If you care about performance then explicit dictionary passing is going to be worse than using type classes.

Of course! But explicit dictionary passing works great for tests: the code size is reduced from O(#modules * #tests) when using the module system to O(#modules + #tests) when using dictionaries. For example, in the algebraic-graphs library, I have 500+ generic tests and around 10 modules. I don't want to write 5000 tests! Here is an example generic test which uses explicit dictionary passing: https://github.com/snowleopard/alga/blob/master/test/Algebra/Graph/Test/Gene.... I don't think it would be possible to reuse this test for different graph data types by using the module system instead of dictionaries. (Perhaps, Backpack could help? I don't know it very well.) Cheers, Andrey -----Original Message----- From: Matthew Pickering [mailto:matthewtpickering@gmail.com] Sent: 30 May 2019 18:26 To: Andrey Mokhov Cc: ghc-devs@haskell.org; Andreas Klebinger Subject: Re: Container type classes If you care about performance then explicit dictionary passing is going to be worse than using type classes. At that point though, what do you gain from using the module system as you are just going to pass the same dictionaries into every function and never change them. So, for me, keep using modules but make the APIs of each module more consistent if you think it's worthwhile. On Thu, May 30, 2019 at 6:11 PM Andrey Mokhov wrote:

Hi all,

I tried to use type classes for unifying APIs of several similar data structures and it didn't work well. (In my case I was working with graphs, instead of sets or maps.)

First, you rarely want to be polymorphic over the set representation, because you care about performance. You really want to use that Very.Special.Set.insert because it has the right performance characteristics for your task at hand. I found only *one* use-case for writing polymorphic functions operating on something like IsSet: the testsuite. Of course, it is very nice to write a single property test like

memberInsertProperty x set = (member x (insert x set) == True)

and then use it for testing all set data structures that implement `member` and `insert`. Here you don't care about performance, only about correctness!

However, this approach leads to problems with type inference, confusing error messages, and complexity. I found that it is much nicer to use explicit dictionary passing and write something like this instead:

memberInsertProperty SetAPI{..} x set = (member x (insert x set) == True)

where `member` and `insert` come from the SetAPI record via RecordWildCards.

Finally, I'm not even sure how to create a type class covering Set and IntSet with the following two methods:

singleton :: a -> Set a map :: Ord b => (a -> b) -> Set a -> Set b

singleton :: Int -> IntSet map :: (Int -> Int) -> IntSet -> IntSet

Could anyone please enlighten me about the right way to abstract over this using type classes?

I tried a few approaches, for example:

class IsSet s where type Elem s singleton :: Elem s -> s map :: Ord (Elem t) => (Elem s -> Elem t) -> s -> t

Looks nice, but I can't define the IntSet instance:

instance IsSet IntSet where type Elem IntSet = Int singleton = IntSet.singleton map = IntSet.map

This fails with: Couldn't match type `t' with `IntSet' -- and indeed, how do I tell the compiler that in the IntSet case s ~ t in the map signature? Shall I add more associated types, or "associated constraints" using ConstraintKinds? I tried and failed, at various stages, repeatedly.

...And then you discover that there is Set.cartesianProduct :: Set a -> Set b -> Set (a, b), but no equivalent in IntSet and things get even more grim.

Cheers, Andrey

_______________________________________________ ghc-devs mailing list ghc-devs@haskell.org http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs