This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [lno] 50% runtime performance regression since yesterday


On Wed, 23 Jun 2004, Sebastian Pop wrote:

> On Wed, Jun 23, 2004 at 03:45:45PM +0200, Sebastian Pop wrote:
> > On Wed, Jun 23, 2004 at 03:26:41PM +0200, Sebastian Pop wrote:
> > >
> > > Does the following patch fix the optimization regression?
> > >
> >
> > Not yet.
> >
>
> Or maybe yes, could you check please?

Didn't change anything.  Diffstat on a diff -cb from yesterdays .s to
todays .s with your patch says (uh, ia64 asm is ugly!)

 tramp3d-v3.s |457785 +++----!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
 1 files changed, 26717 insertions(+), 30096 deletions(-), 400972 modifications(!)

where yesterday:
-rw-r-----    1 rguenth  tat      11804627 Jun 23 17:07 ../tramp3d-v3.s
and now:
-rw-r-----    1 rguenth  tat      11902703 Jun 23 17:06 tramp3d-v3.s

Output of yesterday's binary is (tramp3d-v3 -n 10):
...
i = 2    t = 0.00679089  dt = 0.00271169 (2.67931s/it)
i = 3    t = 0.00937923  dt = 0.00258833 (2.67298s/it)
i = 4    t = 0.0120634   dt = 0.0026842 (2.67131s/it)
i = 5    t = 0.0150383   dt = 0.00297491 (2.6758s/it)
i = 6    t = 0.0183952   dt = 0.00335687 (2.67446s/it)
i = 7    t = 0.0217122   dt = 0.00331702 (2.67459s/it)
i = 8    t = 0.0250681   dt = 0.00335591 (2.67274s/it)
...

And now we have:
...
i = 2	 t = 0.00679089	 dt = 0.00271169 (4.24201s/it)
i = 3	 t = 0.00937923	 dt = 0.00258833 (4.22157s/it)
i = 4	 t = 0.0120634	 dt = 0.0026842 (4.22877s/it)
i = 5	 t = 0.0150383	 dt = 0.00297491 (4.23612s/it)
i = 6	 t = 0.0183952	 dt = 0.00335687 (4.23283s/it)
i = 7	 t = 0.0217122	 dt = 0.00331702 (4.21988s/it)
i = 8	 t = 0.0250681	 dt = 0.00335591 (4.23152s/it)
...


>
> There is no difference between the code generated by the compiler as
> of 2004-06-18 and the version after the patch.  The command line for
> compiling the testcase was: objdir/gcc/cc1plus -O2 -funroll-loops tramp3d-v3.cpp

Oh, I use -ffast-math, too - but that shouldn't change much.

>
> $ diff -dup ./20040618/tramp3d-v3.s ./20040623/tramp3d-v3.s
> --- ./20040618/tramp3d-v3.s     2004-06-23 10:30:24.000000000 -0400
> +++ ./20040623/tramp3d-v3.s     2004-06-23 09:53:11.000000000 -0400
> @@ -1028,5 +1028,5 @@ _ZTI11CreateEvent:
>  _ZTS11CreateEvent:
>         .string "11CreateEvent"
>  .globl memmove
> -       .ident  "GCC: (GNU) 3.5-tree-ssa-lno 20040618 (merged 20040612)"
> +       .ident  "GCC: (GNU) 3.5-tree-ssa-lno 20040623 (merged 20040612)"
>         .section        .note.GNU-stack,"",@progbits
>

--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]