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 panic("Fix backedge handling first");
777 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
784 __)|_| | \_/ | \_/(/_ |_/\__|__
786 The following stuff implements a facility that automatically patches
787 registered ir_node pointers to the new node when a dead node elimination occurs.
790 struct _survive_dce_t {
794 hook_entry_t dead_node_elim;
795 hook_entry_t dead_node_elim_subst;
798 typedef struct _survive_dce_list_t {
799 struct _survive_dce_list_t *next;
801 } survive_dce_list_t;
803 static void dead_node_hook(void *context, ir_graph *irg, int start) {
804 survive_dce_t *sd = context;
807 /* Create a new map before the dead node elimination is performed. */
809 sd->new_places = pmap_create_ex(pmap_count(sd->places));
811 /* Patch back all nodes if dead node elimination is over and something is to be done. */
812 pmap_destroy(sd->places);
813 sd->places = sd->new_places;
814 sd->new_places = NULL;
819 * Hook called when dead node elimination replaces old by nw.
821 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
822 survive_dce_t *sd = context;
823 survive_dce_list_t *list = pmap_get(sd->places, old);
826 /* If the node is to be patched back, write the new address to all registered locations. */
828 survive_dce_list_t *p;
830 for (p = list; p; p = p->next)
833 pmap_insert(sd->new_places, nw, list);
838 * Make a new Survive DCE environment.
840 survive_dce_t *new_survive_dce(void) {
841 survive_dce_t *res = xmalloc(sizeof(res[0]));
842 obstack_init(&res->obst);
843 res->places = pmap_create();
844 res->new_places = NULL;
846 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
847 res->dead_node_elim.context = res;
848 res->dead_node_elim.next = NULL;
850 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
851 res->dead_node_elim_subst.context = res;
852 res->dead_node_elim_subst.next = NULL;
854 #ifndef FIRM_ENABLE_HOOKS
855 assert(0 && "need hooks enabled");
858 register_hook(hook_dead_node_elim, &res->dead_node_elim);
859 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
864 * Free a Survive DCE environment.
866 void free_survive_dce(survive_dce_t *sd) {
867 obstack_free(&sd->obst, NULL);
868 pmap_destroy(sd->places);
869 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
870 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
875 * Register a node pointer to be patched upon DCE.
876 * When DCE occurs, the node pointer specified by @p place will be
877 * patched to the new address of the node it is pointing to.
879 * @param sd The Survive DCE environment.
880 * @param place The address of the node pointer.
882 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
883 if (*place != NULL) {
884 ir_node *irn = *place;
885 survive_dce_list_t *curr = pmap_get(sd->places, irn);
886 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
891 pmap_insert(sd->places, irn, nw);
895 /*--------------------------------------------------------------------*/
896 /* Functionality for inlining */
897 /*--------------------------------------------------------------------*/
900 * Copy node for inlineing. Updates attributes that change when
901 * inlineing but not for dead node elimination.
903 * Copies the node by calling copy_node() and then updates the entity if
904 * it's a local one. env must be a pointer of the frame type of the
905 * inlined procedure. The new entities must be in the link field of
909 copy_node_inline(ir_node *n, void *env) {
911 ir_type *frame_tp = (ir_type *)env;
915 nn = get_new_node (n);
917 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
918 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
920 } else if (is_Block(n)) {
921 nn = get_new_node (n);
922 nn->attr.block.irg = current_ir_graph;
927 * Walker: checks if P_value_arg_base is used.
929 static void find_addr(ir_node *node, void *env) {
930 int *allow_inline = env;
932 is_Start(get_Proj_pred(node)) &&
933 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
939 * Check if we can inline a given call.
940 * Currently, we cannot inline two cases:
941 * - call with compound arguments
942 * - graphs that take the address of a parameter
944 * check these conditions here
946 static int can_inline(ir_node *call, ir_graph *called_graph) {
947 ir_type *call_type = get_Call_type(call);
948 int params, ress, i, res;
949 assert(is_Method_type(call_type));
951 params = get_method_n_params(call_type);
952 ress = get_method_n_ress(call_type);
954 /* check parameters for compound arguments */
955 for (i = 0; i < params; ++i) {
956 ir_type *p_type = get_method_param_type(call_type, i);
958 if (is_compound_type(p_type))
962 /* check results for compound arguments */
963 for (i = 0; i < ress; ++i) {
964 ir_type *r_type = get_method_res_type(call_type, i);
966 if (is_compound_type(r_type))
971 irg_walk_graph(called_graph, find_addr, NULL, &res);
977 exc_handler = 0, /**< There is a handler. */
978 exc_to_end = 1, /**< Branches to End. */
979 exc_no_handler = 2 /**< Exception handling not represented. */
982 /* Inlines a method at the given call site. */
983 int inline_method(ir_node *call, ir_graph *called_graph) {
985 ir_node *post_call, *post_bl;
986 ir_node *in[pn_Start_max];
987 ir_node *end, *end_bl;
991 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
992 enum exc_mode exc_handling;
993 ir_type *called_frame;
994 irg_inline_property prop = get_irg_inline_property(called_graph);
996 if ( (prop < irg_inline_forced) || (prop == irg_inline_forbidden))
999 /* Do not inline variadic functions. */
1000 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
1003 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
1004 get_method_n_params(get_Call_type(call)));
1007 * currently, we cannot inline two cases:
1008 * - call with compound arguments
1009 * - graphs that take the address of a parameter
1011 if (! can_inline(call, called_graph))
1014 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
1015 rem_opt = get_opt_optimize();
1018 /* Handle graph state */
1019 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1020 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
1021 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
1022 set_irg_outs_inconsistent(current_ir_graph);
1023 set_irg_extblk_inconsistent(current_ir_graph);
1024 set_irg_doms_inconsistent(current_ir_graph);
1025 set_irg_loopinfo_inconsistent(current_ir_graph);
1026 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
1028 /* -- Check preconditions -- */
1029 assert(is_Call(call));
1030 /* @@@ does not work for InterfaceIII.java after cgana
1031 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
1032 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
1033 get_Call_type(call)));
1035 if (called_graph == current_ir_graph) {
1036 set_optimize(rem_opt);
1040 /* here we know we WILL inline, so inform the statistics */
1041 hook_inline(call, called_graph);
1043 /* -- Decide how to handle exception control flow: Is there a handler
1044 for the Call node, or do we branch directly to End on an exception?
1046 0 There is a handler.
1048 2 Exception handling not represented in Firm. -- */
1050 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
1051 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
1052 long proj_nr = get_Proj_proj(proj);
1053 if (proj_nr == pn_Call_X_except) Xproj = proj;
1054 if (proj_nr == pn_Call_M_except) Mproj = proj;
1056 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
1057 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
1058 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
1062 the procedure and later replaces the Start node of the called graph.
1063 Post_call is the old Call node and collects the results of the called
1064 graph. Both will end up being a tuple. -- */
1065 post_bl = get_nodes_block(call);
1066 set_irg_current_block(current_ir_graph, post_bl);
1067 /* XxMxPxPxPxT of Start + parameter of Call */
1068 in[pn_Start_X_initial_exec] = new_Jmp();
1069 in[pn_Start_M] = get_Call_mem(call);
1070 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1071 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1072 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1073 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1074 /* in[pn_Start_P_value_arg_base] = ??? */
1075 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1076 pre_call = new_Tuple(pn_Start_max - 1, in);
1080 The new block gets the ins of the old block, pre_call and all its
1081 predecessors and all Phi nodes. -- */
1082 part_block(pre_call);
1084 /* -- Prepare state for dead node elimination -- */
1085 /* Visited flags in calling irg must be >= flag in called irg.
1086 Else walker and arity computation will not work. */
1087 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1088 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1089 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1090 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1091 /* Set pre_call as new Start node in link field of the start node of
1092 calling graph and pre_calls block as new block for the start block
1094 Further mark these nodes so that they are not visited by the
1096 set_irn_link(get_irg_start(called_graph), pre_call);
1097 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1098 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1099 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1100 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1101 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1103 /* Initialize for compaction of in arrays */
1104 inc_irg_block_visited(current_ir_graph);
1106 /* -- Replicate local entities of the called_graph -- */
1107 /* copy the entities. */
1108 called_frame = get_irg_frame_type(called_graph);
1109 for (i = 0; i < get_class_n_members(called_frame); i++) {
1110 ir_entity *new_ent, *old_ent;
1111 old_ent = get_class_member(called_frame, i);
1112 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1113 set_entity_link(old_ent, new_ent);
1116 /* visited is > than that of called graph. With this trick visited will
1117 remain unchanged so that an outer walker, e.g., searching the call nodes
1118 to inline, calling this inline will not visit the inlined nodes. */
1119 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1121 /* -- Performing dead node elimination inlines the graph -- */
1122 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1124 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1125 get_irg_frame_type(called_graph));
1127 /* Repair called_graph */
1128 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1129 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1130 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1132 /* -- Merge the end of the inlined procedure with the call site -- */
1133 /* We will turn the old Call node into a Tuple with the following
1136 0: Phi of all Memories of Return statements.
1137 1: Jmp from new Block that merges the control flow from all exception
1138 predecessors of the old end block.
1139 2: Tuple of all arguments.
1140 3: Phi of Exception memories.
1141 In case the old Call directly branches to End on an exception we don't
1142 need the block merging all exceptions nor the Phi of the exception
1146 /* -- Precompute some values -- */
1147 end_bl = get_new_node(get_irg_end_block(called_graph));
1148 end = get_new_node(get_irg_end(called_graph));
1149 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1150 n_res = get_method_n_ress(get_Call_type(call));
1152 res_pred = xmalloc(n_res * sizeof(*res_pred));
1153 cf_pred = xmalloc(arity * sizeof(*res_pred));
1155 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1157 /* -- archive keepalives -- */
1158 irn_arity = get_irn_arity(end);
1159 for (i = 0; i < irn_arity; i++) {
1160 ir_node *ka = get_End_keepalive(end, i);
1162 add_End_keepalive(get_irg_end(current_ir_graph), ka);
1165 /* The new end node will die. We need not free as the in array is on the obstack:
1166 copy_node() only generated 'D' arrays. */
1168 /* -- Replace Return nodes by Jump nodes. -- */
1170 for (i = 0; i < arity; i++) {
1172 ret = get_irn_n(end_bl, i);
1173 if (is_Return(ret)) {
1174 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1178 set_irn_in(post_bl, n_ret, cf_pred);
1180 /* -- Build a Tuple for all results of the method.
1181 Add Phi node if there was more than one Return. -- */
1182 turn_into_tuple(post_call, pn_Call_max);
1183 /* First the Memory-Phi */
1185 for (i = 0; i < arity; i++) {
1186 ret = get_irn_n(end_bl, i);
1187 if (is_Return(ret)) {
1188 cf_pred[n_ret] = get_Return_mem(ret);
1192 phi = new_Phi(n_ret, cf_pred, mode_M);
1193 set_Tuple_pred(call, pn_Call_M_regular, phi);
1194 /* Conserve Phi-list for further inlinings -- but might be optimized */
1195 if (get_nodes_block(phi) == post_bl) {
1196 set_irn_link(phi, get_irn_link(post_bl));
1197 set_irn_link(post_bl, phi);
1199 /* Now the real results */
1201 for (j = 0; j < n_res; j++) {
1203 for (i = 0; i < arity; i++) {
1204 ret = get_irn_n(end_bl, i);
1205 if (is_Return(ret)) {
1206 cf_pred[n_ret] = get_Return_res(ret, j);
1211 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1215 /* Conserve Phi-list for further inlinings -- but might be optimized */
1216 if (get_nodes_block(phi) == post_bl) {
1217 set_irn_link(phi, get_irn_link(post_bl));
1218 set_irn_link(post_bl, phi);
1221 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1223 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1225 /* handle the regular call */
1226 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1228 /* For now, we cannot inline calls with value_base */
1229 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1231 /* Finally the exception control flow.
1232 We have two (three) possible situations:
1233 First if the Call branches to an exception handler: We need to add a Phi node to
1234 collect the memory containing the exception objects. Further we need
1235 to add another block to get a correct representation of this Phi. To
1236 this block we add a Jmp that resolves into the X output of the Call
1237 when the Call is turned into a tuple.
1238 Second the Call branches to End, the exception is not handled. Just
1239 add all inlined exception branches to the End node.
1240 Third: there is no Exception edge at all. Handle as case two. */
1241 if (exc_handling == exc_handler) {
1243 for (i = 0; i < arity; i++) {
1245 ret = get_irn_n(end_bl, i);
1246 irn = skip_Proj(ret);
1247 if (is_fragile_op(irn) || is_Raise(irn)) {
1248 cf_pred[n_exc] = ret;
1253 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1254 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1255 /* The Phi for the memories with the exception objects */
1257 for (i = 0; i < arity; i++) {
1259 ret = skip_Proj(get_irn_n(end_bl, i));
1261 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1263 } else if (is_fragile_op(ret)) {
1264 /* We rely that all cfops have the memory output at the same position. */
1265 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1267 } else if (is_Raise(ret)) {
1268 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1272 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1274 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1275 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1278 ir_node *main_end_bl;
1279 int main_end_bl_arity;
1280 ir_node **end_preds;
1282 /* assert(exc_handling == 1 || no exceptions. ) */
1284 for (i = 0; i < arity; i++) {
1285 ir_node *ret = get_irn_n(end_bl, i);
1286 ir_node *irn = skip_Proj(ret);
1288 if (is_fragile_op(irn) || is_Raise(irn)) {
1289 cf_pred[n_exc] = ret;
1293 main_end_bl = get_irg_end_block(current_ir_graph);
1294 main_end_bl_arity = get_irn_arity(main_end_bl);
1295 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1297 for (i = 0; i < main_end_bl_arity; ++i)
1298 end_preds[i] = get_irn_n(main_end_bl, i);
1299 for (i = 0; i < n_exc; ++i)
1300 end_preds[main_end_bl_arity + i] = cf_pred[i];
1301 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1302 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1303 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1309 /* -- Turn CSE back on. -- */
1310 set_optimize(rem_opt);
1315 /********************************************************************/
1316 /* Apply inlineing to small methods. */
1317 /********************************************************************/
1319 /** Represents a possible inlinable call in a graph. */
1320 typedef struct _call_entry call_entry;
1321 struct _call_entry {
1322 ir_node *call; /**< the Call */
1323 ir_graph *callee; /**< the callee called here */
1324 call_entry *next; /**< for linking the next one */
1328 * environment for inlining small irgs
1330 typedef struct _inline_env_t {
1331 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1332 call_entry *head; /**< the head of the call entry list */
1333 call_entry *tail; /**< the tail of the call entry list */
1337 * Returns the irg called from a Call node. If the irg is not
1338 * known, NULL is returned.
1340 static ir_graph *get_call_called_irg(ir_node *call) {
1342 ir_graph *called_irg = NULL;
1344 addr = get_Call_ptr(call);
1345 if (is_SymConst(addr) && get_SymConst_kind(addr) == symconst_addr_ent) {
1346 called_irg = get_entity_irg(get_SymConst_entity(addr));
1353 * Walker: Collect all calls to known graphs inside a graph.
1355 static void collect_calls(ir_node *call, void *env) {
1356 if (is_Call(call)) {
1357 ir_graph *called_irg = get_call_called_irg(call);
1359 /* The Call node calls a locally defined method. Remember to inline. */
1360 inline_env_t *ienv = env;
1361 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1363 entry->callee = called_irg;
1366 if (ienv->tail == NULL)
1369 ienv->tail->next = entry;
1376 * Inlines all small methods at call sites where the called address comes
1377 * from a Const node that references the entity representing the called
1379 * The size argument is a rough measure for the code size of the method:
1380 * Methods where the obstack containing the firm graph is smaller than
1383 void inline_small_irgs(ir_graph *irg, int size) {
1384 ir_graph *rem = current_ir_graph;
1387 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1389 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1391 current_ir_graph = irg;
1392 /* Handle graph state */
1393 assert(get_irg_phase_state(irg) != phase_building);
1394 free_callee_info(irg);
1396 /* Find Call nodes to inline.
1397 (We can not inline during a walk of the graph, as inlineing the same
1398 method several times changes the visited flag of the walked graph:
1399 after the first inlineing visited of the callee equals visited of
1400 the caller. With the next inlineing both are increased.) */
1401 obstack_init(&env.obst);
1402 env.head = env.tail = NULL;
1403 irg_walk_graph(irg, NULL, collect_calls, &env);
1405 if (env.head != NULL) {
1406 /* There are calls to inline */
1407 collect_phiprojs(irg);
1408 for (entry = env.head; entry != NULL; entry = entry->next) {
1409 ir_graph *callee = entry->callee;
1410 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1411 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1412 inline_method(entry->call, callee);
1416 obstack_free(&env.obst, NULL);
1417 current_ir_graph = rem;
1421 * Environment for inlining irgs.
1424 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1425 int n_nodes_orig; /**< for statistics */
1426 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1427 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1428 int n_call_nodes; /**< Number of Call nodes in the graph. */
1429 int n_call_nodes_orig; /**< for statistics */
1430 int n_callers; /**< Number of known graphs that call this graphs. */
1431 int n_callers_orig; /**< for statistics */
1432 int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
1436 * Allocate a new environment for inlining.
1438 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1439 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1440 env->n_nodes = -2; /* do not count count Start, End */
1441 env->n_nodes_orig = -2; /* do not count Start, End */
1442 env->call_head = NULL;
1443 env->call_tail = NULL;
1444 env->n_call_nodes = 0;
1445 env->n_call_nodes_orig = 0;
1447 env->n_callers_orig = 0;
1448 env->got_inline = 0;
1452 typedef struct walker_env {
1453 struct obstack *obst; /**< the obstack for allocations. */
1454 inline_irg_env *x; /**< the inline environment */
1455 int ignore_runtime; /**< the ignore runtime flag */
1459 * post-walker: collect all calls in the inline-environment
1460 * of a graph and sum some statistics.
1462 static void collect_calls2(ir_node *call, void *ctx) {
1464 inline_irg_env *x = env->x;
1465 ir_op *op = get_irn_op(call);
1469 /* count meaningful nodes in irg */
1470 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1475 if (op != op_Call) return;
1477 /* check, if it's a runtime call */
1478 if (env->ignore_runtime) {
1479 ir_node *symc = get_Call_ptr(call);
1481 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1482 ir_entity *ent = get_SymConst_entity(symc);
1484 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1489 /* collect all call nodes */
1491 ++x->n_call_nodes_orig;
1493 callee = get_call_called_irg(call);
1495 inline_irg_env *callee_env = get_irg_link(callee);
1496 /* count all static callers */
1497 ++callee_env->n_callers;
1498 ++callee_env->n_callers_orig;
1500 /* link it in the list of possible inlinable entries */
1501 entry = obstack_alloc(env->obst, sizeof(*entry));
1503 entry->callee = callee;
1505 if (x->call_tail == NULL)
1506 x->call_head = entry;
1508 x->call_tail->next = entry;
1509 x->call_tail = entry;
1514 * Returns TRUE if the number of callers is 0 in the irg's environment,
1515 * hence this irg is a leave.
1517 INLINE static int is_leave(ir_graph *irg) {
1518 inline_irg_env *env = get_irg_link(irg);
1519 return env->n_call_nodes == 0;
1523 * Returns TRUE if the number of nodes in the callee is
1524 * smaller then size in the irg's environment.
1526 INLINE static int is_smaller(ir_graph *callee, int size) {
1527 inline_irg_env *env = get_irg_link(callee);
1528 return env->n_nodes < size;
1532 * Append the nodes of the list src to the nodes of the list in environment dst.
1534 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1535 call_entry *entry, *nentry;
1537 /* Note that the src list points to Call nodes in the inlined graph, but
1538 we need Call nodes in our graph. Luckily the inliner leaves this information
1539 in the link field. */
1540 for (entry = src; entry != NULL; entry = entry->next) {
1541 nentry = obstack_alloc(obst, sizeof(*nentry));
1542 nentry->call = get_irn_link(entry->call);
1543 nentry->callee = entry->callee;
1544 nentry->next = NULL;
1545 dst->call_tail->next = nentry;
1546 dst->call_tail = nentry;
1551 * Inlines small leave methods at call sites where the called address comes
1552 * from a Const node that references the entity representing the called
1554 * The size argument is a rough measure for the code size of the method:
1555 * Methods where the obstack containing the firm graph is smaller than
1558 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1559 inline_irg_env *env;
1565 call_entry *entry, *tail;
1566 const call_entry *centry;
1567 struct obstack obst;
1568 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1570 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1571 rem = current_ir_graph;
1572 obstack_init(&obst);
1574 /* extend all irgs by a temporary data structure for inlining. */
1575 n_irgs = get_irp_n_irgs();
1576 for (i = 0; i < n_irgs; ++i)
1577 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1579 /* Precompute information in temporary data structure. */
1581 wenv.ignore_runtime = ignore_runtime;
1582 for (i = 0; i < n_irgs; ++i) {
1583 ir_graph *irg = get_irp_irg(i);
1585 assert(get_irg_phase_state(irg) != phase_building);
1586 free_callee_info(irg);
1588 wenv.x = get_irg_link(irg);
1589 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1592 /* -- and now inline. -- */
1594 /* Inline leaves recursively -- we might construct new leaves. */
1598 for (i = 0; i < n_irgs; ++i) {
1600 int phiproj_computed = 0;
1602 current_ir_graph = get_irp_irg(i);
1603 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1606 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1609 if (env->n_nodes > maxsize) break;
1612 callee = entry->callee;
1614 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1615 if (!phiproj_computed) {
1616 phiproj_computed = 1;
1617 collect_phiprojs(current_ir_graph);
1619 did_inline = inline_method(call, callee);
1622 /* Do some statistics */
1623 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1625 env->got_inline = 1;
1626 --env->n_call_nodes;
1627 env->n_nodes += callee_env->n_nodes;
1628 --callee_env->n_callers;
1630 /* remove this call from the list */
1632 tail->next = entry->next;
1634 env->call_head = entry->next;
1640 env->call_tail = tail;
1642 } while (did_inline);
1644 /* inline other small functions. */
1645 for (i = 0; i < n_irgs; ++i) {
1647 int phiproj_computed = 0;
1649 current_ir_graph = get_irp_irg(i);
1650 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1652 /* note that the list of possible calls is updated during the process */
1654 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1658 callee = entry->callee;
1660 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1661 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1662 if (!phiproj_computed) {
1663 phiproj_computed = 1;
1664 collect_phiprojs(current_ir_graph);
1666 if (inline_method(call, callee)) {
1667 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1669 /* callee was inline. Append it's call list. */
1670 env->got_inline = 1;
1671 --env->n_call_nodes;
1672 append_call_list(&obst, env, callee_env->call_head);
1673 env->n_call_nodes += callee_env->n_call_nodes;
1674 env->n_nodes += callee_env->n_nodes;
1675 --callee_env->n_callers;
1677 /* after we have inlined callee, all called methods inside callee
1678 are now called once more */
1679 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1680 inline_irg_env *penv = get_irg_link(centry->callee);
1684 /* remove this call from the list */
1686 tail->next = entry->next;
1688 env->call_head = entry->next;
1694 env->call_tail = tail;
1697 for (i = 0; i < n_irgs; ++i) {
1698 irg = get_irp_irg(i);
1699 env = (inline_irg_env *)get_irg_link(irg);
1701 if (env->got_inline) {
1702 /* this irg got calls inlined */
1703 set_irg_outs_inconsistent(irg);
1704 set_irg_doms_inconsistent(irg);
1706 optimize_graph_df(irg);
1709 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1710 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1711 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1712 env->n_callers_orig, env->n_callers,
1713 get_entity_name(get_irg_entity(irg))));
1717 obstack_free(&obst, NULL);
1718 current_ir_graph = rem;
1721 /*******************************************************************/
1722 /* Code Placement. Pins all floating nodes to a block where they */
1723 /* will be executed only if needed. */
1724 /*******************************************************************/
1727 * Returns non-zero, is a block is not reachable from Start.
1729 * @param block the block to test
1732 is_Block_unreachable(ir_node *block) {
1733 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1737 * Find the earliest correct block for node n. --- Place n into the
1738 * same Block as its dominance-deepest Input.
1740 * We have to avoid calls to get_nodes_block() here
1741 * because the graph is floating.
1743 * move_out_of_loops() expects that place_floats_early() have placed
1744 * all "living" nodes into a living block. That's why we must
1745 * move nodes in dead block with "live" successors into a valid
1747 * We move them just into the same block as it's successor (or
1748 * in case of a Phi into the effective use block). For Phi successors,
1749 * this may still be a dead block, but then there is no real use, as
1750 * the control flow will be dead later.
1752 * @param n the node to be placed
1753 * @param worklist a worklist, predecessors of non-floating nodes are placed here
1756 place_floats_early(ir_node *n, waitq *worklist) {
1759 /* we must not run into an infinite loop */
1760 assert(irn_not_visited(n));
1761 mark_irn_visited(n);
1763 /* Place floating nodes. */
1764 if (get_irn_pinned(n) == op_pin_state_floats) {
1765 ir_node *curr_block = get_nodes_block(n);
1766 int in_dead_block = is_Block_unreachable(curr_block);
1768 ir_node *b = NULL; /* The block to place this node in */
1770 assert(is_no_Block(n));
1772 if (is_irn_start_block_placed(n)) {
1773 /* These nodes will not be placed by the loop below. */
1774 b = get_irg_start_block(current_ir_graph);
1778 /* find the block for this node. */
1779 irn_arity = get_irn_arity(n);
1780 for (i = 0; i < irn_arity; i++) {
1781 ir_node *pred = get_irn_n(n, i);
1782 ir_node *pred_block;
1784 if ((irn_not_visited(pred))
1785 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1788 * If the current node is NOT in a dead block, but one of its
1789 * predecessors is, we must move the predecessor to a live block.
1790 * Such thing can happen, if global CSE chose a node from a dead block.
1791 * We move it simply to our block.
1792 * Note that neither Phi nor End nodes are floating, so we don't
1793 * need to handle them here.
1795 if (! in_dead_block) {
1796 if (get_irn_pinned(pred) == op_pin_state_floats &&
1797 is_Block_unreachable(get_nodes_block(pred)))
1798 set_nodes_block(pred, curr_block);
1800 place_floats_early(pred, worklist);
1804 * A node in the Bad block must stay in the bad block,
1805 * so don't compute a new block for it.
1810 /* Because all loops contain at least one op_pin_state_pinned node, now all
1811 our inputs are either op_pin_state_pinned or place_early() has already
1812 been finished on them. We do not have any unfinished inputs! */
1813 pred_block = get_nodes_block(pred);
1814 if ((!is_Block_dead(pred_block)) &&
1815 (get_Block_dom_depth(pred_block) > depth)) {
1817 depth = get_Block_dom_depth(pred_block);
1819 /* Avoid that the node is placed in the Start block */
1821 get_Block_dom_depth(get_nodes_block(n)) > 1 &&
1822 get_irg_phase_state(current_ir_graph) != phase_backend) {
1823 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1824 assert(b != get_irg_start_block(current_ir_graph));
1829 set_nodes_block(n, b);
1833 * Add predecessors of non floating nodes and non-floating predecessors
1834 * of floating nodes to worklist and fix their blocks if the are in dead block.
1836 irn_arity = get_irn_arity(n);
1840 * Simplest case: End node. Predecessors are keep-alives,
1841 * no need to move out of dead block.
1843 for (i = -1; i < irn_arity; ++i) {
1844 ir_node *pred = get_irn_n(n, i);
1845 if (irn_not_visited(pred))
1846 waitq_put(worklist, pred);
1848 } else if (is_Block(n)) {
1850 * Blocks: Predecessors are control flow, no need to move
1851 * them out of dead block.
1853 for (i = irn_arity - 1; i >= 0; --i) {
1854 ir_node *pred = get_irn_n(n, i);
1855 if (irn_not_visited(pred))
1856 waitq_put(worklist, pred);
1858 } else if (is_Phi(n)) {
1860 ir_node *curr_block = get_nodes_block(n);
1861 int in_dead_block = is_Block_unreachable(curr_block);
1864 * Phi nodes: move nodes from dead blocks into the effective use
1865 * of the Phi-input if the Phi is not in a bad block.
1867 pred = get_nodes_block(n);
1868 if (irn_not_visited(pred))
1869 waitq_put(worklist, pred);
1871 for (i = irn_arity - 1; i >= 0; --i) {
1872 ir_node *pred = get_irn_n(n, i);
1874 if (irn_not_visited(pred)) {
1875 if (! in_dead_block &&
1876 get_irn_pinned(pred) == op_pin_state_floats &&
1877 is_Block_unreachable(get_nodes_block(pred))) {
1878 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1880 waitq_put(worklist, pred);
1885 ir_node *curr_block = get_nodes_block(n);
1886 int in_dead_block = is_Block_unreachable(curr_block);
1889 * All other nodes: move nodes from dead blocks into the same block.
1891 pred = get_nodes_block(n);
1892 if (irn_not_visited(pred))
1893 waitq_put(worklist, pred);
1895 for (i = irn_arity - 1; i >= 0; --i) {
1896 ir_node *pred = get_irn_n(n, i);
1898 if (irn_not_visited(pred)) {
1899 if (! in_dead_block &&
1900 get_irn_pinned(pred) == op_pin_state_floats &&
1901 is_Block_unreachable(get_nodes_block(pred))) {
1902 set_nodes_block(pred, curr_block);
1904 waitq_put(worklist, pred);
1911 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1912 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1913 * places all floating nodes reachable from its argument through floating
1914 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1916 * @param worklist a worklist, used for the algorithm, empty on in/output
1918 static void place_early(waitq *worklist) {
1920 inc_irg_visited(current_ir_graph);
1922 /* this inits the worklist */
1923 place_floats_early(get_irg_end(current_ir_graph), worklist);
1925 /* Work the content of the worklist. */
1926 while (!waitq_empty(worklist)) {
1927 ir_node *n = waitq_get(worklist);
1928 if (irn_not_visited(n))
1929 place_floats_early(n, worklist);
1932 set_irg_outs_inconsistent(current_ir_graph);
1933 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1937 * Compute the deepest common ancestor of block and dca.
1939 static ir_node *calc_dca(ir_node *dca, ir_node *block) {
1942 /* we do not want to place nodes in dead blocks */
1943 if (is_Block_dead(block))
1946 /* We found a first legal placement. */
1947 if (!dca) return block;
1949 /* Find a placement that is dominates both, dca and block. */
1950 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1951 block = get_Block_idom(block);
1953 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1954 dca = get_Block_idom(dca);
1957 while (block != dca) {
1958 block = get_Block_idom(block); dca = get_Block_idom(dca);
1964 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1965 * I.e., DCA is the block where we might place PRODUCER.
1966 * A data flow edge points from producer to consumer.
1968 static ir_node *consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1970 /* Compute the last block into which we can place a node so that it is
1972 if (is_Phi(consumer)) {
1973 /* our consumer is a Phi-node, the effective use is in all those
1974 blocks through which the Phi-node reaches producer */
1975 ir_node *phi_block = get_nodes_block(consumer);
1976 int arity = get_irn_arity(consumer);
1979 for (i = 0; i < arity; i++) {
1980 if (get_Phi_pred(consumer, i) == producer) {
1981 ir_node *new_block = get_Block_cfgpred_block(phi_block, i);
1983 if (!is_Block_unreachable(new_block))
1984 dca = calc_dca(dca, new_block);
1988 dca = calc_dca(dca, get_nodes_block(consumer));
1994 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1996 static INLINE int get_irn_loop_depth(ir_node *n) {
1997 return get_loop_depth(get_irn_loop(n));
2001 * Move n to a block with less loop depth than it's current block. The
2002 * new block must be dominated by early.
2004 * @param n the node that should be moved
2005 * @param early the earliest block we can n move to
2007 static void move_out_of_loops(ir_node *n, ir_node *early) {
2008 ir_node *best, *dca;
2012 /* Find the region deepest in the dominator tree dominating
2013 dca with the least loop nesting depth, but still dominated
2014 by our early placement. */
2015 dca = get_nodes_block(n);
2018 while (dca != early) {
2019 dca = get_Block_idom(dca);
2020 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
2021 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
2025 if (best != get_nodes_block(n)) {
2027 printf("Moving out of loop: "); DDMN(n);
2028 printf(" Outermost block: "); DDMN(early);
2029 printf(" Best block: "); DDMN(best);
2030 printf(" Innermost block: "); DDMN(get_nodes_block(n));
2032 set_nodes_block(n, best);
2036 /* deepest common ancestor in the dominator tree of all nodes'
2037 blocks depending on us; our final placement has to dominate DCA. */
2038 static ir_node *get_deepest_common_ancestor(ir_node *node, ir_node *dca)
2042 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2043 ir_node *succ = get_irn_out(node, i);
2047 * This consumer is the End node, a keep alive edge.
2048 * This is not a real consumer, so we ignore it
2053 if (is_Proj(succ)) {
2054 dca = get_deepest_common_ancestor(succ, dca);
2056 /* ignore if succ is in dead code */
2057 ir_node *succ_blk = get_nodes_block(succ);
2058 if (is_Block_unreachable(succ_blk))
2060 dca = consumer_dom_dca(dca, succ, node);
2067 static void set_projs_block(ir_node *node, ir_node *block)
2071 for (i = get_irn_n_outs(node) - 1; i >= 0; --i) {
2072 ir_node *succ = get_irn_out(node, i);
2074 assert(is_Proj(succ));
2076 if(get_irn_mode(succ) == mode_T) {
2077 set_projs_block(succ, block);
2079 set_nodes_block(succ, block);
2084 * Find the latest legal block for N and place N into the
2085 * `optimal' Block between the latest and earliest legal block.
2086 * The `optimal' block is the dominance-deepest block of those
2087 * with the least loop-nesting-depth. This places N out of as many
2088 * loops as possible and then makes it as control dependent as
2091 * @param n the node to be placed
2092 * @param worklist a worklist, all successors of non-floating nodes are
2095 static void place_floats_late(ir_node *n, pdeq *worklist) {
2099 assert(irn_not_visited(n)); /* no multiple placement */
2101 mark_irn_visited(n);
2103 /* no need to place block nodes, control nodes are already placed. */
2106 (get_irn_mode(n) != mode_X)) {
2107 /* Remember the early_blk placement of this block to move it
2108 out of loop no further than the early_blk placement. */
2109 early_blk = get_nodes_block(n);
2112 * BEWARE: Here we also get code, that is live, but
2113 * was in a dead block. If the node is life, but because
2114 * of CSE in a dead block, we still might need it.
2117 /* Assure that our users are all placed, except the Phi-nodes.
2118 --- Each data flow cycle contains at least one Phi-node. We
2119 have to break the `user has to be placed before the
2120 producer' dependence cycle and the Phi-nodes are the
2121 place to do so, because we need to base our placement on the
2122 final region of our users, which is OK with Phi-nodes, as they
2123 are op_pin_state_pinned, and they never have to be placed after a
2124 producer of one of their inputs in the same block anyway. */
2125 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2126 ir_node *succ = get_irn_out(n, i);
2127 if (irn_not_visited(succ) && !is_Phi(succ))
2128 place_floats_late(succ, worklist);
2131 if (! is_Block_dead(early_blk)) {
2132 /* do only move things that where not dead */
2133 ir_op *op = get_irn_op(n);
2135 /* We have to determine the final block of this node... except for
2136 constants and Projs */
2137 if ((get_irn_pinned(n) == op_pin_state_floats) &&
2139 (op != op_SymConst) &&
2142 /* deepest common ancestor in the dominator tree of all nodes'
2143 blocks depending on us; our final placement has to dominate
2145 ir_node *dca = get_deepest_common_ancestor(n, NULL);
2147 set_nodes_block(n, dca);
2148 move_out_of_loops(n, early_blk);
2149 if(get_irn_mode(n) == mode_T) {
2150 set_projs_block(n, get_nodes_block(n));
2157 /* Add successors of all non-floating nodes on list. (Those of floating
2158 nodes are placed already and therefore are marked.) */
2159 for (i = 0; i < get_irn_n_outs(n); i++) {
2160 ir_node *succ = get_irn_out(n, i);
2161 if (irn_not_visited(get_irn_out(n, i))) {
2162 pdeq_putr(worklist, succ);
2168 * Place floating nodes on the given worklist as late as possible using
2169 * the dominance tree.
2171 * @param worklist the worklist containing the nodes to place
2173 static void place_late(waitq *worklist) {
2175 inc_irg_visited(current_ir_graph);
2177 /* This fills the worklist initially. */
2178 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2180 /* And now empty the worklist again... */
2181 while (!waitq_empty(worklist)) {
2182 ir_node *n = waitq_get(worklist);
2183 if (irn_not_visited(n))
2184 place_floats_late(n, worklist);
2188 /* Code Placement. */
2189 void place_code(ir_graph *irg) {
2191 ir_graph *rem = current_ir_graph;
2193 current_ir_graph = irg;
2195 /* Handle graph state */
2196 assert(get_irg_phase_state(irg) != phase_building);
2199 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2200 free_loop_information(irg);
2201 construct_cf_backedges(irg);
2204 /* Place all floating nodes as early as possible. This guarantees
2205 a legal code placement. */
2206 worklist = new_waitq();
2207 place_early(worklist);
2209 /* place_early() invalidates the outs, place_late needs them. */
2210 compute_irg_outs(irg);
2212 /* Now move the nodes down in the dominator tree. This reduces the
2213 unnecessary executions of the node. */
2214 place_late(worklist);
2216 set_irg_outs_inconsistent(current_ir_graph);
2217 set_irg_loopinfo_inconsistent(current_ir_graph);
2218 del_waitq(worklist);
2219 current_ir_graph = rem;
2222 typedef struct cf_env {
2223 char changed; /**< flag indicates that the cf graphs has changed. */
2227 * Called by walker of remove_critical_cf_edges().
2229 * Place an empty block to an edge between a blocks of multiple
2230 * predecessors and a block of multiple successors.
2233 * @param env Environment of walker.
2235 static void walk_critical_cf_edges(ir_node *n, void *env) {
2237 ir_node *pre, *block, *jmp;
2239 ir_graph *irg = get_irn_irg(n);
2241 /* Block has multiple predecessors */
2242 arity = get_irn_arity(n);
2244 if (n == get_irg_end_block(irg))
2245 return; /* No use to add a block here. */
2247 for (i = 0; i < arity; ++i) {
2250 pre = get_irn_n(n, i);
2251 cfop = get_irn_op(skip_Proj(pre));
2253 if (is_op_fragile(cfop)) {
2254 if (cfop != op_Raise)
2258 /* we don't want place nodes in the start block, so handle it like forking */
2259 if (is_op_forking(cfop) || cfop == op_Start) {
2260 /* Predecessor has multiple successors. Insert new control flow edge edges. */
2262 /* set predecessor of new block */
2263 block = new_r_Block(irg, 1, &pre);
2264 /* insert new jmp node to new block */
2265 jmp = new_r_Jmp(irg, block);
2266 /* set successor of new block */
2267 set_irn_n(n, i, jmp);
2269 } /* predecessor has multiple successors */
2270 } /* for all predecessors */
2271 } /* n is a multi-entry block */
2274 void remove_critical_cf_edges(ir_graph *irg) {
2279 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &env);
2281 /* control flow changed */
2282 set_irg_outs_inconsistent(irg);
2283 set_irg_extblk_inconsistent(irg);
2284 set_irg_doms_inconsistent(irg);
2285 set_irg_loopinfo_inconsistent(irg);