The SIGCHLD signal is a little funny and doesn’t work like you’d expect: we think we should get one signal per child death, but that’s not it.

Instead, it’s a kind of level-triggered thing where at some unknown intervals it sends the signal if there are any un-waited-for children.

In the loop you provided that burns through the wait(), this loop is consuming multiple children before the signal handler gets around to it, hence less trips through the handler.

Others have pointed out that you should be using a volatile sig_atomic_t variable, and though this is a good idea, it’s not why you’re seeing this behavior.

I believe the only way to get a guaranteed one-signal-per-child is to actually wait for the child in the signal handler – this makes it appear more like an edge-triggered signal.

#include <signal.h>
#include <unistd.h>
#include <stdio.h>
#include <sys/wait.h>
#include <stdlib.h>

static volatile sig_atomic_t cpt = 0;

static void handler(int sig) {
     cpt++;
     wait(NULL);  // ADD ME
}

int main() {
  int i;
  signal(SIGCHLD, handler);
  for (i = 0; i < 5; i++) {
       if (fork() == 0) {
           exit(0);
       }
   }
   while (wait(NULL) != -1) ;
   printf("cpt=%d\n", cpt);
   return 0;
}