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

static volatile int flag0 __attribute__((aligned(64)));
static volatile int flag1 __attribute__((aligned(64)));
static volatile int turn __attribute__((aligned(64)));

static void
lock(volatile int *my_flag, volatile int *other_flag, int other_id)
{
  *my_flag = 1;
  turn = other_id;
  while (*other_flag && turn == other_id) {
  }
  asm volatile ("" : : : "memory");
}

static void
unlock(volatile int *my_flag)
{
  asm volatile ("" : : : "memory");
  *my_flag = 0;
}

#define LOOP_COUNT 100000000
static int counter = 0;

static void*
task0(void *arg)
{
  int i;
  for (i = LOOP_COUNT; i != 0; i--) {
    lock(&flag0, &flag1, 1);
    counter++;
    unlock(&flag0);
  }
  return NULL;
}

static void*
task1(void *arg)
{
  int i;
  for (i = LOOP_COUNT; i != 0; i--) {
    lock(&flag1, &flag0, 0);
    counter++;
    unlock(&flag1);
  }
  return NULL;
}

int
main(int argc, char *argv[])
{
  flag0 = 0;
  flag1 = 0;

  pthread_t t0, t1;
  if (pthread_create(&t0, NULL, task0, NULL) ||
      pthread_create(&t1, NULL, task1, NULL)) {
    fprintf(stderr, "pthread_create\n");
    exit(1);
  }

  if (pthread_join(t0, NULL) ||
      pthread_join(t1, NULL)) {
    fprintf(stderr, "pthread_join\n");
    exit(1);
  }

  printf("counter = %d\n", counter);
  printf("difference = %d\n", LOOP_COUNT * 2 - counter);
  return 0;
}