Re: [patch] single_succ/pred

[Jeffrey A Law <law at redhat dot com>]

On Thu, 2005-03-10 at 17:07 +0100, Zdenek Dvorak wrote:
> Hello,
> 
> on many places we want to work only with basic blocks that have a single
> predecessor or successor, or access the single predecessor/successor of
> a block.  The idioms used to express these:
> 
> EDGE_COUNT (bb->succs) == 1
> EDGE_SUCC (bb, 0)->dest
> EDGE_PRED (bb, 0)->src
> 
> are somewhat cumbersome and the later two have the disadvantage of
> making the code harder to understand, since it is not possible to
> distinguish between access to a single successor of a block, or
> access to the first successor of a block with possibly more successors.
> 
> This patch introduces functions single_succ_p, single_succ and
> single_succ_edge to overcome these problems.
> 
> Bootstrapped & regtested on i686, x86_64, ia64 and ppc.
> 
> Zdenek
> 
> 	* basic-block.h (single_succ_p, single_pred_p, single_succ_edge,
> 	single_pred_edge, single_succ, single_pred): New inline functions.
> 	* bb-reorder.c (rotate_loop, find_traces_1_round,
> 	add_labels_and_missing_jumps, fix_up_fall_thru_edges,
> 	duplicate_computed_gotos): Use the single_succ/pred functions.
> 	* cfganal.c (forwarder_block_p): Ditto.
> 	* cfgbuild.c (compute_outgoing_frequencies): Ditto.
> 	* cfgcleanup.c (try_simplify_condjump, try_forward_edges,
> 	outgoing_edges_match, try_crossjump_to_edge, try_optimize_cfg,
> 	merge_seq_blocks): Ditto.
> 	* cfghooks.c (split_edge, tidy_fallthru_edges): Ditto.
> 	* cfglayout.c (fixup_reorder_chain): Ditto.
> 	* cfgloop.c (mark_single_exit_loops, update_latch_info,
> 	canonicalize_loop_headers, verify_loop_structure): Ditto.
> 	* cfgloopmanip.c (remove_path, unloop, loop_delete_branch_edge,
> 	mfb_update_loops, create_preheader, force_single_succ_latches,
> 	create_loop_notes): Ditto.
> 	* cfgrtl.c (rtl_can_merge_blocks, try_redirect_by_replacing_jump,
> 	force_nonfallthru_and_redirect, rtl_tidy_fallthru_edge,
> 	commit_one_edge_insertion, purge_dead_edges,
> 	cfg_layout_can_merge_blocks_p): Ditto.
> 	* except.c (sjlj_emit_function_enter): Ditto.
> 	* flow.c (init_propagate_block_info): Ditto.
> 	* function.c (thread_prologue_and_epilogue_insns): Ditto.
> 	* gcse.c (find_implicit_sets, bypass_conditional_jumps,
> 	insert_insn_end_bb): Ditto.
> 	* ifcvt.c (merge_if_block, find_if_block, find_if_case_1,
> 	find_if_case_2): Ditto.
> 	* lambda-code.c (perfect_nestify): Ditto.
> 	* lcm.c (optimize_mode_switching): Ditto.
> 	* loop-doloop.c (doloop_modify): Ditto.
> 	* loop-init.c (loop_optimizer_init): Ditto.
> 	* loop-iv.c (simplify_using_initial_values): Ditto.
> 	* loop-unroll.c (unroll_loop_runtime_iterations): Ditto.
> 	* loop-unswitch.c (unswitch_loop): Ditto.
> 	* modulo-sched.c (generate_prolog_epilog): Ditto.
> 	* predict.c (combine_predictions_for_insn, estimate_probability,
> 	tree_estimate_probability, last_basic_block_p,
> 	estimate_bb_frequencies): Ditto.
> 	* profile.c (branch_prob): Ditto.
> 	* regrename.c (copyprop_hardreg_forward): Ditto.
> 	* sched-rgn.c (is_cfg_nonregular, find_rgns, update_live): Ditto.
> 	* tracer.c (layout_superblocks): Ditto.
> 	* tree-cfg.c (tree_can_merge_blocks_p, tree_merge_blocks,
> 	cfg_remove_useless_stmts_bb, cleanup_control_flow,
> 	cleanup_control_expr_graph, disband_implicit_edges,
> 	tree_find_edge_insert_loc, bsi_commit_edge_inserts,
> 	tree_verify_flow_info, tree_make_forwarder_block,
> 	tree_forwarder_block_p, remove_forwarder_block,
> 	remove_forwarder_block_with_phi, merge_phi_nodes): Ditto.
> 	* tree-if-conv.c (tree_if_conversion): Ditto.
> 	* tree-mudflap.c (mf_build_check_statement_for): Ditto.
> 	* tree-ssa-dce.c (remove_dead_stmt): Ditto.
> 	* tree-ssa-dom.c (dom_opt_finalize_block): Ditto.
> 	* tree-ssa-loop-ch.c (should_duplicate_loop_header_p,
> 	copy_loop_headers): Ditto.
> 	* tree-ssa-loop-im.c (loop_commit_inserts): Ditto.
> 	* tree-ssa-loop-ivopts.c (compute_phi_arg_on_exit): Ditto.
> 	* tree-ssa-loop-manip.c (split_loop_exit_edge, ip_normal_pos,
> 	lv_adjust_loop_entry_edge, tree_ssa_loop_version): Ditto.
> 	* tree-ssa-loop-niter.c (simplify_using_initial_conditions): Ditto.
> 	* tree-ssa-loop-unswitch.c (simplify_using_entry_checks): Ditto.
> 	* tree-ssa-phiopt.c (tree_ssa_phiopt, value_replacement): Ditto.
> 	* tree-ssa-pre.c (compute_antic_aux, insert_aux, init_pre): Ditto.
> 	* tree-ssa-threadupdate.c (redirect_edges): Ditto.
> 	* tree-tailcall.c (independent_of_stmt_p, find_tail_calls,
> 	eliminate_tail_call, tree_optimize_tail_calls_1): Ditto.
> 	* tree-vect-analyze.c (vect_analyze_loop_form): Ditto.
> 	* tree-vect-transform.c (vect_update_ivs_after_vectorizer): Ditto.
> 	* tree-vectorizer.c (slpeel_update_phi_nodes_for_guard,
> 	slpeel_add_loop_guard): Ditto.
I only spot checked the changes to see if there were any obvious
goofs -- I didn't find any.  There were a couple places which initially
looked wrong, but further investigation showed the changes were correct.

What those sites have in common is checks for 0 or 1 preds/succs rather
than checking for precisely 1 pred/succ.  For example:

!         || EDGE_COUNT(EDGE_SUCC(bb,0)->dest->preds) <= 1)

The edge count can't be zero in this case.  So translating that
into the following is safe:

!         || single_pred_p (single_succ (bb)))


In other case we have:

!             if (EDGE_COUNT (first->succs) > 1)

Which got turned into:

!             if (!single_succ_p (first))


Without reading the nearby code we wouldn't know that first must
have at least one successor.

Others are of a similar nature.

Please install.  Thanks,

Jeff