Question

I wrote a program and I’m quite amused by the fact that I’d run it dozens of times — I’ve even wrote down results from multiple executions –, and now it does not work.

You might think I’m high as a kite to say so, or that I might have simply changed some lines, but I really do not recall doing any alterations whatsoever on the program.

The problem is a SIGFPE, that raises up in different executions of the program, according to the input. Happens though the signal is raised while inserting a value in an std::unordered_set<Point<T> *>.

Here is a snippet of the code where I do such insertion:

std::vector<Point<T> *> _point, _centroid;
std::vector<std::unordered_set<Point<T> *> > _cluster;

// Main procedure -- pseudocode
for (point in _point) {
    cluster_id = centroid_with_min_distance(point, _centroid);
    has_changed = _change_cluster(point, cluster_id);
}

// Changes from one "point->_cluster's unordered_set" to "c's unordered_set"
bool _change_cluster(Point<T> *point, const unsigned int& c) {

    if ((point->_cluster == c) &&
        (_cluster[c].find(point) != _cluster[c].end())) {
        return false;
    }

    _cluster[point->_cluster].erase(point);
    _cluster[c].insert(point); // Insertion that raises the SIGFPE exception
    point->_cluster = c;
    return true;

}

Here is a considered-important section of valgrind’s output:

==17636== Invalid read of size 8
==17636==    at 0x40A758: std::pair<std::__detail::_Hashtable_iterator<
                Point<unsigned int>*, true, false>, bool>
                std::_Hashtable<Point<unsigned int>*, Point<unsigned int>*,
                std::allocator<Point<unsigned int>*>,
                std::_Identity<Point<unsigned int>*>,
                std::equal_to<Point<unsigned int>*>,
                std::hash<Point<unsigned int>*>,
                std::__detail::_Mod_range_hashing,
                std::__detail::_Default_ranged_hash,
                std::__detail::_Prime_rehash_policy, false, true, true>::
                    _M_insert<Point<unsigned int>* const&>(
                        Point<unsigned int>* const&&&,
                        std::integral_constant<bool, true>) (hashtable.h:966)
==17636==    by 0x408EDA: std::_Hashtable<Point<unsigned int>*,
                 Point<unsigned int>*, std::allocator<Point<unsigned int>*>,
                 std::_Identity<Point<unsigned int>*>,
                 std::equal_to<Point<unsigned int>*>,
                 std::hash<Point<unsigned int>*>,
                 std::__detail::_Mod_range_hashing,
                 std::__detail::_Default_ranged_hash,
                 std::__detail::_Prime_rehash_policy, false, true, true>::
                     insert(Point<unsigned int>* const&) (hashtable.h:400)
==17636==  ... (calls from my program) ...
==17636==  Address 0x620a028 is not stack'd, malloc'd or (recently) free'd
==17636== 
==17636== 
==17636== Process terminating with default action of signal 8 (SIGFPE)
==17636==  Integer divide by zero at address 0x402D07D7C
==17636==    at 0x40252F: std::__detail::_Mod_range_hashing::operator()(
                 unsigned long, unsigned long) const (hashtable_policy.h:376)
==17636==    by 0x40A66C: std::__detail::_Hash_code_base<Point<unsigned int>*,
                 Point<unsigned int>*, std::_Identity<Point<unsigned int>*>,
                 std::equal_to<Point<unsigned int>*>,
                 std::hash<Point<unsigned int>*>,
                 std::__detail::_Mod_range_hashing,
                 std::__detail::_Default_ranged_hash, false>::_M_bucket_index(
                     Point<unsigned int>* const&, unsigned long,
                     unsigned long) const (hashtable_policy.h:758)
==17636==    by 0x40A772: std::pair<std::__detail::_Hashtable_iterator<
                 Point<unsigned int>*, true, false>, bool>
                 std::_Hashtable<Point<unsigned int>*, Point<unsigned int>*,
                 std::allocator<Point<unsigned int>*>,
                 std::_Identity<Point<unsigned int>*>,
                 std::equal_to<Point<unsigned int>*>,
                 std::hash<Point<unsigned int>*>,
                 std::__detail::_Mod_range_hashing,
                 std::__detail::_Default_ranged_hash,
                 std::__detail::_Prime_rehash_policy, false, true, true>::
                     _M_insert<Point<unsigned int>* const&>(
                     Point<unsigned int>* const&&&,
                     std::integral_constant<bool, true>) (hashtable.h:966)
==17636==    by 0x408EDA: std::_Hashtable<Point<unsigned int>*,
                 Point<unsigned int>*, std::allocator<Point<unsigned int>*>,
                 std::_Identity<Point<unsigned int>*>,
                 std::equal_to<Point<unsigned int>*>,
                 std::hash<Point<unsigned int>*>,
                 std::__detail::_Mod_range_hashing,
                 std::__detail::_Default_ranged_hash,
                 std::__detail::_Prime_rehash_policy, false, true, true>::
                     insert(Point<unsigned int>* const&) (hashtable.h:400)
==17636==    ... (calls from my program) ...

The question here is: as I have calculations in my program that may result in divisions by zero — they are not directly related to this process, though –, is it possible that the bug is being shadowed by the insertion? Or should I do some extra treatment while inserting pointers in an std::unordered_set<T>?

I’m compiling the program under x86_64 GNU/Linux, and I’m using g++ (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3.

Answer 1

I suspect that ‘c’ is out of the range of _cluster. Can you replace all of the _cluster[c] with _cluster.at(c), and see if it finds an out of range error.

The valgrind output indicates that it’s an invalid read followed by an integer modulus error. Since it’s doing hash_code % bucket_count, I suspect that you accessed out of the range of the vector.

As an aside, naming things that start with _ can be risky, as if it’s in the global namespace, it is technically reserved for the implementation.