std::negation requires Visual Studio 2015 Update 3

[SylvainCorlay]

As reported here.

[wjakob]

I think that we generally require that compiler because earlier versions may ICE when compiling pybind11 code.

[SylvainCorlay]

OK, it was just a FYI as I have a build machine that complained about this on the update to Pybind11 2.0.0.

In other places of the code base, you check for the specific msvc version.

[jagerman]

We should probably put a cmake fatal error for unsupported compilers. (MSVC before 2015u3, GCC before 4.8, etc.)

[dean0x7d]

VS 2015 Update 2 should be able to compile pybind11 with just some minor workarounds, but I think the bigger issue is that it's difficult to test. AFAIK AppVeyor only has a "Visual Studio 2015" option and it's always kept up to date, so there's only Update 3 to test against.

We should probably put a cmake fatal error for unsupported compilers. (MSVC before 2015u3, GCC before 4.8, etc.)

Making the compiler support explicit sounds good, but I think CMake skimps on the minor version numbers. Perhaps doing the check in common.h?

[jagerman]

Somewhat related: the latest MSVC 2017 RC also ICEs when building pybind as well.

What a wonderful compiler.

Edit: this is caused by the decltype() in the template parameter of the second cpp_function constructor, added in PR #677. I'll just ignore it for now, and add a workaround if the final 2017 release (due in the next week or two, IIRC) still hits it.

[jcelerier]

Another problem with std::negation is that it was introduced in C++17 while pybind11 advertises C++11 compat...

[jagerman]

It's only used (along with std::conjunction and std::disjunction) on MSVC, because MSVC has it in the stl (as of 2015u3), and has trouble with some of the template metaprogramming in common.h.

Edit: this is the part that doesn't work on MSVC. But if you look at the surrounding code, you can see we have various differences based on whether under a C++14, C++11, or C++17 compiler, plus special MSVC handling to avoid compiler bugs.

[jcelerier]

@jagerman when my compiler (clang on macOS ) is in -std=c++1z mode, it seems that it goes through the std::negationcode path which isn't implemented yet in apple's clang.
Would a pull request that would put it behind a feature test macro (__cpp_lib_experimental_logical_traits and __cpp_lib_logical_traits) be accepted ?