2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Optimizations for a whole ir graph, i.e., a procedure.
23 * @author Christian Schaefer, Goetz Lindenmaier, Sebastian Felis,
34 #include "irgraph_t.h"
37 #include "iroptimize.h"
47 #include "pdeq.h" /* Fuer code placement */
52 #include "irbackedge_t.h"
59 #include "iredges_t.h"
62 /*------------------------------------------------------------------*/
63 /* apply optimizations of iropt to all nodes. */
64 /*------------------------------------------------------------------*/
67 * A wrapper around optimize_inplace_2() to be called from a walker.
69 static void optimize_in_place_wrapper (ir_node *n, void *env) {
70 ir_node *optimized = optimize_in_place_2(n);
74 exchange (n, optimized);
79 * Do local optimizations for a node.
81 * @param n the IR-node where to start. Typically the End node
84 * @note current_ir_graph must be set
86 static INLINE void do_local_optimize(ir_node *n) {
87 /* Handle graph state */
88 assert(get_irg_phase_state(current_ir_graph) != phase_building);
90 if (get_opt_global_cse())
91 set_irg_pinned(current_ir_graph, op_pin_state_floats);
92 set_irg_outs_inconsistent(current_ir_graph);
93 set_irg_doms_inconsistent(current_ir_graph);
94 set_irg_loopinfo_inconsistent(current_ir_graph);
96 /* Clean the value_table in irg for the CSE. */
97 del_identities(current_ir_graph->value_table);
98 current_ir_graph->value_table = new_identities();
100 /* walk over the graph */
101 irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
104 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
105 void local_optimize_node(ir_node *n) {
106 ir_graph *rem = current_ir_graph;
107 current_ir_graph = get_irn_irg(n);
109 do_local_optimize(n);
111 current_ir_graph = rem;
115 * Block-Walker: uses dominance depth to mark dead blocks.
117 static void kill_dead_blocks(ir_node *block, void *env) {
120 if (get_Block_dom_depth(block) < 0) {
122 * Note that the new dominance code correctly handles
123 * the End block, i.e. it is always reachable from Start
125 set_Block_dead(block);
129 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */
130 void local_optimize_graph(ir_graph *irg) {
131 ir_graph *rem = current_ir_graph;
132 current_ir_graph = irg;
134 if (get_irg_dom_state(irg) == dom_consistent)
135 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
137 do_local_optimize(get_irg_end(irg));
139 current_ir_graph = rem;
143 * Enqueue all users of a node to a wait queue.
144 * Handles mode_T nodes.
146 static void enqueue_users(ir_node *n, pdeq *waitq) {
147 const ir_edge_t *edge;
149 foreach_out_edge(n, edge) {
150 ir_node *succ = get_edge_src_irn(edge);
152 if (get_irn_link(succ) != waitq) {
153 pdeq_putr(waitq, succ);
154 set_irn_link(succ, waitq);
156 if (get_irn_mode(succ) == mode_T) {
157 /* A mode_T node has Proj's. Because most optimizations
158 run on the Proj's we have to enqueue them also. */
159 enqueue_users(succ, waitq);
165 * Data flow optimization walker.
166 * Optimizes all nodes and enqueue it's users
169 static void opt_walker(ir_node *n, void *env) {
173 optimized = optimize_in_place_2(n);
174 set_irn_link(optimized, NULL);
176 if (optimized != n) {
177 enqueue_users(n, waitq);
178 exchange(n, optimized);
182 /* Applies local optimizations to all nodes in the graph until fixpoint. */
183 void optimize_graph_df(ir_graph *irg) {
184 pdeq *waitq = new_pdeq();
185 ir_graph *rem = current_ir_graph;
189 current_ir_graph = irg;
191 state = edges_assure(irg);
193 if (get_opt_global_cse())
194 set_irg_pinned(current_ir_graph, op_pin_state_floats);
196 /* Clean the value_table in irg for the CSE. */
197 del_identities(irg->value_table);
198 irg->value_table = new_identities();
200 if (get_irg_dom_state(irg) == dom_consistent)
201 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
203 /* invalidate info */
204 set_irg_outs_inconsistent(irg);
205 set_irg_doms_inconsistent(irg);
206 set_irg_loopinfo_inconsistent(irg);
208 set_using_irn_link(irg);
210 /* walk over the graph, but don't touch keep-alives */
211 irg_walk(get_irg_end_block(irg), NULL, opt_walker, waitq);
213 end = get_irg_end(irg);
215 /* optimize keep-alives by removing superfluous ones */
216 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
217 ir_node *ka = get_End_keepalive(end, i);
219 if (irn_visited(ka) && !is_irn_keep(ka)) {
220 /* this node can be regularly visited, no need to keep it */
221 set_End_keepalive(end, i, get_irg_bad(irg));
224 /* now walk again and visit all not yet visited nodes */
225 set_irg_visited(current_ir_graph, get_irg_visited(irg) - 1);
226 irg_walk(get_irg_end(irg), NULL, opt_walker, waitq);
228 /* finish the wait queue */
229 while (! pdeq_empty(waitq)) {
230 ir_node *n = pdeq_getl(waitq);
232 opt_walker(n, waitq);
237 clear_using_irn_link(irg);
240 edges_deactivate(irg);
242 current_ir_graph = rem;
246 /*------------------------------------------------------------------*/
247 /* Routines for dead node elimination / copying garbage collection */
248 /* of the obstack. */
249 /*------------------------------------------------------------------*/
252 * Remember the new node in the old node by using a field all nodes have.
254 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
257 * Get this new node, before the old node is forgotten.
259 #define get_new_node(oldn) get_irn_link(oldn)
262 * Check if a new node was set.
264 #define has_new_node(n) (get_new_node(n) != NULL)
267 * We use the block_visited flag to mark that we have computed the
268 * number of useful predecessors for this block.
269 * Further we encode the new arity in this flag in the old blocks.
270 * Remembering the arity is useful, as it saves a lot of pointer
271 * accesses. This function is called for all Phi and Block nodes
275 compute_new_arity(ir_node *b) {
276 int i, res, irn_arity;
279 irg_v = get_irg_block_visited(current_ir_graph);
280 block_v = get_Block_block_visited(b);
281 if (block_v >= irg_v) {
282 /* we computed the number of preds for this block and saved it in the
284 return block_v - irg_v;
286 /* compute the number of good predecessors */
287 res = irn_arity = get_irn_arity(b);
288 for (i = 0; i < irn_arity; i++)
289 if (is_Bad(get_irn_n(b, i))) res--;
290 /* save it in the flag. */
291 set_Block_block_visited(b, irg_v + res);
297 * Copies the node to the new obstack. The Ins of the new node point to
298 * the predecessors on the old obstack. For block/phi nodes not all
299 * predecessors might be copied. n->link points to the new node.
300 * For Phi and Block nodes the function allocates in-arrays with an arity
301 * only for useful predecessors. The arity is determined by counting
302 * the non-bad predecessors of the block.
304 * @param n The node to be copied
305 * @param env if non-NULL, the node number attribute will be copied to the new node
307 * Note: Also used for loop unrolling.
309 static void copy_node(ir_node *n, void *env) {
312 ir_op *op = get_irn_op(n);
315 /* The end node looses it's flexible in array. This doesn't matter,
316 as dead node elimination builds End by hand, inlineing doesn't use
318 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
321 /* node copied already */
323 } else if (op == op_Block) {
325 new_arity = compute_new_arity(n);
326 n->attr.block.graph_arr = NULL;
328 block = get_nodes_block(n);
330 new_arity = compute_new_arity(block);
332 new_arity = get_irn_arity(n);
335 nn = new_ir_node(get_irn_dbg_info(n),
342 /* Copy the attributes. These might point to additional data. If this
343 was allocated on the old obstack the pointers now are dangling. This
344 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
345 copy_node_attr(n, nn);
349 int copy_node_nr = env != NULL;
351 /* for easier debugging, we want to copy the node numbers too */
352 nn->node_nr = n->node_nr;
358 hook_dead_node_elim_subst(current_ir_graph, n, nn);
362 * Copies new predecessors of old node to new node remembered in link.
363 * Spare the Bad predecessors of Phi and Block nodes.
365 static void copy_preds(ir_node *n, void *env) {
370 nn = get_new_node(n);
373 /* copy the macro block header */
374 ir_node *mbh = get_Block_MacroBlock(n);
377 /* this block is a macroblock header */
378 set_irn_n(nn, -1, nn);
380 /* get the macro block header */
381 set_irn_n(nn, -1, get_new_node(mbh));
384 /* Don't copy Bad nodes. */
386 irn_arity = get_irn_arity(n);
387 for (i = 0; i < irn_arity; i++) {
388 if (! is_Bad(get_irn_n(n, i))) {
389 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
390 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
394 /* repair the block visited flag from above misuse. Repair it in both
395 graphs so that the old one can still be used. */
396 set_Block_block_visited(nn, 0);
397 set_Block_block_visited(n, 0);
398 /* Local optimization could not merge two subsequent blocks if
399 in array contained Bads. Now it's possible.
400 We don't call optimize_in_place as it requires
401 that the fields in ir_graph are set properly. */
402 if ((get_opt_control_flow_straightening()) &&
403 (get_Block_n_cfgpreds(nn) == 1) &&
404 is_Jmp(get_Block_cfgpred(nn, 0))) {
405 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
407 /* Jmp jumps into the block it is in -- deal self cycle. */
408 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
409 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
414 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
415 /* Don't copy node if corresponding predecessor in block is Bad.
416 The Block itself should not be Bad. */
417 block = get_nodes_block(n);
418 set_irn_n(nn, -1, get_new_node(block));
420 irn_arity = get_irn_arity(n);
421 for (i = 0; i < irn_arity; i++) {
422 if (! is_Bad(get_irn_n(block, i))) {
423 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
424 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
428 /* If the pre walker reached this Phi after the post walker visited the
429 block block_visited is > 0. */
430 set_Block_block_visited(get_nodes_block(n), 0);
431 /* Compacting the Phi's ins might generate Phis with only one
433 if (get_irn_arity(nn) == 1)
434 exchange(nn, get_irn_n(nn, 0));
436 irn_arity = get_irn_arity(n);
437 for (i = -1; i < irn_arity; i++)
438 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
440 /* Now the new node is complete. We can add it to the hash table for CSE.
441 @@@ inlining aborts if we identify End. Why? */
443 add_identities(current_ir_graph->value_table, nn);
447 * Copies the graph recursively, compacts the keep-alives of the end node.
449 * @param irg the graph to be copied
450 * @param copy_node_nr If non-zero, the node number will be copied
452 static void copy_graph(ir_graph *irg, int copy_node_nr) {
453 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
454 ir_node *ka; /* keep alive */
458 /* Some nodes must be copied by hand, sigh */
459 vfl = get_irg_visited(irg);
460 set_irg_visited(irg, vfl + 1);
462 oe = get_irg_end(irg);
463 mark_irn_visited(oe);
464 /* copy the end node by hand, allocate dynamic in array! */
465 ne = new_ir_node(get_irn_dbg_info(oe),
472 /* Copy the attributes. Well, there might be some in the future... */
473 copy_node_attr(oe, ne);
474 set_new_node(oe, ne);
476 /* copy the Bad node */
477 ob = get_irg_bad(irg);
478 mark_irn_visited(ob);
479 nb = new_ir_node(get_irn_dbg_info(ob),
486 copy_node_attr(ob, nb);
487 set_new_node(ob, nb);
489 /* copy the NoMem node */
490 om = get_irg_no_mem(irg);
491 mark_irn_visited(om);
492 nm = new_ir_node(get_irn_dbg_info(om),
499 copy_node_attr(om, nm);
500 set_new_node(om, nm);
502 /* copy the live nodes */
503 set_irg_visited(irg, vfl);
504 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
506 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
508 /* visit the anchors as well */
509 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
510 ir_node *n = get_irg_anchor(irg, i);
512 if (n && (get_irn_visited(n) <= vfl)) {
513 set_irg_visited(irg, vfl);
514 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
518 /* copy_preds for the end node ... */
519 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
521 /*- ... and now the keep alives. -*/
522 /* First pick the not marked block nodes and walk them. We must pick these
523 first as else we will oversee blocks reachable from Phis. */
524 irn_arity = get_End_n_keepalives(oe);
525 for (i = 0; i < irn_arity; i++) {
526 ka = get_End_keepalive(oe, i);
528 if (get_irn_visited(ka) <= vfl) {
529 /* We must keep the block alive and copy everything reachable */
530 set_irg_visited(irg, vfl);
531 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
533 add_End_keepalive(ne, get_new_node(ka));
537 /* Now pick other nodes. Here we will keep all! */
538 irn_arity = get_End_n_keepalives(oe);
539 for (i = 0; i < irn_arity; i++) {
540 ka = get_End_keepalive(oe, i);
542 if (get_irn_visited(ka) <= vfl) {
543 /* We didn't copy the node yet. */
544 set_irg_visited(irg, vfl);
545 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
547 add_End_keepalive(ne, get_new_node(ka));
551 /* start block sometimes only reached after keep alives */
552 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
553 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
557 * Copies the graph reachable from current_ir_graph->end to the obstack
558 * in current_ir_graph and fixes the environment.
559 * Then fixes the fields in current_ir_graph containing nodes of the
562 * @param copy_node_nr If non-zero, the node number will be copied
565 copy_graph_env(int copy_node_nr) {
566 ir_graph *irg = current_ir_graph;
567 ir_node *old_end, *new_anchor;
570 /* remove end_except and end_reg nodes */
571 old_end = get_irg_end(irg);
572 set_irg_end_except (irg, old_end);
573 set_irg_end_reg (irg, old_end);
575 /* Not all nodes remembered in irg might be reachable
576 from the end node. Assure their link is set to NULL, so that
577 we can test whether new nodes have been computed. */
578 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
579 ir_node *n = get_irg_anchor(irg, i);
581 set_new_node(n, NULL);
583 /* we use the block walk flag for removing Bads from Blocks ins. */
584 inc_irg_block_visited(irg);
587 copy_graph(irg, copy_node_nr);
590 old_end = get_irg_end(irg);
591 new_anchor = new_Anchor(irg);
593 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
594 ir_node *n = get_irg_anchor(irg, i);
596 set_irn_n(new_anchor, i, get_new_node(n));
599 irg->anchor = new_anchor;
601 /* ensure the new anchor is placed in the endblock */
602 set_irn_n(new_anchor, -1, get_irg_end_block(irg));
606 * Copies all reachable nodes to a new obstack. Removes bad inputs
607 * from block nodes and the corresponding inputs from Phi nodes.
608 * Merges single exit blocks with single entry blocks and removes
610 * Adds all new nodes to a new hash table for CSE. Does not
611 * perform CSE, so the hash table might contain common subexpressions.
613 void dead_node_elimination(ir_graph *irg) {
615 #ifdef INTERPROCEDURAL_VIEW
616 int rem_ipview = get_interprocedural_view();
618 struct obstack *graveyard_obst = NULL;
619 struct obstack *rebirth_obst = NULL;
620 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
622 /* inform statistics that we started a dead-node elimination run */
623 hook_dead_node_elim(irg, 1);
625 /* Remember external state of current_ir_graph. */
626 rem = current_ir_graph;
627 current_ir_graph = irg;
628 #ifdef INTERPROCEDURAL_VIEW
629 set_interprocedural_view(0);
632 assert(get_irg_phase_state(irg) != phase_building);
634 /* Handle graph state */
635 free_callee_info(irg);
639 /* @@@ so far we loose loops when copying */
640 free_loop_information(irg);
642 set_irg_doms_inconsistent(irg);
644 /* A quiet place, where the old obstack can rest in peace,
645 until it will be cremated. */
646 graveyard_obst = irg->obst;
648 /* A new obstack, where the reachable nodes will be copied to. */
649 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
650 irg->obst = rebirth_obst;
651 obstack_init(irg->obst);
652 irg->last_node_idx = 0;
654 /* We also need a new value table for CSE */
655 del_identities(irg->value_table);
656 irg->value_table = new_identities();
658 /* Copy the graph from the old to the new obstack */
659 copy_graph_env(/*copy_node_nr=*/1);
661 /* Free memory from old unoptimized obstack */
662 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
663 xfree(graveyard_obst); /* ... then free it. */
665 /* inform statistics that the run is over */
666 hook_dead_node_elim(irg, 0);
668 current_ir_graph = rem;
669 #ifdef INTERPROCEDURAL_VIEW
670 set_interprocedural_view(rem_ipview);
675 * Relink bad predecessors of a block and store the old in array to the
676 * link field. This function is called by relink_bad_predecessors().
677 * The array of link field starts with the block operand at position 0.
678 * If block has bad predecessors, create a new in array without bad preds.
679 * Otherwise let in array untouched.
681 static void relink_bad_block_predecessors(ir_node *n, void *env) {
682 ir_node **new_in, *irn;
683 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
686 /* if link field of block is NULL, look for bad predecessors otherwise
687 this is already done */
688 if (is_Block(n) && get_irn_link(n) == NULL) {
689 /* save old predecessors in link field (position 0 is the block operand)*/
690 set_irn_link(n, get_irn_in(n));
692 /* count predecessors without bad nodes */
693 old_irn_arity = get_irn_arity(n);
694 for (i = 0; i < old_irn_arity; i++)
695 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
697 /* arity changing: set new predecessors without bad nodes */
698 if (new_irn_arity < old_irn_arity) {
699 /* Get new predecessor array. We do not resize the array, as we must
700 keep the old one to update Phis. */
701 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
703 /* set new predecessors in array */
706 for (i = 0; i < old_irn_arity; i++) {
707 irn = get_irn_n(n, i);
709 new_in[new_irn_n] = irn;
710 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
714 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
715 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
717 } /* ir node has bad predecessors */
718 } /* Block is not relinked */
722 * Relinks Bad predecessors from Blocks and Phis called by walker
723 * remove_bad_predecesors(). If n is a Block, call
724 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
725 * function of Phi's Block. If this block has bad predecessors, relink preds
728 static void relink_bad_predecessors(ir_node *n, void *env) {
729 ir_node *block, **old_in;
730 int i, old_irn_arity, new_irn_arity;
732 /* relink bad predecessors of a block */
734 relink_bad_block_predecessors(n, env);
736 /* If Phi node relink its block and its predecessors */
738 /* Relink predecessors of phi's block */
739 block = get_nodes_block(n);
740 if (get_irn_link(block) == NULL)
741 relink_bad_block_predecessors(block, env);
743 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
744 old_irn_arity = ARR_LEN(old_in);
746 /* Relink Phi predecessors if count of predecessors changed */
747 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
748 /* set new predecessors in array
749 n->in[0] remains the same block */
751 for(i = 1; i < old_irn_arity; i++)
752 if (!is_Bad((ir_node *)old_in[i])) {
753 n->in[new_irn_arity] = n->in[i];
754 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
758 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
759 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
761 } /* n is a Phi node */
765 * Removes Bad Bad predecessors from Blocks and the corresponding
766 * inputs to Phi nodes as in dead_node_elimination but without
768 * On walking up set the link field to NULL, on walking down call
769 * relink_bad_predecessors() (This function stores the old in array
770 * to the link field and sets a new in array if arity of predecessors
773 void remove_bad_predecessors(ir_graph *irg) {
774 panic("Fix backedge handling first");
775 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
782 __)|_| | \_/ | \_/(/_ |_/\__|__
784 The following stuff implements a facility that automatically patches
785 registered ir_node pointers to the new node when a dead node elimination occurs.
788 struct _survive_dce_t {
792 hook_entry_t dead_node_elim;
793 hook_entry_t dead_node_elim_subst;
796 typedef struct _survive_dce_list_t {
797 struct _survive_dce_list_t *next;
799 } survive_dce_list_t;
801 static void dead_node_hook(void *context, ir_graph *irg, int start) {
802 survive_dce_t *sd = context;
805 /* Create a new map before the dead node elimination is performed. */
807 sd->new_places = pmap_create_ex(pmap_count(sd->places));
809 /* Patch back all nodes if dead node elimination is over and something is to be done. */
810 pmap_destroy(sd->places);
811 sd->places = sd->new_places;
812 sd->new_places = NULL;
817 * Hook called when dead node elimination replaces old by nw.
819 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
820 survive_dce_t *sd = context;
821 survive_dce_list_t *list = pmap_get(sd->places, old);
824 /* If the node is to be patched back, write the new address to all registered locations. */
826 survive_dce_list_t *p;
828 for (p = list; p; p = p->next)
831 pmap_insert(sd->new_places, nw, list);
836 * Make a new Survive DCE environment.
838 survive_dce_t *new_survive_dce(void) {
839 survive_dce_t *res = xmalloc(sizeof(res[0]));
840 obstack_init(&res->obst);
841 res->places = pmap_create();
842 res->new_places = NULL;
844 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
845 res->dead_node_elim.context = res;
846 res->dead_node_elim.next = NULL;
848 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
849 res->dead_node_elim_subst.context = res;
850 res->dead_node_elim_subst.next = NULL;
852 #ifndef FIRM_ENABLE_HOOKS
853 assert(0 && "need hooks enabled");
856 register_hook(hook_dead_node_elim, &res->dead_node_elim);
857 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
862 * Free a Survive DCE environment.
864 void free_survive_dce(survive_dce_t *sd) {
865 obstack_free(&sd->obst, NULL);
866 pmap_destroy(sd->places);
867 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
868 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
873 * Register a node pointer to be patched upon DCE.
874 * When DCE occurs, the node pointer specified by @p place will be
875 * patched to the new address of the node it is pointing to.
877 * @param sd The Survive DCE environment.
878 * @param place The address of the node pointer.
880 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
881 if (*place != NULL) {
882 ir_node *irn = *place;
883 survive_dce_list_t *curr = pmap_get(sd->places, irn);
884 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
889 pmap_insert(sd->places, irn, nw);
893 /*--------------------------------------------------------------------*/
894 /* Functionality for inlining */
895 /*--------------------------------------------------------------------*/
898 * Copy node for inlineing. Updates attributes that change when
899 * inlineing but not for dead node elimination.
901 * Copies the node by calling copy_node() and then updates the entity if
902 * it's a local one. env must be a pointer of the frame type of the
903 * inlined procedure. The new entities must be in the link field of
907 copy_node_inline(ir_node *n, void *env) {
909 ir_type *frame_tp = (ir_type *)env;
913 nn = get_new_node (n);
915 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
916 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
918 } else if (is_Block(n)) {
919 nn = get_new_node (n);
920 nn->attr.block.irg = current_ir_graph;
925 * Walker: checks if P_value_arg_base is used.
927 static void find_addr(ir_node *node, void *env) {
928 int *allow_inline = env;
930 is_Start(get_Proj_pred(node)) &&
931 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
937 * Check if we can inline a given call.
938 * Currently, we cannot inline two cases:
939 * - call with compound arguments
940 * - graphs that take the address of a parameter
942 * check these conditions here
944 static int can_inline(ir_node *call, ir_graph *called_graph) {
945 ir_type *call_type = get_Call_type(call);
946 int params, ress, i, res;
947 assert(is_Method_type(call_type));
949 params = get_method_n_params(call_type);
950 ress = get_method_n_ress(call_type);
952 /* check parameters for compound arguments */
953 for (i = 0; i < params; ++i) {
954 ir_type *p_type = get_method_param_type(call_type, i);
956 if (is_compound_type(p_type))
960 /* check results for compound arguments */
961 for (i = 0; i < ress; ++i) {
962 ir_type *r_type = get_method_res_type(call_type, i);
964 if (is_compound_type(r_type))
969 irg_walk_graph(called_graph, find_addr, NULL, &res);
975 exc_handler = 0, /**< There is a handler. */
976 exc_to_end = 1, /**< Branches to End. */
977 exc_no_handler = 2 /**< Exception handling not represented. */
980 /* Inlines a method at the given call site. */
981 int inline_method(ir_node *call, ir_graph *called_graph) {
983 ir_node *post_call, *post_bl;
984 ir_node *in[pn_Start_max];
985 ir_node *end, *end_bl;
989 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
990 enum exc_mode exc_handling;
991 ir_type *called_frame;
992 irg_inline_property prop = get_irg_inline_property(called_graph);
994 if ( (prop < irg_inline_forced) || (prop == irg_inline_forbidden))
997 /* Do not inline variadic functions. */
998 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
1001 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
1002 get_method_n_params(get_Call_type(call)));
1005 * currently, we cannot inline two cases:
1006 * - call with compound arguments
1007 * - graphs that take the address of a parameter
1009 if (! can_inline(call, called_graph))
1012 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
1013 rem_opt = get_opt_optimize();
1016 /* Handle graph state */
1017 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1018 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
1019 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
1020 set_irg_outs_inconsistent(current_ir_graph);
1021 set_irg_extblk_inconsistent(current_ir_graph);
1022 set_irg_doms_inconsistent(current_ir_graph);
1023 set_irg_loopinfo_inconsistent(current_ir_graph);
1024 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
1026 /* -- Check preconditions -- */
1027 assert(is_Call(call));
1028 /* @@@ does not work for InterfaceIII.java after cgana
1029 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
1030 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
1031 get_Call_type(call)));
1033 if (called_graph == current_ir_graph) {
1034 set_optimize(rem_opt);
1038 /* here we know we WILL inline, so inform the statistics */
1039 hook_inline(call, called_graph);
1041 /* -- Decide how to handle exception control flow: Is there a handler
1042 for the Call node, or do we branch directly to End on an exception?
1044 0 There is a handler.
1046 2 Exception handling not represented in Firm. -- */
1048 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
1049 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
1050 long proj_nr = get_Proj_proj(proj);
1051 if (proj_nr == pn_Call_X_except) Xproj = proj;
1052 if (proj_nr == pn_Call_M_except) Mproj = proj;
1054 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
1055 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
1056 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
1060 the procedure and later replaces the Start node of the called graph.
1061 Post_call is the old Call node and collects the results of the called
1062 graph. Both will end up being a tuple. -- */
1063 post_bl = get_nodes_block(call);
1064 set_irg_current_block(current_ir_graph, post_bl);
1065 /* XxMxPxPxPxT of Start + parameter of Call */
1066 in[pn_Start_X_initial_exec] = new_Jmp();
1067 in[pn_Start_M] = get_Call_mem(call);
1068 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1069 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1070 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1071 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1072 /* in[pn_Start_P_value_arg_base] = ??? */
1073 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1074 pre_call = new_Tuple(pn_Start_max - 1, in);
1078 The new block gets the ins of the old block, pre_call and all its
1079 predecessors and all Phi nodes. -- */
1080 part_block(pre_call);
1082 /* -- Prepare state for dead node elimination -- */
1083 /* Visited flags in calling irg must be >= flag in called irg.
1084 Else walker and arity computation will not work. */
1085 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1086 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1087 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1088 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1089 /* Set pre_call as new Start node in link field of the start node of
1090 calling graph and pre_calls block as new block for the start block
1092 Further mark these nodes so that they are not visited by the
1094 set_irn_link(get_irg_start(called_graph), pre_call);
1095 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1096 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1097 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1098 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1099 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1101 /* Initialize for compaction of in arrays */
1102 inc_irg_block_visited(current_ir_graph);
1104 /* -- Replicate local entities of the called_graph -- */
1105 /* copy the entities. */
1106 called_frame = get_irg_frame_type(called_graph);
1107 for (i = 0; i < get_class_n_members(called_frame); i++) {
1108 ir_entity *new_ent, *old_ent;
1109 old_ent = get_class_member(called_frame, i);
1110 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1111 set_entity_link(old_ent, new_ent);
1114 /* visited is > than that of called graph. With this trick visited will
1115 remain unchanged so that an outer walker, e.g., searching the call nodes
1116 to inline, calling this inline will not visit the inlined nodes. */
1117 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1119 /* -- Performing dead node elimination inlines the graph -- */
1120 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1122 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1123 get_irg_frame_type(called_graph));
1125 /* Repair called_graph */
1126 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1127 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1128 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1130 /* -- Merge the end of the inlined procedure with the call site -- */
1131 /* We will turn the old Call node into a Tuple with the following
1134 0: Phi of all Memories of Return statements.
1135 1: Jmp from new Block that merges the control flow from all exception
1136 predecessors of the old end block.
1137 2: Tuple of all arguments.
1138 3: Phi of Exception memories.
1139 In case the old Call directly branches to End on an exception we don't
1140 need the block merging all exceptions nor the Phi of the exception
1144 /* -- Precompute some values -- */
1145 end_bl = get_new_node(get_irg_end_block(called_graph));
1146 end = get_new_node(get_irg_end(called_graph));
1147 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1148 n_res = get_method_n_ress(get_Call_type(call));
1150 res_pred = xmalloc(n_res * sizeof(*res_pred));
1151 cf_pred = xmalloc(arity * sizeof(*res_pred));
1153 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1155 /* -- archive keepalives -- */
1156 irn_arity = get_irn_arity(end);
1157 for (i = 0; i < irn_arity; i++) {
1158 ir_node *ka = get_End_keepalive(end, i);
1160 add_End_keepalive(get_irg_end(current_ir_graph), ka);
1163 /* The new end node will die. We need not free as the in array is on the obstack:
1164 copy_node() only generated 'D' arrays. */
1166 /* -- Replace Return nodes by Jump nodes. -- */
1168 for (i = 0; i < arity; i++) {
1170 ret = get_irn_n(end_bl, i);
1171 if (is_Return(ret)) {
1172 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1176 set_irn_in(post_bl, n_ret, cf_pred);
1178 /* -- Build a Tuple for all results of the method.
1179 Add Phi node if there was more than one Return. -- */
1180 turn_into_tuple(post_call, pn_Call_max);
1181 /* First the Memory-Phi */
1183 for (i = 0; i < arity; i++) {
1184 ret = get_irn_n(end_bl, i);
1185 if (is_Return(ret)) {
1186 cf_pred[n_ret] = get_Return_mem(ret);
1190 phi = new_Phi(n_ret, cf_pred, mode_M);
1191 set_Tuple_pred(call, pn_Call_M_regular, phi);
1192 /* Conserve Phi-list for further inlinings -- but might be optimized */
1193 if (get_nodes_block(phi) == post_bl) {
1194 set_irn_link(phi, get_irn_link(post_bl));
1195 set_irn_link(post_bl, phi);
1197 /* Now the real results */
1199 for (j = 0; j < n_res; j++) {
1201 for (i = 0; i < arity; i++) {
1202 ret = get_irn_n(end_bl, i);
1203 if (is_Return(ret)) {
1204 cf_pred[n_ret] = get_Return_res(ret, j);
1209 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1213 /* Conserve Phi-list for further inlinings -- but might be optimized */
1214 if (get_nodes_block(phi) == post_bl) {
1215 set_irn_link(phi, get_irn_link(post_bl));
1216 set_irn_link(post_bl, phi);
1219 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1221 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1223 /* handle the regular call */
1224 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1226 /* For now, we cannot inline calls with value_base */
1227 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1229 /* Finally the exception control flow.
1230 We have two (three) possible situations:
1231 First if the Call branches to an exception handler: We need to add a Phi node to
1232 collect the memory containing the exception objects. Further we need
1233 to add another block to get a correct representation of this Phi. To
1234 this block we add a Jmp that resolves into the X output of the Call
1235 when the Call is turned into a tuple.
1236 Second the Call branches to End, the exception is not handled. Just
1237 add all inlined exception branches to the End node.
1238 Third: there is no Exception edge at all. Handle as case two. */
1239 if (exc_handling == exc_handler) {
1241 for (i = 0; i < arity; i++) {
1243 ret = get_irn_n(end_bl, i);
1244 irn = skip_Proj(ret);
1245 if (is_fragile_op(irn) || is_Raise(irn)) {
1246 cf_pred[n_exc] = ret;
1251 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1252 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1253 /* The Phi for the memories with the exception objects */
1255 for (i = 0; i < arity; i++) {
1257 ret = skip_Proj(get_irn_n(end_bl, i));
1259 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1261 } else if (is_fragile_op(ret)) {
1262 /* We rely that all cfops have the memory output at the same position. */
1263 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1265 } else if (is_Raise(ret)) {
1266 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1270 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1272 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1273 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1276 ir_node *main_end_bl;
1277 int main_end_bl_arity;
1278 ir_node **end_preds;
1280 /* assert(exc_handling == 1 || no exceptions. ) */
1282 for (i = 0; i < arity; i++) {
1283 ir_node *ret = get_irn_n(end_bl, i);
1284 ir_node *irn = skip_Proj(ret);
1286 if (is_fragile_op(irn) || is_Raise(irn)) {
1287 cf_pred[n_exc] = ret;
1291 main_end_bl = get_irg_end_block(current_ir_graph);
1292 main_end_bl_arity = get_irn_arity(main_end_bl);
1293 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1295 for (i = 0; i < main_end_bl_arity; ++i)
1296 end_preds[i] = get_irn_n(main_end_bl, i);
1297 for (i = 0; i < n_exc; ++i)
1298 end_preds[main_end_bl_arity + i] = cf_pred[i];
1299 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1300 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1301 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1307 /* -- Turn CSE back on. -- */
1308 set_optimize(rem_opt);
1313 /********************************************************************/
1314 /* Apply inlineing to small methods. */
1315 /********************************************************************/
1317 /** Represents a possible inlinable call in a graph. */
1318 typedef struct _call_entry call_entry;
1319 struct _call_entry {
1320 ir_node *call; /**< the Call */
1321 ir_graph *callee; /**< the callee called here */
1322 call_entry *next; /**< for linking the next one */
1326 * environment for inlining small irgs
1328 typedef struct _inline_env_t {
1329 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1330 call_entry *head; /**< the head of the call entry list */
1331 call_entry *tail; /**< the tail of the call entry list */
1335 * Returns the irg called from a Call node. If the irg is not
1336 * known, NULL is returned.
1338 static ir_graph *get_call_called_irg(ir_node *call) {
1340 ir_graph *called_irg = NULL;
1342 addr = get_Call_ptr(call);
1343 if (is_SymConst_addr_ent(addr)) {
1344 called_irg = get_entity_irg(get_SymConst_entity(addr));
1351 * Walker: Collect all calls to known graphs inside a graph.
1353 static void collect_calls(ir_node *call, void *env) {
1354 if (is_Call(call)) {
1355 ir_graph *called_irg = get_call_called_irg(call);
1357 /* The Call node calls a locally defined method. Remember to inline. */
1358 inline_env_t *ienv = env;
1359 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1361 entry->callee = called_irg;
1364 if (ienv->tail == NULL)
1367 ienv->tail->next = entry;
1374 * Inlines all small methods at call sites where the called address comes
1375 * from a Const node that references the entity representing the called
1377 * The size argument is a rough measure for the code size of the method:
1378 * Methods where the obstack containing the firm graph is smaller than
1381 void inline_small_irgs(ir_graph *irg, int size) {
1382 ir_graph *rem = current_ir_graph;
1385 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1387 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1389 current_ir_graph = irg;
1390 /* Handle graph state */
1391 assert(get_irg_phase_state(irg) != phase_building);
1392 free_callee_info(irg);
1394 /* Find Call nodes to inline.
1395 (We can not inline during a walk of the graph, as inlineing the same
1396 method several times changes the visited flag of the walked graph:
1397 after the first inlineing visited of the callee equals visited of
1398 the caller. With the next inlineing both are increased.) */
1399 obstack_init(&env.obst);
1400 env.head = env.tail = NULL;
1401 irg_walk_graph(irg, NULL, collect_calls, &env);
1403 if (env.head != NULL) {
1404 /* There are calls to inline */
1405 collect_phiprojs(irg);
1406 for (entry = env.head; entry != NULL; entry = entry->next) {
1407 ir_graph *callee = entry->callee;
1408 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1409 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1410 inline_method(entry->call, callee);
1414 obstack_free(&env.obst, NULL);
1415 current_ir_graph = rem;
1419 * Environment for inlining irgs.
1422 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1423 int n_nodes_orig; /**< for statistics */
1424 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1425 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1426 int n_call_nodes; /**< Number of Call nodes in the graph. */
1427 int n_call_nodes_orig; /**< for statistics */
1428 int n_callers; /**< Number of known graphs that call this graphs. */
1429 int n_callers_orig; /**< for statistics */
1430 int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
1434 * Allocate a new environment for inlining.
1436 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1437 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1438 env->n_nodes = -2; /* do not count count Start, End */
1439 env->n_nodes_orig = -2; /* do not count Start, End */
1440 env->call_head = NULL;
1441 env->call_tail = NULL;
1442 env->n_call_nodes = 0;
1443 env->n_call_nodes_orig = 0;
1445 env->n_callers_orig = 0;
1446 env->got_inline = 0;
1450 typedef struct walker_env {
1451 struct obstack *obst; /**< the obstack for allocations. */
1452 inline_irg_env *x; /**< the inline environment */
1453 int ignore_runtime; /**< the ignore runtime flag */
1457 * post-walker: collect all calls in the inline-environment
1458 * of a graph and sum some statistics.
1460 static void collect_calls2(ir_node *call, void *ctx) {
1462 inline_irg_env *x = env->x;
1463 ir_op *op = get_irn_op(call);
1467 /* count meaningful nodes in irg */
1468 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1473 if (op != op_Call) return;
1475 /* check, if it's a runtime call */
1476 if (env->ignore_runtime) {
1477 ir_node *symc = get_Call_ptr(call);
1479 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1480 ir_entity *ent = get_SymConst_entity(symc);
1482 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1487 /* collect all call nodes */
1489 ++x->n_call_nodes_orig;
1491 callee = get_call_called_irg(call);
1493 inline_irg_env *callee_env = get_irg_link(callee);
1494 /* count all static callers */
1495 ++callee_env->n_callers;
1496 ++callee_env->n_callers_orig;
1498 /* link it in the list of possible inlinable entries */
1499 entry = obstack_alloc(env->obst, sizeof(*entry));
1501 entry->callee = callee;
1503 if (x->call_tail == NULL)
1504 x->call_head = entry;
1506 x->call_tail->next = entry;
1507 x->call_tail = entry;
1512 * Returns TRUE if the number of callers is 0 in the irg's environment,
1513 * hence this irg is a leave.
1515 INLINE static int is_leave(ir_graph *irg) {
1516 inline_irg_env *env = get_irg_link(irg);
1517 return env->n_call_nodes == 0;
1521 * Returns TRUE if the number of nodes in the callee is
1522 * smaller then size in the irg's environment.
1524 INLINE static int is_smaller(ir_graph *callee, int size) {
1525 inline_irg_env *env = get_irg_link(callee);
1526 return env->n_nodes < size;
1530 * Append the nodes of the list src to the nodes of the list in environment dst.
1532 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1533 call_entry *entry, *nentry;
1535 /* Note that the src list points to Call nodes in the inlined graph, but
1536 we need Call nodes in our graph. Luckily the inliner leaves this information
1537 in the link field. */
1538 for (entry = src; entry != NULL; entry = entry->next) {
1539 nentry = obstack_alloc(obst, sizeof(*nentry));
1540 nentry->call = get_irn_link(entry->call);
1541 nentry->callee = entry->callee;
1542 nentry->next = NULL;
1543 dst->call_tail->next = nentry;
1544 dst->call_tail = nentry;
1549 * Inlines small leave methods at call sites where the called address comes
1550 * from a Const node that references the entity representing the called
1552 * The size argument is a rough measure for the code size of the method:
1553 * Methods where the obstack containing the firm graph is smaller than
1556 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1557 inline_irg_env *env;
1563 call_entry *entry, *tail;
1564 const call_entry *centry;
1565 struct obstack obst;
1566 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1568 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1569 rem = current_ir_graph;
1570 obstack_init(&obst);
1572 /* extend all irgs by a temporary data structure for inlining. */
1573 n_irgs = get_irp_n_irgs();
1574 for (i = 0; i < n_irgs; ++i)
1575 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1577 /* Precompute information in temporary data structure. */
1579 wenv.ignore_runtime = ignore_runtime;
1580 for (i = 0; i < n_irgs; ++i) {
1581 ir_graph *irg = get_irp_irg(i);
1583 assert(get_irg_phase_state(irg) != phase_building);
1584 free_callee_info(irg);
1586 wenv.x = get_irg_link(irg);
1587 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1590 /* -- and now inline. -- */
1592 /* Inline leaves recursively -- we might construct new leaves. */
1596 for (i = 0; i < n_irgs; ++i) {
1598 int phiproj_computed = 0;
1600 current_ir_graph = get_irp_irg(i);
1601 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1604 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1607 if (env->n_nodes > maxsize) break;
1610 callee = entry->callee;
1612 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1613 if (!phiproj_computed) {
1614 phiproj_computed = 1;
1615 collect_phiprojs(current_ir_graph);
1617 did_inline = inline_method(call, callee);
1620 /* Do some statistics */
1621 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1623 env->got_inline = 1;
1624 --env->n_call_nodes;
1625 env->n_nodes += callee_env->n_nodes;
1626 --callee_env->n_callers;
1628 /* remove this call from the list */
1630 tail->next = entry->next;
1632 env->call_head = entry->next;
1638 env->call_tail = tail;
1640 } while (did_inline);
1642 /* inline other small functions. */
1643 for (i = 0; i < n_irgs; ++i) {
1645 int phiproj_computed = 0;
1647 current_ir_graph = get_irp_irg(i);
1648 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1650 /* note that the list of possible calls is updated during the process */
1652 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1656 callee = entry->callee;
1658 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1659 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1660 if (!phiproj_computed) {
1661 phiproj_computed = 1;
1662 collect_phiprojs(current_ir_graph);
1664 if (inline_method(call, callee)) {
1665 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1667 /* callee was inline. Append it's call list. */
1668 env->got_inline = 1;
1669 --env->n_call_nodes;
1670 append_call_list(&obst, env, callee_env->call_head);
1671 env->n_call_nodes += callee_env->n_call_nodes;
1672 env->n_nodes += callee_env->n_nodes;
1673 --callee_env->n_callers;
1675 /* after we have inlined callee, all called methods inside callee
1676 are now called once more */
1677 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1678 inline_irg_env *penv = get_irg_link(centry->callee);
1682 /* remove this call from the list */
1684 tail->next = entry->next;
1686 env->call_head = entry->next;
1692 env->call_tail = tail;
1695 for (i = 0; i < n_irgs; ++i) {
1696 irg = get_irp_irg(i);
1697 env = (inline_irg_env *)get_irg_link(irg);
1699 if (env->got_inline) {
1700 /* this irg got calls inlined */
1701 set_irg_outs_inconsistent(irg);
1702 set_irg_doms_inconsistent(irg);
1704 optimize_graph_df(irg);
1707 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1708 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1709 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1710 env->n_callers_orig, env->n_callers,
1711 get_entity_name(get_irg_entity(irg))));
1715 obstack_free(&obst, NULL);
1716 current_ir_graph = rem;
1719 /*******************************************************************/
1720 /* Code Placement. Pins all floating nodes to a block where they */
1721 /* will be executed only if needed. */
1722 /*******************************************************************/
1725 * Returns non-zero, is a block is not reachable from Start.
1727 * @param block the block to test
1730 is_Block_unreachable(ir_node *block) {
1731 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1735 * Find the earliest correct block for node n. --- Place n into the
1736 * same Block as its dominance-deepest Input.
1738 * We have to avoid calls to get_nodes_block() here
1739 * because the graph is floating.
1741 * move_out_of_loops() expects that place_floats_early() have placed
1742 * all "living" nodes into a living block. That's why we must
1743 * move nodes in dead block with "live" successors into a valid
1745 * We move them just into the same block as it's successor (or
1746 * in case of a Phi into the effective use block). For Phi successors,
1747 * this may still be a dead block, but then there is no real use, as
1748 * the control flow will be dead later.
1750 * @param n the node to be placed
1751 * @param worklist a worklist, predecessors of non-floating nodes are placed here
1754 place_floats_early(ir_node *n, waitq *worklist) {
1757 /* we must not run into an infinite loop */
1758 assert(irn_not_visited(n));
1759 mark_irn_visited(n);
1761 /* Place floating nodes. */
1762 if (get_irn_pinned(n) == op_pin_state_floats) {
1763 ir_node *curr_block = get_nodes_block(n);
1764 int in_dead_block = is_Block_unreachable(curr_block);
1766 ir_node *b = NULL; /* The block to place this node in */
1768 assert(is_no_Block(n));
1770 if (is_irn_start_block_placed(n)) {
1771 /* These nodes will not be placed by the loop below. */
1772 b = get_irg_start_block(current_ir_graph);
1776 /* find the block for this node. */
1777 irn_arity = get_irn_arity(n);
1778 for (i = 0; i < irn_arity; i++) {
1779 ir_node *pred = get_irn_n(n, i);
1780 ir_node *pred_block;
1782 if ((irn_not_visited(pred))
1783 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1786 * If the current node is NOT in a dead block, but one of its
1787 * predecessors is, we must move the predecessor to a live block.
1788 * Such thing can happen, if global CSE chose a node from a dead block.
1789 * We move it simply to our block.
1790 * Note that neither Phi nor End nodes are floating, so we don't
1791 * need to handle them here.
1793 if (! in_dead_block) {
1794 if (get_irn_pinned(pred) == op_pin_state_floats &&
1795 is_Block_unreachable(get_nodes_block(pred)))
1796 set_nodes_block(pred, curr_block);
1798 place_floats_early(pred, worklist);
1802 * A node in the Bad block must stay in the bad block,
1803 * so don't compute a new block for it.
1808 /* Because all loops contain at least one op_pin_state_pinned node, now all
1809 our inputs are either op_pin_state_pinned or place_early() has already
1810 been finished on them. We do not have any unfinished inputs! */
1811 pred_block = get_nodes_block(pred);
1812 if ((!is_Block_dead(pred_block)) &&
1813 (get_Block_dom_depth(pred_block) > depth)) {
1815 depth = get_Block_dom_depth(pred_block);
1817 /* Avoid that the node is placed in the Start block */
1819 get_Block_dom_depth(get_nodes_block(n)) > 1 &&
1820 get_irg_phase_state(current_ir_graph) != phase_backend) {
1821 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1822 assert(b != get_irg_start_block(current_ir_graph));
1827 set_nodes_block(n, b);
1831 * Add predecessors of non floating nodes and non-floating predecessors
1832 * of floating nodes to worklist and fix their blocks if the are in dead block.
1834 irn_arity = get_irn_arity(n);
1838 * Simplest case: End node. Predecessors are keep-alives,
1839 * no need to move out of dead block.
1841 for (i = -1; i < irn_arity; ++i) {
1842 ir_node *pred = get_irn_n(n, i);
1843 if (irn_not_visited(pred))
1844 waitq_put(worklist, pred);
1846 } else if (is_Block(n)) {
1848 * Blocks: Predecessors are control flow, no need to move
1849 * them out of dead block.
1851 for (i = irn_arity - 1; i >= 0; --i) {
1852 ir_node *pred = get_irn_n(n, i);
1853 if (irn_not_visited(pred))
1854 waitq_put(worklist, pred);
1856 } else if (is_Phi(n)) {
1858 ir_node *curr_block = get_nodes_block(n);
1859 int in_dead_block = is_Block_unreachable(curr_block);
1862 * Phi nodes: move nodes from dead blocks into the effective use
1863 * of the Phi-input if the Phi is not in a bad block.
1865 pred = get_nodes_block(n);
1866 if (irn_not_visited(pred))
1867 waitq_put(worklist, pred);
1869 for (i = irn_arity - 1; i >= 0; --i) {
1870 ir_node *pred = get_irn_n(n, i);
1872 if (irn_not_visited(pred)) {
1873 if (! in_dead_block &&
1874 get_irn_pinned(pred) == op_pin_state_floats &&
1875 is_Block_unreachable(get_nodes_block(pred))) {
1876 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1878 waitq_put(worklist, pred);
1883 ir_node *curr_block = get_nodes_block(n);
1884 int in_dead_block = is_Block_unreachable(curr_block);
1887 * All other nodes: move nodes from dead blocks into the same block.
1889 pred = get_nodes_block(n);
1890 if (irn_not_visited(pred))
1891 waitq_put(worklist, pred);
1893 for (i = irn_arity - 1; i >= 0; --i) {
1894 ir_node *pred = get_irn_n(n, i);
1896 if (irn_not_visited(pred)) {
1897 if (! in_dead_block &&
1898 get_irn_pinned(pred) == op_pin_state_floats &&
1899 is_Block_unreachable(get_nodes_block(pred))) {
1900 set_nodes_block(pred, curr_block);
1902 waitq_put(worklist, pred);
1909 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1910 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1911 * places all floating nodes reachable from its argument through floating
1912 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1914 * @param worklist a worklist, used for the algorithm, empty on in/output
1916 static void place_early(waitq *worklist) {
1918 inc_irg_visited(current_ir_graph);
1920 /* this inits the worklist */
1921 place_floats_early(get_irg_end(current_ir_graph), worklist);
1923 /* Work the content of the worklist. */
1924 while (!waitq_empty(worklist)) {
1925 ir_node *n = waitq_get(worklist);
1926 if (irn_not_visited(n))
1927 place_floats_early(n, worklist);
1930 set_irg_outs_inconsistent(current_ir_graph);
1931 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1935 * Compute the deepest common ancestor of block and dca.
1937 static ir_node *calc_dca(ir_node *dca, ir_node *block) {
1940 /* we do not want to place nodes in dead blocks */
1941 if (is_Block_dead(block))
1944 /* We found a first legal placement. */
1945 if (!dca) return block;
1947 /* Find a placement that is dominates both, dca and block. */
1948 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1949 block = get_Block_idom(block);
1951 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1952 dca = get_Block_idom(dca);
1955 while (block != dca) {
1956 block = get_Block_idom(block); dca = get_Block_idom(dca);
1962 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1963 * I.e., DCA is the block where we might place PRODUCER.
1964 * A data flow edge points from producer to consumer.
1966 static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1968 /* Compute the last block into which we can place a node so that it is
1970 if (is_Phi(consumer)) {
1971 /* our consumer is a Phi-node, the effective use is in all those
1972 blocks through which the Phi-node reaches producer */
1973 ir_node *phi_block = get_nodes_block(consumer);
1974 int arity = get_irn_arity(consumer);
1977 for (i = 0; i < arity; i++) {
1978 if (get_Phi_pred(consumer, i) == producer) {
1979 ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
1981 if (!is_Block_unreachable(new_block))
1982 dca = calc_dca(dca, new_block);
1986 dca = calc_dca(dca, get_nodes_block(consumer));
1992 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1994 static INLINE int get_irn_loop_depth(ir_node *n) {
1995 return get_loop_depth(get_irn_loop(n));
1999 * Move n to a block with less loop depth than it's current block. The
2000 * new block must be dominated by early.
2002 * @param n the node that should be moved
2003 * @param early the earliest block we can n move to
2005 static void move_out_of_loops(ir_node *n, ir_node *early) {
2006 ir_node *best, *dca;
2010 /* Find the region deepest in the dominator tree dominating
2011 dca with the least loop nesting depth, but still dominated
2012 by our early placement. */
2013 dca = get_nodes_block(n);
2016 while (dca != early) {
2017 dca = get_Block_idom(dca);
2018 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
2019 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
2023 if (best != get_nodes_block(n)) {
2025 printf("Moving out of loop: "); DDMN(n);
2026 printf(" Outermost block: "); DDMN(early);
2027 printf(" Best block: "); DDMN(best);
2028 printf(" Innermost block: "); DDMN(get_nodes_block(n));
2030 set_nodes_block(n, best);
2034 /* deepest common ancestor in the dominator tree of all nodes'
2035 blocks depending on us; our final placement has to dominate DCA. */
2036 static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
2040 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2041 ir_node *succ = get_irn_out(node, i);
2045 * This consumer is the End node, a keep alive edge.
2046 * This is not a real consumer, so we ignore it
2051 if (is_Proj(succ)) {
2052 dca = get_deepest_common_ancestor(succ, dca);
2054 /* ignore if succ is in dead code */
2055 ir_node *succ_blk = get_nodes_block(succ);
2056 if (is_Block_unreachable(succ_blk))
2058 dca = consumer_dom_dca(dca, succ, node);
2065 static void set_projs_block(ir_node *node, ir_node *block)
2069 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2070 ir_node *succ = get_irn_out(node, i);
2072 assert(is_Proj(succ));
2074 if(get_irn_mode(succ) == mode_T) {
2075 set_projs_block(succ, block);
2077 set_nodes_block(succ, block);
2082 * Find the latest legal block for N and place N into the
2083 * `optimal' Block between the latest and earliest legal block.
2084 * The `optimal' block is the dominance-deepest block of those
2085 * with the least loop-nesting-depth. This places N out of as many
2086 * loops as possible and then makes it as control dependent as
2089 * @param n the node to be placed
2090 * @param worklist a worklist, all successors of non-floating nodes are
2093 static void place_floats_late(ir_node *n, pdeq *worklist) {
2097 assert(irn_not_visited(n)); /* no multiple placement */
2099 mark_irn_visited(n);
2101 /* no need to place block nodes, control nodes are already placed. */
2104 (get_irn_mode(n) != mode_X)) {
2105 /* Remember the early_blk placement of this block to move it
2106 out of loop no further than the early_blk placement. */
2107 early_blk = get_nodes_block(n);
2110 * BEWARE: Here we also get code, that is live, but
2111 * was in a dead block. If the node is life, but because
2112 * of CSE in a dead block, we still might need it.
2115 /* Assure that our users are all placed, except the Phi-nodes.
2116 --- Each data flow cycle contains at least one Phi-node. We
2117 have to break the `user has to be placed before the
2118 producer' dependence cycle and the Phi-nodes are the
2119 place to do so, because we need to base our placement on the
2120 final region of our users, which is OK with Phi-nodes, as they
2121 are op_pin_state_pinned, and they never have to be placed after a
2122 producer of one of their inputs in the same block anyway. */
2123 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2124 ir_node *succ = get_irn_out(n, i);
2125 if (irn_not_visited(succ) && !is_Phi(succ))
2126 place_floats_late(succ, worklist);
2129 if (! is_Block_dead(early_blk)) {
2130 /* do only move things that where not dead */
2131 ir_op *op = get_irn_op(n);
2133 /* We have to determine the final block of this node... except for
2134 constants and Projs */
2135 if ((get_irn_pinned(n) == op_pin_state_floats) &&
2137 (op != op_SymConst) &&
2140 /* deepest common ancestor in the dominator tree of all nodes'
2141 blocks depending on us; our final placement has to dominate
2143 ir_node *dca = get_deepest_common_ancestor(n, NULL);
2145 set_nodes_block(n, dca);
2146 move_out_of_loops(n, early_blk);
2147 if(get_irn_mode(n) == mode_T) {
2148 set_projs_block(n, get_nodes_block(n));
2155 /* Add successors of all non-floating nodes on list. (Those of floating
2156 nodes are placed already and therefore are marked.) */
2157 for (i = 0; i < get_irn_n_outs(n); i++) {
2158 ir_node *succ = get_irn_out(n, i);
2159 if (irn_not_visited(get_irn_out(n, i))) {
2160 pdeq_putr(worklist, succ);
2166 * Place floating nodes on the given worklist as late as possible using
2167 * the dominance tree.
2169 * @param worklist the worklist containing the nodes to place
2171 static void place_late(waitq *worklist) {
2173 inc_irg_visited(current_ir_graph);
2175 /* This fills the worklist initially. */
2176 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2178 /* And now empty the worklist again... */
2179 while (!waitq_empty(worklist)) {
2180 ir_node *n = waitq_get(worklist);
2181 if (irn_not_visited(n))
2182 place_floats_late(n, worklist);
2186 /* Code Placement. */
2187 void place_code(ir_graph *irg) {
2189 ir_graph *rem = current_ir_graph;
2191 current_ir_graph = irg;
2193 /* Handle graph state */
2194 assert(get_irg_phase_state(irg) != phase_building);
2197 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2198 free_loop_information(irg);
2199 construct_cf_backedges(irg);
2202 /* Place all floating nodes as early as possible. This guarantees
2203 a legal code placement. */
2204 worklist = new_waitq();
2205 place_early(worklist);
2207 /* place_early() invalidates the outs, place_late needs them. */
2208 compute_irg_outs(irg);
2210 /* Now move the nodes down in the dominator tree. This reduces the
2211 unnecessary executions of the node. */
2212 place_late(worklist);
2214 set_irg_outs_inconsistent(current_ir_graph);
2215 set_irg_loopinfo_inconsistent(current_ir_graph);
2216 del_waitq(worklist);
2217 current_ir_graph = rem;
2220 typedef struct cf_env {
2221 char ignore_exc_edges; /**< set if exception edges should be ignored. */
2222 char changed; /**< flag indicates that the cf graphs has changed. */
2226 * Called by walker of remove_critical_cf_edges().
2228 * Place an empty block to an edge between a blocks of multiple
2229 * predecessors and a block of multiple successors.
2232 * @param env Environment of walker.
2234 static void walk_critical_cf_edges(ir_node *n, void *env) {
2236 ir_node *pre, *block, *jmp;
2238 ir_graph *irg = get_irn_irg(n);
2240 /* Block has multiple predecessors */
2241 arity = get_irn_arity(n);
2243 if (n == get_irg_end_block(irg))
2244 return; /* No use to add a block here. */
2246 for (i = 0; i < arity; ++i) {
2249 pre = get_irn_n(n, i);
2250 /* don't count Bad's */
2254 cfop = get_irn_op(skip_Proj(pre));
2255 if (is_op_fragile(cfop)) {
2256 if (cenv->ignore_exc_edges && get_Proj_proj(pre) == pn_Generic_X_except)
2260 /* we don't want place nodes in the start block, so handle it like forking */
2261 if (is_op_forking(cfop) || cfop == op_Start) {
2262 /* Predecessor has multiple successors. Insert new control flow edge edges. */
2264 /* set predecessor of new block */
2265 block = new_r_Block(irg, 1, &pre);
2266 /* insert new jmp node to new block */
2267 jmp = new_r_Jmp(irg, block);
2268 /* set successor of new block */
2269 set_irn_n(n, i, jmp);
2271 } /* predecessor has multiple successors */
2272 } /* for all predecessors */
2273 } /* n is a multi-entry block */
2276 void remove_critical_cf_edges(ir_graph *irg) {
2279 env.ignore_exc_edges = 1;
2282 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &env);
2284 /* control flow changed */
2285 set_irg_outs_inconsistent(irg);
2286 set_irg_extblk_inconsistent(irg);
2287 set_irg_doms_inconsistent(irg);
2288 set_irg_loopinfo_inconsistent(irg);