I have a function which computes a sum and depending on the result of the sum does a certain task or gives a warning message. I need to run this function millions of times in a particle trajectory simulation (it calculates the time and position dependent force) and noticed that my code was very slow.
This is my MWE:
#include <stdio.h>
#include <math.h>
int foo()
{
double sum =0;
double dummy_sum = 0;
for (size_t i=0; i<40; i++)
{
sum+=1e-2;
dummy_sum += 1e-2;
}
if (sum>5.) // normally this should not happen
{
printf("warning message!\n");
return(-1);
}
else if(sum >0)
{
// normal behavior
}
return(0);
}
int main()
{
int fooint;
for(size_t i=0; i<5e9; i++)
{
fooint = foo();
if(fooint!=0)
{
return(fooint);
}
}
return 0;
}
I compile with gcc -Ofast -std=c99 test.c -o test.exe using gcc version 4.8.5.
In order to find a way to optimize my function I ran time ./test.exe which indicated a run time of about 38s on my machine.
When I removed the lines
printf("warning message!\n");
return(-1);
or the line
sum+=1e-2;
the run time was reduced to about 6s.
Also when I changed the number of iterations in the function loop to 35 the speed is increased and the run time is 6s. Increasing it to 36 gives a run time of 33s.
I compiled and ran the code on a different machine with the same linux and gcc versions and got similar results (run times were slightly longer as the machine is older).
What may cause this strange behavior?
It's a GCC bug. Clang optimises your code properly to
foo: # @foo
xor eax, eax
ret
main: # @main
xor eax, eax
ret
at near-zero run time.