Re: Sync Rep v17

Marti Raudsepp <marti@juffo.org> writes:
> On Tue, Feb 22, 2011 at 07:38, Fujii Masao <masao.fujii@gmail.com> wrote:
>> +       SpinLockAcquire(&WalSndCtl->ctlmutex);
>> +       result = WalSndCtl->sync_rep_service_available;
>> +       SpinLockRelease(&WalSndCtl->ctlmutex);

>> volatile pointer needs to be used to prevent code rearrangement.

> I don't think that's necessary. Spinlock functions already prevent
> reordering using __asm__ __volatile__

You're mistaken.  We started using that volatile-pointer convention
after noting that some compilers would misoptimize the code otherwise.

It's not a problem with LWLock-protected stuff because the LWLock calls
are actual out-of-line function calls, and the compiler knows it doesn't
know what those functions might do.  But gcc is a lot more willing to
reorder stuff around asm operations, so you can't assume that
SpinLockAcquire/SpinLockRelease are equally safe.  The way to prevent
optimization bugs is to make sure that the fetches/stores protected by a
spinlock are done through volatile pointers.
        regards, tom lane