vectorisation code?

Geoffrey Mainland mainland at apeiron.net
Fri Jan 22 14:57:59 UTC 2016


On Fri, Jan 22, 2016 at 03:23:56PM +0100, Thomas Miedema wrote:
> On Fri, Jan 22, 2016 at 3:17 PM, Geoffrey Mainland <mainland at apeiron.net> wrote:
>> On 1/22/16 8:05 AM, Ben Gamari wrote:
>>> Manuel M T Chakravarty <chak at cse.unsw.edu.au> writes:
>>>> The way I see it, the main cost of keeping DPH around is to handle
>>>> breakages such as that with vector. I can't promise to address those
>>>> in a timely manner, which is why I agreed to disable/remove DPH.
>>>> However, as Geoff stepped forward, this issue is solved. As for the
>>>> overhead in compile time etc, I don't think, it is that much of a
>>>> deal. During development, most compiles runs are incremental anyway.
>>>
>>> Judging by the VCS history it seems that nothing happened in response to
>>> this thread. Geoff, do you see yourself having time to pick this up in
>>> the near future? If not, perhaps we should pick up this matter again and seriously
>>> consider parking this code in a branch until someone is able to pick it
>>> up again.
>>> Cheers,
>>> - Ben
>>
>> Yes, I am willing to do the work to get DPH back into the build in the
>> near future. However, that only makes sense if we are willing to build
>> DPH regularly. Also, I can't be solely responsible for all breakage
>> resulting from DPH; DPH has regularly exposed bugs in the past, which is
>> one reason to get it back into the regular build, but I can't promise to
>> fix all problems that might be exposed by DPH in the future :)
>> 
>> If I put a patch on Phab that updates DPH, are we willing to make DPH
>> part of the regular validation script again?
>> 
>> Cheers,
>> Geoff
>
> We could make all of hackage be part of the ghc build, and it would turn
> up bugs. But we don't do that either. Why is dph special?

Manuel and Simon can say more, but DPH has in the past been very good at
exposing, for example, regressions in the inliner. It exercises GHC in a
way that few other packages do.

DPH is intimately tied to GHC, so it's not something that can be
maintained separately as a package. If we aren't willing to make DPH
part of the regular build, then it will just bitrot again quickly, and
there's little point in doing the work to get it running again.

I'm of the opinion that DPH still has value and that it would be a shame
to lose it forever, which is effectively what will happen if we relegate
the vectorizer to a branch. I am willing to get DPH working again, but
only if there is general agreement that DPH is worth having---and that
we are willing to once again make it part of the regular build.

Geoff


More information about the ghc-devs mailing list