I am working on the Trees that Grow stuff, and hit a small problem I have type family XIB x thing type family XNewImplicitBndrs x thing type ForallXImplicitBndrs (c :: * -> Constraint) (x :: *) (thing :: *) = ( c (XIB x thing) , c (XNewImplicitBndrs x thing) ) and I want to extend the DataId constraint type DataId p = ( Data p , ForallXImplicitBndrs Data p thing ) But the problem is I do not have `thing` at this point, and to get it in the code will involve some hs-boot nastiness. Is there any way to require "forall thing. Data thing" inside the DataId constraint? Alan
Hi Alan, One way is to define a new typeclass. type ImplicitBndrs c x thing = (c (XIB x thing), c (XNewImplicitBndrs x thing)) class ForallImplicitBndrs c x where withImplicitBndrs :: forall thing. (ImplicitBndrs c x thing => t) -> t Although that requires you to write one instance for every combination of (c, x). The "universal instance" of ImplicitBndrs can also be wrapped in the Dict type in the constraints library. In fact, constraints has an even more general mechanism for all of this in Data.Constraint.Forall. Pro: no extra instances are necessary, so it is more easily extensible. Con: internally relies heavily on unsafeCoerce, although the above is already stretching the language a lot. We start by defining a "class synonym" instead of a type synonym, to enable partial application. class (c (XIB x thing), c (XNewImplicitBndrs x thing)) => ImplicitBndrs c x thing instance (c (XIB x thing), c (XNewImplicitBndrs x thing)) => ImplicitBndrs c x thing -- "forall thing. ImplicitBndrs c x thing" type ForallImplicitBndrs c x = Forall (ImplicitBndrs c x) withImplicitBndrs :: forall c x thing . ForallImplicitBndrs c x => (ImplicitBndrs c x thing => t) -> t withImplicitBndrs t = case inst @(ImplicitBndrs c x) @thing of Sub Dict -> t I think the ICFP paper "Quantified class constraints" talks about a more principled future solution, and there is an old GHC ticket or two about it. Cheers, Li-yao On 10/23/2017 11:06 AM, Alan & Kim Zimmerman wrote:
I am working on the Trees that Grow stuff, and hit a small problem
I have
type family XIB x thing type family XNewImplicitBndrs x thing
type ForallXImplicitBndrs (c :: * -> Constraint) (x :: *) (thing :: *) = ( c (XIB x thing) , c (XNewImplicitBndrs x thing) )
and I want to extend the DataId constraint
type DataId p = ( Data p
, ForallXImplicitBndrs Data p thing )
But the problem is I do not have `thing` at this point, and to get it in the code will involve some hs-boot nastiness.
Is there any way to require "forall thing. Data thing" inside the DataId constraint?
Alan
I’m lost. Could you give me a bit more context? I’m deeply suspicious about that ForallXImplicitBndrs thing with strange higher kinded parameters. Smells all wrong to me. Simon From: ghc-devs [mailto:ghc-devs-bounces@haskell.org] On Behalf Of Alan & Kim Zimmerman Sent: 23 October 2017 16:07 To: ghc-devs@haskell.org Subject: forall in constraint I am working on the Trees that Grow stuff, and hit a small problem I have type family XIB x thing type family XNewImplicitBndrs x thing type ForallXImplicitBndrs (c :: * -> Constraint) (x :: *) (thing :: *) = ( c (XIB x thing) , c (XNewImplicitBndrs x thing) ) and I want to extend the DataId constraint type DataId p = ( Data p , ForallXImplicitBndrs Data p thing ) But the problem is I do not have `thing` at this point, and to get it in the code will involve some hs-boot nastiness. Is there any way to require "forall thing. Data thing" inside the DataId constraint? Alan
In Shayan's implementation he has [1]
data ImplicitBndrs x thing
= IB
(XIB x thing)
thing
| NewImplicitBndrs
(XNewImplicitBndrs x thing)
type family XIB x thing
type family XNewImplicitBndrs x thing
type ForallXImplicitBndrs c x thing =
( c (XIB x thing)
, c (XNewImplicitBndrs x thing)
)
This gets used, in the same file as
type LSigType x = ImplicitBndrs x (LType x)
where `thing` is resolved to a specific type.
Because it is all in the same file, there is no problem making a
constraint on anything using LSigType, that mentions LHsType.
But in the approach I am taking[2], the type families are defined in
HsExtension, which is compiled early in the cycle, and imported by
HsTypes, HsBinds, HsDecl etc.
In order to derive a Data instance for anything using `LSigType x`, we
need to be able to specify that a Data instance exists for `LHsType x`.
So we can either do that directly in HsBinds, or try to add it to the
existing
DataId constraint in HsExtension.
The first approach leads to a spread of the constraint throughout the AST,
which gets very messy.
The second approach requires being able to specify a
`forall thing. Data thing` constraint in HsExtension.
I tried an intermediate approach, introducing a constraint in HsDecls[3] to
capture this,
but it eventually runs into needing it in the HsExpr.hs-boot file, which
means I need
LHsType in that file.
Perhaps the simplest way forward is to get rid of the `thing` parameter
completely,
and introduce the three or so ImplicitBinders variants that are used.
I hope this does not just confuse things even more.
Alan
[1]
https://github.com/ghc/ghc/blob/wip/GrowableAST/compiler/hsSyn/AST.hs#L475
[2] https://github.com/ghc/ghc/tree/wip/ttg/2017-10-21
[3]
https://github.com/ghc/ghc/commit/22812296818fe955752fa4762cf72250abd09bf9#d...
On 23 October 2017 at 23:04, Simon Peyton Jones
I’m lost. Could you give me a bit more context?
I’m deeply suspicious about that ForallXImplicitBndrs thing with strange higher kinded parameters. Smells all wrong to me.
Simon
* Like I say I am DEEPLY suspicious of ForallXImplicitBndrs. I can’t make head or tail of it. Is see in you patch you define
type ForallXPat (c :: * -> Constraint) (x :: *) =…
What is this? Why do we need it? What goes wrong if we remove it altogether?
* Likewise `forall thing. Data thing` makes no sense to me as a constraint, even with quantified context. Shayan and I discussed this at ICFP and agreed it made no sense. The hoped-for quantified-context extension is NOT a solution.
Overall I’ve lost track of these enormous constraint tuples that seem to be associated with Data instances. Can you give a small artificial (ie not full GHC) example of why they are necessary? Perhaps it’s this
data T p = T1 (XT1 p) Int
| T1 (XT2 p) (IdP p)
I suppose that a Data instance would need to have
instance (Data (XT1 p), Data (XT2p), Data (IdP p)) => Data IT p) where…
Is that right? In which case why do you have all this PostRn stuff in the DataId type? (And why is it called DataId?)
* “The first approach leads to a spread of the constraint throughout the AST, which gets very messy.” I don’t understand what the first approach is, or why it gets messy. Could you be more concrete?
* “Perhaps the simplest way forward is to get rid of the `thing` parameter completely, and introduce the three or so ImplicitBinders variants that are used.” I don’t’ think it could possibly make anything simpler to have three separate data types. Can you illustrate concretely?
It’d be great to explore these issues with small, concrete examples, rather than the full glory of GHC, both for our own benefit and the benefit of those who will review the patch and (in future) understand the code.
Sorry to be slow
Simon
From: Alan & Kim Zimmerman [mailto:alan.zimm@gmail.com]
Sent: 23 October 2017 22:36
To: Simon Peyton Jones
participants (3)
-
Alan & Kim Zimmerman -
Li-yao Xia -
Simon Peyton Jones