Re: [C++ PATCH] Diagnose static data members in anon ns which are used, but weren't defined (PR c++/34094)

[Jason Merrill <jason at redhat dot com>]

Jakub Jelinek wrote:
> Before the visibility changes for anon namespaces if a static data member
> wasn't defined (just declared), but used, there would be a link time
> failure.  But as constrain_visibility (decl, VISIBILITY_ANON) makes
> now the static data members !TREE_PUBLIC and DECL_NOR_REALLY_EXTERN,
> even when they are not defined g++ emits them in .bss.

That seems like the bug.  Why are we emitting a definition of something 
which hasn't been defined?

> I'm not sure whether the && DECL_INITIAL (decl) == NULL_TREE should
> be there, given that [class.static.data]/4 says that the definition should
> be present even for const qualified static data members with initializers
> in the class, but removing that line breaks a bunch of testcases in the
> testsuite.

The problem is that decl_needed_p isn't a strong enough test; I think we 
set TREE_USED even if we just pull out the constant value of a static 
data member.  We only need a definition for it if we try to use its address.

Jason