2 * Copyright (C) 1995-2007 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 int state = edges_activated(irg);
186 ir_graph *rem = current_ir_graph;
190 current_ir_graph = irg;
195 if (get_opt_global_cse())
196 set_irg_pinned(current_ir_graph, op_pin_state_floats);
198 /* Clean the value_table in irg for the CSE. */
199 del_identities(irg->value_table);
200 irg->value_table = new_identities();
202 if (get_irg_dom_state(irg) == dom_consistent)
203 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
205 /* invalidate info */
206 set_irg_outs_inconsistent(irg);
207 set_irg_doms_inconsistent(irg);
208 set_irg_loopinfo_inconsistent(irg);
210 set_using_irn_link(irg);
212 /* walk over the graph, but don't touch keep-alives */
213 irg_walk(get_irg_end_block(irg), NULL, opt_walker, waitq);
215 end = get_irg_end(irg);
217 /* optimize keep-alives by removing superfluous ones */
218 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
219 ir_node *ka = get_End_keepalive(end, i);
221 if (irn_visited(ka) && !is_irn_keep(ka)) {
222 /* this node can be regularly visited, no need to keep it */
223 set_End_keepalive(end, i, get_irg_bad(irg));
226 /* now walk again and visit all not yet visited nodes */
227 set_irg_visited(current_ir_graph, get_irg_visited(irg) - 1);
228 irg_walk(get_irg_end(irg), NULL, opt_walker, waitq);
230 /* finish the wait queue */
231 while (! pdeq_empty(waitq)) {
232 ir_node *n = pdeq_getl(waitq);
234 opt_walker(n, waitq);
239 clear_using_irn_link(irg);
242 edges_deactivate(irg);
244 current_ir_graph = rem;
248 /*------------------------------------------------------------------*/
249 /* Routines for dead node elimination / copying garbage collection */
250 /* of the obstack. */
251 /*------------------------------------------------------------------*/
254 * Remember the new node in the old node by using a field all nodes have.
256 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
259 * Get this new node, before the old node is forgotten.
261 #define get_new_node(oldn) get_irn_link(oldn)
264 * Check if a new node was set.
266 #define has_new_node(n) (get_new_node(n) != NULL)
269 * We use the block_visited flag to mark that we have computed the
270 * number of useful predecessors for this block.
271 * Further we encode the new arity in this flag in the old blocks.
272 * Remembering the arity is useful, as it saves a lot of pointer
273 * accesses. This function is called for all Phi and Block nodes
277 compute_new_arity(ir_node *b) {
278 int i, res, irn_arity;
281 irg_v = get_irg_block_visited(current_ir_graph);
282 block_v = get_Block_block_visited(b);
283 if (block_v >= irg_v) {
284 /* we computed the number of preds for this block and saved it in the
286 return block_v - irg_v;
288 /* compute the number of good predecessors */
289 res = irn_arity = get_irn_arity(b);
290 for (i = 0; i < irn_arity; i++)
291 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
292 /* save it in the flag. */
293 set_Block_block_visited(b, irg_v + res);
299 * Copies the node to the new obstack. The Ins of the new node point to
300 * the predecessors on the old obstack. For block/phi nodes not all
301 * predecessors might be copied. n->link points to the new node.
302 * For Phi and Block nodes the function allocates in-arrays with an arity
303 * only for useful predecessors. The arity is determined by counting
304 * the non-bad predecessors of the block.
306 * @param n The node to be copied
307 * @param env if non-NULL, the node number attribute will be copied to the new node
309 * Note: Also used for loop unrolling.
311 static void copy_node(ir_node *n, void *env) {
314 ir_op *op = get_irn_op(n);
316 /* The end node looses it's flexible in array. This doesn't matter,
317 as dead node elimination builds End by hand, inlineing doesn't use
319 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
322 /* node copied already */
324 } else if (op == op_Block) {
326 new_arity = compute_new_arity(n);
327 n->attr.block.graph_arr = NULL;
329 block = get_nodes_block(n);
331 new_arity = compute_new_arity(block);
333 new_arity = get_irn_arity(n);
336 nn = new_ir_node(get_irn_dbg_info(n),
343 /* Copy the attributes. These might point to additional data. If this
344 was allocated on the old obstack the pointers now are dangling. This
345 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
346 copy_node_attr(n, nn);
350 int copy_node_nr = env != NULL;
352 /* for easier debugging, we want to copy the node numbers too */
353 nn->node_nr = n->node_nr;
359 hook_dead_node_elim_subst(current_ir_graph, n, nn);
363 * Copies new predecessors of old node to new node remembered in link.
364 * Spare the Bad predecessors of Phi and Block nodes.
366 static void copy_preds(ir_node *n, void *env) {
371 nn = get_new_node(n);
374 /* copy the macro block header */
375 ir_node *mbh = get_Block_MacroBlock(n);
378 set_irn_n(nn, -1, get_new_node(mbh));
380 /* Don't copy Bad nodes. */
382 irn_arity = get_irn_arity(n);
383 for (i = 0; i < irn_arity; i++) {
384 if (! is_Bad(get_irn_n(n, i))) {
385 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
386 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
390 /* repair the block visited flag from above misuse. Repair it in both
391 graphs so that the old one can still be used. */
392 set_Block_block_visited(nn, 0);
393 set_Block_block_visited(n, 0);
394 /* Local optimization could not merge two subsequent blocks if
395 in array contained Bads. Now it's possible.
396 We don't call optimize_in_place as it requires
397 that the fields in ir_graph are set properly. */
398 if ((get_opt_control_flow_straightening()) &&
399 (get_Block_n_cfgpreds(nn) == 1) &&
400 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
401 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
403 /* Jmp jumps into the block it is in -- deal self cycle. */
404 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
405 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
410 } else if (is_Phi(n)) {
411 /* Don't copy node if corresponding predecessor in block is Bad.
412 The Block itself should not be Bad. */
413 block = get_nodes_block(n);
414 set_irn_n(nn, -1, get_new_node(block));
416 irn_arity = get_irn_arity(n);
417 for (i = 0; i < irn_arity; i++) {
418 if (! is_Bad(get_irn_n(block, i))) {
419 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
420 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
424 /* If the pre walker reached this Phi after the post walker visited the
425 block block_visited is > 0. */
426 set_Block_block_visited(get_nodes_block(n), 0);
427 /* Compacting the Phi's ins might generate Phis with only one
429 if (get_irn_arity(nn) == 1)
430 exchange(nn, get_irn_n(nn, 0));
432 irn_arity = get_irn_arity(n);
433 for (i = -1; i < irn_arity; i++)
434 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
436 /* Now the new node is complete. We can add it to the hash table for CSE.
437 @@@ inlining aborts if we identify End. Why? */
438 if (get_irn_op(nn) != op_End)
439 add_identities(current_ir_graph->value_table, nn);
443 * Copies the graph recursively, compacts the keep-alives of the end node.
445 * @param irg the graph to be copied
446 * @param copy_node_nr If non-zero, the node number will be copied
448 static void copy_graph(ir_graph *irg, int copy_node_nr) {
449 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
450 ir_node *ka; /* keep alive */
454 /* Some nodes must be copied by hand, sigh */
455 vfl = get_irg_visited(irg);
456 set_irg_visited(irg, vfl + 1);
458 oe = get_irg_end(irg);
459 mark_irn_visited(oe);
460 /* copy the end node by hand, allocate dynamic in array! */
461 ne = new_ir_node(get_irn_dbg_info(oe),
468 /* Copy the attributes. Well, there might be some in the future... */
469 copy_node_attr(oe, ne);
470 set_new_node(oe, ne);
472 /* copy the Bad node */
473 ob = get_irg_bad(irg);
474 mark_irn_visited(ob);
475 nb = new_ir_node(get_irn_dbg_info(ob),
482 copy_node_attr(ob, nb);
483 set_new_node(ob, nb);
485 /* copy the NoMem node */
486 om = get_irg_no_mem(irg);
487 mark_irn_visited(om);
488 nm = new_ir_node(get_irn_dbg_info(om),
495 copy_node_attr(om, nm);
496 set_new_node(om, nm);
498 /* copy the live nodes */
499 set_irg_visited(irg, vfl);
500 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
502 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
504 /* visit the anchors as well */
505 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
506 ir_node *n = get_irg_anchor(irg, i);
508 if (n && (get_irn_visited(n) <= vfl)) {
509 set_irg_visited(irg, vfl);
510 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
514 /* copy_preds for the end node ... */
515 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
517 /*- ... and now the keep alives. -*/
518 /* First pick the not marked block nodes and walk them. We must pick these
519 first as else we will oversee blocks reachable from Phis. */
520 irn_arity = get_End_n_keepalives(oe);
521 for (i = 0; i < irn_arity; i++) {
522 ka = get_End_keepalive(oe, i);
524 if (get_irn_visited(ka) <= vfl) {
525 /* We must keep the block alive and copy everything reachable */
526 set_irg_visited(irg, vfl);
527 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
529 add_End_keepalive(ne, get_new_node(ka));
533 /* Now pick other nodes. Here we will keep all! */
534 irn_arity = get_End_n_keepalives(oe);
535 for (i = 0; i < irn_arity; i++) {
536 ka = get_End_keepalive(oe, i);
538 if (get_irn_visited(ka) <= vfl) {
539 /* We didn't copy the node yet. */
540 set_irg_visited(irg, vfl);
541 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
543 add_End_keepalive(ne, get_new_node(ka));
547 /* start block sometimes only reached after keep alives */
548 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
549 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
553 * Copies the graph reachable from current_ir_graph->end to the obstack
554 * in current_ir_graph and fixes the environment.
555 * Then fixes the fields in current_ir_graph containing nodes of the
558 * @param copy_node_nr If non-zero, the node number will be copied
561 copy_graph_env(int copy_node_nr) {
562 ir_graph *irg = current_ir_graph;
563 ir_node *old_end, *new_anchor;
566 /* remove end_except and end_reg nodes */
567 old_end = get_irg_end(irg);
568 set_irg_end_except (irg, old_end);
569 set_irg_end_reg (irg, old_end);
571 /* Not all nodes remembered in irg might be reachable
572 from the end node. Assure their link is set to NULL, so that
573 we can test whether new nodes have been computed. */
574 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
575 ir_node *n = get_irg_anchor(irg, i);
577 set_new_node(n, NULL);
579 /* we use the block walk flag for removing Bads from Blocks ins. */
580 inc_irg_block_visited(irg);
583 copy_graph(irg, copy_node_nr);
586 old_end = get_irg_end(irg);
587 new_anchor = new_Anchor(irg);
589 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
590 ir_node *n = get_irg_anchor(irg, i);
592 set_irn_n(new_anchor, i, get_new_node(n));
595 irg->anchor = new_anchor;
597 /* ensure the new anchor is placed in the endblock */
598 set_irn_n(new_anchor, -1, get_irg_end_block(irg));
602 * Copies all reachable nodes to a new obstack. Removes bad inputs
603 * from block nodes and the corresponding inputs from Phi nodes.
604 * Merges single exit blocks with single entry blocks and removes
606 * Adds all new nodes to a new hash table for CSE. Does not
607 * perform CSE, so the hash table might contain common subexpressions.
610 dead_node_elimination(ir_graph *irg) {
611 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
613 int rem_ipview = get_interprocedural_view();
614 struct obstack *graveyard_obst = NULL;
615 struct obstack *rebirth_obst = NULL;
616 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
618 /* inform statistics that we started a dead-node elimination run */
619 hook_dead_node_elim(irg, 1);
621 /* Remember external state of current_ir_graph. */
622 rem = current_ir_graph;
623 current_ir_graph = irg;
624 set_interprocedural_view(0);
626 assert(get_irg_phase_state(irg) != phase_building);
628 /* Handle graph state */
629 free_callee_info(irg);
633 /* @@@ so far we loose loops when copying */
634 free_loop_information(irg);
636 set_irg_doms_inconsistent(irg);
638 /* A quiet place, where the old obstack can rest in peace,
639 until it will be cremated. */
640 graveyard_obst = irg->obst;
642 /* A new obstack, where the reachable nodes will be copied to. */
643 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
644 irg->obst = rebirth_obst;
645 obstack_init(irg->obst);
646 irg->last_node_idx = 0;
648 /* We also need a new value table for CSE */
649 del_identities(irg->value_table);
650 irg->value_table = new_identities();
652 /* Copy the graph from the old to the new obstack */
653 copy_graph_env(/*copy_node_nr=*/1);
655 /* Free memory from old unoptimized obstack */
656 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
657 xfree (graveyard_obst); /* ... then free it. */
659 /* inform statistics that the run is over */
660 hook_dead_node_elim(irg, 0);
662 current_ir_graph = rem;
663 set_interprocedural_view(rem_ipview);
668 * Relink bad predecessors of a block and store the old in array to the
669 * link field. This function is called by relink_bad_predecessors().
670 * The array of link field starts with the block operand at position 0.
671 * If block has bad predecessors, create a new in array without bad preds.
672 * Otherwise let in array untouched.
674 static void relink_bad_block_predecessors(ir_node *n, void *env) {
675 ir_node **new_in, *irn;
676 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
679 /* if link field of block is NULL, look for bad predecessors otherwise
680 this is already done */
681 if (get_irn_op(n) == op_Block &&
682 get_irn_link(n) == NULL) {
684 /* save old predecessors in link field (position 0 is the block operand)*/
685 set_irn_link(n, get_irn_in(n));
687 /* count predecessors without bad nodes */
688 old_irn_arity = get_irn_arity(n);
689 for (i = 0; i < old_irn_arity; i++)
690 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
692 /* arity changing: set new predecessors without bad nodes */
693 if (new_irn_arity < old_irn_arity) {
694 /* Get new predecessor array. We do not resize the array, as we must
695 keep the old one to update Phis. */
696 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
698 /* set new predecessors in array */
701 for (i = 0; i < old_irn_arity; i++) {
702 irn = get_irn_n(n, i);
704 new_in[new_irn_n] = irn;
705 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
709 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
710 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
712 } /* ir node has bad predecessors */
713 } /* Block is not relinked */
717 * Relinks Bad predecessors from Blocks and Phis called by walker
718 * remove_bad_predecesors(). If n is a Block, call
719 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
720 * function of Phi's Block. If this block has bad predecessors, relink preds
723 static void relink_bad_predecessors(ir_node *n, void *env) {
724 ir_node *block, **old_in;
725 int i, old_irn_arity, new_irn_arity;
727 /* relink bad predecessors of a block */
728 if (get_irn_op(n) == op_Block)
729 relink_bad_block_predecessors(n, env);
731 /* If Phi node relink its block and its predecessors */
732 if (get_irn_op(n) == op_Phi) {
734 /* Relink predecessors of phi's block */
735 block = get_nodes_block(n);
736 if (get_irn_link(block) == NULL)
737 relink_bad_block_predecessors(block, env);
739 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
740 old_irn_arity = ARR_LEN(old_in);
742 /* Relink Phi predecessors if count of predecessors changed */
743 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
744 /* set new predecessors in array
745 n->in[0] remains the same block */
747 for(i = 1; i < old_irn_arity; i++)
748 if (!is_Bad((ir_node *)old_in[i])) {
749 n->in[new_irn_arity] = n->in[i];
750 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
754 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
755 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
757 } /* n is a Phi node */
761 * Removes Bad Bad predecessors from Blocks and the corresponding
762 * inputs to Phi nodes as in dead_node_elimination but without
764 * On walking up set the link field to NULL, on walking down call
765 * relink_bad_predecessors() (This function stores the old in array
766 * to the link field and sets a new in array if arity of predecessors
769 void remove_bad_predecessors(ir_graph *irg) {
770 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
777 __)|_| | \_/ | \_/(/_ |_/\__|__
779 The following stuff implements a facility that automatically patches
780 registered ir_node pointers to the new node when a dead node elimination occurs.
783 struct _survive_dce_t {
787 hook_entry_t dead_node_elim;
788 hook_entry_t dead_node_elim_subst;
791 typedef struct _survive_dce_list_t {
792 struct _survive_dce_list_t *next;
794 } survive_dce_list_t;
796 static void dead_node_hook(void *context, ir_graph *irg, int start) {
797 survive_dce_t *sd = context;
800 /* Create a new map before the dead node elimination is performed. */
802 sd->new_places = pmap_create_ex(pmap_count(sd->places));
804 /* Patch back all nodes if dead node elimination is over and something is to be done. */
805 pmap_destroy(sd->places);
806 sd->places = sd->new_places;
807 sd->new_places = NULL;
812 * Hook called when dead node elimination replaces old by nw.
814 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
815 survive_dce_t *sd = context;
816 survive_dce_list_t *list = pmap_get(sd->places, old);
819 /* If the node is to be patched back, write the new address to all registered locations. */
821 survive_dce_list_t *p;
823 for (p = list; p; p = p->next)
826 pmap_insert(sd->new_places, nw, list);
831 * Make a new Survive DCE environment.
833 survive_dce_t *new_survive_dce(void) {
834 survive_dce_t *res = xmalloc(sizeof(res[0]));
835 obstack_init(&res->obst);
836 res->places = pmap_create();
837 res->new_places = NULL;
839 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
840 res->dead_node_elim.context = res;
841 res->dead_node_elim.next = NULL;
843 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
844 res->dead_node_elim_subst.context = res;
845 res->dead_node_elim_subst.next = NULL;
847 #ifndef FIRM_ENABLE_HOOKS
848 assert(0 && "need hooks enabled");
851 register_hook(hook_dead_node_elim, &res->dead_node_elim);
852 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
857 * Free a Survive DCE environment.
859 void free_survive_dce(survive_dce_t *sd) {
860 obstack_free(&sd->obst, NULL);
861 pmap_destroy(sd->places);
862 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
863 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
868 * Register a node pointer to be patched upon DCE.
869 * When DCE occurs, the node pointer specified by @p place will be
870 * patched to the new address of the node it is pointing to.
872 * @param sd The Survive DCE environment.
873 * @param place The address of the node pointer.
875 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
876 if (*place != NULL) {
877 ir_node *irn = *place;
878 survive_dce_list_t *curr = pmap_get(sd->places, irn);
879 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
884 pmap_insert(sd->places, irn, nw);
888 /*--------------------------------------------------------------------*/
889 /* Functionality for inlining */
890 /*--------------------------------------------------------------------*/
893 * Copy node for inlineing. Updates attributes that change when
894 * inlineing but not for dead node elimination.
896 * Copies the node by calling copy_node() and then updates the entity if
897 * it's a local one. env must be a pointer of the frame type of the
898 * inlined procedure. The new entities must be in the link field of
902 copy_node_inline(ir_node *n, void *env) {
904 ir_type *frame_tp = (ir_type *)env;
907 if (get_irn_op(n) == op_Sel) {
908 nn = get_new_node (n);
910 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
911 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
913 } else if (get_irn_op(n) == op_Block) {
914 nn = get_new_node (n);
915 nn->attr.block.irg = current_ir_graph;
920 * Walker: checks if P_value_arg_base is used.
922 static void find_addr(ir_node *node, void *env) {
923 int *allow_inline = env;
924 if (is_Proj(node) && get_irn_op(get_Proj_pred(node)) == op_Start) {
925 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
931 * Check if we can inline a given call.
932 * Currently, we cannot inline two cases:
933 * - call with compound arguments
934 * - graphs that take the address of a parameter
936 * check these conditions here
938 static int can_inline(ir_node *call, ir_graph *called_graph) {
939 ir_type *call_type = get_Call_type(call);
940 int params, ress, i, res;
941 assert(is_Method_type(call_type));
943 params = get_method_n_params(call_type);
944 ress = get_method_n_ress(call_type);
946 /* check parameters for compound arguments */
947 for (i = 0; i < params; ++i) {
948 ir_type *p_type = get_method_param_type(call_type, i);
950 if (is_compound_type(p_type))
954 /* check results for compound arguments */
955 for (i = 0; i < ress; ++i) {
956 ir_type *r_type = get_method_res_type(call_type, i);
958 if (is_compound_type(r_type))
963 irg_walk_graph(called_graph, find_addr, NULL, &res);
969 exc_handler = 0, /**< There is a handler. */
970 exc_to_end = 1, /**< Branches to End. */
971 exc_no_handler = 2 /**< Exception handling not represented. */
974 /* Inlines a method at the given call site. */
975 int inline_method(ir_node *call, ir_graph *called_graph) {
977 ir_node *post_call, *post_bl;
978 ir_node *in[pn_Start_max];
979 ir_node *end, *end_bl;
983 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
984 enum exc_mode exc_handling;
985 ir_type *called_frame;
986 irg_inline_property prop = get_irg_inline_property(called_graph);
988 if ( (prop < irg_inline_forced) &&
989 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
991 /* Do not inline variadic functions. */
992 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
995 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
996 get_method_n_params(get_Call_type(call)));
999 * currently, we cannot inline two cases:
1000 * - call with compound arguments
1001 * - graphs that take the address of a parameter
1003 if (! can_inline(call, called_graph))
1006 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
1007 rem_opt = get_opt_optimize();
1010 /* Handle graph state */
1011 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1012 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
1013 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
1014 set_irg_outs_inconsistent(current_ir_graph);
1015 set_irg_extblk_inconsistent(current_ir_graph);
1016 set_irg_doms_inconsistent(current_ir_graph);
1017 set_irg_loopinfo_inconsistent(current_ir_graph);
1018 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
1020 /* -- Check preconditions -- */
1021 assert(is_Call(call));
1022 /* @@@ does not work for InterfaceIII.java after cgana
1023 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
1024 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
1025 get_Call_type(call)));
1027 if (called_graph == current_ir_graph) {
1028 set_optimize(rem_opt);
1032 /* here we know we WILL inline, so inform the statistics */
1033 hook_inline(call, called_graph);
1035 /* -- Decide how to handle exception control flow: Is there a handler
1036 for the Call node, or do we branch directly to End on an exception?
1038 0 There is a handler.
1040 2 Exception handling not represented in Firm. -- */
1042 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
1043 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
1044 long proj_nr = get_Proj_proj(proj);
1045 if (proj_nr == pn_Call_X_except) Xproj = proj;
1046 if (proj_nr == pn_Call_M_except) Mproj = proj;
1048 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
1049 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
1050 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
1054 the procedure and later replaces the Start node of the called graph.
1055 Post_call is the old Call node and collects the results of the called
1056 graph. Both will end up being a tuple. -- */
1057 post_bl = get_nodes_block(call);
1058 set_irg_current_block(current_ir_graph, post_bl);
1059 /* XxMxPxPxPxT of Start + parameter of Call */
1060 in[pn_Start_X_initial_exec] = new_Jmp();
1061 in[pn_Start_M] = get_Call_mem(call);
1062 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1063 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1064 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1065 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1066 /* in[pn_Start_P_value_arg_base] = ??? */
1067 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1068 pre_call = new_Tuple(pn_Start_max - 1, in);
1072 The new block gets the ins of the old block, pre_call and all its
1073 predecessors and all Phi nodes. -- */
1074 part_block(pre_call);
1076 /* -- Prepare state for dead node elimination -- */
1077 /* Visited flags in calling irg must be >= flag in called irg.
1078 Else walker and arity computation will not work. */
1079 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1080 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1081 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1082 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1083 /* Set pre_call as new Start node in link field of the start node of
1084 calling graph and pre_calls block as new block for the start block
1086 Further mark these nodes so that they are not visited by the
1088 set_irn_link(get_irg_start(called_graph), pre_call);
1089 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1090 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1091 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1092 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1093 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1095 /* Initialize for compaction of in arrays */
1096 inc_irg_block_visited(current_ir_graph);
1098 /* -- Replicate local entities of the called_graph -- */
1099 /* copy the entities. */
1100 called_frame = get_irg_frame_type(called_graph);
1101 for (i = 0; i < get_class_n_members(called_frame); i++) {
1102 ir_entity *new_ent, *old_ent;
1103 old_ent = get_class_member(called_frame, i);
1104 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1105 set_entity_link(old_ent, new_ent);
1108 /* visited is > than that of called graph. With this trick visited will
1109 remain unchanged so that an outer walker, e.g., searching the call nodes
1110 to inline, calling this inline will not visit the inlined nodes. */
1111 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1113 /* -- Performing dead node elimination inlines the graph -- */
1114 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1116 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1117 get_irg_frame_type(called_graph));
1119 /* Repair called_graph */
1120 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1121 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1122 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1124 /* -- Merge the end of the inlined procedure with the call site -- */
1125 /* We will turn the old Call node into a Tuple with the following
1128 0: Phi of all Memories of Return statements.
1129 1: Jmp from new Block that merges the control flow from all exception
1130 predecessors of the old end block.
1131 2: Tuple of all arguments.
1132 3: Phi of Exception memories.
1133 In case the old Call directly branches to End on an exception we don't
1134 need the block merging all exceptions nor the Phi of the exception
1138 /* -- Precompute some values -- */
1139 end_bl = get_new_node(get_irg_end_block(called_graph));
1140 end = get_new_node(get_irg_end(called_graph));
1141 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1142 n_res = get_method_n_ress(get_Call_type(call));
1144 res_pred = xmalloc(n_res * sizeof(*res_pred));
1145 cf_pred = xmalloc(arity * sizeof(*res_pred));
1147 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1149 /* -- archive keepalives -- */
1150 irn_arity = get_irn_arity(end);
1151 for (i = 0; i < irn_arity; i++) {
1152 ir_node *ka = get_End_keepalive(end, i);
1154 add_End_keepalive(get_irg_end(current_ir_graph), ka);
1157 /* The new end node will die. We need not free as the in array is on the obstack:
1158 copy_node() only generated 'D' arrays. */
1160 /* -- Replace Return nodes by Jump nodes. -- */
1162 for (i = 0; i < arity; i++) {
1164 ret = get_irn_n(end_bl, i);
1165 if (is_Return(ret)) {
1166 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1170 set_irn_in(post_bl, n_ret, cf_pred);
1172 /* -- Build a Tuple for all results of the method.
1173 Add Phi node if there was more than one Return. -- */
1174 turn_into_tuple(post_call, pn_Call_max);
1175 /* First the Memory-Phi */
1177 for (i = 0; i < arity; i++) {
1178 ret = get_irn_n(end_bl, i);
1179 if (is_Return(ret)) {
1180 cf_pred[n_ret] = get_Return_mem(ret);
1184 phi = new_Phi(n_ret, cf_pred, mode_M);
1185 set_Tuple_pred(call, pn_Call_M_regular, phi);
1186 /* Conserve Phi-list for further inlinings -- but might be optimized */
1187 if (get_nodes_block(phi) == post_bl) {
1188 set_irn_link(phi, get_irn_link(post_bl));
1189 set_irn_link(post_bl, phi);
1191 /* Now the real results */
1193 for (j = 0; j < n_res; j++) {
1195 for (i = 0; i < arity; i++) {
1196 ret = get_irn_n(end_bl, i);
1197 if (get_irn_op(ret) == op_Return) {
1198 cf_pred[n_ret] = get_Return_res(ret, j);
1203 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1207 /* Conserve Phi-list for further inlinings -- but might be optimized */
1208 if (get_nodes_block(phi) == post_bl) {
1209 set_irn_link(phi, get_irn_link(post_bl));
1210 set_irn_link(post_bl, phi);
1213 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1215 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1218 /* For now, we cannot inline calls with value_base */
1219 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1221 /* Finally the exception control flow.
1222 We have two (three) possible situations:
1223 First if the Call branches to an exception handler: We need to add a Phi node to
1224 collect the memory containing the exception objects. Further we need
1225 to add another block to get a correct representation of this Phi. To
1226 this block we add a Jmp that resolves into the X output of the Call
1227 when the Call is turned into a tuple.
1228 Second the Call branches to End, the exception is not handled. Just
1229 add all inlined exception branches to the End node.
1230 Third: there is no Exception edge at all. Handle as case two. */
1231 if (exc_handling == exc_handler) {
1233 for (i = 0; i < arity; i++) {
1235 ret = get_irn_n(end_bl, i);
1236 irn = skip_Proj(ret);
1237 if (is_fragile_op(irn) || (get_irn_op(irn) == op_Raise)) {
1238 cf_pred[n_exc] = ret;
1243 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1244 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1245 /* The Phi for the memories with the exception objects */
1247 for (i = 0; i < arity; i++) {
1249 ret = skip_Proj(get_irn_n(end_bl, i));
1251 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1253 } else if (is_fragile_op(ret)) {
1254 /* We rely that all cfops have the memory output at the same position. */
1255 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1257 } else if (get_irn_op(ret) == op_Raise) {
1258 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1262 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1264 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1265 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1267 set_Tuple_pred(call, pn_Call_X_regular, new_Bad());
1269 ir_node *main_end_bl;
1270 int main_end_bl_arity;
1271 ir_node **end_preds;
1273 /* assert(exc_handling == 1 || no exceptions. ) */
1275 for (i = 0; i < arity; i++) {
1276 ir_node *ret = get_irn_n(end_bl, i);
1277 ir_node *irn = skip_Proj(ret);
1279 if (is_fragile_op(irn) || (get_irn_op(irn) == op_Raise)) {
1280 cf_pred[n_exc] = ret;
1284 main_end_bl = get_irg_end_block(current_ir_graph);
1285 main_end_bl_arity = get_irn_arity(main_end_bl);
1286 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1288 for (i = 0; i < main_end_bl_arity; ++i)
1289 end_preds[i] = get_irn_n(main_end_bl, i);
1290 for (i = 0; i < n_exc; ++i)
1291 end_preds[main_end_bl_arity + i] = cf_pred[i];
1292 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1293 set_Tuple_pred(call, pn_Call_X_regular, new_Bad());
1294 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1295 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1301 /* -- Turn CSE back on. -- */
1302 set_optimize(rem_opt);
1307 /********************************************************************/
1308 /* Apply inlineing to small methods. */
1309 /********************************************************************/
1311 /** Represents a possible inlinable call in a graph. */
1312 typedef struct _call_entry call_entry;
1313 struct _call_entry {
1314 ir_node *call; /**< the Call */
1315 ir_graph *callee; /**< the callee called here */
1316 call_entry *next; /**< for linking the next one */
1320 * environment for inlining small irgs
1322 typedef struct _inline_env_t {
1323 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1324 call_entry *head; /**< the head of the call entry list */
1325 call_entry *tail; /**< the tail of the call entry list */
1329 * Returns the irg called from a Call node. If the irg is not
1330 * known, NULL is returned.
1332 static ir_graph *get_call_called_irg(ir_node *call) {
1334 ir_graph *called_irg = NULL;
1336 addr = get_Call_ptr(call);
1337 if (is_SymConst(addr) && get_SymConst_kind(addr) == symconst_addr_ent) {
1338 called_irg = get_entity_irg(get_SymConst_entity(addr));
1345 * Walker: Collect all calls to known graphs inside a graph.
1347 static void collect_calls(ir_node *call, void *env) {
1348 if (is_Call(call)) {
1349 ir_graph *called_irg = get_call_called_irg(call);
1351 /* The Call node calls a locally defined method. Remember to inline. */
1352 inline_env_t *ienv = env;
1353 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1355 entry->callee = called_irg;
1358 if (ienv->tail == NULL)
1361 ienv->tail->next = entry;
1368 * Inlines all small methods at call sites where the called address comes
1369 * from a Const node that references the entity representing the called
1371 * The size argument is a rough measure for the code size of the method:
1372 * Methods where the obstack containing the firm graph is smaller than
1375 void inline_small_irgs(ir_graph *irg, int size) {
1376 ir_graph *rem = current_ir_graph;
1379 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1381 if (!(get_opt_optimize() && get_opt_inline())) return;
1383 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1385 current_ir_graph = irg;
1386 /* Handle graph state */
1387 assert(get_irg_phase_state(irg) != phase_building);
1388 free_callee_info(irg);
1390 /* Find Call nodes to inline.
1391 (We can not inline during a walk of the graph, as inlineing the same
1392 method several times changes the visited flag of the walked graph:
1393 after the first inlineing visited of the callee equals visited of
1394 the caller. With the next inlineing both are increased.) */
1395 obstack_init(&env.obst);
1396 env.head = env.tail = NULL;
1397 irg_walk_graph(irg, NULL, collect_calls, &env);
1399 if (env.head != NULL) {
1400 /* There are calls to inline */
1401 collect_phiprojs(irg);
1402 for (entry = env.head; entry != NULL; entry = entry->next) {
1403 ir_graph *callee = entry->callee;
1404 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1405 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1406 inline_method(entry->call, callee);
1410 obstack_free(&env.obst, NULL);
1411 current_ir_graph = rem;
1415 * Environment for inlining irgs.
1418 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1419 int n_nodes_orig; /**< for statistics */
1420 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1421 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1422 int n_call_nodes; /**< Number of Call nodes in the graph. */
1423 int n_call_nodes_orig; /**< for statistics */
1424 int n_callers; /**< Number of known graphs that call this graphs. */
1425 int n_callers_orig; /**< for statistics */
1426 int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
1430 * Allocate a new environment for inlining.
1432 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1433 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1434 env->n_nodes = -2; /* do not count count Start, End */
1435 env->n_nodes_orig = -2; /* do not count Start, End */
1436 env->call_head = NULL;
1437 env->call_tail = NULL;
1438 env->n_call_nodes = 0;
1439 env->n_call_nodes_orig = 0;
1441 env->n_callers_orig = 0;
1442 env->got_inline = 0;
1446 typedef struct walker_env {
1447 struct obstack *obst; /**< the obstack for allocations. */
1448 inline_irg_env *x; /**< the inline environment */
1449 int ignore_runtime; /**< the ignore runtime flag */
1453 * post-walker: collect all calls in the inline-environment
1454 * of a graph and sum some statistics.
1456 static void collect_calls2(ir_node *call, void *ctx) {
1458 inline_irg_env *x = env->x;
1459 ir_op *op = get_irn_op(call);
1463 /* count meaningful nodes in irg */
1464 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1469 if (op != op_Call) return;
1471 /* check, if it's a runtime call */
1472 if (env->ignore_runtime) {
1473 ir_node *symc = get_Call_ptr(call);
1475 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1476 ir_entity *ent = get_SymConst_entity(symc);
1478 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1483 /* collect all call nodes */
1485 ++x->n_call_nodes_orig;
1487 callee = get_call_called_irg(call);
1489 inline_irg_env *callee_env = get_irg_link(callee);
1490 /* count all static callers */
1491 ++callee_env->n_callers;
1492 ++callee_env->n_callers_orig;
1494 /* link it in the list of possible inlinable entries */
1495 entry = obstack_alloc(env->obst, sizeof(*entry));
1497 entry->callee = callee;
1499 if (x->call_tail == NULL)
1500 x->call_head = entry;
1502 x->call_tail->next = entry;
1503 x->call_tail = entry;
1508 * Returns TRUE if the number of callers in 0 in the irg's environment,
1509 * hence this irg is a leave.
1511 INLINE static int is_leave(ir_graph *irg) {
1512 inline_irg_env *env = get_irg_link(irg);
1513 return env->n_call_nodes == 0;
1517 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1519 INLINE static int is_smaller(ir_graph *callee, int size) {
1520 inline_irg_env *env = get_irg_link(callee);
1521 return env->n_nodes < size;
1525 * Append the nodes of the list src to the nodes of the list in environment dst.
1527 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1528 call_entry *entry, *nentry;
1530 /* Note that the src list points to Call nodes in the inlined graph, but
1531 we need Call nodes in our graph. Luckily the inliner leaves this information
1532 in the link field. */
1533 for (entry = src; entry != NULL; entry = entry->next) {
1534 nentry = obstack_alloc(obst, sizeof(*nentry));
1535 nentry->call = get_irn_link(entry->call);
1536 nentry->callee = entry->callee;
1537 nentry->next = NULL;
1538 dst->call_tail->next = nentry;
1539 dst->call_tail = nentry;
1544 * Inlines small leave methods at call sites where the called address comes
1545 * from a Const node that references the entity representing the called
1547 * The size argument is a rough measure for the code size of the method:
1548 * Methods where the obstack containing the firm graph is smaller than
1551 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1552 inline_irg_env *env;
1558 call_entry *entry, *tail;
1559 const call_entry *centry;
1560 struct obstack obst;
1561 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1563 if (!(get_opt_optimize() && get_opt_inline())) return;
1565 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1566 rem = current_ir_graph;
1567 obstack_init(&obst);
1569 /* extend all irgs by a temporary data structure for inlining. */
1570 n_irgs = get_irp_n_irgs();
1571 for (i = 0; i < n_irgs; ++i)
1572 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1574 /* Precompute information in temporary data structure. */
1576 wenv.ignore_runtime = ignore_runtime;
1577 for (i = 0; i < n_irgs; ++i) {
1578 ir_graph *irg = get_irp_irg(i);
1580 assert(get_irg_phase_state(irg) != phase_building);
1581 free_callee_info(irg);
1583 wenv.x = get_irg_link(irg);
1584 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1587 /* -- and now inline. -- */
1589 /* Inline leaves recursively -- we might construct new leaves. */
1593 for (i = 0; i < n_irgs; ++i) {
1595 int phiproj_computed = 0;
1597 current_ir_graph = get_irp_irg(i);
1598 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1601 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1604 if (env->n_nodes > maxsize) break;
1607 callee = entry->callee;
1609 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1610 if (!phiproj_computed) {
1611 phiproj_computed = 1;
1612 collect_phiprojs(current_ir_graph);
1614 did_inline = inline_method(call, callee);
1617 /* Do some statistics */
1618 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1620 env->got_inline = 1;
1621 --env->n_call_nodes;
1622 env->n_nodes += callee_env->n_nodes;
1623 --callee_env->n_callers;
1625 /* remove this call from the list */
1627 tail->next = entry->next;
1629 env->call_head = entry->next;
1635 env->call_tail = tail;
1637 } while (did_inline);
1639 /* inline other small functions. */
1640 for (i = 0; i < n_irgs; ++i) {
1642 int phiproj_computed = 0;
1644 current_ir_graph = get_irp_irg(i);
1645 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1647 /* note that the list of possible calls is updated during the process */
1649 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1653 callee = entry->callee;
1655 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1656 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1657 if (!phiproj_computed) {
1658 phiproj_computed = 1;
1659 collect_phiprojs(current_ir_graph);
1661 if (inline_method(call, callee)) {
1662 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1664 /* callee was inline. Append it's call list. */
1665 env->got_inline = 1;
1666 --env->n_call_nodes;
1667 append_call_list(&obst, env, callee_env->call_head);
1668 env->n_call_nodes += callee_env->n_call_nodes;
1669 env->n_nodes += callee_env->n_nodes;
1670 --callee_env->n_callers;
1672 /* after we have inlined callee, all called methods inside callee
1673 are now called once more */
1674 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1675 inline_irg_env *penv = get_irg_link(centry->callee);
1679 /* remove this call from the list */
1681 tail->next = entry->next;
1683 env->call_head = entry->next;
1689 env->call_tail = tail;
1692 for (i = 0; i < n_irgs; ++i) {
1693 irg = get_irp_irg(i);
1694 env = (inline_irg_env *)get_irg_link(irg);
1696 if (env->got_inline) {
1697 /* this irg got calls inlined */
1698 set_irg_outs_inconsistent(irg);
1699 set_irg_doms_inconsistent(irg);
1701 optimize_graph_df(irg);
1704 if (env->got_inline || (env->n_callers_orig != env->n_callers))
1705 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1706 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1707 env->n_callers_orig, env->n_callers,
1708 get_entity_name(get_irg_entity(irg))));
1711 obstack_free(&obst, NULL);
1712 current_ir_graph = rem;
1715 /*******************************************************************/
1716 /* Code Placement. Pins all floating nodes to a block where they */
1717 /* will be executed only if needed. */
1718 /*******************************************************************/
1721 * Returns non-zero, is a block is not reachable from Start.
1723 * @param block the block to test
1726 is_Block_unreachable(ir_node *block) {
1727 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1731 * Find the earliest correct block for node n. --- Place n into the
1732 * same Block as its dominance-deepest Input.
1734 * We have to avoid calls to get_nodes_block() here
1735 * because the graph is floating.
1737 * move_out_of_loops() expects that place_floats_early() have placed
1738 * all "living" nodes into a living block. That's why we must
1739 * move nodes in dead block with "live" successors into a valid
1741 * We move them just into the same block as it's successor (or
1742 * in case of a Phi into the effective use block). For Phi successors,
1743 * this may still be a dead block, but then there is no real use, as
1744 * the control flow will be dead later.
1746 * @param n the node to be placed
1747 * @param worklist a worklist, predecessors of non-floating nodes are placed here
1750 place_floats_early(ir_node *n, waitq *worklist) {
1753 /* we must not run into an infinite loop */
1754 assert(irn_not_visited(n));
1755 mark_irn_visited(n);
1757 /* Place floating nodes. */
1758 if (get_irn_pinned(n) == op_pin_state_floats) {
1759 ir_node *curr_block = get_irn_n(n, -1);
1760 int in_dead_block = is_Block_unreachable(curr_block);
1762 ir_node *b = NULL; /* The block to place this node in */
1764 assert(is_no_Block(n));
1766 if (is_irn_start_block_placed(n)) {
1767 /* These nodes will not be placed by the loop below. */
1768 b = get_irg_start_block(current_ir_graph);
1772 /* find the block for this node. */
1773 irn_arity = get_irn_arity(n);
1774 for (i = 0; i < irn_arity; i++) {
1775 ir_node *pred = get_irn_n(n, i);
1776 ir_node *pred_block;
1778 if ((irn_not_visited(pred))
1779 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1782 * If the current node is NOT in a dead block, but one of its
1783 * predecessors is, we must move the predecessor to a live block.
1784 * Such thing can happen, if global CSE chose a node from a dead block.
1785 * We move it simply to our block.
1786 * Note that neither Phi nor End nodes are floating, so we don't
1787 * need to handle them here.
1789 if (! in_dead_block) {
1790 if (get_irn_pinned(pred) == op_pin_state_floats &&
1791 is_Block_unreachable(get_irn_n(pred, -1)))
1792 set_nodes_block(pred, curr_block);
1794 place_floats_early(pred, worklist);
1798 * A node in the Bad block must stay in the bad block,
1799 * so don't compute a new block for it.
1804 /* Because all loops contain at least one op_pin_state_pinned node, now all
1805 our inputs are either op_pin_state_pinned or place_early() has already
1806 been finished on them. We do not have any unfinished inputs! */
1807 pred_block = get_irn_n(pred, -1);
1808 if ((!is_Block_dead(pred_block)) &&
1809 (get_Block_dom_depth(pred_block) > depth)) {
1811 depth = get_Block_dom_depth(pred_block);
1813 /* Avoid that the node is placed in the Start block */
1814 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)
1815 && get_irg_phase_state(current_ir_graph) != phase_backend) {
1816 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1817 assert(b != get_irg_start_block(current_ir_graph));
1822 set_nodes_block(n, b);
1826 * Add predecessors of non floating nodes and non-floating predecessors
1827 * of floating nodes to worklist and fix their blocks if the are in dead block.
1829 irn_arity = get_irn_arity(n);
1831 if (get_irn_op(n) == op_End) {
1833 * Simplest case: End node. Predecessors are keep-alives,
1834 * no need to move out of dead block.
1836 for (i = -1; i < irn_arity; ++i) {
1837 ir_node *pred = get_irn_n(n, i);
1838 if (irn_not_visited(pred))
1839 waitq_put(worklist, pred);
1841 } else if (is_Block(n)) {
1843 * Blocks: Predecessors are control flow, no need to move
1844 * them out of dead block.
1846 for (i = irn_arity - 1; i >= 0; --i) {
1847 ir_node *pred = get_irn_n(n, i);
1848 if (irn_not_visited(pred))
1849 waitq_put(worklist, pred);
1851 } else if (is_Phi(n)) {
1853 ir_node *curr_block = get_irn_n(n, -1);
1854 int in_dead_block = is_Block_unreachable(curr_block);
1857 * Phi nodes: move nodes from dead blocks into the effective use
1858 * of the Phi-input if the Phi is not in a bad block.
1860 pred = get_irn_n(n, -1);
1861 if (irn_not_visited(pred))
1862 waitq_put(worklist, pred);
1864 for (i = irn_arity - 1; i >= 0; --i) {
1865 ir_node *pred = get_irn_n(n, i);
1867 if (irn_not_visited(pred)) {
1868 if (! in_dead_block &&
1869 get_irn_pinned(pred) == op_pin_state_floats &&
1870 is_Block_unreachable(get_irn_n(pred, -1))) {
1871 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1873 waitq_put(worklist, pred);
1878 ir_node *curr_block = get_irn_n(n, -1);
1879 int in_dead_block = is_Block_unreachable(curr_block);
1882 * All other nodes: move nodes from dead blocks into the same block.
1884 pred = get_irn_n(n, -1);
1885 if (irn_not_visited(pred))
1886 waitq_put(worklist, pred);
1888 for (i = irn_arity - 1; i >= 0; --i) {
1889 ir_node *pred = get_irn_n(n, i);
1891 if (irn_not_visited(pred)) {
1892 if (! in_dead_block &&
1893 get_irn_pinned(pred) == op_pin_state_floats &&
1894 is_Block_unreachable(get_irn_n(pred, -1))) {
1895 set_nodes_block(pred, curr_block);
1897 waitq_put(worklist, pred);
1904 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1905 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1906 * places all floating nodes reachable from its argument through floating
1907 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1909 * @param worklist a worklist, used for the algorithm, empty on in/output
1911 static void place_early(waitq *worklist) {
1913 inc_irg_visited(current_ir_graph);
1915 /* this inits the worklist */
1916 place_floats_early(get_irg_end(current_ir_graph), worklist);
1918 /* Work the content of the worklist. */
1919 while (!waitq_empty(worklist)) {
1920 ir_node *n = waitq_get(worklist);
1921 if (irn_not_visited(n))
1922 place_floats_early(n, worklist);
1925 set_irg_outs_inconsistent(current_ir_graph);
1926 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1930 * Compute the deepest common ancestor of block and dca.
1932 static ir_node *calc_dca(ir_node *dca, ir_node *block) {
1935 /* we do not want to place nodes in dead blocks */
1936 if (is_Block_dead(block))
1939 /* We found a first legal placement. */
1940 if (!dca) return block;
1942 /* Find a placement that is dominates both, dca and block. */
1943 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1944 block = get_Block_idom(block);
1946 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1947 dca = get_Block_idom(dca);
1950 while (block != dca) {
1951 block = get_Block_idom(block); dca = get_Block_idom(dca);
1957 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1958 * I.e., DCA is the block where we might place PRODUCER.
1959 * A data flow edge points from producer to consumer.
1962 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer) {
1963 ir_node *block = NULL;
1965 /* Compute the latest block into which we can place a node so that it is
1967 if (get_irn_op(consumer) == op_Phi) {
1968 /* our consumer is a Phi-node, the effective use is in all those
1969 blocks through which the Phi-node reaches producer */
1971 ir_node *phi_block = get_nodes_block(consumer);
1972 irn_arity = get_irn_arity(consumer);
1974 for (i = 0; i < irn_arity; i++) {
1975 if (get_irn_n(consumer, i) == producer) {
1976 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1978 if (! is_Block_unreachable(new_block))
1979 block = calc_dca(block, new_block);
1984 block = get_irn_n(producer, -1);
1986 assert(is_no_Block(consumer));
1987 block = get_nodes_block(consumer);
1990 /* Compute the deepest common ancestor of block and dca. */
1991 return calc_dca(dca, block);
1994 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1996 static INLINE int get_irn_loop_depth(ir_node *n) {
1997 return get_loop_depth(get_irn_loop(n));
2001 * Move n to a block with less loop depth than it's current block. The
2002 * new block must be dominated by early.
2004 * @param n the node that should be moved
2005 * @param early the earliest block we can n move to
2007 static void move_out_of_loops(ir_node *n, ir_node *early) {
2008 ir_node *best, *dca;
2012 /* Find the region deepest in the dominator tree dominating
2013 dca with the least loop nesting depth, but still dominated
2014 by our early placement. */
2015 dca = get_nodes_block(n);
2018 while (dca != early) {
2019 dca = get_Block_idom(dca);
2020 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
2021 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
2025 if (best != get_nodes_block(n)) {
2027 printf("Moving out of loop: "); DDMN(n);
2028 printf(" Outermost block: "); DDMN(early);
2029 printf(" Best block: "); DDMN(best);
2030 printf(" Innermost block: "); DDMN(get_nodes_block(n));
2032 set_nodes_block(n, best);
2036 /* deepest common ancestor in the dominator tree of all nodes'
2037 blocks depending on us; our final placement has to dominate DCA. */
2038 static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
2042 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2043 ir_node *succ = get_irn_out(node, i);
2048 * This consumer is the End node, a keep alive edge.
2049 * This is not a real consumer, so we ignore it
2055 dca = get_deepest_common_ancestor(succ, dca);
2057 /* ignore if succ is in dead code */
2058 succ_blk = get_irn_n(succ, -1);
2059 if (is_Block_unreachable(succ_blk))
2061 dca = consumer_dom_dca(dca, succ, node);
2068 static void set_projs_block(ir_node *node, ir_node *block)
2072 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2073 ir_node *succ = get_irn_out(node, i);
2075 assert(is_Proj(succ));
2077 if(get_irn_mode(succ) == mode_T) {
2078 set_projs_block(succ, block);
2080 set_nodes_block(succ, block);
2085 * Find the latest legal block for N and place N into the
2086 * `optimal' Block between the latest and earliest legal block.
2087 * The `optimal' block is the dominance-deepest block of those
2088 * with the least loop-nesting-depth. This places N out of as many
2089 * loops as possible and then makes it as control dependent as
2092 * @param n the node to be placed
2093 * @param worklist a worklist, all successors of non-floating nodes are
2096 static void place_floats_late(ir_node *n, pdeq *worklist) {
2100 assert(irn_not_visited(n)); /* no multiple placement */
2102 mark_irn_visited(n);
2104 /* no need to place block nodes, control nodes are already placed. */
2105 if ((get_irn_op(n) != op_Block) &&
2107 (get_irn_mode(n) != mode_X)) {
2108 /* Remember the early_blk placement of this block to move it
2109 out of loop no further than the early_blk placement. */
2110 early_blk = get_irn_n(n, -1);
2113 * BEWARE: Here we also get code, that is live, but
2114 * was in a dead block. If the node is life, but because
2115 * of CSE in a dead block, we still might need it.
2118 /* Assure that our users are all placed, except the Phi-nodes.
2119 --- Each data flow cycle contains at least one Phi-node. We
2120 have to break the `user has to be placed before the
2121 producer' dependence cycle and the Phi-nodes are the
2122 place to do so, because we need to base our placement on the
2123 final region of our users, which is OK with Phi-nodes, as they
2124 are op_pin_state_pinned, and they never have to be placed after a
2125 producer of one of their inputs in the same block anyway. */
2126 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2127 ir_node *succ = get_irn_out(n, i);
2128 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
2129 place_floats_late(succ, worklist);
2132 if (! is_Block_dead(early_blk)) {
2133 /* do only move things that where not dead */
2134 ir_op *op = get_irn_op(n);
2136 /* We have to determine the final block of this node... except for
2137 constants and Projs */
2138 if ((get_irn_pinned(n) == op_pin_state_floats) &&
2140 (op != op_SymConst) &&
2143 /* deepest common ancestor in the dominator tree of all nodes'
2144 blocks depending on us; our final placement has to dominate
2146 ir_node *dca = get_deepest_common_ancestor(n, NULL);
2148 set_nodes_block(n, dca);
2149 move_out_of_loops(n, early_blk);
2150 if(get_irn_mode(n) == mode_T) {
2151 set_projs_block(n, get_nodes_block(n));
2158 /* Add successors of all non-floating nodes on list. (Those of floating
2159 nodes are placed already and therefore are marked.) */
2160 for (i = 0; i < get_irn_n_outs(n); i++) {
2161 ir_node *succ = get_irn_out(n, i);
2162 if (irn_not_visited(get_irn_out(n, i))) {
2163 pdeq_putr(worklist, succ);
2169 * Place floating nodes on the given worklist as late as possible using
2170 * the dominance tree.
2172 * @param worklist the worklist containing the nodes to place
2174 static void place_late(waitq *worklist) {
2176 inc_irg_visited(current_ir_graph);
2178 /* This fills the worklist initially. */
2179 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2181 /* And now empty the worklist again... */
2182 while (!waitq_empty(worklist)) {
2183 ir_node *n = waitq_get(worklist);
2184 if (irn_not_visited(n))
2185 place_floats_late(n, worklist);
2189 /* Code Placement. */
2190 void place_code(ir_graph *irg) {
2192 ir_graph *rem = current_ir_graph;
2194 current_ir_graph = irg;
2196 /* Handle graph state */
2197 assert(get_irg_phase_state(irg) != phase_building);
2200 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2201 free_loop_information(irg);
2202 construct_backedges(irg);
2205 /* Place all floating nodes as early as possible. This guarantees
2206 a legal code placement. */
2207 worklist = new_waitq();
2208 place_early(worklist);
2210 /* place_early() invalidates the outs, place_late needs them. */
2211 compute_irg_outs(irg);
2213 /* Now move the nodes down in the dominator tree. This reduces the
2214 unnecessary executions of the node. */
2215 place_late(worklist);
2217 set_irg_outs_inconsistent(current_ir_graph);
2218 set_irg_loopinfo_inconsistent(current_ir_graph);
2219 del_waitq(worklist);
2220 current_ir_graph = rem;
2224 * Called by walker of remove_critical_cf_edges().
2226 * Place an empty block to an edge between a blocks of multiple
2227 * predecessors and a block of multiple successors.
2230 * @param env Environment of walker. The changed field.
2232 static void walk_critical_cf_edges(ir_node *n, void *env) {
2234 ir_node *pre, *block, *jmp;
2236 ir_graph *irg = get_irn_irg(n);
2238 /* Block has multiple predecessors */
2239 arity = get_irn_arity(n);
2241 if (n == get_irg_end_block(irg))
2242 return; /* No use to add a block here. */
2244 for (i = 0; i < arity; ++i) {
2247 pre = get_irn_n(n, i);
2248 cfop = get_irn_op(skip_Proj(pre));
2249 /* Predecessor has multiple successors. Insert new control flow edge but
2250 ignore exception edges. */
2251 if (! is_op_fragile(cfop) && is_op_forking(cfop)) {
2252 /* set predecessor of new block */
2253 block = new_r_Block(irg, 1, &pre);
2254 /* insert new jmp node to new block */
2255 jmp = new_r_Jmp(irg, block);
2256 /* set successor of new block */
2257 set_irn_n(n, i, jmp);
2259 } /* predecessor has multiple successors */
2260 } /* for all predecessors */
2261 } /* n is a multi-entry block */
2264 void remove_critical_cf_edges(ir_graph *irg) {
2267 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &changed);
2269 /* control flow changed */
2270 set_irg_outs_inconsistent(irg);
2271 set_irg_extblk_inconsistent(irg);
2272 set_irg_doms_inconsistent(irg);
2273 set_irg_loopinfo_inconsistent(irg);