
as an additional point of data,
https://github.com/ekmett/bytes/pull/49#issuecomment-580924670 and the
indirectly linked ticket
https://gitlab.haskell.org/ghc/ghc/issues/17767#note_251123
is an example of how not trivial a pretty simple use of quantified
constraints / related machinery can be!
a good "guiding light" for core libraries in haskell perhaps should be
"what choices jointly give the best type inference, composability,
generality, performance, and usability" and when we hit a trade off for
these parameters, make sure we understand those and have very well
understood tradeoffs.
@daniel for these sorts of instance renovation proposals, its probably
best to show case "heres the type and instance and uses" in each of the K
different flavors and what gets better/easier/harder simpler. I honestly
dont do it as often as I should myself!
In this case, putting together a tiny repo with different module doing the
various flavors and how they work might be best for grounding this. bonus
points if its cabalized so that folks on different lagging rates of ghc
versions can poke around :)
(one idea i've been thinking about is how to best make various phases of
possibly nontrivial proposals easy to evaluate and compare for impact and
benefit, but thats a discussion for another time)
as ever, be well all on this bizarre spring we're all experiencing
-Carter
On Sat, Mar 14, 2020 at 11:44 AM Oleg Grenrus
QuantifiedConstraints are not buggy, but they are not _complete_ (I do mean that as "buggy" = not sound, properly expressive = complete).
There are at least three issues:
1. Instance definition need UndecidableInstances (in my opinion this is big deal) 2. Instances are not elegant (easy to write, but not elegant). 3. QuantifiedConstraints resists to be abstracted over
See https://gist.github.com/phadej/266d68cf5cc1229c3548b7965f4335f8 for the standalone code file.
For the reasons below I wouldn't recommend using QuantifiedConstraints. I very like them, but I'm not convinced the feature is ready for "prime time". To put into perspective, I think code classes of `singletons` are more ready to be included in `base` than changing instances of Data.Functor.Sum and .Product. I do use singletons in my code more than QuantifiedConstraints. :)
I'm very worried how this change will affect libraries like `free` and `recursion-schemes` and what builds on top of them. This is not only change to `base`, it strongly guides how downstream libraries should be written (or changed) as well.
On the other hand, I don't feel strongly about
instance (Eq (f a), Eq (g a)) => Eq (Product f g a)
Yet, in the light of `free` it is "a step backwards". See https://hackage.haskell.org/package/free-5/changelog
---
Simple example with `newtype Fix f = Fix (f (Fix f))`
class Eq1 f where liftEq :: (a -> b -> Bool) -> f a -> f b -> Bool
class Eq1 f => Ord1 f where liftCompare :: (a -> b -> Ordering) -> f a -> f b -> Ordering
instance Eq1 f => Eq (Fix f) where (==) = eq where eq (Fix x) (Fix y) = liftEq eq x y
instance Ord1 f => Ord (Fix f) where compare = cmp where cmp (Fix x) (Fix y) = liftCompare cmp x y
works. It's boilerplate, but it's already written.
However, if we want to use QuantifiedConstraints, then we
1. need UndecidableInstances 2. and then the code turns out to be less elegant:
instance (forall x. Eq x => Eq (f x)) => Eq (Fix f) where (==) = eq where eq (Fix x) (Fix y) = x == y
instance (forall x. Ord x => Ord (f x)) => Ord (Fix f) where compare = cmp where cmp (Fix x) (Fix y) = compare x y
fails to compile with
GHCi, version 8.8.3: https://www.haskell.org/ghc/ :? for help [1 of 1] Compiling Main ( Ord1.hs, interpreted )
Ord1.hs:28:10: error: • Could not deduce (Ord x) arising from the superclasses of an instance declaration
we need to write Ord instance differently
instance (forall x. Ord x => Ord (f x), forall x. Eq x => Eq (f x)) => Ord (Fix f) where compare = cmp where cmp (Fix x) (Fix y) = compare x y
This _cannot_ be made to work:
forall x. Ord x => Ord (f x)
doesn't entail
forall x. Eq x => Eq (f x)
If you try to write defaultLiftEq using liftCompare
defaultLiftEq :: Ord1 f => (a -> b -> Bool) -> f a -> f b -> Bool defaultLiftEq eq x y = EQ == liftCompare _problem_ x y
then you will se a problem.
---
Third issue is that We cannot abstract over QuantifiedConstraints.
Take an example `Dict`. We can use `Dict` for various things.
data Dict :: (k -> Constraint) -> k -> * where Dict :: c a => Dict c a
It nicely uses PolyKinds extension so we can write:
eqInt :: Dict Eq Int eqInt = Dict
eq1List :: Dict Eq1 [] eq1List = Dict
And we can use Dict to *manually* thread information
entail :: Dict Ord1 f -> Dict Eq1 f entail Dict = Dict
The selling pitch of QuantifiedConstraints, that we could get this for free. Above Ord (Fix) example however makes me suspicious. Let's try:
eqQList :: Dict (forall x. Eq x => Eq (f x)) [] eqQList = undefined
But it doesn't work!
Ord1.hs:53:28: error: • Expected kind ‘(* -> *) -> Constraint’, but ‘Eq (f x)’ has kind ‘*’ • In the first argument of ‘Dict’, namely ‘(forall x. Eq x => Eq (f x))’ In the type signature: eqQList :: Dict (forall x. Eq x => Eq (f x)) [] | 53 | eqQList :: Dict (forall x. Eq x => Eq (f x)) [] |
Ord1.hs:53:36: error: • Expected a type, but ‘Eq (f x)’ has kind ‘Constraint’ • In the first argument of ‘Dict’, namely ‘(forall x. Eq x => Eq (f x))’ In the type signature: eqQList :: Dict (forall x. Eq x => Eq (f x)) [] | 53 | eqQList :: Dict (forall x. Eq x => Eq (f x)) [] | ^^^^^^^^
Then we remember that we have seen that, we **cannot define** type synonyms for quantified constraints
type Eq1' f = forall x. Eq x => Eq (f x)
errors with
Ord1.hs:56:33: error: • Expected a type, but ‘Eq (f x)’ has kind ‘Constraint’
Luckily GHC-8.10.1 (which is not released at the moment of writing) will give us ability to say
type Eq1' :: (* -> *) -> Constraint type Eq1' f = forall x. Eq x => Eq (f x)
This is promising! Let's try to fix eqQList
eqQList2 :: Dict Eq1' [] eqQList2 = undefined
But that doesn't work. Type-aliases have to be fully applied!
How in Haskell we fix issues when type-aliases (of classes) need to be partially evaluated? We defined
class ... => Example a b c instance ... => Example a b c
Third try
eqQList3 :: Dict Eq1'' f eqQList3 = Dict
The type signature is accepted, but the implementation is not
Ord1.hs:67:12: error: • Could not deduce (Eq (f x)) arising from a use of ‘Dict’
At this point I'm clueless.
---
Best regards, Oleg
P.S. If we would like to take QuantifiedConstraints somewhere into use, then IMO we should start with MonadTrans class. IIRC it's well motivated in the paper. But UndecidableInstances is very unfortunate. On 14.3.2020 16.10, chessai . wrote:
I can second Richard's estimation of QuantifiedConstraints, I have used them a lot in my own code since they were in HEAD. I consider it a sufficiently stable feature to include in base or any library.
On Sat, Mar 14, 2020, 4:16 AM Richard Eisenberg
wrote: On Mar 14, 2020, at 4:14 AM, Eric Mertens
wrote: The last thing I'd heard about quantified constraints was that they were buggy and I've been avoiding relying on them. (I should probably review that assumptions at some point.)
Without expressing an opinion about chessai's proposal (which I have not really thought about): quantified constraints are in good shape and ready for prime time. They have limitations (e.g. you can't mention a type family to the right of the =>), but when they are valid, they work well. I'll never swear that a feature is bug-free, but I think it's reasonable to consider using quantified constraints in `base`.
Richard
_______________________________________________ Libraries mailing listLibraries@haskell.orghttp://mail.haskell.org/cgi-bin/mailman/listinfo/libraries
_______________________________________________ Libraries mailing list Libraries@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/libraries