[Bug c/49148] New: Certain expressions take an extremely long time for no apparent reason

[Bug c/49148] New: Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

           Summary: Certain expressions take an extremely long time for no
                    apparent reason
           Product: gcc
           Version: 4.4.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: jchthys@yahoo.com


Created attachment 24349
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24349
This is a minimal test case of an expression that takes 100 times longer to
evaluate than it should. See comments for explanations.

I have been doing work on 1-D Jacobi stencils. Basically, this involves a lot
of repeated operations. For some reason, however, I have been getting extremely
long run times for certain expressions. What is weird is that if I TAKE OUT a
multiply operation, the function takes about 100 times longer to run than if I
left it in.

The line that causes the problems is marked in the source code. It reads:

     new[i] = (A[i-1] + A[i] + A[i+1]) / 4.0;

If I change this to the following, there is no problem:

     new[i] = (A[i-1] + 2*A[i] + A[i+1]) / 4.0;

As another example, the following line runs slowly:

     new[i] = (A[i-1] + A[i] + A[i+1]) * 0.3;

while this line runs quickly:

     new[i] = (A[i-1] + 2*A[i] + A[i+1]) * 0.333;

The reason I think this is a bug is that it does not happen with older versions
of GCC (for example, 4.1.2), nor with the Intel C Compiler.

[Bug c/49148] Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

John Peyton <jchthys at yahoo dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization

--- Comment #1 from John Peyton <jchthys at yahoo dot com> 2011-05-24 18:56:52 UTC ---
The following command was used to compile and execute the test case:

     gcc gcc_bug.c -O3 -o gcc_bug && ./gcc_bug

Note that -O3 was used. In addition, -lm affects performance minimally if it is
used.

[Bug c/49148] Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

--- Comment #2 from John Peyton <jchthys at yahoo dot com> 2011-05-24 19:16:20 UTC ---
Created attachment 24350
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24350
This is a minimal example of an expression that takes longer to evaluate than
it should. Run the slightly longer file for an example of the speed difference.

[Bug c/49148] Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

--- Comment #3 from John Peyton <jchthys at yahoo dot com> 2011-05-24 19:16:51 UTC ---
Created attachment 24351
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24351
Finally, this is a preprocessed version of the minimal case.

[Bug c/49148] Certain expressions take an extremely long time for no apparent reason

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> 2011-05-24 20:27:39 UTC ---
Maybe you are getting subnormals which causes a huge performance decrease.

[Bug rtl-optimization/49148] Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

John Peyton <jchthys at yahoo dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |minor

[Bug rtl-optimization/49148] Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

--- Comment #5 from John Peyton <jchthys at yahoo dot com> 2011-05-24 21:26:49 UTC ---
> Maybe you are getting subnormals which causes a huge performance decrease.

It looks like that might indeed be the case. It didn't seem to happen with ICC
though. Is it possible that other compilers have different ways of handling
subnormals?

[Bug rtl-optimization/49148] Certain expressions take an extremely long time for no apparent reason

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

--- Comment #6 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-05-25 09:26:56 UTC ---
Try -ffast-math (the default for ICC) which disables subnormals.

[Bug rtl-optimization/49148] Certain expressions take an extremely long time for no apparent reason

[jchthys at yahoo dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

--- Comment #7 from John Peyton <jchthys at yahoo dot com> 2011-05-25 14:40:56 UTC ---
> Try -ffast-math (the default for ICC) which disables subnormals.

Thanks, this helped it run much faster!

I noticed it still takes longer to divide by 3.0 than by 4.0. More
specifically, the following line was fast:

     new[i] = (A[i-1] + 2*A[i] + A[i+1]) / 4.0;

The following line took about 4 times as long to run without -ffast-math:

     new[i] = (A[i-1] + A[i] + A[i+1]) / 3.0;

With -ffast-math, it was faster; however, it was still took about 2 times as
long as the division by 4.0 version. In addition, changing division by 3.0 to
multiplication by 0.333 or even 0.3 did not affect the speed.

Could this also be related to subnormal numbers, or is this just a different
issue?

[Bug rtl-optimization/49148] Certain expressions take an extremely long time for no apparent reason

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=49148

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |INVALID
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #8 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to John Peyton from comment #7)
> > Try -ffast-math (the default for ICC) which disables subnormals.
> 
> Thanks, this helped it run much faster!

Yes so you confirm this is a subnormal issue.

> 
> Could this also be related to subnormal numbers, or is this just a different
> issue?

Different issue related to /4.0 can be done as *0.25 but I think that might be
solved in newer versions of GCC; I didn't try.