From the disassembler view, it seems the implementation of weak_ptr::lock in Windows SGX SDK is similar to the Linux one :
__shared_weak_count*
__shared_weak_count::lock() _NOEXCEPT
{
long object_owners = __libcpp_atomic_load(&__shared_owners_);
while (object_owners != -1)
{
if (__libcpp_atomic_compare_exchange(&__shared_owners_,
&object_owners,
object_owners+1))
return this;
}
return nullptr;
}
However the compare_exchange function ends up being this piece of code from include\libc++\support\win32\atomic.h :
template<class _Tp>
static inline bool __atomic_compare_exchange_strong_explicit(volatile _Tp *__val, volatile _Tp *__expected, _Tp __after, int, int)
{
return _atomic_msvc_wrapper::FetchClass<sizeof(_Tp)>::compare_exchange<_Tp>(__val, *__expected, __after);
}
This implementation is incorrect, as it loses the feature that if the exchange doesn't happen, the current value of __val is written to __expected. As a result, if indeed there has been a change between the load on line 4 and the exchange on line 7, object_owners is not modified, and the loop continues, possibly forever (if the value doesn't come back to what it was).
This happens sometimes in our code, during shutdown, when many shared or weak pointers are released concurrently.
