Hardly any discussion of type class hierarchy goes without the common gripe that Applicative ought to be the superclass of Monad. However, restructuring this hierarchy would break a great deal of code, particularly if we removed redundant methods such as "pure".
This Haskell extension proposal provides a simple solution for refactoring type-class hierarchies while maintaining backwards compatibility. However, it can also do more than this, allowing boilerplate declarations to be generated in terms of some common pattern of usage. Functions in Haskell enjoy the reasoning and abstraction benefits of referential transparency. This extension gives us this same power, but for instance declarations, allowing for the abstraction of common instantiation patterns.
This repository contains some other documentation about this extension, as well as a work-in-progress TH prototype of the feature. The file "Rewriter.hs" uses the haskell-src-exts package to rewrite Haskell code to allow regular classes and instances to use this framework.
This proposal shares some goals with Default Superclass Instances, but achieves these goals in a fashion that's easier to understand and implement, of course making a different set of trade-offs. We sacrifice the ability to restructure hierarchies without touching the actual code, in exchange for being capable of much more.
deriving class PreOrder a where -- A list of parameters to the instance template. These get substituted into -- the instances below. They're not exported - and this symbol will only -- exist externally if one of the instances re-exports it. (<=) :: a -> a -> Bool -- The instances generated by the instance template. instance Eq a where x == y = (x <= y) && (y <= x) instance Ord a where (<=) = (<=)
(<=) nested inside
instance Ord is not ambiguous with the outer one,
because the template parameters shadow the methods defined within.
In order to use this
deriving class PreOrder declaration, we write code that
looks like this:
instance PreOrder Bool where False <= _ = True True <= x = x
The methods defined in the instance are used as the parameters to the instance template, expanding into:
instance Eq a where x == y = (x <= y) && (y <= x) where False <= _ = True True <= x = x instance Ord a where (<=) = (<=) where False <= _ = True True <= x = x
This is the trivial, definitional desugaring - the compiler could certainly do something more clever.
It'd be nice to still have the property that instance heads can be used as constraints in polymorphic types. As a result, it's reasonable for the above definition to implicitly create the following constraint-kind synonym:
type PreOrder a = (Eq a, Ord a)
On the topic of these more recent type-system enhancements, while I have not
yet implemented it in the TH prototype, it seems like having type and data
family declarations will be relatively straightforward. This is true of both
the head of the
deriving class (which resemble class syntax), and the
generated instances (which use instance syntax).
This is because the parameters are directly substituted into the generated instance. So, a type family instance declaration will be substituted into all usages of that type family in the generated instances. It seems like data family declarations could be more problematic because it seems like their usage should usually be linear, as it makes less sense to define a data-type multiple times (the names of the constructors would conflict).
We can now make instance declarations generate more class instances than before, which leads to some things to consider:
The generated instances might conflict with those already defined in the module. This is a good primary mechanism for "opt-ing out" of a particular generated instance. Since this is adding some "spooky action at distance", it's reasonable to generate WARNINGs in order to notify the user that something unexpected might happen.
"Hiding" declarations would be a good way to opt out. These hiding declarations could refer to constraint synonyms (which might have been generated by other instance templates), in order to succinctly suppress a bunch of the generated instances.
This particular design decision has been much discussed much within the context of Superclass Default Instances. Within that proposal, this is called "Option 2", and seems to be the most popular / the consensus.
The instances a particular module exports is now dependent on the way the classes are defined in its imports. This is not as much of an issue as it sounds - an instance of
Monadwill still mean we have an instance of
Monad- be it a normal or compound class constraint.
However, this is an issue when it comes to orphan instances, as identical library + client code could have an instance clash, given a different set of definitions for the library's dependencies. I think that this is acceptable, as orphan instances are known to be dangerous, and many instance templates would not have this sort of behavior.
The generated instances might conflict with those produced by other template invocations. This is something that we'd quite reasonably want to do, while re-working entire hierarchies. The rest of this section is devoted to this issue.
For example, we could write an instance template for the old version of
Applicative, that generated the new versions of
Applicative. We could then also write an instance template for
Monadthat generated all three.
The problem with this is that code using the current hierarchy would quite often reasonably have instances for all three, so two versions of
Applicativewould be generated. So, the question becomes how we can tell which definition supersedes, without hackily defining an arbitrary priority order.
A reasonable way to do this is by allowing an instance template that generates a subset of another to have priority. "Subset" here means that all of the instances generated by one template overlap with some instance in another template (causing those to be suppressed).
It's simple. We're just supplying values to a generic instance, to create a specific one, and these parameters are referentially transparent.
More powerful instance derivation allows us to mitigate the impact of historical decisions.
Being able to rework, say, the Numeric class hierarchy, is the main goal of this proposal (and those that came before): http://hackage.haskell.org/trac/ghc/wiki/DefaultSuperclassInstances
As mentioned in that page, default superclass instances have been a "matter of consternation" for some time, as no approach to the problem has been satisfying enough to be implemented. By forcing the decision of how to implement a class to be per-datatype, we avoid attempting to define typeclass instances which "always" can be implemented universally in terms of some other.
In order to create configurable instances (see the "weak typing" section) with superclass instances, we need to create a new typeclass. This is odd, because it doesn't make much sense to use it in any other context than the superclass of the instance. I think it's preferable to be able to "hide" this class so that it can't be depended on.
Avoidance of Template Haskell.
The TH implementation of this library generates TH code. In other words, it's making it convenient to write code that was already quite possible to write before. However, the error messages and potential for analysis by tools are impaired. By making it a language feature, we can conquer a good deal of the macro-expansion design space for typeclasses.
People have observed many things that are wrong with TH / mis-aligned with Haskell philosophy. I find it to be an extremely useful and invaluable tool despite, but many of these points are valid.
Compared to the power and complexity of TH, this feature is very simple, and we can rely on the reasoning power of referential transparency.
No potential solution to the "default superclass instances" is without its trade-offs. I think that the described solution is a straightforward, understandable solution to the problem, that buys a lot of power, with comparatively minimal issues:
We cannot create instances that are automagically derived in the presence of some set of constraints. This was the intent of the problematic "superclass- defaults" proposals, and is a huge can of worms. I think that wanting this feature is a little over-reaching, if you want to maintain Haskell's current typeclass semantics.
Ideally we'd be able to seamlessly use old code with our new typeclasses, in concordance with "Design goal 1" mentioned in the "Other Proposals" section. However, this means that this feature would need to only have a LANGUAGE extension option (e.g. -XInstanceTemplates) for the modules defining instance templates.
Are we comfortable changing the meaning of code without additional pragma, in the event that the dependencies specify this pragma? Is there any precedent for language extensions doing this?
[Niklas Broberg's take on this, in the context of superclass default instances] (http://firstname.lastname@example.org/msg20351.html)
If ConstraintKinds is used as the mechanism that allows for naming the set of derived instances, then this would also mean that -XConstraintKinds would need to be implicit in users of instance templates. This is not so disastrous, though - as this could be restricted to usage of constraint kinds, and not actual declaration of them / usage of "Constraint".
One bit of ugliness is that now we can't move an instance declaration that is overlapped by an instance template into another file, without having a declaration of the form
hiding instance ...that suppresses the generated instance.
One alternative is to have instance templates be a part of the exported signature, and then later do whole-program instantiation of these templates. This is the same sort of resolution that would need to take place for superclass instance defaults to work out while preserving this mobility-of-instances property.
The difference in behavior found when moving an instance out of a module can actually be beneficial - we can use an instance template which overlaps with those defined in the module, even if its definition doesn't know about them! All of the generated instances that overlap with the ones that already exist will be suppressed.
There's a bit more to say about this idea!
[A few examples of how it's useful] (https://github.com/mgsloan/instance-templates/blob/master/Examples.md)
[We gain additional ability to represent API differences] (https://github.com/mgsloan/instance-templates/blob/master/Deltas.md)
[Some more ideas and notes] (https://github.com/mgsloan/instance-templates/blob/master/Extras.md)
The following people (and a few others!) all gave me invaluable encouragement and feedback in the process of working on this proposal and prototype:
- Edward Kmett
- Luite Stegeman
- Drew Day
- William Cauchois