C++: Static Members can't be Defined at Declaration, but Static Function Variables can?

Question

Here are two variables declared with the keyword static:

void fcn() {
    static int x = 2;
}

class cls() {
    static int y;
};

We all know that in order for cls to link properly, int cls::y needs to be explicitly defined by the programmer exactly once.

Based on the answers to static variables in an inlined function , it seems that even though no out-of-class definition is required for fcn::x , it is guaranteed that even inlined versions of fcn from different compilation units will reference the same fcn::x. If this is true, then the linker has to be smart enough to reach between compilation units and connect multiple instances of "the same" variable to ensure that static function variables perform as expected.

If this is possible for static function variables, it seems to me that it should also be possible for static class members... so why does the standard require a single out-of-class definition of static class members?

Answer 1

Yes, the linker will indeed have to merge different instances of fcn::x. In other words, even though formally the language says that fcn::x has no linkage, physically it will have to be exposed as an external symbol in all object files that contain it. This is how it is typically implemented in practice: your compiler will expose fcx::x as some sort of heavily mangled external name @#$%^&_fcx_x or such (to ensure it can never clash with "real" external names). This is what the linker will use to merge all instances of fcn::x into one.

As for class members... Firstly, it is not really about what is "possible". It is about the language-level concepts of declarations and definitions. It is about One Definition Rule, which is a higher-level concept than what is "possible" based on raw linker features. According to that rule, objects with external linkage shall be defined by the user and shall have one and only one definition. Static data members of the class are objects with external linkage. The rest follows.

Secondly, and more practically, there's another serious issue with static data members. It is their order of initialization. Static data members are [guaranteed to be] initialized no later than when the first function from the containing translation unit is called (which refers to the translation unit that contains the data member definition). And static objects declared in a single translation unit are initialized the order of their definition, top-to bottom. This is an important property of static data member initialization process. Allowing static data members to be defined "automatically" would defy this part of the specification and would require massive changes to this part of the language.

In other words, when you provide a dedicated definition for a static data member of the class, you are not just doing it for ODR compliance, you are actually expressing your desired initialization order for that object.

Meanwhile, static variables inside functions are objects with no linkage. Hence they receive a different treatment at the conceptual level. And they have well-defined order-of-initialization semantics that is completely unaffected by the need to merge multiple definitions into one.