Re: [powerpc64le] seq_cst memory order possibly not honored

[Andrey Semashev <andrey dot semashev at gmail dot com>]

On 14.08.2015 11:51, Jonathan Wakely wrote:
> On 14 August 2015 at 01:37, Andrey Semashev wrote:
> > 1. Is my test valid or is there a flaw that I'm missing?
>
> The cppmem tool at http://svr-pes20-cppmem.cl.cam.ac.uk/cppmem/ shows
> that there are consistent executions where (x==0 && y==0) is true. I
> used this code:
>
> int main() {
>    atomic_int a = 0;
>    atomic_int b = 0;
>    int x, y;
>    {{{
>      {
>        a.store(1,mo_seq_cst);
>        a.load(mo_relaxed);
>        x = b.load(mo_relaxed);
>      }
>    |||
>      {
>        b.store(1, mo_seq_cst);
>        b.load(mo_relaxed);
>        y = a.load(mo_relaxed);
>      }
>    }}}
>    return 0;
> }

I'm not familiar with that tool and not sure how to interpret its 
output, but I think it misses one point that is present in the original 
test and I failed to mention. The Boost.Atomic test uses Boost.Thread 
barriers to ensure that both threads have executed the store-load-load 
region before checking for x and y.

Otherwise I cannot see how (x==0 && y==0) could happen. The last load in 
each thread is sequenced after the first seq_cst store and both stores 
are ordered with respect to each other, so one of the threads should 
produce 1.