Build time regressions

Simon Peyton Jones simonpj at microsoft.com
Wed Oct 1 08:57:19 UTC 2014


It sounds as if there are two issues here:


·         Should GHC unpack a !’d constructor argument if the constructor’s argument has a lot of fields?  It probably isn’t profitable to unbox very large products, because it doesn’t save much allocation, and might cause extra allocation at pattern-match sites.  So I think the answer is yes.  I’ll open a ticket.



·         Is some library (binary? blaze?) creating far too much code in some circumstances?  I have no idea about this, but it sounds fishy.  Simply creating the large worker function should not make things go bad.



Incidentally, John, using {-# NOUNPACK #-} !Bar would prevent the unpacking while still allowing the field to be strict.  It’s manually controllable.


Simon



From: ghc-devs [mailto:ghc-devs-bounces at haskell.org] On Behalf Of John Lato
Sent: 01 October 2014 00:45
To: Edward Z. Yang
Cc: Joachim Breitner; ghc-devs at haskell.org
Subject: Re: Build time regressions

Hi Edward,

This is possibly unrelated, but the setup seems almost identical to a very similar problem we had in some code, i.e. very long compile times (6+ minutes for 1 module) and excessive memory usage when compiling generic serialization instances for some data structures.

In our case, I also thought that INLINE functions were the cause of the problem, but it turns out they were not.  We had a nested data structure, e.g.

> data Foo { fooBar :: !Bar, ... }

with Bar very large (~150 records).

even when we explicitly NOINLINE'd the function that serialized Bar, GHC still created a very large helper function of the form:

> serialize_foo :: Int# -> Int#  -> ...

where the arguments were the unboxed fields of the Bar structure, along with the other fields within Foo.  It appears that even though the serialization function was NOINLINE'd, it simply created a Builder, and while combining the Builder's ghc saw the full structure.  Our serializer uses blaze, but perhaps Binary's builder is similar enough the same thing could happen.

Anyway, in our case the fix was to simply remove the bang pattern from the 'fooBar' record field.  Then the serialize_foo function takes a Bar as an argument and serializes that.  I'm not entirely sure why compilation takes so much longer otherwise.  I've tried dumping the output of each simplifier phase and it clearly gets stuck at a certain point, but I didn't really debug in much detail so I don't recall the details.

If you think this is related, I can investigate more thoroughly.

Cheers,
John L.

On Wed, Oct 1, 2014 at 4:54 AM, Edward Z. Yang <ezyang at mit.edu<mailto:ezyang at mit.edu>> wrote:
Hello Joachim,

This was halfway known, but it sounds like we haven't solved
it completely.

The beginning of the sordid tale was when Cabal HEAD switched
to using derived binary instances:
https://ghc.haskell.org/trac/ghc/ticket/9583

SPJ fixed the infinite loop bug in the simplifier, but apparently
the deriving binary generates a lot of code, meaning a lot of
memory. https://ghc.haskell.org/trac/ghc/ticket/9630
hvr's fix was specifically to solve this problem.

But it sounds like it didn't eliminate the regression entirely?
If there's an unrelated regression, we should suss it out.  It would
be helpful if someone could revert just the deriving changes,
and see if this reverts the compilation time.

Edward

Excerpts from Joachim Breitner's message of 2014-09-30 13:36:27 -0700:
> Hi,
>
> the attached graph shows a noticable increase in build time caused by
>
> Update Cabal submodule & ghc-pkg to use new module re-export types
> author    Edward Z. Yang <ezyang at cs.stanford.edu<mailto:ezyang at cs.stanford.edu>>
> https://git.haskell.org/ghc.git/commit/4b648be19c75e6c6a8e6f9f93fa12c7a4176f0ae
>
> and only halfway mitigated by
>
> Update `binary` submodule in an attempt to address #9630
> author    Herbert Valerio Riedel <hvr at gnu.org<mailto:hvr at gnu.org>>
> https://git.haskell.org/ghc.git/commit/3ecca02516af5de803e4ff667c8c969c5bffb35f
>
>
> I am not sure if the improvement is related to the regression, but in
> any case: Edward, was such an increase expected by you? If not, can you
> explain it? Can it be avoided?
>
> Or maybe Cabal just became much larger... +38% in allocations when
> running haddock on it seems to confirm this.
>
> Greetings,
> Joachim
>
_______________________________________________
ghc-devs mailing list
ghc-devs at haskell.org<mailto:ghc-devs at haskell.org>
http://www.haskell.org/mailman/listinfo/ghc-devs

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.haskell.org/pipermail/ghc-devs/attachments/20141001/a12cd0fb/attachment-0001.html>


More information about the ghc-devs mailing list