Re: autopar reduction and OMP_ATOMIC gimplify/expand changes - final patch

[Razya Ladelsky <RAZYA at il dot ibm dot com>]

Zdenek Dvorak <rakdver@kam.mff.cuni.cz> wrote on 31/10/2007 14:08:43:

> Hi,
> 
> looking at the example, there seem to be two problems with the
> transformation:
> 
> > after reduction transformation (only relevant parts):
> > 
> > parloop
> > {
> > 
> > ....
> > 
> >   # Two new variables are created for each reduction:
> >   "reduction" is the variable holding the neutral element for the
> >   particular operation, e.g. 0 for PLUS_EXPR, 1 for MULT_EXPR, etc.
> >   "reduction_initial" is the initial value given by the user.
> >   It is kept and will be used after the parallel computing is done.  #
> > 
> >   reduction_initial.24_46 = 1;
> >   reduction.23_45 = 0;
> >   .paral_data_store.32.reduction.23 = reduction.23_45;
> > 
> >   #pragma omp parallel num_threads(4) 
> >   reduction.28_48 = .paral_data_load.33_51->reduction.23;
> 
> You cannot do this -- another thread could have already altered the
> value of reduction.23.  You need a separate variable holding the initial
> value of reduction.28_48 (or preferably, as this initial value is always
> a constant, just put the constant here directly).

Very true.
Thanks for noticing.
I'll send the patch to fix it.

> 
> >   #pragma omp for schedule(static)
> >   # sum.27_29 = PHI <sum.27_11, reduction.28_48>
> >   sum.27_11 = D.1827_8 + sum.27_29;
> >   OMP_CONTINUE
> > 
> >   # Adding this reduction phi is done at create_phi_for_local_result() 
#
> >   # reduction.23_58 = PHI <sum.27_11(5), 0(23)>
> >   OMP_RETURN
> > 
> >   # Creating the atomic operation is done at 
> >   create_call_for_reduction_1()  #
> > 
> >   #pragma omp atomic_load
> >   D.1839_59 = *&.paral_data_load.33_51->reduction.23;
> >   D.1840_60 = reduction.23_58 + D.1839_59;
> >   #pragma omp atomic_store (D.1840_60);
> > 
> >   OMP_RETURN
> > 
> >  # collecting the result after the join of the threads is done at
> >   create_loads_for_reductions().
> >   a new variable "reduction_final" is created.  It calculates the 
final
> >   value from the initial value and the value computed by the threads #
> > 
> >   reduction_final.34_53 = .paral_data_load.33_52->reduction.23;
> >   sum_37 = reduction_initial.24_46 + reduction_final.34_53;
> >   sum_43 = D.1795_41 + sum_37;
> 
> There seems to be another error -- paral_data_load is no longer 
accessible
> at this point; you need to load the value from paral_data_store.
> 

Actually the complete code is:

  .paral_data_load.33_52 = &.paral_data_store.32;
  reduction_final.34_53 = .paral_data_load.33_52->reduction.23;
  sum_37 = reduction_initial.24_46 + reduction_final.34_53;

So .paral_data_load.33_52 is initialized with .paral_data_store.32
Ok?

> Zdenek