Simplify <optional> macro

[jagerman]

This simplifies the std::optional/std::experimental::optional macros by
having just one macro that provides the namespace (either std or
std::experimental), which removes some code duplication.

It also enables the std::optional test which was previously only
testing optional under c++14 mode with std::e::o, but skipping it under
c++17 mode with std::o.

[wjakob]

Hmm, I think the argument for keeping both is that they may be different types in an actual implementation of libstdc++. So partial overloading just on std::optional may break std::experimental::optional code.

[jagerman]

It still has both--they just get referenced via PYBIND11_OPTIONAL_NS::optional now instead of having two different chunks of code with the only difference being that one writes std::optional and the other std::experimental::optional.

[jagerman]

(And we don't ever support both at once).

[wjakob]

Yes -- but what if we're compiling code on a c++17 compilant compiler that was still using std::experimental::optional?

[wjakob]

(I admit that I don't really get the policies about std::experimental namespace. For instance, will aliases there always exist or only temporarily?)

[wjakob]

cc @aldanor

[aldanor]

But I think that we do want to support both at once, that was one of the points. GCC has both and the experimental one is not going away in GCC 6. Maybe next major GCC or may be not. If you look at your compiler's experimental include folder, you'll find a bunch of stuff there that has become stable since.

[jagerman]

Yes -- but what if we're compiling code on a c++17 compilant compiler that was still using std::experimental::optional?

At least for libstdc++, this is impossible: including <optional> in C++14 or below is an (intentional) compilation error, and including <experimental/optional> in C++17 is likewise a compilation error. (Edit: never mind, I'm clearly wrong about that.)

I took a look at clang with libc++, and, at least as of llvm 3.9, the current c++1z support doesn't have <optional>, so I can't really say what will happen there.

[jagerman]

Hmm, perhaps I'm wrong about it, but I'm pretty sure it was the case with the gcc-7 snapshot when I submitted the other PR, but I can't reproduce that now. With the current snapshot, including both seems fine--I'll amend the PR.

And indeed they are not the same:

#include <experimental/optional>
#include <optional>
#include <iostream>

int main() {
    std::cout << "same: " << std::is_same<std::optional<int>, std::experimental::optional<int>>::value << std::endl;
}

outputs: same: 0

[jagerman]

Never mind, I'm just misremembering the logic here; including both was always allowed. I'll fix the PR to just fix the test case.

[jagerman]

My bad on that one (especially considering I wrote that logic in the first place), sorry for the noise.

I amended the PR to just add testing for std::optional when available.

[jagerman]

And though we don't currently have any tests in C++17 mode, I tested it locally with a g++ 6.2 with -std=c++17 and it runs and passes both tests.

[wjakob]

Thanks!