RFC: Organising .inline.hpp and .hpp includes

[David Holmes]

I'm trying to tackle the incorrect inclusion of .hpp files instead of 
.inline.hpp starting with the use of atomic.inline.hpp ref:

https://bugs.openjdk.java.net/browse/JDK-8157907

The basic rules would seem to be:

If class Foo has inline methods that need to call the inline Atomic 
methods then:

- Those methods should be defined in foo.inline.hpp

    inline void Foo::a() { /* do stuff with atomics */ }

- foo.hpp declares those methods:

    inline void a();

- foo.inline.hpp includes foo.hpp
- foo.inline.hpp includes atomic.inline.hpp
- foo.cpp includes foo.inline.hpp

Do have agreement so far?

Now suppose we have class Bar which uses Foo.

- bar.cpp should include foo.inline.hpp
- If bar.hpp only references Foo but doesn't need to see a full 
definition then bar.hpp can forward declare Foo
- If bar.hpp defines inline methods that require full access to Foo then 
either:

a) bar.hpp includes foo.inline.hpp; or
b) The inline functions of Bar are themselves moved to bar.inline.hpp 
which includes foo.inline.hpp

I think there was general opinion that (a) is undesirable, and that (b) 
is preferable - is that right?

What I am finding is that the fan out of applying (b) is rather large - 
to the point where we may end up having x.hpp, x.inline.hpp and x.cpp 
variants for very many x!

Just to add another dimension to this, what should precompiled.hpp 
include? I would assume the .inline.hpp files right?

Thanks,
David