This looks like inline variable semantics leaking through...

MSVC treats an in-class definition of a static data member as being a definition. For example,

struct T { static const bool value = true; };
extern const bool &b = T::value;

does not require a separate definition of T::value under the MS ABI. Because this affects the ABI, Clang implements the same behavior.

However, this means that Clang thinks the initializer of the static data member in your testcase is part of the definition of the variable, not part of the declaration, and instantiating a class doesn't instantiate the definition portions of any static data members (http://eel.is/c++draft/temp.inst#3.1), so Clang doesn't instantiate the initializer with the class definition and instead delays the instantiation until the variable is used (like it would for an inline static data member).

Clang does diagnose if you later mention Invalid::value (more specifically, if you odr-use it or if it's needed for constant evaluation), because that triggers the instantiation of the definition.

Given that MSVC instantiates the initializer with the class declaration in this case, and doing so is more consistent with the language rules, we should probably do the same.

Sorry for the delay in responding. Thanks for all that information, it was super helpful. I did some digging and it seems like this stems from Clang incorrectly marking constexpr var decls as implicitly inline on Windows. I think I'll be able to put up a fix for this later today.

Thanks again for the help and quick reply!