It exploits how dictionary passing works in a pretty robust way that ghc is likely to at some point codify officially.
On Saturday, April 25, 2015, Ertugrul Söylemez <
ertesx@gmx.de> wrote:
> Could that specialization be accomplished today using eds reflection
> pkg? I guess not quite in terms of that pre apply pattern.
I'm not familiar with that package and couldn't find it on Hackage by a
quick search. But I believe that it can only be done with compiler
support, although with enough hackery you can probably get an ugly
version of it using TH.
> This is interesting. And it's a good example of a larger problem of
> not enough support for composable specialization with good sharing
> across the use sites that doesn't require egregious Inlining. At
> least for code that isn't Type class driven.
Indeed. Specialisation is a really good way to get very fast code
without making your executable size explode. I believe that support for
more fine-grained specialisation should and will improve. I'm not sure
how to make it more composable though.
>> At the definition and instantiation sites I mostly miss defaults. At
>> the application sites I would love to have specialisation for certain
>> arguments. For example I would like to be able to tell GHC that I
>> would like to have a version of my function `f` with a certain
>> argument inlined. Note that I don't want to inline `f` itself.
>> Rather I'd like to preapply certain arguments:
>>
>> f :: X -> Y -> Z
>>
>> {-# SPECIALISE f SomeX #-}
>> {-# SPECIALISE f SomeOtherX #-}
>>
>> This would generate two specialised versions of `f` with exactly the
>> given arguments inlined. That way I can get a very efficient `f`
>> without having to inline it at the application sites. And as long as
>> `f` is INLINABLE I can put those pragmas pretty much everywhere. I
>> believe this is exactly what happens for type class dictionaries.
>>
>> This can (and probably should) be a separate feature though. For some
>> of my applications I need to inline a huge chunk of code multiple times
>> to compensate for the lack of this feature.