Re: [PATCH] Optimize inserting value_type into std::vector

[Jonathan Wakely <jwakely at redhat dot com>]

On 15/06/16 11:15 +0100, Jonathan Wakely wrote:
> 	* include/bits/stl_vector.h (vector::_S_insert_aux_assign): Define
> 	new overloaded functions.
> 	* include/bits/vector.tcc (vector::_M_insert_aux): Use new functions
> 	to avoid creating a redundant temporary.
> 	* testsuite/23_containers/vector/modifiers/insert_vs_emplace.cc: New
> 	test.
>
> Tested x86_64-linux.
>
> This improves our performance on Howard Hinnant's "insert vs emplace"
> experiment at
> http://htmlpreview.github.io/?https://github.com/HowardHinnant/papers/blob/master/insert_vs_emplace.html
>
> With this small change there is no difference between emplacing or
> using the relevant insert / push_back function. That also means we
> beat libc++ in some cases, making us the bestest, whoo!

We still lose to libc++ in one case. For the "lvalue no reallocation"
test our insert and emplace are equal to libc++'s emplace, but
libc++'s insert is even better.

For "xvalue no reallocation" and "rvalue no reallocation" our insert
and emplace are equal to libc++'s insert, but libc++'s emplace is
worse. So either my patch is wrong or we've got a new minimum for the
emplace operations.

In all other cases our results are unchanged by this patch.

So if this patch goes in we fix the case that made it difficult for
Howard to give consistent advice, where emplace was better than insert
for libstdc++ but not the other implementations. With this patch
insert is always at least as good as emplace, for all implementations.