I have a similar problem as this one:
"undefined reference" to static field template specialization
but the workaround they use won't work for me.
I have a CRTP class with static data members, one of which is a std::mutex. Unfortunately, GCC's (4.8.2) linker is giving me an "undefined reference" error for this mutex. Clang (3.4) does not. Is there a workaround? The original question (linked above) invoked the copy constructor on the static data member that forced GCC to emit a symbol, but since my data member is a std::mutex, that's not an option -- the copy constructor is deleted, and there are no argument constructors. Am I just hosed?
I don't believe the problem is with std::mutex, I think the problem is with how GCC handles static data members in template classes that rely on default constructors.
Thanks for any help!
Here's a slim version of my problem: test.hh
#include <mutex>
template < class T >
class CRTP_class {
public:
T * ptr_;
static std::mutex mutex_; // linker error here
static int clearly_a_problem_with_mutex_; // no linker error here
};
class Foo : public CRTP_class< Foo >
{
public:
void set_bar( int setting );
int bar_;
};
test.cc
#include <test.hh>
template<> std::mutex CRTP_class< Foo >::mutex_;
template<> int CRTP_class< Foo >::clearly_a_problem_with_mutex_( 0 );
void Foo::set_bar( int setting ) {
std::lock_guard< std::mutex > locker( mutex_ );
++clearly_a_problem_with_mutex_;
bar_ = setting;
}
main.cc
#include <test.hh>
int main() {
Foo foo;
foo.set_bar( 5 );
}
And then I compile with this command:
g++ -std=c++0x main.cc test.cc -I.
to get the error
/tmp/cclyxUfC.o: In function `Foo::set_bar(int)':
test.cc:(.text+0x86): undefined reference to `CRTP_class<Foo>::mutex_'
collect2: error: ld returned 1 exit status
(Edit 1: In response to the commenter who suggested that this is a duplicate bug with "why do templates have to be in header files" -- There is nothing odd about putting the separate template specializations into the .cc files instead of into the .hh files -- which is really what you need in the case where you've got a mutex and you want, for obvious reasons, only a single copy of that mutex. If you declare a static data member in a header file, then each translation unit that #includes the header will end up with their own copy of the mutex in which case it's not serving its job of ensuring mutual exclusion)
(Edit 2: Ooops! I linked to the wrong prior bug.)
