7.8.1 plan

[Mateusz Kowalczyk]

On 25/03/14 16:18, kyra wrote:
> On 3/25/2014 19:11, Mateusz Kowalczyk wrote:
>> That commit is not in 2.14.1. Do you have any benchmarks to show the 
>> speedup? If the commit does some significant speedup, I'm not against 
>> backporting it into Haddock released with 7.8.2 (and there is at least 
>> 1 other fix I want to get into 7.8.2) but we pretty much closed the 
>> 2.14.1, it's up on Hackage and everything and that's what's planned to 
>> ship with 7.8.1. 
> 
> Hmm, now, when I'm trying to reproduce things I don't see haddock 
> producing any assembly output let alone split it when using 7.8rc2 haddock.
> 
> It seemed to me some time ago haddock became slow when processing a 
> package built with --enable-split-objs and I've decided to look into 
> things and discovered haddock wants .hi interface files and produces 
> assembly output to produce these interface files.
> 
> I was extremely surprised, rechecked things several times and saw the 
> same picture.
> 
> Now I can't reproduce this at all.
> 
> If haddock never produced .hi interface files and/or assembly output 
> then that was some mental aberration and the whole story can be dropped.
> 
> Regards,
> Kyra
> 
> 
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org
> http://www.haskell.org/mailman/listinfo/ghc-devs
> 

As far as I know (and this is mostly guessing), Haddock will ask GHC to
produce .hi files when they do not already exist. I think you might have
some luck if you simply try cabal configure && cabal haddock but I don't
know.

I took the patch because I saw no harm in it and all tests were passing
but I do admit that I did not check the performance gain (we don't have
benchmarks, perhaps we should hook some up) or loss. I think GHC has
some Haddock perf tests but I did not check the numbers and I don't
think it would even fall under that.

If you can come up with something then let me know. If you can't, I
think we'll keep the patch just in case unless someone can show it
breaks something.

The only instances of Haddock becoming really slow that I can think of
is some rather old ticket (#101 on Haddock Trac) in presence of Template
Haskell but I have closed it a while ago due to lack of information to
go on and inability to replicate without the reporter's help. Perhaps
this was your use case?

-- 
Mateusz K.