Subj : Re: Memory visibility with Pthreads
To   : comp.programming.threads
From : Maciej Sobczak
Date : Mon Aug 15 2005 11:12 am

Peter Dimov wrote:

>>lock
>>A = 10;
>>B = 10;
>>unlock
> 
> and this:
> 
> lock
> B = 10;
> A = 10;
> unlock

Of course, but the point was to ensure what is the meaning of lock and 
unlock in this context. From the visibility point of view, anything that 
happens *between two consecutive calls* to any synchronization functions 
cannot be relied upon with regard to their order - from the point of 
view of other threads, the two versions above make no difference, or at 
least no other thread should expect and rely on any. Even worse - it is 
not even sure whether there are two stores or maybe just only one (think 
of A and B being "neighbours" within what could be stored as a single 
longer "word"). That's why my concern was only whether the assignment to 
A happens "before" or "after" lock. And since lock itself does not 
influence the visibility of A in other threads, then the assignment to A 
can be moved into the critical section without disrupting anything that 
is otherwise guaranteed.


-- 
Maciej Sobczak : http://www.msobczak.com/
Programming    : http://www.msobczak.com/prog/

.