6.5 Examples

The following program will be used to demonstrate the effects of different optimization levels:

#include <stdio.h>

double
powern (double d, unsigned n)
{
  double x = 1.0;
  unsigned j;

  for (j = 1; j <= n; j++)
    x *= d;

  return x;
}

int
main (void)
{
  double sum = 0.0;
  unsigned i;
  
  for (i = 1; i <= 100000000; i++)
    {
      sum += powern (i, i % 5);
    }

  printf ("sum = %g\n", sum);
  return 0;
}

The main program contains a loop calling the powern function. This function computes the n-th power of a floating point number by repeated multiplication--it has been chosen because it is suitable for both inlining and loop-unrolling. The run-time of the program can be measured using the time command in the GNU Bash shell.

Here are some results for the program above, compiled on a 566MHz Intel Celeron with 16KB L1-cache and 128KB L2-cache, using GCC 3.3.1 on a GNU/Linux system:

$ gcc -Wall -O0 test.c -lm
$ time ./a.out 
real    0m13.388s
user    0m13.370s
sys     0m0.010s

$ gcc -Wall -O1 test.c -lm
$ time ./a.out
real    0m10.030s
user    0m10.030s
sys     0m0.000s

$ gcc -Wall -O2 test.c -lm
$ time ./a.out
real    0m8.388s
user    0m8.380s
sys     0m0.000s

$ gcc -Wall -O3 test.c -lm
$ time ./a.out
real    0m6.742s
user    0m6.730s
sys     0m0.000s

$ gcc -Wall -O3 -funroll-loops test.c -lm
$ time ./a.out
real    0m5.412s
user    0m5.390s
sys     0m0.000s

The relevant entry in the output for comparing the speed of the resulting executables is the 'user' time, which gives the actual CPU time spent running the process. The other rows, 'real' and 'sys', record the total real time for the process to run (including times where other processes were using the CPU) and the time spent waiting for operating system calls. Although only one run is shown for each case above, the benchmarks were executed several times to confirm the results.

From the results it can be seen in this case that increasing the optimization level with -O1, -O2 and -O3 produces an increasing speedup, relative to the unoptimized code compiled with -O0. The additional option -funroll-loops produces a further speedup. The speed of the program is more than doubled overall, when going from unoptimized code to the highest level of optimization.

Note that for a small program such as this there can be considerable variation between systems and compiler versions. For example, on a Mobile 2.0GHz Intel Pentium 4M system the trend of the results using the same version of GCC is similar except that the performance with -O2 is slightly worse than with -O1. This illustrates an important point: optimizations may not necessarily make a program faster in every case.