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);
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 /* this block is a macroblock header */
379 set_irn_n(nn, -1, nn);
381 /* get the macro block header */
382 set_irn_n(nn, -1, get_new_node(mbh));
385 /* Don't copy Bad nodes. */
387 irn_arity = get_irn_arity(n);
388 for (i = 0; i < irn_arity; i++) {
389 if (! is_Bad(get_irn_n(n, i))) {
390 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
391 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
395 /* repair the block visited flag from above misuse. Repair it in both
396 graphs so that the old one can still be used. */
397 set_Block_block_visited(nn, 0);
398 set_Block_block_visited(n, 0);
399 /* Local optimization could not merge two subsequent blocks if
400 in array contained Bads. Now it's possible.
401 We don't call optimize_in_place as it requires
402 that the fields in ir_graph are set properly. */
403 if ((get_opt_control_flow_straightening()) &&
404 (get_Block_n_cfgpreds(nn) == 1) &&
405 is_Jmp(get_Block_cfgpred(nn, 0))) {
406 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
408 /* Jmp jumps into the block it is in -- deal self cycle. */
409 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
410 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
415 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
416 /* Don't copy node if corresponding predecessor in block is Bad.
417 The Block itself should not be Bad. */
418 block = get_nodes_block(n);
419 set_irn_n(nn, -1, get_new_node(block));
421 irn_arity = get_irn_arity(n);
422 for (i = 0; i < irn_arity; i++) {
423 if (! is_Bad(get_irn_n(block, i))) {
424 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
425 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
429 /* If the pre walker reached this Phi after the post walker visited the
430 block block_visited is > 0. */
431 set_Block_block_visited(get_nodes_block(n), 0);
432 /* Compacting the Phi's ins might generate Phis with only one
434 if (get_irn_arity(nn) == 1)
435 exchange(nn, get_irn_n(nn, 0));
437 irn_arity = get_irn_arity(n);
438 for (i = -1; i < irn_arity; i++)
439 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
441 /* Now the new node is complete. We can add it to the hash table for CSE.
442 @@@ inlining aborts if we identify End. Why? */
444 add_identities(current_ir_graph->value_table, nn);
448 * Copies the graph recursively, compacts the keep-alives of the end node.
450 * @param irg the graph to be copied
451 * @param copy_node_nr If non-zero, the node number will be copied
453 static void copy_graph(ir_graph *irg, int copy_node_nr) {
454 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
455 ir_node *ka; /* keep alive */
459 /* Some nodes must be copied by hand, sigh */
460 vfl = get_irg_visited(irg);
461 set_irg_visited(irg, vfl + 1);
463 oe = get_irg_end(irg);
464 mark_irn_visited(oe);
465 /* copy the end node by hand, allocate dynamic in array! */
466 ne = new_ir_node(get_irn_dbg_info(oe),
473 /* Copy the attributes. Well, there might be some in the future... */
474 copy_node_attr(oe, ne);
475 set_new_node(oe, ne);
477 /* copy the Bad node */
478 ob = get_irg_bad(irg);
479 mark_irn_visited(ob);
480 nb = new_ir_node(get_irn_dbg_info(ob),
487 copy_node_attr(ob, nb);
488 set_new_node(ob, nb);
490 /* copy the NoMem node */
491 om = get_irg_no_mem(irg);
492 mark_irn_visited(om);
493 nm = new_ir_node(get_irn_dbg_info(om),
500 copy_node_attr(om, nm);
501 set_new_node(om, nm);
503 /* copy the live nodes */
504 set_irg_visited(irg, vfl);
505 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
507 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
509 /* visit the anchors as well */
510 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
511 ir_node *n = get_irg_anchor(irg, i);
513 if (n && (get_irn_visited(n) <= vfl)) {
514 set_irg_visited(irg, vfl);
515 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
519 /* copy_preds for the end node ... */
520 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
522 /*- ... and now the keep alives. -*/
523 /* First pick the not marked block nodes and walk them. We must pick these
524 first as else we will oversee blocks reachable from Phis. */
525 irn_arity = get_End_n_keepalives(oe);
526 for (i = 0; i < irn_arity; i++) {
527 ka = get_End_keepalive(oe, i);
529 if (get_irn_visited(ka) <= vfl) {
530 /* We must keep the block alive and copy everything reachable */
531 set_irg_visited(irg, vfl);
532 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
534 add_End_keepalive(ne, get_new_node(ka));
538 /* Now pick other nodes. Here we will keep all! */
539 irn_arity = get_End_n_keepalives(oe);
540 for (i = 0; i < irn_arity; i++) {
541 ka = get_End_keepalive(oe, i);
543 if (get_irn_visited(ka) <= vfl) {
544 /* We didn't copy the node yet. */
545 set_irg_visited(irg, vfl);
546 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
548 add_End_keepalive(ne, get_new_node(ka));
552 /* start block sometimes only reached after keep alives */
553 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
554 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
558 * Copies the graph reachable from current_ir_graph->end to the obstack
559 * in current_ir_graph and fixes the environment.
560 * Then fixes the fields in current_ir_graph containing nodes of the
563 * @param copy_node_nr If non-zero, the node number will be copied
566 copy_graph_env(int copy_node_nr) {
567 ir_graph *irg = current_ir_graph;
568 ir_node *old_end, *new_anchor;
571 /* remove end_except and end_reg nodes */
572 old_end = get_irg_end(irg);
573 set_irg_end_except (irg, old_end);
574 set_irg_end_reg (irg, old_end);
576 /* Not all nodes remembered in irg might be reachable
577 from the end node. Assure their link is set to NULL, so that
578 we can test whether new nodes have been computed. */
579 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
580 ir_node *n = get_irg_anchor(irg, i);
582 set_new_node(n, NULL);
584 /* we use the block walk flag for removing Bads from Blocks ins. */
585 inc_irg_block_visited(irg);
588 copy_graph(irg, copy_node_nr);
591 old_end = get_irg_end(irg);
592 new_anchor = new_Anchor(irg);
594 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
595 ir_node *n = get_irg_anchor(irg, i);
597 set_irn_n(new_anchor, i, get_new_node(n));
600 irg->anchor = new_anchor;
602 /* ensure the new anchor is placed in the endblock */
603 set_irn_n(new_anchor, -1, get_irg_end_block(irg));
607 * Copies all reachable nodes to a new obstack. Removes bad inputs
608 * from block nodes and the corresponding inputs from Phi nodes.
609 * Merges single exit blocks with single entry blocks and removes
611 * Adds all new nodes to a new hash table for CSE. Does not
612 * perform CSE, so the hash table might contain common subexpressions.
615 dead_node_elimination(ir_graph *irg) {
616 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
618 #ifdef INTERPROCEDURAL_VIEW
619 int rem_ipview = get_interprocedural_view();
621 struct obstack *graveyard_obst = NULL;
622 struct obstack *rebirth_obst = NULL;
623 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
625 /* inform statistics that we started a dead-node elimination run */
626 hook_dead_node_elim(irg, 1);
628 /* Remember external state of current_ir_graph. */
629 rem = current_ir_graph;
630 current_ir_graph = irg;
631 #ifdef INTERPROCEDURAL_VIEW
632 set_interprocedural_view(0);
635 assert(get_irg_phase_state(irg) != phase_building);
637 /* Handle graph state */
638 free_callee_info(irg);
642 /* @@@ so far we loose loops when copying */
643 free_loop_information(irg);
645 set_irg_doms_inconsistent(irg);
647 /* A quiet place, where the old obstack can rest in peace,
648 until it will be cremated. */
649 graveyard_obst = irg->obst;
651 /* A new obstack, where the reachable nodes will be copied to. */
652 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
653 irg->obst = rebirth_obst;
654 obstack_init(irg->obst);
655 irg->last_node_idx = 0;
657 /* We also need a new value table for CSE */
658 del_identities(irg->value_table);
659 irg->value_table = new_identities();
661 /* Copy the graph from the old to the new obstack */
662 copy_graph_env(/*copy_node_nr=*/1);
664 /* Free memory from old unoptimized obstack */
665 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
666 xfree(graveyard_obst); /* ... then free it. */
668 /* inform statistics that the run is over */
669 hook_dead_node_elim(irg, 0);
671 current_ir_graph = rem;
672 #ifdef INTERPROCEDURAL_VIEW
673 set_interprocedural_view(rem_ipview);
679 * Relink bad predecessors of a block and store the old in array to the
680 * link field. This function is called by relink_bad_predecessors().
681 * The array of link field starts with the block operand at position 0.
682 * If block has bad predecessors, create a new in array without bad preds.
683 * Otherwise let in array untouched.
685 static void relink_bad_block_predecessors(ir_node *n, void *env) {
686 ir_node **new_in, *irn;
687 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
690 /* if link field of block is NULL, look for bad predecessors otherwise
691 this is already done */
692 if (is_Block(n) && get_irn_link(n) == NULL) {
693 /* save old predecessors in link field (position 0 is the block operand)*/
694 set_irn_link(n, get_irn_in(n));
696 /* count predecessors without bad nodes */
697 old_irn_arity = get_irn_arity(n);
698 for (i = 0; i < old_irn_arity; i++)
699 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
701 /* arity changing: set new predecessors without bad nodes */
702 if (new_irn_arity < old_irn_arity) {
703 /* Get new predecessor array. We do not resize the array, as we must
704 keep the old one to update Phis. */
705 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
707 /* set new predecessors in array */
710 for (i = 0; i < old_irn_arity; i++) {
711 irn = get_irn_n(n, i);
713 new_in[new_irn_n] = irn;
714 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
718 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
719 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
721 } /* ir node has bad predecessors */
722 } /* Block is not relinked */
726 * Relinks Bad predecessors from Blocks and Phis called by walker
727 * remove_bad_predecesors(). If n is a Block, call
728 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
729 * function of Phi's Block. If this block has bad predecessors, relink preds
732 static void relink_bad_predecessors(ir_node *n, void *env) {
733 ir_node *block, **old_in;
734 int i, old_irn_arity, new_irn_arity;
736 /* relink bad predecessors of a block */
738 relink_bad_block_predecessors(n, env);
740 /* If Phi node relink its block and its predecessors */
742 /* Relink predecessors of phi's block */
743 block = get_nodes_block(n);
744 if (get_irn_link(block) == NULL)
745 relink_bad_block_predecessors(block, env);
747 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
748 old_irn_arity = ARR_LEN(old_in);
750 /* Relink Phi predecessors if count of predecessors changed */
751 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
752 /* set new predecessors in array
753 n->in[0] remains the same block */
755 for(i = 1; i < old_irn_arity; i++)
756 if (!is_Bad((ir_node *)old_in[i])) {
757 n->in[new_irn_arity] = n->in[i];
758 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
762 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
763 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
765 } /* n is a Phi node */
769 * Removes Bad Bad predecessors from Blocks and the corresponding
770 * inputs to Phi nodes as in dead_node_elimination but without
772 * On walking up set the link field to NULL, on walking down call
773 * relink_bad_predecessors() (This function stores the old in array
774 * to the link field and sets a new in array if arity of predecessors
777 void remove_bad_predecessors(ir_graph *irg) {
778 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
785 __)|_| | \_/ | \_/(/_ |_/\__|__
787 The following stuff implements a facility that automatically patches
788 registered ir_node pointers to the new node when a dead node elimination occurs.
791 struct _survive_dce_t {
795 hook_entry_t dead_node_elim;
796 hook_entry_t dead_node_elim_subst;
799 typedef struct _survive_dce_list_t {
800 struct _survive_dce_list_t *next;
802 } survive_dce_list_t;
804 static void dead_node_hook(void *context, ir_graph *irg, int start) {
805 survive_dce_t *sd = context;
808 /* Create a new map before the dead node elimination is performed. */
810 sd->new_places = pmap_create_ex(pmap_count(sd->places));
812 /* Patch back all nodes if dead node elimination is over and something is to be done. */
813 pmap_destroy(sd->places);
814 sd->places = sd->new_places;
815 sd->new_places = NULL;
820 * Hook called when dead node elimination replaces old by nw.
822 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
823 survive_dce_t *sd = context;
824 survive_dce_list_t *list = pmap_get(sd->places, old);
827 /* If the node is to be patched back, write the new address to all registered locations. */
829 survive_dce_list_t *p;
831 for (p = list; p; p = p->next)
834 pmap_insert(sd->new_places, nw, list);
839 * Make a new Survive DCE environment.
841 survive_dce_t *new_survive_dce(void) {
842 survive_dce_t *res = xmalloc(sizeof(res[0]));
843 obstack_init(&res->obst);
844 res->places = pmap_create();
845 res->new_places = NULL;
847 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
848 res->dead_node_elim.context = res;
849 res->dead_node_elim.next = NULL;
851 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
852 res->dead_node_elim_subst.context = res;
853 res->dead_node_elim_subst.next = NULL;
855 #ifndef FIRM_ENABLE_HOOKS
856 assert(0 && "need hooks enabled");
859 register_hook(hook_dead_node_elim, &res->dead_node_elim);
860 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
865 * Free a Survive DCE environment.
867 void free_survive_dce(survive_dce_t *sd) {
868 obstack_free(&sd->obst, NULL);
869 pmap_destroy(sd->places);
870 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
871 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
876 * Register a node pointer to be patched upon DCE.
877 * When DCE occurs, the node pointer specified by @p place will be
878 * patched to the new address of the node it is pointing to.
880 * @param sd The Survive DCE environment.
881 * @param place The address of the node pointer.
883 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
884 if (*place != NULL) {
885 ir_node *irn = *place;
886 survive_dce_list_t *curr = pmap_get(sd->places, irn);
887 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
892 pmap_insert(sd->places, irn, nw);
896 /*--------------------------------------------------------------------*/
897 /* Functionality for inlining */
898 /*--------------------------------------------------------------------*/
901 * Copy node for inlineing. Updates attributes that change when
902 * inlineing but not for dead node elimination.
904 * Copies the node by calling copy_node() and then updates the entity if
905 * it's a local one. env must be a pointer of the frame type of the
906 * inlined procedure. The new entities must be in the link field of
910 copy_node_inline(ir_node *n, void *env) {
912 ir_type *frame_tp = (ir_type *)env;
916 nn = get_new_node (n);
918 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
919 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
921 } else if (is_Block(n)) {
922 nn = get_new_node (n);
923 nn->attr.block.irg = current_ir_graph;
928 * Walker: checks if P_value_arg_base is used.
930 static void find_addr(ir_node *node, void *env) {
931 int *allow_inline = env;
933 is_Start(get_Proj_pred(node)) &&
934 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
940 * Check if we can inline a given call.
941 * Currently, we cannot inline two cases:
942 * - call with compound arguments
943 * - graphs that take the address of a parameter
945 * check these conditions here
947 static int can_inline(ir_node *call, ir_graph *called_graph) {
948 ir_type *call_type = get_Call_type(call);
949 int params, ress, i, res;
950 assert(is_Method_type(call_type));
952 params = get_method_n_params(call_type);
953 ress = get_method_n_ress(call_type);
955 /* check parameters for compound arguments */
956 for (i = 0; i < params; ++i) {
957 ir_type *p_type = get_method_param_type(call_type, i);
959 if (is_compound_type(p_type))
963 /* check results for compound arguments */
964 for (i = 0; i < ress; ++i) {
965 ir_type *r_type = get_method_res_type(call_type, i);
967 if (is_compound_type(r_type))
972 irg_walk_graph(called_graph, find_addr, NULL, &res);
978 exc_handler = 0, /**< There is a handler. */
979 exc_to_end = 1, /**< Branches to End. */
980 exc_no_handler = 2 /**< Exception handling not represented. */
983 /* Inlines a method at the given call site. */
984 int inline_method(ir_node *call, ir_graph *called_graph) {
986 ir_node *post_call, *post_bl;
987 ir_node *in[pn_Start_max];
988 ir_node *end, *end_bl;
992 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
993 enum exc_mode exc_handling;
994 ir_type *called_frame;
995 irg_inline_property prop = get_irg_inline_property(called_graph);
997 if ( (prop < irg_inline_forced) &&
998 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
1000 /* Do not inline variadic functions. */
1001 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
1004 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
1005 get_method_n_params(get_Call_type(call)));
1008 * currently, we cannot inline two cases:
1009 * - call with compound arguments
1010 * - graphs that take the address of a parameter
1012 if (! can_inline(call, called_graph))
1015 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
1016 rem_opt = get_opt_optimize();
1019 /* Handle graph state */
1020 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1021 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
1022 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
1023 set_irg_outs_inconsistent(current_ir_graph);
1024 set_irg_extblk_inconsistent(current_ir_graph);
1025 set_irg_doms_inconsistent(current_ir_graph);
1026 set_irg_loopinfo_inconsistent(current_ir_graph);
1027 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
1029 /* -- Check preconditions -- */
1030 assert(is_Call(call));
1031 /* @@@ does not work for InterfaceIII.java after cgana
1032 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
1033 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
1034 get_Call_type(call)));
1036 if (called_graph == current_ir_graph) {
1037 set_optimize(rem_opt);
1041 /* here we know we WILL inline, so inform the statistics */
1042 hook_inline(call, called_graph);
1044 /* -- Decide how to handle exception control flow: Is there a handler
1045 for the Call node, or do we branch directly to End on an exception?
1047 0 There is a handler.
1049 2 Exception handling not represented in Firm. -- */
1051 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
1052 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
1053 long proj_nr = get_Proj_proj(proj);
1054 if (proj_nr == pn_Call_X_except) Xproj = proj;
1055 if (proj_nr == pn_Call_M_except) Mproj = proj;
1057 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
1058 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
1059 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
1063 the procedure and later replaces the Start node of the called graph.
1064 Post_call is the old Call node and collects the results of the called
1065 graph. Both will end up being a tuple. -- */
1066 post_bl = get_nodes_block(call);
1067 set_irg_current_block(current_ir_graph, post_bl);
1068 /* XxMxPxPxPxT of Start + parameter of Call */
1069 in[pn_Start_X_initial_exec] = new_Jmp();
1070 in[pn_Start_M] = get_Call_mem(call);
1071 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1072 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1073 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1074 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1075 /* in[pn_Start_P_value_arg_base] = ??? */
1076 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1077 pre_call = new_Tuple(pn_Start_max - 1, in);
1081 The new block gets the ins of the old block, pre_call and all its
1082 predecessors and all Phi nodes. -- */
1083 part_block(pre_call);
1085 /* -- Prepare state for dead node elimination -- */
1086 /* Visited flags in calling irg must be >= flag in called irg.
1087 Else walker and arity computation will not work. */
1088 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1089 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1090 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1091 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1092 /* Set pre_call as new Start node in link field of the start node of
1093 calling graph and pre_calls block as new block for the start block
1095 Further mark these nodes so that they are not visited by the
1097 set_irn_link(get_irg_start(called_graph), pre_call);
1098 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1099 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1100 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1101 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1102 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1104 /* Initialize for compaction of in arrays */
1105 inc_irg_block_visited(current_ir_graph);
1107 /* -- Replicate local entities of the called_graph -- */
1108 /* copy the entities. */
1109 called_frame = get_irg_frame_type(called_graph);
1110 for (i = 0; i < get_class_n_members(called_frame); i++) {
1111 ir_entity *new_ent, *old_ent;
1112 old_ent = get_class_member(called_frame, i);
1113 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1114 set_entity_link(old_ent, new_ent);
1117 /* visited is > than that of called graph. With this trick visited will
1118 remain unchanged so that an outer walker, e.g., searching the call nodes
1119 to inline, calling this inline will not visit the inlined nodes. */
1120 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1122 /* -- Performing dead node elimination inlines the graph -- */
1123 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1125 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1126 get_irg_frame_type(called_graph));
1128 /* Repair called_graph */
1129 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1130 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1131 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1133 /* -- Merge the end of the inlined procedure with the call site -- */
1134 /* We will turn the old Call node into a Tuple with the following
1137 0: Phi of all Memories of Return statements.
1138 1: Jmp from new Block that merges the control flow from all exception
1139 predecessors of the old end block.
1140 2: Tuple of all arguments.
1141 3: Phi of Exception memories.
1142 In case the old Call directly branches to End on an exception we don't
1143 need the block merging all exceptions nor the Phi of the exception
1147 /* -- Precompute some values -- */
1148 end_bl = get_new_node(get_irg_end_block(called_graph));
1149 end = get_new_node(get_irg_end(called_graph));
1150 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1151 n_res = get_method_n_ress(get_Call_type(call));
1153 res_pred = xmalloc(n_res * sizeof(*res_pred));
1154 cf_pred = xmalloc(arity * sizeof(*res_pred));
1156 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1158 /* -- archive keepalives -- */
1159 irn_arity = get_irn_arity(end);
1160 for (i = 0; i < irn_arity; i++) {
1161 ir_node *ka = get_End_keepalive(end, i);
1163 add_End_keepalive(get_irg_end(current_ir_graph), ka);
1166 /* The new end node will die. We need not free as the in array is on the obstack:
1167 copy_node() only generated 'D' arrays. */
1169 /* -- Replace Return nodes by Jump nodes. -- */
1171 for (i = 0; i < arity; i++) {
1173 ret = get_irn_n(end_bl, i);
1174 if (is_Return(ret)) {
1175 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1179 set_irn_in(post_bl, n_ret, cf_pred);
1181 /* -- Build a Tuple for all results of the method.
1182 Add Phi node if there was more than one Return. -- */
1183 turn_into_tuple(post_call, pn_Call_max);
1184 /* First the Memory-Phi */
1186 for (i = 0; i < arity; i++) {
1187 ret = get_irn_n(end_bl, i);
1188 if (is_Return(ret)) {
1189 cf_pred[n_ret] = get_Return_mem(ret);
1193 phi = new_Phi(n_ret, cf_pred, mode_M);
1194 set_Tuple_pred(call, pn_Call_M_regular, phi);
1195 /* Conserve Phi-list for further inlinings -- but might be optimized */
1196 if (get_nodes_block(phi) == post_bl) {
1197 set_irn_link(phi, get_irn_link(post_bl));
1198 set_irn_link(post_bl, phi);
1200 /* Now the real results */
1202 for (j = 0; j < n_res; j++) {
1204 for (i = 0; i < arity; i++) {
1205 ret = get_irn_n(end_bl, i);
1206 if (is_Return(ret)) {
1207 cf_pred[n_ret] = get_Return_res(ret, j);
1212 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1216 /* Conserve Phi-list for further inlinings -- but might be optimized */
1217 if (get_nodes_block(phi) == post_bl) {
1218 set_irn_link(phi, get_irn_link(post_bl));
1219 set_irn_link(post_bl, phi);
1222 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1224 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1226 /* handle the regular call */
1227 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1229 /* For now, we cannot inline calls with value_base */
1230 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1232 /* Finally the exception control flow.
1233 We have two (three) possible situations:
1234 First if the Call branches to an exception handler: We need to add a Phi node to
1235 collect the memory containing the exception objects. Further we need
1236 to add another block to get a correct representation of this Phi. To
1237 this block we add a Jmp that resolves into the X output of the Call
1238 when the Call is turned into a tuple.
1239 Second the Call branches to End, the exception is not handled. Just
1240 add all inlined exception branches to the End node.
1241 Third: there is no Exception edge at all. Handle as case two. */
1242 if (exc_handling == exc_handler) {
1244 for (i = 0; i < arity; i++) {
1246 ret = get_irn_n(end_bl, i);
1247 irn = skip_Proj(ret);
1248 if (is_fragile_op(irn) || is_Raise(irn)) {
1249 cf_pred[n_exc] = ret;
1254 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1255 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1256 /* The Phi for the memories with the exception objects */
1258 for (i = 0; i < arity; i++) {
1260 ret = skip_Proj(get_irn_n(end_bl, i));
1262 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1264 } else if (is_fragile_op(ret)) {
1265 /* We rely that all cfops have the memory output at the same position. */
1266 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1268 } else if (is_Raise(ret)) {
1269 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1273 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1275 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1276 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1279 ir_node *main_end_bl;
1280 int main_end_bl_arity;
1281 ir_node **end_preds;
1283 /* assert(exc_handling == 1 || no exceptions. ) */
1285 for (i = 0; i < arity; i++) {
1286 ir_node *ret = get_irn_n(end_bl, i);
1287 ir_node *irn = skip_Proj(ret);
1289 if (is_fragile_op(irn) || is_Raise(irn)) {
1290 cf_pred[n_exc] = ret;
1294 main_end_bl = get_irg_end_block(current_ir_graph);
1295 main_end_bl_arity = get_irn_arity(main_end_bl);
1296 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1298 for (i = 0; i < main_end_bl_arity; ++i)
1299 end_preds[i] = get_irn_n(main_end_bl, i);
1300 for (i = 0; i < n_exc; ++i)
1301 end_preds[main_end_bl_arity + i] = cf_pred[i];
1302 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1303 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1304 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1310 /* -- Turn CSE back on. -- */
1311 set_optimize(rem_opt);
1316 /********************************************************************/
1317 /* Apply inlineing to small methods. */
1318 /********************************************************************/
1320 /** Represents a possible inlinable call in a graph. */
1321 typedef struct _call_entry call_entry;
1322 struct _call_entry {
1323 ir_node *call; /**< the Call */
1324 ir_graph *callee; /**< the callee called here */
1325 call_entry *next; /**< for linking the next one */
1329 * environment for inlining small irgs
1331 typedef struct _inline_env_t {
1332 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1333 call_entry *head; /**< the head of the call entry list */
1334 call_entry *tail; /**< the tail of the call entry list */
1338 * Returns the irg called from a Call node. If the irg is not
1339 * known, NULL is returned.
1341 static ir_graph *get_call_called_irg(ir_node *call) {
1343 ir_graph *called_irg = NULL;
1345 addr = get_Call_ptr(call);
1346 if (is_SymConst(addr) && get_SymConst_kind(addr) == symconst_addr_ent) {
1347 called_irg = get_entity_irg(get_SymConst_entity(addr));
1354 * Walker: Collect all calls to known graphs inside a graph.
1356 static void collect_calls(ir_node *call, void *env) {
1357 if (is_Call(call)) {
1358 ir_graph *called_irg = get_call_called_irg(call);
1360 /* The Call node calls a locally defined method. Remember to inline. */
1361 inline_env_t *ienv = env;
1362 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1364 entry->callee = called_irg;
1367 if (ienv->tail == NULL)
1370 ienv->tail->next = entry;
1377 * Inlines all small methods at call sites where the called address comes
1378 * from a Const node that references the entity representing the called
1380 * The size argument is a rough measure for the code size of the method:
1381 * Methods where the obstack containing the firm graph is smaller than
1384 void inline_small_irgs(ir_graph *irg, int size) {
1385 ir_graph *rem = current_ir_graph;
1388 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1390 if (!(get_opt_optimize() && get_opt_inline())) return;
1392 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1394 current_ir_graph = irg;
1395 /* Handle graph state */
1396 assert(get_irg_phase_state(irg) != phase_building);
1397 free_callee_info(irg);
1399 /* Find Call nodes to inline.
1400 (We can not inline during a walk of the graph, as inlineing the same
1401 method several times changes the visited flag of the walked graph:
1402 after the first inlineing visited of the callee equals visited of
1403 the caller. With the next inlineing both are increased.) */
1404 obstack_init(&env.obst);
1405 env.head = env.tail = NULL;
1406 irg_walk_graph(irg, NULL, collect_calls, &env);
1408 if (env.head != NULL) {
1409 /* There are calls to inline */
1410 collect_phiprojs(irg);
1411 for (entry = env.head; entry != NULL; entry = entry->next) {
1412 ir_graph *callee = entry->callee;
1413 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1414 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1415 inline_method(entry->call, callee);
1419 obstack_free(&env.obst, NULL);
1420 current_ir_graph = rem;
1424 * Environment for inlining irgs.
1427 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1428 int n_nodes_orig; /**< for statistics */
1429 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1430 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1431 int n_call_nodes; /**< Number of Call nodes in the graph. */
1432 int n_call_nodes_orig; /**< for statistics */
1433 int n_callers; /**< Number of known graphs that call this graphs. */
1434 int n_callers_orig; /**< for statistics */
1435 int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
1439 * Allocate a new environment for inlining.
1441 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1442 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1443 env->n_nodes = -2; /* do not count count Start, End */
1444 env->n_nodes_orig = -2; /* do not count Start, End */
1445 env->call_head = NULL;
1446 env->call_tail = NULL;
1447 env->n_call_nodes = 0;
1448 env->n_call_nodes_orig = 0;
1450 env->n_callers_orig = 0;
1451 env->got_inline = 0;
1455 typedef struct walker_env {
1456 struct obstack *obst; /**< the obstack for allocations. */
1457 inline_irg_env *x; /**< the inline environment */
1458 int ignore_runtime; /**< the ignore runtime flag */
1462 * post-walker: collect all calls in the inline-environment
1463 * of a graph and sum some statistics.
1465 static void collect_calls2(ir_node *call, void *ctx) {
1467 inline_irg_env *x = env->x;
1468 ir_op *op = get_irn_op(call);
1472 /* count meaningful nodes in irg */
1473 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1478 if (op != op_Call) return;
1480 /* check, if it's a runtime call */
1481 if (env->ignore_runtime) {
1482 ir_node *symc = get_Call_ptr(call);
1484 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1485 ir_entity *ent = get_SymConst_entity(symc);
1487 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1492 /* collect all call nodes */
1494 ++x->n_call_nodes_orig;
1496 callee = get_call_called_irg(call);
1498 inline_irg_env *callee_env = get_irg_link(callee);
1499 /* count all static callers */
1500 ++callee_env->n_callers;
1501 ++callee_env->n_callers_orig;
1503 /* link it in the list of possible inlinable entries */
1504 entry = obstack_alloc(env->obst, sizeof(*entry));
1506 entry->callee = callee;
1508 if (x->call_tail == NULL)
1509 x->call_head = entry;
1511 x->call_tail->next = entry;
1512 x->call_tail = entry;
1517 * Returns TRUE if the number of callers is 0 in the irg's environment,
1518 * hence this irg is a leave.
1520 INLINE static int is_leave(ir_graph *irg) {
1521 inline_irg_env *env = get_irg_link(irg);
1522 return env->n_call_nodes == 0;
1526 * Returns TRUE if the number of nodes in the callee is
1527 * smaller then size in the irg's environment.
1529 INLINE static int is_smaller(ir_graph *callee, int size) {
1530 inline_irg_env *env = get_irg_link(callee);
1531 return env->n_nodes < size;
1535 * Append the nodes of the list src to the nodes of the list in environment dst.
1537 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1538 call_entry *entry, *nentry;
1540 /* Note that the src list points to Call nodes in the inlined graph, but
1541 we need Call nodes in our graph. Luckily the inliner leaves this information
1542 in the link field. */
1543 for (entry = src; entry != NULL; entry = entry->next) {
1544 nentry = obstack_alloc(obst, sizeof(*nentry));
1545 nentry->call = get_irn_link(entry->call);
1546 nentry->callee = entry->callee;
1547 nentry->next = NULL;
1548 dst->call_tail->next = nentry;
1549 dst->call_tail = nentry;
1554 * Inlines small leave methods at call sites where the called address comes
1555 * from a Const node that references the entity representing the called
1557 * The size argument is a rough measure for the code size of the method:
1558 * Methods where the obstack containing the firm graph is smaller than
1561 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1562 inline_irg_env *env;
1568 call_entry *entry, *tail;
1569 const call_entry *centry;
1570 struct obstack obst;
1571 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1573 if (!(get_opt_optimize() && get_opt_inline())) return;
1575 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1576 rem = current_ir_graph;
1577 obstack_init(&obst);
1579 /* extend all irgs by a temporary data structure for inlining. */
1580 n_irgs = get_irp_n_irgs();
1581 for (i = 0; i < n_irgs; ++i)
1582 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1584 /* Precompute information in temporary data structure. */
1586 wenv.ignore_runtime = ignore_runtime;
1587 for (i = 0; i < n_irgs; ++i) {
1588 ir_graph *irg = get_irp_irg(i);
1590 assert(get_irg_phase_state(irg) != phase_building);
1591 free_callee_info(irg);
1593 wenv.x = get_irg_link(irg);
1594 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1597 /* -- and now inline. -- */
1599 /* Inline leaves recursively -- we might construct new leaves. */
1603 for (i = 0; i < n_irgs; ++i) {
1605 int phiproj_computed = 0;
1607 current_ir_graph = get_irp_irg(i);
1608 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1611 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1614 if (env->n_nodes > maxsize) break;
1617 callee = entry->callee;
1619 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1620 if (!phiproj_computed) {
1621 phiproj_computed = 1;
1622 collect_phiprojs(current_ir_graph);
1624 did_inline = inline_method(call, callee);
1627 /* Do some statistics */
1628 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1630 env->got_inline = 1;
1631 --env->n_call_nodes;
1632 env->n_nodes += callee_env->n_nodes;
1633 --callee_env->n_callers;
1635 /* remove this call from the list */
1637 tail->next = entry->next;
1639 env->call_head = entry->next;
1645 env->call_tail = tail;
1647 } while (did_inline);
1649 /* inline other small functions. */
1650 for (i = 0; i < n_irgs; ++i) {
1652 int phiproj_computed = 0;
1654 current_ir_graph = get_irp_irg(i);
1655 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1657 /* note that the list of possible calls is updated during the process */
1659 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1663 callee = entry->callee;
1665 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1666 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1667 if (!phiproj_computed) {
1668 phiproj_computed = 1;
1669 collect_phiprojs(current_ir_graph);
1671 if (inline_method(call, callee)) {
1672 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1674 /* callee was inline. Append it's call list. */
1675 env->got_inline = 1;
1676 --env->n_call_nodes;
1677 append_call_list(&obst, env, callee_env->call_head);
1678 env->n_call_nodes += callee_env->n_call_nodes;
1679 env->n_nodes += callee_env->n_nodes;
1680 --callee_env->n_callers;
1682 /* after we have inlined callee, all called methods inside callee
1683 are now called once more */
1684 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1685 inline_irg_env *penv = get_irg_link(centry->callee);
1689 /* remove this call from the list */
1691 tail->next = entry->next;
1693 env->call_head = entry->next;
1699 env->call_tail = tail;
1702 for (i = 0; i < n_irgs; ++i) {
1703 irg = get_irp_irg(i);
1704 env = (inline_irg_env *)get_irg_link(irg);
1706 if (env->got_inline) {
1707 /* this irg got calls inlined */
1708 set_irg_outs_inconsistent(irg);
1709 set_irg_doms_inconsistent(irg);
1711 optimize_graph_df(irg);
1714 if (env->got_inline || (env->n_callers_orig != env->n_callers))
1715 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1716 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1717 env->n_callers_orig, env->n_callers,
1718 get_entity_name(get_irg_entity(irg))));
1721 obstack_free(&obst, NULL);
1722 current_ir_graph = rem;
1725 /*******************************************************************/
1726 /* Code Placement. Pins all floating nodes to a block where they */
1727 /* will be executed only if needed. */
1728 /*******************************************************************/
1731 * Returns non-zero, is a block is not reachable from Start.
1733 * @param block the block to test
1736 is_Block_unreachable(ir_node *block) {
1737 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1741 * Find the earliest correct block for node n. --- Place n into the
1742 * same Block as its dominance-deepest Input.
1744 * We have to avoid calls to get_nodes_block() here
1745 * because the graph is floating.
1747 * move_out_of_loops() expects that place_floats_early() have placed
1748 * all "living" nodes into a living block. That's why we must
1749 * move nodes in dead block with "live" successors into a valid
1751 * We move them just into the same block as it's successor (or
1752 * in case of a Phi into the effective use block). For Phi successors,
1753 * this may still be a dead block, but then there is no real use, as
1754 * the control flow will be dead later.
1756 * @param n the node to be placed
1757 * @param worklist a worklist, predecessors of non-floating nodes are placed here
1760 place_floats_early(ir_node *n, waitq *worklist) {
1763 /* we must not run into an infinite loop */
1764 assert(irn_not_visited(n));
1765 mark_irn_visited(n);
1767 /* Place floating nodes. */
1768 if (get_irn_pinned(n) == op_pin_state_floats) {
1769 ir_node *curr_block = get_nodes_block(n);
1770 int in_dead_block = is_Block_unreachable(curr_block);
1772 ir_node *b = NULL; /* The block to place this node in */
1774 assert(is_no_Block(n));
1776 if (is_irn_start_block_placed(n)) {
1777 /* These nodes will not be placed by the loop below. */
1778 b = get_irg_start_block(current_ir_graph);
1782 /* find the block for this node. */
1783 irn_arity = get_irn_arity(n);
1784 for (i = 0; i < irn_arity; i++) {
1785 ir_node *pred = get_irn_n(n, i);
1786 ir_node *pred_block;
1788 if ((irn_not_visited(pred))
1789 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1792 * If the current node is NOT in a dead block, but one of its
1793 * predecessors is, we must move the predecessor to a live block.
1794 * Such thing can happen, if global CSE chose a node from a dead block.
1795 * We move it simply to our block.
1796 * Note that neither Phi nor End nodes are floating, so we don't
1797 * need to handle them here.
1799 if (! in_dead_block) {
1800 if (get_irn_pinned(pred) == op_pin_state_floats &&
1801 is_Block_unreachable(get_nodes_block(pred)))
1802 set_nodes_block(pred, curr_block);
1804 place_floats_early(pred, worklist);
1808 * A node in the Bad block must stay in the bad block,
1809 * so don't compute a new block for it.
1814 /* Because all loops contain at least one op_pin_state_pinned node, now all
1815 our inputs are either op_pin_state_pinned or place_early() has already
1816 been finished on them. We do not have any unfinished inputs! */
1817 pred_block = get_nodes_block(pred);
1818 if ((!is_Block_dead(pred_block)) &&
1819 (get_Block_dom_depth(pred_block) > depth)) {
1821 depth = get_Block_dom_depth(pred_block);
1823 /* Avoid that the node is placed in the Start block */
1825 get_Block_dom_depth(get_nodes_block(n)) > 1 &&
1826 get_irg_phase_state(current_ir_graph) != phase_backend) {
1827 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1828 assert(b != get_irg_start_block(current_ir_graph));
1833 set_nodes_block(n, b);
1837 * Add predecessors of non floating nodes and non-floating predecessors
1838 * of floating nodes to worklist and fix their blocks if the are in dead block.
1840 irn_arity = get_irn_arity(n);
1844 * Simplest case: End node. Predecessors are keep-alives,
1845 * no need to move out of dead block.
1847 for (i = -1; i < irn_arity; ++i) {
1848 ir_node *pred = get_irn_n(n, i);
1849 if (irn_not_visited(pred))
1850 waitq_put(worklist, pred);
1852 } else if (is_Block(n)) {
1854 * Blocks: Predecessors are control flow, no need to move
1855 * them out of dead block.
1857 for (i = irn_arity - 1; i >= 0; --i) {
1858 ir_node *pred = get_irn_n(n, i);
1859 if (irn_not_visited(pred))
1860 waitq_put(worklist, pred);
1862 } else if (is_Phi(n)) {
1864 ir_node *curr_block = get_nodes_block(n);
1865 int in_dead_block = is_Block_unreachable(curr_block);
1868 * Phi nodes: move nodes from dead blocks into the effective use
1869 * of the Phi-input if the Phi is not in a bad block.
1871 pred = get_nodes_block(n);
1872 if (irn_not_visited(pred))
1873 waitq_put(worklist, pred);
1875 for (i = irn_arity - 1; i >= 0; --i) {
1876 ir_node *pred = get_irn_n(n, i);
1878 if (irn_not_visited(pred)) {
1879 if (! in_dead_block &&
1880 get_irn_pinned(pred) == op_pin_state_floats &&
1881 is_Block_unreachable(get_nodes_block(pred))) {
1882 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1884 waitq_put(worklist, pred);
1889 ir_node *curr_block = get_nodes_block(n);
1890 int in_dead_block = is_Block_unreachable(curr_block);
1893 * All other nodes: move nodes from dead blocks into the same block.
1895 pred = get_nodes_block(n);
1896 if (irn_not_visited(pred))
1897 waitq_put(worklist, pred);
1899 for (i = irn_arity - 1; i >= 0; --i) {
1900 ir_node *pred = get_irn_n(n, i);
1902 if (irn_not_visited(pred)) {
1903 if (! in_dead_block &&
1904 get_irn_pinned(pred) == op_pin_state_floats &&
1905 is_Block_unreachable(get_nodes_block(pred))) {
1906 set_nodes_block(pred, curr_block);
1908 waitq_put(worklist, pred);
1915 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1916 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1917 * places all floating nodes reachable from its argument through floating
1918 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1920 * @param worklist a worklist, used for the algorithm, empty on in/output
1922 static void place_early(waitq *worklist) {
1924 inc_irg_visited(current_ir_graph);
1926 /* this inits the worklist */
1927 place_floats_early(get_irg_end(current_ir_graph), worklist);
1929 /* Work the content of the worklist. */
1930 while (!waitq_empty(worklist)) {
1931 ir_node *n = waitq_get(worklist);
1932 if (irn_not_visited(n))
1933 place_floats_early(n, worklist);
1936 set_irg_outs_inconsistent(current_ir_graph);
1937 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1941 * Compute the deepest common ancestor of block and dca.
1943 static ir_node *calc_dca(ir_node *dca, ir_node *block) {
1946 /* we do not want to place nodes in dead blocks */
1947 if (is_Block_dead(block))
1950 /* We found a first legal placement. */
1951 if (!dca) return block;
1953 /* Find a placement that is dominates both, dca and block. */
1954 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1955 block = get_Block_idom(block);
1957 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1958 dca = get_Block_idom(dca);
1961 while (block != dca) {
1962 block = get_Block_idom(block); dca = get_Block_idom(dca);
1968 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1969 * I.e., DCA is the block where we might place PRODUCER.
1970 * A data flow edge points from producer to consumer.
1972 static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1974 /* Compute the last block into which we can place a node so that it is
1976 if (is_Phi(consumer)) {
1977 /* our consumer is a Phi-node, the effective use is in all those
1978 blocks through which the Phi-node reaches producer */
1979 ir_node *phi_block = get_nodes_block(consumer);
1980 int arity = get_irn_arity(consumer);
1983 for (i = 0; i < arity; i++) {
1984 if (get_Phi_pred(consumer, i) == producer) {
1985 ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
1987 if (!is_Block_unreachable(new_block))
1988 dca = calc_dca(dca, new_block);
1992 dca = calc_dca(dca, get_nodes_block(consumer));
1998 /* FIXME: the name clashes here with the function from ana/field_temperature.c
2000 static INLINE int get_irn_loop_depth(ir_node *n) {
2001 return get_loop_depth(get_irn_loop(n));
2005 * Move n to a block with less loop depth than it's current block. The
2006 * new block must be dominated by early.
2008 * @param n the node that should be moved
2009 * @param early the earliest block we can n move to
2011 static void move_out_of_loops(ir_node *n, ir_node *early) {
2012 ir_node *best, *dca;
2016 /* Find the region deepest in the dominator tree dominating
2017 dca with the least loop nesting depth, but still dominated
2018 by our early placement. */
2019 dca = get_nodes_block(n);
2022 while (dca != early) {
2023 dca = get_Block_idom(dca);
2024 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
2025 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
2029 if (best != get_nodes_block(n)) {
2031 printf("Moving out of loop: "); DDMN(n);
2032 printf(" Outermost block: "); DDMN(early);
2033 printf(" Best block: "); DDMN(best);
2034 printf(" Innermost block: "); DDMN(get_nodes_block(n));
2036 set_nodes_block(n, best);
2040 /* deepest common ancestor in the dominator tree of all nodes'
2041 blocks depending on us; our final placement has to dominate DCA. */
2042 static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
2046 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2047 ir_node *succ = get_irn_out(node, i);
2051 * This consumer is the End node, a keep alive edge.
2052 * This is not a real consumer, so we ignore it
2057 if (is_Proj(succ)) {
2058 dca = get_deepest_common_ancestor(succ, dca);
2060 /* ignore if succ is in dead code */
2061 ir_node *succ_blk = get_nodes_block(succ);
2062 if (is_Block_unreachable(succ_blk))
2064 dca = consumer_dom_dca(dca, succ, node);
2071 static void set_projs_block(ir_node *node, ir_node *block)
2075 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2076 ir_node *succ = get_irn_out(node, i);
2078 assert(is_Proj(succ));
2080 if(get_irn_mode(succ) == mode_T) {
2081 set_projs_block(succ, block);
2083 set_nodes_block(succ, block);
2088 * Find the latest legal block for N and place N into the
2089 * `optimal' Block between the latest and earliest legal block.
2090 * The `optimal' block is the dominance-deepest block of those
2091 * with the least loop-nesting-depth. This places N out of as many
2092 * loops as possible and then makes it as control dependent as
2095 * @param n the node to be placed
2096 * @param worklist a worklist, all successors of non-floating nodes are
2099 static void place_floats_late(ir_node *n, pdeq *worklist) {
2103 assert(irn_not_visited(n)); /* no multiple placement */
2105 mark_irn_visited(n);
2107 /* no need to place block nodes, control nodes are already placed. */
2110 (get_irn_mode(n) != mode_X)) {
2111 /* Remember the early_blk placement of this block to move it
2112 out of loop no further than the early_blk placement. */
2113 early_blk = get_nodes_block(n);
2116 * BEWARE: Here we also get code, that is live, but
2117 * was in a dead block. If the node is life, but because
2118 * of CSE in a dead block, we still might need it.
2121 /* Assure that our users are all placed, except the Phi-nodes.
2122 --- Each data flow cycle contains at least one Phi-node. We
2123 have to break the `user has to be placed before the
2124 producer' dependence cycle and the Phi-nodes are the
2125 place to do so, because we need to base our placement on the
2126 final region of our users, which is OK with Phi-nodes, as they
2127 are op_pin_state_pinned, and they never have to be placed after a
2128 producer of one of their inputs in the same block anyway. */
2129 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2130 ir_node *succ = get_irn_out(n, i);
2131 if (irn_not_visited(succ) && !is_Phi(succ))
2132 place_floats_late(succ, worklist);
2135 if (! is_Block_dead(early_blk)) {
2136 /* do only move things that where not dead */
2137 ir_op *op = get_irn_op(n);
2139 /* We have to determine the final block of this node... except for
2140 constants and Projs */
2141 if ((get_irn_pinned(n) == op_pin_state_floats) &&
2143 (op != op_SymConst) &&
2146 /* deepest common ancestor in the dominator tree of all nodes'
2147 blocks depending on us; our final placement has to dominate
2149 ir_node *dca = get_deepest_common_ancestor(n, NULL);
2151 set_nodes_block(n, dca);
2152 move_out_of_loops(n, early_blk);
2153 if(get_irn_mode(n) == mode_T) {
2154 set_projs_block(n, get_nodes_block(n));
2161 /* Add successors of all non-floating nodes on list. (Those of floating
2162 nodes are placed already and therefore are marked.) */
2163 for (i = 0; i < get_irn_n_outs(n); i++) {
2164 ir_node *succ = get_irn_out(n, i);
2165 if (irn_not_visited(get_irn_out(n, i))) {
2166 pdeq_putr(worklist, succ);
2172 * Place floating nodes on the given worklist as late as possible using
2173 * the dominance tree.
2175 * @param worklist the worklist containing the nodes to place
2177 static void place_late(waitq *worklist) {
2179 inc_irg_visited(current_ir_graph);
2181 /* This fills the worklist initially. */
2182 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2184 /* And now empty the worklist again... */
2185 while (!waitq_empty(worklist)) {
2186 ir_node *n = waitq_get(worklist);
2187 if (irn_not_visited(n))
2188 place_floats_late(n, worklist);
2192 /* Code Placement. */
2193 void place_code(ir_graph *irg) {
2195 ir_graph *rem = current_ir_graph;
2197 current_ir_graph = irg;
2199 /* Handle graph state */
2200 assert(get_irg_phase_state(irg) != phase_building);
2203 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2204 free_loop_information(irg);
2205 construct_cf_backedges(irg);
2208 /* Place all floating nodes as early as possible. This guarantees
2209 a legal code placement. */
2210 worklist = new_waitq();
2211 place_early(worklist);
2213 /* place_early() invalidates the outs, place_late needs them. */
2214 compute_irg_outs(irg);
2216 /* Now move the nodes down in the dominator tree. This reduces the
2217 unnecessary executions of the node. */
2218 place_late(worklist);
2220 set_irg_outs_inconsistent(current_ir_graph);
2221 set_irg_loopinfo_inconsistent(current_ir_graph);
2222 del_waitq(worklist);
2223 current_ir_graph = rem;
2226 typedef struct cf_env {
2227 char changed; /**< flag indicates that the cf graphs has changed. */
2231 * Called by walker of remove_critical_cf_edges().
2233 * Place an empty block to an edge between a blocks of multiple
2234 * predecessors and a block of multiple successors.
2237 * @param env Environment of walker.
2239 static void walk_critical_cf_edges(ir_node *n, void *env) {
2241 ir_node *pre, *block, *jmp;
2243 ir_graph *irg = get_irn_irg(n);
2245 /* Block has multiple predecessors */
2246 arity = get_irn_arity(n);
2248 if (n == get_irg_end_block(irg))
2249 return; /* No use to add a block here. */
2251 for (i = 0; i < arity; ++i) {
2254 pre = get_irn_n(n, i);
2255 cfop = get_irn_op(skip_Proj(pre));
2257 if (is_op_fragile(cfop)) {
2258 if (cfop != op_Raise)
2262 /* we don't want place nodes in the start block, so handle it like forking */
2263 if (is_op_forking(cfop) || cfop == op_Start) {
2264 /* Predecessor has multiple successors. Insert new control flow edge edges. */
2266 /* set predecessor of new block */
2267 block = new_r_Block(irg, 1, &pre);
2268 /* insert new jmp node to new block */
2269 jmp = new_r_Jmp(irg, block);
2270 /* set successor of new block */
2271 set_irn_n(n, i, jmp);
2273 } /* predecessor has multiple successors */
2274 } /* for all predecessors */
2275 } /* n is a multi-entry block */
2278 void remove_critical_cf_edges(ir_graph *irg) {
2283 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &env);
2285 /* control flow changed */
2286 set_irg_outs_inconsistent(irg);
2287 set_irg_extblk_inconsistent(irg);
2288 set_irg_doms_inconsistent(irg);
2289 set_irg_loopinfo_inconsistent(irg);