Hi Cesar! On Fri, 18 Jul 2014 14:07:00 -0700, Cesar Philippidis <ce...@codesourcery.com> wrote: > This patch enables acc constructs to be used inside nested functions.
Thanks!
> I
> doubt nested functions will be used much in c, but some of the openacc
> fortran tutorials I've seen online make use of internal subroutines in
> fortran. Those internal subroutines are implemented as nested functions.
> Does this look OK to commit to gomp-4_0-branch?
I think we first need to resolve the following issue:
> --- a/gcc/gimple.h
> +++ b/gcc/gimple.h
> @@ -576,22 +576,6 @@ struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> tree data_arg;
> };
>
> -/* GIMPLE_OACC_KERNELS */
> -struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> - gimple_statement_oacc_kernels : public gimple_statement_omp_parallel_layout
> -{
> - /* No extra fields; adds invariant:
> - stmt->code == GIMPLE_OACC_KERNELS. */
> -};
> -
> -/* GIMPLE_OACC_PARALLEL */
> -struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> - gimple_statement_oacc_parallel : public
> gimple_statement_omp_parallel_layout
> -{
> - /* No extra fields; adds invariant:
> - stmt->code == GIMPLE_OACC_PARALLEL. */
> -};
> -
> /* GIMPLE_OMP_PARALLEL or GIMPLE_TASK */
> struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> gimple_statement_omp_taskreg : public gimple_statement_omp_parallel_layout
> @@ -617,6 +601,22 @@ struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> stmt->code == GIMPLE_OMP_TARGET. */
> };
>
> +/* GIMPLE_OACC_KERNELS */
> +struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> + gimple_statement_oacc_kernels : public gimple_statement_omp_taskreg
> +{
> + /* No extra fields; adds invariant:
> + stmt->code == GIMPLE_OACC_KERNELS. */
> +};
> +
> +/* GIMPLE_OACC_PARALLEL */
> +struct GTY((tag("GSS_OMP_PARALLEL_LAYOUT")))
> + gimple_statement_oacc_parallel : public gimple_statement_omp_taskreg
> +{
> + /* No extra fields; adds invariant:
> + stmt->code == GIMPLE_OACC_PARALLEL. */
> +};
> +
> /* GIMPLE_OMP_TASK */
>
> struct GTY((tag("GSS_OMP_TASK")))
> @@ -927,7 +927,8 @@ template <>
> inline bool
> is_a_helper <gimple_statement_omp_taskreg *>::test (gimple gs)
> {
> - return gs->code == GIMPLE_OMP_PARALLEL || gs->code == GIMPLE_OMP_TASK;
> + return gs->code == GIMPLE_OMP_PARALLEL || gs->code == GIMPLE_OMP_TASK
> + || gs->code == GIMPLE_OACC_PARALLEL || gs->code == GIMPLE_OACC_KERNELS;
> }
>
> template <>
> @@ -1135,7 +1136,8 @@ template <>
> inline bool
> is_a_helper <const gimple_statement_omp_taskreg *>::test (const_gimple gs)
> {
> - return gs->code == GIMPLE_OMP_PARALLEL || gs->code == GIMPLE_OMP_TASK;
> + return gs->code == GIMPLE_OMP_PARALLEL || gs->code == GIMPLE_OMP_TASK
> + || gs->code == GIMPLE_OACC_PARALLEL || gs->code == GIMPLE_OACC_KERNELS;
> }
Doing it this way has been "disapproved" before, and I raised the issue
in
<http://news.gmane.org/find-root.php?message_id=%3C87egxcjkmc.fsf%40kepler.schwinge.homeip.net%3E>.
If that makes sense to you, please continue this discussion until I'm
back next week.
> --- a/gcc/tree-nested.c
> +++ b/gcc/tree-nested.c
> @@ -1322,7 +1320,22 @@ convert_nonlocal_reference_stmt (gimple_stmt_iterator
> *gsi, bool *handled_ops_p,
>
> case GIMPLE_OACC_KERNELS:
> case GIMPLE_OACC_PARALLEL:
> - gcc_unreachable ();
> + save_suppress = info->suppress_expansion;
> + convert_nonlocal_omp_clauses (gimple_omp_taskreg_clauses_ptr (stmt),
> + wi);
> + save_local_var_chain = info->new_local_var_chain;
> + info->new_local_var_chain = NULL;
> +
> + walk_body (convert_nonlocal_reference_stmt,
> convert_nonlocal_reference_op,
> + info, gimple_omp_body_ptr (stmt));
> +
> + if (info->new_local_var_chain)
> + declare_vars (info->new_local_var_chain,
> + gimple_seq_first_stmt (gimple_omp_body (stmt)),
> + false);
> + info->new_local_var_chain = save_local_var_chain;
> + info->suppress_expansion = save_suppress;
> + break;
>
> case GIMPLE_OMP_PARALLEL:
> case GIMPLE_OMP_TASK:
Wouldn't we rather group the GIMPLE_OACC_* with GIMPLE_OMP_TARGET
(handling all these together)? (Either way, please avoid the code
duplication for GIMPLE_OACC_KERNELS/GIMPLE_OACC_PARALLEL even if grouping
with GIMPLE_OMP_PARALLEL/GIMPLE_OMP_TASK; I think the latter's existing
code can basically be re-used as is?
Grüße,
Thomas
pgp5PQ38mZ5g7.pgp
Description: PGP signature
