A question about run-time errors when class members are undefined

[Simon Peyton Jones]

Anthony

You may be interested in Carlos Camarao’s interesting work.  For a long time now he has advocated (in effect) making each function into its own type class, rather that grouping them into classes.   Perhaps that is in line with your thinking.
https://homepages.dcc.ufmg.br/~camarao/

Simon

From: Haskell-prime <haskell-prime-bounces at haskell.org> On Behalf Of Anthony Clayden
Sent: 06 October 2018 04:19
To: Petr Pudlák <petr.mvd at gmail.com>
Cc: haskell-prime at haskell.org
Subject: Re: A question about run-time errors when class members are undefined


On Sat, 6 Oct 2018 at 9:47 AM, Petr Pudlák <redirect at vodafone.co.nz<mailto:redirect at vodafone.co.nz>> wrote:

IIRC one of the arguments against having many separate classes is that a class is not a just set of methods, it's also the relations between them,

Hi Petr, I was talking about splitting out Haskell's current class hierarchy as a step towards doing away with classes altogether. If your language insists on methods being held in classes, that's just tedious bureacracy to invent class names.

The relations between classes (including between single-method classes) can be captured through superclass constraints. For example, in the Haskell 2010 report

class (Eq a, Show a) => Num a where ...

such as the important laws between `return` and `>>=`. And then for example a class with just `return` doesn't give any information what `return x` means or what should be its properties.

Then make Bind a superclass constraint on `return` (or vice versa, or both ways).

Just as the laws for Num's methods are defined in terms of equality

x + negate x == fromInteger 0          -- for example

Talking about laws is a red herring: you can't declare the laws/the compiler doesn't enforce them or rely on them in any way. Indeed the Lensaholics seem to take pleasure in building lenses that break the (van Laarhoven) laws.



That said, one of really painful points of Haskell is that refactoring a hierarchy of type-classes means breaking all the code that implements them. This was also one of the main reasons why reason making Applicative a superclass of Monad took so long. It'd be much nicer to design type-classes in such a way that an implementation doesn't have to really care about the exact hierarchy.

Yes that's what I was saying. Unfortunately for Haskell's Num class, I think it's just too hard. So a new language has an opportunity to avoid that. If OTOH Helium wants to slavishly follow Haskell, I'm wondering what is the point of Helium.

With Applicative, IIRC, refactoring had to wait until we got Constraint kinds and type families that could produce them. Would Helium want to put all that into a language aimed at beginners?


 For example, in Haskell we could have

class (Return m, Bind m) => Monad m where

without any methods specified. But instances of `Monad` should be only such types for which `return` and `>>=` satisfy the monad laws.

First: what does "satisfy the xxx laws" mean? The Haskell report and GHC's Prelude documentation state a bunch of laws; and it's a good discipline to write down laws if you're creating a class; but it's only documentation. Arguably IO, the most commonly used Monad, breaks the Monad laws in rather serious ways because it imposes sequence of execution; and it would be unfit for purpose if it were pure/lazy function application.

Then: what do you think a language could do to detect if some instance satisfies the laws? (Even supposing you could declare them.)


And this would distinguish them from types that have both `Return` and `Bind` instances, but don't satisfy the laws.

You could have distinct classes/distinct operators. Oh, but then `do` dotation would break.


Unfortunately I'm not sure if there is a good solution for achieving both these directions.

I don't think there's any solution for achieving "satisfy the xxx laws".


AntC


čt 4. 10. 2018 v 3:56 odesílatel Anthony Clayden <anthony_clayden at clear.net.nz<mailto:anthony_clayden at clear.net.nz>> napsal:
> We are adding classes and instances to Helium.

> We wondered about the aspect that it is allowed to have a class instance

> of which not all fields have a piece of code/value associated with them, ...



I have a suggestion for that. But first let me understand where you're going with Helium. Are you aiming to slavishly reproduce Haskell's classes/instances, or is this a chance for a rethink?



Will you want to include associated types and associated datatypes in the classes? Note those are just syntactic sugar for top-level type families and data families. It does aid readability to put them within the class.



I would certainly rethink the current grouping of methods into classes. Number purists have long wanted to split class Num into Additive vs Multiplicative. (Additive would be a superclass of Multiplicative.) For the Naturals perhaps we want Presburger arithmetic then Additive just contains (+), with `negate` certainly in a different class, perhaps (-) subtract also in a dedicated class. Also there's people wanting Monads with just `bind` not `return`. But restructuring the Prelude classes/methods is just too hard with all that legacy code. Even though you should be able to do:



class (Additive a, Subtractive a, Negative a, Multiplicative a, Divisive a) => Num a



Note there's a lot of classes with a single method, and that seems to be an increasing trend. Historically it wasn't so easy in Haskell to do that superclass constraints business; if it had been perhaps there would be more classes with a single method. Then there's some disadvantages to classes holding multiple methods:

* the need to provide an overloading for every method, even though it may not make sense

  (or suffer a run-time error, as you say)

* the inability to 'fine tune' methods for a specific datatype [**]

* an internal compiler/object code cost of passing a group of methods in a dictionary as tuple

  (as apposed to directly selecting a single method)



[**] Nats vs Integrals vs Fractionals for `Num`; and (this will be controversial, but ...) Some people want to/some languages do use (+) for concatenating Strings/lists. But the other methods in `Num` don't make any sense.



If all your classes have a single method, the class name would seem to be superfluous, and the class/instance decl syntax seems too verbose.



So here's a suggestion. I'll need to illustrate with some definite syntax, but there's nothing necessary about it. (I'll borrow the Explicit Type Application `@`.) To give an instance overloading for method `show` or (==)



show @Int = primShowInt                     -- in effect pattern matching on the type

(==) @Int = primEqInt                       -- so see showList below

That is: I'm giving an overloading for those methods on type `Int`. How do I declare those methods are overloadable? In their signature:



show @a :: a -> String                      -- compare show :: Show a => a -> String

(==) @a :: a -> a -> Bool

Non-overladable functions don't have `@a` to the left of `::`.

How do I show that a class has a superclass constraint? That is: a method has a supermethod constraint, we'll still use `=>`:



show @a :: showsPrec @a => a -> String      -- supermethod constraint

show @[a] :: show a => [a] -> String        -- instance decl, because not bare a, with constraint =>

show @[a] xss = showList xss

(*) @a :: (+) @a => a -> a -> a



Is this idea completely off the wall? Take a look at Wadler's original 1988 memo introducing what became type classes.

http://homepages.inf.ed.ac.uk/wadler/papers/class-letter/class-letter.txt<https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fhomepages.inf.ed.ac.uk%2Fwadler%2Fpapers%2Fclass-letter%2Fclass-letter.txt&data=02%7C01%7Csimonpj%40microsoft.com%7C5ec7c3a23a9746bb154b08d62b3a7ba2%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636743927613528630&sdata=vXgR1YLqw8ERzEbhVZBGr%2FpB5fMLMmYtnwt6Bpp4wGs%3D&reserved=0>



It reviews several possible designs, but not all those possibilities made it into his paper (with Stephen Blott) later in 1988/January 1989. In particular look at Section 1's 'Simple overloading'. It's what I'm suggesting above (modulo a bit of syntax). At the end of Section 1, Wadler rejects this design because of "potential blow-ups". But he should have pushed the idea a bit further. Perhaps he was scared to allow function/method names into type signatures? (I've already sneaked that in above with constraints.) These days Haskell is getting more relaxed about namespaces: the type `@`pplication exactly allows type names appearing in terms. So to counter his example, the programmer writes:



square x = x * x                             -- no explicit signature given

square :: (*) @a => a -> a                   -- signature inferred, because (*) is overloaded

rms = sqrt . square                          -- no explicit signature

rms :: sqrt @a => a -> a                     -- signature inferred



Note the inferred signature for `rms` doesn't need `(*) @a` even though it's inferred from `square`. Because (*) is a supermethod of `sqrt`. `sqrt` might also have other supermethods, that amount to `Floating`.



> ... a run-time error results.

>

> Does anyone know of a rationale for this choice, since it seems rather unhaskell-like.

If you allow default method implementations (in the class, as Cale points out), then I guess you have to allow instance decls that don't mention all the methods. I think there should at least be a warning if there's no default method. Also beware the default method might have a more specific signature, which means it can't be applied for some particular instance.

Altogether, I'd say, the culprit is the strong bias in early Haskell to bunch methods together into classes. These days with Haskell's richer/more fine-tuned typeclass features: what do typeclasses do that can't be done more precisely at method level -- indeed that would _better_ be done at method level?


AntC
_______________________________________________
Haskell-prime mailing list
Haskell-prime at haskell.org<mailto:Haskell-prime at haskell.org>
http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-prime<https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fmail.haskell.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fhaskell-prime&data=02%7C01%7Csimonpj%40microsoft.com%7C5ec7c3a23a9746bb154b08d62b3a7ba2%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636743927613528630&sdata=%2FVDtNLEzevef69Qs9rzrd5V2lBJzMLjRAxBBYvbsPAs%3D&reserved=0>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/haskell-prime/attachments/20181008/d6a2e733/attachment-0001.html>