This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch][graphite] Rewrite of "out of graphite"
- From: Richard Guenther <rguenther at suse dot de>
- To: Sebastian Pop <sebpop at gmail dot com>
- Cc: GCC Patches <gcc-patches at gcc dot gnu dot org>, Diego Novillo <dnovillo at google dot com>, Daniel Berlin <dberlin at dberlin dot org>, "Sjodin, Jan" <Jan dot Sjodin at amd dot com>, "Jagasia, Harsha" <harsha dot jagasia at amd dot com>, Christophe Harle <christophe dot harle at amd dot com>, Tobias Grosser <grosser at fim dot uni-passau dot de>, Albert Cohen <Albert dot Cohen at inria dot fr>, Konrad Trifunovic <konrad dot trifunovic at gmail dot com>
- Date: Wed, 3 Dec 2008 11:22:19 +0100 (CET)
- Subject: Re: [patch][graphite] Rewrite of "out of graphite"
- References: <cb9d34b20812022354w2a41bcfeg796ca952c7006e40@mail.gmail.com>
On Wed, 3 Dec 2008, Sebastian Pop wrote:
> Hi,
>
> I committed the attached patch to the graphite branch: it fixes all
> the problems that were reported against the graphite code generation:
>
> PR middle-end/37852
> PR middle-end/37883
> PR middle-end/37928
> PR middle-end/37980
> PR middle-end/38038
> PR middle-end/38039
> PR middle-end/38073
> PR middle-end/38083
> PR middle-end/38125
>
> I have reduced all the bugs left in the polyhedron benchmarks: they do
> not occur in the code generation part. I separately committed all the
> reduced testcases to the graphite branch such that we can address them
> more easily. These are the following testcases:
>
> * testsuite/gcc.dg/graphite/pr38084.c: New.
> * testsuite/gfortran.dg/graphite/id-1.f90: New.
> * testsuite/gfortran.dg/graphite/id-2.f90: New.
> * testsuite/gfortran.dg/graphite/id-3.f90: New.
> * testsuite/gfortran.dg/graphite/id-4.f90: New.
> * testsuite/gfortran.dg/graphite/pr37857.f90: New.
>
> I'm testing the attached patch on amd64-linux. Okay to commit to
> trunk without splitting the patch into smaller pieces: all the pieces
> are interconnected, and splitting the patch in smaller pieces is
> a painful experience already (yes, I have tried).
The middle-end parts could have been splitted ;) The middle-end parts
are ok, the graphite parts as well, but you may want to wait for more
comments from other graphite reviewers.
Thanks,
Richard.
>
> Thanks,
> Sebastian Pop
> --
> AMD - GNU Tools
>
>
> 2008-12-02 Sebastian Pop <sebastian.pop@amd.com>
> Jan Sjodin <jan.sjodin@amd.com>
> Harsha Jagasia <harsha.jagasia@amd.com>
>
> PR middle-end/37852
> PR middle-end/37883
> PR middle-end/37928
> PR middle-end/37980
> PR middle-end/38038
> PR middle-end/38039
> PR middle-end/38073
> PR middle-end/38083
> PR middle-end/38125
>
> * testsuite/gcc.dg/graphite/pr38073.c: New.
> * testsuite/gcc.dg/graphite/pr37928.c: New.
> * testsuite/gcc.dg/graphite/pr37883.c: New.
> * testsuite/gcc.dg/graphite/pr38125.c: New.
> * testsuite/gfortran.dg/graphite/pr38083.f90: New.
> * testsuite/gfortran.dg/graphite/pr37852.f90: New.
> * testsuite/gfortran.dg/graphite/pr37980.f90: New.
>
> * testsuite/gcc.dg/graphite/scop-18.c: Remove reduction, test for
> the number of detected scops. Copy exact same test for loop blocking...
> * testsuite/gcc.dg/graphite/block-1.c: Fix the number of expected loops
> to be blocked as reductions are not handled.
> * testsuite/gcc.dg/graphite/block-4.c: ...here. New.
>
> * tree-phinodes.c (remove_phi_nodes): New, extracted from...
> * tree-cfg.c (remove_phi_nodes_and_edges_for_unreachable_block): ...here.
> * tree-flow.h (remove_phi_nodes, canonicalize_loop_ivs): Declared.
> * Makefile.in (graphite.o): Depend on value-prof.h.
> (graphite.o-warn): Removed -Wno-error.
> * tree-parloops.c (canonicalize_loop_ivs): Allow reduction_list
> to be a NULL pointer. Call update_stmt. Return the newly created
> cannonical induction variable.
>
> * graphite.h (debug_rename_map): Declared. Fix some comments.
>
> * graphite.c: Reimplement the code generation from graphite to gimple.
> Include value-prof.h.
> (loop_iv_stack_get_iv): Do not return NULL for constant substitutions.
> (get_old_iv_from_ssa_name): Removed.
> (graphite_stmt_p): New.
> (new_graphite_bb): Test for useful statements before building a
> graphite statement for the basic block.
> (free_graphite_bb): Do not free GBB_DATA_REFS: this is a bug
> in free_data_ref that calls BITMAP_FREE (DR_VOPS (dr)) without
> reason.
> (recompute_all_dominators, graphite_verify,
> nb_reductions_in_loop, graphite_loop_normal_form): New.
> (scop_record_loop): Call graphite_loop_normal_form.
> (build_scop_loop_nests): Iterate over all the blocks of the
> function instead of relying on the incomplete information from
> SCOP_BBS. Return the success of the operation.
> (find_params_in_bb): Use the data from GBB_DATA_REFS.
> (add_bb_domains): Removed.
> (build_loop_iteration_domains): Don't call add_bb_domains.
> Add the iteration domain only to the basic blocks that have been
> translated to graphite.
> (build_scop_conditions_1): Add constraints only if the basic
> block have been translated to graphite.
> (build_scop_data_accesses): Completely disabled until data
> dependence is correctly implemented.
> (debug_rename_elt, debug_rename_map_1, debug_rename_map): New.
> (remove_all_edges_1, remove_all_edges): Removed.
> (get_new_name_from_old_name): New.
> (graphite_rename_variables_in_stmt): Renamed
> rename_variables_in_stmt. Call get_new_name_from_old_name.
> Use replace_exp and update_stmt.
> (is_old_iv): Renamed is_iv.
> (expand_scalar_variables_stmt): Extra parameter for renaming map.
> Use replace_exp and update_stmt.
> (expand_scalar_variables_expr): Same. Use the map to get the
> new names for the renaming of induction variables and for the
> renaming of variables after a basic block has been copied.
> (expand_scalar_variables): Same.
> (graphite_rename_variables): Renamed rename_variables.
> (move_phi_nodes): Removed.
> (get_false_edge_from_guard_bb): New.
> (build_iv_mapping): Do not insert the induction variable of a
> loop in the renaming iv map if the basic block does not belong
> to that loop.
> (register_old_new_names, graphite_copy_stmts_from_block,
> copy_bb_and_scalar_dependences): New.
> (translate_clast): Heavily reimplemented: copy basic blocks,
> do not move them. Finally, in call cleanup_tree_cfg in gloog.
> At each translation step call graphite_verify ensuring the
> consistency of the SSA, loops and dominators information.
> (collect_virtual_phis, find_vdef_for_var_in_bb,
> find_vdef_for_var_1, find_vdef_for_var,
> patch_phis_for_virtual_defs): Removed huge hack.
> (mark_old_loops, remove_dead_loops, skip_phi_defs,
> collect_scop_exit_phi_args, patch_scop_exit_phi_args,
> gbb_can_be_ignored, scop_remove_ignoreable_gbbs, ): Removed.
> (remove_sese_region, ifsese, if_region_entry, if_region_exit,
> if_region_get_condition_block, if_region_set_false_region,
> create_if_region_on_edge, move_sese_in_condition, bb_in_sese_p,
> sese_find_uses_to_rename_use, sese_find_uses_to_rename_bb,
> sese_add_exit_phis_edge, sese_add_exit_phis_var,
> rewrite_into_sese_closed_ssa): New.
> (gloog): Remove dead code. Early return if code cannot be
> generated. Call cleanup_tree_cfg once the scop has been code
> generated.
> (graphite_trans_scop_block, graphite_trans_loop_block): Do not
> block loops with less than two loops.
> (graphite_apply_transformations): Remove the call to
> scop_remove_ignoreable_gbbs.
> (limit_scops): When build_scop_loop_nests fails, continue on next scop.
> Fix open_scop.entry.
> (graphite_transform_loops): Call recompute_all_dominators: force the
> recomputation of correct CDI_DOMINATORS and CDI_POST_DOMINATORS.
> Call initialize_original_copy_tables and free_original_copy_tables
> to be able to copy basic blocks during code generation.
> When build_scop_loop_nests fails, continue on next scop.
> (value_clast): New union.
> (clast_to_gcc_expression): Fix type cast warning.
>
--
Richard Guenther <rguenther@suse.de>
Novell / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746 - GF: Markus Rex