cas.c

#include <stdio.h>
#include <pthread.h>

struct data {
  int c1;
  int padding[1024*1024*64];
  int c2;
} data;

int repeats = 1;
int count = 100000000;

void* t1(void* arg)
{
  int reps = repeats;
  int cnt = count;
  int i, j;
  for (j = 0; j < reps; ++j)
    for (i = 0; i < cnt; ++i)
#ifdef BARE
      data.c1++;
#endif
#ifdef FNA
  __sync_fetch_and_add(&data.c1, 1);
#endif
#ifdef ANF
  __sync_add_and_fetch(&data.c1, 1);
#endif
  fprintf(stderr, "%d\n", data.c1);
}

void* t2(void* arg)
{
  int reps = repeats;
  int cnt = count;
  int i, j;
  for (j = 0; j < reps; ++j)
    for (i = 0; i < cnt; ++i)
#ifdef BARE
      data.c2++;
#endif
#ifdef FNA
  __sync_fetch_and_add(&data.c2, 1);
#endif
#ifdef ANF
  __sync_add_and_fetch(&data.c2, 1);
#endif
  fprintf(stderr, "%d\n", data.c2);
}

int main(int argc, char * argv[])
{
  pthread_t th1, th2;
  void *rv;
  pthread_create(&th1, NULL, t1, NULL);
  pthread_create(&th2, NULL, t2, NULL);

  pthread_join(th1, &rv);
  pthread_join(th2, &rv);
  return 0;
}

console output

dvk@dvk-laptop ~/tmp $ gcc -O0 -DFNA cas.c -o cas -lpthread
dvk@dvk-laptop ~/tmp $ time ./cas
100000000
100000000

real	0m2.445s
user	0m4.755s
sys	0m0.001s
dvk@dvk-laptop ~/tmp $ gcc -O0 -DANF cas.c -o cas -lpthread
dvk@dvk-laptop ~/tmp $ time ./cas
100000000
100000000

real	0m2.448s
user	0m4.768s
sys	0m0.004s
dvk@dvk-laptop ~/tmp $ gcc -O0 -DBARE cas.c -o cas -lpthread
dvk@dvk-laptop ~/tmp $ time ./cas
100000000
100000000

real	0m0.432s
user	0m0.656s
sys	0m0.001s