Make dmsort work with non-default-constructible types (fixes #1)

[Morwenn]

See issue #1 for the detail.

[adrian17]

Nice find, and silly resize. Thanks.

I'll repeat the benchmarks with this change later (in ~12h) (although it's far from the critical part of the algorithm, so it shouldn't matter anyway) and most likely merge it then.

[adrian17]

Hm, the second overload (with-trivial-copies) may also need this, because a type may be trivially copyable and not be default constructible.

[adrian17]

This can most likely be merged with the loop above:

                for (int i = 0; i < num_dropped_in_row; ++i) {
                    --read;
                    *read = std::move(*(dropped.end() - 1));
                    dropped.pop_back();
                }

[Morwenn]

Sure, we could try that too, that might be more sensible.

[Morwenn]

Eh, it's true that I didn't think of trivially copyable types that are not default-constructible, but I guess that such types can indeed exist. Good catch.

[Morwenn]

By the way, it looks like the algorithm could easily be made to work with bidirectional iterators instead of random-access ones. The changes I see to make the magic happen:

• Replace read -= num_dropped_in_row; by std::advance(read, -num_dropped_in_row), first chaging num_dropped_in_row to a signed type (std::iterator_traits<Iter>::difference_type looks like a good choice for that)
• Replace a few it - 1 by std::prev(it) and it + 1 by std::next(it)

Apparently, the algorithm almost never performs big jumps in the collection, so making it work with bidirectional iterators could be worth (unlike e.g. heapsort). You may want to consider it since there aren't many good sorting algorithms for bidirectional iterators :)

---

My small tests on one million integer lists show that it's more than decent for bidirectional iterators too (ok, the more efficient would still probably be to move everything to a vector, sort it, and move everything back):

[adrian17]

Thank you.

I'll take a look at bidirectional iterator changes too, thanks for that advice too!