx of
       Just tr -> Just (Dyn tr (unsafeCoerce f x))

where funResultTy :: TypeRep -> TypeRep -> Maybe TypeRep does the decomposition of tf, checking that it is of form tx -> tr.

The unsafeCoerce makes dynApply part of the Trusted Code Base (TCB). That might not be so bad, but it is a poster-child for a raft of other similar functions (e.g. in Cloud Haskell).

So our main goal is to be able to write dynApply in a type-safe way so that it is not part of the TCB.

Step 1: Type-indexed type representations

The first obvious thing is that we must make a connection between the type-rep arg of Dyn and the value itself. Something like this:

data Dynamic where
  Dyn :: TTypeRep a -> a -> Dynamic

Here we are using a type-indexed type representation, TTypeRep. Now the connection betweeen the two is visible to the type system.

We can re-define the Typeable class and TypeRep thus:

class Typeable a where
  tTypeRep :: TTypeRep a     -- No need for a proxy argument!

data TypeRep where
  TypeRep :: TTypeRep a -> TypeRep

typeRep :: Typeable a => proxy a -> TypeRep
-- The current typeRep function
typeRep (p :: proxy a) = TypeRep (tTypeRep :: TTypeRep a)

It is helpful to have both Typeable a (the class) and TTypeRep a (the value).

  • It is sometimes convenient to pass a TTypeRep around implicitly (via a Typeable a => constraint).
  • But in tricky cases, it is also often much clearer (and less laden with proxy arguments) to pass it around as an ordinary, named value.

We can get from Typeable a to TTypeRep by using the class method tTypeRep. But what about the other way round? We need to add the following primitive:

withTypeable :: TTypeRep a -> (Typeable a => b) -> b

(This seems both simpler and more useful than making the Typeable class recursive through TypeRep data declaration.) See more on the sub-page.

We can also compare two TTypeReps to give a statically-usable proof of equality:

eqTT :: TTypeRep a -> TTypeRep b -> Maybe (a :~: b)
eqT  :: (Typeable a, Typeable b) => Maybe (a :~: b)

data a :~: b where  -- Defined in Data.Typeable.Equality
   Refl :: a :~: a

(eqT and :~: exist already as part of the exports of Data.Typeable.)

Step 2: Decomposing functions

If we want to decompose TTypable, at least in the function arrow case, we need a function like this:

decomposeFun :: TTypeRep fun
             -> r
             -> (forall arg res. (fun ~ (arg->res)) 
                              => TTypeRep arg -> TTypeRep res -> r)
             -> r
-- (decomposeFun tf def k) sees if 'tf' is a function type
-- If so, it applies 'k' to the argument and result type
-- representations; if not, it returns 'def'

This function is part of Typeable, and replaces funResultTy.

Now we can write dynApply, in a completely type-safe way, outside the TCB:

dynApply :: Dynamic -> Dynamic -> Maybe Dynamic
dynApply (Dyn tf f) (Dyn tx x)
  = decomposeFun tf Nothing $ \ ta tr ->
    case eqTT ta tx of
      Nothing   -> Nothing
      Just Refl -> Just (Dyn tr (f x))

Pattern synonyms. An alternative, rather nicer interface for decomoposeFun would use a pattern synonym instead of continuation-passing style. Here is the signature for the pattern synonym:

pattern TRFun :: fun ~ (arg -> res)
              => TTypeRep arg 
              -> TTypeRep res 
              -> TTypeRep fun

which looks (by design) very like the signature for a GADT data constructor. Now we can use TRFun in a pattern, thus:

dynApply :: Dynamic -> Dynamic -> Maybe Dynamic
dynApply (Dyn (TRFun ta tr) f) (Dyn tx x)
  = case eqTT ta tx of
      Nothing   -> Nothing
      Just Refl -> Just (Dyn tr (f x))
dynApply _ _ = Nothing

Is that not beautiful? The second equation for dynApply is needed in case the TRFun pattern does not match.

Step 3: Extracting kind equalities

Suppose we have a TTypeRep (a :: k). It might be nice to get a TTypeRep (k :: *) out of it. (Here, we assume * :: *, but the idea actually works without that assumption.) An example is an attempt to typecheck the expression

typeOf :: Typeable (a :: k) => Proxy a -> TypeRep

where typeOf :: Typeable a => a -> TypeRep is a long-standing part of the Typeable API. To typecheck that expression, we need a Typeable (Proxy a) dictionary. Of course, to build such a thing, we need Typeable k, which we have no way of getting.

So, we now include

kindRep :: TTypeRep (a :: k) -> TTypeRep k

in the API. I (Richard) conjecture that this is all possible without kind equalities, but would be rather awkward due to GHC's insistence that kind parameters be implicit. See #10343, which inspired this bit.

Step 4: Decomposing arbitrary types

It is all very well being able to decompose functions, but what about decomposing other types, like Maybe Int?

To do this it is natural to regard types as built from type constructors and binary application, like this:

data TTypeRep (a :: k) :: * where
    TRApp :: TTypeRep a -> TTypeRep b -> TTypeRep (a b)
    TRCon :: TTyCon a -> TTypeRep a

-- TTyCon must be polykinded, so that we can have
--    (TTyCon Maybe), (TTyCon Either), (TTyCon Int) etc
-- TTyCon :: forall k. k -> *

(We could, and ultimately should, use pattern synonyms again, but it's more concrete to use a GADT for now. Perhaps surprisingly, it is actually fine to expose TTypeRep concretely, with its constructors; we can't construct ill-formed TTypeReps.)

While this GADT is expressible in GHC now (note the existential kind in TRApp), it is not very useful without kind equalities. (GHC does not currently support kind equalities, but Richard Eisenberg is working that very question.) Why? Here is the type of TRApp in its full glory, with normally-invisible kind args in angle brackets:

TRApp :: forall k1 k2. forall (a :: k1 -> k2) (b :: k1).
         TTypeRep <k1->k2> a -> TTypeRep <k1> b -> TTypeRep <k2> (a b)

Or, to be really explicit about the existentials:

TRApp :: forall k2 (c:k2).                      -- Universal
         forall k1 (a :: k1 -> k2) (b :: k1).   -- Existential
         (c ~ a b)
      => TTypeRep <k1->k2> a 
      -> TTypeRep <k2>     b 
      -> TTypeRep <k2>     c

Now suppose we want to implement decomposeFun. We should be able to do this outside the TCB, i.e. without unsafeCoerce:

arrowCon :: TTyCon (->)   -- The type rep for (->)

decomposeFun :: TTypeRep fun
             -> r
             -> (forall arg res. (fun ~ (arg->res)) 
                              => TypeRep arg -> TTypeRep res -> r)
             -> r
decomposeFun tr def kont
  = case tr of
      TRApp (TRApp (TRCon c) r1) r2 -> case eqTyCon arrowCon c of 
          Just HRefl -> kont r1 r2
          Nothing    -> def
      _ -> default 

But look at the arguments of eqTyCon:

   arrowCon :: TTyCon <*->*->*>   (->)
   c        :: TTyCon <k1->k2->*> tc

where k1 and k2 are existential kinds bound by the two nested TRApp constructors, and tc the existential bound by the inner TRApp. But kont is expecting arg and res to have kind *! So we need proofs that k1 ~ * and k2 ~ *.

The real work is done by eqTyCon:

eqTyCon :: forall (k1 k2 :: *). 
           forall (a :: k1) (b :: k2). 
           TTyCon <k1> a -> TTyCon <k2> b -> Maybe (a :~~: b)

where :~~: is a kind-heterogeneous version of :~::

data (a::k1) :~~: (b::k2) where
  HRefl :: forall (a::k).  a :~~: a

Or, to write the type of HRefl with its constraints explicit:

HRefl :: forall k1 k2. forall (a::k1) (b::k2).
         (k1 ~ k2, a ~ b) 
      => a :~~: b

That is, HRefl encapsulates a proof of kind equality as well as one of type equality.

So Step 3 (allowing TypeRep to be fully decomposed in a type safe way) absolutely requires kind equalities.

Other points

Fast comparison of TypeRep

The current implementation of TypeRep allows constant-time comparison based on fingerprints. To support this in the new scheme we would want to add a fingerprint to every TypeRep node. But we would not want clients to see those fingerprints, lest they forge them.

Conclusion: make TypeRep abstract. But then how can we pattern-match on it? Pattern synonyms seem to be exactly what we need.

Trusted computing base

With all of this, what is left in the TCB? The TTyCon type is a new abstract type, and the comparison (based on fingerprints) must be in the TCB.

TTyCon a  --- abstract type
eqTyCon :: forall k1 k2. forall (a :: k1) (b :: k2). 
           TTyCon a -> TTyCon b -> Maybe (a :~~: b)
tyConKind :: TTyCon (a :: k) -> TTypeRep (k :: *)

Note that TTyCon represents type constructors already applied to any kind arguments. Even with kind equalities, we have no way of creating a representation for Proxy :: forall k. k -> *, as that would require an impredicative kind for the implicit kind argument to TTyCon.

withTypeable could be written in core (and perhaps could be generalized to other constraints) but not in source Haskell:

withTypeable :: TypeRep a -> (Typeable a => b) -> b

See more here.

Implementing the "fast comparison" idea above will require putting more code in the TCB because of interaction with fingerprints. But the functions above are morally the only ones that need be trusted.

What about structure information?

Related but different issue:

Step by step

We can do steps 1 and 2 (and probably 3) without kind equalities, although the implementation of decomposeFun will use unsafeCoerce and will be part of the TCB. However, this will all likely happen after merging with kind equalities.

Proposed API

This version assumes homoegeneous equality (as GHC is today, July 2, 2015). Below is the version with heterogeneous equality.

data TTypeRep (a :: k)        -- abstract
data TTyCon (a :: k)          -- abstract

-- Costructors
trCon :: forall k (a :: k). TTyCon a -> TTypeRep a  -- in TCB
trApp :: forall k k1 (a :: k1 -> k) (b :: k1). 
         TTypeRep a -> TTypeRep b -> TTypeRep (a b)  -- in TCB

-- Destructors
pattern type TRCon :: forall k (a :: k). TyCon a -> TTypeRep a
pattern type TRApp :: forall k. exists k1 (a :: k1 -> k) (b :: k1). 
                      TTypeRep a -> TTypeRep b -> TTypeRep (a b)
pattern type TRFun :: TTypeRep arg 
                   -> TTypeRep res 
                   -> TypeRep (TTypeRep (arg->res))

-- For now, homogeneous equalities
eqTyCon :: forall k (a :: k) (b :: k). 
           TTyCon a -> TTyCon b -> Maybe (a :~: b)
eqTT    :: forall k (a :: k) (b :: k). 
           TTypeRep a ->T TypeRep b -> Maybe (a :~: b)

This assumes heterogeneous equality (that is, with kind equalities).

data TTypeRep (a :: k)        -- abstract
data TTyCon (a :: k)          -- abstract

tyConKind :: forall k (a :: k). TTyCon a -> TTypeRep k

-- Costructors
trCon :: forall k (a :: k). TTyCon a -> TTypeRep a
trApp :: forall k k1 (a :: k1 -> k) (b :: k1). 
         TTypeRep a -> TTypeRep b -> TTypeRep (a b)

-- Destructors
pattern type TRCon :: forall k (a :: k). TyCon a -> TTypeRep a
pattern type TRApp :: forall k. exists k1 (a :: k1 -> k) (b :: k1). 
                      TTypeRep a -> TTypeRep b -> TTypeRep (a b)

-- this definition to go into Data.Type.Equality
data (a :: k1) :~~: (b :: k2) where
   HRefl :: a :~~: a

eqTyCon :: forall k1 k2 (a :: k1) (b :: k2). 
           TTyCon a -> TTyCon b -> Maybe (a :~~: b)

-- the following definitions can be defined on top of the interface
-- above, but would be included for convenience
pattern type TRFun :: TTypeRep arg -> TTypeRep res
                   -> TTypeRep (arg -> res)

eqTT    :: forall k1 k2 (a :: k1) (b :: k2). 
           TTypeRep a -> TTypeRep b -> Maybe (a :~~: b)

typeRepKind :: forall k (a :: k). TTypeRep a -> TTypeRep k

class forall k (a :: k). Typeable k => Typeable a where
  tTypeRep :: TTypeRep a

SLPJ: Do we need both :~: and :~~:?

RAE: I don't think it's strictly necessary, but keeping the distinction might be useful. :~: will unify the kinds of its arguments during type inference, and that might be what the programmer wants. I don't feel strongly here.

Some of this design is motivated by the desire to allow flexibility in the implementation to allow for fingerprinting for fast equality comparisons. Naturally, the fingerprints have to be in the TCB. If it weren't for them, though, the TypeRep type could be exported concretely.

Could we simplify this a bit by removing TyCon? RAE: No.

The class declaration for Typeable is highly suspect, as it is manifestly cyclic. However, (forgetting about implementation) this doesn't cause a problem here, because all kinds have kind *. Thus, we know the cycle terminates. Implementation is an open question, but I (RAE) surmise that this barrier is surmountable.

RAE: How can we get a TTyCon for a known-at-compile-time tycon? I want something like tyConRep @(->) that will give something of type TTyCon (->). The problem is that the type argument to tyConRep might not be a bare type constructor... but I would hate to have such a function return a Maybe, as it would be very annoying in practice, and it could be checked at compile time. I almost wonder if a new highly-magical language primitive would be helpful here.

Iceland_jack: I recall a conversation about making :~: a pattern synonym for :~~:


unsafeCoerce can be written

As painfully demonstrated (painful in the conclusion, not the demonstration!) in comment:16:ticket:9858, Typeable can now (7.10.1 RC1) be abused to write unsafeCoerce. The problem is that today's TypeReps ignore kind parameters.

Drawing this out somewhat: we allow only Typeable instances for unapplied constructors. That is, we have

deriving instance Typeable Maybe


deriving instance Typeable (Maybe Int)

However, what if we have

data PK (a :: k)

? PK properly has two parameters: the kind k and the type a :: k. However, whenever we write PK in source code, the k parameter is implicitly provided. Thus, when we write

deriving instance Typeable PK

we actually supply a kind parameter k0. GHC tries to find a value for k0, fails, and leaves it as a skolem variable. We get an instance forall k. Typeable (PK k). But, of course, PK * and PK (* -> *) should have different TypeReps. Disaster!

A stop-gap solution

NB: This was not implemented. The "medium-term solution" below was, as of 7.10.1.

comment:16:ticket:9858 is a demonstration that the boat is taking on water! Fix the leak, fast!

The "obvious" answer -- don't supply the k here -- doesn't work. The instance for PK would become Typeable <forall k. k -> *> PK, where a normally-implicit kind parameter is supplied in angle brackets. This is an impredicative kind, certainly beyond GHC's type system's abilities at the moment, especially in a type-class argument.

One (somewhat unpleasant) way forward is to allow kind parameters to be supplied in Typeable instances, but still restrict type parameters. So, we would have

deriving instance Typeable (PK :: * -> *)
deriving instance Typeable (PK :: (* -> *) -> *)

which would yield instances Typeable <* -> *> (PK <*>) and Typeable <(* -> *) -> *> (PK <* -> *>) -- two separate instances with different, and distinguishable TypeReps. What's painful here is that, clearly, there are an infinite number of such instantiations of PK. We could maybe provide a handful of useful ones, but we could never provide them all.

That, then, is the current plan of attack. (No, it's not.) Typeable instance heads must include concrete instantiations of all kind parameters but no type parameters. base will provide several instantiations for poly-kinded datatypes, but users may have to write orphan instances to get others. AutoDeriveTypeable will ignore poly-kinded datatype definitions -- users must write explicit instances if they want to.

Question: With this design, orphan instances will be unavoidable. Given that two different authors may introduce the same orphan instances, would it work to mark every Typeable instance as {-# INCOHERENT #-}? For this to work out, we would need a guarantee that the fingerprints of the instances are the same regardless of originating module.

Medium-term solution

This is implemented in 7.10.

Although it is impossible to create all necessary Typeable instances for poly-kinded constructors at the definition site (there's an infinite number), it is possible to create Typeable "instances" on demand at use sites. The idea (originally proposed in comment:20:ticket:9858) is to add some custom logic to the solver to invent Typeable evidence on demand. Then, whenever the solver needs to satisfy a Typeable constraint, it will just recur down to the type's leaves and invent evidence there.

For poly-kinded type constructors, we still need to worry about kind parameters. This is easy, actually: we just come up with some mechanism with which to incoporate kind parameters into a TypeReps fingerprint. One such mechanism would be to use the current fingerprint-generation algorithm (used on types) and just apply it to the kinds. Of course, kinds are different from types, but there's no reason we can't use the same algorithm. Problem solved.

Some drawbacks:

  • There is no way to inspect a TypeRep's kind or its kind parameters. But that's OK for now.
  • Under this proposal, everything is Typeable. But maybe that's not so bad.
  • It seems much more efficient to generate type constructor fingerprints at definition sites. Doing so would add to the code size of every type-declaring module. Worse, we would have to add data constructor fingerprints, too, considering the possibility of promotion.
  • More complexity in the solver.

Long-term solution

We actually don't have a good long-term solution available. We thought that * :: * and kind equalities would fix this, but they don't. To see why, let's examine the type of trApp:

trApp :: forall k k1 (a :: k1 -> k) (b :: k1). 
         TTypeRep a -> TTypeRep b -> TTypeRep (a b)

In the * :: * world, it would be reasonable to have TTypeReps for kinds, assuming we have TTypeReps always take explicit kind parameters. So, we might imagine having a TTypeRep for PK (call it pkRep) and a TTypeRep for Bool (call it boolRep). Does pkRep `trApp` boolRep` type-check? Unfortunately, no. We have

pkRep   :: TTypeRep <forall k. k -> *> PK
boolRep :: TTypeRep <*> Bool

but trApp says that a must have type k1 -> k for some k1 and k. Here PK would be the value for a, but PK's kind is forall k. k -> *, which doesn't fit k1 -> k at all! We would need to generalize trApp to

trApp2 :: forall (k1 :: *) (k :: k1 -> *)
                 (a :: pi (b :: k1). k b) (b :: k1).
          TTypeRep <pi (b :: k1). k b> a -> TTypeRep <k1> b
       -> TTypeRep <k b> (a b)

Note that the kind of a is a Π-type, dependent on the choice of b, and that x -> y would be considered shorthand for pi (_ :: x). y. While Richard's branch includes support for such types, there are no plans for type-level lambda or higher-order unification, both of which would be necessary to make this definition usable. For example, calling trApp2 on PK would still fail, because we try to match pi (b :: k1). k b with forall k2. k2 -> *. The k2 is not the last parameter of the body of the forall, so straightforward unification fails. We must choose k1 := *, k := \x. x -> * to get the types to line up. Urgh.

Why kind equalities, then?

Given the fact that Richard's branch doesn't solve this problem, what is it good for? It still works wonders in the mono-kinded case, such as for decomposing ->. It's just that poly-kinded constructors are still a pain.

Last modified 5 days ago Last modified on Feb 22, 2017 11:58:09 AM