[cfe-dev] OpenMP support in CLANG: A proposal

[Ronan Keryell]

>>>>> On Wed, 07 Nov 2012 17:02:23 +0100, Olaf Krzikalla <Olaf.Krzikalla at tu-dresden.de> said:

    Olaf> Am 07.11.2012 16:00, schrieb Mahesha HS:
    >> Regarding your question about *scope*, I meant following
    >> example. In OpenMP annotated C++ source, following code snippet
    >> is valid. In this particular example, I meant about handling the
    >> scope of variable "VarA" .
    >> 
> =======================================
> #pragma omp parallel if (int VarA = 2)
> {
>     // parallel region
> }
>
> VarA = 3;
>
> #pragma omp parallel if (VarA = 2)
> {
>     // parallel region
> }
> ========================================

    Olaf> This is not backed up by the OpenMP standard. The Grammar
    Olaf> mentions only
    Olaf> if (expression)

    Olaf> And the C++ standard has its own grammar rules for expressions
    Olaf> in conditions in order to make declarations possible.  Of
    Olaf> course if the above extension is consensus among all other
    Olaf> compilers then we have to adopt to it. However I doubt it. It
    Olaf> begs for trouble if you e.g. compile with OpenMP
    Olaf> disabled. Thus, if this is just an gcc extension I'd just
    Olaf> leave it out.

I was puzzled by this too and I quickly browse again the OpenMP 3.1
standard.

I think we should get back to the "OpenMP spirit", but of course it is
very subjective and I had some passionate debates on this (hi Antoniu
Pop! :-) ).

For me, the first example violates the property that if a compiler
ignores the #pragma omp, the program should remain a syntax-correct
program here. If we have an example where VarA is not defined before but
used inside the parallel block, it would not even compile.

The second example is that we have a side effect on VarA = 2.
This makes me uncomfortable since I try usually to write OpenMP programs
with the same semantics as the sequential program resulting of ignoring
the #pragma. It is where we can argument on what is the "OpenMP spirit",
since I do not remember having seen this stated as an axiom in the norm.

The only thing I can relate to it in the "OpenMP Application Program
Interface Version 3.1 July 2011" norm:

p. 35: Restrictions
lines 24-25:

    A program must not depend on any ordering of the evaluations of the
    clauses of the parallel directive, or on any side effects of the
    evaluations of the clauses.

p. 36: 2.4.1 Determining the Number of Threads for a parallel Region
lines 16-19:

    The if clause expression and the num_threads clause expression are
    evaluated in the context outside of the parallel construct, and no
    ordering of those evaluations is specified. It is also unspecified
    whether, in what order, or how many times any side-effects of the
    evaluation of the num_threads or if clause expressions occur.

So these not preclude the use of side effect (well, I had not imagined
to be able to put side effects here. Thank you for this Halloween
suggestion :-) ). I just keep thinking side effects in the OpenMP pragma
is against the OpenMP spirit and depends too much on implementation
details.

So clearly
#pragma omp parallel num_threads(++i) if(++i)
is forbidden :-)

If we go back to:

#pragma omp parallel if (VarA = 2)
{
    // parallel region
}
this is not forbidden by the norm but I choke on the fact that if we
have an OpenMP compiler, the value of VarA is 2 in the continuation of the program
after the #pragma, but if we compile without an OpenMP compiler, the
value of VarA is unchanged neither even accessed.

So to summarize: I vote for Olaf Krzikalla on this point. :-)
-- 
  Ronan KERYELL                            |\/  Phone:  +1 650 386 6482
  SILKAN Wild Systems                      |/)
  4962 El Camino Real #201                 K    Ronan.Keryell at silkan.com
  Los Altos, CA 94022                      |\   skype:keryell
  USA                                      | \  http://silkan.com