X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;ds=sidebyside;f=ir%2Fir%2Firgopt.c;h=1b280f2ec1caa123e86b0d0daadd2f8a258fbc62;hb=bd019d8cd785c1a433c19dee44bf3f9e82397286;hp=7f7d6d7dfd3b36218393afcc6a2b8efc782cb791;hpb=ddfcdcb1227bd6fb6720f8485dc62995f7bfd98f;p=libfirm diff --git a/ir/ir/irgopt.c b/ir/ir/irgopt.c index 7f7d6d7df..1b280f2ec 100644 --- a/ir/ir/irgopt.c +++ b/ir/ir/irgopt.c @@ -22,7 +22,6 @@ * @brief Optimizations for a whole ir graph, i.e., a procedure. * @author Christian Schaefer, Goetz Lindenmaier, Sebastian Felis, * Michael Beck - * @version $Id$ */ #include "config.h" @@ -51,7 +50,8 @@ /** * A wrapper around optimize_inplace_2() to be called from a walker. */ -static void optimize_in_place_wrapper(ir_node *n, void *env) { +static void optimize_in_place_wrapper(ir_node *n, void *env) +{ ir_node *optimized = optimize_in_place_2(n); (void) env; @@ -68,26 +68,27 @@ static void optimize_in_place_wrapper(ir_node *n, void *env) { * * @note current_ir_graph must be set */ -static inline void do_local_optimize(ir_node *n) { +static inline void do_local_optimize(ir_node *n) +{ + ir_graph *irg = get_irn_irg(n); + /* Handle graph state */ - assert(get_irg_phase_state(current_ir_graph) != phase_building); + assert(get_irg_phase_state(irg) != phase_building); if (get_opt_global_cse()) - set_irg_pinned(current_ir_graph, op_pin_state_floats); - set_irg_outs_inconsistent(current_ir_graph); - set_irg_doms_inconsistent(current_ir_graph); - set_irg_loopinfo_inconsistent(current_ir_graph); + set_irg_pinned(irg, op_pin_state_floats); + clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE); /* Clean the value_table in irg for the CSE. */ - del_identities(current_ir_graph->value_table); - current_ir_graph->value_table = new_identities(); + new_identities(irg); /* walk over the graph */ irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL); } /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */ -void local_optimize_node(ir_node *n) { +void local_optimize_node(ir_node *n) +{ ir_graph *rem = current_ir_graph; current_ir_graph = get_irn_irg(n); @@ -96,49 +97,38 @@ void local_optimize_node(ir_node *n) { current_ir_graph = rem; } -/** - * Block-Walker: uses dominance depth to mark dead blocks. - */ -static void kill_dead_blocks(ir_node *block, void *env) { - (void) env; - - if (get_Block_dom_depth(block) < 0) { - /* - * Note that the new dominance code correctly handles - * the End block, i.e. it is always reachable from Start - */ - set_Block_dead(block); - } -} - -/* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */ -void local_optimize_graph(ir_graph *irg) { - ir_graph *rem = current_ir_graph; - current_ir_graph = irg; - - if (get_irg_dom_state(irg) == dom_consistent) - irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL); - - do_local_optimize(get_irg_end(irg)); - - current_ir_graph = rem; +static void enqueue_node(ir_node *node, pdeq *waitq) +{ + if (get_irn_link(node) == waitq) + return; + pdeq_putr(waitq, node); + set_irn_link(node, waitq); } /** * Enqueue all users of a node to a wait queue. * Handles mode_T nodes. */ -static void enqueue_users(ir_node *n, pdeq *waitq) { +static void enqueue_users(ir_node *n, pdeq *waitq) +{ const ir_edge_t *edge; foreach_out_edge(n, edge) { - ir_node *succ = get_edge_src_irn(edge); + ir_node *succ = get_edge_src_irn(edge); + const ir_edge_t *edge2; - if (get_irn_link(succ) != waitq) { - pdeq_putr(waitq, succ); - set_irn_link(succ, waitq); - } - if (get_irn_mode(succ) == mode_T) { + enqueue_node(succ, waitq); + + /* Also enqueue Phis to prevent inconsistencies. */ + if (is_Block(succ)) { + foreach_out_edge(succ, edge2) { + ir_node *succ2 = get_edge_src_irn(edge2); + + if (is_Phi(succ2)) { + enqueue_node(succ2, waitq); + } + } + } else if (get_irn_mode(succ) == mode_T) { /* A mode_T node has Proj's. Because most optimizations run on the Proj's we have to enqueue them also. */ enqueue_users(succ, waitq); @@ -146,13 +136,51 @@ static void enqueue_users(ir_node *n, pdeq *waitq) { } } +/** + * Block-Walker: uses dominance depth to mark dead blocks. + */ +static void find_unreachable_blocks(ir_node *block, void *env) +{ + pdeq *waitq = (pdeq*) env; + + if (get_Block_dom_depth(block) < 0) { + ir_graph *irg = get_irn_irg(block); + ir_node *end = get_irg_end(irg); + + const ir_edge_t *edge; + foreach_block_succ(block, edge) { + const ir_edge_t *edge2; + ir_node *succ_block = get_edge_src_irn(edge); + enqueue_node(succ_block, waitq); + foreach_out_edge(succ_block, edge2) { + ir_node *succ = get_edge_src_irn(edge2); + if (is_Phi(succ)) + enqueue_node(succ, waitq); + } + } + enqueue_node(end, waitq); + } +} + +/* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */ +void local_optimize_graph(ir_graph *irg) +{ + ir_graph *rem = current_ir_graph; + current_ir_graph = irg; + + do_local_optimize(get_irg_end(irg)); + + current_ir_graph = rem; +} + /** * Data flow optimization walker. - * Optimizes all nodes and enqueue it's users + * Optimizes all nodes and enqueue its users * if done. */ -static void opt_walker(ir_node *n, void *env) { - pdeq *waitq = env; +static void opt_walker(ir_node *n, void *env) +{ + pdeq *waitq = (pdeq*)env; ir_node *optimized; optimized = optimize_in_place_2(n); @@ -164,66 +192,69 @@ static void opt_walker(ir_node *n, void *env) { } } -/* Applies local optimizations to all nodes in the graph until fixpoint. */ -int optimize_graph_df(ir_graph *irg) { +int optimize_graph_df(ir_graph *irg) +{ pdeq *waitq = new_pdeq(); ir_graph *rem = current_ir_graph; ir_node *end; - int state, changed; current_ir_graph = irg; - state = edges_assure(irg); - if (get_opt_global_cse()) - set_irg_pinned(current_ir_graph, op_pin_state_floats); + set_irg_pinned(irg, op_pin_state_floats); - /* Clean the value_table in irg for the CSE. */ - del_identities(irg->value_table); - irg->value_table = new_identities(); + /* enable unreachable code elimination */ + assert(!is_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE)); + set_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE); - if (get_irg_dom_state(irg) == dom_consistent) - irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL); + new_identities(irg); + edges_assure(irg); + assure_doms(irg); - /* invalidate info */ - set_irg_outs_inconsistent(irg); - set_irg_doms_inconsistent(irg); - set_irg_loopinfo_inconsistent(irg); ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK); - - /* walk over the graph, but don't touch keep-alives */ irg_walk_graph(irg, NULL, opt_walker, waitq); /* any optimized nodes are stored in the wait queue, * so if it's not empty, the graph has been changed */ - changed = !pdeq_empty(waitq); - - /* finish the wait queue */ - while (! pdeq_empty(waitq)) { - ir_node *n = pdeq_getl(waitq); - if (! is_Bad(n)) + while (!pdeq_empty(waitq)) { + /* finish the wait queue */ + while (! pdeq_empty(waitq)) { + ir_node *n = (ir_node*)pdeq_getl(waitq); opt_walker(n, waitq); + } + /* Calculate dominance so we can kill unreachable code + * We want this intertwined with localopts for better optimization (phase coupling) */ + compute_doms(irg); + irg_block_walk_graph(irg, NULL, find_unreachable_blocks, waitq); } - del_pdeq(waitq); - ir_free_resources(irg, IR_RESOURCE_IRN_LINK); - if (! state) - edges_deactivate(irg); + /* disable unreachable code elimination */ + clear_irg_state(irg, IR_GRAPH_STATE_OPTIMIZE_UNREACHABLE_CODE); + set_irg_state(irg, IR_GRAPH_STATE_NO_UNREACHABLE_CODE); + + /* invalidate infos */ + clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_DOMINANCE); + clear_irg_state(irg, IR_GRAPH_STATE_CONSISTENT_LOOPINFO); + clear_irg_state(irg, IR_GRAPH_STATE_VALID_EXTENDED_BLOCKS); + edges_deactivate(irg); /* Finally kill BAD and doublets from the keep alives. - Doing this AFTER edges where deactivated saves cycles */ - end = get_irg_end(irg); + * Doing this AFTER edges where deactivated saves cycles */ + end = get_irg_end(irg); remove_End_Bads_and_doublets(end); current_ir_graph = rem; - return changed; + + /* Note we do not have a reliable way to detect changes, since some + * localopt rules change the inputs of a node and do not return a new + * node, so we conservatively say true here */ + return true; } -/* Creates an ir_graph pass for optimize_graph_df. */ ir_graph_pass_t *optimize_graph_df_pass(const char *name) { return def_graph_pass_ret(name ? name : "optimize_graph_df", optimize_graph_df); -} /* optimize_graph_df_pass */ +}