[Bug c++/59845] New: loop optimization problem when profiling

[Bug c++/59845] New: loop optimization problem when profiling

[arshamidi at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59845

            Bug ID: 59845
           Summary: loop optimization problem when profiling
           Product: gcc
           Version: 4.8.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: arshamidi at gmail dot com

I'm using g++ with GCC 4.8.1; when I use fprofile-generate too optimize my
code, it actually slows down the process time: (constant folding problem within
loop?)

int main(){

    int n;
    int i;
    //int n2;
    printf("Enter a number: ");
    scanf("%d", &n);
    //n2 = n/2;
    printf("divisors are:\n\n");

    for (i=1; i <= /*n2*/ n/2; i++)
        if (n%i == 0)
            printf("\n%d", i);

    printf("\n\n end.\n");

    return 0;
}

Here I must use the constant 'n2' (commented out) too actually get the desired
speed -in combination with -fprofile-generate. 
I'm using the following switches:
-Ofast;-fomit-frame-pointer;-march=core2; -mfpmath=sse; -fprofile-generate;
-Wall

no problem without -fprofile-generate and -fprofile-use

[Bug c++/59845] loop optimization problem when profiling

[arshamidi at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59845

--- Comment #1 from ArshaGCC <arshamidi at gmail dot com> ---
I found that only when I use very high vales for input the problem persists.
for example when using 123456 as input for profiling it generates incorrect
code but when using 12345 as input the correct code will be generated.

Input is an integer and I can specify very long values for input like:
123456789.

[Bug c++/59845] loop optimization problem when profiling

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59845

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
>(constant folding problem within loop?)

No it is doing the following:

if (n>=1)
{
  i = 1;
  while (true)
    {
      if (n%i == 0)
        {
          printf("\n%d", i);
          i++;
          if (n/2 < i) break;
        }
      else
        {
          i++;
          if (n/2 < i) break;
        }
    }

}


Notice the load from n is not pulled out of the loop and that is because the
compiler thinks n escapes via scanf.

[Bug c++/59845] loop optimization problem when profiling

[arshamidi at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59845

--- Comment #3 from ArshaGCC <arshamidi at gmail dot com> ---
I further analyzed and found an interesting issue:

The optimizer bug (profile optimizer) occurs when input > 32768. When input is
less than or equal to 32768 (2^15) it always generate the correct optimized
code.

Could it be an optimization bug found in profile optimizer?

[Bug c++/59845] loop optimization problem when profiling

[arshamidi at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59845

ArshaGCC <arshamidi at gmail dot com> changed:

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

--- Comment #4 from ArshaGCC <arshamidi at gmail dot com> ---
With the new GCC 4.9 (GCC 4.9.2 currently i'm using), the problem seems to be
resolved.

Thanks.