On Mon, Jul 04, 2022 at 06:09:31PM +0200, Tobias Burnus wrote:
> thanks for the comment & spec referral. I have now updated the patch –
> and included the new C/Fortran testcase.
Thanks.
> + while (true)
> + {
> + old_loc = gfc_current_locus;
> + if (gfc_match ("val )") == MATCH_YES)
> + {
> + if (linear_op != OMP_LINEAR_DEFAULT)
> + {
> + duplicate_mod = true;
> + break;
> + }
> + linear_op = OMP_LINEAR_VAL;
> + has_modifiers = true;
> + break;
> + }
> + else if (gfc_match ("val , ") == MATCH_YES)
> + {
> + if (linear_op != OMP_LINEAR_DEFAULT)
> + {
> + duplicate_mod = true;
> + break;
> + }
> + linear_op = OMP_LINEAR_VAL;
> + has_modifiers = true;
> + continue;
> + }
Perhaps you could avoid some code duplication by doing it like:
bool close_paren = gfc_match ("val )") == MATCH_YES;
if (close_paren || gfc_match ("val , ") == MATCH_YES)
{
if (linear_op != OMP_LINEAR_DEFAULT)
{
duplicate_mod = true;
break;
}
linear_op = OMP_LINEAR_VAL;
has_modifiers = true;
if (close_paren)
break;
else
continue;
}
and similarly for uval and ref.
> + else if (!has_modifiers
> + && gfc_match ("%e )", &step) == MATCH_YES)
> + {
> + if ((step->expr_type == EXPR_FUNCTION
> + || step->expr_type == EXPR_VARIABLE)
> + && strcmp (step->symtree->name, "step") == 0)
> + {
> + gfc_current_locus = old_loc;
> + gfc_match ("step (");
> + has_error = true;
> + }
I think the above should accept even
linear (v : step (1) + 0) or linear (v : step (1, 2, 3) * 1)
which is desirable, because step then will be some other operation (hope
folding isn't performed yet).
> --- /dev/null
> +++ b/gcc/testsuite/gfortran.dg/gomp/linear-4.f90
> @@ -0,0 +1,102 @@
> +! { dg-do compile }
> +! { dg-options "-fopenmp" }
> +
> +module m
> +implicit none
> +
> +integer :: i
> +
> +interface
> + integer function bar (x, y, z)
> + integer :: x, y
> + integer, value :: z
> + !!$omp declare simd linear (x : ref, step (1)) linear (y : step (2),
> uval)
Are all these !! intentional?
The test then doesn't test much.
Or is that a FIXME?
Jakub
