I'm writing a little collection of sorting algorithms and some benchmarks (using Google benchmark). I wrote the following heap sort but I'm struggling in understanding why it is so slower than std::sort_heap implemented in the STL C++ Stardard Library. The code is inspired to the algorithm described in Cormen book "Introduction to algorithms"

Am I doing something silly?
I have also profiled the code using perf and it seems that the hottest instruction is the comparison function.

Here it goes my implementation:

template <typename Iterator>
inline Iterator left_child(const Iterator begin, Iterator idx) {
  return idx + std::distance(begin, idx) + 1;  

template <class Iterator>
inline Iterator right_child(const Iterator begin, Iterator idx) {
  return left_child(begin, idx) + 1;

template <class Iterator>
inline Iterator parent(Iterator begin, Iterator idx) {
  return begin + std::distance(begin + 1, idx) / 2;

template <typename Iterator, typename CMP_FN>
inline void heapify(const Iterator begin, const Iterator end, Iterator idx,
                       CMP_FN cmp) {
  bool go = true;
  while (go) {
    const Iterator left(left_child(begin, idx));
    const Iterator right(right_child(begin, idx));
    Iterator candidate(idx);
    // this if relies on if short circuiting cmp has to be guarded for segfault
    if (left < end && !cmp(*left, *idx)) candidate = left;
    //this if relies on if short circuiting cmp has to be guarded for segfault
    if (right < end && !cmp(*right, *candidate)) candidate = right;

    if (candidate != idx) {
      std::swap(*candidate, *idx);
      // go=true;
      idx = candidate;
    } else
      go = false;

template <typename Iterator, typename CMP_FN>
inline void build_heap(const Iterator begin, const Iterator end, CMP_FN cmp) {
  const auto d = distance(begin, end) / 2;
  for (Iterator s = begin + d; s >= begin; s--)
    heapify(begin, end, s, cmp);


/// @Brief Heap sort implementation (inspiration from Cormen et al Book)
// CMP_FN has type: D -> D -> bool
template <typename Iterator, typename CMP_FN>
void heap_sort(Iterator s, Iterator e, CMP_FN cmp) {
  auto d = std::distance(s,e);
  if(d <= 1)

  build_heap(s, e, cmp);
    std::swap(*s, *e);
    heapify(s, e, s, cmp);

When benchmarked those are the timings I get:

Benchmark                                                          Time           CPU Iterations
benchmark_random_values<quicksorter_hoare>/32768             2 ms          2 ms        154
benchmark_random_values<quicksorter_hoare>/65536             4 ms          4 ms         73
benchmark_random_values<quicksorter_hoare>/131072            8 ms          8 ms         34
benchmark_random_values<quicksorter_hoare>/262144           17 ms         17 ms         16
benchmark_random_values<quicksorter_hoare>/524288           36 ms         36 ms          8
benchmark_random_values<quicksorter_hoare>/1048576          77 ms         77 ms          4
benchmark_random_values<quicksorter_hoare>/2097152         159 ms        159 ms          2
benchmark_random_values<quicksorter_hoare>/4194304         336 ms        336 ms          1
benchmark_random_values<quicksorter_hoare>/8388608         701 ms        701 ms          1
benchmark_random_values<quicksorter_hoare>/16777216       1466 ms       1466 ms          1
benchmark_random_values<heap_sorter>/32768                   2 ms          2 ms        111
benchmark_random_values<heap_sorter>/65536                   5 ms          5 ms         53
benchmark_random_values<heap_sorter>/131072                 12 ms         12 ms         24
benchmark_random_values<heap_sorter>/262144                 25 ms         25 ms         11
benchmark_random_values<heap_sorter>/524288                 55 ms         55 ms          5
benchmark_random_values<heap_sorter>/1048576               120 ms        120 ms          2
benchmark_random_values<heap_sorter>/2097152               264 ms        264 ms          1
benchmark_random_values<heap_sorter>/4194304               659 ms        659 ms          1
benchmark_random_values<heap_sorter>/8388608              1657 ms       1657 ms          1
benchmark_random_values<heap_sorter>/16777216             4038 ms       4038 ms          1
benchmark_random_values<sort_heap_std>/32768                 2 ms          2 ms        108
benchmark_random_values<sort_heap_std>/65536                 5 ms          5 ms         60
benchmark_random_values<sort_heap_std>/131072                9 ms          9 ms         30
benchmark_random_values<sort_heap_std>/262144               20 ms         20 ms         12
benchmark_random_values<sort_heap_std>/524288               40 ms         40 ms          7
benchmark_random_values<sort_heap_std>/1048576              87 ms         87 ms          3
benchmark_random_values<sort_heap_std>/2097152             175 ms        175 ms          2
benchmark_random_values<sort_heap_std>/4194304             365 ms        365 ms          1
benchmark_random_values<sort_heap_std>/8388608             757 ms        757 ms          1
benchmark_random_values<sort_heap_std>/16777216           1736 ms       1736 ms          1

As you can see from the timing the heap sorter is much slower than the std::sort_heap and also than a quicksort that I wrote.


Well, before we even consider efficiency, lets look at correctness.
Your code invokes Undefined Behavior:

Specifically, you create iterators which will refer far past the range you got in left_child() and right_child() after you swapped a leaf-node in heapify.

That probably also answers for your programs speed.

Or maybe you compared heap_sort() performance with std::sort_heap(), which is completely unfair as the former corresponds to the combination of std::make_heap()+std::sort_heap().

Why do you split heapsort() into two words like heap_sort()? That at best confuses.

You would normally call the function which builds the whole heap heapify(), though build_heap() isn't that uncommon. I would keep to make_heap() like the standard uses though.

The helper-function you unaccountably called heapify() even though it's just a small step in building a heap would normally be called sift_down(), down_heap() or a variant thereof.

I would expect a (preprocessor-) constant under CMP_FN instead of a template parameter. Why don't you use the more common Compare, Comp or Cmp, especially as it need not be a function, only a callable?

A qualified call to std::swap(a, b) cannot pick up specializations in other namespaces using ADL. Use using std::swap; swap(a, b); or std::iter_swap(pa, pb);.

Regarding "something silly": Avoid using flags for flow-control, dedicated control-structures are easier to follow.

  • \$\begingroup\$ thanks for your answer. I fixed the left_child functions by checking if (distance(idx, end-2) < distance(begin, idx) and if that fails returning end(the right_child similarly). Is that what you were refereeing to as UB? Regarding the performance, the problem is still there. What do you mean that the comparison is unfair? shouldn't make_heap correspond to my build_heap and sort_heap to my heap_sort? Plus thanks for the stylistic comments. I fill fix them up right away. \$\endgroup\$ – Davide Spataro Jul 4 '17 at 16:01
  • \$\begingroup\$ Obviously I am benchmarking make_heap+sort_heap vs my heap_sort \$\endgroup\$ – Davide Spataro Jul 4 '17 at 16:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.