obscure problem with sigsetjmp on linux with C

Question

For some reason when I return from the signal handler for SIGFPE, the variable i is one less than I expect it to be, to get the correct results from my program (iterate through all of the pairs in the pairs array) I must check the return value of sigsetjmp and increment i if I am returning from a signal "catch". Why is this? Why is the incrementation that happens during the iteration of the floating point exception lost?

#include <stdio.h>
#include <signal.h>
#include <stdlib.h>
#include <setjmp.h>                                                               

jmp_buf e;

int i;
void float_exception ();
int main ()
{
    int pairs[][2] = {
        {10, -5}, {10, -3}, {-10, -3}, {-10, -5}, {-10, 3}, {-10, 5}, {-10, 0},
        {10, 0}, {0, 10}, {2, 3}, {3, 3}, {5, 10}
    };
    int npairs;
    /* handle sigfpe so /0 doesn't interrupt the rest of the program */
    signal (SIGFPE, float_exception);
    printf ("Seeing what my C implementation does with negative or 0 modulo"
            "\narithmetic.\n");
    npairs =  sizeof (pairs) / sizeof (int) / 2;
    i = 0;

    if (sigsetjmp (e, 1) != 0) {
        i++;    /* without this line, i is one less than I expect it to be */
    }
    for (; i < npairs - 1; i++) {
        printf ("%d:   %d %% %d\t= ", i, pairs[i][0], pairs[i][1]);
        fflush (stdout);
        printf ("%d\n", pairs[i][0] %pairs[i][1]);
        fflush (stdout);
    }
    return 0;
}



void float_exception ()
{
    printf ("fpe\n");
    fflush (stdout);
    longjmp (e, 1);
}

Answer 1

That looks right to me. The "normal" i++ happens at the end of the for() loop - and if you call longjmp(e,1) you skip that.

Answer 2

caf's answer seems to be right.

What I actually expected initially, was, that i needs to be declared volatile. At least that's what the author of NSBlog claims.

One other thing to note: it might be dangerous to call printf and fflush inside a signal handler. Given the fact that SIGFPE is a synchronous signal, it's not completely out of the question, but you should know what you are doing.

And finally, from the Linux man-page: "longjmp() and siglongjmp() make programs hard to understand and maintain. If possible an alternative should be used."