https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59767
--- Comment #3 from mikulas at artax dot karlin.mff.cuni.cz --- The problem here is that we don't really know what does the standard specify. People often suggest the Batty's paper Mathematizing C++ Concurrency ( http://www.cl.cam.ac.uk/~pes20/cpp/popl085ap-sewell.pdf ) as the explanation, but the paper really describes a different memory model than the C++ standard (citing section 4: "Unfortunately N3092 allows the following nonsequentially consistent execution of the SB example with SC atomics (initialisation writes, such as (a) and (b), are non-atomic so that they need not be compiled with memory fences): - We devised a stronger restriction on the values that may be read by SC atomics, stated in ยง2.7, that does provide sequential consistency here." - so it doesn't describe the standard, but something stronger that authors have "devised") You can look at this example https://gcc.gnu.org/ml/gcc/2014-02/msg00053.html . The assert can fail - so it implies that __ATOMIC_SEQ_CST is not a full barrier, it is somehow weaker. But how much weaker is it? Who knows? I look at https://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html and I would really like to know "why is it so?" "Where does the standard specify this mapping?" I couldn't really find an answer for that.
