No namespace awareness of static template fields

[llvmbot]

Bugzilla Link	8084
Resolution	FIXED
Resolved on	Sep 12, 2010 00:25
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor
CC	@DougGregor,@efriedma-quic

Extended Description

cat foo.cxx
namespace F {
template struct Foo {
static A *field;
};
}

using namespace F;
struct Bar : Foo {};
template <> Bar* Foo::field = 0;

clang++ -fsyntax-only foo.cxx
foo.cxx:9:28: error: static data member specialization of 'field' must originally be declared in namespace 'F'
template <> Bar* Foo::field = 0;
^
foo.cxx:3:19: note: explicitly specialized declaration is here
static A *field;

================================================================
system: Linux 2.6.32-24-generic amd64 Ubuntu 10.04
clang version 2.8 (trunk 113084)
Target: x86_64-unknown-linux-gnu
Thread model: posix

[llvmbot]

works:
namespace F {
template struct Foo {
static A *field;
};
}

namespace F {
struct Bar : Foo {};
template <> Bar* Foo::field = 0;
}

---

same error:
namespace F {
template struct Foo {
static A *field;
};
}

struct Bar : F::Foo {};
template <> Bar* F::Foo::field = 0;

[efriedma-quic]

I'm pretty sure gcc is wrongly accepting this... where does the testcase come from?

[llvmbot]

These code snippets are the short version of the problem when I compile ogre3d. Here the original problem in ogre3d 1.7.1 (http://www.ogre3d.org/):

File: ogreSrc/OgreMain/include/OgreSingleton.h
...
namespace Ogre {
template class Singleton {
protected:
static T* ms_Singleton;
};
}

File: ogreSrc/Samples/DeferredShading/include/SharedData.h
...
class SharedData : public Ogre::Singleton { ... }

File: ogreSrc/Samples/DeferredShading/src/DeferredShadingDemo.cpp
...
using namespace Ogre;
template<> SharedData* Singleton::ms_Singleton = 0;

In my opinion this c++ code is valid.

[DougGregor]

Clang is technically correct, per C++98 [temp.expl.spec]p2:

An explicit specialization of a member function, member class or static data member of a class template shall be declared in the namespace of which the class template is a member.

C++0x actually loosens this restriction to allow the code in question, so it's probably reasonable to downgrade this to an "extension warning" in C++98/03 mode. Besides, GCC doesn't diagnose it and EDG only diagnoses it in strict mode, so giving an error is rather harsh.

[DougGregor]

When the code obeys the C++0x semantics but not the C++98/03 semantics (as the examples described here do), I've downgraded the error to a warning in Clang r113718.