Version 13 (modified by 6 years ago) (diff) | ,
---|

# Default superclass instances

A matter of much consternation, here is a proposal to allow type class declarations to include default instance declarations for their superclasses. It's based on Jón Fairbairn's proposal, but it has a more explicit 'off switch' and the policy on corner-cases is rejection. Credit is due also to the class system extension proposal and its ancestors, in particular, John Meacham's class alias proposal.

We may distinguish two uses of superclasses (not necessarily exclusive).

- A class can
*widen*its superclass, extending its interface with new functionality (e.g., adding an inverse to a monoid to obtain a group -- inversion seldom provides an implementation of composition). The subclass methods provide extra functionality, but do not induce a standard implementation of the superclass. Num provides arithmetic operations, widening Eq, but does not induce an implementation of`(==)`

. - A class can
*deepen*its superclass, providing methods at least as expressive, admitting a default implementation of superclass functionality (e.g., the`compare`

function from an Ord instance can be used to give an Eq instance;`traverse`

from Traversable f can be instantiated to`foldMapDefault`

for Foldable f and`fmapDefault`

for Functor f).

(**SLPJ** I don't understand this distinction clearly.)

This proposal concerns the latter phenomenon, which is currently such a nuisance that Functor and Applicative are not superclasses of Monad. Nobody wants to be forced to write Functor and Applicative instances, just to access the Monad interface. Moreover, any proposal to refine the library by splitting a type class into depth-layers is (rightly!) greeted with howls of protest as an absence of superclass instances gives rise to breakage of the existing codebase.

Default superclass instances are implemented in the Strathclyde Haskell Enhancement. They should enable some tidying of the library, with relatively few tears. Moreover, they should allow us to deepen type class hierarchies as we learn. Retaining backward compatibility in relative silence is the motivation for an opt-in default.

## The proposal

Concretely, the proposal is as follows.

### Default superclass instances

First, we allow a class declaration to include a **default superclass instance declaration** for some, none, or all of its superclass constraints. We say that superclasses with default implementations are **intrinsic** superclasses. Example:

class Functor f => Applicative f where return :: x -> f x (<*>) :: f (s -> t) -> f s -> f t (>>) :: f s -> f t -> f t fs >> ft = return (flip const) <*> fs <*> ft instance Functor f where fmap = (<*>) . return

Note the `instance`

declaration nested inside the `class`

declaration. This is the default superclass instance declaration, and `Functor`

thereby becomes an intrinsic superclass of `Applicative`

. Moreover, note that the definition of `fmap`

uses the `<*>`

operation of `Applicative`

; that is the whole point!

Here is another example:

class Applicative f => Monad f where (>>=) :: f a -> (a -> f b) -> f b instance Applicative f where ff <*> fs = ff >>= \ f -> fs >>= \ s -> return (f s)

Here, `Applicative`

is an intrinsic superclass of `Monad`

.

### Instance declarations

A default superclass instance in a class declaration for class C has an effect on the instance declarations for C.

Specifically:

- An instance declaration
instance Q => C ty where ...defs...

for class C generates an extra instance declarationinstance Q => Si ty where ....

for each intrinsic superclass Si of C

- The method definitions in
`...defs...`

are distributed to the appropriate instance declaration, according to which class the method belongs to.

- Any methods that are not specified explicitly are "filled in"
from the default definition given in the default superclass instance.
(If there is no default definition, then a warning is produced,
and a definition that calls
`error`

is used instead.)

For example, assume the class declaration for `Monad`

above. Then
this instance declaration:

instance Monad m where (>>=) = ...blah... (<*) = ...bleh...

would generate an extra instance declaration for the instrinsic superclass `Applicative`

,
with the methods distributed appropriately:

instance Monad m where (>>=) = ...blah... instance Applicative m where (<*) = ...bleh... -- Programmer specified ff <*> fs = ff >>= \ f -> fs >>= \ s -> return (f s) -- From the default superclass instance

We call these extra instance declarations an **intrinsic instance declaration**.
(The term "derived instance" is already taken!)

This process is recursive. Since `Functor`

is an intrinsic superclass of `Applicative`

,
the intrinsic instance for `Applicative`

recursively
generates an intrinsic instance for `Functor`

:

instance Functor m where fmap = (<*>) . return -- From default superclass instance

## The opt-out mechanism

Just because you can make default instances, they are not always the instances you want. A key example is

instance Monad m => Monad (ReaderT r m) where ...

which would give us by default the intrinsic instance

instance Monad m => Applicative (ReaderT r m) where ...

thus preventing us adding the more general

instance Applicative m => Applicative (ReaderT r m) where ...

To inhibit the generation of an intrinsic instance declaration, one can use a
`hiding`

clause in the instance declaration:

instance Sub x where ... hiding instance Super

which acts to prevent the generation of instances for Super and all of Super's intrinsic superclasses in turn. For example: write

instance Monad m => Monad (ReaderT r m) where ... return x = ... ba >>= bf = ... hiding instance Applicative

The `hiding`

clause applies to all the intrinsic instances generated
from an instance declaration. For example, we might write

instance Monad T where return x = ... ba >>= bf = ... hiding instance Functor

Note that `Functor`

is only an indirect intrinsic superclass of `Monad`

, via `Applicative`

.
So the above instance would generate an intrinsic instance for `Applicative`

but not for `Functor`

.

Jón's proposal had a more subtle opt-out policy, namely that an intrinsic superclass can be quietly pre-empted by an instance for the superclass from a prior or the present module. Note that to declare an instance of the subclass, one must produce an instance of the superclass by the same module at the latest.

This quiet exclusion
policy is not enough to handle the case of multiple candidate
intrinsic instances arising from multiple intrinsic superclasses (e.g.,
`Traversable`

and `Monad`

giving competing `Functor`

instances), so some
explicit form is required. The question for us, then, is what should
happen if an intrinsic superclass not explicitly hidden were to clash
with an explicit instance from the same or a prior module. We could

- Reject this as a duplicate instance declaration, which indeed it is.
- Allow the explicit to supersede the intrinsic default, but issue a warning suggesting to either remove the explicit instance or add an explicit opt-out, or
- Allow the explicit to supersede the intrinsic default silently.

As it stands, we propose option 1 as somehow the principled thing to do. We acknowledge that it creates an issue with legacy code, precisely because there are plenty of places where we have written the full stack of instances, often just doing the obvious default thing: these should be cleaned up, sooner or later.

Option 3 avoids that
problem but risks perplexity: if I make use of some cool package which
introduces some `Foo :: * -> *`

, I might notice that `Foo`

is
a monad and add a `Monad Foo`

instance in my own code, expecting
the `Applicative Foo`

instance to be generated in concert; to my
horror, I find my code has subtle bugs because the package introduced
a different, non-monadic, `Applicative Foo`

instance which I'm
accidentally using instead. Option 2 is certainly worth considering as
a pragmatic transitional compromise, although the 'transitional' has a
dangerous tendency to be permanent.

## Multi-headed instance declarations

While we're about it, to allow multi-headed instance declarations for class-disjoint conjunctions, with the same semantics for constraint duplication and method distribution as for the defaults, so

instance S => (C x, C' x) where methodOfC = ... methodOfC' = ...

is short for

instance S => C x where methodOfC = ... instance S => C' x where methodOfC' = ...

This proposal fits handily with the kind Fact proposal, which allows multiple constraints to be abbreviated by ordinary type synonyms. So we might write

type Stringy x = (Read x, Show s) instance Stringy Int where read = ... show = ...

The common factor is that one instance declaration is expanded into several with the method definitions distributed appropriately among them.

## Details

Each default superclass instance declaration in a `class`

declaration must be for
a distinct class. So one of these is OK and the other is not:

-- This is ILLEGAL class (Tweedle dum, Tweedle dee) => Rum dum dee where instance Tweedle dum where ... instance Tweedle dee where ... -- But this is OK class (Tweedle dum, Tweedle dee) => Rum dum dee where instance Tweedle dee where ...

By requiring that intrinsic superclasses be class-distinct, we ensure that the distribution of methods to spawned instances is unambiguous.