std::sort does not always call std::swap

For small ranges, std::sort implementations in GCC’s stdlibc++ (and other standard library implementations) recurs to insertion sort for performance reasons (it’s faster than quicksort / introsort on small ranges).

GCC’s insertion sort implementation in turn doesn’t swap via std::swap – instead, it moves whole ranges of values at a time, instead of swapping individually, thus potentially saving performance. The relevant part is here (bits/stl_algo.h:2187, GCC 4.7.2):

typename iterator_traits<_RandomAccessIterator>::value_type
  __val = _GLIBCXX_MOVE(*__i);
_GLIBCXX_MOVE_BACKWARD3(__first, __i, __i + 1);
*__first = _GLIBCXX_MOVE(__val);

_GLIBCXX_MOVE is the same as std::move from C++11 and _GLIBCXX_MOVE_BACKWARD3 is std::move_backward – however, this is only the case if __GXX_EXPERIMENTAL_CXX0X__ is defined; if not, then these operations resort to copying instead of moving!

What this does is move the value at the current position (__i) to a temporary storage, then move all previous values from __first to __i one up, and then re-insert the temporary value at __first. So this performs n swaps in one operation instead having to move n values to a temporary location:

  first           i
+---+---+---+---+---+---+
| b | c | d | e | a | f |
+---+---+---+---+---+---+
                  |
  <---------------+


  first           i
+---+---+---+---+---+---+
| --> b-> c-> d-> e-> f |
+---+---+---+---+---+---+


  first           i
+---+---+---+---+---+---+
| a | b | c | d | e | f |
+---+---+---+---+---+---+
  ^

Leave a Comment