> On 5. Dec 2024, at 14.38, Martin Storsjö <mar...@martin.st> wrote: > > Within winpthreads, a native posix thread is a thread that has > been created via pthread_create. > > Previously, pthread key destructors were executed immediately when > the thread exits, for native posix threads. For non-posix native > threads (i.e. threads created via some other API than pthreads), > the pthread key destructors were executed later, via the > __dyn_tls_pthread callback. > > When a thread has got TLS objects, where the TLS implementation is > provided via emutls (compiled with the GCC posix thread model, which > uses winpthreads), the memory for those TLS objects is allocated > via emutls and gets deallocated via a pthread key destructor. > > When the destructors of the TLS objects are executed via > mingw-w64-crt's __cxa_thread_atexit, those destructors are executed > via a TLS callback, after the point when emutls has deallocated > their storage memory. > > To fix this, don't run the pthread key destructors immediately, > but later via the __dyn_tls_pthread callback, just like for non-posix > native threads. This makes sure the memory for the TLS objects isn't > freed at the time when the destructors are executed. > > See also: https://bugreports.qt.io/browse/QTBUG-131476
Also see https://bugreports.qt.io/browse/QTBUG-131892 for more discussion on the same, which also contains a direct request about whether we can try to fix winpthreads for this issue. // Martin _______________________________________________ Mingw-w64-public mailing list Mingw-w64-public@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
