[Omp] atomic directive semantics

[Sven Karlsson]

Hi,

Richard Henderson wrote:
> Hmm.  But since readers don't use #pragma omp atomic, critical
> sections for writers only are by themselves are not sufficient.
> The platform must have an atomic store instruction for the data
> type; otherwise a reader can read a partial update.

This is up to the programmer. The atomic directive support atomic 
_updates_, see 2.7.4 in 2.5. There are no guarantees for reads of the 
atomically updated entity unless you use extra synchronization. The 
specification is very specific about atomic only providing protection 
for the update.

Naturally, this can be implemented in a number of ways. Some 
implementations, like mine, revert to locks if there is no hardware 
support for a particular data type. This is allowed.

More refined implementations can use load linked/store conditional 
instruction pairs, compare-and-swap or some other means of achieving 
atomic operations. My implementation use such methods for the types that 
  have support in the architecture/hardware.

Naturally, I can only speak for my implementation but I would assume 
most implementations do the same. At least the implementations I have 
studied does something along these lines.

Sanjiv or Larry, step in if the ARB thinks otherwise. This is my 
perception of atomic.

Best regards
  Sven