On Wed, Jun 24, 2020 at 05:47:06PM +0100, Kwok Cheung Yeung wrote:
> There appears to be a bug in the handling of the 'if' clause (without a
> directive name modifier) for combined OpenMP constructs in the Fortran
> front-end:
>
> static void
> gfc_split_omp_clauses (gfc_code *code,
> gfc_omp_clauses clausesa[GFC_OMP_SPLIT_NUM])
> {
> ...
> if (code->ext.omp_clauses != NULL)
> {
> if (mask & GFC_OMP_MASK_TARGET)
> {
> /* And this is copied to all. */
> clausesa[GFC_OMP_SPLIT_PARALLEL].if_expr
> = code->ext.omp_clauses->if_expr;
> }
Yeah, looks like a pasto.
> Currently, if 'target' is in the combined contruct, then the 'if' is applied
> to the 'parallel' construct, but there are combined constructs with 'target'
> but not 'parallel' (e.g. target teams distribute), which result in the 'if'
> not getting applied at all. This is also redundant, as the unmodified if is
> always applied to the 'parallel' construct if there is one.
>
> The patch changes the behaviour to match what the common C/C++ FE does,
> which is to apply the 'if' to every applicable sub-construct in the combined
> construct. I have included a testcase to check that the if clauses have been
> applied correctly by the time it gets to the ME. I have also found a case
> that results in an ICE (using 'target parallel' with an 'if' clause) - I
> have commented out this out for now and filed it as PR 95869.
>
> I have tested for regressions in the gfortran and libgomp testsuites. Okay
> for master/OG10?
Ok, thanks.
Though, while you are at it, if you'd like to also add the related, but
still missing support for if (simd: ...), or if (cancel: ...) that C/C++
already do support, it would be greatly appreciated.
Jakub
