Le 16/08/2017 à 01:13, Peppard via Boost a écrit :
However I maybe found another regression introduced after the 1.64 release, this time in Boost.Thread (with pthread implementation and BOOST_HAS_NANOSLEEP). A minimal example:
#include
int main() { boost::this_thread::sleep(boost::posix_time::seconds(1)); }
This sleeps "forever", however a slightly more complicated program might not. I bisected it to this commit: https://github.com/boostorg/thread/commit/d4cff01c724f1355fef0990f3d6d678e52...
The problem seems to be the BOOST_THREAD_USEFIXES_TIMESPEC define, which alters some internal behavior, effectively changing
nanosleep({0, 999366710}, NULL) into
nanosleep({1501988898, 350057596}, [stuck for a very long time] . If the sleep is called inside a boost::thread, hidden::sleep_until is implemented via condition.do_wait_until instead of nanosleep, this works. For me a slightly extended test code works as expected if compiled in debug mode but is stuck in release mode. I'm testing under Linux with GCC 7.1 and a Clang++ 5.0.0 snapshot. This bug was fixed in the develop branch a couple weeks ago. https://github.com/boostorg/thread/pull/127 (At least it looks like the same bug. I never tracked down the commit
On 8/15/2017 6:10 AM, Peppard via Boost wrote: that introduced the bug, or figured out whether or not it was a regression.) Looks like this fixes the bug, but as you said, only in develop. And even with the fix, Thanks for checking. I get the impression that the library is in a experimental/unstable state with all the time/clock changes over the
On 2017-08-15 19:45, David Olsen via Boost wrote: past months. There is even commented out code in some places. And we have timespec_now() vs timespec_now_realtime(), the first one being used only when BOOST_THREAD_USEFIXES_TIMESPEC is not defined, but if it's not defined, the code of timespec_now and timespec_now_realtime is the same, right? Do I miss something? You are right that I have had a lot of trouble lastly with timeout issues. Iwill see what can I do this evening.
Anyway, as it stands, it looks like rc3 thread lib does not pass its own test suite (stuck e.g. in ex_tennis). Isn't passing the regression test suite (at least for major platforms) a requirement for an official release? Vicente