source: CIVL/doc/omp/independence.txt@ e898536

In general the structure of an OMP Parallel statement looks like this:

#pragma omp parallel
 { 
  S1
#pragma omp ...
  S2
  S3
 }

Where the parallel contains regions of C code, e.g., S1, S3, that
are executed by each thread, and OMP statements, e.g., S2, which
control how threads are executed by threads.  OMP worksharing and
synchronization statements have different semantics from parallel
(and enclosing workshares), but the principle is the same.

To perform an independence analysis each region of the program which
may execute in parallel must be subjected to analysis.  

This requires that the barrier semantics of OMP constructs be
made explicit and that an analysis collects all sets of potentially
parallel regions.  See a summary of implicit barriers and other 
thread-ordering related OMP statements below.

An interesting challenge arises due to the nesting of OMP statements
within control flow constructs and the "shape" of the regions that
may arise.   For example:

#pragma omp parallel
  if (c) 
#pragma omp for    
    S1
  else
    S2


Independence comes down to whether two potentially parallel regions
may exhibit a read-write dependence.   These dependences arise from
accesses to shared variables including shared arrays where the index
expressions cannot be determined to be disjoint.

--------------------------------------------------
Barrier semantics for OMP statements:

parallel
 end barrier

critical
  no barrier, but special semantics should be considered to improve
  precision of independence analysis

barrier
  end barrier

atomic
  no barrier, but special semantics should be considered to improve
  precision of independence analysis

ordered
  no barrier, but special semantics should be considered to improve
  precision of independence analysis

for
 end barrier

for nowait

sections
 end barrier

sections nowait

section

single
 end barrier

single nowait