Hi, On Tue, Dec 13, 2016 at 11:35:37AM +0100, Jakub Jelinek wrote: > On Tue, Dec 13, 2016 at 11:15:43AM +0100, Martin Jambor wrote: > > I have bootstrapped the two patches on aarch64-linux and bootstrapped > > and tested them on x86_64-linux. What do you think? > > Thanks a lot for the work. If you wouldn't mind doing a couple of further > changes (see below), I'd appreciate it, but if you want to commit it right > away, I'm ok with that too. I have rebased the patches on current trunk and: 1) updated the first patch to rename omp-device.[ch] to omp-offload.[ch] as suggested by Thomas and fixed the mistakes in ChangeLog discovered by Alex, and 2) updated the second patch to deal with most of the issues below and to remove spaces before tabs that Alex found. I am bootstrapping the result now and because they have been approved and since it seems that a number of people have had a chance to look at and even test them, I will commit them after the bootstrap and testing finishes. The updated patches are attached, ChangeLogs are at the end of this message. > > @@ -4321,7 +4322,7 @@ expand_cilk_for (struct omp_region *region, struct omp_for_data *fd) > > tree child_fndecl > = gimple_omp_parallel_child_fn ( > - as_a (last_stmt (region->outer->entry))); > + as_a (last_stmt (region->outer->entry))); > tree t, low_val = NULL_TREE, high_val = NULL_TREE; > for (t = DECL_ARGUMENTS (child_fndecl); t; t = TREE_CHAIN (t)) > { > > My preference for the above would be > > gomp_parallel *par_stmt > = as_a (last_stmt (region->outer->entry)); > tree child_fndecl = gimple_omp_parallel_child_fn (par_stmt); OK, I have changed it so. > > @@ -6428,7 +6429,7 @@ expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb, > floating point. This allows the atomic operation to properly > succeed even with NaNs and -0.0. */ > stmt = gimple_build_cond_empty > - (build2 (NE_EXPR, boolean_type_node, > + (build2 (NE_EXPR, boolean_type_node, > new_storedi, old_vali)); > gsi_insert_before (&si, stmt, GSI_SAME_STMT); > > And here > > tree ne = build2 (NE_EXPR, boolean_type_node, new_storedi, old_vali); > stmt = gimple_build_cond_empty (ne); Likewise. > > @@ -442,7 +442,7 @@ omp_max_simt_vf (void) > if (!optimize) > return 0; > if (ENABLE_OFFLOADING) > - for (const char *c = getenv ("OFFLOAD_TARGET_NAMES"); c; ) > + for (const char *c = getenv ("OFFLOAD_TARGET_NAMES"); c;) > > I admit I don't know what the coding style we have in this case. Ok > either way. I have made this change in order to silence a complaint from check_GNU_style.sh. I think the only sane way to write the above is as a while loop :-) But I have left it as it was, I really wanted to be changing things only slightly. > > @@ -5860,7 +5860,7 @@ lower_omp_sections (gimple_stmt_iterator *gsi_p, omp_context *ctx) > new_body = maybe_catch_exception (new_body); > > t = gimple_build_omp_return > - (!!omp_find_clause (gimple_omp_sections_clauses (stmt), > + (!!omp_find_clause (gimple_omp_sections_clauses (stmt), > OMP_CLAUSE_NOWAIT)); > gimple_seq_add_stmt (&new_body, t); > maybe_add_implicit_barrier_cancel (ctx, &new_body); > @@ -6023,7 +6023,7 @@ lower_omp_single (gimple_stmt_iterator *gsi_p, omp_context *ctx) > bind_body = maybe_catch_exception (bind_body); > > t = gimple_build_omp_return > - (!!omp_find_clause (gimple_omp_single_clauses (single_stmt), > + (!!omp_find_clause (gimple_omp_single_clauses (single_stmt), > OMP_CLAUSE_NOWAIT)); > gimple_seq_add_stmt (&bind_body_tail, t); > maybe_add_implicit_barrier_cancel (ctx, &bind_body_tail); > > And in the above 2 spots something like: > bool nowait = omp_find_clause (gimple_omp_single_clauses (single_stmt), > OMP_CLAUSE_NOWAIT) != NULL_TREE; > t = gimple_build_omp_return (nowait); I have changed it so too. > > (wonder why we use t for gimple, renaming t var to g would be also nice). But > ok either way. I have changed the second occurrence to use g because it was the only use of t, but left t in the first one because it had other uses. > > @@ -8668,7 +8669,7 @@ lower_omp_1 (gimple_stmt_iterator *gsi_p, omp_context *ctx) > break; > case GIMPLE_TRANSACTION: > lower_omp (gimple_transaction_body_ptr ( > - as_a (stmt)), > + as_a (stmt)), > ctx); > > Here it fits, so no need to wrap: > lower_omp (gimple_transaction_body_ptr (as_a (stmt)), > ctx); > > I have changed this as you requested too. Thanks, Martin 2016-12-14 Martin Jambor * omp-general.h: New file. * omp-general.c: New file. * omp-expand.h: Likewise. * omp-expand.c: Likewise. * omp-offload.h: Likewise. * omp-offload.c: Likewise. * omp-grid.c: Likewise. * omp-grid.c: Likewise. * omp-low.h: Include omp-general.h and omp-grid.h. Removed includes of params.h, symbol-summary.h, lto-section-names.h, cilk.h, tree-eh.h, ipa-prop.h, tree-cfgcleanup.h, cfgloop.h, except.h, expr.h, stmt.h, varasm.h, calls.h, explow.h, dojump.h, flags.h, tree-into-ssa.h, tree-cfg.h, cfganal.h, alias.h, emit-rtl.h, optabs.h, expmed.h, alloc-pool.h, cfghooks.h, rtl.h and memmodel.h. (omp_find_combined_for): Declare. (find_omp_clause): Renamed to omp_find_clause and moved to omp-general.h. (free_omp_regions): Renamed to omp_free_regions and moved to omp-expand.h. (replace_oacc_fn_attrib): Renamed to oacc_replace_fn_attrib and moved to omp-general.h. (set_oacc_fn_attrib): Renamed to oacc_set_fn_attrib and moved to omp-general.h. (build_oacc_routine_dims): Renamed to oacc_build_routine_dims and moved to omp-general.h. (get_oacc_fn_attrib): Renamed to oacc_get_fn_attrib and moved to omp-general.h. (oacc_fn_attrib_kernels_p): Moved to omp-general.h. (get_oacc_fn_dim_size): Renamed to oacc_get_fn_dim_size and moved to omp-general.c. (omp_expand_local): Moved to omp-expand.h. (make_gimple_omp_edges): Renamed to omp_make_gimple_edges and moved to omp-expand.h. (omp_finish_file): Moved to omp-offload.h. (default_goacc_validate_dims): Renamed to oacc_default_goacc_validate_dims and moved to omp-offload.h. (offload_funcs, offload_vars): Moved to omp-offload.h. * omp-low.c: Include omp-general.h, omp-offload.h and omp-grid.h. (omp_region): Moved to omp-expand.c. (omp_for_data_loop): Moved to omp-general.h. (omp_for_data): Likewise. (oacc_loop): Moved to omp-offload.c. (oacc_loop_flags): Moved to omp-general.h. (offload_funcs, offload_vars): Moved to omp-offload.c. (root_omp_region): Moved to omp-expand.c. (omp_any_child_fn_dumped): Likewise. (find_omp_clause): Renamed to omp_find_clause and moved to omp-general.c. (is_combined_parallel): Moved to omp-expand.c. (is_reference): Renamed to omp_is_reference and and moved to omp-general.c. (adjust_for_condition): Renamed to omp_adjust_for_condition and moved to omp-general.c. (get_omp_for_step_from_incr): Renamed to omp_get_for_step_from_incr and moved to omp-general.c. (extract_omp_for_data): Renamed to omp_extract_for_data and moved to omp-general.c. (workshare_safe_to_combine_p): Moved to omp-expand.c. (omp_adjust_chunk_size): Likewise. (get_ws_args_for): Likewise. (get_base_type): Removed. (dump_omp_region): Moved to omp-expand.c. (debug_omp_region): Likewise. (debug_all_omp_regions): Likewise. (new_omp_region): Likewise. (free_omp_region_1): Likewise. (free_omp_regions): Renamed to omp_free_regions and moved to omp-expand.c. (find_combined_for): Renamed to omp_find_combined_for, made global. (build_omp_barrier): Renamed to omp_build_barrier and moved to omp-general.c. (omp_max_vf): Moved to omp-general.c. (omp_max_simt_vf): Likewise. (gimple_build_cond_empty): Moved to omp-expand.c. (parallel_needs_hsa_kernel_p): Likewise. (expand_omp_build_assign): Moved declaration to omp-expand.c. (expand_parallel_call): Moved to omp-expand.c. (expand_cilk_for_call): Likewise. (expand_task_call): Likewise. (vec2chain): Likewise. (remove_exit_barrier): Likewise. (remove_exit_barriers): Likewise. (optimize_omp_library_calls): Likewise. (expand_omp_regimplify_p): Likewise. (expand_omp_build_assign): Likewise. (expand_omp_taskreg): Likewise. (oacc_collapse): Likewise. (expand_oacc_collapse_init): Likewise. (expand_oacc_collapse_vars): Likewise. (expand_omp_for_init_counts): Likewise. (expand_omp_for_init_vars): Likewise. (extract_omp_for_update_vars): Likewise. (expand_omp_ordered_source): Likewise. (expand_omp_ordered_sink): Likewise. (expand_omp_ordered_source_sink): Likewise. (expand_omp_for_ordered_loops): Likewise. (expand_omp_for_generic): Likewise. (expand_omp_for_static_nochunk): Likewise. (find_phi_with_arg_on_edge): Likewise. (expand_omp_for_static_chunk): Likewise. (expand_cilk_for): Likewise. (expand_omp_simd): Likewise. (expand_omp_taskloop_for_outer): Likewise. (expand_omp_taskloop_for_inner): Likewise. (expand_oacc_for): Likewise. (expand_omp_for): Likewise. (expand_omp_sections): Likewise. (expand_omp_single): Likewise. (expand_omp_synch): Likewise. (expand_omp_atomic_load): Likewise. (expand_omp_atomic_store): Likewise. (expand_omp_atomic_fetch_op): Likewise. (expand_omp_atomic_pipeline): Likewise. (expand_omp_atomic_mutex): Likewise. (expand_omp_atomic): Likewise. (oacc_launch_pack): and moved to omp-general.c, made public. (OACC_FN_ATTRIB): Likewise. (replace_oacc_fn_attrib): Renamed to oacc_replace_fn_attrib and moved to omp-general.c. (set_oacc_fn_attrib): Renamed to oacc_set_fn_attrib and moved to omp-general.c. (build_oacc_routine_dims): Renamed to oacc_build_routine_dims and moved to omp-general.c. (get_oacc_fn_attrib): Renamed to oacc_get_fn_attrib and moved to omp-general.c. (oacc_fn_attrib_kernels_p): Moved to omp-general.c. (oacc_fn_attrib_level): Moved to omp-offload.c. (get_oacc_fn_dim_size): Renamed to oacc_get_fn_dim_size and moved to omp-general.c. (get_oacc_ifn_dim_arg): Renamed to oacc_get_ifn_dim_arg and moved to omp-general.c. (mark_loops_in_oacc_kernels_region): Moved to omp-expand.c. (grid_launch_attributes_trees): Likewise. (grid_attr_trees): Likewise. (grid_create_kernel_launch_attr_types): Likewise. (grid_insert_store_range_dim): Likewise. (grid_get_kernel_launch_attributes): Likewise. (get_target_argument_identifier_1): Likewise. (get_target_argument_identifier): Likewise. (get_target_argument_value): Likewise. (push_target_argument_according_to_value): Likewise. (get_target_arguments): Likewise. (expand_omp_target): Likewise. (grid_expand_omp_for_loop): Moved to omp-grid.c. (grid_arg_decl_map): Likewise. (grid_remap_kernel_arg_accesses): Likewise. (grid_expand_target_grid_body): Likewise. (expand_omp): Renamed to omp_expand and moved to omp-expand.c. (build_omp_regions_1): Moved to omp-expand.c. (build_omp_regions_root): Likewise. (omp_expand_local): Likewise. (build_omp_regions): Likewise. (execute_expand_omp): Likewise. (pass_data_expand_omp): Likewise. (pass_expand_omp): Likewise. (make_pass_expand_omp): Likewise. (pass_data_expand_omp_ssa): Likewise. (pass_expand_omp_ssa): Likewise. (make_pass_expand_omp_ssa): Likewise. (grid_lastprivate_predicate): Renamed to omp_grid_lastprivate_predicate and moved to omp-grid.c, made public. (grid_prop): Moved to omp-grid.c. (GRID_MISSED_MSG_PREFIX): Likewise. (grid_safe_assignment_p): Likewise. (grid_seq_only_contains_local_assignments): Likewise. (grid_find_single_omp_among_assignments_1): Likewise. (grid_find_single_omp_among_assignments): Likewise. (grid_find_ungridifiable_statement): Likewise. (grid_parallel_clauses_gridifiable): Likewise. (grid_inner_loop_gridifiable_p): Likewise. (grid_dist_follows_simple_pattern): Likewise. (grid_gfor_follows_tiling_pattern): Likewise. (grid_call_permissible_in_distribute_p): Likewise. (grid_handle_call_in_distribute): Likewise. (grid_dist_follows_tiling_pattern): Likewise. (grid_target_follows_gridifiable_pattern): Likewise. (grid_remap_prebody_decls): Likewise. (grid_var_segment): Likewise. (grid_mark_variable_segment): Likewise. (grid_copy_leading_local_assignments): Likewise. (grid_process_grid_body): Likewise. (grid_eliminate_combined_simd_part): Likewise. (grid_mark_tiling_loops): Likewise. (grid_mark_tiling_parallels_and_loops): Likewise. (grid_process_kernel_body_copy): Likewise. (grid_attempt_target_gridification): Likewise. (grid_gridify_all_targets_stmt): Likewise. (grid_gridify_all_targets): Renamed to omp_grid_gridify_all_targets and moved to omp-grid.c, made public. (make_gimple_omp_edges): Renamed to omp_make_gimple_edges and moved to omp-expand.c. (add_decls_addresses_to_decl_constructor): Moved to omp-offload.c. (omp_finish_file): Likewise. (oacc_thread_numbers): Likewise. (oacc_xform_loop): Likewise. (oacc_default_dims, oacc_min_dims): Likewise. (oacc_parse_default_dims): Likewise. (oacc_validate_dims): Likewise. (new_oacc_loop_raw): Likewise. (new_oacc_loop_outer): Likewise. (new_oacc_loop): Likewise. (new_oacc_loop_routine): Likewise. (finish_oacc_loop): Likewise. (free_oacc_loop): Likewise. (dump_oacc_loop_part): Likewise. (dump_oacc_loop): Likewise. (debug_oacc_loop): Likewise. (oacc_loop_discover_walk): Likewise. (oacc_loop_sibling_nreverse): Likewise. (oacc_loop_discovery): Likewise. (oacc_loop_xform_head_tail): Likewise. (oacc_loop_xform_loop): Likewise. (oacc_loop_process): Likewise. (oacc_loop_fixed_partitions): Likewise. (oacc_loop_auto_partitions): Likewise. (oacc_loop_partition): Likewise. (default_goacc_fork_join): Likewise. (default_goacc_reduction): Likewise. (execute_oacc_device_lower): Likewise. (default_goacc_validate_dims): Likewise. (default_goacc_dim_limit): Likewise. (pass_data_oacc_device_lower): Likewise. (pass_oacc_device_lower): Likewise. (make_pass_oacc_device_lower): Likewise. (execute_omp_device_lower): Likewise. (pass_data_omp_device_lower): Likewise. (pass_omp_device_lower): Likewise. (make_pass_omp_device_lower): Likewise. (pass_data_omp_target_link): Likewise. (pass_omp_target_link): Likewise. (find_link_var_op): Likewise. (pass_omp_target_link::execute): Likewise. (make_pass_omp_target_link): Likewise. * Makefile.in (OBJS): Added omp-offload.o, omp-expand.o, omp-general.o and omp-grid.o. (GTFILES): Added omp-offload.h, omp-offload.c and omp-expand.c, removed omp-low.h. * gimple-fold.c: Include omp-general.h instead of omp-low.h. (fold_internal_goacc_dim): Adjusted calls to get_oacc_ifn_dim_arg and get_oacc_fn_dim_size to use their new names. * gimplify.c: Include omp-low.h. (omp_notice_variable): Adjust the call to get_oacc_fn_attrib to use its new name. (gimplify_omp_task): Adjusted calls to find_omp_clause to use its new name. (gimplify_omp_for): Likewise. * lto-cgraph.c: Include omp-offload.h instead of omp-low.h. * toplev.c: Include omp-offload.h instead of omp-low.h. * tree-cfg.c: Include omp-general.h instead of omp-low.h. Also include omp-expand.h. (make_edges_bb): Adjusted the call to make_gimple_omp_edges to use its new name. (make_edges): Adjust the call to free_omp_regions to use its new name. * tree-parloops.c: Include omp-general.h. (create_parallel_loop): Adjusted the call to set_oacc_fn_attrib to use its new name. (parallelize_loops): Adjusted the call to get_oacc_fn_attrib to use its new name. * tree-ssa-loop.c: Include omp-general.h instead of omp-low.h. (gate_oacc_kernels): Adjusted the call to get_oacc_fn_attrib to use its new name. * tree-vrp.c: Include omp-general.h instead of omp-low.h. (extract_range_basic): Adjusted calls to get_oacc_ifn_dim_arg and get_oacc_fn_dim_size to use their new names. * varpool.c: Include omp-offload.h instead of omp-low.h. * gengtype.c (open_base_files): Replace omp-low.h with omp-offload.h in ifiles. * config/nvptx/nvptx.c: Include omp-general.c. (nvptx_expand_call): Adjusted the call to get_oacc_fn_attrib to use its new name. (nvptx_reorg): Likewise. (nvptx_record_offload_symbol): Likewise. gcc/c-family: * c-omp.c: Include omp-general.h instead of omp-low.h. (c_finish_oacc_wait): Adjusted call to find_omp_clause to use its new name. gcc/c/ * c-parser.c: Include omp-general.h and omp-offload.h instead of omp-low.h. (c_finish_oacc_routine): Adjusted call to get_oacc_fn_attrib, build_oacc_routine_dims and replace_oacc_fn_attrib to use their new names. (c_parser_oacc_enter_exit_data): Adjusted call to find_omp_clause to use its new name. (c_parser_oacc_update): Likewise. (c_parser_omp_simd): Likewise. (c_parser_omp_target_update): Likewise. * c-typeck.c: Include omp-general.h instead of omp-low.h. (c_finish_omp_cancel): Adjusted call to find_omp_clause to use its new name. (c_finish_omp_cancellation_point): Likewise. * gimple-parser.c: Do not include omp-low.h gcc/cp/ * parser.c: Include omp-general.h and omp-offload.h instead of omp-low.h. (cp_parser_omp_simd): Adjusted calls to find_omp_clause to use its new name. (cp_parser_omp_target_update): Likewise. (cp_parser_oacc_declare): Likewise. (cp_parser_oacc_enter_exit_data): Likewise. (cp_parser_oacc_update): Likewise. (cp_finalize_oacc_routine): Adjusted call to get_oacc_fn_attrib, build_oacc_routine_dims and replace_oacc_fn_attrib to use their new names. * semantics.c: Include omp-general insteda of omp-low.h. (finish_omp_for): Adjusted calls to find_omp_clause to use its new name. (finish_omp_cancel): Likewise. (finish_omp_cancellation_point): Likewise. fortran/ * trans-openmp.c: Include omp-general.h. 2016-12-14 Martin Jambor * omp-device.c: Fix coding style. * omp-expand.c: Likewise. * omp-general.c: Likewise. * omp-grid.c: Likewise. * omp-low.c: Fix coding style of parts touched by the previous splitting patch.