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

volatile uint64_t glob1;
volatile uint64_t glob2;
volatile int stop;

#define HIST_SIZE 100
void *
ThreadWriter(void *x)
{
  uint32_t i = 0;
  uint32_t delay_dist[HIST_SIZE] = {0};
  uint32_t h, l, tsc1, tsc2;
  volatile uint32_t delay = 0;

  __asm__ __volatile__("rdtsc" : "=a" (l), "=d" (h));
  tsc1 = ((uint64_t)l) | (((uint64_t)h) << 32);

  while(!stop)
  {
    glob1 = i;

    while(glob2 < glob1){};

    __asm__ __volatile__("rdtsc" : "=a" (l), "=d" (h));
    tsc2 = ((uint64_t)l) | (((uint64_t)h) << 32);
    delay = (tsc2 - tsc1)/10;
    tsc1 = tsc2;

    if (delay < HIST_SIZE)
      delay_dist[delay]++;

    i++;
  }

  printf("Ping-pong distribution by amount of cycles:\n");
  for(int d = 0; d < HIST_SIZE; d++)
    printf("%3d: %d\n", 10*d, delay_dist[d]);

  return x;
}

void *
ThreadReader(void *x)
{
  volatile uint32_t my_glob1 = glob1;

  while(!stop)
  {
    if (my_glob1 != glob1)
    {
      glob2 = glob1;
      my_glob1 = glob1;
    }
  }

  return x;
}

void
inthandler(int sig)
{
  stop = 1;
  glob2++;
}

#define THR 2

int main()
{
  pthread_t t[THR];
  int i;
  struct sigaction act;

  act.sa_handler = inthandler;
  sigaction(SIGINT, &act, NULL);

  pthread_create(&(t[0]), NULL, ThreadWriter, NULL);
  pthread_create(&(t[1]), NULL, ThreadReader, NULL);

  for (i = 0; i < THR; i++)
    pthread_join(t[i], NULL);

  return 0;
}