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, Michael Beck
9 * Copyright: (c) 1998-2006 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
19 #include "irgraph_t.h"
31 #include "pdeq.h" /* Fuer code placement */
36 #include "irbackedge_t.h"
43 #include "iredges_t.h"
46 /*------------------------------------------------------------------*/
47 /* apply optimizations of iropt to all nodes. */
48 /*------------------------------------------------------------------*/
51 * A wrapper around optimize_inplace_2() to be called from a walker.
53 static void optimize_in_place_wrapper (ir_node *n, void *env) {
54 ir_node *optimized = optimize_in_place_2(n);
55 if (optimized != n) exchange (n, optimized);
59 * Do local optimizations for a node.
61 * @param n the IR-node where to start. Typically the End node
64 * @note current_ir_graph must be set
66 static INLINE void do_local_optimize(ir_node *n) {
67 /* Handle graph state */
68 assert(get_irg_phase_state(current_ir_graph) != phase_building);
70 if (get_opt_global_cse())
71 set_irg_pinned(current_ir_graph, op_pin_state_floats);
72 set_irg_outs_inconsistent(current_ir_graph);
73 set_irg_doms_inconsistent(current_ir_graph);
74 set_irg_loopinfo_inconsistent(current_ir_graph);
76 /* Clean the value_table in irg for the CSE. */
77 del_identities(current_ir_graph->value_table);
78 current_ir_graph->value_table = new_identities();
80 /* walk over the graph */
81 irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
84 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
85 void local_optimize_node(ir_node *n) {
86 ir_graph *rem = current_ir_graph;
87 current_ir_graph = get_irn_irg(n);
91 current_ir_graph = rem;
95 * Block-Walker: uses dominance depth to mark dead blocks.
97 static void kill_dead_blocks(ir_node *block, void *env)
99 if (get_Block_dom_depth(block) < 0) {
101 * Note that the new dominance code correctly handles
102 * the End block, i.e. it is always reachable from Start
104 set_Block_dead(block);
108 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n. */
109 void local_optimize_graph(ir_graph *irg) {
110 ir_graph *rem = current_ir_graph;
111 current_ir_graph = irg;
113 if (get_irg_dom_state(irg) == dom_consistent)
114 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
116 do_local_optimize(get_irg_end(irg));
118 current_ir_graph = rem;
122 * Enqueue all users of a node to a wait queue.
123 * Handles mode_T nodes.
125 static void enqueue_users(ir_node *n, pdeq *waitq) {
126 const ir_edge_t *edge;
128 foreach_out_edge(n, edge) {
129 ir_node *succ = get_edge_src_irn(edge);
131 if (get_irn_link(succ) != waitq) {
132 pdeq_putr(waitq, succ);
133 set_irn_link(succ, waitq);
135 if (get_irn_mode(succ) == mode_T) {
136 /* A mode_T node has Proj's. Because most optimizations
137 run on the Proj's we have to enqueue them also. */
138 enqueue_users(succ, waitq);
144 * Data flow optimization walker.
145 * Optimizes all nodes and enqueue it's users
148 static void opt_walker(ir_node *n, void *env) {
152 optimized = optimize_in_place_2(n);
153 set_irn_link(optimized, NULL);
155 if (optimized != n) {
156 enqueue_users(n, waitq);
157 exchange(n, optimized);
161 /* Applies local optimizations to all nodes in the graph until fixpoint. */
162 void optimize_graph_df(ir_graph *irg) {
163 pdeq *waitq = new_pdeq();
164 int state = edges_activated(irg);
165 ir_graph *rem = current_ir_graph;
167 current_ir_graph = irg;
172 if (get_opt_global_cse())
173 set_irg_pinned(current_ir_graph, op_pin_state_floats);
175 /* Clean the value_table in irg for the CSE. */
176 del_identities(irg->value_table);
177 irg->value_table = new_identities();
179 if (get_irg_dom_state(irg) == dom_consistent)
180 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
182 /* invalidate info */
183 set_irg_outs_inconsistent(irg);
184 set_irg_doms_inconsistent(irg);
185 set_irg_loopinfo_inconsistent(irg);
187 /* walk over the graph */
188 irg_walk_graph(irg, NULL, opt_walker, waitq);
190 /* finish the wait queue */
191 while (! pdeq_empty(waitq)) {
192 ir_node *n = pdeq_getl(waitq);
194 opt_walker(n, waitq);
200 edges_deactivate(irg);
202 current_ir_graph = rem;
206 /*------------------------------------------------------------------*/
207 /* Routines for dead node elimination / copying garbage collection */
208 /* of the obstack. */
209 /*------------------------------------------------------------------*/
212 * Remember the new node in the old node by using a field all nodes have.
214 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
217 * Get this new node, before the old node is forgotten.
219 #define get_new_node(oldn) get_irn_link(oldn)
222 * Check if a new node was set.
224 #define has_new_node(n) (get_new_node(n) != NULL)
227 * We use the block_visited flag to mark that we have computed the
228 * number of useful predecessors for this block.
229 * Further we encode the new arity in this flag in the old blocks.
230 * Remembering the arity is useful, as it saves a lot of pointer
231 * accesses. This function is called for all Phi and Block nodes
235 compute_new_arity(ir_node *b) {
236 int i, res, irn_arity;
239 irg_v = get_irg_block_visited(current_ir_graph);
240 block_v = get_Block_block_visited(b);
241 if (block_v >= irg_v) {
242 /* we computed the number of preds for this block and saved it in the
244 return block_v - irg_v;
246 /* compute the number of good predecessors */
247 res = irn_arity = get_irn_arity(b);
248 for (i = 0; i < irn_arity; i++)
249 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
250 /* save it in the flag. */
251 set_Block_block_visited(b, irg_v + res);
257 * Copies the node to the new obstack. The Ins of the new node point to
258 * the predecessors on the old obstack. For block/phi nodes not all
259 * predecessors might be copied. n->link points to the new node.
260 * For Phi and Block nodes the function allocates in-arrays with an arity
261 * only for useful predecessors. The arity is determined by counting
262 * the non-bad predecessors of the block.
264 * @param n The node to be copied
265 * @param env if non-NULL, the node number attribute will be copied to the new node
267 * Note: Also used for loop unrolling.
269 static void copy_node(ir_node *n, void *env) {
272 ir_op *op = get_irn_op(n);
273 int copy_node_nr = env != NULL;
275 /* The end node looses it's flexible in array. This doesn't matter,
276 as dead node elimination builds End by hand, inlineing doesn't use
278 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
281 /* node copied already */
283 } else if (op == op_Block) {
285 new_arity = compute_new_arity(n);
286 n->attr.block.graph_arr = NULL;
288 block = get_nodes_block(n);
290 new_arity = compute_new_arity(block);
292 new_arity = get_irn_arity(n);
295 nn = new_ir_node(get_irn_dbg_info(n),
302 /* Copy the attributes. These might point to additional data. If this
303 was allocated on the old obstack the pointers now are dangling. This
304 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
305 copy_node_attr(n, nn);
306 new_backedge_info(nn);
310 /* for easier debugging, we want to copy the node numbers too */
311 nn->node_nr = n->node_nr;
316 hook_dead_node_elim_subst(current_ir_graph, n, nn);
320 * Copies new predecessors of old node to new node remembered in link.
321 * Spare the Bad predecessors of Phi and Block nodes.
324 copy_preds(ir_node *n, void *env) {
328 nn = get_new_node(n);
330 /* printf("\n old node: "); DDMSG2(n);
331 printf(" new node: "); DDMSG2(nn);
332 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
335 /* Don't copy Bad nodes. */
337 irn_arity = get_irn_arity(n);
338 for (i = 0; i < irn_arity; i++)
339 if (! is_Bad(get_irn_n(n, i))) {
340 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
341 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
344 /* repair the block visited flag from above misuse. Repair it in both
345 graphs so that the old one can still be used. */
346 set_Block_block_visited(nn, 0);
347 set_Block_block_visited(n, 0);
348 /* Local optimization could not merge two subsequent blocks if
349 in array contained Bads. Now it's possible.
350 We don't call optimize_in_place as it requires
351 that the fields in ir_graph are set properly. */
352 if ((get_opt_control_flow_straightening()) &&
353 (get_Block_n_cfgpreds(nn) == 1) &&
354 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
355 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
357 /* Jmp jumps into the block it is in -- deal self cycle. */
358 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
359 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
364 } else if (get_irn_op(n) == op_Phi) {
365 /* Don't copy node if corresponding predecessor in block is Bad.
366 The Block itself should not be Bad. */
367 block = get_nodes_block(n);
368 set_irn_n(nn, -1, get_new_node(block));
370 irn_arity = get_irn_arity(n);
371 for (i = 0; i < irn_arity; i++)
372 if (! is_Bad(get_irn_n(block, i))) {
373 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
374 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
377 /* If the pre walker reached this Phi after the post walker visited the
378 block block_visited is > 0. */
379 set_Block_block_visited(get_nodes_block(n), 0);
380 /* Compacting the Phi's ins might generate Phis with only one
382 if (get_irn_arity(nn) == 1)
383 exchange(nn, get_irn_n(nn, 0));
385 irn_arity = get_irn_arity(n);
386 for (i = -1; i < irn_arity; i++)
387 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
389 /* Now the new node is complete. We can add it to the hash table for CSE.
390 @@@ inlining aborts if we identify End. Why? */
391 if (get_irn_op(nn) != op_End)
392 add_identities(current_ir_graph->value_table, nn);
396 * Copies the graph recursively, compacts the keep-alives of the end node.
398 * @param irg the graph to be copied
399 * @param copy_node_nr If non-zero, the node number will be copied
401 static void copy_graph(ir_graph *irg, int copy_node_nr) {
402 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
403 ir_node *ka; /* keep alive */
407 /* Some nodes must be copied by hand, sigh */
408 vfl = get_irg_visited(irg);
409 set_irg_visited(irg, vfl + 1);
411 oe = get_irg_end(irg);
412 mark_irn_visited(oe);
413 /* copy the end node by hand, allocate dynamic in array! */
414 ne = new_ir_node(get_irn_dbg_info(oe),
421 /* Copy the attributes. Well, there might be some in the future... */
422 copy_node_attr(oe, ne);
423 set_new_node(oe, ne);
425 /* copy the Bad node */
426 ob = get_irg_bad(irg);
427 mark_irn_visited(ob);
428 nb = new_ir_node(get_irn_dbg_info(ob),
435 copy_node_attr(ob, nb);
436 set_new_node(ob, nb);
438 /* copy the NoMem node */
439 om = get_irg_no_mem(irg);
440 mark_irn_visited(om);
441 nm = new_ir_node(get_irn_dbg_info(om),
448 copy_node_attr(om, nm);
449 set_new_node(om, nm);
451 /* copy the live nodes */
452 set_irg_visited(irg, vfl);
453 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
455 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
457 /* visit the anchors as well */
458 for (i = anchor_max - 1; i >= 0; --i) {
459 ir_node *n = irg->anchors[i];
461 if (n && (get_irn_visited(n) <= vfl)) {
462 set_irg_visited(irg, vfl);
463 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
467 /* copy_preds for the end node ... */
468 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
470 /*- ... and now the keep alives. -*/
471 /* First pick the not marked block nodes and walk them. We must pick these
472 first as else we will oversee blocks reachable from Phis. */
473 irn_arity = get_End_n_keepalives(oe);
474 for (i = 0; i < irn_arity; i++) {
475 ka = get_End_keepalive(oe, i);
477 if (get_irn_visited(ka) <= vfl) {
478 /* We must keep the block alive and copy everything reachable */
479 set_irg_visited(irg, vfl);
480 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
482 add_End_keepalive(ne, get_new_node(ka));
486 /* Now pick other nodes. Here we will keep all! */
487 irn_arity = get_End_n_keepalives(oe);
488 for (i = 0; i < irn_arity; i++) {
489 ka = get_End_keepalive(oe, i);
491 if (get_irn_visited(ka) <= vfl) {
492 /* We didn't copy the node yet. */
493 set_irg_visited(irg, vfl);
494 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
496 add_End_keepalive(ne, get_new_node(ka));
500 /* start block sometimes only reached after keep alives */
501 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
502 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
506 * Copies the graph reachable from current_ir_graph->end to the obstack
507 * in current_ir_graph and fixes the environment.
508 * Then fixes the fields in current_ir_graph containing nodes of the
511 * @param copy_node_nr If non-zero, the node number will be copied
514 copy_graph_env(int copy_node_nr) {
515 ir_graph *irg = current_ir_graph;
516 ir_node *old_end, *n;
519 /* remove end_except and end_reg nodes */
520 old_end = get_irg_end(irg);
521 set_irg_end_except (irg, old_end);
522 set_irg_end_reg (irg, old_end);
524 /* Not all nodes remembered in irg might be reachable
525 from the end node. Assure their link is set to NULL, so that
526 we can test whether new nodes have been computed. */
527 for (i = anchor_max - 1; i >= 0; --i) {
529 set_new_node(irg->anchors[i], NULL);
531 /* we use the block walk flag for removing Bads from Blocks ins. */
532 inc_irg_block_visited(irg);
535 copy_graph(irg, copy_node_nr);
537 /* fix the fields in irg */
538 old_end = get_irg_end(irg);
539 for (i = anchor_max - 1; i >= 0; --i) {
542 irg->anchors[i] = get_new_node(n);
548 * Copies all reachable nodes to a new obstack. Removes bad inputs
549 * from block nodes and the corresponding inputs from Phi nodes.
550 * Merges single exit blocks with single entry blocks and removes
552 * Adds all new nodes to a new hash table for CSE. Does not
553 * perform CSE, so the hash table might contain common subexpressions.
556 dead_node_elimination(ir_graph *irg) {
557 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
559 int rem_ipview = get_interprocedural_view();
560 struct obstack *graveyard_obst = NULL;
561 struct obstack *rebirth_obst = NULL;
562 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
564 /* inform statistics that we started a dead-node elimination run */
565 hook_dead_node_elim(irg, 1);
567 /* Remember external state of current_ir_graph. */
568 rem = current_ir_graph;
569 current_ir_graph = irg;
570 set_interprocedural_view(0);
572 assert(get_irg_phase_state(irg) != phase_building);
574 /* Handle graph state */
575 free_callee_info(irg);
579 /* @@@ so far we loose loops when copying */
580 free_loop_information(irg);
582 set_irg_doms_inconsistent(irg);
584 /* A quiet place, where the old obstack can rest in peace,
585 until it will be cremated. */
586 graveyard_obst = irg->obst;
588 /* A new obstack, where the reachable nodes will be copied to. */
589 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
590 irg->obst = rebirth_obst;
591 obstack_init(irg->obst);
592 irg->last_node_idx = 0;
594 /* We also need a new value table for CSE */
595 del_identities(irg->value_table);
596 irg->value_table = new_identities();
598 /* Copy the graph from the old to the new obstack */
599 copy_graph_env(/*copy_node_nr=*/1);
601 /* Free memory from old unoptimized obstack */
602 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
603 xfree (graveyard_obst); /* ... then free it. */
605 /* inform statistics that the run is over */
606 hook_dead_node_elim(irg, 0);
608 current_ir_graph = rem;
609 set_interprocedural_view(rem_ipview);
614 * Relink bad predecessors of a block and store the old in array to the
615 * link field. This function is called by relink_bad_predecessors().
616 * The array of link field starts with the block operand at position 0.
617 * If block has bad predecessors, create a new in array without bad preds.
618 * Otherwise let in array untouched.
620 static void relink_bad_block_predecessors(ir_node *n, void *env) {
621 ir_node **new_in, *irn;
622 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
624 /* if link field of block is NULL, look for bad predecessors otherwise
625 this is already done */
626 if (get_irn_op(n) == op_Block &&
627 get_irn_link(n) == NULL) {
629 /* save old predecessors in link field (position 0 is the block operand)*/
630 set_irn_link(n, get_irn_in(n));
632 /* count predecessors without bad nodes */
633 old_irn_arity = get_irn_arity(n);
634 for (i = 0; i < old_irn_arity; i++)
635 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
637 /* arity changing: set new predecessors without bad nodes */
638 if (new_irn_arity < old_irn_arity) {
639 /* Get new predecessor array. We do not resize the array, as we must
640 keep the old one to update Phis. */
641 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
643 /* set new predecessors in array */
646 for (i = 0; i < old_irn_arity; i++) {
647 irn = get_irn_n(n, i);
649 new_in[new_irn_n] = irn;
650 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
654 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
655 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
658 } /* ir node has bad predecessors */
660 } /* Block is not relinked */
664 * Relinks Bad predecessors from Blocks and Phis called by walker
665 * remove_bad_predecesors(). If n is a Block, call
666 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
667 * function of Phi's Block. If this block has bad predecessors, relink preds
670 static void relink_bad_predecessors(ir_node *n, void *env) {
671 ir_node *block, **old_in;
672 int i, old_irn_arity, new_irn_arity;
674 /* relink bad predecessors of a block */
675 if (get_irn_op(n) == op_Block)
676 relink_bad_block_predecessors(n, env);
678 /* If Phi node relink its block and its predecessors */
679 if (get_irn_op(n) == op_Phi) {
681 /* Relink predecessors of phi's block */
682 block = get_nodes_block(n);
683 if (get_irn_link(block) == NULL)
684 relink_bad_block_predecessors(block, env);
686 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
687 old_irn_arity = ARR_LEN(old_in);
689 /* Relink Phi predecessors if count of predecessors changed */
690 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
691 /* set new predecessors in array
692 n->in[0] remains the same block */
694 for(i = 1; i < old_irn_arity; i++)
695 if (!is_Bad((ir_node *)old_in[i])) {
696 n->in[new_irn_arity] = n->in[i];
697 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
701 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
702 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
705 } /* n is a Phi node */
709 * Removes Bad Bad predecessors from Blocks and the corresponding
710 * inputs to Phi nodes as in dead_node_elimination but without
712 * On walking up set the link field to NULL, on walking down call
713 * relink_bad_predecessors() (This function stores the old in array
714 * to the link field and sets a new in array if arity of predecessors
717 void remove_bad_predecessors(ir_graph *irg) {
718 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
725 __)|_| | \_/ | \_/(/_ |_/\__|__
727 The following stuff implements a facility that automatically patches
728 registered ir_node pointers to the new node when a dead node elimination occurs.
731 struct _survive_dce_t {
735 hook_entry_t dead_node_elim;
736 hook_entry_t dead_node_elim_subst;
739 typedef struct _survive_dce_list_t {
740 struct _survive_dce_list_t *next;
742 } survive_dce_list_t;
744 static void dead_node_hook(void *context, ir_graph *irg, int start)
746 survive_dce_t *sd = context;
748 /* Create a new map before the dead node elimination is performed. */
750 sd->new_places = pmap_create_ex(pmap_count(sd->places));
753 /* Patch back all nodes if dead node elimination is over and something is to be done. */
755 pmap_destroy(sd->places);
756 sd->places = sd->new_places;
757 sd->new_places = NULL;
762 * Hook called when dead node elimination replaces old by nw.
764 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw)
766 survive_dce_t *sd = context;
767 survive_dce_list_t *list = pmap_get(sd->places, old);
769 /* If the node is to be patched back, write the new address to all registered locations. */
771 survive_dce_list_t *p;
773 for(p = list; p; p = p->next)
776 pmap_insert(sd->new_places, nw, list);
781 * Make a new Survive DCE environment.
783 survive_dce_t *new_survive_dce(void)
785 survive_dce_t *res = xmalloc(sizeof(res[0]));
786 obstack_init(&res->obst);
787 res->places = pmap_create();
788 res->new_places = NULL;
790 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
791 res->dead_node_elim.context = res;
792 res->dead_node_elim.next = NULL;
794 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
795 res->dead_node_elim_subst.context = res;
796 res->dead_node_elim_subst.next = NULL;
798 register_hook(hook_dead_node_elim, &res->dead_node_elim);
799 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
804 * Free a Survive DCE environment.
806 void free_survive_dce(survive_dce_t *sd)
808 obstack_free(&sd->obst, NULL);
809 pmap_destroy(sd->places);
810 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
811 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
816 * Register a node pointer to be patched upon DCE.
817 * When DCE occurs, the node pointer specified by @p place will be
818 * patched to the new address of the node it is pointing to.
820 * @param sd The Survive DCE environment.
821 * @param place The address of the node pointer.
823 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place)
826 ir_node *irn = *place;
827 survive_dce_list_t *curr = pmap_get(sd->places, irn);
828 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw));
833 pmap_insert(sd->places, irn, nw);
837 /*--------------------------------------------------------------------*/
838 /* Functionality for inlining */
839 /*--------------------------------------------------------------------*/
842 * Copy node for inlineing. Updates attributes that change when
843 * inlineing but not for dead node elimination.
845 * Copies the node by calling copy_node() and then updates the entity if
846 * it's a local one. env must be a pointer of the frame type of the
847 * inlined procedure. The new entities must be in the link field of
851 copy_node_inline (ir_node *n, void *env) {
853 ir_type *frame_tp = (ir_type *)env;
856 if (get_irn_op(n) == op_Sel) {
857 nn = get_new_node (n);
859 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
860 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
862 } else if (get_irn_op(n) == op_Block) {
863 nn = get_new_node (n);
864 nn->attr.block.irg = current_ir_graph;
869 * Walker: checks if P_value_arg_base is used.
871 static void find_addr(ir_node *node, void *env) {
872 int *allow_inline = env;
873 if (is_Proj(node) && get_irn_op(get_Proj_pred(node)) == op_Start) {
874 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
880 * currently, we cannot inline two cases:
881 * - call with compound arguments
882 * - graphs that take the address of a parameter
884 * check these conditions here
886 static int can_inline(ir_node *call, ir_graph *called_graph)
888 ir_type *call_type = get_Call_type(call);
889 int params, ress, i, res;
890 assert(is_Method_type(call_type));
892 params = get_method_n_params(call_type);
893 ress = get_method_n_ress(call_type);
895 /* check parameters for compound arguments */
896 for (i = 0; i < params; ++i) {
897 ir_type *p_type = get_method_param_type(call_type, i);
899 if (is_compound_type(p_type))
903 /* check results for compound arguments */
904 for (i = 0; i < ress; ++i) {
905 ir_type *r_type = get_method_res_type(call_type, i);
907 if (is_compound_type(r_type))
912 irg_walk_graph(called_graph, find_addr, NULL, &res);
917 /* Inlines a method at the given call site. */
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 if (called_graph == current_ir_graph) {
971 set_optimize(rem_opt);
975 /* here we know we WILL inline, so inform the statistics */
976 hook_inline(call, called_graph);
978 /* -- Decide how to handle exception control flow: Is there a handler
979 for the Call node, or do we branch directly to End on an exception?
981 0 There is a handler.
983 2 Exception handling not represented in Firm. -- */
985 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
986 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
987 assert(is_Proj(proj));
988 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
989 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
991 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
992 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
993 else { exc_handling = 2; } /* !Mproj && !Xproj */
998 the procedure and later replaces the Start node of the called graph.
999 Post_call is the old Call node and collects the results of the called
1000 graph. Both will end up being a tuple. -- */
1001 post_bl = get_nodes_block(call);
1002 set_irg_current_block(current_ir_graph, post_bl);
1003 /* XxMxPxPxPxT of Start + parameter of Call */
1004 in[pn_Start_X_initial_exec] = new_Jmp();
1005 in[pn_Start_M] = get_Call_mem(call);
1006 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1007 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1008 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1009 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1010 /* in[pn_Start_P_value_arg_base] = ??? */
1011 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1012 pre_call = new_Tuple(pn_Start_max - 1, in);
1016 The new block gets the ins of the old block, pre_call and all its
1017 predecessors and all Phi nodes. -- */
1018 part_block(pre_call);
1020 /* -- Prepare state for dead node elimination -- */
1021 /* Visited flags in calling irg must be >= flag in called irg.
1022 Else walker and arity computation will not work. */
1023 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1024 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1025 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1026 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1027 /* Set pre_call as new Start node in link field of the start node of
1028 calling graph and pre_calls block as new block for the start block
1030 Further mark these nodes so that they are not visited by the
1032 set_irn_link(get_irg_start(called_graph), pre_call);
1033 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1034 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1035 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1036 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1037 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1039 /* Initialize for compaction of in arrays */
1040 inc_irg_block_visited(current_ir_graph);
1042 /* -- Replicate local entities of the called_graph -- */
1043 /* copy the entities. */
1044 called_frame = get_irg_frame_type(called_graph);
1045 for (i = 0; i < get_class_n_members(called_frame); i++) {
1046 entity *new_ent, *old_ent;
1047 old_ent = get_class_member(called_frame, i);
1048 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1049 set_entity_link(old_ent, new_ent);
1052 /* visited is > than that of called graph. With this trick visited will
1053 remain unchanged so that an outer walker, e.g., searching the call nodes
1054 to inline, calling this inline will not visit the inlined nodes. */
1055 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1057 /* -- Performing dead node elimination inlines the graph -- */
1058 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1060 /* @@@ endless loops are not copied!! -- they should be, I think... */
1061 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1062 get_irg_frame_type(called_graph));
1064 /* Repair called_graph */
1065 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1066 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1067 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1069 /* -- Merge the end of the inlined procedure with the call site -- */
1070 /* We will turn the old Call node into a Tuple with the following
1073 0: Phi of all Memories of Return statements.
1074 1: Jmp from new Block that merges the control flow from all exception
1075 predecessors of the old end block.
1076 2: Tuple of all arguments.
1077 3: Phi of Exception memories.
1078 In case the old Call directly branches to End on an exception we don't
1079 need the block merging all exceptions nor the Phi of the exception
1083 /* -- Precompute some values -- */
1084 end_bl = get_new_node(get_irg_end_block(called_graph));
1085 end = get_new_node(get_irg_end(called_graph));
1086 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1087 n_res = get_method_n_ress(get_Call_type(call));
1089 res_pred = xmalloc (n_res * sizeof(*res_pred));
1090 cf_pred = xmalloc (arity * sizeof(*res_pred));
1092 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1094 /* -- archive keepalives -- */
1095 irn_arity = get_irn_arity(end);
1096 for (i = 0; i < irn_arity; i++)
1097 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
1099 /* The new end node will die. We need not free as the in array is on the obstack:
1100 copy_node() only generated 'D' arrays. */
1102 /* -- Replace Return nodes by Jump nodes. -- */
1104 for (i = 0; i < arity; i++) {
1106 ret = get_irn_n(end_bl, i);
1107 if (is_Return(ret)) {
1108 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1112 set_irn_in(post_bl, n_ret, cf_pred);
1114 /* -- Build a Tuple for all results of the method.
1115 Add Phi node if there was more than one Return. -- */
1116 turn_into_tuple(post_call, 4);
1117 /* First the Memory-Phi */
1119 for (i = 0; i < arity; i++) {
1120 ret = get_irn_n(end_bl, i);
1121 if (is_Return(ret)) {
1122 cf_pred[n_ret] = get_Return_mem(ret);
1126 phi = new_Phi(n_ret, cf_pred, mode_M);
1127 set_Tuple_pred(call, pn_Call_M_regular, phi);
1128 /* Conserve Phi-list for further inlinings -- but might be optimized */
1129 if (get_nodes_block(phi) == post_bl) {
1130 set_irn_link(phi, get_irn_link(post_bl));
1131 set_irn_link(post_bl, phi);
1133 /* Now the real results */
1135 for (j = 0; j < n_res; j++) {
1137 for (i = 0; i < arity; i++) {
1138 ret = get_irn_n(end_bl, i);
1139 if (get_irn_op(ret) == op_Return) {
1140 cf_pred[n_ret] = get_Return_res(ret, j);
1145 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1149 /* Conserve Phi-list for further inlinings -- but might be optimized */
1150 if (get_nodes_block(phi) == post_bl) {
1151 set_irn_link(phi, get_irn_link(post_bl));
1152 set_irn_link(post_bl, phi);
1155 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1157 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1159 /* Finally the exception control flow.
1160 We have two (three) possible situations:
1161 First if the Call branches to an exception handler: We need to add a Phi node to
1162 collect the memory containing the exception objects. Further we need
1163 to add another block to get a correct representation of this Phi. To
1164 this block we add a Jmp that resolves into the X output of the Call
1165 when the Call is turned into a tuple.
1166 Second the Call branches to End, the exception is not handled. Just
1167 add all inlined exception branches to the End node.
1168 Third: there is no Exception edge at all. Handle as case two. */
1169 if (exc_handling == 0) {
1171 for (i = 0; i < arity; i++) {
1173 ret = get_irn_n(end_bl, i);
1174 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1175 cf_pred[n_exc] = ret;
1180 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1181 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1182 /* The Phi for the memories with the exception objects */
1184 for (i = 0; i < arity; i++) {
1186 ret = skip_Proj(get_irn_n(end_bl, i));
1188 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1190 } else if (is_fragile_op(ret)) {
1191 /* We rely that all cfops have the memory output at the same position. */
1192 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1194 } else if (get_irn_op(ret) == op_Raise) {
1195 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1199 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1201 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1202 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1205 ir_node *main_end_bl;
1206 int main_end_bl_arity;
1207 ir_node **end_preds;
1209 /* assert(exc_handling == 1 || no exceptions. ) */
1211 for (i = 0; i < arity; i++) {
1212 ir_node *ret = get_irn_n(end_bl, i);
1214 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1215 cf_pred[n_exc] = ret;
1219 main_end_bl = get_irg_end_block(current_ir_graph);
1220 main_end_bl_arity = get_irn_arity(main_end_bl);
1221 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1223 for (i = 0; i < main_end_bl_arity; ++i)
1224 end_preds[i] = get_irn_n(main_end_bl, i);
1225 for (i = 0; i < n_exc; ++i)
1226 end_preds[main_end_bl_arity + i] = cf_pred[i];
1227 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1228 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1229 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1235 /* -- Turn CSE back on. -- */
1236 set_optimize(rem_opt);
1241 /********************************************************************/
1242 /* Apply inlineing to small methods. */
1243 /********************************************************************/
1245 /** Represents a possible inlinable call in a graph. */
1246 typedef struct _call_entry call_entry;
1247 struct _call_entry {
1248 ir_node *call; /**< the Call */
1249 ir_graph *callee; /**< the callee called here */
1250 call_entry *next; /**< for linking the next one */
1254 * environment for inlining small irgs
1256 typedef struct _inline_env_t {
1257 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1258 call_entry *head; /**< the head of the call entry list */
1259 call_entry *tail; /**< the tail of the call entry list */
1263 * Returns the irg called from a Call node. If the irg is not
1264 * known, NULL is returned.
1266 static ir_graph *get_call_called_irg(ir_node *call) {
1268 ir_graph *called_irg = NULL;
1270 addr = get_Call_ptr(call);
1271 if (is_SymConst(addr) && get_SymConst_kind(addr) == symconst_addr_ent) {
1272 called_irg = get_entity_irg(get_SymConst_entity(addr));
1279 * Walker: Collect all calls to known graphs inside a graph.
1281 static void collect_calls(ir_node *call, void *env) {
1282 if (is_Call(call)) {
1283 ir_graph *called_irg = get_call_called_irg(call);
1285 /* The Call node calls a locally defined method. Remember to inline. */
1286 inline_env_t *ienv = env;
1287 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1289 entry->callee = called_irg;
1292 if (ienv->tail == NULL)
1295 ienv->tail->next = entry;
1302 * Inlines all small methods at call sites where the called address comes
1303 * from a Const node that references the entity representing the called
1305 * The size argument is a rough measure for the code size of the method:
1306 * Methods where the obstack containing the firm graph is smaller than
1309 void inline_small_irgs(ir_graph *irg, int size) {
1310 ir_graph *rem = current_ir_graph;
1313 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1315 if (!(get_opt_optimize() && get_opt_inline())) return;
1317 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1319 current_ir_graph = irg;
1320 /* Handle graph state */
1321 assert(get_irg_phase_state(irg) != phase_building);
1322 free_callee_info(irg);
1324 /* Find Call nodes to inline.
1325 (We can not inline during a walk of the graph, as inlineing the same
1326 method several times changes the visited flag of the walked graph:
1327 after the first inlineing visited of the callee equals visited of
1328 the caller. With the next inlineing both are increased.) */
1329 obstack_init(&env.obst);
1330 env.head = env.tail = NULL;
1331 irg_walk_graph(irg, NULL, collect_calls, &env);
1333 if (env.head != NULL) {
1334 /* There are calls to inline */
1335 collect_phiprojs(irg);
1336 for (entry = env.head; entry != NULL; entry = entry->next) {
1337 ir_graph *callee = entry->callee;
1338 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1339 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1340 inline_method(entry->call, callee);
1344 obstack_free(&env.obst, NULL);
1345 current_ir_graph = rem;
1349 * Environment for inlining irgs.
1352 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1353 int n_nodes_orig; /**< for statistics */
1354 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1355 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1356 int n_call_nodes; /**< Number of Call nodes in the graph. */
1357 int n_call_nodes_orig; /**< for statistics */
1358 int n_callers; /**< Number of known graphs that call this graphs. */
1359 int n_callers_orig; /**< for statistics */
1363 * Allocate a new environment for inlining.
1365 static inline_irg_env *alloc_inline_irg_env(struct obstack *obst) {
1366 inline_irg_env *env = obstack_alloc(obst, sizeof(*env));
1367 env->n_nodes = -2; /* do not count count Start, End */
1368 env->n_nodes_orig = -2; /* do not count Start, End */
1369 env->call_head = NULL;
1370 env->call_tail = NULL;
1371 env->n_call_nodes = 0;
1372 env->n_call_nodes_orig = 0;
1374 env->n_callers_orig = 0;
1378 typedef struct walker_env {
1379 struct obstack *obst; /**< the obstack for allocations. */
1380 inline_irg_env *x; /**< the inline environment */
1381 int ignore_runtime; /**< the ignore runtime flag */
1385 * post-walker: collect all calls in the inline-environment
1386 * of a graph and sum some statistics.
1388 static void collect_calls2(ir_node *call, void *ctx) {
1390 inline_irg_env *x = env->x;
1391 ir_op *op = get_irn_op(call);
1395 /* count meaningful nodes in irg */
1396 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1401 if (op != op_Call) return;
1403 /* check, if it's a runtime call */
1404 if (env->ignore_runtime) {
1405 ir_node *symc = get_Call_ptr(call);
1407 if (is_SymConst(symc) && get_SymConst_kind(symc) == symconst_addr_ent) {
1408 entity *ent = get_SymConst_entity(symc);
1410 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1415 /* collect all call nodes */
1417 ++x->n_call_nodes_orig;
1419 callee = get_call_called_irg(call);
1421 inline_irg_env *callee_env = get_irg_link(callee);
1422 /* count all static callers */
1423 ++callee_env->n_callers;
1424 ++callee_env->n_callers_orig;
1426 /* link it in the list of possible inlinable entries */
1427 entry = obstack_alloc(env->obst, sizeof(*entry));
1429 entry->callee = callee;
1431 if (x->call_tail == NULL)
1432 x->call_head = entry;
1434 x->call_tail->next = entry;
1435 x->call_tail = entry;
1440 * Returns TRUE if the number of callers in 0 in the irg's environment,
1441 * hence this irg is a leave.
1443 INLINE static int is_leave(ir_graph *irg) {
1444 inline_irg_env *env = get_irg_link(irg);
1445 return env->n_call_nodes == 0;
1449 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1451 INLINE static int is_smaller(ir_graph *callee, int size) {
1452 inline_irg_env *env = get_irg_link(callee);
1453 return env->n_nodes < size;
1457 * Append the nodes of the list src to the nodes of the list in environment dst.
1459 static void append_call_list(struct obstack *obst, inline_irg_env *dst, call_entry *src) {
1460 call_entry *entry, *nentry;
1462 /* Note that the src list points to Call nodes in the inlined graph, but
1463 we need Call nodes in our graph. Luckily the inliner leaves this information
1464 in the link field. */
1465 for (entry = src; entry != NULL; entry = entry->next) {
1466 nentry = obstack_alloc(obst, sizeof(*nentry));
1467 nentry->call = get_irn_link(entry->call);
1468 nentry->callee = entry->callee;
1469 nentry->next = NULL;
1470 dst->call_tail->next = nentry;
1471 dst->call_tail = nentry;
1476 * Inlines small leave methods at call sites where the called address comes
1477 * from a Const node that references the entity representing the called
1479 * The size argument is a rough measure for the code size of the method:
1480 * Methods where the obstack containing the firm graph is smaller than
1483 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1484 inline_irg_env *env;
1490 call_entry *entry, *tail;
1491 const call_entry *centry;
1492 struct obstack obst;
1493 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1495 if (!(get_opt_optimize() && get_opt_inline())) return;
1497 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1498 rem = current_ir_graph;
1499 obstack_init(&obst);
1501 /* extend all irgs by a temporary data structure for inlining. */
1502 n_irgs = get_irp_n_irgs();
1503 for (i = 0; i < n_irgs; ++i)
1504 set_irg_link(get_irp_irg(i), alloc_inline_irg_env(&obst));
1506 /* Precompute information in temporary data structure. */
1508 wenv.ignore_runtime = ignore_runtime;
1509 for (i = 0; i < n_irgs; ++i) {
1510 ir_graph *irg = get_irp_irg(i);
1512 assert(get_irg_phase_state(irg) != phase_building);
1513 free_callee_info(irg);
1515 wenv.x = get_irg_link(irg);
1516 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1519 /* -- and now inline. -- */
1521 /* Inline leaves recursively -- we might construct new leaves. */
1525 for (i = 0; i < n_irgs; ++i) {
1527 int phiproj_computed = 0;
1529 current_ir_graph = get_irp_irg(i);
1530 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1533 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1536 if (env->n_nodes > maxsize) break;
1539 callee = entry->callee;
1541 if (is_leave(callee) && is_smaller(callee, leavesize)) {
1542 if (!phiproj_computed) {
1543 phiproj_computed = 1;
1544 collect_phiprojs(current_ir_graph);
1546 did_inline = inline_method(call, callee);
1549 /* Do some statistics */
1550 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1551 --env->n_call_nodes;
1552 env->n_nodes += callee_env->n_nodes;
1553 --callee_env->n_callers;
1555 /* remove this call from the list */
1557 tail->next = entry->next;
1559 env->call_head = entry->next;
1565 env->call_tail = tail;
1567 } while (did_inline);
1569 /* inline other small functions. */
1570 for (i = 0; i < n_irgs; ++i) {
1572 int phiproj_computed = 0;
1574 current_ir_graph = get_irp_irg(i);
1575 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1577 /* note that the list of possible calls is updated during the process */
1579 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1583 callee = entry->callee;
1585 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1586 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1587 if (!phiproj_computed) {
1588 phiproj_computed = 1;
1589 collect_phiprojs(current_ir_graph);
1591 if (inline_method(call, callee)) {
1592 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1593 /* callee was inline. Append it's call list. */
1594 --env->n_call_nodes;
1595 append_call_list(&obst, env, callee_env->call_head);
1596 env->n_call_nodes += callee_env->n_call_nodes;
1597 env->n_nodes += callee_env->n_nodes;
1598 --callee_env->n_callers;
1600 /* after we have inlined callee, all called methods inside callee
1601 are now called once more */
1602 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1603 inline_irg_env *penv = get_irg_link(centry->callee);
1607 /* remove this call from the list */
1609 tail->next = entry->next;
1611 env->call_head = entry->next;
1617 env->call_tail = tail;
1620 for (i = 0; i < n_irgs; ++i) {
1621 irg = get_irp_irg(i);
1622 env = (inline_irg_env *)get_irg_link(irg);
1623 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1624 (env->n_callers_orig != env->n_callers))
1625 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1626 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1627 env->n_callers_orig, env->n_callers,
1628 get_entity_name(get_irg_entity(irg))));
1631 obstack_free(&obst, NULL);
1632 current_ir_graph = rem;
1635 /*******************************************************************/
1636 /* Code Placement. Pins all floating nodes to a block where they */
1637 /* will be executed only if needed. */
1638 /*******************************************************************/
1641 * Returns non-zero, is a block is not reachable from Start.
1643 * @param block the block to test
1646 is_Block_unreachable(ir_node *block) {
1647 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1651 * Find the earliest correct block for N. --- Place N into the
1652 * same Block as its dominance-deepest Input.
1654 * We have to avoid calls to get_nodes_block() here
1655 * because the graph is floating.
1657 * move_out_of_loops() expects that place_floats_early() have placed
1658 * all "living" nodes into a living block. That's why we must
1659 * move nodes in dead block with "live" successors into a valid
1661 * We move them just into the same block as it's successor (or
1662 * in case of a Phi into the effective use block). For Phi successors,
1663 * this may still be a dead block, but then there is no real use, as
1664 * the control flow will be dead later.
1667 place_floats_early(ir_node *n, waitq *worklist)
1671 /* we must not run into an infinite loop */
1672 assert(irn_not_visited(n));
1673 mark_irn_visited(n);
1675 /* Place floating nodes. */
1676 if (get_irn_pinned(n) == op_pin_state_floats) {
1677 ir_node *curr_block = get_irn_n(n, -1);
1678 int in_dead_block = is_Block_unreachable(curr_block);
1680 ir_node *b = NULL; /* The block to place this node in */
1682 assert(is_no_Block(n));
1684 if (is_irn_start_block_placed(n)) {
1685 /* These nodes will not be placed by the loop below. */
1686 b = get_irg_start_block(current_ir_graph);
1690 /* find the block for this node. */
1691 irn_arity = get_irn_arity(n);
1692 for (i = 0; i < irn_arity; i++) {
1693 ir_node *pred = get_irn_n(n, i);
1694 ir_node *pred_block;
1696 if ((irn_not_visited(pred))
1697 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1700 * If the current node is NOT in a dead block, but one of its
1701 * predecessors is, we must move the predecessor to a live block.
1702 * Such thing can happen, if global CSE chose a node from a dead block.
1703 * We move it simply to our block.
1704 * Note that neither Phi nor End nodes are floating, so we don't
1705 * need to handle them here.
1707 if (! in_dead_block) {
1708 if (get_irn_pinned(pred) == op_pin_state_floats &&
1709 is_Block_unreachable(get_irn_n(pred, -1)))
1710 set_nodes_block(pred, curr_block);
1712 place_floats_early(pred, worklist);
1716 * A node in the Bad block must stay in the bad block,
1717 * so don't compute a new block for it.
1722 /* Because all loops contain at least one op_pin_state_pinned node, now all
1723 our inputs are either op_pin_state_pinned or place_early() has already
1724 been finished on them. We do not have any unfinished inputs! */
1725 pred_block = get_irn_n(pred, -1);
1726 if ((!is_Block_dead(pred_block)) &&
1727 (get_Block_dom_depth(pred_block) > depth)) {
1729 depth = get_Block_dom_depth(pred_block);
1731 /* Avoid that the node is placed in the Start block */
1732 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1733 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1734 assert(b != get_irg_start_block(current_ir_graph));
1739 set_nodes_block(n, b);
1743 * Add predecessors of non floating nodes and non-floating predecessors
1744 * of floating nodes to worklist and fix their blocks if the are in dead block.
1746 irn_arity = get_irn_arity(n);
1748 if (get_irn_op(n) == op_End) {
1750 * Simplest case: End node. Predecessors are keep-alives,
1751 * no need to move out of dead block.
1753 for (i = -1; i < irn_arity; ++i) {
1754 ir_node *pred = get_irn_n(n, i);
1755 if (irn_not_visited(pred))
1756 waitq_put(worklist, pred);
1759 else if (is_Block(n)) {
1761 * Blocks: Predecessors are control flow, no need to move
1762 * them out of dead block.
1764 for (i = irn_arity - 1; i >= 0; --i) {
1765 ir_node *pred = get_irn_n(n, i);
1766 if (irn_not_visited(pred))
1767 waitq_put(worklist, pred);
1770 else if (is_Phi(n)) {
1772 ir_node *curr_block = get_irn_n(n, -1);
1773 int in_dead_block = is_Block_unreachable(curr_block);
1776 * Phi nodes: move nodes from dead blocks into the effective use
1777 * of the Phi-input if the Phi is not in a bad block.
1779 pred = get_irn_n(n, -1);
1780 if (irn_not_visited(pred))
1781 waitq_put(worklist, pred);
1783 for (i = irn_arity - 1; i >= 0; --i) {
1784 ir_node *pred = get_irn_n(n, i);
1786 if (irn_not_visited(pred)) {
1787 if (! in_dead_block &&
1788 get_irn_pinned(pred) == op_pin_state_floats &&
1789 is_Block_unreachable(get_irn_n(pred, -1))) {
1790 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1792 waitq_put(worklist, pred);
1798 ir_node *curr_block = get_irn_n(n, -1);
1799 int in_dead_block = is_Block_unreachable(curr_block);
1802 * All other nodes: move nodes from dead blocks into the same block.
1804 pred = get_irn_n(n, -1);
1805 if (irn_not_visited(pred))
1806 waitq_put(worklist, pred);
1808 for (i = irn_arity - 1; i >= 0; --i) {
1809 ir_node *pred = get_irn_n(n, i);
1811 if (irn_not_visited(pred)) {
1812 if (! in_dead_block &&
1813 get_irn_pinned(pred) == op_pin_state_floats &&
1814 is_Block_unreachable(get_irn_n(pred, -1))) {
1815 set_nodes_block(pred, curr_block);
1817 waitq_put(worklist, pred);
1824 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1825 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1826 * places all floating nodes reachable from its argument through floating
1827 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1829 static INLINE void place_early(waitq *worklist) {
1831 inc_irg_visited(current_ir_graph);
1833 /* this inits the worklist */
1834 place_floats_early(get_irg_end(current_ir_graph), worklist);
1836 /* Work the content of the worklist. */
1837 while (!waitq_empty(worklist)) {
1838 ir_node *n = waitq_get(worklist);
1839 if (irn_not_visited(n))
1840 place_floats_early(n, worklist);
1843 set_irg_outs_inconsistent(current_ir_graph);
1844 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1848 * Compute the deepest common ancestor of block and dca.
1850 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1854 /* we do not want to place nodes in dead blocks */
1855 if (is_Block_dead(block))
1858 /* We found a first legal placement. */
1859 if (!dca) return block;
1861 /* Find a placement that is dominates both, dca and block. */
1862 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1863 block = get_Block_idom(block);
1865 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1866 dca = get_Block_idom(dca);
1869 while (block != dca)
1870 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1875 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1876 * I.e., DCA is the block where we might place PRODUCER.
1877 * A data flow edge points from producer to consumer.
1880 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1882 ir_node *block = NULL;
1884 /* Compute the latest block into which we can place a node so that it is
1886 if (get_irn_op(consumer) == op_Phi) {
1887 /* our consumer is a Phi-node, the effective use is in all those
1888 blocks through which the Phi-node reaches producer */
1890 ir_node *phi_block = get_nodes_block(consumer);
1891 irn_arity = get_irn_arity(consumer);
1893 for (i = 0; i < irn_arity; i++) {
1894 if (get_irn_n(consumer, i) == producer) {
1895 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1897 if (! is_Block_unreachable(new_block))
1898 block = calc_dca(block, new_block);
1903 block = get_irn_n(producer, -1);
1906 assert(is_no_Block(consumer));
1907 block = get_nodes_block(consumer);
1910 /* Compute the deepest common ancestor of block and dca. */
1911 return calc_dca(dca, block);
1914 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1916 static INLINE int get_irn_loop_depth(ir_node *n) {
1917 return get_loop_depth(get_irn_loop(n));
1921 * Move n to a block with less loop depth than it's current block. The
1922 * new block must be dominated by early.
1924 * @param n the node that should be moved
1925 * @param early the earliest block we can n move to
1927 static void move_out_of_loops(ir_node *n, ir_node *early)
1929 ir_node *best, *dca;
1933 /* Find the region deepest in the dominator tree dominating
1934 dca with the least loop nesting depth, but still dominated
1935 by our early placement. */
1936 dca = get_nodes_block(n);
1939 while (dca != early) {
1940 dca = get_Block_idom(dca);
1941 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1942 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1946 if (best != get_nodes_block(n)) {
1948 printf("Moving out of loop: "); DDMN(n);
1949 printf(" Outermost block: "); DDMN(early);
1950 printf(" Best block: "); DDMN(best);
1951 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1953 set_nodes_block(n, best);
1958 * Find the latest legal block for N and place N into the
1959 * `optimal' Block between the latest and earliest legal block.
1960 * The `optimal' block is the dominance-deepest block of those
1961 * with the least loop-nesting-depth. This places N out of as many
1962 * loops as possible and then makes it as control dependent as
1965 static void place_floats_late(ir_node *n, pdeq *worklist)
1970 assert(irn_not_visited(n)); /* no multiple placement */
1972 mark_irn_visited(n);
1974 /* no need to place block nodes, control nodes are already placed. */
1975 if ((get_irn_op(n) != op_Block) &&
1977 (get_irn_mode(n) != mode_X)) {
1978 /* Remember the early_blk placement of this block to move it
1979 out of loop no further than the early_blk placement. */
1980 early_blk = get_irn_n(n, -1);
1983 * BEWARE: Here we also get code, that is live, but
1984 * was in a dead block. If the node is life, but because
1985 * of CSE in a dead block, we still might need it.
1988 /* Assure that our users are all placed, except the Phi-nodes.
1989 --- Each data flow cycle contains at least one Phi-node. We
1990 have to break the `user has to be placed before the
1991 producer' dependence cycle and the Phi-nodes are the
1992 place to do so, because we need to base our placement on the
1993 final region of our users, which is OK with Phi-nodes, as they
1994 are op_pin_state_pinned, and they never have to be placed after a
1995 producer of one of their inputs in the same block anyway. */
1996 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1997 ir_node *succ = get_irn_out(n, i);
1998 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1999 place_floats_late(succ, worklist);
2002 if (! is_Block_dead(early_blk)) {
2003 /* do only move things that where not dead */
2005 /* We have to determine the final block of this node... except for
2007 if ((get_irn_pinned(n) == op_pin_state_floats) &&
2008 (get_irn_op(n) != op_Const) &&
2009 (get_irn_op(n) != op_SymConst)) {
2010 ir_node *dca = NULL; /* deepest common ancestor in the
2011 dominator tree of all nodes'
2012 blocks depending on us; our final
2013 placement has to dominate DCA. */
2014 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
2015 ir_node *succ = get_irn_out(n, i);
2018 if (get_irn_op(succ) == op_End) {
2020 * This consumer is the End node, a keep alive edge.
2021 * This is not a real consumer, so we ignore it
2026 /* ignore if succ is in dead code */
2027 succ_blk = get_irn_n(succ, -1);
2028 if (is_Block_unreachable(succ_blk))
2030 dca = consumer_dom_dca(dca, succ, n);
2033 set_nodes_block(n, dca);
2034 move_out_of_loops(n, early_blk);
2040 /* Add predecessors of all non-floating nodes on list. (Those of floating
2041 nodes are placed already and therefore are marked.) */
2042 for (i = 0; i < get_irn_n_outs(n); i++) {
2043 ir_node *succ = get_irn_out(n, i);
2044 if (irn_not_visited(get_irn_out(n, i))) {
2045 pdeq_putr(worklist, succ);
2050 static INLINE void place_late(waitq *worklist) {
2052 inc_irg_visited(current_ir_graph);
2054 /* This fills the worklist initially. */
2055 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2057 /* And now empty the worklist again... */
2058 while (!waitq_empty(worklist)) {
2059 ir_node *n = waitq_get(worklist);
2060 if (irn_not_visited(n))
2061 place_floats_late(n, worklist);
2065 void place_code(ir_graph *irg) {
2067 ir_graph *rem = current_ir_graph;
2069 current_ir_graph = irg;
2071 if (!(get_opt_optimize() && get_opt_global_cse())) return;
2073 /* Handle graph state */
2074 assert(get_irg_phase_state(irg) != phase_building);
2077 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2078 free_loop_information(irg);
2079 construct_backedges(irg);
2082 /* Place all floating nodes as early as possible. This guarantees
2083 a legal code placement. */
2084 worklist = new_waitq();
2085 place_early(worklist);
2087 /* place_early() invalidates the outs, place_late needs them. */
2088 compute_irg_outs(irg);
2090 /* Now move the nodes down in the dominator tree. This reduces the
2091 unnecessary executions of the node. */
2092 place_late(worklist);
2094 set_irg_outs_inconsistent(current_ir_graph);
2095 set_irg_loopinfo_inconsistent(current_ir_graph);
2096 del_waitq(worklist);
2097 current_ir_graph = rem;
2101 * Called by walker of remove_critical_cf_edges().
2103 * Place an empty block to an edge between a blocks of multiple
2104 * predecessors and a block of multiple successors.
2107 * @param env Environment of walker. The changed field.
2109 static void walk_critical_cf_edges(ir_node *n, void *env) {
2111 ir_node *pre, *block, *jmp;
2113 ir_graph *irg = get_irn_irg(n);
2115 /* Block has multiple predecessors */
2116 arity = get_irn_arity(n);
2118 if (n == get_irg_end_block(irg))
2119 return; /* No use to add a block here. */
2121 for (i = 0; i < arity; ++i) {
2124 pre = get_irn_n(n, i);
2125 cfop = get_irn_op(skip_Proj(pre));
2126 /* Predecessor has multiple successors. Insert new control flow edge but
2127 ignore exception edges. */
2128 if (! is_op_fragile(cfop) && is_op_forking(cfop)) {
2129 /* set predecessor of new block */
2130 block = new_r_Block(irg, 1, &pre);
2131 /* insert new jmp node to new block */
2132 jmp = new_r_Jmp(irg, block);
2133 /* set successor of new block */
2134 set_irn_n(n, i, jmp);
2136 } /* predecessor has multiple successors */
2137 } /* for all predecessors */
2138 } /* n is a multi-entry block */
2141 void remove_critical_cf_edges(ir_graph *irg) {
2144 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &changed);
2146 /* control flow changed */
2147 set_irg_outs_inconsistent(irg);
2148 set_irg_extblk_inconsistent(irg);
2149 set_irg_doms_inconsistent(irg);
2150 set_irg_loopinfo_inconsistent(irg);