Support class-specific operator new and delete

[dean0x7d]

Fixes #754.

void_t was needed to work around a GCC 4.8 issue, but it's pretty useful to have anyway.

[jagerman]

Looks good to me.

[dean0x7d]

I'd recommend this waits for v2.2 since it does affect a pretty key part of class support. Also might need to see about future-proofing for C++17 alignment-aware new.

[jagerman]

C++17 support can definitely wait for 2.2, but I think the current PR fixes an important issue that is likely going to come up again in 2.1's stable lifetime (given the improved Eigen support in 2.1).

[wjakob]

I've merged this now so that it can be part of 2.1

[dean0x7d]

Note: Clang 4.0 doesn't compile in C++17mode due to #ifdef __cpp_lib_void_t. I'll push a minor fix shortly.

[jagerman]

Whoops, didn't notice that.

Should we promote the gcc7/clang4 builds to non-optional? (clang 4.0 is building with the stable released version now; gcc7 is still a snapshot, so perhaps just clang 4?)

[dean0x7d]

I just pushed a fix for clang 4.0: f0e58a6

[wjakob]

Just out of curiosity: what's the issue with __cpp_lib_void_t? Does Clang define it but then not offer void_t?

[jagerman]

No, it isn't working as its supposed to. http://bugs.llvm.org/show_bug.cgi?id=26086 is the issue, I believe.

[jagerman]

... which basically means LLVM libc++'s void_t is broken, since doing exactly this is the entire point of std::void_t in the first place.

[dean0x7d]

Exactly that. The new-style implementation of void_t is template<class...> using void_t = void;, i.e. no need for a helper struct. The new libc++ implements it that way, but clang 4.0 doesn't implement CWG1558 so it just ignores unused alias template arguments. The interesting thing is that this did work before, sometime in clang 3.x.

Anyway, I just removed __cpp_lib_void_t rather than adding extra compiler version checks since both gcc and clang are in c++1z mode for now.