[PROPOSAL] Adding Generics-based DefaultSignature to `deepseq` package

Thu Oct 16 15:03:46 UTC 2014

The important problem, as Edward Kmett would put it, is that Haskell is not
good at dealing with lots of abstractions. In particular, making the
typeclass hierarchy too fine-grained makes it painful to work with, because
programmers have to satisfy the tower of superclass constraints in order to
write an instance for a class. DefaultSignatures addresses this in a very
limited way: If

A a => B a => C a => D a

then I may be able to give A, B, and C methods defaults with signatures so
that I can declare an instance of D without needing to declare all the
superclass instances. Unfortunately, this breaks down as soon as things
branch:

A a => B a => C a => D a

||
V

E a => F a => G a

Both E and B may offer perfectly reasonable default definitions of a method
in A, but I can only choose *one* of them. It also fails when class A is in
someone else's module, and I'm doing a ton of work with subclasses of B and
would like very much to add a default definition of a method in A, but
simply can't. The current common use of DefaultSignatures is to use it
*only* to provide defaults for Generic instances. While this single
use-case works reasonably well, it effectively privileges Generic over
everything else and leaves the general problem unsolved.

The sort of general solution I'd hope for would probably look something
vaguely like this, but I imagine the type gurus might see problems:

Allow a *subclass* of a class to define (and override) default methods for
the superclass. There is, of course, an immediate challenge: a single type
could be a member of two subclasses, each of which defines a default for
the same superclass method. The best solution I can think of to this is to
require that such incoherent defaults be resolved manually by giving an
explicit superclass instance declaration; ideally, that declaration would
be able to access and choose from one of the available defaults, but that
might be more trouble than it's worth.

On Thu, Oct 16, 2014 at 10:39 AM, José Pedro Magalhães <dreixel at gmail.com>
wrote:

> I'd like to know exactly what is the important problem, and how
> DefaultSignatures are insufficiently general. Perhaps we can improve them,
> or come up with something better!
>
> On Thu, Oct 16, 2014 at 2:36 PM, David Feuer <david.feuer at gmail.com>
> wrote:
>
>> I'm generally opposed to DefaultSignatures as an upside-down,
>> insufficiently-general attempt to solve an important problem, and generally
>> think the less relies on them the better.
>> On Oct 16, 2014 6:40 AM, "Herbert Valerio Riedel" <hvr at gnu.org> wrote:
>>
>>>
>>> The Proposal
>>> ============
>>>
>>> I hereby propose to merge `deepseq-generics`[2] into `deepseq`[1] in
>>> order to add Generics support to the `NFData` class based on the
>>> `-XDeriveGenerics` and `-XDefaultSignature` language extensions.
>>>
>>> A concrete patch is available for bike-review at [3]
>>>
>>>
>>> Prior Proposal & What's changed
>>> ===============================
>>>
>>> About 2 years ago, I already proposed something similar[4].  Back then
>>> the major concern was avoiding a conditionally exported API as using the
>>> (back then) rather young `Generics` extension would leave the Haskell98
>>> domain.
>>>
>>> This lead to me release Generics support as a companion package[2] which
>>> turns out to have become a rather popular package (judging from the
>>> Hackage download-count stats).
>>>
>>> I only realized after the discussion was effectively finished, that
>>> having a separate `deepseq-generics` actually does have an IMO
>>> non-neglectable downside:
>>>
>>>   You can't support a `DefaultSignature`-based default implementation,
>>> as those need to be backed into the `NFData` class.
>>>
>>> Missing out on `DefaultSignature` would be a shame IMO, because
>>>
>>>  * There's a chance that starting with GHC 7.10 `deriving` may work for
>>>    arbitrary classes[5], putting `NFData` on equal footing as built-in
>>>    classes such as `Eq` or `Show`. Specifically, you would be able to
>>>    write
>>>
>>>       data Foo = Foo [Int] String (Bool,Char) | Bar (Maybe Char)
>>>                  deriving (Show, Generic, NFData)
>>>
>>>    instead of having to manually write the following boilerplate
>>>
>>>       instance NFData Foo where
>>>          rnf (Foo x y z) = rnf x `seq` rnf y `seq` rnf z
>>>          rnf (Bar x)     = rnf x
>>>
>>>    which gets tedious rather soon if you have many (and more complex)
>>>    types and tend to refactor regularly (with a risk of failing to adapt
>>>    your manual instances if you change the strictness of fields)
>>>
>>>
>>>  * The current default `rnf` implementation, i.e.
>>>
>>>      rnf a = a `seq` ()
>>>
>>>    is rather error-prone, as it's *very* easy to end up with an
>>>    incorrect instance. Especially after refactoring a type for which the
>>>    NF=WHNF assumption was broken after refactoring by adding new fields,
>>>    or changing the strictness of existing fields.
>>>
>>>    The Generics-derived `rnf` implementation does not have such a
>>>    problem.
>>>
>>>
>>> Moreover, popular packages are starting adopt (and even recommend) the
>>> use of Generics in combination with `DefaultSignature` to provide
>>> automatically derived default instances, most notably `hashable`[6],
>>> `binary`[7], or `aeson`[8] just to name a few. In addition to providing
>>> a precedence for the use of Generics, I consider those packages evidence
>>> for Generics to have proven itself to the point of replacing
>>> TemplateHaskell in these use-cases.
>>>
>>>
>>> Compatibility & Breakage Considerations
>>> =======================================
>>>
>>>  * This change requires a major version bump to deepseq-1.4.0
>>>
>>>  * `deepseq` needs to drop GHC 7.0.* support as GHC 7.2 is the first
>>>    version to support Generics & `DefaultSignature`.
>>>
>>>  * Code relying on the current `rnf` default-implementation will most
>>>    likely break (unless a `Generics` instance happens to be in-place)
>>>
>>>    However, it's easy to provide forward/backward-compatibility w/o any
>>>    CPP, by simply explicitly defining
>>>
>>>      instance NFData XYZ where rnf = seq x ()
>>>
>>>
>>>
>>> Discussion Period: 2 weeks
>>>
>>>
>>>
>>>  [1]: http://hackage.haskell.org/package/deepseq
>>>  [2]: http://hackage.haskell.org/package/deepseq-generics
>>>  [3]: https://github.com/haskell/deepseq/pull/1
>>>  [4]: http://thread.gmane.org/gmane.comp.lang.haskell.libraries/17940
>>>  [5]: https://ghc.haskell.org/trac/ghc/ticket/5462
>>>  [6]: http://hackage.haskell.org/package/hashable
>>>  [7]: http://hackage.haskell.org/package/binary
>>>  [8]: http://hackage.haskell.org/package/aeson
>>> _______________________________________________
>>> Libraries mailing list
>>> Libraries at haskell.org
>>> http://www.haskell.org/mailman/listinfo/libraries
>>>
>>
>> _______________________________________________
>> Libraries mailing list
>> Libraries at haskell.org
>> http://www.haskell.org/mailman/listinfo/libraries
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/libraries/attachments/20141016/e843dbd3/attachment-0001.html>