[Bug fortran/53950] New: 1.5 times slowdown from 4.4.0 to 4.7.0

[Bug fortran/53950] New: 1.5 times slowdown from 4.4.0 to 4.7.0

[kapojko at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

             Bug #: 53950
           Summary: 1.5 times slowdown from 4.4.0 to 4.7.0
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: kapojko@yahoo.com


Created attachment 27784
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27784
Demonstration of slowdown

I'm using MinGW GCC 4.4.0 to work with C/C++/Fortran project. Some days ago
I've downloaded web installer and switched to GCC 4.7.0. Unfortunately I
noticed a dramatic performance regression.

I attach a minimal program that demonstrates the slowdown to the ticket.

Compiled with "gfortran -o test_self.exe -O3 test_self.f90" I get on my Core
i7:
  GCC 4.4.0:     Time of operation was    1.2012070      seconds
  GCC 4.7.0:     Time of operation was    1.85641098      seconds


Yury Kapoyko, Russia

[Bug fortran/53950] 1.5 times slowdown from 4.4.0 to 4.7.0

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |*-mingw
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2012-07-13
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-07-13 10:47:58 UTC ---
Are you building a 64bit or a 32bit executable?  Are you sure this is not
an effect of whoever provides random_number() or log()?

[Bug fortran/53950] 1.5 times slowdown from 4.4.0 to 4.7.0

[kapojko at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

--- Comment #2 from Yury Kapoyko <kapojko at yahoo dot com> 2012-07-13 10:52:27 UTC ---
Both targets are 32 bit.

Yes, after deeper profiling it seems a problem to be inside log/cos/random. For
example, fragments from gprof:
  4.4.0:  cosf - self seconds 0.77
  4.7.0:  cosf - self seconds 0.94
and so on.

Is there a way to fix it?

[Bug fortran/53950] 1.5 times slowdown from 4.4.0 to 4.7.0

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|*-mingw                     |i?86-mingw

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-07-13 11:00:46 UTC ---
These functions are not part of GCC but eventually part of the mingw runtime,
so I suggest to file a bug there.

[Bug fortran/53950] 1.5 times slowdown from 4.4.0 to 4.7.0

[kapojko at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

--- Comment #4 from Yury Kapoyko <kapojko at yahoo dot com> 2012-07-13 12:39:05 UTC ---
I agree with you. This buf refers to mingw, I think, it could be closed.

Unfortunately, real application slows down to 30 times! But it's very hard to
"extract" bug from there because it is rather complicated and multi-threaded.
Also I have no profilers for Windows except gprof which is useless here.
Compiled on Fedora (GCC 4.6.3) application works fine.

I assume that the problem is with libgfortran, because if I run old application
(compiled with GCC 4.4.0), but with new gfortran.dll in PATH it works exactly
as bad, as new one.

[Bug fortran/53950] [4.7/4.8/4.9 Regression] 1.5 times slowdown from 4.4.0 to 4.7.0 with -m32

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

Dominique d'Humieres <dominiq at lps dot ens.fr> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
            Summary|1.5 times slowdown from     |[4.7/4.8/4.9 Regression]
                   |4.4.0 to 4.7.0              |1.5 times slowdown from
                   |                            |4.4.0 to 4.7.0 with -m32

--- Comment #5 from Dominique d'Humieres <dominiq at lps dot ens.fr> ---
On a Core2Duo 2.5Ghz under x86_64-apple-darwin10, I get

            -m64           -m32
4.9.0    1.9743720      2.2397862
4.8.1    2.0542710      2.2539010
4.7.3    1.9538360      2.2104399
4.6.4    1.9587610      2.1613581
4.5.4    2.0137279      2.3153970
4.4.7    1.9963471      1.9965631

Note that with 4.8 and trunk, I had to change slightly the code otherwise the
loop is optimized out.

[Bug fortran/53950] [4.7/4.8/4.9 Regression] 1.5 times slowdown from 4.4.0 to 4.7.0 with -m32

[janus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53950

janus at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
                 CC|                            |janus at gcc dot gnu.org
         Resolution|---                         |INVALID

--- Comment #6 from janus at gcc dot gnu.org ---
(In reply to Yury Kapoyko from comment #4)
> I agree with you. This buf refers to mingw, I think, it could be closed.

Doing so.