Allow unconstrained existential contexts in newtypes

changed weight to 5

added Tfeature request Trac import labels

Why is this useful?

In GHC it's disallowed because an existential is modeled as a data constructor with a field that captures the type in the same way that it captures the value fields. But since the representation of a newtype is supposed to be the same as the representation of the (single) field, you can't do that.

So, it'd be quite difficult to make GHC do this (i.e. it'd affect GHC's typed intermediate language); and I can't see any useful applications.

Simon

closed

Please reopen this ticket if you have a useful application for it.

Trac metadata

Trac field	Value
Resolution	Unresolved → ResolvedWon'tFix

Trac metadata

Trac field	Value
Architecture	Unknown → Unknown/Multiple

Trac metadata

Trac field	Value
Operating system	Unknown → Unknown/Multiple

reopened

This would actually be a useful feature with GADTs, since matching on one can tell us what the existential field was, even without storing a context for it in our newtype.

It's definitely not essential to anything I'm doing, but perhaps the recent changes to GHC since the last time people looked at this ticket make this feature easier to implement?

Trac metadata

Trac field	Value
TypeOfFailure	- → OtherFailure
Resolution	ResolvedWon'tFix → Unresolved

Okay, so this ticket-revival stemmed from a question I asked on IRC, so here's what I wanted to use this for.

class (Monad (m context)) => MonadContext m context where
  getContext
    :: m context context
  withContext
    :: MonadContext m context'
    => context'
    -> m context' a
    -> m context a


data FallibleContextualOperation failure context a =
  FallibleContextualOperation {
      fallibleContextualOperationAction :: context -> IO (Either failure a)
    }


instance Monad (FallibleContextualOperation failure context) where
  return a = FallibleContextualOperation {
                 fallibleContextualOperationAction = \_ -> return $ Right a
               }
  x >>= f =
    FallibleContextualOperation {
        fallibleContextualOperationAction = \context -> do
          v <- fallibleContextualOperationAction x context
          case v of
            failure@(Left _) -> return failure
            Right y -> fallibleContextualOperationAction (f y) context
      }


instance MonadContext (FallibleContextualOperation failure) context where
  getContext =
    FallibleContextualOperation {
        fallibleContextualOperationAction =
          \context -> return $ Right context
      }
  withContext context' x =
    FallibleContextualOperation {
        fallibleContextualOperationAction =
          \context -> fallibleContextualOperationAction x context'
      }

I can provide more details of what this is motivated by upon request, but basically, I want to have my "master implementation" of this class based on FallibleContextualOperation, and then I want to make two types that get that behavior for free:

newtype Serialization context a =
  forall backend
  . Serialization {
        serializationOperation :: FallibleContextualOperation
                                    (SerializationFailure backend)
                                    context
                                    a
      }
deriving instance Monad (Serialization context)
deriving instance MonadContext Serialization context


newtype Deserialization context a =
  forall backend
  . Deserialization {
        deserializationOperation :: FallibleContextualOperation
                                    (DeserializationFailure backend)
                                    context
                                    a
      }
deriving instance Monad (Deserialization context)
deriving instance MonadContext Deserialization context

The point is so that I can then do:

class Serializable context a where
  serialize :: Serialization context a
  deserialize :: Deserialization context a

.. and write many instances of this, elsewhere in my program, which don't need to know or care about the fact that there are two backend types (ByteStrings and open files), or any of this other nasty plumbing detail.

I don't know if this is sufficient motivation to justify the work the ticket is requesting, since I don't know just how much work it is. But it's presented here for everyone's edification. :)

I also have run into cases where I'd like to make a "trivial" GADT into a newtype. In particular, when programming in a "dependent"-like style using GADTs to encode the tags of dependent sums, it's often nice to have a wrapper that does nothing but alter the type index - a very general example would be "data Map f t a where Map :: t a -> Map f t (f a)" - which allows transforming the last type parameter by an arbitrary type-level function. Most of the time this usage could be avoided by including 'f' in the dependent sum type itself, but that complicates other things and even then I've found you occasionally want to modify the types of tags.

Another situation where I've wanted something like this to forget a phantom type - and I've seen the same pattern in other people's code I've read on hackage.

There is a different possible resolution to this request, which would probably be more work but not break portability - guarantee that any "data" declaration satisfying certain criteria will be compiled to a newtype-like representation. Here's my stab at that set of criteria:

Only one constructor
Only one field with nonzero width in that constructor (counting constraints as fields)
That field is marked strict

It seems like those requirements should be sufficient to justify special-case handling to compile them to something effectively the same as a newtype. Or if some mechanism already causes this to effectively by the case, then I'd be happy with that being documented and test-cases added to ensure it continues to be a stated goal to cover situations like these.

changed milestone to %7.6.1

Anything involving existentials is going to be hard to implement using newtype directly. But as 'mokus' says, it might be possible to make a guarantee, right in the code generator, that certain sorts of data types do not allocate a box. The conditions are, I think, very nearly as 'mokus' says:

Only one constructor
Only one field with nonzero width in that constructor (counting constraints as fields)
That field is marked strict
That field has a boxed (or polymorphic) type

I think this'd be do-able. The question is how important it is in practice; it's one more thing to maintain.

Simon

changed milestone to %7.6.2

Just a minor bump on this. The more fancy GADT and Poly/DataKind programming I do, the more this bothers me. It seems like a real pity to be penalized (with an extra box) for choosing to encode invariants in indices, especially with all the new delicious features GHC has been getting recently.

To be clear about what I'm looking for:

data IntList where
  Nil :: IntList
  Cons :: Int -> IntList -> IntList

data Nat = Z | S Nat

data IntVec :: Nat -> * where
   NilV :: IntVec Z
   ConsV :: Int -> IntVec n -> IntVec (S n)

-- N.B: not data
newtype Exists (f :: k -> *) = forall x. Exists (f x)

type IntVecList = Exists IntVec

-- IntList and IntVecList should be isomorphic! If Exists can't be a newtype, I have to pay a penalty for adding indices to my type :(

Replying to [ticket:1965#comment:84463 pumpkin]:

To be clear about what I'm looking for:

Yes, I always wondered how existentially quantifying over a type constructor with non-* domain kinds can ever necessitate extra information at runtime. I fully support the reopen proposal because of data Hidden :: k -> * where Hide :: t a -> Hidden t wants to be a newtype.

changed milestone to %7.10.1

Moving to 7.10.1.

Dan Doel pointed out that as useful as this could be, it would allow this fairly odd use case:

data Nat = Z | S Nat

data Fin n where
  Zero :: Fin (S n)
  Suc  :: Fin n -> Fin (S n)

newtype SomeFin where
  SomeFin :: Fin n -> SomeFin

suc :: SomeFin -> SomeFin
suc (SomeFin n) = SomeFin (Suc n)

inf :: SomeFin
inf = suc inf

Which leads to inf containing a Fin whose existential index can't actually be any valid (finite) type. Fin thus becomes a bit of a misnomer in this context, since it's not finite. Making SomeFin into data rather than newtype fixes the problem by making inf into a bottom.

removed milestone

Moving to 7.12.1 milestone; if you feel this is an error and should be addressed sooner, please move it back to the 7.10.1 milestone.

mentioned in issue #10016

Replying to [ticket:1965#comment:57127 simonpj]:

Only one field with nonzero width in that constructor (counting constraints as fields)

I'd like to point out (as may already have been indented), that read literally, this should include things like

data Dict :: Constraint -> * where
  Dict :: a => Dict a

from the constraints package.

Trac metadata

Trac field	Value
CC	ireney.knapp@gmail.com, jake.mcarthur@gmail.com, mokus@deepbondi.net, pumpkingod@gmail.com, sweirich@cis.upenn.edu → ireney.knapp@gmail.com, jake.mcarthur@gmail.com, mokus@deepbondi.net, oerjan, pumpkingod@gmail.com, sweirich@cis.upenn.edu

changed milestone to %8.0.1

Milestone renamed

Use cases like pumpkin's Some have been cropping up more and more frequently for me; I think this would be pretty useful.

Edit: here's a link to an actual Some in the wild: https://hackage.haskell.org/package/dependent-map-0.2.1.0/docs/Data-Dependent-Map.html#t:Some

But the Some at the link is this:

data Some tag :: (k -> *) -> * where
  This :: !(tag t) -> Some k tag

There's no reason why this can't be a newtype today, is there?

So it doesn't look relevant to this ticket. (Except I suppose that if ticket:1965#comment:57127 was implemented, it'd be efficient already.)

Hmm, when I try to convert it to a newtype, like this:

newtype Some tag = forall t. This (tag t)

I get this error:

src/Data/Some.hs:17:20:
    A newtype constructor cannot have existential type variables
    This :: forall (k :: BOX) (tag :: k -> *) (t :: k).
            tag t -> Some tag
    In the definition of data constructor ‘This’
    In the newtype declaration for ‘Some’

Is there a different way of converting it to a newtype that I'm overlooking?

Here's a link to the original source of that Some: https://hackage.haskell.org/package/dependent-sum-0.3.2.1/docs/src/Data-Some.html#Some . The Hoogle output is more confusing to me than the original.

Oh sorry, my fault. I stupidly missed the fact that t was existential. Just ignore ticket:1965#comment:118559. So this is an example of the ticket.

Back to ticket:1965#comment:57127 then, I think.

Replying to [ticket:1965#comment:57127 simonpj]:

Anything involving existentials is going to be hard to implement using newtype directly. But as 'mokus' says, it might be possible to make a guarantee, right in the code generator, that certain sorts of data types do not allocate a box. The conditions are, I think, very nearly as 'mokus' says:

Only one constructor

Only one field with nonzero width in that constructor (counting constraints as fields)

That field is marked strict

That field has a boxed (or polymorphic) type

I think this'd be do-able. The question is how important it is in practice; it's one more thing to maintain.

I would like to have something like this very much! Among other things, it's one possible way to make IntMap nicer. One potential extension: I think constraints only need to count as fields if any of them are classes that have methods, or whose superclasses have methods. In particular, it could be very useful to have equality constraints involving type families.

I've started a wiki page to give examples of why we want this.

changed the description

Possible example from A Foundation for GADTs and Inductive Families, encoding GADTs as left Kan extensions can be reworked to:

data N = O | S N

data LanO j where 
  LanO :: LanO (S i)

data LanS j where 
  LanS :: Fin i -> LanS (S i)

Fin defined as

data Fin a where
  FinO :: LanO i -> Fin i
  FinS :: LanS i -> Fin i

pattern FinO'   = FinO LanO
pattern FinS' n = FinS (LanS n)

Do LanO, LanS satisfy the conditions to be newtypes?

Original code===

data Fin :: * -> * where
  NZero :: Lan S One j -> Fin j
  NSucc :: Lan S Fin j -> Fin j

data One j      = One
data Lan h _X j = forall i. Lan (Eql j (h i), _X i)
data Eql i j    where Refl :: Eql i i

Replying to [ticket:1965#comment:121781 Iceland_jack]:

data LanO j where 
LanO :: LanO (S i)

data LanS j where 
LanS :: Fin i -> LanS (S i)

Fin defined as

Do LanO, LanS satisfy the conditions to be newtypes?

LanO does not, because its constructor takes no arguments. But that's perfectly fine; there's only one LanO value anyway. LanS does not either, because its constructor is lazy. If you wrote LanS :: !(Fin i) -> LanS (S i) then it would satisfy the conditions.

LanO might, however, be a candidate for being zero width when in a strict field, which could help other types satisfy the newtype condition.

mentioned in issue #15590 (closed)

Note that data A = A !Int and newtype A = A Int have subtly different surface language semantics when they're pattern matched on. What would case undefined of A _ -> 1 evaluate to? For newtypes, it's 1 whereas for strict data types this would blow up.

I guess what I'm saying is: The suggested lowering must preserve data semantics and can never behave the same as actual newtypes. Semantically, data is data and newtype is newtype. It would get rid of the performance implications, though.

See this reddit thread for a longer discussion: https://www.reddit.com/r/haskell/comments/6xri4d/whats_the_difference_between_newtype_type_and_data/dmi19pd

Yes for @sgraf's reason I wish we used an equivalence relation on these internal notion of deep representations, and also RuntimeReps. For example, given matching deep structure (not tracked in the kind/RuntimeRep):

forall a t. exists t' t'. (t :: Kind a) ~R (t' :: Kind TupleRep [a]) ~R (t'' :: Kind SumRep [a])

Maybe it's not worth the extra coercions in practice, but certainly I think its good teaching to pretend newtypes are always:

  newtype Foo (a :: k) :: TupleRep [k] = Foo (a :: k)

to explain

newtype A :: TupleRep [LiftedRep] = A (Int :: LiftedRep)
case undefined of A _ -> 1

contrasted with pretend elaborated unpacked !Int:

data A' :: LiftedRep = A' (!Int :: IntRep)
case undefined of A' _ -> undefined

That the shallow representations of A and A' have different strictness (opposite of their interior's shallow strictness), cleanly justifies how the case-expressions have different semantics despite the overall isomorphism of A and A'.

added Pnormal label

I know there are a lot of open questions about how to implement the newtype optimization optimally for these types. But it seems fairly clear that just doing the simplest possible thing in STG will improve matters, taking the nasty work-arounds out of user code. Can we maybe just get that now, and see if someone comes up with a better implementation later?

Allow unconstrained existential contexts in newtypes

Child items ...

Activity

Original code===

Allow unconstrained existential contexts in newtypes

Relates to

Activity

Original code===