2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Optimizations for a whole ir graph, i.e., a procedure.
23 * @author Christian Schaefer, Goetz Lindenmaier, Sebastian Felis,
34 #include "irgraph_t.h"
37 #include "iroptimize.h"
47 #include "pdeq.h" /* Fuer code placement */
52 #include "irbackedge_t.h"
59 #include "iredges_t.h"
62 /*------------------------------------------------------------------*/
63 /* apply optimizations of iropt to all nodes. */
64 /*------------------------------------------------------------------*/
67 * A wrapper around optimize_inplace_2() to be called from a walker.
69 static void optimize_in_place_wrapper (ir_node *n, void *env) {
70 ir_node *optimized = optimize_in_place_2(n);
74 exchange (n, optimized);
79 * Do local optimizations for a node.
81 * @param n the IR-node where to start. Typically the End node
84 * @note current_ir_graph must be set
86 static INLINE void do_local_optimize(ir_node *n) {
87 /* Handle graph state */
88 assert(get_irg_phase_state(current_ir_graph) != phase_building);
90 if (get_opt_global_cse())
91 set_irg_pinned(current_ir_graph, op_pin_state_floats);
92 set_irg_outs_inconsistent(current_ir_graph);
93 set_irg_doms_inconsistent(current_ir_graph);
94 set_irg_loopinfo_inconsistent(current_ir_graph);
96 /* Clean the value_table in irg for the CSE. */
97 del_identities(current_ir_graph->value_table);
98 current_ir_graph->value_table = new_identities();
100 /* walk over the graph */
101 irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
104 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
105 void local_optimize_node(ir_node *n) {
106 ir_graph *rem = current_ir_graph;
107 current_ir_graph = get_irn_irg(n);
109 do_local_optimize(n);
111 current_ir_graph = rem;
115 * Block-Walker: uses dominance depth to mark dead blocks.
117 static void kill_dead_blocks(ir_node *block, void *env) {
120 if (get_Block_dom_depth(block) < 0) {
122 * Note that the new dominance code correctly handles
123 * the End block, i.e. it is always reachable from Start
125 set_Block_dead(block);
129 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */
130 void local_optimize_graph(ir_graph *irg) {
131 ir_graph *rem = current_ir_graph;
132 current_ir_graph = irg;
134 if (get_irg_dom_state(irg) == dom_consistent)
135 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
137 do_local_optimize(get_irg_end(irg));
139 current_ir_graph = rem;
143 * Enqueue all users of a node to a wait queue.
144 * Handles mode_T nodes.
146 static void enqueue_users(ir_node *n, pdeq *waitq) {
147 const ir_edge_t *edge;
149 foreach_out_edge(n, edge) {
150 ir_node *succ = get_edge_src_irn(edge);
152 if (get_irn_link(succ) != waitq) {
153 pdeq_putr(waitq, succ);
154 set_irn_link(succ, waitq);
156 if (get_irn_mode(succ) == mode_T) {
157 /* A mode_T node has Proj's. Because most optimizations
158 run on the Proj's we have to enqueue them also. */
159 enqueue_users(succ, waitq);
165 * Data flow optimization walker.
166 * Optimizes all nodes and enqueue it's users
169 static void opt_walker(ir_node *n, void *env) {
173 optimized = optimize_in_place_2(n);
174 set_irn_link(optimized, NULL);
176 if (optimized != n) {
177 enqueue_users(n, waitq);
178 exchange(n, optimized);
182 /* Applies local optimizations to all nodes in the graph until fixpoint. */
183 void optimize_graph_df(ir_graph *irg) {
184 pdeq *waitq = new_pdeq();
185 ir_graph *rem = current_ir_graph;
189 current_ir_graph = irg;
191 state = edges_assure(irg);
193 if (get_opt_global_cse())
194 set_irg_pinned(current_ir_graph, op_pin_state_floats);
196 /* Clean the value_table in irg for the CSE. */
197 del_identities(irg->value_table);
198 irg->value_table = new_identities();
200 if (get_irg_dom_state(irg) == dom_consistent)
201 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
203 /* invalidate info */
204 set_irg_outs_inconsistent(irg);
205 set_irg_doms_inconsistent(irg);
206 set_irg_loopinfo_inconsistent(irg);
208 set_using_irn_link(irg);
210 /* walk over the graph, but don't touch keep-alives */
211 irg_walk(get_irg_end_block(irg), NULL, opt_walker, waitq);
213 end = get_irg_end(irg);
215 /* optimize keep-alives by removing superfluous ones */
216 for (i = get_End_n_keepalives(end) - 1; i >= 0; --i) {
217 ir_node *ka = get_End_keepalive(end, i);
219 if (irn_visited(ka) && !is_irn_keep(ka)) {
220 /* this node can be regularly visited, no need to keep it */
221 set_End_keepalive(end, i, get_irg_bad(irg));
224 /* now walk again and visit all not yet visited nodes */
225 set_irg_visited(current_ir_graph, get_irg_visited(irg) - 1);
226 irg_walk(get_irg_end(irg), NULL, opt_walker, waitq);
228 /* finish the wait queue */
229 while (! pdeq_empty(waitq)) {
230 ir_node *n = pdeq_getl(waitq);
232 opt_walker(n, waitq);
237 clear_using_irn_link(irg);
240 edges_deactivate(irg);
242 current_ir_graph = rem;
246 /*------------------------------------------------------------------*/
247 /* Routines for dead node elimination / copying garbage collection */
248 /* of the obstack. */
249 /*------------------------------------------------------------------*/
252 * Remember the new node in the old node by using a field all nodes have.
254 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
257 * Get this new node, before the old node is forgotten.
259 #define get_new_node(oldn) get_irn_link(oldn)
262 * Check if a new node was set.
264 #define has_new_node(n) (get_new_node(n) != NULL)
267 * We use the block_visited flag to mark that we have computed the
268 * number of useful predecessors for this block.
269 * Further we encode the new arity in this flag in the old blocks.
270 * Remembering the arity is useful, as it saves a lot of pointer
271 * accesses. This function is called for all Phi and Block nodes
275 compute_new_arity(ir_node *b) {
276 int i, res, irn_arity;
279 irg_v = get_irg_block_visited(current_ir_graph);
280 block_v = get_Block_block_visited(b);
281 if (block_v >= irg_v) {
282 /* we computed the number of preds for this block and saved it in the
284 return block_v - irg_v;
286 /* compute the number of good predecessors */
287 res = irn_arity = get_irn_arity(b);
288 for (i = 0; i < irn_arity; i++)
289 if (is_Bad(get_irn_n(b, i))) res--;
290 /* save it in the flag. */
291 set_Block_block_visited(b, irg_v + res);
297 * Copies the node to the new obstack. The Ins of the new node point to
298 * the predecessors on the old obstack. For block/phi nodes not all
299 * predecessors might be copied. n->link points to the new node.
300 * For Phi and Block nodes the function allocates in-arrays with an arity
301 * only for useful predecessors. The arity is determined by counting
302 * the non-bad predecessors of the block.
304 * @param n The node to be copied
305 * @param env if non-NULL, the node number attribute will be copied to the new node
307 * Note: Also used for loop unrolling.
309 static void copy_node(ir_node *n, void *env) {
312 ir_op *op = get_irn_op(n);
315 /* The end node looses it's flexible in array. This doesn't matter,
316 as dead node elimination builds End by hand, inlineing doesn't use
318 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
321 /* node copied already */
323 } else if (op == op_Block) {
325 new_arity = compute_new_arity(n);
326 n->attr.block.graph_arr = NULL;
328 block = get_nodes_block(n);
330 new_arity = compute_new_arity(block);
332 new_arity = get_irn_arity(n);
335 nn = new_ir_node(get_irn_dbg_info(n),
342 /* Copy the attributes. These might point to additional data. If this
343 was allocated on the old obstack the pointers now are dangling. This
344 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
345 if (op == op_Block) {
346 /* we cannot allow blocks WITHOUT macroblock input */
347 set_irn_n(nn, -1, get_irn_n(n, -1));
349 copy_node_attr(n, nn);
353 int copy_node_nr = env != NULL;
355 /* for easier debugging, we want to copy the node numbers too */
356 nn->node_nr = n->node_nr;
362 hook_dead_node_elim_subst(current_ir_graph, n, nn);
366 * Copies new predecessors of old node to new node remembered in link.
367 * Spare the Bad predecessors of Phi and Block nodes.
369 static void copy_preds(ir_node *n, void *env) {
374 nn = get_new_node(n);
377 /* copy the macro block header */
378 ir_node *mbh = get_Block_MacroBlock(n);
381 /* this block is a macroblock header */
382 set_irn_n(nn, -1, nn);
384 /* get the macro block header */
385 ir_node *nmbh = get_new_node(mbh);
386 assert(nmbh != NULL);
387 set_irn_n(nn, -1, nmbh);
390 /* Don't copy Bad nodes. */
392 irn_arity = get_irn_arity(n);
393 for (i = 0; i < irn_arity; i++) {
394 if (! is_Bad(get_irn_n(n, i))) {
395 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
396 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
400 /* repair the block visited flag from above misuse. Repair it in both
401 graphs so that the old one can still be used. */
402 set_Block_block_visited(nn, 0);
403 set_Block_block_visited(n, 0);
404 /* Local optimization could not merge two subsequent blocks if
405 in array contained Bads. Now it's possible.
406 We don't call optimize_in_place as it requires
407 that the fields in ir_graph are set properly. */
408 if ((get_opt_control_flow_straightening()) &&
409 (get_Block_n_cfgpreds(nn) == 1) &&
410 is_Jmp(get_Block_cfgpred(nn, 0))) {
411 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
413 /* Jmp jumps into the block it is in -- deal self cycle. */
414 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
415 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
420 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
421 /* Don't copy node if corresponding predecessor in block is Bad.
422 The Block itself should not be Bad. */
423 block = get_nodes_block(n);
424 set_irn_n(nn, -1, get_new_node(block));
426 irn_arity = get_irn_arity(n);
427 for (i = 0; i < irn_arity; i++) {
428 if (! is_Bad(get_irn_n(block, i))) {
429 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
430 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
434 /* If the pre walker reached this Phi after the post walker visited the
435 block block_visited is > 0. */
436 set_Block_block_visited(get_nodes_block(n), 0);
437 /* Compacting the Phi's ins might generate Phis with only one
439 if (get_irn_arity(nn) == 1)
440 exchange(nn, get_irn_n(nn, 0));
442 irn_arity = get_irn_arity(n);
443 for (i = -1; i < irn_arity; i++)
444 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
446 /* Now the new node is complete. We can add it to the hash table for CSE.
447 @@@ inlining aborts if we identify End. Why? */
449 add_identities(current_ir_graph->value_table, nn);
453 * Copies the graph recursively, compacts the keep-alives of the end node.
455 * @param irg the graph to be copied
456 * @param copy_node_nr If non-zero, the node number will be copied
458 static void copy_graph(ir_graph *irg, int copy_node_nr) {
459 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
460 ir_node *ka; /* keep alive */
464 /* Some nodes must be copied by hand, sigh */
465 vfl = get_irg_visited(irg);
466 set_irg_visited(irg, vfl + 1);
468 oe = get_irg_end(irg);
469 mark_irn_visited(oe);
470 /* copy the end node by hand, allocate dynamic in array! */
471 ne = new_ir_node(get_irn_dbg_info(oe),
478 /* Copy the attributes. Well, there might be some in the future... */
479 copy_node_attr(oe, ne);
480 set_new_node(oe, ne);
482 /* copy the Bad node */
483 ob = get_irg_bad(irg);
484 mark_irn_visited(ob);
485 nb = new_ir_node(get_irn_dbg_info(ob),
492 copy_node_attr(ob, nb);
493 set_new_node(ob, nb);
495 /* copy the NoMem node */
496 om = get_irg_no_mem(irg);
497 mark_irn_visited(om);
498 nm = new_ir_node(get_irn_dbg_info(om),
505 copy_node_attr(om, nm);
506 set_new_node(om, nm);
508 /* copy the live nodes */
509 set_irg_visited(irg, vfl);
510 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
512 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
514 /* visit the anchors as well */
515 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
516 ir_node *n = get_irg_anchor(irg, i);
518 if (n && (get_irn_visited(n) <= vfl)) {
519 set_irg_visited(irg, vfl);
520 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
524 /* copy_preds for the end node ... */
525 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
527 /*- ... and now the keep alives. -*/
528 /* First pick the not marked block nodes and walk them. We must pick these
529 first as else we will oversee blocks reachable from Phis. */
530 irn_arity = get_End_n_keepalives(oe);
531 for (i = 0; i < irn_arity; i++) {
532 ka = get_End_keepalive(oe, i);
534 if (get_irn_visited(ka) <= vfl) {
535 /* We must keep the block alive and copy everything reachable */
536 set_irg_visited(irg, vfl);
537 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
539 add_End_keepalive(ne, get_new_node(ka));
543 /* Now pick other nodes. Here we will keep all! */
544 irn_arity = get_End_n_keepalives(oe);
545 for (i = 0; i < irn_arity; i++) {
546 ka = get_End_keepalive(oe, i);
548 if (get_irn_visited(ka) <= vfl) {
549 /* We didn't copy the node yet. */
550 set_irg_visited(irg, vfl);
551 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
553 add_End_keepalive(ne, get_new_node(ka));
557 /* start block sometimes only reached after keep alives */
558 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
559 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
563 * Copies the graph reachable from current_ir_graph->end to the obstack
564 * in current_ir_graph and fixes the environment.
565 * Then fixes the fields in current_ir_graph containing nodes of the
568 * @param copy_node_nr If non-zero, the node number will be copied
571 copy_graph_env(int copy_node_nr) {
572 ir_graph *irg = current_ir_graph;
573 ir_node *old_end, *new_anchor;
576 /* remove end_except and end_reg nodes */
577 old_end = get_irg_end(irg);
578 set_irg_end_except (irg, old_end);
579 set_irg_end_reg (irg, old_end);
581 /* Not all nodes remembered in irg might be reachable
582 from the end node. Assure their link is set to NULL, so that
583 we can test whether new nodes have been computed. */
584 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
585 ir_node *n = get_irg_anchor(irg, i);
587 set_new_node(n, NULL);
589 /* we use the block walk flag for removing Bads from Blocks ins. */
590 inc_irg_block_visited(irg);
593 copy_graph(irg, copy_node_nr);
596 old_end = get_irg_end(irg);
597 new_anchor = new_Anchor(irg);
599 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
600 ir_node *n = get_irg_anchor(irg, i);
602 set_irn_n(new_anchor, i, get_new_node(n));
605 irg->anchor = new_anchor;
607 /* ensure the new anchor is placed in the endblock */
608 set_irn_n(new_anchor, -1, get_irg_end_block(irg));
612 * Copies all reachable nodes to a new obstack. Removes bad inputs
613 * from block nodes and the corresponding inputs from Phi nodes.
614 * Merges single exit blocks with single entry blocks and removes
616 * Adds all new nodes to a new hash table for CSE. Does not
617 * perform CSE, so the hash table might contain common subexpressions.
619 void dead_node_elimination(ir_graph *irg) {
621 #ifdef INTERPROCEDURAL_VIEW
622 int rem_ipview = get_interprocedural_view();
624 struct obstack *graveyard_obst = NULL;
625 struct obstack *rebirth_obst = NULL;
626 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
628 /* inform statistics that we started a dead-node elimination run */
629 hook_dead_node_elim(irg, 1);
631 /* Remember external state of current_ir_graph. */
632 rem = current_ir_graph;
633 current_ir_graph = irg;
634 #ifdef INTERPROCEDURAL_VIEW
635 set_interprocedural_view(0);
638 assert(get_irg_phase_state(irg) != phase_building);
640 /* Handle graph state */
641 free_callee_info(irg);
645 /* @@@ so far we loose loops when copying */
646 free_loop_information(irg);
648 set_irg_doms_inconsistent(irg);
650 /* A quiet place, where the old obstack can rest in peace,
651 until it will be cremated. */
652 graveyard_obst = irg->obst;
654 /* A new obstack, where the reachable nodes will be copied to. */
655 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
656 irg->obst = rebirth_obst;
657 obstack_init(irg->obst);
658 irg->last_node_idx = 0;
660 /* We also need a new value table for CSE */
661 del_identities(irg->value_table);
662 irg->value_table = new_identities();
664 /* Copy the graph from the old to the new obstack */
665 copy_graph_env(/*copy_node_nr=*/1);
667 /* Free memory from old unoptimized obstack */
668 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
669 xfree(graveyard_obst); /* ... then free it. */
671 /* inform statistics that the run is over */
672 hook_dead_node_elim(irg, 0);
674 current_ir_graph = rem;
675 #ifdef INTERPROCEDURAL_VIEW
676 set_interprocedural_view(rem_ipview);
681 * Relink bad predecessors of a block and store the old in array to the
682 * link field. This function is called by relink_bad_predecessors().
683 * The array of link field starts with the block operand at position 0.
684 * If block has bad predecessors, create a new in array without bad preds.
685 * Otherwise let in array untouched.
687 static void relink_bad_block_predecessors(ir_node *n, void *env) {
688 ir_node **new_in, *irn;
689 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
692 /* if link field of block is NULL, look for bad predecessors otherwise
693 this is already done */
694 if (is_Block(n) && get_irn_link(n) == NULL) {
695 /* save old predecessors in link field (position 0 is the block operand)*/
696 set_irn_link(n, get_irn_in(n));
698 /* count predecessors without bad nodes */
699 old_irn_arity = get_irn_arity(n);
700 for (i = 0; i < old_irn_arity; i++)
701 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
703 /* arity changing: set new predecessors without bad nodes */
704 if (new_irn_arity < old_irn_arity) {
705 /* Get new predecessor array. We do not resize the array, as we must
706 keep the old one to update Phis. */
707 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
709 /* set new predecessors in array */
712 for (i = 0; i < old_irn_arity; i++) {
713 irn = get_irn_n(n, i);
715 new_in[new_irn_n] = irn;
716 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
720 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
721 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
723 } /* ir node has bad predecessors */
724 } /* Block is not relinked */
728 * Relinks Bad predecessors from Blocks and Phis called by walker
729 * remove_bad_predecesors(). If n is a Block, call
730 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
731 * function of Phi's Block. If this block has bad predecessors, relink preds
734 static void relink_bad_predecessors(ir_node *n, void *env) {
735 ir_node *block, **old_in;
736 int i, old_irn_arity, new_irn_arity;
738 /* relink bad predecessors of a block */
740 relink_bad_block_predecessors(n, env);
742 /* If Phi node relink its block and its predecessors */
744 /* Relink predecessors of phi's block */
745 block = get_nodes_block(n);
746 if (get_irn_link(block) == NULL)
747 relink_bad_block_predecessors(block, env);
749 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
750 old_irn_arity = ARR_LEN(old_in);
752 /* Relink Phi predecessors if count of predecessors changed */
753 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
754 /* set new predecessors in array
755 n->in[0] remains the same block */
757 for(i = 1; i < old_irn_arity; i++)
758 if (!is_Bad((ir_node *)old_in[i])) {
759 n->in[new_irn_arity] = n->in[i];
760 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
764 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
765 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
767 } /* n is a Phi node */
771 * Removes Bad Bad predecessors from Blocks and the corresponding
772 * inputs to Phi nodes as in dead_node_elimination but without
774 * On walking up set the link field to NULL, on walking down call
775 * relink_bad_predecessors() (This function stores the old in array
776 * to the link field and sets a new in array if arity of predecessors
779 void remove_bad_predecessors(ir_graph *irg) {
780 panic("Fix backedge handling first");
781 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
788 __)|_| | \_/ | \_/(/_ |_/\__|__
790 The following stuff implements a facility that automatically patches
791 registered ir_node pointers to the new node when a dead node elimination occurs.
794 struct _survive_dce_t {
798 hook_entry_t dead_node_elim;
799 hook_entry_t dead_node_elim_subst;
802 typedef struct _survive_dce_list_t {
803 struct _survive_dce_list_t *next;
805 } survive_dce_list_t;
807 static void dead_node_hook(void *context, ir_graph *irg, int start) {
808 survive_dce_t *sd = context;
811 /* Create a new map before the dead node elimination is performed. */
813 sd->new_places = pmap_create_ex(pmap_count(sd->places));
815 /* Patch back all nodes if dead node elimination is over and something is to be done. */
816 pmap_destroy(sd->places);
817 sd->places = sd->new_places;
818 sd->new_places = NULL;
823 * Hook called when dead node elimination replaces old by nw.
825 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
826 survive_dce_t *sd = context;
827 survive_dce_list_t *list = pmap_get(sd->places, old);
830 /* If the node is to be patched back, write the new address to all registered locations. */
832 survive_dce_list_t *p;
834 for (p = list; p; p = p->next)
837 pmap_insert(sd->new_places, nw, list);
842 * Make a new Survive DCE environment.
844 survive_dce_t *new_survive_dce(void) {
845 survive_dce_t *res = xmalloc(sizeof(res[0]));
846 obstack_init(&res->obst);
847 res->places = pmap_create();
848 res->new_places = NULL;
850 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
851 res->dead_node_elim.context = res;
852 res->dead_node_elim.next = NULL;
854 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
855 res->dead_node_elim_subst.context = res;
856 res->dead_node_elim_subst.next = NULL;
858 #ifndef FIRM_ENABLE_HOOKS
859 assert(0 && "need hooks enabled");
862 register_hook(hook_dead_node_elim, &res->dead_node_elim);
863 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
868 * Free a Survive DCE environment.
870 void free_survive_dce(survive_dce_t *sd) {
871 obstack_free(&sd->obst, NULL);
872 pmap_destroy(sd->places);
873 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
874 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
879 * Register a node pointer to be patched upon DCE.
880 * When DCE occurs, the node pointer specified by @p place will be
881 * patched to the new address of the node it is pointing to.
883 * @param sd The Survive DCE environment.
884 * @param place The address of the node pointer.
886 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
887 if (*place != NULL) {
888 ir_node *irn = *place;
889 survive_dce_list_t *curr = pmap_get(sd->places, irn);
890 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
895 pmap_insert(sd->places, irn, nw);
899 /*--------------------------------------------------------------------*/
900 /* Functionality for inlining */
901 /*--------------------------------------------------------------------*/
904 * Copy node for inlineing. Updates attributes that change when
905 * inlineing but not for dead node elimination.
907 * Copies the node by calling copy_node() and then updates the entity if
908 * it's a local one. env must be a pointer of the frame type of the
909 * inlined procedure. The new entities must be in the link field of
913 copy_node_inline(ir_node *n, void *env) {
915 ir_type *frame_tp = (ir_type *)env;
919 nn = get_new_node (n);
921 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
922 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
924 } else if (is_Block(n)) {
925 nn = get_new_node (n);
926 nn->attr.block.irg = current_ir_graph;
931 * Walker: checks if P_value_arg_base is used.
933 static void find_addr(ir_node *node, void *env) {
934 int *allow_inline = env;
936 is_Start(get_Proj_pred(node)) &&
937 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
943 * Check if we can inline a given call.
944 * Currently, we cannot inline two cases:
945 * - call with compound arguments
946 * - graphs that take the address of a parameter
948 * check these conditions here
950 static int can_inline(ir_node *call, ir_graph *called_graph) {
951 ir_type *call_type = get_Call_type(call);
952 int params, ress, i, res;
953 assert(is_Method_type(call_type));
955 params = get_method_n_params(call_type);
956 ress = get_method_n_ress(call_type);
958 /* check parameters for compound arguments */
959 for (i = 0; i < params; ++i) {
960 ir_type *p_type = get_method_param_type(call_type, i);
962 if (is_compound_type(p_type))
966 /* check results for compound arguments */
967 for (i = 0; i < ress; ++i) {
968 ir_type *r_type = get_method_res_type(call_type, i);
970 if (is_compound_type(r_type))
975 irg_walk_graph(called_graph, find_addr, NULL, &res);
981 exc_handler = 0, /**< There is a handler. */
982 exc_to_end = 1, /**< Branches to End. */
983 exc_no_handler = 2 /**< Exception handling not represented. */
986 /* Inlines a method at the given call site. */
987 int inline_method(ir_node *call, ir_graph *called_graph) {
989 ir_node *post_call, *post_bl;
990 ir_node *in[pn_Start_max];
991 ir_node *end, *end_bl;
995 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
996 enum exc_mode exc_handling;
997 ir_type *called_frame;
998 irg_inline_property prop = get_irg_inline_property(called_graph);
1000 if ( (prop < irg_inline_forced) || (prop == irg_inline_forbidden))
1003 /* Do not inline variadic functions. */
1004 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
1007 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
1008 get_method_n_params(get_Call_type(call)));
1011 * currently, we cannot inline two cases:
1012 * - call with compound arguments
1013 * - graphs that take the address of a parameter
1015 if (! can_inline(call, called_graph))
1018 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
1019 rem_opt = get_opt_optimize();
1022 /* Handle graph state */
1023 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1024 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
1025 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
1026 set_irg_outs_inconsistent(current_ir_graph);
1027 set_irg_extblk_inconsistent(current_ir_graph);
1028 set_irg_doms_inconsistent(current_ir_graph);
1029 set_irg_loopinfo_inconsistent(current_ir_graph);
1030 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
1032 /* -- Check preconditions -- */
1033 assert(is_Call(call));
1034 /* @@@ does not work for InterfaceIII.java after cgana
1035 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
1036 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
1037 get_Call_type(call)));
1039 if (called_graph == current_ir_graph) {
1040 set_optimize(rem_opt);
1044 /* here we know we WILL inline, so inform the statistics */
1045 hook_inline(call, called_graph);
1047 /* -- Decide how to handle exception control flow: Is there a handler
1048 for the Call node, or do we branch directly to End on an exception?
1050 0 There is a handler.
1052 2 Exception handling not represented in Firm. -- */
1054 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
1055 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
1056 long proj_nr = get_Proj_proj(proj);
1057 if (proj_nr == pn_Call_X_except) Xproj = proj;
1058 if (proj_nr == pn_Call_M_except) Mproj = proj;
1060 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
1061 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
1062 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
1066 the procedure and later replaces the Start node of the called graph.
1067 Post_call is the old Call node and collects the results of the called
1068 graph. Both will end up being a tuple. -- */
1069 post_bl = get_nodes_block(call);
1070 set_irg_current_block(current_ir_graph, post_bl);
1071 /* XxMxPxPxPxT of Start + parameter of Call */
1072 in[pn_Start_X_initial_exec] = new_Jmp();
1073 in[pn_Start_M] = get_Call_mem(call);
1074 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1075 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1076 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1077 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1078 /* in[pn_Start_P_value_arg_base] = ??? */
1079 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1080 pre_call = new_Tuple(pn_Start_max - 1, in);
1084 The new block gets the ins of the old block, pre_call and all its
1085 predecessors and all Phi nodes. -- */
1086 part_block(pre_call);
1088 /* -- Prepare state for dead node elimination -- */
1089 /* Visited flags in calling irg must be >= flag in called irg.
1090 Else walker and arity computation will not work. */
1091 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1092 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1093 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1094 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1095 /* Set pre_call as new Start node in link field of the start node of
1096 calling graph and pre_calls block as new block for the start block
1098 Further mark these nodes so that they are not visited by the
1100 set_irn_link(get_irg_start(called_graph), pre_call);
1101 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1102 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1103 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1104 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1105 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1107 /* Initialize for compaction of in arrays */
1108 inc_irg_block_visited(current_ir_graph);
1110 /* -- Replicate local entities of the called_graph -- */
1111 /* copy the entities. */
1112 called_frame = get_irg_frame_type(called_graph);
1113 for (i = 0; i < get_class_n_members(called_frame); i++) {
1114 ir_entity *new_ent, *old_ent;
1115 old_ent = get_class_member(called_frame, i);
1116 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1117 set_entity_link(old_ent, new_ent);
1120 /* visited is > than that of called graph. With this trick visited will
1121 remain unchanged so that an outer walker, e.g., searching the call nodes
1122 to inline, calling this inline will not visit the inlined nodes. */
1123 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1125 /* -- Performing dead node elimination inlines the graph -- */
1126 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1128 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1129 get_irg_frame_type(called_graph));
1131 /* Repair called_graph */
1132 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1133 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1134 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1136 /* -- Merge the end of the inlined procedure with the call site -- */
1137 /* We will turn the old Call node into a Tuple with the following
1140 0: Phi of all Memories of Return statements.
1141 1: Jmp from new Block that merges the control flow from all exception
1142 predecessors of the old end block.
1143 2: Tuple of all arguments.
1144 3: Phi of Exception memories.
1145 In case the old Call directly branches to End on an exception we don't
1146 need the block merging all exceptions nor the Phi of the exception
1150 /* -- Precompute some values -- */
1151 end_bl = get_new_node(get_irg_end_block(called_graph));
1152 end = get_new_node(get_irg_end(called_graph));
1153 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1154 n_res = get_method_n_ress(get_Call_type(call));
1156 res_pred = xmalloc(n_res * sizeof(*res_pred));
1157 cf_pred = xmalloc(arity * sizeof(*res_pred));
1159 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1161 /* -- archive keepalives -- */
1162 irn_arity = get_irn_arity(end);
1163 for (i = 0; i < irn_arity; i++) {
1164 ir_node *ka = get_End_keepalive(end, i);
1166 add_End_keepalive(get_irg_end(current_ir_graph), ka);
1169 /* The new end node will die. We need not free as the in array is on the obstack:
1170 copy_node() only generated 'D' arrays. */
1172 /* -- Replace Return nodes by Jump nodes. -- */
1174 for (i = 0; i < arity; i++) {
1176 ret = get_irn_n(end_bl, i);
1177 if (is_Return(ret)) {
1178 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1182 set_irn_in(post_bl, n_ret, cf_pred);
1184 /* -- Build a Tuple for all results of the method.
1185 Add Phi node if there was more than one Return. -- */
1186 turn_into_tuple(post_call, pn_Call_max);
1187 /* First the Memory-Phi */
1189 for (i = 0; i < arity; i++) {
1190 ret = get_irn_n(end_bl, i);
1191 if (is_Return(ret)) {
1192 cf_pred[n_ret] = get_Return_mem(ret);
1196 phi = new_Phi(n_ret, cf_pred, mode_M);
1197 set_Tuple_pred(call, pn_Call_M_regular, phi);
1198 /* Conserve Phi-list for further inlinings -- but might be optimized */
1199 if (get_nodes_block(phi) == post_bl) {
1200 set_irn_link(phi, get_irn_link(post_bl));
1201 set_irn_link(post_bl, phi);
1203 /* Now the real results */
1205 for (j = 0; j < n_res; j++) {
1207 for (i = 0; i < arity; i++) {
1208 ret = get_irn_n(end_bl, i);
1209 if (is_Return(ret)) {
1210 cf_pred[n_ret] = get_Return_res(ret, j);
1215 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1219 /* Conserve Phi-list for further inlinings -- but might be optimized */
1220 if (get_nodes_block(phi) == post_bl) {
1221 set_irn_link(phi, get_irn_link(post_bl));
1222 set_irn_link(post_bl, phi);
1225 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1227 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1229 /* handle the regular call */
1230 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1232 /* For now, we cannot inline calls with value_base */
1233 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1235 /* Finally the exception control flow.
1236 We have two (three) possible situations:
1237 First if the Call branches to an exception handler: We need to add a Phi node to
1238 collect the memory containing the exception objects. Further we need
1239 to add another block to get a correct representation of this Phi. To
1240 this block we add a Jmp that resolves into the X output of the Call
1241 when the Call is turned into a tuple.
1242 Second the Call branches to End, the exception is not handled. Just
1243 add all inlined exception branches to the End node.
1244 Third: there is no Exception edge at all. Handle as case two. */
1245 if (exc_handling == exc_handler) {
1247 for (i = 0; i < arity; i++) {
1249 ret = get_irn_n(end_bl, i);
1250 irn = skip_Proj(ret);
1251 if (is_fragile_op(irn) || is_Raise(irn)) {
1252 cf_pred[n_exc] = ret;
1257 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1258 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1259 /* The Phi for the memories with the exception objects */
1261 for (i = 0; i < arity; i++) {
1263 ret = skip_Proj(get_irn_n(end_bl, i));
1265 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1267 } else if (is_fragile_op(ret)) {
1268 /* We rely that all cfops have the memory output at the same position. */
1269 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1271 } else if (is_Raise(ret)) {
1272 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1276 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1278 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1279 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1282 ir_node *main_end_bl;
1283 int main_end_bl_arity;
1284 ir_node **end_preds;
1286 /* assert(exc_handling == 1 || no exceptions. ) */
1288 for (i = 0; i < arity; i++) {
1289 ir_node *ret = get_irn_n(end_bl, i);
1290 ir_node *irn = skip_Proj(ret);
1292 if (is_fragile_op(irn) || is_Raise(irn)) {
1293 cf_pred[n_exc] = ret;
1297 main_end_bl = get_irg_end_block(current_ir_graph);
1298 main_end_bl_arity = get_irn_arity(main_end_bl);
1299 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1301 for (i = 0; i < main_end_bl_arity; ++i)
1302 end_preds[i] = get_irn_n(main_end_bl, i);
1303 for (i = 0; i < n_exc; ++i)
1304 end_preds[main_end_bl_arity + i] = cf_pred[i];
1305 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1306 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1307 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1313 /* -- Turn CSE back on. -- */
1314 set_optimize(rem_opt);
1319 /********************************************************************/
1320 /* Apply inlineing to small methods. */
1321 /********************************************************************/
1323 /** Represents a possible inlinable call in a graph. */
1324 typedef struct _call_entry call_entry;
1325 struct _call_entry {
1326 ir_node *call; /**< the Call */
1327 ir_graph *callee; /**< the callee called here */
1328 call_entry *next; /**< for linking the next one */
1332 * environment for inlining small irgs
1334 typedef struct _inline_env_t {
1335 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1336 call_entry *head; /**< the head of the call entry list */
1337 call_entry *tail; /**< the tail of the call entry list */
1341 * Returns the irg called from a Call node. If the irg is not
1342 * known, NULL is returned.
1344 static ir_graph *get_call_called_irg(ir_node *call) {
1346 ir_graph *called_irg = NULL;
1348 addr = get_Call_ptr(call);
1349 if (is_SymConst_addr_ent(addr)) {
1350 called_irg = get_entity_irg(get_SymConst_entity(addr));
1357 * Walker: Collect all calls to known graphs inside a graph.
1359 static void collect_calls(ir_node *call, void *env) {
1360 if (is_Call(call)) {
1361 ir_graph *called_irg = get_call_called_irg(call);
1363 /* The Call node calls a locally defined method. Remember to inline. */
1364 inline_env_t *ienv = env;
1365 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1367 entry->callee = called_irg;
1370 if (ienv->tail == NULL)
1373 ienv->tail->next = entry;
1380 * Inlines all small methods at call sites where the called address comes
1381 * from a Const node that references the entity representing the called
1383 * The size argument is a rough measure for the code size of the method:
1384 * Methods where the obstack containing the firm graph is smaller than
1387 void inline_small_irgs(ir_graph *irg, int size) {
1388 ir_graph *rem = current_ir_graph;
1391 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1393 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1395 current_ir_graph = irg;
1396 /* Handle graph state */
1397 assert(get_irg_phase_state(irg) != phase_building);
1398 free_callee_info(irg);
1400 /* Find Call nodes to inline.
1401 (We can not inline during a walk of the graph, as inlineing the same
1402 method several times changes the visited flag of the walked graph:
1403 after the first inlineing visited of the callee equals visited of
1404 the caller. With the next inlineing both are increased.) */
1405 obstack_init(&env.obst);
1406 env.head = env.tail = NULL;
1407 irg_walk_graph(irg, NULL, collect_calls, &env);
1409 if (env.head != NULL) {
1410 /* There are calls to inline */
1411 collect_phiprojs(irg);
1412 for (entry = env.head; entry != NULL; entry = entry->next) {
1413 ir_graph *callee = entry->callee;
1414 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1415 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1416 inline_method(entry->call, callee);
1420 obstack_free(&env.obst, NULL);
1421 current_ir_graph = rem;
1425 * Environment for inlining irgs.
1428 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1429 int n_nodes_orig; /**< for statistics */
1430 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1431 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1432 int n_call_nodes; /**< Number of Call nodes in the graph. */
1433 int n_call_nodes_orig; /**< for statistics */
1434 int n_callers; /**< Number of known graphs that call this graphs. */
1435 int n_callers_orig; /**< for statistics */
1436 int got_inline; /**< Set, if at leat one call inside this graph was inlined. */
1440 * Allocate a new environment for inlining.
1442 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1443 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1444 env->n_nodes = -2; /* do not count count Start, End */
1445 env->n_nodes_orig = -2; /* do not count Start, End */
1446 env->call_head = NULL;
1447 env->call_tail = NULL;
1448 env->n_call_nodes = 0;
1449 env->n_call_nodes_orig = 0;
1451 env->n_callers_orig = 0;
1452 env->got_inline = 0;
1456 typedef struct walker_env {
1457 struct obstack *obst; /**< the obstack for allocations. */
1458 inline_irg_env *x; /**< the inline environment */
1459 int ignore_runtime; /**< the ignore runtime flag */
1463 * post-walker: collect all calls in the inline-environment
1464 * of a graph and sum some statistics.
1466 static void collect_calls2(ir_node *call, void *ctx) {
1468 inline_irg_env *x = env->x;
1469 ir_op *op = get_irn_op(call);
1473 /* count meaningful nodes in irg */
1474 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1479 if (op != op_Call) return;
1481 /* check, if it's a runtime call */
1482 if (env->ignore_runtime) {
1483 ir_node *symc = get_Call_ptr(call);
1485 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1486 ir_entity *ent = get_SymConst_entity(symc);
1488 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1493 /* collect all call nodes */
1495 ++x->n_call_nodes_orig;
1497 callee = get_call_called_irg(call);
1499 inline_irg_env *callee_env = get_irg_link(callee);
1500 /* count all static callers */
1501 ++callee_env->n_callers;
1502 ++callee_env->n_callers_orig;
1504 /* link it in the list of possible inlinable entries */
1505 entry = obstack_alloc(env->obst, sizeof(*entry));
1507 entry->callee = callee;
1509 if (x->call_tail == NULL)
1510 x->call_head = entry;
1512 x->call_tail->next = entry;
1513 x->call_tail = entry;
1518 * Returns TRUE if the number of callers is 0 in the irg's environment,
1519 * hence this irg is a leave.
1521 INLINE static int is_leave(ir_graph *irg) {
1522 inline_irg_env *env = get_irg_link(irg);
1523 return env->n_call_nodes == 0;
1527 * Returns TRUE if the number of nodes in the callee is
1528 * smaller then size in the irg's environment.
1530 INLINE static int is_smaller(ir_graph *callee, int size) {
1531 inline_irg_env *env = get_irg_link(callee);
1532 return env->n_nodes < size;
1536 * Append the nodes of the list src to the nodes of the list in environment dst.
1538 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1539 call_entry *entry, *nentry;
1541 /* Note that the src list points to Call nodes in the inlined graph, but
1542 we need Call nodes in our graph. Luckily the inliner leaves this information
1543 in the link field. */
1544 for (entry = src; entry != NULL; entry = entry->next) {
1545 nentry = obstack_alloc(obst, sizeof(*nentry));
1546 nentry->call = get_irn_link(entry->call);
1547 nentry->callee = entry->callee;
1548 nentry->next = NULL;
1549 dst->call_tail->next = nentry;
1550 dst->call_tail = nentry;
1555 * Inlines small leave methods at call sites where the called address comes
1556 * from a Const node that references the entity representing the called
1558 * The size argument is a rough measure for the code size of the method:
1559 * Methods where the obstack containing the firm graph is smaller than
1562 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1563 inline_irg_env *env;
1569 call_entry *entry, *tail;
1570 const call_entry *centry;
1571 struct obstack obst;
1572 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1574 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1575 rem = current_ir_graph;
1576 obstack_init(&obst);
1578 /* extend all irgs by a temporary data structure for inlining. */
1579 n_irgs = get_irp_n_irgs();
1580 for (i = 0; i < n_irgs; ++i)
1581 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1583 /* Precompute information in temporary data structure. */
1585 wenv.ignore_runtime = ignore_runtime;
1586 for (i = 0; i < n_irgs; ++i) {
1587 ir_graph *irg = get_irp_irg(i);
1589 assert(get_irg_phase_state(irg) != phase_building);
1590 free_callee_info(irg);
1592 wenv.x = get_irg_link(irg);
1593 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1596 /* -- and now inline. -- */
1598 /* Inline leaves recursively -- we might construct new leaves. */
1602 for (i = 0; i < n_irgs; ++i) {
1604 int phiproj_computed = 0;
1606 current_ir_graph = get_irp_irg(i);
1607 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1610 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1613 if (env->n_nodes > maxsize) break;
1616 callee = entry->callee;
1618 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1619 if (!phiproj_computed) {
1620 phiproj_computed = 1;
1621 collect_phiprojs(current_ir_graph);
1623 did_inline = inline_method(call, callee);
1626 /* Do some statistics */
1627 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1629 env->got_inline = 1;
1630 --env->n_call_nodes;
1631 env->n_nodes += callee_env->n_nodes;
1632 --callee_env->n_callers;
1634 /* remove this call from the list */
1636 tail->next = entry->next;
1638 env->call_head = entry->next;
1644 env->call_tail = tail;
1646 } while (did_inline);
1648 /* inline other small functions. */
1649 for (i = 0; i < n_irgs; ++i) {
1651 int phiproj_computed = 0;
1653 current_ir_graph = get_irp_irg(i);
1654 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1656 /* note that the list of possible calls is updated during the process */
1658 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1662 callee = entry->callee;
1664 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1665 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1666 if (!phiproj_computed) {
1667 phiproj_computed = 1;
1668 collect_phiprojs(current_ir_graph);
1670 if (inline_method(call, callee)) {
1671 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1673 /* callee was inline. Append it's call list. */
1674 env->got_inline = 1;
1675 --env->n_call_nodes;
1676 append_call_list(&obst, env, callee_env->call_head);
1677 env->n_call_nodes += callee_env->n_call_nodes;
1678 env->n_nodes += callee_env->n_nodes;
1679 --callee_env->n_callers;
1681 /* after we have inlined callee, all called methods inside callee
1682 are now called once more */
1683 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1684 inline_irg_env *penv = get_irg_link(centry->callee);
1688 /* remove this call from the list */
1690 tail->next = entry->next;
1692 env->call_head = entry->next;
1698 env->call_tail = tail;
1701 for (i = 0; i < n_irgs; ++i) {
1702 irg = get_irp_irg(i);
1703 env = (inline_irg_env *)get_irg_link(irg);
1705 if (env->got_inline) {
1706 /* this irg got calls inlined */
1707 set_irg_outs_inconsistent(irg);
1708 set_irg_doms_inconsistent(irg);
1710 optimize_graph_df(irg);
1713 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1714 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1715 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1716 env->n_callers_orig, env->n_callers,
1717 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 ignore_exc_edges; /**< set if exception edges should be ignored. */
2228 char changed; /**< flag indicates that the cf graphs has changed. */
2232 * Called by walker of remove_critical_cf_edges().
2234 * Place an empty block to an edge between a blocks of multiple
2235 * predecessors and a block of multiple successors.
2238 * @param env Environment of walker.
2240 static void walk_critical_cf_edges(ir_node *n, void *env) {
2242 ir_node *pre, *block, *jmp;
2244 ir_graph *irg = get_irn_irg(n);
2246 /* Block has multiple predecessors */
2247 arity = get_irn_arity(n);
2249 if (n == get_irg_end_block(irg))
2250 return; /* No use to add a block here. */
2252 for (i = 0; i < arity; ++i) {
2255 pre = get_irn_n(n, i);
2256 /* don't count Bad's */
2260 cfop = get_irn_op(skip_Proj(pre));
2261 if (is_op_fragile(cfop)) {
2262 if (cenv->ignore_exc_edges && get_Proj_proj(pre) == pn_Generic_X_except)
2266 /* we don't want place nodes in the start block, so handle it like forking */
2267 if (is_op_forking(cfop) || cfop == op_Start) {
2268 /* Predecessor has multiple successors. Insert new control flow edge edges. */
2270 /* set predecessor of new block */
2271 block = new_r_Block(irg, 1, &pre);
2272 /* insert new jmp node to new block */
2273 jmp = new_r_Jmp(irg, block);
2274 /* set successor of new block */
2275 set_irn_n(n, i, jmp);
2277 } /* predecessor has multiple successors */
2278 } /* for all predecessors */
2279 } /* n is a multi-entry block */
2282 void remove_critical_cf_edges(ir_graph *irg) {
2285 env.ignore_exc_edges = 1;
2288 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &env);
2290 /* control flow changed */
2291 set_irg_outs_inconsistent(irg);
2292 set_irg_extblk_inconsistent(irg);
2293 set_irg_doms_inconsistent(irg);
2294 set_irg_loopinfo_inconsistent(irg);