Data.Foldable causes missed foldr/build opportunities

[Gabriel Gonzalez]

On 07/25/2013 10:41 AM, Edward Kmett wrote:
> One of the open concerns about it is definitely ensuring that we get 
> the fusion opportunities we can.
>
> If you put an INLINE pragma on your Foldable version of each do the 
> fusion rules fire after it gets inlined into a call site that uses it 
> as a list?
>

I tried both INLINE and INLINABLE and neither causes the fusion rules to 
fire.  I also tried:

* adding an orphan SPECIALIZE rule for `Data.Foldable.mapM_` in the 
module where I defined `each`

* Specializing the type of `each` to consume lists, but still using the 
`Foldable` `mapM_`

* Defining a new copy of `each` (using the `Foldable` version) in the 
same module as the code that uses it, specializing the type signature to 
lists, and trying out INLINE/INLINABLE or no pragma.

None of those causes the rule to fire, either.

> -Edward
>
> On Thu, Jul 25, 2013 at 1:22 PM, Gabriel Gonzalez 
> <gabriel439 at gmail.com <mailto:gabriel439 at gmail.com>> wrote:
>
>     I'm now in favor of the `Data.Foldable` proposal, but I just
>     wanted to mention that the proposal needs to include some extra
>     pragma work to ensure that build/foldr optimizations fire.  I was
>     just experimenting with the following combinator for `pipes`
>     trying out the following two versions:
>
>         each :: (Monad m) => [a] -> Producer a m ()
>         each = mapM yield
>
>         each :: (Monad m, Foldable f) => f a -> Producer a m ()
>         each = Data.Foldable.mapM yield
>
>     When I do a pure `pipes`-based fold over both `Producers`s, the
>     version specialized to lists triggers a firing of the build/foldr
>     fusion rule and runs about 20% faster.  The true improvement for
>     `mapM` by itself is probably even greater than that because I
>     haven't optimized the folding code yet.  The latter version does
>     not trigger the rule firing.  Either way I'm going to include the
>     latter `Foldable` version but I just wanted to mention this
>     because I remember people were asking if this would impact fusion
>     or not.
>
>     _______________________________________________
>     Libraries mailing list
>     Libraries at haskell.org <mailto: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/20130725/13602cec/attachment.htm>