------- Additional Comments From dberlin at gcc dot gnu dot org 2005-07-15
00:41 -------
Subject: Re: [4.1 Regression] corrupted
profile info with -O3 -fprofile-use
On Fri, 2005-07-15 at 00:07 +0000, janis at gcc dot gnu dot org wrote:
> ------- Additional Comments From janis at gcc dot gnu dot org 2005-07-15
> 00:07 -------
> I've got diverging scripts on various systems, and when I thought I was
> testing
> with a compiler from before 20050630 I was actually using today's mainline.
>
> I have verified (and triple-checked) that none of the failures I currently see
> in SPEC CPU2000 when using profile-generate/profile-use show up just before
> the
> patch at http://gcc.gnu.org/ml/gcc-cvs/2005-06/msg01108.html. Many of them
> show
> up immediately after that patch, and others creep in later. Among the
> failures
> immediately after the patch are
>
> gap with profopts -O3 -m64
> gcc with profopts -O3 -m32/-m64
> twolf with profopts -O3 -m32/-m64
> vpr with profopts -O3 -m32/-m64
>
Okay, well.
I can eliminate 1/3rd of the patch's actual changes if you try the above
with -fno-tree-salias added.
Then, if that isn't better, can you also try the following patch and see
if it either triggers an assert or fixes the bug?
I never noticed before, but there are places we are looping over the
referenced variables and *adding things to the array at the same time*.
If the hashtable resizes at just the wrong point, i imagine this could
cause some bad things to happen.
> No minimized testcase yet.
>
------- Additional Comments From dberlin at gcc dot gnu dot org 2005-07-15
00:41 -------
Created an attachment (id=9277)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=9277&action=view)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22471
