3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
21 #include "irgraph_t.h"
34 #include "pdeq.h" /* Fuer code placement */
39 #include "irbackedge_t.h"
46 #include "iredges_t.h"
49 /*------------------------------------------------------------------*/
50 /* apply optimizations of iropt to all nodes. */
51 /*------------------------------------------------------------------*/
54 * A wrapper around optimize_inplace_2() to be called from a walker.
56 static void optimize_in_place_wrapper (ir_node *n, void *env) {
57 ir_node *optimized = optimize_in_place_2(n);
58 if (optimized != n) exchange (n, optimized);
62 * Do local optimizations for a node.
64 * @param n the IR-node where to start. Typically the End node
67 * @note current_ir_graph must be set
69 static INLINE void do_local_optimize(ir_node *n) {
70 /* Handle graph state */
71 assert(get_irg_phase_state(current_ir_graph) != phase_building);
73 if (get_opt_global_cse())
74 set_irg_pinned(current_ir_graph, op_pin_state_floats);
75 set_irg_outs_inconsistent(current_ir_graph);
76 set_irg_doms_inconsistent(current_ir_graph);
77 set_irg_loopinfo_inconsistent(current_ir_graph);
79 /* Clean the value_table in irg for the CSE. */
80 del_identities(current_ir_graph->value_table);
81 current_ir_graph->value_table = new_identities();
83 /* walk over the graph */
84 irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
87 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
88 void local_optimize_node(ir_node *n) {
89 ir_graph *rem = current_ir_graph;
90 current_ir_graph = get_irn_irg(n);
94 current_ir_graph = rem;
98 * Block-Walker: uses dominance depth to mark dead blocks.
100 static void kill_dead_blocks(ir_node *block, void *env)
102 if (get_Block_dom_depth(block) < 0) {
104 * Note that the new dominance code correctly handles
105 * the End block, i.e. it is always reachable from Start
107 set_Block_dead(block);
112 local_optimize_graph(ir_graph *irg) {
113 ir_graph *rem = current_ir_graph;
114 current_ir_graph = irg;
116 if (get_irg_dom_state(irg) == dom_consistent)
117 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
119 do_local_optimize(get_irg_end(irg));
121 current_ir_graph = rem;
125 * Enqueue all users of a node to a wait queue.
126 * Handles mode_T nodes.
128 static void enqueue_users(ir_node *n, pdeq *waitq) {
129 const ir_edge_t *edge;
131 foreach_out_edge(n, edge) {
132 ir_node *succ = get_edge_src_irn(edge);
134 if (get_irn_link(succ) != waitq) {
135 pdeq_putr(waitq, succ);
136 set_irn_link(succ, waitq);
138 if (get_irn_mode(succ) == mode_T) {
139 /* A mode_T node has Proj's. Because most optimizations
140 run on the Proj's we have to enqueue them also. */
141 enqueue_users(succ, waitq);
147 * Data flow optimization walker.
148 * Optimizes all nodes and enqueue it's users
151 static void opt_walker(ir_node *n, void *env) {
155 optimized = optimize_in_place_2(n);
156 set_irn_link(optimized, NULL);
158 if (optimized != n) {
159 enqueue_users(n, waitq);
160 exchange(n, optimized);
164 /* Applies local optimizations to all nodes in the graph until fixpoint. */
165 void optimize_graph_df(ir_graph *irg) {
166 pdeq *waitq = new_pdeq();
167 int state = edges_activated(irg);
168 ir_graph *rem = current_ir_graph;
170 current_ir_graph = irg;
175 if (get_opt_global_cse())
176 set_irg_pinned(current_ir_graph, op_pin_state_floats);
178 /* Clean the value_table in irg for the CSE. */
179 del_identities(irg->value_table);
180 irg->value_table = new_identities();
182 if (get_irg_dom_state(irg) == dom_consistent)
183 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
185 /* invalidate info */
186 set_irg_outs_inconsistent(irg);
187 set_irg_doms_inconsistent(irg);
188 set_irg_loopinfo_inconsistent(irg);
190 /* walk over the graph */
191 irg_walk_graph(irg, NULL, opt_walker, waitq);
193 /* finish the wait queue */
194 while (! pdeq_empty(waitq)) {
195 ir_node *n = pdeq_getl(waitq);
197 opt_walker(n, waitq);
203 edges_deactivate(irg);
205 current_ir_graph = rem;
209 /*------------------------------------------------------------------*/
210 /* Routines for dead node elimination / copying garbage collection */
211 /* of the obstack. */
212 /*------------------------------------------------------------------*/
215 * Remember the new node in the old node by using a field all nodes have.
217 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
220 * Get this new node, before the old node is forgotten.
222 #define get_new_node(oldn) get_irn_link(oldn)
225 * Check if a new node was set.
227 #define has_new_node(n) (get_new_node(n) != NULL)
230 * We use the block_visited flag to mark that we have computed the
231 * number of useful predecessors for this block.
232 * Further we encode the new arity in this flag in the old blocks.
233 * Remembering the arity is useful, as it saves a lot of pointer
234 * accesses. This function is called for all Phi and Block nodes
238 compute_new_arity(ir_node *b) {
239 int i, res, irn_arity;
242 irg_v = get_irg_block_visited(current_ir_graph);
243 block_v = get_Block_block_visited(b);
244 if (block_v >= irg_v) {
245 /* we computed the number of preds for this block and saved it in the
247 return block_v - irg_v;
249 /* compute the number of good predecessors */
250 res = irn_arity = get_irn_arity(b);
251 for (i = 0; i < irn_arity; i++)
252 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
253 /* save it in the flag. */
254 set_Block_block_visited(b, irg_v + res);
260 * Copies the node to the new obstack. The Ins of the new node point to
261 * the predecessors on the old obstack. For block/phi nodes not all
262 * predecessors might be copied. n->link points to the new node.
263 * For Phi and Block nodes the function allocates in-arrays with an arity
264 * only for useful predecessors. The arity is determined by counting
265 * the non-bad predecessors of the block.
267 * @param n The node to be copied
268 * @param env if non-NULL, the node number attribute will be copied to the new node
270 * Note: Also used for loop unrolling.
272 static void copy_node(ir_node *n, void *env) {
275 ir_op *op = get_irn_op(n);
276 int copy_node_nr = env != NULL;
278 /* The end node looses it's flexible in array. This doesn't matter,
279 as dead node elimination builds End by hand, inlineing doesn't use
281 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
284 /* node copied already */
286 } else if (op == op_Block) {
288 new_arity = compute_new_arity(n);
289 n->attr.block.graph_arr = NULL;
291 block = get_nodes_block(n);
293 new_arity = compute_new_arity(block);
295 new_arity = get_irn_arity(n);
298 nn = new_ir_node(get_irn_dbg_info(n),
305 /* Copy the attributes. These might point to additional data. If this
306 was allocated on the old obstack the pointers now are dangling. This
307 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
308 copy_node_attr(n, nn);
309 new_backedge_info(nn);
313 /* for easier debugging, we want to copy the node numbers too */
314 nn->node_nr = n->node_nr;
319 hook_dead_node_elim_subst(current_ir_graph, n, nn);
323 * Copies new predecessors of old node to new node remembered in link.
324 * Spare the Bad predecessors of Phi and Block nodes.
327 copy_preds (ir_node *n, void *env) {
331 nn = get_new_node(n);
333 /* printf("\n old node: "); DDMSG2(n);
334 printf(" new node: "); DDMSG2(nn);
335 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
338 /* Don't copy Bad nodes. */
340 irn_arity = get_irn_arity(n);
341 for (i = 0; i < irn_arity; i++)
342 if (! is_Bad(get_irn_n(n, i))) {
343 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
344 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
347 /* repair the block visited flag from above misuse. Repair it in both
348 graphs so that the old one can still be used. */
349 set_Block_block_visited(nn, 0);
350 set_Block_block_visited(n, 0);
351 /* Local optimization could not merge two subsequent blocks if
352 in array contained Bads. Now it's possible.
353 We don't call optimize_in_place as it requires
354 that the fields in ir_graph are set properly. */
355 if ((get_opt_control_flow_straightening()) &&
356 (get_Block_n_cfgpreds(nn) == 1) &&
357 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
358 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
360 /* Jmp jumps into the block it is in -- deal self cycle. */
361 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
362 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
367 } else if (get_irn_op(n) == op_Phi) {
368 /* Don't copy node if corresponding predecessor in block is Bad.
369 The Block itself should not be Bad. */
370 block = get_nodes_block(n);
371 set_irn_n (nn, -1, get_new_node(block));
373 irn_arity = get_irn_arity(n);
374 for (i = 0; i < irn_arity; i++)
375 if (! is_Bad(get_irn_n(block, i))) {
376 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
377 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
380 /* If the pre walker reached this Phi after the post walker visited the
381 block block_visited is > 0. */
382 set_Block_block_visited(get_nodes_block(n), 0);
383 /* Compacting the Phi's ins might generate Phis with only one
385 if (get_irn_arity(nn) == 1)
386 exchange(nn, get_irn_n(nn, 0));
388 irn_arity = get_irn_arity(n);
389 for (i = -1; i < irn_arity; i++)
390 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
392 /* Now the new node is complete. We can add it to the hash table for CSE.
393 @@@ inlining aborts if we identify End. Why? */
394 if (get_irn_op(nn) != op_End)
395 add_identities (current_ir_graph->value_table, nn);
399 * Copies the graph recursively, compacts the keep-alives of the end node.
401 * @param irg the graph to be copied
402 * @param copy_node_nr If non-zero, the node number will be copied
404 static void copy_graph(ir_graph *irg, int copy_node_nr) {
405 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
406 ir_node *ka; /* keep alive */
410 /* Some nodes must be copied by hand, sigh */
411 vfl = get_irg_visited(irg);
412 set_irg_visited(irg, vfl + 1);
414 oe = get_irg_end(irg);
415 mark_irn_visited(oe);
416 /* copy the end node by hand, allocate dynamic in array! */
417 ne = new_ir_node(get_irn_dbg_info(oe),
424 /* Copy the attributes. Well, there might be some in the future... */
425 copy_node_attr(oe, ne);
426 set_new_node(oe, ne);
428 /* copy the Bad node */
429 ob = get_irg_bad(irg);
430 mark_irn_visited(ob);
431 nb = new_ir_node(get_irn_dbg_info(ob),
438 copy_node_attr(ob, nb);
439 set_new_node(ob, nb);
441 /* copy the NoMem node */
442 om = get_irg_no_mem(irg);
443 mark_irn_visited(om);
444 nm = new_ir_node(get_irn_dbg_info(om),
451 copy_node_attr(om, nm);
452 set_new_node(om, nm);
454 /* copy the live nodes */
455 set_irg_visited(irg, vfl);
456 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
458 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
460 /* visit the anchors as well */
461 for (i = anchor_max - 1; i >= 0; --i) {
462 ir_node *n = irg->anchors[i];
464 if (n && (get_irn_visited(n) <= vfl)) {
465 set_irg_visited(irg, vfl);
466 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
470 /* copy_preds for the end node ... */
471 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
473 /*- ... and now the keep alives. -*/
474 /* First pick the not marked block nodes and walk them. We must pick these
475 first as else we will oversee blocks reachable from Phis. */
476 irn_arity = get_End_n_keepalives(oe);
477 for (i = 0; i < irn_arity; i++) {
478 ka = get_End_keepalive(oe, i);
480 if (get_irn_visited(ka) <= vfl) {
481 /* We must keep the block alive and copy everything reachable */
482 set_irg_visited(irg, vfl);
483 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
485 add_End_keepalive(ne, get_new_node(ka));
489 /* Now pick other nodes. Here we will keep all! */
490 irn_arity = get_End_n_keepalives(oe);
491 for (i = 0; i < irn_arity; i++) {
492 ka = get_End_keepalive(oe, i);
494 if (get_irn_visited(ka) <= vfl) {
495 /* We didn't copy the node yet. */
496 set_irg_visited(irg, vfl);
497 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
499 add_End_keepalive(ne, get_new_node(ka));
503 /* start block sometimes only reached after keep alives */
504 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
505 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
509 * Copies the graph reachable from current_ir_graph->end to the obstack
510 * in current_ir_graph and fixes the environment.
511 * Then fixes the fields in current_ir_graph containing nodes of the
514 * @param copy_node_nr If non-zero, the node number will be copied
517 copy_graph_env (int copy_node_nr) {
518 ir_graph *irg = current_ir_graph;
519 ir_node *old_end, *n;
522 /* remove end_except and end_reg nodes */
523 old_end = get_irg_end(irg);
524 set_irg_end_except (irg, old_end);
525 set_irg_end_reg (irg, old_end);
527 /* Not all nodes remembered in irg might be reachable
528 from the end node. Assure their link is set to NULL, so that
529 we can test whether new nodes have been computed. */
530 for (i = anchor_max - 1; i >= 0; --i)
532 set_new_node(irg->anchors[i], NULL);
534 /* we use the block walk flag for removing Bads from Blocks ins. */
535 inc_irg_block_visited(irg);
538 copy_graph(irg, copy_node_nr);
540 /* fix the fields in irg */
541 old_end = get_irg_end(irg);
542 for (i = anchor_max - 1; i >= 0; --i) {
545 irg->anchors[i] = get_new_node(n);
551 * Copies all reachable nodes to a new obstack. Removes bad inputs
552 * from block nodes and the corresponding inputs from Phi nodes.
553 * Merges single exit blocks with single entry blocks and removes
555 * Adds all new nodes to a new hash table for CSE. Does not
556 * perform CSE, so the hash table might contain common subexpressions.
559 dead_node_elimination(ir_graph *irg) {
561 int rem_ipview = get_interprocedural_view();
562 struct obstack *graveyard_obst = NULL;
563 struct obstack *rebirth_obst = NULL;
565 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
566 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
568 /* inform statistics that we started a dead-node elimination run */
569 hook_dead_node_elim(irg, 1);
571 /* Remember external state of current_ir_graph. */
572 rem = current_ir_graph;
573 current_ir_graph = irg;
574 set_interprocedural_view(0);
576 assert(get_irg_phase_state(current_ir_graph) != phase_building);
578 /* Handle graph state */
579 free_callee_info(current_ir_graph);
580 free_irg_outs(current_ir_graph);
583 /* @@@ so far we loose loops when copying */
584 free_loop_information(current_ir_graph);
586 set_irg_doms_inconsistent(irg);
588 /* A quiet place, where the old obstack can rest in peace,
589 until it will be cremated. */
590 graveyard_obst = irg->obst;
592 /* A new obstack, where the reachable nodes will be copied to. */
593 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
594 current_ir_graph->obst = rebirth_obst;
595 obstack_init (current_ir_graph->obst);
596 current_ir_graph->last_node_idx = 0;
598 /* We also need a new value table for CSE */
599 del_identities(irg->value_table);
600 irg->value_table = new_identities();
602 /* Copy the graph from the old to the new obstack */
603 copy_graph_env(/*copy_node_nr=*/1);
605 /* Free memory from old unoptimized obstack */
606 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
607 xfree (graveyard_obst); /* ... then free it. */
609 /* inform statistics that the run is over */
610 hook_dead_node_elim(irg, 0);
612 current_ir_graph = rem;
613 set_interprocedural_view(rem_ipview);
618 * Relink bad predecessors of a block and store the old in array to the
619 * link field. This function is called by relink_bad_predecessors().
620 * The array of link field starts with the block operand at position 0.
621 * If block has bad predecessors, create a new in array without bad preds.
622 * Otherwise let in array untouched.
624 static void relink_bad_block_predecessors(ir_node *n, void *env) {
625 ir_node **new_in, *irn;
626 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
628 /* if link field of block is NULL, look for bad predecessors otherwise
629 this is already done */
630 if (get_irn_op(n) == op_Block &&
631 get_irn_link(n) == NULL) {
633 /* save old predecessors in link field (position 0 is the block operand)*/
634 set_irn_link(n, get_irn_in(n));
636 /* count predecessors without bad nodes */
637 old_irn_arity = get_irn_arity(n);
638 for (i = 0; i < old_irn_arity; i++)
639 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
641 /* arity changing: set new predecessors without bad nodes */
642 if (new_irn_arity < old_irn_arity) {
643 /* Get new predecessor array. We do not resize the array, as we must
644 keep the old one to update Phis. */
645 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
647 /* set new predecessors in array */
650 for (i = 0; i < old_irn_arity; i++) {
651 irn = get_irn_n(n, i);
653 new_in[new_irn_n] = irn;
654 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
658 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
659 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
662 } /* ir node has bad predecessors */
664 } /* Block is not relinked */
668 * Relinks Bad predecessors from Blocks and Phis called by walker
669 * remove_bad_predecesors(). If n is a Block, call
670 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
671 * function of Phi's Block. If this block has bad predecessors, relink preds
674 static void relink_bad_predecessors(ir_node *n, void *env) {
675 ir_node *block, **old_in;
676 int i, old_irn_arity, new_irn_arity;
678 /* relink bad predecessors of a block */
679 if (get_irn_op(n) == op_Block)
680 relink_bad_block_predecessors(n, env);
682 /* If Phi node relink its block and its predecessors */
683 if (get_irn_op(n) == op_Phi) {
685 /* Relink predecessors of phi's block */
686 block = get_nodes_block(n);
687 if (get_irn_link(block) == NULL)
688 relink_bad_block_predecessors(block, env);
690 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
691 old_irn_arity = ARR_LEN(old_in);
693 /* Relink Phi predecessors if count of predecessors changed */
694 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
695 /* set new predecessors in array
696 n->in[0] remains the same block */
698 for(i = 1; i < old_irn_arity; i++)
699 if (!is_Bad((ir_node *)old_in[i])) {
700 n->in[new_irn_arity] = n->in[i];
701 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
705 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
706 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
709 } /* n is a Phi node */
713 * Removes Bad Bad predecessors from Blocks and the corresponding
714 * inputs to Phi nodes as in dead_node_elimination but without
716 * On walking up set the link field to NULL, on walking down call
717 * relink_bad_predecessors() (This function stores the old in array
718 * to the link field and sets a new in array if arity of predecessors
721 void remove_bad_predecessors(ir_graph *irg) {
722 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
729 __)|_| | \_/ | \_/(/_ |_/\__|__
731 The following stuff implements a facility that automatically patches
732 registered ir_node pointers to the new node when a dead node elimination occurs.
735 struct _survive_dce_t {
739 hook_entry_t dead_node_elim;
740 hook_entry_t dead_node_elim_subst;
743 typedef struct _survive_dce_list_t {
744 struct _survive_dce_list_t *next;
746 } survive_dce_list_t;
748 static void dead_node_hook(void *context, ir_graph *irg, int start)
750 survive_dce_t *sd = context;
752 /* Create a new map before the dead node elimination is performed. */
754 sd->new_places = pmap_create_ex(pmap_count(sd->places));
757 /* Patch back all nodes if dead node elimination is over and something is to be done. */
759 pmap_destroy(sd->places);
760 sd->places = sd->new_places;
761 sd->new_places = NULL;
766 * Hook called when dead node elimination replaces old by nw.
768 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw)
770 survive_dce_t *sd = context;
771 survive_dce_list_t *list = pmap_get(sd->places, old);
773 /* If the node is to be patched back, write the new address to all registered locations. */
775 survive_dce_list_t *p;
777 for(p = list; p; p = p->next)
780 pmap_insert(sd->new_places, nw, list);
785 * Make a new Survive DCE environment.
787 survive_dce_t *new_survive_dce(void)
789 survive_dce_t *res = xmalloc(sizeof(res[0]));
790 obstack_init(&res->obst);
791 res->places = pmap_create();
792 res->new_places = NULL;
794 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
795 res->dead_node_elim.context = res;
796 res->dead_node_elim.next = NULL;
798 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
799 res->dead_node_elim_subst.context = res;
800 res->dead_node_elim_subst.next = NULL;
802 register_hook(hook_dead_node_elim, &res->dead_node_elim);
803 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
808 * Free a Survive DCE environment.
810 void free_survive_dce(survive_dce_t *sd)
812 obstack_free(&sd->obst, NULL);
813 pmap_destroy(sd->places);
814 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
815 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
820 * Register a node pointer to be patched upon DCE.
821 * When DCE occurs, the node pointer specified by @p place will be
822 * patched to the new address of the node it is pointing to.
824 * @param sd The Survive DCE environment.
825 * @param place The address of the node pointer.
827 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place)
830 ir_node *irn = *place;
831 survive_dce_list_t *curr = pmap_get(sd->places, irn);
832 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw));
837 pmap_insert(sd->places, irn, nw);
841 /*--------------------------------------------------------------------*/
842 /* Functionality for inlining */
843 /*--------------------------------------------------------------------*/
846 * Copy node for inlineing. Updates attributes that change when
847 * inlineing but not for dead node elimination.
849 * Copies the node by calling copy_node() and then updates the entity if
850 * it's a local one. env must be a pointer of the frame type of the
851 * inlined procedure. The new entities must be in the link field of
855 copy_node_inline (ir_node *n, void *env) {
857 ir_type *frame_tp = (ir_type *)env;
860 if (get_irn_op(n) == op_Sel) {
861 nn = get_new_node (n);
863 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
864 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
866 } else if (get_irn_op(n) == op_Block) {
867 nn = get_new_node (n);
868 nn->attr.block.irg = current_ir_graph;
872 static void find_addr(ir_node *node, void *env)
874 if (get_irn_opcode(node) == iro_Proj) {
875 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
881 * currently, we cannot inline two cases:
882 * - call with compound arguments
883 * - graphs that take the address of a parameter
885 * check these conditions here
887 static int can_inline(ir_node *call, ir_graph *called_graph)
889 ir_type *call_type = get_Call_type(call);
890 int params, ress, i, res;
891 assert(is_Method_type(call_type));
893 params = get_method_n_params(call_type);
894 ress = get_method_n_ress(call_type);
897 for (i = 0; i < params; ++i) {
898 ir_type *p_type = get_method_param_type(call_type, i);
900 if (is_compound_type(p_type))
905 for (i = 0; i < ress; ++i) {
906 ir_type *r_type = get_method_res_type(call_type, i);
908 if (is_compound_type(r_type))
913 irg_walk_graph(called_graph, find_addr, NULL, &res);
918 int inline_method(ir_node *call, ir_graph *called_graph) {
920 ir_node *post_call, *post_bl;
921 ir_node *in[pn_Start_max];
922 ir_node *end, *end_bl;
926 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
928 ir_type *called_frame;
929 irg_inline_property prop = get_irg_inline_property(called_graph);
931 if ( (prop < irg_inline_forced) &&
932 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
934 /* Do not inline variadic functions. */
935 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
938 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
939 get_method_n_params(get_Call_type(call)));
942 * currently, we cannot inline two cases:
943 * - call with compound arguments
944 * - graphs that take the address of a parameter
946 if (! can_inline(call, called_graph))
949 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
950 rem_opt = get_opt_optimize();
953 /* Handle graph state */
954 assert(get_irg_phase_state(current_ir_graph) != phase_building);
955 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
956 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
957 set_irg_outs_inconsistent(current_ir_graph);
958 set_irg_extblk_inconsistent(current_ir_graph);
959 set_irg_doms_inconsistent(current_ir_graph);
960 set_irg_loopinfo_inconsistent(current_ir_graph);
961 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
963 /* -- Check preconditions -- */
964 assert(is_Call(call));
965 /* @@@ does not work for InterfaceIII.java after cgana
966 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
967 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
968 get_Call_type(call)));
970 assert(get_type_tpop(get_Call_type(call)) == type_method);
971 if (called_graph == current_ir_graph) {
972 set_optimize(rem_opt);
976 /* here we know we WILL inline, so inform the statistics */
977 hook_inline(call, called_graph);
979 /* -- Decide how to handle exception control flow: Is there a handler
980 for the Call node, or do we branch directly to End on an exception?
982 0 There is a handler.
984 2 Exception handling not represented in Firm. -- */
986 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
987 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
988 assert(get_irn_op(proj) == op_Proj);
989 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
990 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
992 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
993 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
994 else { exc_handling = 2; } /* !Mproj && !Xproj */
999 the procedure and later replaces the Start node of the called graph.
1000 Post_call is the old Call node and collects the results of the called
1001 graph. Both will end up being a tuple. -- */
1002 post_bl = get_nodes_block(call);
1003 set_irg_current_block(current_ir_graph, post_bl);
1004 /* XxMxPxPxPxT of Start + parameter of Call */
1005 in[pn_Start_X_initial_exec] = new_Jmp();
1006 in[pn_Start_M] = get_Call_mem(call);
1007 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1008 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1009 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1010 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1011 /* in[pn_Start_P_value_arg_base] = ??? */
1012 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1013 pre_call = new_Tuple(pn_Start_max - 1, in);
1017 The new block gets the ins of the old block, pre_call and all its
1018 predecessors and all Phi nodes. -- */
1019 part_block(pre_call);
1021 /* -- Prepare state for dead node elimination -- */
1022 /* Visited flags in calling irg must be >= flag in called irg.
1023 Else walker and arity computation will not work. */
1024 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1025 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1026 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1027 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1028 /* Set pre_call as new Start node in link field of the start node of
1029 calling graph and pre_calls block as new block for the start block
1031 Further mark these nodes so that they are not visited by the
1033 set_irn_link(get_irg_start(called_graph), pre_call);
1034 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1035 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1036 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1037 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1038 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1040 /* Initialize for compaction of in arrays */
1041 inc_irg_block_visited(current_ir_graph);
1043 /* -- Replicate local entities of the called_graph -- */
1044 /* copy the entities. */
1045 called_frame = get_irg_frame_type(called_graph);
1046 for (i = 0; i < get_class_n_members(called_frame); i++) {
1047 entity *new_ent, *old_ent;
1048 old_ent = get_class_member(called_frame, i);
1049 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1050 set_entity_link(old_ent, new_ent);
1053 /* visited is > than that of called graph. With this trick visited will
1054 remain unchanged so that an outer walker, e.g., searching the call nodes
1055 to inline, calling this inline will not visit the inlined nodes. */
1056 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1058 /* -- Performing dead node elimination inlines the graph -- */
1059 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1061 /* @@@ endless loops are not copied!! -- they should be, I think... */
1062 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1063 get_irg_frame_type(called_graph));
1065 /* Repair called_graph */
1066 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1067 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1068 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1070 /* -- Merge the end of the inlined procedure with the call site -- */
1071 /* We will turn the old Call node into a Tuple with the following
1074 0: Phi of all Memories of Return statements.
1075 1: Jmp from new Block that merges the control flow from all exception
1076 predecessors of the old end block.
1077 2: Tuple of all arguments.
1078 3: Phi of Exception memories.
1079 In case the old Call directly branches to End on an exception we don't
1080 need the block merging all exceptions nor the Phi of the exception
1084 /* -- Precompute some values -- */
1085 end_bl = get_new_node(get_irg_end_block(called_graph));
1086 end = get_new_node(get_irg_end(called_graph));
1087 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1088 n_res = get_method_n_ress(get_Call_type(call));
1090 res_pred = xmalloc (n_res * sizeof(*res_pred));
1091 cf_pred = xmalloc (arity * sizeof(*res_pred));
1093 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1095 /* -- archive keepalives -- */
1096 irn_arity = get_irn_arity(end);
1097 for (i = 0; i < irn_arity; i++)
1098 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
1100 /* The new end node will die. We need not free as the in array is on the obstack:
1101 copy_node() only generated 'D' arrays. */
1103 /* -- Replace Return nodes by Jump nodes. -- */
1105 for (i = 0; i < arity; i++) {
1107 ret = get_irn_n(end_bl, i);
1108 if (is_Return(ret)) {
1109 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1113 set_irn_in(post_bl, n_ret, cf_pred);
1115 /* -- Build a Tuple for all results of the method.
1116 Add Phi node if there was more than one Return. -- */
1117 turn_into_tuple(post_call, 4);
1118 /* First the Memory-Phi */
1120 for (i = 0; i < arity; i++) {
1121 ret = get_irn_n(end_bl, i);
1122 if (is_Return(ret)) {
1123 cf_pred[n_ret] = get_Return_mem(ret);
1127 phi = new_Phi(n_ret, cf_pred, mode_M);
1128 set_Tuple_pred(call, pn_Call_M_regular, phi);
1129 /* Conserve Phi-list for further inlinings -- but might be optimized */
1130 if (get_nodes_block(phi) == post_bl) {
1131 set_irn_link(phi, get_irn_link(post_bl));
1132 set_irn_link(post_bl, phi);
1134 /* Now the real results */
1136 for (j = 0; j < n_res; j++) {
1138 for (i = 0; i < arity; i++) {
1139 ret = get_irn_n(end_bl, i);
1140 if (get_irn_op(ret) == op_Return) {
1141 cf_pred[n_ret] = get_Return_res(ret, j);
1146 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1150 /* Conserve Phi-list for further inlinings -- but might be optimized */
1151 if (get_nodes_block(phi) == post_bl) {
1152 set_irn_link(phi, get_irn_link(post_bl));
1153 set_irn_link(post_bl, phi);
1156 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1158 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1160 /* Finally the exception control flow.
1161 We have two (three) possible situations:
1162 First if the Call branches to an exception handler: We need to add a Phi node to
1163 collect the memory containing the exception objects. Further we need
1164 to add another block to get a correct representation of this Phi. To
1165 this block we add a Jmp that resolves into the X output of the Call
1166 when the Call is turned into a tuple.
1167 Second the Call branches to End, the exception is not handled. Just
1168 add all inlined exception branches to the End node.
1169 Third: there is no Exception edge at all. Handle as case two. */
1170 if (exc_handling == 0) {
1172 for (i = 0; i < arity; i++) {
1174 ret = get_irn_n(end_bl, i);
1175 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1176 cf_pred[n_exc] = ret;
1181 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1182 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1183 /* The Phi for the memories with the exception objects */
1185 for (i = 0; i < arity; i++) {
1187 ret = skip_Proj(get_irn_n(end_bl, i));
1189 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1191 } else if (is_fragile_op(ret)) {
1192 /* We rely that all cfops have the memory output at the same position. */
1193 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1195 } else if (get_irn_op(ret) == op_Raise) {
1196 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1200 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1202 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1203 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1206 ir_node *main_end_bl;
1207 int main_end_bl_arity;
1208 ir_node **end_preds;
1210 /* assert(exc_handling == 1 || no exceptions. ) */
1212 for (i = 0; i < arity; i++) {
1213 ir_node *ret = get_irn_n(end_bl, i);
1215 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1216 cf_pred[n_exc] = ret;
1220 main_end_bl = get_irg_end_block(current_ir_graph);
1221 main_end_bl_arity = get_irn_arity(main_end_bl);
1222 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1224 for (i = 0; i < main_end_bl_arity; ++i)
1225 end_preds[i] = get_irn_n(main_end_bl, i);
1226 for (i = 0; i < n_exc; ++i)
1227 end_preds[main_end_bl_arity + i] = cf_pred[i];
1228 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1229 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1230 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1236 #if 0 /* old. now better, correcter, faster implementation. */
1238 /* -- If the exception control flow from the inlined Call directly
1239 branched to the end block we now have the following control
1240 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1241 remove the Jmp along with it's empty block and add Jmp's
1242 predecessors as predecessors of this end block. No problem if
1243 there is no exception, because then branches Bad to End which
1245 @@@ can't we know this beforehand: by getting the Proj(1) from
1246 the Call link list and checking whether it goes to Proj. */
1247 /* find the problematic predecessor of the end block. */
1248 end_bl = get_irg_end_block(current_ir_graph);
1249 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1250 cf_op = get_Block_cfgpred(end_bl, i);
1251 if (get_irn_op(cf_op) == op_Proj) {
1252 cf_op = get_Proj_pred(cf_op);
1253 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1254 /* There are unoptimized tuples from inlineing before when no exc */
1255 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1256 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1257 assert(get_irn_op(cf_op) == op_Jmp);
1263 if (i < get_Block_n_cfgpreds(end_bl)) {
1264 bl = get_nodes_block(cf_op);
1265 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1266 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1267 for (j = 0; j < i; j++)
1268 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1269 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1270 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1271 for (j = j; j < arity; j++)
1272 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1273 set_irn_in(end_bl, arity, cf_pred);
1275 /* Remove the exception pred from post-call Tuple. */
1276 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1281 /* -- Turn CSE back on. -- */
1282 set_optimize(rem_opt);
1287 /********************************************************************/
1288 /* Apply inlineing to small methods. */
1289 /********************************************************************/
1291 /* It makes no sense to inline too many calls in one procedure. Anyways,
1292 I didn't get a version with NEW_ARR_F to run. */
1293 #define MAX_INLINE 1024
1296 * environment for inlining small irgs
1298 typedef struct _inline_env_t {
1300 ir_node *calls[MAX_INLINE];
1304 * Returns the irg called from a Call node. If the irg is not
1305 * known, NULL is returned.
1307 static ir_graph *get_call_called_irg(ir_node *call) {
1309 ir_graph *called_irg = NULL;
1311 assert(is_Call(call));
1313 addr = get_Call_ptr(call);
1314 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1315 called_irg = get_entity_irg(get_SymConst_entity(addr));
1321 static void collect_calls(ir_node *call, void *env) {
1324 if (! is_Call(call)) return;
1326 addr = get_Call_ptr(call);
1328 if (get_irn_op(addr) == op_SymConst) {
1329 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1330 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1331 inline_env_t *ienv = (inline_env_t *)env;
1332 if (called_irg && ienv->pos < MAX_INLINE) {
1333 /* The Call node calls a locally defined method. Remember to inline. */
1334 ienv->calls[ienv->pos++] = call;
1341 * Inlines all small methods at call sites where the called address comes
1342 * from a Const node that references the entity representing the called
1344 * The size argument is a rough measure for the code size of the method:
1345 * Methods where the obstack containing the firm graph is smaller than
1348 void inline_small_irgs(ir_graph *irg, int size) {
1350 ir_graph *rem = current_ir_graph;
1351 inline_env_t env /* = {0, NULL}*/;
1353 if (!(get_opt_optimize() && get_opt_inline())) return;
1355 current_ir_graph = irg;
1356 /* Handle graph state */
1357 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1358 free_callee_info(current_ir_graph);
1360 /* Find Call nodes to inline.
1361 (We can not inline during a walk of the graph, as inlineing the same
1362 method several times changes the visited flag of the walked graph:
1363 after the first inlineing visited of the callee equals visited of
1364 the caller. With the next inlineing both are increased.) */
1366 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1368 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1369 /* There are calls to inline */
1370 collect_phiprojs(irg);
1371 for (i = 0; i < env.pos; i++) {
1373 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1374 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1375 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1376 inline_method(env.calls[i], callee);
1381 current_ir_graph = rem;
1385 * Environment for inlining irgs.
1388 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1389 int n_nodes_orig; /**< for statistics */
1390 eset *call_nodes; /**< All call nodes in this graph */
1392 int n_call_nodes_orig; /**< for statistics */
1393 int n_callers; /**< Number of known graphs that call this graphs. */
1394 int n_callers_orig; /**< for statistics */
1398 * Allocate a new environment for inlining.
1400 static inline_irg_env *new_inline_irg_env(void) {
1401 inline_irg_env *env = xmalloc(sizeof(*env));
1402 env->n_nodes = -2; /* do not count count Start, End */
1403 env->n_nodes_orig = -2; /* do not count Start, End */
1404 env->call_nodes = eset_create();
1405 env->n_call_nodes = 0;
1406 env->n_call_nodes_orig = 0;
1408 env->n_callers_orig = 0;
1413 * destroy an environment for inlining.
1415 static void free_inline_irg_env(inline_irg_env *env) {
1416 eset_destroy(env->call_nodes);
1421 * post-walker: collect all calls in the inline-environment
1422 * of a graph and sum some statistics.
1424 static void collect_calls2(ir_node *call, void *env) {
1425 inline_irg_env *x = (inline_irg_env *)env;
1426 ir_op *op = get_irn_op(call);
1429 /* count meaningful nodes in irg */
1430 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1435 if (op != op_Call) return;
1437 /* collect all call nodes */
1438 eset_insert(x->call_nodes, call);
1440 x->n_call_nodes_orig++;
1442 /* count all static callers */
1443 callee = get_call_called_irg(call);
1445 inline_irg_env *callee_env = get_irg_link(callee);
1446 callee_env->n_callers++;
1447 callee_env->n_callers_orig++;
1452 * Returns TRUE if the number of callers in 0 in the irg's environment,
1453 * hence this irg is a leave.
1455 INLINE static int is_leave(ir_graph *irg) {
1456 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1460 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1462 INLINE static int is_smaller(ir_graph *callee, int size) {
1463 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1468 * Inlines small leave methods at call sites where the called address comes
1469 * from a Const node that references the entity representing the called
1471 * The size argument is a rough measure for the code size of the method:
1472 * Methods where the obstack containing the firm graph is smaller than
1475 void inline_leave_functions(int maxsize, int leavesize, int size) {
1476 inline_irg_env *env;
1477 int i, n_irgs = get_irp_n_irgs();
1478 ir_graph *rem = current_ir_graph;
1481 if (!(get_opt_optimize() && get_opt_inline())) return;
1483 /* extend all irgs by a temporary data structure for inlining. */
1484 for (i = 0; i < n_irgs; ++i)
1485 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1487 /* Precompute information in temporary data structure. */
1488 for (i = 0; i < n_irgs; ++i) {
1489 current_ir_graph = get_irp_irg(i);
1490 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1491 free_callee_info(current_ir_graph);
1493 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1494 get_irg_link(current_ir_graph));
1497 /* -- and now inline. -- */
1499 /* Inline leaves recursively -- we might construct new leaves. */
1500 while (did_inline) {
1503 for (i = 0; i < n_irgs; ++i) {
1505 int phiproj_computed = 0;
1507 current_ir_graph = get_irp_irg(i);
1508 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1510 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1513 if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
1514 callee = get_call_called_irg(call);
1516 if (env->n_nodes > maxsize) continue; // break;
1518 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1519 if (!phiproj_computed) {
1520 phiproj_computed = 1;
1521 collect_phiprojs(current_ir_graph);
1523 did_inline = inline_method(call, callee);
1526 /* Do some statistics */
1527 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1528 env->n_call_nodes --;
1529 env->n_nodes += callee_env->n_nodes;
1530 callee_env->n_callers--;
1537 /* inline other small functions. */
1538 for (i = 0; i < n_irgs; ++i) {
1541 int phiproj_computed = 0;
1543 current_ir_graph = get_irp_irg(i);
1544 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1546 /* we can not walk and change a set, nor remove from it.
1548 walkset = env->call_nodes;
1549 env->call_nodes = eset_create();
1550 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1553 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1554 callee = get_call_called_irg(call);
1557 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1558 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1559 if (!phiproj_computed) {
1560 phiproj_computed = 1;
1561 collect_phiprojs(current_ir_graph);
1563 if (inline_method(call, callee)) {
1564 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1565 env->n_call_nodes--;
1566 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1567 env->n_call_nodes += callee_env->n_call_nodes;
1568 env->n_nodes += callee_env->n_nodes;
1569 callee_env->n_callers--;
1572 eset_insert(env->call_nodes, call);
1575 eset_destroy(walkset);
1578 for (i = 0; i < n_irgs; ++i) {
1579 current_ir_graph = get_irp_irg(i);
1581 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1582 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1583 (env->n_callers_orig != env->n_callers))
1584 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1585 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1586 env->n_callers_orig, env->n_callers,
1587 get_entity_name(get_irg_entity(current_ir_graph)));
1589 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1592 current_ir_graph = rem;
1595 /*******************************************************************/
1596 /* Code Placement. Pins all floating nodes to a block where they */
1597 /* will be executed only if needed. */
1598 /*******************************************************************/
1601 * Returns non-zero, is a block is not reachable from Start.
1603 * @param block the block to test
1606 is_Block_unreachable(ir_node *block) {
1607 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1611 * Find the earliest correct block for N. --- Place N into the
1612 * same Block as its dominance-deepest Input.
1614 * We have to avoid calls to get_nodes_block() here
1615 * because the graph is floating.
1617 * move_out_of_loops() expects that place_floats_early() have placed
1618 * all "living" nodes into a living block. That's why we must
1619 * move nodes in dead block with "live" successors into a valid
1621 * We move them just into the same block as it's successor (or
1622 * in case of a Phi into the effective use block). For Phi successors,
1623 * this may still be a dead block, but then there is no real use, as
1624 * the control flow will be dead later.
1627 place_floats_early(ir_node *n, pdeq *worklist)
1631 /* we must not run into an infinite loop */
1632 assert(irn_not_visited(n));
1633 mark_irn_visited(n);
1635 /* Place floating nodes. */
1636 if (get_irn_pinned(n) == op_pin_state_floats) {
1637 ir_node *curr_block = get_irn_n(n, -1);
1638 int in_dead_block = is_Block_unreachable(curr_block);
1640 ir_node *b = NULL; /* The block to place this node in */
1642 assert(get_irn_op(n) != op_Block);
1644 if ((get_irn_op(n) == op_Const) ||
1645 (get_irn_op(n) == op_SymConst) ||
1647 (get_irn_op(n) == op_Unknown)) {
1648 /* These nodes will not be placed by the loop below. */
1649 b = get_irg_start_block(current_ir_graph);
1653 /* find the block for this node. */
1654 irn_arity = get_irn_arity(n);
1655 for (i = 0; i < irn_arity; i++) {
1656 ir_node *pred = get_irn_n(n, i);
1657 ir_node *pred_block;
1659 if ((irn_not_visited(pred))
1660 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1663 * If the current node is NOT in a dead block, but one of its
1664 * predecessors is, we must move the predecessor to a live block.
1665 * Such thing can happen, if global CSE chose a node from a dead block.
1666 * We move it simple to our block.
1667 * Note that neither Phi nor End nodes are floating, so we don't
1668 * need to handle them here.
1670 if (! in_dead_block) {
1671 if (get_irn_pinned(pred) == op_pin_state_floats &&
1672 is_Block_unreachable(get_irn_n(pred, -1)))
1673 set_nodes_block(pred, curr_block);
1675 place_floats_early(pred, worklist);
1679 * A node in the Bad block must stay in the bad block,
1680 * so don't compute a new block for it.
1685 /* Because all loops contain at least one op_pin_state_pinned node, now all
1686 our inputs are either op_pin_state_pinned or place_early() has already
1687 been finished on them. We do not have any unfinished inputs! */
1688 pred_block = get_irn_n(pred, -1);
1689 if ((!is_Block_dead(pred_block)) &&
1690 (get_Block_dom_depth(pred_block) > depth)) {
1692 depth = get_Block_dom_depth(pred_block);
1694 /* Avoid that the node is placed in the Start block */
1695 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1696 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1697 assert(b != get_irg_start_block(current_ir_graph));
1702 set_nodes_block(n, b);
1706 * Add predecessors of non floating nodes and non-floating predecessors
1707 * of floating nodes to worklist and fix their blocks if the are in dead block.
1709 irn_arity = get_irn_arity(n);
1711 if (get_irn_op(n) == op_End) {
1713 * Simplest case: End node. Predecessors are keep-alives,
1714 * no need to move out of dead block.
1716 for (i = -1; i < irn_arity; ++i) {
1717 ir_node *pred = get_irn_n(n, i);
1718 if (irn_not_visited(pred))
1719 pdeq_putr(worklist, pred);
1722 else if (is_Block(n)) {
1724 * Blocks: Predecessors are control flow, no need to move
1725 * them out of dead block.
1727 for (i = irn_arity - 1; i >= 0; --i) {
1728 ir_node *pred = get_irn_n(n, i);
1729 if (irn_not_visited(pred))
1730 pdeq_putr(worklist, pred);
1733 else if (is_Phi(n)) {
1735 ir_node *curr_block = get_irn_n(n, -1);
1736 int in_dead_block = is_Block_unreachable(curr_block);
1739 * Phi nodes: move nodes from dead blocks into the effective use
1740 * of the Phi-input if the Phi is not in a bad block.
1742 pred = get_irn_n(n, -1);
1743 if (irn_not_visited(pred))
1744 pdeq_putr(worklist, pred);
1746 for (i = irn_arity - 1; i >= 0; --i) {
1747 ir_node *pred = get_irn_n(n, i);
1749 if (irn_not_visited(pred)) {
1750 if (! in_dead_block &&
1751 get_irn_pinned(pred) == op_pin_state_floats &&
1752 is_Block_unreachable(get_irn_n(pred, -1))) {
1753 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1755 pdeq_putr(worklist, pred);
1761 ir_node *curr_block = get_irn_n(n, -1);
1762 int in_dead_block = is_Block_unreachable(curr_block);
1765 * All other nodes: move nodes from dead blocks into the same block.
1767 pred = get_irn_n(n, -1);
1768 if (irn_not_visited(pred))
1769 pdeq_putr(worklist, pred);
1771 for (i = irn_arity - 1; i >= 0; --i) {
1772 ir_node *pred = get_irn_n(n, i);
1774 if (irn_not_visited(pred)) {
1775 if (! in_dead_block &&
1776 get_irn_pinned(pred) == op_pin_state_floats &&
1777 is_Block_unreachable(get_irn_n(pred, -1))) {
1778 set_nodes_block(pred, curr_block);
1780 pdeq_putr(worklist, pred);
1787 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1788 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1789 * places all floating nodes reachable from its argument through floating
1790 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1792 static INLINE void place_early(pdeq *worklist) {
1794 inc_irg_visited(current_ir_graph);
1796 /* this inits the worklist */
1797 place_floats_early(get_irg_end(current_ir_graph), worklist);
1799 /* Work the content of the worklist. */
1800 while (!pdeq_empty(worklist)) {
1801 ir_node *n = pdeq_getl(worklist);
1802 if (irn_not_visited(n))
1803 place_floats_early(n, worklist);
1806 set_irg_outs_inconsistent(current_ir_graph);
1807 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1811 * Compute the deepest common ancestor of block and dca.
1813 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1817 /* we do not want to place nodes in dead blocks */
1818 if (is_Block_dead(block))
1821 /* We found a first legal placement. */
1822 if (!dca) return block;
1824 /* Find a placement that is dominates both, dca and block. */
1825 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1826 block = get_Block_idom(block);
1828 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1829 dca = get_Block_idom(dca);
1832 while (block != dca)
1833 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1838 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1839 * I.e., DCA is the block where we might place PRODUCER.
1840 * A data flow edge points from producer to consumer.
1843 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1845 ir_node *block = NULL;
1847 /* Compute the latest block into which we can place a node so that it is
1849 if (get_irn_op(consumer) == op_Phi) {
1850 /* our consumer is a Phi-node, the effective use is in all those
1851 blocks through which the Phi-node reaches producer */
1853 ir_node *phi_block = get_nodes_block(consumer);
1854 irn_arity = get_irn_arity(consumer);
1856 for (i = 0; i < irn_arity; i++) {
1857 if (get_irn_n(consumer, i) == producer) {
1858 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1860 if (! is_Block_unreachable(new_block))
1861 block = calc_dca(block, new_block);
1866 block = get_irn_n(producer, -1);
1869 assert(is_no_Block(consumer));
1870 block = get_nodes_block(consumer);
1873 /* Compute the deepest common ancestor of block and dca. */
1874 return calc_dca(dca, block);
1877 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1879 static INLINE int get_irn_loop_depth(ir_node *n) {
1880 return get_loop_depth(get_irn_loop(n));
1884 * Move n to a block with less loop depth than it's current block. The
1885 * new block must be dominated by early.
1887 * @param n the node that should be moved
1888 * @param early the earliest block we can n move to
1891 move_out_of_loops (ir_node *n, ir_node *early)
1893 ir_node *best, *dca;
1897 /* Find the region deepest in the dominator tree dominating
1898 dca with the least loop nesting depth, but still dominated
1899 by our early placement. */
1900 dca = get_nodes_block(n);
1903 while (dca != early) {
1904 dca = get_Block_idom(dca);
1905 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1906 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1910 if (best != get_nodes_block(n)) {
1912 printf("Moving out of loop: "); DDMN(n);
1913 printf(" Outermost block: "); DDMN(early);
1914 printf(" Best block: "); DDMN(best);
1915 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1917 set_nodes_block(n, best);
1922 * Find the latest legal block for N and place N into the
1923 * `optimal' Block between the latest and earliest legal block.
1924 * The `optimal' block is the dominance-deepest block of those
1925 * with the least loop-nesting-depth. This places N out of as many
1926 * loops as possible and then makes it as control dependent as
1930 place_floats_late(ir_node *n, pdeq *worklist)
1935 assert(irn_not_visited(n)); /* no multiple placement */
1937 mark_irn_visited(n);
1939 /* no need to place block nodes, control nodes are already placed. */
1940 if ((get_irn_op(n) != op_Block) &&
1942 (get_irn_mode(n) != mode_X)) {
1943 /* Remember the early_blk placement of this block to move it
1944 out of loop no further than the early_blk placement. */
1945 early_blk = get_irn_n(n, -1);
1948 * BEWARE: Here we also get code, that is live, but
1949 * was in a dead block. If the node is life, but because
1950 * of CSE in a dead block, we still might need it.
1953 /* Assure that our users are all placed, except the Phi-nodes.
1954 --- Each data flow cycle contains at least one Phi-node. We
1955 have to break the `user has to be placed before the
1956 producer' dependence cycle and the Phi-nodes are the
1957 place to do so, because we need to base our placement on the
1958 final region of our users, which is OK with Phi-nodes, as they
1959 are op_pin_state_pinned, and they never have to be placed after a
1960 producer of one of their inputs in the same block anyway. */
1961 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1962 ir_node *succ = get_irn_out(n, i);
1963 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1964 place_floats_late(succ, worklist);
1967 if (! is_Block_dead(early_blk)) {
1968 /* do only move things that where not dead */
1970 /* We have to determine the final block of this node... except for
1972 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1973 (get_irn_op(n) != op_Const) &&
1974 (get_irn_op(n) != op_SymConst)) {
1975 ir_node *dca = NULL; /* deepest common ancestor in the
1976 dominator tree of all nodes'
1977 blocks depending on us; our final
1978 placement has to dominate DCA. */
1979 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1980 ir_node *succ = get_irn_out(n, i);
1983 if (get_irn_op(succ) == op_End) {
1985 * This consumer is the End node, a keep alive edge.
1986 * This is not a real consumer, so we ignore it
1991 /* ignore if succ is in dead code */
1992 succ_blk = get_irn_n(succ, -1);
1993 if (is_Block_unreachable(succ_blk))
1995 dca = consumer_dom_dca(dca, succ, n);
1998 set_nodes_block(n, dca);
1999 move_out_of_loops(n, early_blk);
2005 /* Add predecessors of all non-floating nodes on list. (Those of floating
2006 nodes are placed already and therefore are marked.) */
2007 for (i = 0; i < get_irn_n_outs(n); i++) {
2008 ir_node *succ = get_irn_out(n, i);
2009 if (irn_not_visited(get_irn_out(n, i))) {
2010 pdeq_putr(worklist, succ);
2015 static INLINE void place_late(pdeq *worklist) {
2017 inc_irg_visited(current_ir_graph);
2019 /* This fills the worklist initially. */
2020 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2022 /* And now empty the worklist again... */
2023 while (!pdeq_empty(worklist)) {
2024 ir_node *n = pdeq_getl(worklist);
2025 if (irn_not_visited(n))
2026 place_floats_late(n, worklist);
2030 void place_code(ir_graph *irg) {
2032 ir_graph *rem = current_ir_graph;
2034 current_ir_graph = irg;
2036 if (!(get_opt_optimize() && get_opt_global_cse())) return;
2038 /* Handle graph state */
2039 assert(get_irg_phase_state(irg) != phase_building);
2042 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2043 free_loop_information(irg);
2044 construct_backedges(irg);
2047 /* Place all floating nodes as early as possible. This guarantees
2048 a legal code placement. */
2049 worklist = new_pdeq();
2050 place_early(worklist);
2052 /* place_early() invalidates the outs, place_late needs them. */
2053 compute_irg_outs(irg);
2055 /* Now move the nodes down in the dominator tree. This reduces the
2056 unnecessary executions of the node. */
2057 place_late(worklist);
2059 set_irg_outs_inconsistent(current_ir_graph);
2060 set_irg_loopinfo_inconsistent(current_ir_graph);
2062 current_ir_graph = rem;
2066 * Called by walker of remove_critical_cf_edges().
2068 * Place an empty block to an edge between a blocks of multiple
2069 * predecessors and a block of multiple successors.
2072 * @param env Environment of walker. The changed field.
2074 static void walk_critical_cf_edges(ir_node *n, void *env) {
2076 ir_node *pre, *block, *jmp;
2079 /* Block has multiple predecessors */
2080 if (is_Block(n) && (get_irn_arity(n) > 1)) {
2081 if (n == get_irg_end_block(current_ir_graph))
2082 return; /* No use to add a block here. */
2084 arity = get_irn_arity(n);
2085 for (i=0; i<arity; i++) {
2086 pre = get_irn_n(n, i);
2087 /* Predecessor has multiple successors. Insert new control flow edge. */
2088 if (op_Raise != get_irn_op(skip_Proj(pre))) {
2089 /* set predecessor of new block */
2090 block = new_Block(1, &pre);
2091 /* insert new jmp node to new block */
2092 set_cur_block(block);
2095 /* set successor of new block */
2096 set_irn_n(n, i, jmp);
2098 } /* predecessor has multiple successors */
2099 } /* for all predecessors */
2100 } /* n is a block */
2103 void remove_critical_cf_edges(ir_graph *irg) {
2105 irg_walk_graph(irg, NULL, walk_critical_cf_edges, &changed);
2108 /* control flow changed */
2109 set_irg_outs_inconsistent(irg);
2110 set_irg_extblk_inconsistent(irg);
2111 set_irg_doms_inconsistent(current_ir_graph);
2112 set_irg_loopinfo_inconsistent(current_ir_graph);