ghc-7.8-rc2 in -O2 mode eats all stack and RAM on pandoc-citeproc / highlighting-kate

[Sergei Trofimovich]

On Sat, 22 Mar 2014 22:21:42 +0300
Sergei Trofimovich <slyich at gmail.com> wrote:

> Hello!
> 
> I have noticed the problem in ghc-7.6.3 first
> when tried to build all haskell userland with -O2 opt level.
> 
> It led to amazing bugs!
> 
> Here is one of those (highlighting-kate hackage package):
> > [281 of 452] Compiling Text.Highlighting.Kate.Syntax.Asp ( highlighting-kate-0.5.6.1/Text/Highlighting/Kate/Syntax/Asp.hs, highlighting-kate-0.5.6.1/Text/Highlighting/Kate/Syntax/Asp.o )
> > stack overflow: use +RTS -K<size> to increase it
> 
> How to reproduce it:
> 1. Download a bundled file (6.6MB):
>     http://code.haskell.org/~slyfox/selfcontained-eater-ghc-7.8-rc2.tar.gz
> 2. Unpack and run there:
>   ./mk.sh
> 
> The script is designed to plug any built ghc version w/o external depends.
> 
> Command will fail as:
>     $ ./mk.sh
>     ...
>     [281 of 452] Compiling Text.Highlighting.Kate.Syntax.Asp ( highlighting-kate-0.5.6.1/Text/Highlighting/Kate/Syntax/Asp.hs, highlighting-kate-0.5.6.1/Text/Highlighting/Kate/Syntax/Asp.o )
>     stack overflow: use +RTS -K<size> to increase it
> 
> On ghc-7.6.3 it will progress a bit more: down to 452 file and will crash there
> similar way.
> 
> I've 'cabal unpack'-ed all sources and configured/de-.hsc-ed them
> manually/added needed -DWhatever / added {-# LANGUAGE CPP #-}
> around. Nothing else.
> 
> It's very hard to shrink such large thing manually down to 2-3 files.
> Would be cool if ghc (and cabal) would be able to spit something
> self-sufficient (like 'gcc -i' does) for devs to reproduce.
> 
> Adding '-v' shows such log:
>     ...
>     *** Simplifier:
>     Result size of Simplifier iteration=1
>       = {terms: 21,973, types: 21,838, coercions: 1,842}
>     Result size of Simplifier iteration=2
>       = {terms: 21,952, types: 21,819, coercions: 1,842}
>     Result size of Simplifier
>       = {terms: 21,950, types: 21,817, coercions: 1,842}
>     *** SpecConstr:
>     Result size of SpecConstr***<CRASH>

Nobody interested? Is it too scary?

Such inliner blowups are hard to shrink down from real
examples down to toy ones. I could try to but I need
a bit of guidance.

Maybe you need only an intermediate core step right
before an OOM, whatever?

Thanks!

-- 

  Sergei
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://www.haskell.org/pipermail/ghc-devs/attachments/20140403/3e77b8ac/attachment.sig>