On 7 November 2011 21:51, Jonathan Wakely wrote:
>
> Aha, this is a problem with all platforms where _GLIBCXX_HAVE_TLS is
> not defined, std::call_once uses std:function which assumes a copyable
> target object.
>
> I'm working on it ..
>

Fixed like so ...

        * include/std/mutex (call_once): Store closure in __once_functor
        as bound function wrapper might not be copyable.

The fix would be simpler but the lambda can't capture the parameter
pack directly because of PR c++/41933. Unfortunately the non-TLS
std::call_once implementation is quite wasteful, creating a function
wrapper, which is captured by a closure, which is stored in a
std::function.  IIRC the std::function has to be kept for backwards
compatibility, otherwise we could just store the closure and a pointer
to a function, like the TLS code but using a mutex to serialize calls.

Tested x86_64-linux and x86_64-netbsd5.1, committed to trunk.
Index: include/std/mutex
===================================================================
--- include/std/mutex	(revision 181123)
+++ include/std/mutex	(working copy)
@@ -810,8 +810,9 @@
       __once_call = &__once_call_impl<decltype(__bound_functor)>;
 #else
       unique_lock<mutex> __functor_lock(__get_once_mutex());
-      __once_functor = std::__bind_simple(std::forward<_Callable>(__f),
+      auto __callable = std::__bind_simple(std::forward<_Callable>(__f),
           std::forward<_Args>(__args)...);
+      __once_functor = [&]() { __callable(); };
       __set_once_functor_lock_ptr(&__functor_lock);
 #endif
 

Reply via email to