[Omp] A question about OpenMP 2.5

[Tim Prince]

anmey at rz.rwth-aachen.de wrote:
> Dear  all,
> 
> I don't think that the question was related to the OpenMP memory model, 
> but to the wording and understanding of the current OpenMP specification.
> 
> Brad looked at a perfectly legal piece of code
> (I slightly modified it using default(none) etc.)
>  >     void a1(int n, float *a, float *b)
>  >     {
>  >         int i;
>  >     #pragma omp parallel for default(none) shared(a,b) private(i,n)
>  >         for (i=1; i<n; i++) /* i is private by default */
>  >             b[i] = (a[i] + a[i-1]) / 2.0;
>  >     }
> 
> and thinks that when reading the specs, the result of this loop shall be 
> undefined.
> 


> if the memory regions to which a and b are pointing to do not overlap 
> (aliasing) multiple threads are writing to different memory locations 
> and not to the same variable as b[i] is not a variable.
> This is not related to atomicity etc.
Appropriate use of restrict and (possibly) const qualifiers would be 
required to permit this loop to be optimized, in the absence of omp 
directive.   As written, overlap is not excluded.
I don't think the OpenMP standard says so explicitly, but the omp 
directive is likely to be taken as an assertion of no aliasing, thus 
permitting threading, but possibly not permitting full optimization of 
code executed by each thread.