[cpp-threads] Whence Sequential Consistency?

[Peter Dimov]

Lawrence Crowl wrote:

> But Raul Silvera (above) seems to imply that not all uses of locks
> result in a sequentially consistent program.

I assume that you are talking about programs that have no data races (apart 
from the lock locations). Even then, consider this equivalent of one of the 
classic non-SC examples:

// a, b initially locked

// T1

unlock( a );
r1 = try_lock( b );

// T2

unlock( b );
r2 = try_lock( a );

I think that it's perfectly possible for both try_locks to fail, something 
that can't happen in a sequentially consistent execution.

> How can non-sequentially consistent atomics build
> a lock that enforces sequentially consistent behavior on programs?

Why not? For one thing, a lock is allowed to wait until SC is achieved, 
while the atomics are non-blocking.