[cfe-dev] Eager instantiation of static data member

[Stephan Bergmann via cfe-dev]

On 04/01/2019 11:17, Stephan Bergmann wrote:
> I'm puzzled why Clang (up through recent trunk) rejects
> 
>> $ cat test.cc
>> template<typename T> struct S { static const int n = T::n; };
>> S<int> s;
>>
>> $ clang++ -std=c++17 -c test.cc
>> test.cc:1:54: error: type 'int' cannot be used prior to '::' because 
>> it has no members
>> template<typename T> struct S { static const int n = T::n; };
>>                                                      ^
>> test.cc:2:8: note: in instantiation of template class 'S<int>' 
>> requested here
>> S<int> s;
>>        ^
>> 1 error generated.
> 
> when C++17 [temp.inst]/3 states that "the initialization (and any 
> associated side effects) of a static data member does not occur unless 
> the static data member is itself used in a way that requires the 
> definition of the static data member to exist."
> 
> Is the initializer instantiated to check whether it is a constant 
> expression (as required by [class.static.data]/3)?  If yes, is the 
> compiler indeed allowed/required to check that when instantiating S<int> 
> here?
> 
> (A similar
> 
>> template<typename T> struct S { static constexpr int n = T::n; };
>> S<int> s;
> 
> is accepted by Clang, FWIW.)

Upon closer inspection,

   static const int n = T::n;

is a declaration but not a definition, so C++17 [temp.inst]/2 "The 
implicit instantiation of a class template specialization causes the 
implicit instantiation of the declarations, but not of the definitions 
[...] of the [...] static data members [...]" arguably doesn't kick in 
to avoid instantiation of the initializer.  (Whereas the implicitly inline

   static constexpr int n = T::n;

/is/ a definition.)

While for the initializers of non-static data members the intent is 
apparently to not instantiate them eagerly (see 
<http://www.open-std.org/jtc1/sc22/wg21/docs/cwg_active.html#1396> 
"Deferred instantiation and checking of non-static data member 
initializers", and

> template<typename T> struct S { int n = T::n; };
> S<int> s{0};

is indeed accepted by Clang), this appears to not carry over to 
initializers of (non-inline, in-class) static data member declarations.

The resulting question is whether that is a conscious decision or an 
oversight in the standard.  (Note that at least recent GCC appears to 
not eagerly instantiate such initializers; it accepts the original test.cc.)