https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110016
--- Comment #13 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
I suspect if you change the lambda/call to substrate::threadPool_t::waitWork to
be:
inline std::pair<bool, std::tuple<args_t...>> waitWork()
noexcept
{
std::unique_lock<std::mutex> lock{workMutex};
++waitingThreads;
// wait, but protect ourselves from accidental
wake-ups..
auto b = [this]() noexcept -> bool { bool t = finished;
for(volatile int i = 0 ; i < 10000;i ++); return t || !work.empty(); };
//if (!b())
haveWork.wait(lock, b);
--waitingThreads;
if (!work.empty())
return {true, work.pop()};
return {false, {}};
}
you might hit the issue with more C++ implementations. This should simulates
the issue I was mentioning by adding a slight delay between the load of
finished before the call to wait(lock).