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 (is_Bad(get_irn_n(b, i))) 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);
317 /* The end node looses it's flexible in array. This doesn't matter,
318 as dead node elimination builds End by hand, inlineing doesn't use
320 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
323 /* node copied already */
325 } else if (op == op_Block) {
327 new_arity = compute_new_arity(n);
328 n->attr.block.graph_arr = NULL;
330 block = get_nodes_block(n);
332 new_arity = compute_new_arity(block);
334 new_arity = get_irn_arity(n);
337 nn = new_ir_node(get_irn_dbg_info(n),
344 /* Copy the attributes. These might point to additional data. If this
345 was allocated on the old obstack the pointers now are dangling. This
346 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
347 copy_node_attr(n, nn);
351 int copy_node_nr = env != NULL;
353 /* for easier debugging, we want to copy the node numbers too */
354 nn->node_nr = n->node_nr;
360 hook_dead_node_elim_subst(current_ir_graph, n, nn);
364 * Copies new predecessors of old node to new node remembered in link.
365 * Spare the Bad predecessors of Phi and Block nodes.
367 static void copy_preds(ir_node *n, void *env) {
372 nn = get_new_node(n);
375 /* copy the macro block header */
376 ir_node *mbh = get_Block_MacroBlock(n);
379 /* this block is a macroblock header */
380 set_irn_n(nn, -1, nn);
382 /* get the macro block header */
383 set_irn_n(nn, -1, get_new_node(mbh));
386 /* Don't copy Bad nodes. */
388 irn_arity = get_irn_arity(n);
389 for (i = 0; i < irn_arity; i++) {
390 if (! is_Bad(get_irn_n(n, i))) {
391 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
392 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
396 /* repair the block visited flag from above misuse. Repair it in both
397 graphs so that the old one can still be used. */
398 set_Block_block_visited(nn, 0);
399 set_Block_block_visited(n, 0);
400 /* Local optimization could not merge two subsequent blocks if
401 in array contained Bads. Now it's possible.
402 We don't call optimize_in_place as it requires
403 that the fields in ir_graph are set properly. */
404 if ((get_opt_control_flow_straightening()) &&
405 (get_Block_n_cfgpreds(nn) == 1) &&
406 is_Jmp(get_Block_cfgpred(nn, 0))) {
407 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
409 /* Jmp jumps into the block it is in -- deal self cycle. */
410 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
411 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
416 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
417 /* Don't copy node if corresponding predecessor in block is Bad.
418 The Block itself should not be Bad. */
419 block = get_nodes_block(n);
420 set_irn_n(nn, -1, get_new_node(block));
422 irn_arity = get_irn_arity(n);
423 for (i = 0; i < irn_arity; i++) {
424 if (! is_Bad(get_irn_n(block, i))) {
425 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
426 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
430 /* If the pre walker reached this Phi after the post walker visited the
431 block block_visited is > 0. */
432 set_Block_block_visited(get_nodes_block(n), 0);
433 /* Compacting the Phi's ins might generate Phis with only one
435 if (get_irn_arity(nn) == 1)
436 exchange(nn, get_irn_n(nn, 0));
438 irn_arity = get_irn_arity(n);
439 for (i = -1; i < irn_arity; i++)
440 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
442 /* Now the new node is complete. We can add it to the hash table for CSE.
443 @@@ inlining aborts if we identify End. Why? */
445 add_identities(current_ir_graph->value_table, nn);
449 * Copies the graph recursively, compacts the keep-alives of the end node.
451 * @param irg the graph to be copied
452 * @param copy_node_nr If non-zero, the node number will be copied
454 static void copy_graph(ir_graph *irg, int copy_node_nr) {
455 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
456 ir_node *ka; /* keep alive */
460 /* Some nodes must be copied by hand, sigh */
461 vfl = get_irg_visited(irg);
462 set_irg_visited(irg, vfl + 1);
464 oe = get_irg_end(irg);
465 mark_irn_visited(oe);
466 /* copy the end node by hand, allocate dynamic in array! */
467 ne = new_ir_node(get_irn_dbg_info(oe),
474 /* Copy the attributes. Well, there might be some in the future... */
475 copy_node_attr(oe, ne);
476 set_new_node(oe, ne);
478 /* copy the Bad node */
479 ob = get_irg_bad(irg);
480 mark_irn_visited(ob);
481 nb = new_ir_node(get_irn_dbg_info(ob),
488 copy_node_attr(ob, nb);
489 set_new_node(ob, nb);
491 /* copy the NoMem node */
492 om = get_irg_no_mem(irg);
493 mark_irn_visited(om);
494 nm = new_ir_node(get_irn_dbg_info(om),
501 copy_node_attr(om, nm);
502 set_new_node(om, nm);
504 /* copy the live nodes */
505 set_irg_visited(irg, vfl);
506 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
508 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
510 /* visit the anchors as well */
511 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
512 ir_node *n = get_irg_anchor(irg, i);
514 if (n && (get_irn_visited(n) <= vfl)) {
515 set_irg_visited(irg, vfl);
516 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
520 /* copy_preds for the end node ... */
521 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
523 /*- ... and now the keep alives. -*/
524 /* First pick the not marked block nodes and walk them. We must pick these
525 first as else we will oversee blocks reachable from Phis. */
526 irn_arity = get_End_n_keepalives(oe);
527 for (i = 0; i < irn_arity; i++) {
528 ka = get_End_keepalive(oe, i);
530 if (get_irn_visited(ka) <= vfl) {
531 /* We must keep the block alive and copy everything reachable */
532 set_irg_visited(irg, vfl);
533 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
535 add_End_keepalive(ne, get_new_node(ka));
539 /* Now pick other nodes. Here we will keep all! */
540 irn_arity = get_End_n_keepalives(oe);
541 for (i = 0; i < irn_arity; i++) {
542 ka = get_End_keepalive(oe, i);
544 if (get_irn_visited(ka) <= vfl) {
545 /* We didn't copy the node yet. */
546 set_irg_visited(irg, vfl);
547 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
549 add_End_keepalive(ne, get_new_node(ka));
553 /* start block sometimes only reached after keep alives */
554 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
555 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
559 * Copies the graph reachable from current_ir_graph->end to the obstack
560 * in current_ir_graph and fixes the environment.
561 * Then fixes the fields in current_ir_graph containing nodes of the
564 * @param copy_node_nr If non-zero, the node number will be copied
567 copy_graph_env(int copy_node_nr) {
568 ir_graph *irg = current_ir_graph;
569 ir_node *old_end, *new_anchor;
572 /* remove end_except and end_reg nodes */
573 old_end = get_irg_end(irg);
574 set_irg_end_except (irg, old_end);
575 set_irg_end_reg (irg, old_end);
577 /* Not all nodes remembered in irg might be reachable
578 from the end node. Assure their link is set to NULL, so that
579 we can test whether new nodes have been computed. */
580 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
581 ir_node *n = get_irg_anchor(irg, i);
583 set_new_node(n, NULL);
585 /* we use the block walk flag for removing Bads from Blocks ins. */
586 inc_irg_block_visited(irg);
589 copy_graph(irg, copy_node_nr);
592 old_end = get_irg_end(irg);
593 new_anchor = new_Anchor(irg);
595 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
596 ir_node *n = get_irg_anchor(irg, i);
598 set_irn_n(new_anchor, i, get_new_node(n));
601 irg->anchor = new_anchor;
603 /* ensure the new anchor is placed in the endblock */
604 set_irn_n(new_anchor, -1, get_irg_end_block(irg));
608 * Copies all reachable nodes to a new obstack. Removes bad inputs
609 * from block nodes and the corresponding inputs from Phi nodes.
610 * Merges single exit blocks with single entry blocks and removes
612 * Adds all new nodes to a new hash table for CSE. Does not
613 * perform CSE, so the hash table might contain common subexpressions.
615 void dead_node_elimination(ir_graph *irg) {
617 #ifdef INTERPROCEDURAL_VIEW
618 int rem_ipview = get_interprocedural_view();
620 struct obstack *graveyard_obst = NULL;
621 struct obstack *rebirth_obst = NULL;
622 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
624 /* inform statistics that we started a dead-node elimination run */
625 hook_dead_node_elim(irg, 1);
627 /* Remember external state of current_ir_graph. */
628 rem = current_ir_graph;
629 current_ir_graph = irg;
630 #ifdef INTERPROCEDURAL_VIEW
631 set_interprocedural_view(0);
634 assert(get_irg_phase_state(irg) != phase_building);
636 /* Handle graph state */
637 free_callee_info(irg);
641 /* @@@ so far we loose loops when copying */
642 free_loop_information(irg);
644 set_irg_doms_inconsistent(irg);
646 /* A quiet place, where the old obstack can rest in peace,
647 until it will be cremated. */
648 graveyard_obst = irg->obst;
650 /* A new obstack, where the reachable nodes will be copied to. */
651 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
652 irg->obst = rebirth_obst;
653 obstack_init(irg->obst);
654 irg->last_node_idx = 0;
656 /* We also need a new value table for CSE */
657 del_identities(irg->value_table);
658 irg->value_table = new_identities();
660 /* Copy the graph from the old to the new obstack */
661 copy_graph_env(/*copy_node_nr=*/1);
663 /* Free memory from old unoptimized obstack */
664 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
665 xfree(graveyard_obst); /* ... then free it. */
667 /* inform statistics that the run is over */
668 hook_dead_node_elim(irg, 0);
670 current_ir_graph = rem;
671 #ifdef INTERPROCEDURAL_VIEW
672 set_interprocedural_view(rem_ipview);
677 * Relink bad predecessors of a block and store the old in array to the
678 * link field. This function is called by relink_bad_predecessors().
679 * The array of link field starts with the block operand at position 0.
680 * If block has bad predecessors, create a new in array without bad preds.
681 * Otherwise let in array untouched.
683 static void relink_bad_block_predecessors(ir_node *n, void *env) {
684 ir_node **new_in, *irn;
685 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
688 /* if link field of block is NULL, look for bad predecessors otherwise
689 this is already done */
690 if (is_Block(n) && get_irn_link(n) == NULL) {
691 /* save old predecessors in link field (position 0 is the block operand)*/
692 set_irn_link(n, get_irn_in(n));
694 /* count predecessors without bad nodes */
695 old_irn_arity = get_irn_arity(n);
696 for (i = 0; i < old_irn_arity; i++)
697 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
699 /* arity changing: set new predecessors without bad nodes */
700 if (new_irn_arity < old_irn_arity) {
701 /* Get new predecessor array. We do not resize the array, as we must
702 keep the old one to update Phis. */
703 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
705 /* set new predecessors in array */
708 for (i = 0; i < old_irn_arity; i++) {
709 irn = get_irn_n(n, i);
711 new_in[new_irn_n] = irn;
712 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
716 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
717 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
719 } /* ir node has bad predecessors */
720 } /* Block is not relinked */
724 * Relinks Bad predecessors from Blocks and Phis called by walker
725 * remove_bad_predecesors(). If n is a Block, call
726 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
727 * function of Phi's Block. If this block has bad predecessors, relink preds
730 static void relink_bad_predecessors(ir_node *n, void *env) {
731 ir_node *block, **old_in;
732 int i, old_irn_arity, new_irn_arity;
734 /* relink bad predecessors of a block */
736 relink_bad_block_predecessors(n, env);
738 /* If Phi node relink its block and its predecessors */
740 /* Relink predecessors of phi's block */
741 block = get_nodes_block(n);
742 if (get_irn_link(block) == NULL)
743 relink_bad_block_predecessors(block, env);
745 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
746 old_irn_arity = ARR_LEN(old_in);
748 /* Relink Phi predecessors if count of predecessors changed */
749 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
750 /* set new predecessors in array
751 n->in[0] remains the same block */
753 for(i = 1; i < old_irn_arity; i++)
754 if (!is_Bad((ir_node *)old_in[i])) {
755 n->in[new_irn_arity] = n->in[i];
756 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
760 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
761 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
763 } /* n is a Phi node */
767 * Removes Bad Bad predecessors from Blocks and the corresponding
768 * inputs to Phi nodes as in dead_node_elimination but without
770 * On walking up set the link field to NULL, on walking down call
771 * relink_bad_predecessors() (This function stores the old in array
772 * to the link field and sets a new in array if arity of predecessors
775 void remove_bad_predecessors(ir_graph *irg) {
776 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
783 __)|_| | \_/ | \_/(/_ |_/\__|__
785 The following stuff implements a facility that automatically patches
786 registered ir_node pointers to the new node when a dead node elimination occurs.
789 struct _survive_dce_t {
793 hook_entry_t dead_node_elim;
794 hook_entry_t dead_node_elim_subst;
797 typedef struct _survive_dce_list_t {
798 struct _survive_dce_list_t *next;
800 } survive_dce_list_t;
802 static void dead_node_hook(void *context, ir_graph *irg, int start) {
803 survive_dce_t *sd = context;
806 /* Create a new map before the dead node elimination is performed. */
808 sd->new_places = pmap_create_ex(pmap_count(sd->places));
810 /* Patch back all nodes if dead node elimination is over and something is to be done. */
811 pmap_destroy(sd->places);
812 sd->places = sd->new_places;
813 sd->new_places = NULL;
818 * Hook called when dead node elimination replaces old by nw.
820 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
821 survive_dce_t *sd = context;
822 survive_dce_list_t *list = pmap_get(sd->places, old);
825 /* If the node is to be patched back, write the new address to all registered locations. */
827 survive_dce_list_t *p;
829 for (p = list; p; p = p->next)
832 pmap_insert(sd->new_places, nw, list);
837 * Make a new Survive DCE environment.
839 survive_dce_t *new_survive_dce(void) {
840 survive_dce_t *res = xmalloc(sizeof(res[0]));
841 obstack_init(&res->obst);
842 res->places = pmap_create();
843 res->new_places = NULL;
845 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
846 res->dead_node_elim.context = res;
847 res->dead_node_elim.next = NULL;
849 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
850 res->dead_node_elim_subst.context = res;
851 res->dead_node_elim_subst.next = NULL;
853 #ifndef FIRM_ENABLE_HOOKS
854 assert(0 && "need hooks enabled");
857 register_hook(hook_dead_node_elim, &res->dead_node_elim);
858 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
863 * Free a Survive DCE environment.
865 void free_survive_dce(survive_dce_t *sd) {
866 obstack_free(&sd->obst, NULL);
867 pmap_destroy(sd->places);
868 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
869 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
874 * Register a node pointer to be patched upon DCE.
875 * When DCE occurs, the node pointer specified by @p place will be
876 * patched to the new address of the node it is pointing to.
878 * @param sd The Survive DCE environment.
879 * @param place The address of the node pointer.
881 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
882 if (*place != NULL) {
883 ir_node *irn = *place;
884 survive_dce_list_t *curr = pmap_get(sd->places, irn);
885 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
890 pmap_insert(sd->places, irn, nw);
894 /*--------------------------------------------------------------------*/
895 /* Functionality for inlining */
896 /*--------------------------------------------------------------------*/
899 * Copy node for inlineing. Updates attributes that change when
900 * inlineing but not for dead node elimination.
902 * Copies the node by calling copy_node() and then updates the entity if
903 * it's a local one. env must be a pointer of the frame type of the
904 * inlined procedure. The new entities must be in the link field of
908 copy_node_inline(ir_node *n, void *env) {
910 ir_type *frame_tp = (ir_type *)env;
914 nn = get_new_node (n);
916 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
917 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
919 } else if (is_Block(n)) {
920 nn = get_new_node (n);
921 nn->attr.block.irg = current_ir_graph;
926 * Walker: checks if P_value_arg_base is used.
928 static void find_addr(ir_node *node, void *env) {
929 int *allow_inline = env;
931 is_Start(get_Proj_pred(node)) &&
932 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
938 * Check if we can inline a given call.
939 * Currently, we cannot inline two cases:
940 * - call with compound arguments
941 * - graphs that take the address of a parameter
943 * check these conditions here
945 static int can_inline(ir_node *call, ir_graph *called_graph) {
946 ir_type *call_type = get_Call_type(call);
947 int params, ress, i, res;
948 assert(is_Method_type(call_type));
950 params = get_method_n_params(call_type);
951 ress = get_method_n_ress(call_type);
953 /* check parameters for compound arguments */
954 for (i = 0; i < params; ++i) {
955 ir_type *p_type = get_method_param_type(call_type, i);
957 if (is_compound_type(p_type))
961 /* check results for compound arguments */
962 for (i = 0; i < ress; ++i) {
963 ir_type *r_type = get_method_res_type(call_type, i);
965 if (is_compound_type(r_type))
970 irg_walk_graph(called_graph, find_addr, NULL, &res);
976 exc_handler = 0, /**< There is a handler. */
977 exc_to_end = 1, /**< Branches to End. */
978 exc_no_handler = 2 /**< Exception handling not represented. */
981 /* Inlines a method at the given call site. */
982 int inline_method(ir_node *call, ir_graph *called_graph) {
984 ir_node *post_call, *post_bl;
985 ir_node *in[pn_Start_max];
986 ir_node *end, *end_bl;
990 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
991 enum exc_mode exc_handling;
992 ir_type *called_frame;
993 irg_inline_property prop = get_irg_inline_property(called_graph);
995 if ( (prop < irg_inline_forced) || (prop == irg_inline_forbidden))
998 /* Do not inline variadic functions. */
999 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
1002 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
1003 get_method_n_params(get_Call_type(call)));
1006 * currently, we cannot inline two cases:
1007 * - call with compound arguments
1008 * - graphs that take the address of a parameter
1010 if (! can_inline(call, called_graph))
1013 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
1014 rem_opt = get_opt_optimize();
1017 /* Handle graph state */
1018 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1019 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
1020 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
1021 set_irg_outs_inconsistent(current_ir_graph);
1022 set_irg_extblk_inconsistent(current_ir_graph);
1023 set_irg_doms_inconsistent(current_ir_graph);
1024 set_irg_loopinfo_inconsistent(current_ir_graph);
1025 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
1027 /* -- Check preconditions -- */
1028 assert(is_Call(call));
1029 /* @@@ does not work for InterfaceIII.java after cgana
1030 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
1031 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
1032 get_Call_type(call)));
1034 if (called_graph == current_ir_graph) {
1035 set_optimize(rem_opt);
1039 /* here we know we WILL inline, so inform the statistics */
1040 hook_inline(call, called_graph);
1042 /* -- Decide how to handle exception control flow: Is there a handler
1043 for the Call node, or do we branch directly to End on an exception?
1045 0 There is a handler.
1047 2 Exception handling not represented in Firm. -- */
1049 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
1050 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
1051 long proj_nr = get_Proj_proj(proj);
1052 if (proj_nr == pn_Call_X_except) Xproj = proj;
1053 if (proj_nr == pn_Call_M_except) Mproj = proj;
1055 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
1056 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
1057 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
1061 the procedure and later replaces the Start node of the called graph.
1062 Post_call is the old Call node and collects the results of the called
1063 graph. Both will end up being a tuple. -- */
1064 post_bl = get_nodes_block(call);
1065 set_irg_current_block(current_ir_graph, post_bl);
1066 /* XxMxPxPxPxT of Start + parameter of Call */
1067 in[pn_Start_X_initial_exec] = new_Jmp();
1068 in[pn_Start_M] = get_Call_mem(call);
1069 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1070 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1071 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1072 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1073 /* in[pn_Start_P_value_arg_base] = ??? */
1074 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1075 pre_call = new_Tuple(pn_Start_max - 1, in);
1079 The new block gets the ins of the old block, pre_call and all its
1080 predecessors and all Phi nodes. -- */
1081 part_block(pre_call);
1083 /* -- Prepare state for dead node elimination -- */
1084 /* Visited flags in calling irg must be >= flag in called irg.
1085 Else walker and arity computation will not work. */
1086 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1087 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1088 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1089 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1090 /* Set pre_call as new Start node in link field of the start node of
1091 calling graph and pre_calls block as new block for the start block
1093 Further mark these nodes so that they are not visited by the
1095 set_irn_link(get_irg_start(called_graph), pre_call);
1096 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1097 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1098 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1099 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1100 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1102 /* Initialize for compaction of in arrays */
1103 inc_irg_block_visited(current_ir_graph);
1105 /* -- Replicate local entities of the called_graph -- */
1106 /* copy the entities. */
1107 called_frame = get_irg_frame_type(called_graph);
1108 for (i = 0; i < get_class_n_members(called_frame); i++) {
1109 ir_entity *new_ent, *old_ent;
1110 old_ent = get_class_member(called_frame, i);
1111 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1112 set_entity_link(old_ent, new_ent);
1115 /* visited is > than that of called graph. With this trick visited will
1116 remain unchanged so that an outer walker, e.g., searching the call nodes
1117 to inline, calling this inline will not visit the inlined nodes. */
1118 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1120 /* -- Performing dead node elimination inlines the graph -- */
1121 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1123 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1124 get_irg_frame_type(called_graph));
1126 /* Repair called_graph */
1127 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1128 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1129 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1131 /* -- Merge the end of the inlined procedure with the call site -- */
1132 /* We will turn the old Call node into a Tuple with the following
1135 0: Phi of all Memories of Return statements.
1136 1: Jmp from new Block that merges the control flow from all exception
1137 predecessors of the old end block.
1138 2: Tuple of all arguments.
1139 3: Phi of Exception memories.
1140 In case the old Call directly branches to End on an exception we don't
1141 need the block merging all exceptions nor the Phi of the exception
1145 /* -- Precompute some values -- */
1146 end_bl = get_new_node(get_irg_end_block(called_graph));
1147 end = get_new_node(get_irg_end(called_graph));
1148 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1149 n_res = get_method_n_ress(get_Call_type(call));
1151 res_pred = xmalloc(n_res * sizeof(*res_pred));
1152 cf_pred = xmalloc(arity * sizeof(*res_pred));
1154 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1156 /* -- archive keepalives -- */
1157 irn_arity = get_irn_arity(end);
1158 for (i = 0; i < irn_arity; i++) {
1159 ir_node *ka = get_End_keepalive(end, i);
1161 add_End_keepalive(get_irg_end(current_ir_graph), ka);
1164 /* The new end node will die. We need not free as the in array is on the obstack:
1165 copy_node() only generated 'D' arrays. */
1167 /* -- Replace Return nodes by Jump nodes. -- */
1169 for (i = 0; i < arity; i++) {
1171 ret = get_irn_n(end_bl, i);
1172 if (is_Return(ret)) {
1173 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1177 set_irn_in(post_bl, n_ret, cf_pred);
1179 /* -- Build a Tuple for all results of the method.
1180 Add Phi node if there was more than one Return. -- */
1181 turn_into_tuple(post_call, pn_Call_max);
1182 /* First the Memory-Phi */
1184 for (i = 0; i < arity; i++) {
1185 ret = get_irn_n(end_bl, i);
1186 if (is_Return(ret)) {
1187 cf_pred[n_ret] = get_Return_mem(ret);
1191 phi = new_Phi(n_ret, cf_pred, mode_M);
1192 set_Tuple_pred(call, pn_Call_M_regular, phi);
1193 /* Conserve Phi-list for further inlinings -- but might be optimized */
1194 if (get_nodes_block(phi) == post_bl) {
1195 set_irn_link(phi, get_irn_link(post_bl));
1196 set_irn_link(post_bl, phi);
1198 /* Now the real results */
1200 for (j = 0; j < n_res; j++) {
1202 for (i = 0; i < arity; i++) {
1203 ret = get_irn_n(end_bl, i);
1204 if (is_Return(ret)) {
1205 cf_pred[n_ret] = get_Return_res(ret, j);
1210 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1214 /* Conserve Phi-list for further inlinings -- but might be optimized */
1215 if (get_nodes_block(phi) == post_bl) {
1216 set_irn_link(phi, get_irn_link(post_bl));
1217 set_irn_link(post_bl, phi);
1220 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1222 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1224 /* handle the regular call */
1225 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1227 /* For now, we cannot inline calls with value_base */
1228 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1230 /* Finally the exception control flow.
1231 We have two (three) possible situations:
1232 First if the Call branches to an exception handler: We need to add a Phi node to
1233 collect the memory containing the exception objects. Further we need
1234 to add another block to get a correct representation of this Phi. To
1235 this block we add a Jmp that resolves into the X output of the Call
1236 when the Call is turned into a tuple.
1237 Second the Call branches to End, the exception is not handled. Just
1238 add all inlined exception branches to the End node.
1239 Third: there is no Exception edge at all. Handle as case two. */
1240 if (exc_handling == exc_handler) {
1242 for (i = 0; i < arity; i++) {
1244 ret = get_irn_n(end_bl, i);
1245 irn = skip_Proj(ret);
1246 if (is_fragile_op(irn) || is_Raise(irn)) {
1247 cf_pred[n_exc] = ret;
1252 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1253 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1254 /* The Phi for the memories with the exception objects */
1256 for (i = 0; i < arity; i++) {
1258 ret = skip_Proj(get_irn_n(end_bl, i));
1260 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1262 } else if (is_fragile_op(ret)) {
1263 /* We rely that all cfops have the memory output at the same position. */
1264 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1266 } else if (is_Raise(ret)) {
1267 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1271 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1273 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1274 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1277 ir_node *main_end_bl;
1278 int main_end_bl_arity;
1279 ir_node **end_preds;
1281 /* assert(exc_handling == 1 || no exceptions. ) */
1283 for (i = 0; i < arity; i++) {
1284 ir_node *ret = get_irn_n(end_bl, i);
1285 ir_node *irn = skip_Proj(ret);
1287 if (is_fragile_op(irn) || is_Raise(irn)) {
1288 cf_pred[n_exc] = ret;
1292 main_end_bl = get_irg_end_block(current_ir_graph);
1293 main_end_bl_arity = get_irn_arity(main_end_bl);
1294 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1296 for (i = 0; i < main_end_bl_arity; ++i)
1297 end_preds[i] = get_irn_n(main_end_bl, i);
1298 for (i = 0; i < n_exc; ++i)
1299 end_preds[main_end_bl_arity + i] = cf_pred[i];
1300 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1301 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1302 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1308 /* -- Turn CSE back on. -- */
1309 set_optimize(rem_opt);
1314 /********************************************************************/
1315 /* Apply inlineing to small methods. */
1316 /********************************************************************/
1318 /** Represents a possible inlinable call in a graph. */
1319 typedef struct _call_entry call_entry;
1320 struct _call_entry {
1321 ir_node *call; /**< the Call */
1322 ir_graph *callee; /**< the callee called here */
1323 call_entry *next; /**< for linking the next one */
1327 * environment for inlining small irgs
1329 typedef struct _inline_env_t {
1330 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1331 call_entry *head; /**< the head of the call entry list */
1332 call_entry *tail; /**< the tail of the call entry list */
1336 * Returns the irg called from a Call node. If the irg is not
1337 * known, NULL is returned.
1339 static ir_graph *get_call_called_irg(ir_node *call) {
1341 ir_graph *called_irg = NULL;
1343 addr = get_Call_ptr(call);
1344 if (is_SymConst(addr) && get_SymConst_kind(addr) == symconst_addr_ent) {
1345 called_irg = get_entity_irg(get_SymConst_entity(addr));
1352 * Walker: Collect all calls to known graphs inside a graph.
1354 static void collect_calls(ir_node *call, void *env) {
1355 if (is_Call(call)) {
1356 ir_graph *called_irg = get_call_called_irg(call);
1358 /* The Call node calls a locally defined method. Remember to inline. */
1359 inline_env_t *ienv = env;
1360 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1362 entry->callee = called_irg;
1365 if (ienv->tail == NULL)
1368 ienv->tail->next = entry;
1375 * Inlines all small methods at call sites where the called address comes
1376 * from a Const node that references the entity representing the called
1378 * The size argument is a rough measure for the code size of the method:
1379 * Methods where the obstack containing the firm graph is smaller than
1382 void inline_small_irgs(ir_graph *irg, int size) {
1383 ir_graph *rem = current_ir_graph;
1386 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1388 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1390 current_ir_graph = irg;
1391 /* Handle graph state */
1392 assert(get_irg_phase_state(irg) != phase_building);
1393 free_callee_info(irg);
1395 /* Find Call nodes to inline.
1396 (We can not inline during a walk of the graph, as inlineing the same
1397 method several times changes the visited flag of the walked graph:
1398 after the first inlineing visited of the callee equals visited of
1399 the caller. With the next inlineing both are increased.) */
1400 obstack_init(&env.obst);
1401 env.head = env.tail = NULL;
1402 irg_walk_graph(irg, NULL, collect_calls, &env);
1404 if (env.head != NULL) {
1405 /* There are calls to inline */
1406 collect_phiprojs(irg);
1407 for (entry = env.head; entry != NULL; entry = entry->next) {
1408 ir_graph *callee = entry->callee;
1409 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1410 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1411 inline_method(entry->call, callee);
1415 obstack_free(&env.obst, NULL);
1416 current_ir_graph = rem;
1420 * Environment for inlining irgs.
1423 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1424 int n_nodes_orig; /**< for statistics */
1425 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1426 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1427 int n_call_nodes; /**< Number of Call nodes in the graph. */
1428 int n_call_nodes_orig; /**< for statistics */
1429 int n_callers; /**< Number of known graphs that call this graphs. */
1430 int n_callers_orig; /**< for statistics */
1431 int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
1435 * Allocate a new environment for inlining.
1437 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1438 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1439 env->n_nodes = -2; /* do not count count Start, End */
1440 env->n_nodes_orig = -2; /* do not count Start, End */
1441 env->call_head = NULL;
1442 env->call_tail = NULL;
1443 env->n_call_nodes = 0;
1444 env->n_call_nodes_orig = 0;
1446 env->n_callers_orig = 0;
1447 env->got_inline = 0;
1451 typedef struct walker_env {
1452 struct obstack *obst; /**< the obstack for allocations. */
1453 inline_irg_env *x; /**< the inline environment */
1454 int ignore_runtime; /**< the ignore runtime flag */
1458 * post-walker: collect all calls in the inline-environment
1459 * of a graph and sum some statistics.
1461 static void collect_calls2(ir_node *call, void *ctx) {
1463 inline_irg_env *x = env->x;
1464 ir_op *op = get_irn_op(call);
1468 /* count meaningful nodes in irg */
1469 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1474 if (op != op_Call) return;
1476 /* check, if it's a runtime call */
1477 if (env->ignore_runtime) {
1478 ir_node *symc = get_Call_ptr(call);
1480 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1481 ir_entity *ent = get_SymConst_entity(symc);
1483 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1488 /* collect all call nodes */
1490 ++x->n_call_nodes_orig;
1492 callee = get_call_called_irg(call);
1494 inline_irg_env *callee_env = get_irg_link(callee);
1495 /* count all static callers */
1496 ++callee_env->n_callers;
1497 ++callee_env->n_callers_orig;
1499 /* link it in the list of possible inlinable entries */
1500 entry = obstack_alloc(env->obst, sizeof(*entry));
1502 entry->callee = callee;
1504 if (x->call_tail == NULL)
1505 x->call_head = entry;
1507 x->call_tail->next = entry;
1508 x->call_tail = entry;
1513 * Returns TRUE if the number of callers is 0 in the irg's environment,
1514 * hence this irg is a leave.
1516 INLINE static int is_leave(ir_graph *irg) {
1517 inline_irg_env *env = get_irg_link(irg);
1518 return env->n_call_nodes == 0;
1522 * Returns TRUE if the number of nodes in the callee is
1523 * smaller then size in the irg's environment.
1525 INLINE static int is_smaller(ir_graph *callee, int size) {
1526 inline_irg_env *env = get_irg_link(callee);
1527 return env->n_nodes < size;
1531 * Append the nodes of the list src to the nodes of the list in environment dst.
1533 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1534 call_entry *entry, *nentry;
1536 /* Note that the src list points to Call nodes in the inlined graph, but
1537 we need Call nodes in our graph. Luckily the inliner leaves this information
1538 in the link field. */
1539 for (entry = src; entry != NULL; entry = entry->next) {
1540 nentry = obstack_alloc(obst, sizeof(*nentry));
1541 nentry->call = get_irn_link(entry->call);
1542 nentry->callee = entry->callee;
1543 nentry->next = NULL;
1544 dst->call_tail->next = nentry;
1545 dst->call_tail = nentry;
1550 * Inlines small leave methods at call sites where the called address comes
1551 * from a Const node that references the entity representing the called
1553 * The size argument is a rough measure for the code size of the method:
1554 * Methods where the obstack containing the firm graph is smaller than
1557 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1558 inline_irg_env *env;
1564 call_entry *entry, *tail;
1565 const call_entry *centry;
1566 struct obstack obst;
1567 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1569 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1570 rem = current_ir_graph;
1571 obstack_init(&obst);
1573 /* extend all irgs by a temporary data structure for inlining. */
1574 n_irgs = get_irp_n_irgs();
1575 for (i = 0; i < n_irgs; ++i)
1576 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1578 /* Precompute information in temporary data structure. */
1580 wenv.ignore_runtime = ignore_runtime;
1581 for (i = 0; i < n_irgs; ++i) {
1582 ir_graph *irg = get_irp_irg(i);
1584 assert(get_irg_phase_state(irg) != phase_building);
1585 free_callee_info(irg);
1587 wenv.x = get_irg_link(irg);
1588 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1591 /* -- and now inline. -- */
1593 /* Inline leaves recursively -- we might construct new leaves. */
1597 for (i = 0; i < n_irgs; ++i) {
1599 int phiproj_computed = 0;
1601 current_ir_graph = get_irp_irg(i);
1602 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1605 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1608 if (env->n_nodes > maxsize) break;
1611 callee = entry->callee;
1613 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1614 if (!phiproj_computed) {
1615 phiproj_computed = 1;
1616 collect_phiprojs(current_ir_graph);
1618 did_inline = inline_method(call, callee);
1621 /* Do some statistics */
1622 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1624 env->got_inline = 1;
1625 --env->n_call_nodes;
1626 env->n_nodes += callee_env->n_nodes;
1627 --callee_env->n_callers;
1629 /* remove this call from the list */
1631 tail->next = entry->next;
1633 env->call_head = entry->next;
1639 env->call_tail = tail;
1641 } while (did_inline);
1643 /* inline other small functions. */
1644 for (i = 0; i < n_irgs; ++i) {
1646 int phiproj_computed = 0;
1648 current_ir_graph = get_irp_irg(i);
1649 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1651 /* note that the list of possible calls is updated during the process */
1653 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1657 callee = entry->callee;
1659 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1660 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1661 if (!phiproj_computed) {
1662 phiproj_computed = 1;
1663 collect_phiprojs(current_ir_graph);
1665 if (inline_method(call, callee)) {
1666 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1668 /* callee was inline. Append it's call list. */
1669 env->got_inline = 1;
1670 --env->n_call_nodes;
1671 append_call_list(&obst, env, callee_env->call_head);
1672 env->n_call_nodes += callee_env->n_call_nodes;
1673 env->n_nodes += callee_env->n_nodes;
1674 --callee_env->n_callers;
1676 /* after we have inlined callee, all called methods inside callee
1677 are now called once more */
1678 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1679 inline_irg_env *penv = get_irg_link(centry->callee);
1683 /* remove this call from the list */
1685 tail->next = entry->next;
1687 env->call_head = entry->next;
1693 env->call_tail = tail;
1696 for (i = 0; i < n_irgs; ++i) {
1697 irg = get_irp_irg(i);
1698 env = (inline_irg_env *)get_irg_link(irg);
1700 if (env->got_inline) {
1701 /* this irg got calls inlined */
1702 set_irg_outs_inconsistent(irg);
1703 set_irg_doms_inconsistent(irg);
1705 optimize_graph_df(irg);
1708 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1709 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1710 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1711 env->n_callers_orig, env->n_callers,
1712 get_entity_name(get_irg_entity(irg))));
1716 obstack_free(&obst, NULL);
1717 current_ir_graph = rem;
1720 /*******************************************************************/
1721 /* Code Placement. Pins all floating nodes to a block where they */
1722 /* will be executed only if needed. */
1723 /*******************************************************************/
1726 * Returns non-zero, is a block is not reachable from Start.
1728 * @param block the block to test
1731 is_Block_unreachable(ir_node *block) {
1732 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1736 * Find the earliest correct block for node n. --- Place n into the
1737 * same Block as its dominance-deepest Input.
1739 * We have to avoid calls to get_nodes_block() here
1740 * because the graph is floating.
1742 * move_out_of_loops() expects that place_floats_early() have placed
1743 * all "living" nodes into a living block. That's why we must
1744 * move nodes in dead block with "live" successors into a valid
1746 * We move them just into the same block as it's successor (or
1747 * in case of a Phi into the effective use block). For Phi successors,
1748 * this may still be a dead block, but then there is no real use, as
1749 * the control flow will be dead later.
1751 * @param n the node to be placed
1752 * @param worklist a worklist, predecessors of non-floating nodes are placed here
1755 place_floats_early(ir_node *n, waitq *worklist) {
1758 /* we must not run into an infinite loop */
1759 assert(irn_not_visited(n));
1760 mark_irn_visited(n);
1762 /* Place floating nodes. */
1763 if (get_irn_pinned(n) == op_pin_state_floats) {
1764 ir_node *curr_block = get_nodes_block(n);
1765 int in_dead_block = is_Block_unreachable(curr_block);
1767 ir_node *b = NULL; /* The block to place this node in */
1769 assert(is_no_Block(n));
1771 if (is_irn_start_block_placed(n)) {
1772 /* These nodes will not be placed by the loop below. */
1773 b = get_irg_start_block(current_ir_graph);
1777 /* find the block for this node. */
1778 irn_arity = get_irn_arity(n);
1779 for (i = 0; i < irn_arity; i++) {
1780 ir_node *pred = get_irn_n(n, i);
1781 ir_node *pred_block;
1783 if ((irn_not_visited(pred))
1784 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1787 * If the current node is NOT in a dead block, but one of its
1788 * predecessors is, we must move the predecessor to a live block.
1789 * Such thing can happen, if global CSE chose a node from a dead block.
1790 * We move it simply to our block.
1791 * Note that neither Phi nor End nodes are floating, so we don't
1792 * need to handle them here.
1794 if (! in_dead_block) {
1795 if (get_irn_pinned(pred) == op_pin_state_floats &&
1796 is_Block_unreachable(get_nodes_block(pred)))
1797 set_nodes_block(pred, curr_block);
1799 place_floats_early(pred, worklist);
1803 * A node in the Bad block must stay in the bad block,
1804 * so don't compute a new block for it.
1809 /* Because all loops contain at least one op_pin_state_pinned node, now all
1810 our inputs are either op_pin_state_pinned or place_early() has already
1811 been finished on them. We do not have any unfinished inputs! */
1812 pred_block = get_nodes_block(pred);
1813 if ((!is_Block_dead(pred_block)) &&
1814 (get_Block_dom_depth(pred_block) > depth)) {
1816 depth = get_Block_dom_depth(pred_block);
1818 /* Avoid that the node is placed in the Start block */
1820 get_Block_dom_depth(get_nodes_block(n)) > 1 &&
1821 get_irg_phase_state(current_ir_graph) != phase_backend) {
1822 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1823 assert(b != get_irg_start_block(current_ir_graph));
1828 set_nodes_block(n, b);
1832 * Add predecessors of non floating nodes and non-floating predecessors
1833 * of floating nodes to worklist and fix their blocks if the are in dead block.
1835 irn_arity = get_irn_arity(n);
1839 * Simplest case: End node. Predecessors are keep-alives,
1840 * no need to move out of dead block.
1842 for (i = -1; i < irn_arity; ++i) {
1843 ir_node *pred = get_irn_n(n, i);
1844 if (irn_not_visited(pred))
1845 waitq_put(worklist, pred);
1847 } else if (is_Block(n)) {
1849 * Blocks: Predecessors are control flow, no need to move
1850 * them out of dead block.
1852 for (i = irn_arity - 1; i >= 0; --i) {
1853 ir_node *pred = get_irn_n(n, i);
1854 if (irn_not_visited(pred))
1855 waitq_put(worklist, pred);
1857 } else if (is_Phi(n)) {
1859 ir_node *curr_block = get_nodes_block(n);
1860 int in_dead_block = is_Block_unreachable(curr_block);
1863 * Phi nodes: move nodes from dead blocks into the effective use
1864 * of the Phi-input if the Phi is not in a bad block.
1866 pred = get_nodes_block(n);
1867 if (irn_not_visited(pred))
1868 waitq_put(worklist, pred);
1870 for (i = irn_arity - 1; i >= 0; --i) {
1871 ir_node *pred = get_irn_n(n, i);
1873 if (irn_not_visited(pred)) {
1874 if (! in_dead_block &&
1875 get_irn_pinned(pred) == op_pin_state_floats &&
1876 is_Block_unreachable(get_nodes_block(pred))) {
1877 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1879 waitq_put(worklist, pred);
1884 ir_node *curr_block = get_nodes_block(n);
1885 int in_dead_block = is_Block_unreachable(curr_block);
1888 * All other nodes: move nodes from dead blocks into the same block.
1890 pred = get_nodes_block(n);
1891 if (irn_not_visited(pred))
1892 waitq_put(worklist, pred);
1894 for (i = irn_arity - 1; i >= 0; --i) {
1895 ir_node *pred = get_irn_n(n, i);
1897 if (irn_not_visited(pred)) {
1898 if (! in_dead_block &&
1899 get_irn_pinned(pred) == op_pin_state_floats &&
1900 is_Block_unreachable(get_nodes_block(pred))) {
1901 set_nodes_block(pred, curr_block);
1903 waitq_put(worklist, pred);
1910 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1911 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1912 * places all floating nodes reachable from its argument through floating
1913 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1915 * @param worklist a worklist, used for the algorithm, empty on in/output
1917 static void place_early(waitq *worklist) {
1919 inc_irg_visited(current_ir_graph);
1921 /* this inits the worklist */
1922 place_floats_early(get_irg_end(current_ir_graph), worklist);
1924 /* Work the content of the worklist. */
1925 while (!waitq_empty(worklist)) {
1926 ir_node *n = waitq_get(worklist);
1927 if (irn_not_visited(n))
1928 place_floats_early(n, worklist);
1931 set_irg_outs_inconsistent(current_ir_graph);
1932 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1936 * Compute the deepest common ancestor of block and dca.
1938 static ir_node *calc_dca(ir_node *dca, ir_node *block) {
1941 /* we do not want to place nodes in dead blocks */
1942 if (is_Block_dead(block))
1945 /* We found a first legal placement. */
1946 if (!dca) return block;
1948 /* Find a placement that is dominates both, dca and block. */
1949 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1950 block = get_Block_idom(block);
1952 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1953 dca = get_Block_idom(dca);
1956 while (block != dca) {
1957 block = get_Block_idom(block); dca = get_Block_idom(dca);
1963 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1964 * I.e., DCA is the block where we might place PRODUCER.
1965 * A data flow edge points from producer to consumer.
1967 static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1969 /* Compute the last block into which we can place a node so that it is
1971 if (is_Phi(consumer)) {
1972 /* our consumer is a Phi-node, the effective use is in all those
1973 blocks through which the Phi-node reaches producer */
1974 ir_node *phi_block = get_nodes_block(consumer);
1975 int arity = get_irn_arity(consumer);
1978 for (i = 0; i < arity; i++) {
1979 if (get_Phi_pred(consumer, i) == producer) {
1980 ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
1982 if (!is_Block_unreachable(new_block))
1983 dca = calc_dca(dca, new_block);
1987 dca = calc_dca(dca, get_nodes_block(consumer));
1993 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1995 static INLINE int get_irn_loop_depth(ir_node *n) {
1996 return get_loop_depth(get_irn_loop(n));
2000 * Move n to a block with less loop depth than it's current block. The
2001 * new block must be dominated by early.
2003 * @param n the node that should be moved
2004 * @param early the earliest block we can n move to
2006 static void move_out_of_loops(ir_node *n, ir_node *early) {
2007 ir_node *best, *dca;
2011 /* Find the region deepest in the dominator tree dominating
2012 dca with the least loop nesting depth, but still dominated
2013 by our early placement. */
2014 dca = get_nodes_block(n);
2017 while (dca != early) {
2018 dca = get_Block_idom(dca);
2019 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
2020 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
2024 if (best != get_nodes_block(n)) {
2026 printf("Moving out of loop: "); DDMN(n);
2027 printf(" Outermost block: "); DDMN(early);
2028 printf(" Best block: "); DDMN(best);
2029 printf(" Innermost block: "); DDMN(get_nodes_block(n));
2031 set_nodes_block(n, best);
2035 /* deepest common ancestor in the dominator tree of all nodes'
2036 blocks depending on us; our final placement has to dominate DCA. */
2037 static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
2041 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2042 ir_node *succ = get_irn_out(node, i);
2046 * This consumer is the End node, a keep alive edge.
2047 * This is not a real consumer, so we ignore it
2052 if (is_Proj(succ)) {
2053 dca = get_deepest_common_ancestor(succ, dca);
2055 /* ignore if succ is in dead code */
2056 ir_node *succ_blk = get_nodes_block(succ);
2057 if (is_Block_unreachable(succ_blk))
2059 dca = consumer_dom_dca(dca, succ, node);
2066 static void set_projs_block(ir_node *node, ir_node *block)
2070 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2071 ir_node *succ = get_irn_out(node, i);
2073 assert(is_Proj(succ));
2075 if(get_irn_mode(succ) == mode_T) {
2076 set_projs_block(succ, block);
2078 set_nodes_block(succ, block);
2083 * Find the latest legal block for N and place N into the
2084 * `optimal' Block between the latest and earliest legal block.
2085 * The `optimal' block is the dominance-deepest block of those
2086 * with the least loop-nesting-depth. This places N out of as many
2087 * loops as possible and then makes it as control dependent as
2090 * @param n the node to be placed
2091 * @param worklist a worklist, all successors of non-floating nodes are
2094 static void place_floats_late(ir_node *n, pdeq *worklist) {
2098 assert(irn_not_visited(n)); /* no multiple placement */
2100 mark_irn_visited(n);
2102 /* no need to place block nodes, control nodes are already placed. */
2105 (get_irn_mode(n) != mode_X)) {
2106 /* Remember the early_blk placement of this block to move it
2107 out of loop no further than the early_blk placement. */
2108 early_blk = get_nodes_block(n);
2111 * BEWARE: Here we also get code, that is live, but
2112 * was in a dead block. If the node is life, but because
2113 * of CSE in a dead block, we still might need it.
2116 /* Assure that our users are all placed, except the Phi-nodes.
2117 --- Each data flow cycle contains at least one Phi-node. We
2118 have to break the `user has to be placed before the
2119 producer' dependence cycle and the Phi-nodes are the
2120 place to do so, because we need to base our placement on the
2121 final region of our users, which is OK with Phi-nodes, as they
2122 are op_pin_state_pinned, and they never have to be placed after a
2123 producer of one of their inputs in the same block anyway. */
2124 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2125 ir_node *succ = get_irn_out(n, i);
2126 if (irn_not_visited(succ) && !is_Phi(succ))
2127 place_floats_late(succ, worklist);
2130 if (! is_Block_dead(early_blk)) {
2131 /* do only move things that where not dead */
2132 ir_op *op = get_irn_op(n);
2134 /* We have to determine the final block of this node... except for
2135 constants and Projs */
2136 if ((get_irn_pinned(n) == op_pin_state_floats) &&
2138 (op != op_SymConst) &&
2141 /* deepest common ancestor in the dominator tree of all nodes'
2142 blocks depending on us; our final placement has to dominate
2144 ir_node *dca = get_deepest_common_ancestor(n, NULL);
2146 set_nodes_block(n, dca);
2147 move_out_of_loops(n, early_blk);
2148 if(get_irn_mode(n) == mode_T) {
2149 set_projs_block(n, get_nodes_block(n));
2156 /* Add successors of all non-floating nodes on list. (Those of floating
2157 nodes are placed already and therefore are marked.) */
2158 for (i = 0; i < get_irn_n_outs(n); i++) {
2159 ir_node *succ = get_irn_out(n, i);
2160 if (irn_not_visited(get_irn_out(n, i))) {
2161 pdeq_putr(worklist, succ);
2167 * Place floating nodes on the given worklist as late as possible using
2168 * the dominance tree.
2170 * @param worklist the worklist containing the nodes to place
2172 static void place_late(waitq *worklist) {
2174 inc_irg_visited(current_ir_graph);
2176 /* This fills the worklist initially. */
2177 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2179 /* And now empty the worklist again... */
2180 while (!waitq_empty(worklist)) {
2181 ir_node *n = waitq_get(worklist);
2182 if (irn_not_visited(n))
2183 place_floats_late(n, worklist);
2187 /* Code Placement. */
2188 void place_code(ir_graph *irg) {
2190 ir_graph *rem = current_ir_graph;
2192 current_ir_graph = irg;
2194 /* Handle graph state */
2195 assert(get_irg_phase_state(irg) != phase_building);
2198 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2199 free_loop_information(irg);
2200 construct_cf_backedges(irg);
2203 /* Place all floating nodes as early as possible. This guarantees
2204 a legal code placement. */
2205 worklist = new_waitq();
2206 place_early(worklist);
2208 /* place_early() invalidates the outs, place_late needs them. */
2209 compute_irg_outs(irg);
2211 /* Now move the nodes down in the dominator tree. This reduces the
2212 unnecessary executions of the node. */
2213 place_late(worklist);
2215 set_irg_outs_inconsistent(current_ir_graph);
2216 set_irg_loopinfo_inconsistent(current_ir_graph);
2217 del_waitq(worklist);
2218 current_ir_graph = rem;
2221 typedef struct cf_env {
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 cfop = get_irn_op(skip_Proj(pre));
2252 if (is_op_fragile(cfop)) {
2253 if (cfop != op_Raise)
2257 /* we don't want place nodes in the start block, so handle it like forking */
2258 if (is_op_forking(cfop) || cfop == op_Start) {
2259 /* Predecessor has multiple successors. Insert new control flow edge edges. */
2261 /* set predecessor of new block */
2262 block = new_r_Block(irg, 1, &pre);
2263 /* insert new jmp node to new block */
2264 jmp = new_r_Jmp(irg, block);
2265 /* set successor of new block */
2266 set_irn_n(n, i, jmp);
2268 } /* predecessor has multiple successors */
2269 } /* for all predecessors */
2270 } /* n is a multi-entry block */
2273 void remove_critical_cf_edges(ir_graph *irg) {
2278 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &env);
2280 /* control flow changed */
2281 set_irg_outs_inconsistent(irg);
2282 set_irg_extblk_inconsistent(irg);
2283 set_irg_doms_inconsistent(irg);
2284 set_irg_loopinfo_inconsistent(irg);