[Bug tree-optimization/34683] [4.3 Regression] Fortran FE generated IL pessimizes middle-end IL and analysis

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #37 from rguenther at suse dot de  2008-01-09 09:45 -------
Subject: Re:  [4.3 Regression] Fortran FE
 generated IL pessimizes middle-end IL and analysis

On Wed, 9 Jan 2008, jaydub66 at gmail dot com wrote:

> ------- Comment #36 from jaydub66 at gmail dot com  2008-01-09 09:38 -------
> 
> > How does the trunk compare now to the numbers mentioned in comment #16 ?
> 
> Compiling with rev. 131414 gives:
> 
> -O1 -fstrict-aliasing: 33sec, 438197 kB
> -O2:                   97sec, 669637 kB
> -O3:                   50sec, 392669 kB
> 
> This is of course better than a few days ago (thanks to Richard's patches). But
> it's not even close to the numbers in comment #16.
> 
> What worries me about this is not the 33sec compile time (I could live with
> that, if I'd have to), but the really large amount of memory. On a machine with
> only 256MB compiling takes several minutes, due to massive swapping.
> 
> Richard:
> You said that the remaining issue is probably not gonna be fixed for 4.3. I
> don't quite see the reason for this, since this apparently *was* fixed at some
> point in the 4.3 trunk (see comment 16).

Yeah, but we won't trade wrong-code bugs for better compile-time here.
I have no idea how practical it is to fix basic-block duplication code
(which is the worst offender regarding compile-time in this circumstances
of too many VOPs).

> Also this would really be a shame since it would make GCC 4.3 practically
> unusable for our project (http://gibuu.physik.uni-giessen.de/GiBUU/).

There is a workaround available - you can use
--param max-fields-for-field-sensitive=0
to disable SFTs, but this may disable useful optimization.  This is
the route we want to go with 4.4, addressing the missed optimizations in
different ways.

Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34683