[Haskell-cafe] coercing D a to D b

[Evan Laforge]

Nice, it looks like "Note [Case 2: CSEing case binders]" applies to
what I was talking about.

I assume once I can upgrade to 8.2 (or 8.4 more likely) then both my
examples should wind up looking different at the core level, but then
both be a coerce in STG?  Well, I guess implicitly a coerce since STG
doesn't have types like that.


On Thu, Dec 14, 2017 at 7:16 AM, Joachim Breitner
<mail at joachim-breitner.de> wrote:
> Hi,
>
> Am Mittwoch, den 13.12.2017, 21:34 -0800 schrieb Evan Laforge:
>> Given:
>>
>> data D a = A a | B Int Char
>>
>> dmapNoCoerce :: (a -> b) -> D a -> D b
>> dmapNoCoerce f (A a) = A (f a)
>> dmapNoCoerce _ (B i c) = B i c
>>
>> I have to reconstruct a B change it from D a to D b.  But at the lower level,
>> couldn't this be implemented as a type cast?  What prevents such an
>> optimization?
>>
>> I can write this as
>>
>> dmapCoerce :: (a -> b) -> D a -> D b
>> dmapCoerce f (A a) = A (f a)
>> dmapCoerce _ b@(B {}) = Unsafe.Coerce.unsafeCoerce b
>>
>> From the core, it looks like dmapCoerce indeed has a cast with no allocation,
>> while dmapNoCoerce allocates a new B.
>>
>> It seems to work, but is it safe?  Is there a more principled way to do it?
>> I can't convince Data.Coerce to cooperate, presumably because 'a' and 'b' are
>> not coercible themselves, and it doesn't believe me if I try to tell it the
>> type role is phantom.
>
>
> this coercion is not possible in Core, because Core’s type system is
> not expressive enough (at least now…). But luckily, at the STG level
> this optimization is possible, see
> https://git.haskell.org/ghc.git/commitdiff/19d5c7312bf0ad9ae764168132aecf3696d5410b
>
> This is probably in 8.2.
>
> Joachim
>
> --
> Joachim Breitner
>   mail at joachim-breitner.de
>   http://www.joachim-breitner.de/
>
> _______________________________________________
> Haskell-Cafe mailing list
> To (un)subscribe, modify options or view archives go to:
> http://mail.haskell.org/cgi-bin/mailman/listinfo/haskell-cafe
> Only members subscribed via the mailman list are allowed to post.