Divide by zero error from Colt OpenLongObjectHashMap

Question

I saw this exception from Colt OpenLongObjectHashMap:

java.lang.ArithmeticException: divide by zero
        at cern.colt.map.OpenLongObjectHashMap.indexOfKey(Unknown Source)
        at cern.colt.map.OpenLongObjectHashMap.get(Unknown Source)

It's not reproduceable.

This is indexOfKey:

protected int indexOfKey(long key) {
    final long tab[] = table;
    final byte stat[] = state;
    final int length = tab.length;

    final int hash = HashFunctions.hash(key) & 0x7FFFFFFF;
    int i = hash % length;
    int decrement = hash % (length-2); // double hashing, see http://www.eece.unm.edu/faculty/heileman/hash/node4.html
    //int decrement = (hash / length) % length;
    if (decrement == 0) decrement = 1;

    // stop if we find a free slot, or if we find the key itself.
    // do skip over removed slots (yes, open addressing is like that...)
    while (stat[i] != FREE && (stat[i] == REMOVED || tab[i] != key)) {
        i -= decrement;
        //hashCollisions++;
        if (i<0) i+=length;
    }

    if (stat[i] == FREE) return -1; // not found
    return i; //found, return index where key is contained
}

So the only divisors used are length and (length - 2), where length is table.length, table being an internal array.

However, table is only ever initialised to an array of minimum size 3 (and the default is 277 which is what I am using). Integer wrap around doesn't seem possible either.

So this would seem to be an impossible error.

Any ideas?

Answer 1

This turned out to be a Java compiler optimization error in the IBM JDK JIT compiler that was in use.

See this bug report: IJ06000: UNEXPECTED DIVIDE BY ZERO EXCEPTION

The recommended fix is to disable the LoopVersioner optimisation on the problem methods.