Subject: Re: [boost] storing intrusive_ptr in atomic?!
From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2013-10-25 09:13:40
On Friday 25 October 2013 15:07:36 Oliver Kowalke wrote:
> 2013/10/25 Andrey Semashev <andrey.semashev_at_[hidden]>
>
> > > T * expected1 = 0;
> > >
> > > T * expected2 = 0;
> > > if ( ! a[index].compare_exchange_strong( expected1, expected2) ) {
> > >
> > > // expected1 is not NULL, expected1 removed == a[index] contains NULL
> > >
> > > }
> >
> > No, that's not right. If compare_exchange_strong returns false, the
> > operation
> > have failed and a[index] != NULL (at least, that's how
> > compare_exchange_strong
> > have left it). It should be:
> >
> > T* expected = a[index].load(memory_order_acquire);
> > while (!a[index].compare_exchange_weak(expected, NULL,
> > memory_order_release,
> > memory_order_acquire))
> > {
> > }
> >
> > if (expected)
> >
> > // The expected value was removed from the array
> >
> > else
> >
> > // The array element was removed by some other thread
>
> agreed - but shouldn't we use strong memory-order
> (I remember that Herb S. recommended to use memory_order_strong in one of
> its sessions)
Weak CAS can spuriously fail but also can be cheaper on some platforms.
Spurious failures don't cause any problems in the code above, so weak CAS is
preferred.