[nikosavola]

• tl::nth_element and the heap functions forward to the standard library, the copied nth_element had no introselect fallback
• tl::sort and tl::partial_sort stay because their tie order is observable, but the copied insertion sort helpers move instead of assign

case	before ms	after ms	speedup
std_sort_boxes/100000	9.78	10.9	0.90x
std_sort_boxes/1000000	112	122	0.92x
std_sort_ints/100000	5.67	5.26	1.08x
std_sort_ints/1000000	67.9	64.2	1.06x
tl_nth_element_dups/100000	0.524	0.535	0.98x
tl_nth_element_dups/1000000	7.37	7.28	1.01x
tl_sort_boxes/100000	9.57	10.1	0.95x
tl_sort_boxes/1000000	106	112	0.95x
tl_sort_ints/100000	5.73	5.22	1.10x
tl_sort_ints/1000000	67.8	61.9	1.10x

Measured on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, g++ -O2 (qmake default), median of 7 repetitions, before and after builds interleaved.

[klayoutmatthias]

why?

I don't recall the exact reason I introduced that, but I think it was some MSVC STL implementation.

It did not employ std::swap, so optimized overloads where not considered.

Matthias

[nikosavola]

Thanks, that makes sense. tl::sort still uses its own partition with tl_iter_swap. I did not change that part.

The parts I replaced did not use swap before either. The heap code just copied values. The standard nth_element uses iter_swap, and since C++11 that finds custom swap functions. Also, tl::nth_element is only used by tlKDTree, and that is only used in a test.

So this is mostly cleanup. If you think it is not worth it, I can close it.