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 /* Defined in iropt.c */
50 pset *new_identities (void);
51 void del_identities (pset *value_table);
52 void add_identities (pset *value_table, ir_node *node);
54 /*------------------------------------------------------------------*/
55 /* apply optimizations of iropt to all nodes. */
56 /*------------------------------------------------------------------*/
58 static void init_link (ir_node *n, void *env) {
59 set_irn_link(n, NULL);
62 #if 0 /* Old version. Avoids Ids.
63 This is not necessary: we do a post walk, and get_irn_n
64 removes ids anyways. So it's much cheaper to call the
65 optimization less often and use the exchange() algorithm. */
67 optimize_in_place_wrapper (ir_node *n, void *env) {
69 ir_node *optimized, *old;
71 irn_arity = get_irn_arity(n);
72 for (i = 0; i < irn_arity; i++) {
73 /* get_irn_n skips Id nodes, so comparison old != optimized does not
74 show all optimizations. Therefore always set new predecessor. */
75 old = get_irn_intra_n(n, i);
76 optimized = optimize_in_place_2(old);
77 set_irn_n(n, i, optimized);
80 if (get_irn_op(n) == op_Block) {
81 optimized = optimize_in_place_2(n);
82 if (optimized != n) exchange (n, optimized);
87 optimize_in_place_wrapper (ir_node *n, void *env) {
88 ir_node *optimized = optimize_in_place_2(n);
89 if (optimized != n) exchange (n, optimized);
94 static INLINE void do_local_optimize(ir_node *n) {
95 /* Handle graph state */
96 assert(get_irg_phase_state(current_ir_graph) != phase_building);
98 if (get_opt_global_cse())
99 set_irg_pinned(current_ir_graph, op_pin_state_floats);
100 set_irg_outs_inconsistent(current_ir_graph);
101 set_irg_doms_inconsistent(current_ir_graph);
102 set_irg_loopinfo_inconsistent(current_ir_graph);
104 /* Clean the value_table in irg for the CSE. */
105 del_identities(current_ir_graph->value_table);
106 current_ir_graph->value_table = new_identities();
108 /* walk over the graph */
109 irg_walk(n, init_link, optimize_in_place_wrapper, NULL);
112 void local_optimize_node(ir_node *n) {
113 ir_graph *rem = current_ir_graph;
114 current_ir_graph = get_irn_irg(n);
116 do_local_optimize(n);
118 current_ir_graph = rem;
122 * Block-Walker: uses dominance depth to mark dead blocks.
124 static void kill_dead_blocks(ir_node *block, void *env)
126 if (get_Block_dom_depth(block) < 0)
127 if (block != get_irg_end_block(current_ir_graph)) {
128 /* we don't want that the end block of graphs with
129 endless loops is marked bad (although it is of course */
130 set_Block_dead(block);
135 local_optimize_graph (ir_graph *irg) {
136 ir_graph *rem = current_ir_graph;
137 current_ir_graph = irg;
139 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
140 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
142 do_local_optimize(get_irg_end(irg));
144 current_ir_graph = rem;
148 /*------------------------------------------------------------------*/
149 /* Routines for dead node elimination / copying garbage collection */
150 /* of the obstack. */
151 /*------------------------------------------------------------------*/
154 * Remember the new node in the old node by using a field all nodes have.
156 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
159 * Get this new node, before the old node is forgotten.
161 #define get_new_node(oldn) get_irn_link(oldn)
164 * Check if a new node was set.
166 #define has_new_node(n) (get_new_node(n) != NULL)
169 * We use the block_visited flag to mark that we have computed the
170 * number of useful predecessors for this block.
171 * Further we encode the new arity in this flag in the old blocks.
172 * Remembering the arity is useful, as it saves a lot of pointer
173 * accesses. This function is called for all Phi and Block nodes
177 compute_new_arity(ir_node *b) {
178 int i, res, irn_arity;
181 irg_v = get_irg_block_visited(current_ir_graph);
182 block_v = get_Block_block_visited(b);
183 if (block_v >= irg_v) {
184 /* we computed the number of preds for this block and saved it in the
186 return block_v - irg_v;
188 /* compute the number of good predecessors */
189 res = irn_arity = get_irn_arity(b);
190 for (i = 0; i < irn_arity; i++)
191 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
192 /* save it in the flag. */
193 set_Block_block_visited(b, irg_v + res);
199 * Copies the node to the new obstack. The Ins of the new node point to
200 * the predecessors on the old obstack. For block/phi nodes not all
201 * predecessors might be copied. n->link points to the new node.
202 * For Phi and Block nodes the function allocates in-arrays with an arity
203 * only for useful predecessors. The arity is determined by counting
204 * the non-bad predecessors of the block.
206 * @param n The node to be copied
207 * @param env if non-NULL, the node number attribute will be copied to the new node
209 * Note: Also used for loop unrolling.
211 static void copy_node(ir_node *n, void *env) {
214 ir_op *op = get_irn_op(n);
215 int copy_node_nr = env != NULL;
217 /* The end node looses it's flexible in array. This doesn't matter,
218 as dead node elimination builds End by hand, inlineing doesn't use
220 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
223 /* node copied already */
225 } else if (op == op_Block) {
227 new_arity = compute_new_arity(n);
228 n->attr.block.graph_arr = NULL;
230 block = get_nodes_block(n);
232 new_arity = compute_new_arity(block);
234 new_arity = get_irn_arity(n);
237 nn = new_ir_node(get_irn_dbg_info(n),
244 /* Copy the attributes. These might point to additional data. If this
245 was allocated on the old obstack the pointers now are dangling. This
246 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
247 copy_node_attr(n, nn);
248 new_backedge_info(nn);
252 /* for easier debugging, we want to copy the node numbers too */
253 nn->node_nr = n->node_nr;
258 hook_dead_node_elim_subst(current_ir_graph, n, nn);
262 * Copies new predecessors of old node to new node remembered in link.
263 * Spare the Bad predecessors of Phi and Block nodes.
266 copy_preds (ir_node *n, void *env) {
270 nn = get_new_node(n);
272 /* printf("\n old node: "); DDMSG2(n);
273 printf(" new node: "); DDMSG2(nn);
274 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
277 /* Don't copy Bad nodes. */
279 irn_arity = get_irn_arity(n);
280 for (i = 0; i < irn_arity; i++)
281 if (! is_Bad(get_irn_n(n, i))) {
282 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
283 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
286 /* repair the block visited flag from above misuse. Repair it in both
287 graphs so that the old one can still be used. */
288 set_Block_block_visited(nn, 0);
289 set_Block_block_visited(n, 0);
290 /* Local optimization could not merge two subsequent blocks if
291 in array contained Bads. Now it's possible.
292 We don't call optimize_in_place as it requires
293 that the fields in ir_graph are set properly. */
294 if ((get_opt_control_flow_straightening()) &&
295 (get_Block_n_cfgpreds(nn) == 1) &&
296 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
297 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
299 /* Jmp jumps into the block it is in -- deal self cycle. */
300 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
301 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
306 } else if (get_irn_op(n) == op_Phi) {
307 /* Don't copy node if corresponding predecessor in block is Bad.
308 The Block itself should not be Bad. */
309 block = get_nodes_block(n);
310 set_irn_n (nn, -1, get_new_node(block));
312 irn_arity = get_irn_arity(n);
313 for (i = 0; i < irn_arity; i++)
314 if (! is_Bad(get_irn_n(block, i))) {
315 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
316 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
319 /* If the pre walker reached this Phi after the post walker visited the
320 block block_visited is > 0. */
321 set_Block_block_visited(get_nodes_block(n), 0);
322 /* Compacting the Phi's ins might generate Phis with only one
324 if (get_irn_arity(nn) == 1)
325 exchange(nn, get_irn_n(nn, 0));
327 irn_arity = get_irn_arity(n);
328 for (i = -1; i < irn_arity; i++)
329 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
331 /* Now the new node is complete. We can add it to the hash table for CSE.
332 @@@ inlining aborts if we identify End. Why? */
333 if (get_irn_op(nn) != op_End)
334 add_identities (current_ir_graph->value_table, nn);
338 * Copies the graph recursively, compacts the keep-alives of the end node.
340 * @param irg the graph to be copied
341 * @param copy_node_nr If non-zero, the node number will be copied
343 static void copy_graph(ir_graph *irg, int copy_node_nr) {
344 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
345 ir_node *ka; /* keep alive */
349 /* Some nodes must be copied by hand, sigh */
350 vfl = get_irg_visited(irg);
351 set_irg_visited(irg, vfl + 1);
353 oe = get_irg_end(irg);
354 mark_irn_visited(oe);
355 /* copy the end node by hand, allocate dynamic in array! */
356 ne = new_ir_node(get_irn_dbg_info(oe),
363 /* Copy the attributes. Well, there might be some in the future... */
364 copy_node_attr(oe, ne);
365 set_new_node(oe, ne);
367 /* copy the Bad node */
368 ob = get_irg_bad(irg);
369 mark_irn_visited(ob);
370 nb = new_ir_node(get_irn_dbg_info(ob),
377 copy_node_attr(ob, nb);
378 set_new_node(ob, nb);
380 /* copy the NoMem node */
381 om = get_irg_no_mem(irg);
382 mark_irn_visited(om);
383 nm = new_ir_node(get_irn_dbg_info(om),
390 copy_node_attr(om, nm);
391 set_new_node(om, nm);
393 /* copy the live nodes */
394 set_irg_visited(irg, vfl);
395 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
397 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
399 /* visit the anchors as well */
400 for (i = anchor_max - 1; i >= 0; --i) {
401 ir_node *n = irg->anchors[i];
403 if (n && (get_irn_visited(n) <= vfl)) {
404 set_irg_visited(irg, vfl);
405 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
409 /* copy_preds for the end node ... */
410 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
412 /*- ... and now the keep alives. -*/
413 /* First pick the not marked block nodes and walk them. We must pick these
414 first as else we will oversee blocks reachable from Phis. */
415 irn_arity = get_irn_arity(oe);
416 for (i = 0; i < irn_arity; i++) {
417 ka = get_irn_intra_n(oe, i);
419 (get_irn_visited(ka) <= vfl)) {
420 /* We must keep the block alive and copy everything reachable */
421 set_irg_visited(irg, vfl);
422 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
423 add_End_keepalive(ne, get_new_node(ka));
427 /* Now pick other nodes. Here we will keep all! */
428 irn_arity = get_irn_arity(oe);
429 for (i = 0; i < irn_arity; i++) {
430 ka = get_irn_intra_n(oe, i);
432 if (get_irn_visited(ka) <= vfl) {
433 /* We didn't copy the node yet. */
434 set_irg_visited(irg, vfl);
435 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
437 add_End_keepalive(ne, get_new_node(ka));
441 /* start block sometimes only reached after keep alives */
442 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
443 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
447 * Copies the graph reachable from current_ir_graph->end to the obstack
448 * in current_ir_graph and fixes the environment.
449 * Then fixes the fields in current_ir_graph containing nodes of the
452 * @param copy_node_nr If non-zero, the node number will be copied
455 copy_graph_env (int copy_node_nr) {
456 ir_graph *irg = current_ir_graph;
457 ir_node *old_end, *n;
460 /* remove end_except and end_reg nodes */
461 old_end = get_irg_end(irg);
462 set_irg_end_except (irg, old_end);
463 set_irg_end_reg (irg, old_end);
465 /* Not all nodes remembered in irg might be reachable
466 from the end node. Assure their link is set to NULL, so that
467 we can test whether new nodes have been computed. */
468 for (i = anchor_max - 1; i >= 0; --i)
470 set_new_node(irg->anchors[i], NULL);
472 /* we use the block walk flag for removing Bads from Blocks ins. */
473 inc_irg_block_visited(irg);
476 copy_graph(irg, copy_node_nr);
478 /* fix the fields in irg */
479 old_end = get_irg_end(irg);
480 for (i = anchor_max - 1; i >= 0; --i) {
483 irg->anchors[i] = get_new_node(n);
489 * Copies all reachable nodes to a new obstack. Removes bad inputs
490 * from block nodes and the corresponding inputs from Phi nodes.
491 * Merges single exit blocks with single entry blocks and removes
493 * Adds all new nodes to a new hash table for CSE. Does not
494 * perform CSE, so the hash table might contain common subexpressions.
497 dead_node_elimination(ir_graph *irg) {
499 int rem_ipview = get_interprocedural_view();
500 struct obstack *graveyard_obst = NULL;
501 struct obstack *rebirth_obst = NULL;
503 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
504 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
506 /* inform statistics that we started a dead-node elimination run */
507 hook_dead_node_elim(irg, 1);
509 /* Remember external state of current_ir_graph. */
510 rem = current_ir_graph;
511 current_ir_graph = irg;
512 set_interprocedural_view(0);
514 assert(get_irg_phase_state(current_ir_graph) != phase_building);
516 /* Handle graph state */
517 free_callee_info(current_ir_graph);
518 free_irg_outs(current_ir_graph);
521 /* @@@ so far we loose loops when copying */
522 free_loop_information(current_ir_graph);
524 set_irg_doms_inconsistent(irg);
526 /* A quiet place, where the old obstack can rest in peace,
527 until it will be cremated. */
528 graveyard_obst = irg->obst;
530 /* A new obstack, where the reachable nodes will be copied to. */
531 rebirth_obst = xmalloc (sizeof(*rebirth_obst));
532 current_ir_graph->obst = rebirth_obst;
533 obstack_init (current_ir_graph->obst);
535 /* We also need a new hash table for cse */
536 del_identities (irg->value_table);
537 irg->value_table = new_identities ();
539 /* Copy the graph from the old to the new obstack */
542 /* Free memory from old unoptimized obstack */
543 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
544 xfree (graveyard_obst); /* ... then free it. */
546 /* inform statistics that the run is over */
547 hook_dead_node_elim(irg, 0);
549 current_ir_graph = rem;
550 set_interprocedural_view(rem_ipview);
555 * Relink bad predecessors of a block and store the old in array to the
556 * link field. This function is called by relink_bad_predecessors().
557 * The array of link field starts with the block operand at position 0.
558 * If block has bad predecessors, create a new in array without bad preds.
559 * Otherwise let in array untouched.
561 static void relink_bad_block_predecessors(ir_node *n, void *env) {
562 ir_node **new_in, *irn;
563 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
565 /* if link field of block is NULL, look for bad predecessors otherwise
566 this is already done */
567 if (get_irn_op(n) == op_Block &&
568 get_irn_link(n) == NULL) {
570 /* save old predecessors in link field (position 0 is the block operand)*/
571 set_irn_link(n, get_irn_in(n));
573 /* count predecessors without bad nodes */
574 old_irn_arity = get_irn_arity(n);
575 for (i = 0; i < old_irn_arity; i++)
576 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
578 /* arity changing: set new predecessors without bad nodes */
579 if (new_irn_arity < old_irn_arity) {
580 /* Get new predecessor array. We do not resize the array, as we must
581 keep the old one to update Phis. */
582 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
584 /* set new predecessors in array */
587 for (i = 0; i < old_irn_arity; i++) {
588 irn = get_irn_n(n, i);
590 new_in[new_irn_n] = irn;
591 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
595 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
596 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
599 } /* ir node has bad predecessors */
601 } /* Block is not relinked */
605 * Relinks Bad predecessors from Blocks and Phis called by walker
606 * remove_bad_predecesors(). If n is a Block, call
607 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
608 * function of Phi's Block. If this block has bad predecessors, relink preds
611 static void relink_bad_predecessors(ir_node *n, void *env) {
612 ir_node *block, **old_in;
613 int i, old_irn_arity, new_irn_arity;
615 /* relink bad predecessors of a block */
616 if (get_irn_op(n) == op_Block)
617 relink_bad_block_predecessors(n, env);
619 /* If Phi node relink its block and its predecessors */
620 if (get_irn_op(n) == op_Phi) {
622 /* Relink predecessors of phi's block */
623 block = get_nodes_block(n);
624 if (get_irn_link(block) == NULL)
625 relink_bad_block_predecessors(block, env);
627 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
628 old_irn_arity = ARR_LEN(old_in);
630 /* Relink Phi predecessors if count of predecessors changed */
631 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
632 /* set new predecessors in array
633 n->in[0] remains the same block */
635 for(i = 1; i < old_irn_arity; i++)
636 if (!is_Bad((ir_node *)old_in[i])) {
637 n->in[new_irn_arity] = n->in[i];
638 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
642 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
643 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
646 } /* n is a Phi node */
650 * Removes Bad Bad predecessors from Blocks and the corresponding
651 * inputs to Phi nodes as in dead_node_elimination but without
653 * On walking up set the link field to NULL, on walking down call
654 * relink_bad_predecessors() (This function stores the old in array
655 * to the link field and sets a new in array if arity of predecessors
658 void remove_bad_predecessors(ir_graph *irg) {
659 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
666 __)|_| | \_/ | \_/(/_ |_/\__|__
668 The following stuff implements a facility that automatically patches
669 registered ir_node pointers to the new node when a dead node elimination occurs.
672 struct _survive_dce_t {
676 hook_entry_t dead_node_elim;
677 hook_entry_t dead_node_elim_subst;
680 typedef struct _survive_dce_list_t {
681 struct _survive_dce_list_t *next;
683 } survive_dce_list_t;
685 static void dead_node_hook(void *context, ir_graph *irg, int start)
687 survive_dce_t *sd = context;
689 /* Create a new map before the dead node elimination is performed. */
691 sd->new_places = pmap_create_ex(pmap_count(sd->places));
694 /* Patch back all nodes if dead node elimination is over and something is to be done. */
696 pmap_destroy(sd->places);
697 sd->places = sd->new_places;
698 sd->new_places = NULL;
703 * Hook called when dead node elimination replaces old by nw.
705 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw)
707 survive_dce_t *sd = context;
708 survive_dce_list_t *list = pmap_get(sd->places, old);
710 /* If the node is to be patched back, write the new address to all registered locations. */
712 survive_dce_list_t *p;
714 for(p = list; p; p = p->next)
717 pmap_insert(sd->new_places, nw, list);
722 * Make a new Survive DCE environment.
724 survive_dce_t *new_survive_dce(void)
726 survive_dce_t *res = xmalloc(sizeof(res[0]));
727 obstack_init(&res->obst);
728 res->places = pmap_create();
729 res->new_places = NULL;
731 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
732 res->dead_node_elim.context = res;
733 res->dead_node_elim.next = NULL;
735 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
736 res->dead_node_elim_subst.context = res;
737 res->dead_node_elim_subst.next = NULL;
739 register_hook(hook_dead_node_elim, &res->dead_node_elim);
740 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
745 * Free a Survive DCE environment.
747 void free_survive_dce(survive_dce_t *sd)
749 obstack_free(&sd->obst, NULL);
750 pmap_destroy(sd->places);
751 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
752 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
757 * Register a node pointer to be patched upon DCE.
758 * When DCE occurs, the node pointer specified by @p place will be
759 * patched to the new address of the node it is pointing to.
761 * @param sd The Survive DCE environment.
762 * @param place The address of the node pointer.
764 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place)
767 ir_node *irn = *place;
768 survive_dce_list_t *curr = pmap_get(sd->places, irn);
769 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw));
774 pmap_insert(sd->places, irn, nw);
778 /*--------------------------------------------------------------------*/
779 /* Functionality for inlining */
780 /*--------------------------------------------------------------------*/
783 * Copy node for inlineing. Updates attributes that change when
784 * inlineing but not for dead node elimination.
786 * Copies the node by calling copy_node() and then updates the entity if
787 * it's a local one. env must be a pointer of the frame type of the
788 * inlined procedure. The new entities must be in the link field of
792 copy_node_inline (ir_node *n, void *env) {
794 ir_type *frame_tp = (ir_type *)env;
797 if (get_irn_op(n) == op_Sel) {
798 nn = get_new_node (n);
800 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
801 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
803 } else if (get_irn_op(n) == op_Block) {
804 nn = get_new_node (n);
805 nn->attr.block.irg = current_ir_graph;
809 static void find_addr(ir_node *node, void *env)
811 if (get_irn_opcode(node) == iro_Proj) {
812 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
818 * currently, we cannot inline two cases:
819 * - call with compound arguments
820 * - graphs that take the address of a parameter
822 * check these conditions here
824 static int can_inline(ir_node *call, ir_graph *called_graph)
826 ir_type *call_type = get_Call_type(call);
827 int params, ress, i, res;
828 assert(is_Method_type(call_type));
830 params = get_method_n_params(call_type);
831 ress = get_method_n_ress(call_type);
834 for (i = 0; i < params; ++i) {
835 ir_type *p_type = get_method_param_type(call_type, i);
837 if (is_compound_type(p_type))
842 for (i = 0; i < ress; ++i) {
843 ir_type *r_type = get_method_res_type(call_type, i);
845 if (is_compound_type(r_type))
850 irg_walk_graph(called_graph, find_addr, NULL, &res);
855 int inline_method(ir_node *call, ir_graph *called_graph) {
857 ir_node *post_call, *post_bl;
859 ir_node *end, *end_bl;
863 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
865 ir_type *called_frame;
866 irg_inline_property prop = get_irg_inline_property(called_graph);
868 if ( (prop != irg_inline_forced) &&
869 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
871 /* Do not inline variadic functions. */
872 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
875 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
876 get_method_n_params(get_Call_type(call)));
879 * currently, we cannot inline two cases:
880 * - call with compound arguments
881 * - graphs that take the address of a parameter
883 if (! can_inline(call, called_graph))
886 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
887 rem_opt = get_opt_optimize();
890 /* Handle graph state */
891 assert(get_irg_phase_state(current_ir_graph) != phase_building);
892 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
893 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
894 set_irg_outs_inconsistent(current_ir_graph);
895 set_irg_extblk_inconsistent(current_ir_graph);
896 set_irg_doms_inconsistent(current_ir_graph);
897 set_irg_loopinfo_inconsistent(current_ir_graph);
898 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
900 /* -- Check preconditions -- */
901 assert(is_Call(call));
902 /* @@@ does not work for InterfaceIII.java after cgana
903 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
904 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
905 get_Call_type(call)));
907 assert(get_type_tpop(get_Call_type(call)) == type_method);
908 if (called_graph == current_ir_graph) {
909 set_optimize(rem_opt);
913 /* here we know we WILL inline, so inform the statistics */
914 hook_inline(call, called_graph);
916 /* -- Decide how to handle exception control flow: Is there a handler
917 for the Call node, or do we branch directly to End on an exception?
919 0 There is a handler.
921 2 Exception handling not represented in Firm. -- */
923 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
924 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
925 assert(get_irn_op(proj) == op_Proj);
926 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
927 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
929 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
930 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
931 else { exc_handling = 2; } /* !Mproj && !Xproj */
936 the procedure and later replaces the Start node of the called graph.
937 Post_call is the old Call node and collects the results of the called
938 graph. Both will end up being a tuple. -- */
939 post_bl = get_nodes_block(call);
940 set_irg_current_block(current_ir_graph, post_bl);
941 /* XxMxPxP of Start + parameter of Call */
942 in[pn_Start_X_initial_exec] = new_Jmp();
943 in[pn_Start_M] = get_Call_mem(call);
944 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
945 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
946 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
947 /* in[pn_Start_P_value_arg_base] = ??? */
948 pre_call = new_Tuple(5, in);
952 The new block gets the ins of the old block, pre_call and all its
953 predecessors and all Phi nodes. -- */
954 part_block(pre_call);
956 /* -- Prepare state for dead node elimination -- */
957 /* Visited flags in calling irg must be >= flag in called irg.
958 Else walker and arity computation will not work. */
959 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
960 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
961 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
962 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
963 /* Set pre_call as new Start node in link field of the start node of
964 calling graph and pre_calls block as new block for the start block
966 Further mark these nodes so that they are not visited by the
968 set_irn_link(get_irg_start(called_graph), pre_call);
969 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
970 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
971 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
972 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
973 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
975 /* Initialize for compaction of in arrays */
976 inc_irg_block_visited(current_ir_graph);
978 /* -- Replicate local entities of the called_graph -- */
979 /* copy the entities. */
980 called_frame = get_irg_frame_type(called_graph);
981 for (i = 0; i < get_class_n_members(called_frame); i++) {
982 entity *new_ent, *old_ent;
983 old_ent = get_class_member(called_frame, i);
984 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
985 set_entity_link(old_ent, new_ent);
988 /* visited is > than that of called graph. With this trick visited will
989 remain unchanged so that an outer walker, e.g., searching the call nodes
990 to inline, calling this inline will not visit the inlined nodes. */
991 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
993 /* -- Performing dead node elimination inlines the graph -- */
994 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
996 /* @@@ endless loops are not copied!! -- they should be, I think... */
997 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
998 get_irg_frame_type(called_graph));
1000 /* Repair called_graph */
1001 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1002 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1003 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1005 /* -- Merge the end of the inlined procedure with the call site -- */
1006 /* We will turn the old Call node into a Tuple with the following
1009 0: Phi of all Memories of Return statements.
1010 1: Jmp from new Block that merges the control flow from all exception
1011 predecessors of the old end block.
1012 2: Tuple of all arguments.
1013 3: Phi of Exception memories.
1014 In case the old Call directly branches to End on an exception we don't
1015 need the block merging all exceptions nor the Phi of the exception
1019 /* -- Precompute some values -- */
1020 end_bl = get_new_node(get_irg_end_block(called_graph));
1021 end = get_new_node(get_irg_end(called_graph));
1022 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1023 n_res = get_method_n_ress(get_Call_type(call));
1025 res_pred = xmalloc (n_res * sizeof(*res_pred));
1026 cf_pred = xmalloc (arity * sizeof(*res_pred));
1028 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1030 /* -- archive keepalives -- */
1031 irn_arity = get_irn_arity(end);
1032 for (i = 0; i < irn_arity; i++)
1033 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
1035 /* The new end node will die. We need not free as the in array is on the obstack:
1036 copy_node() only generated 'D' arrays. */
1038 /* -- Replace Return nodes by Jump nodes. -- */
1040 for (i = 0; i < arity; i++) {
1042 ret = get_irn_n(end_bl, i);
1043 if (is_Return(ret)) {
1044 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1048 set_irn_in(post_bl, n_ret, cf_pred);
1050 /* -- Build a Tuple for all results of the method.
1051 Add Phi node if there was more than one Return. -- */
1052 turn_into_tuple(post_call, 4);
1053 /* First the Memory-Phi */
1055 for (i = 0; i < arity; i++) {
1056 ret = get_irn_n(end_bl, i);
1057 if (is_Return(ret)) {
1058 cf_pred[n_ret] = get_Return_mem(ret);
1062 phi = new_Phi(n_ret, cf_pred, mode_M);
1063 set_Tuple_pred(call, pn_Call_M_regular, phi);
1064 /* Conserve Phi-list for further inlinings -- but might be optimized */
1065 if (get_nodes_block(phi) == post_bl) {
1066 set_irn_link(phi, get_irn_link(post_bl));
1067 set_irn_link(post_bl, phi);
1069 /* Now the real results */
1071 for (j = 0; j < n_res; j++) {
1073 for (i = 0; i < arity; i++) {
1074 ret = get_irn_n(end_bl, i);
1075 if (get_irn_op(ret) == op_Return) {
1076 cf_pred[n_ret] = get_Return_res(ret, j);
1081 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1085 /* Conserve Phi-list for further inlinings -- but might be optimized */
1086 if (get_nodes_block(phi) == post_bl) {
1087 set_irn_link(phi, get_irn_link(post_bl));
1088 set_irn_link(post_bl, phi);
1091 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1093 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1095 /* Finally the exception control flow.
1096 We have two (three) possible situations:
1097 First if the Call branches to an exception handler: We need to add a Phi node to
1098 collect the memory containing the exception objects. Further we need
1099 to add another block to get a correct representation of this Phi. To
1100 this block we add a Jmp that resolves into the X output of the Call
1101 when the Call is turned into a tuple.
1102 Second the Call branches to End, the exception is not handled. Just
1103 add all inlined exception branches to the End node.
1104 Third: there is no Exception edge at all. Handle as case two. */
1105 if (exc_handling == 0) {
1107 for (i = 0; i < arity; i++) {
1109 ret = get_irn_n(end_bl, i);
1110 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1111 cf_pred[n_exc] = ret;
1116 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1117 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1118 /* The Phi for the memories with the exception objects */
1120 for (i = 0; i < arity; i++) {
1122 ret = skip_Proj(get_irn_n(end_bl, i));
1124 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1126 } else if (is_fragile_op(ret)) {
1127 /* We rely that all cfops have the memory output at the same position. */
1128 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1130 } else if (get_irn_op(ret) == op_Raise) {
1131 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1135 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1137 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1138 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1141 ir_node *main_end_bl;
1142 int main_end_bl_arity;
1143 ir_node **end_preds;
1145 /* assert(exc_handling == 1 || no exceptions. ) */
1147 for (i = 0; i < arity; i++) {
1148 ir_node *ret = get_irn_n(end_bl, i);
1150 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1151 cf_pred[n_exc] = ret;
1155 main_end_bl = get_irg_end_block(current_ir_graph);
1156 main_end_bl_arity = get_irn_arity(main_end_bl);
1157 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1159 for (i = 0; i < main_end_bl_arity; ++i)
1160 end_preds[i] = get_irn_n(main_end_bl, i);
1161 for (i = 0; i < n_exc; ++i)
1162 end_preds[main_end_bl_arity + i] = cf_pred[i];
1163 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1164 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1165 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1171 #if 0 /* old. now better, correcter, faster implementation. */
1173 /* -- If the exception control flow from the inlined Call directly
1174 branched to the end block we now have the following control
1175 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1176 remove the Jmp along with it's empty block and add Jmp's
1177 predecessors as predecessors of this end block. No problem if
1178 there is no exception, because then branches Bad to End which
1180 @@@ can't we know this beforehand: by getting the Proj(1) from
1181 the Call link list and checking whether it goes to Proj. */
1182 /* find the problematic predecessor of the end block. */
1183 end_bl = get_irg_end_block(current_ir_graph);
1184 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1185 cf_op = get_Block_cfgpred(end_bl, i);
1186 if (get_irn_op(cf_op) == op_Proj) {
1187 cf_op = get_Proj_pred(cf_op);
1188 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1189 /* There are unoptimized tuples from inlineing before when no exc */
1190 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1191 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1192 assert(get_irn_op(cf_op) == op_Jmp);
1198 if (i < get_Block_n_cfgpreds(end_bl)) {
1199 bl = get_nodes_block(cf_op);
1200 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1201 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1202 for (j = 0; j < i; j++)
1203 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1204 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1205 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1206 for (j = j; j < arity; j++)
1207 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1208 set_irn_in(end_bl, arity, cf_pred);
1210 /* Remove the exception pred from post-call Tuple. */
1211 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1216 /* -- Turn CSE back on. -- */
1217 set_optimize(rem_opt);
1222 /********************************************************************/
1223 /* Apply inlineing to small methods. */
1224 /********************************************************************/
1226 /* It makes no sense to inline too many calls in one procedure. Anyways,
1227 I didn't get a version with NEW_ARR_F to run. */
1228 #define MAX_INLINE 1024
1231 * environment for inlining small irgs
1233 typedef struct _inline_env_t {
1235 ir_node *calls[MAX_INLINE];
1239 * Returns the irg called from a Call node. If the irg is not
1240 * known, NULL is returned.
1242 static ir_graph *get_call_called_irg(ir_node *call) {
1244 ir_graph *called_irg = NULL;
1246 assert(is_Call(call));
1248 addr = get_Call_ptr(call);
1249 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1250 called_irg = get_entity_irg(get_SymConst_entity(addr));
1256 static void collect_calls(ir_node *call, void *env) {
1259 if (! is_Call(call)) return;
1261 addr = get_Call_ptr(call);
1263 if (get_irn_op(addr) == op_SymConst) {
1264 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1265 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1266 inline_env_t *ienv = (inline_env_t *)env;
1267 if (called_irg && ienv->pos < MAX_INLINE) {
1268 /* The Call node calls a locally defined method. Remember to inline. */
1269 ienv->calls[ienv->pos++] = call;
1276 * Inlines all small methods at call sites where the called address comes
1277 * from a Const node that references the entity representing the called
1279 * The size argument is a rough measure for the code size of the method:
1280 * Methods where the obstack containing the firm graph is smaller than
1283 void inline_small_irgs(ir_graph *irg, int size) {
1285 ir_graph *rem = current_ir_graph;
1286 inline_env_t env /* = {0, NULL}*/;
1288 if (!(get_opt_optimize() && get_opt_inline())) return;
1290 current_ir_graph = irg;
1291 /* Handle graph state */
1292 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1293 free_callee_info(current_ir_graph);
1295 /* Find Call nodes to inline.
1296 (We can not inline during a walk of the graph, as inlineing the same
1297 method several times changes the visited flag of the walked graph:
1298 after the first inlineing visited of the callee equals visited of
1299 the caller. With the next inlineing both are increased.) */
1301 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1303 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1304 /* There are calls to inline */
1305 collect_phiprojs(irg);
1306 for (i = 0; i < env.pos; i++) {
1308 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1309 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1310 (get_irg_inline_property(callee) == irg_inline_forced)) {
1311 inline_method(env.calls[i], callee);
1316 current_ir_graph = rem;
1320 * Environment for inlining irgs.
1323 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1324 int n_nodes_orig; /**< for statistics */
1325 eset *call_nodes; /**< All call nodes in this graph */
1327 int n_call_nodes_orig; /**< for statistics */
1328 int n_callers; /**< Number of known graphs that call this graphs. */
1329 int n_callers_orig; /**< for statistics */
1333 * Allocate a new environment for inlining.
1335 static inline_irg_env *new_inline_irg_env(void) {
1336 inline_irg_env *env = xmalloc(sizeof(*env));
1337 env->n_nodes = -2; /* do not count count Start, End */
1338 env->n_nodes_orig = -2; /* do not count Start, End */
1339 env->call_nodes = eset_create();
1340 env->n_call_nodes = 0;
1341 env->n_call_nodes_orig = 0;
1343 env->n_callers_orig = 0;
1348 * destroy an environment for inlining.
1350 static void free_inline_irg_env(inline_irg_env *env) {
1351 eset_destroy(env->call_nodes);
1356 * post-walker: collect all calls in the inline-environment
1357 * of a graph and sum some statistics.
1359 static void collect_calls2(ir_node *call, void *env) {
1360 inline_irg_env *x = (inline_irg_env *)env;
1361 ir_op *op = get_irn_op(call);
1364 /* count meaningful nodes in irg */
1365 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1370 if (op != op_Call) return;
1372 /* collect all call nodes */
1373 eset_insert(x->call_nodes, call);
1375 x->n_call_nodes_orig++;
1377 /* count all static callers */
1378 callee = get_call_called_irg(call);
1380 inline_irg_env *callee_env = get_irg_link(callee);
1381 callee_env->n_callers++;
1382 callee_env->n_callers_orig++;
1387 * Returns TRUE if the number of callers in 0 in the irg's environment,
1388 * hence this irg is a leave.
1390 INLINE static int is_leave(ir_graph *irg) {
1391 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1395 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1397 INLINE static int is_smaller(ir_graph *callee, int size) {
1398 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1403 * Inlines small leave methods at call sites where the called address comes
1404 * from a Const node that references the entity representing the called
1406 * The size argument is a rough measure for the code size of the method:
1407 * Methods where the obstack containing the firm graph is smaller than
1410 void inline_leave_functions(int maxsize, int leavesize, int size) {
1411 inline_irg_env *env;
1412 int i, n_irgs = get_irp_n_irgs();
1413 ir_graph *rem = current_ir_graph;
1416 if (!(get_opt_optimize() && get_opt_inline())) return;
1418 /* extend all irgs by a temporary data structure for inlining. */
1419 for (i = 0; i < n_irgs; ++i)
1420 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1422 /* Precompute information in temporary data structure. */
1423 for (i = 0; i < n_irgs; ++i) {
1424 current_ir_graph = get_irp_irg(i);
1425 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1426 free_callee_info(current_ir_graph);
1428 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1429 get_irg_link(current_ir_graph));
1432 /* -- and now inline. -- */
1434 /* Inline leaves recursively -- we might construct new leaves. */
1435 while (did_inline) {
1438 for (i = 0; i < n_irgs; ++i) {
1440 int phiproj_computed = 0;
1442 current_ir_graph = get_irp_irg(i);
1443 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1445 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1448 if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
1449 callee = get_call_called_irg(call);
1451 if (env->n_nodes > maxsize) continue; // break;
1453 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1454 if (!phiproj_computed) {
1455 phiproj_computed = 1;
1456 collect_phiprojs(current_ir_graph);
1458 did_inline = inline_method(call, callee);
1461 /* Do some statistics */
1462 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1463 env->n_call_nodes --;
1464 env->n_nodes += callee_env->n_nodes;
1465 callee_env->n_callers--;
1472 /* inline other small functions. */
1473 for (i = 0; i < n_irgs; ++i) {
1476 int phiproj_computed = 0;
1478 current_ir_graph = get_irp_irg(i);
1479 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1481 /* we can not walk and change a set, nor remove from it.
1483 walkset = env->call_nodes;
1484 env->call_nodes = eset_create();
1485 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1488 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1489 callee = get_call_called_irg(call);
1492 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1493 (get_irg_inline_property(callee) == irg_inline_forced))) {
1494 if (!phiproj_computed) {
1495 phiproj_computed = 1;
1496 collect_phiprojs(current_ir_graph);
1498 if (inline_method(call, callee)) {
1499 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1500 env->n_call_nodes--;
1501 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1502 env->n_call_nodes += callee_env->n_call_nodes;
1503 env->n_nodes += callee_env->n_nodes;
1504 callee_env->n_callers--;
1507 eset_insert(env->call_nodes, call);
1510 eset_destroy(walkset);
1513 for (i = 0; i < n_irgs; ++i) {
1514 current_ir_graph = get_irp_irg(i);
1516 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1517 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1518 (env->n_callers_orig != env->n_callers))
1519 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1520 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1521 env->n_callers_orig, env->n_callers,
1522 get_entity_name(get_irg_entity(current_ir_graph)));
1524 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1527 current_ir_graph = rem;
1530 /*******************************************************************/
1531 /* Code Placement. Pins all floating nodes to a block where they */
1532 /* will be executed only if needed. */
1533 /*******************************************************************/
1536 * Returns non-zero, is a block is not reachable from Start.
1538 * @param block the block to test
1541 is_Block_unreachable(ir_node *block) {
1542 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1546 * Find the earliest correct block for N. --- Place N into the
1547 * same Block as its dominance-deepest Input.
1549 * We have to avoid calls to get_nodes_block() here
1550 * because the graph is floating.
1552 * move_out_of_loops() expects that place_floats_early() have placed
1553 * all "living" nodes into a living block. That's why we must
1554 * move nodes in dead block with "live" successors into a valid
1556 * We move them just into the same block as it's successor (or
1557 * in case of a Phi into the effective use block). For Phi successors,
1558 * this may still be a dead block, but then there is no real use, as
1559 * the control flow will be dead later.
1562 place_floats_early(ir_node *n, pdeq *worklist)
1566 /* we must not run into an infinite loop */
1567 assert(irn_not_visited(n));
1568 mark_irn_visited(n);
1570 /* Place floating nodes. */
1571 if (get_irn_pinned(n) == op_pin_state_floats) {
1572 ir_node *curr_block = get_irn_n(n, -1);
1573 int in_dead_block = is_Block_unreachable(curr_block);
1575 ir_node *b = NULL; /* The block to place this node in */
1577 assert(get_irn_op(n) != op_Block);
1579 if ((get_irn_op(n) == op_Const) ||
1580 (get_irn_op(n) == op_SymConst) ||
1582 (get_irn_op(n) == op_Unknown)) {
1583 /* These nodes will not be placed by the loop below. */
1584 b = get_irg_start_block(current_ir_graph);
1588 /* find the block for this node. */
1589 irn_arity = get_irn_arity(n);
1590 for (i = 0; i < irn_arity; i++) {
1591 ir_node *pred = get_irn_n(n, i);
1592 ir_node *pred_block;
1594 if ((irn_not_visited(pred))
1595 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1598 * If the current node is NOT in a dead block, but one of its
1599 * predecessors is, we must move the predecessor to a live block.
1600 * Such thing can happen, if global CSE chose a node from a dead block.
1601 * We move it simple to our block.
1602 * Note that neither Phi nor End nodes are floating, so we don't
1603 * need to handle them here.
1605 if (! in_dead_block) {
1606 if (get_irn_pinned(pred) == op_pin_state_floats &&
1607 is_Block_unreachable(get_irn_n(pred, -1)))
1608 set_nodes_block(pred, curr_block);
1610 place_floats_early(pred, worklist);
1614 * A node in the Bad block must stay in the bad block,
1615 * so don't compute a new block for it.
1620 /* Because all loops contain at least one op_pin_state_pinned node, now all
1621 our inputs are either op_pin_state_pinned or place_early() has already
1622 been finished on them. We do not have any unfinished inputs! */
1623 pred_block = get_irn_n(pred, -1);
1624 if ((!is_Block_dead(pred_block)) &&
1625 (get_Block_dom_depth(pred_block) > depth)) {
1627 depth = get_Block_dom_depth(pred_block);
1629 /* Avoid that the node is placed in the Start block */
1630 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1631 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1632 assert(b != get_irg_start_block(current_ir_graph));
1637 set_nodes_block(n, b);
1641 * Add predecessors of non floating nodes and non-floating predecessors
1642 * of floating nodes to worklist and fix their blocks if the are in dead block.
1644 irn_arity = get_irn_arity(n);
1646 if (get_irn_op(n) == op_End) {
1648 * Simplest case: End node. Predecessors are keep-alives,
1649 * no need to move out of dead block.
1651 for (i = -1; i < irn_arity; ++i) {
1652 ir_node *pred = get_irn_n(n, i);
1653 if (irn_not_visited(pred))
1654 pdeq_putr(worklist, pred);
1657 else if (is_Block(n)) {
1659 * Blocks: Predecessors are control flow, no need to move
1660 * them out of dead block.
1662 for (i = irn_arity - 1; i >= 0; --i) {
1663 ir_node *pred = get_irn_n(n, i);
1664 if (irn_not_visited(pred))
1665 pdeq_putr(worklist, pred);
1668 else if (is_Phi(n)) {
1670 ir_node *curr_block = get_irn_n(n, -1);
1671 int in_dead_block = is_Block_unreachable(curr_block);
1674 * Phi nodes: move nodes from dead blocks into the effective use
1675 * of the Phi-input if the Phi is not in a bad block.
1677 pred = get_irn_n(n, -1);
1678 if (irn_not_visited(pred))
1679 pdeq_putr(worklist, pred);
1681 for (i = irn_arity - 1; i >= 0; --i) {
1682 ir_node *pred = get_irn_n(n, i);
1684 if (irn_not_visited(pred)) {
1685 if (! in_dead_block &&
1686 get_irn_pinned(pred) == op_pin_state_floats &&
1687 is_Block_unreachable(get_irn_n(pred, -1))) {
1688 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1690 pdeq_putr(worklist, pred);
1696 ir_node *curr_block = get_irn_n(n, -1);
1697 int in_dead_block = is_Block_unreachable(curr_block);
1700 * All other nodes: move nodes from dead blocks into the same block.
1702 pred = get_irn_n(n, -1);
1703 if (irn_not_visited(pred))
1704 pdeq_putr(worklist, pred);
1706 for (i = irn_arity - 1; i >= 0; --i) {
1707 ir_node *pred = get_irn_n(n, i);
1709 if (irn_not_visited(pred)) {
1710 if (! in_dead_block &&
1711 get_irn_pinned(pred) == op_pin_state_floats &&
1712 is_Block_unreachable(get_irn_n(pred, -1))) {
1713 set_nodes_block(pred, curr_block);
1715 pdeq_putr(worklist, pred);
1722 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1723 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1724 * places all floating nodes reachable from its argument through floating
1725 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1727 static INLINE void place_early(pdeq *worklist) {
1729 inc_irg_visited(current_ir_graph);
1731 /* this inits the worklist */
1732 place_floats_early(get_irg_end(current_ir_graph), worklist);
1734 /* Work the content of the worklist. */
1735 while (!pdeq_empty(worklist)) {
1736 ir_node *n = pdeq_getl(worklist);
1737 if (irn_not_visited(n))
1738 place_floats_early(n, worklist);
1741 set_irg_outs_inconsistent(current_ir_graph);
1742 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1746 * Compute the deepest common ancestor of block and dca.
1748 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1752 /* we do not want to place nodes in dead blocks */
1753 if (is_Block_dead(block))
1756 /* We found a first legal placement. */
1757 if (!dca) return block;
1759 /* Find a placement that is dominates both, dca and block. */
1760 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1761 block = get_Block_idom(block);
1763 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1764 dca = get_Block_idom(dca);
1767 while (block != dca)
1768 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1773 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1774 * I.e., DCA is the block where we might place PRODUCER.
1775 * A data flow edge points from producer to consumer.
1778 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1780 ir_node *block = NULL;
1782 /* Compute the latest block into which we can place a node so that it is
1784 if (get_irn_op(consumer) == op_Phi) {
1785 /* our consumer is a Phi-node, the effective use is in all those
1786 blocks through which the Phi-node reaches producer */
1788 ir_node *phi_block = get_nodes_block(consumer);
1789 irn_arity = get_irn_arity(consumer);
1791 for (i = 0; i < irn_arity; i++) {
1792 if (get_irn_n(consumer, i) == producer) {
1793 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1795 if (! is_Block_unreachable(new_block))
1796 block = calc_dca(block, new_block);
1801 block = get_irn_n(producer, -1);
1804 assert(is_no_Block(consumer));
1805 block = get_nodes_block(consumer);
1808 /* Compute the deepest common ancestor of block and dca. */
1809 return calc_dca(dca, block);
1812 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1814 static INLINE int get_irn_loop_depth(ir_node *n) {
1815 return get_loop_depth(get_irn_loop(n));
1819 * Move n to a block with less loop depth than it's current block. The
1820 * new block must be dominated by early.
1822 * @param n the node that should be moved
1823 * @param early the earliest block we can n move to
1826 move_out_of_loops (ir_node *n, ir_node *early)
1828 ir_node *best, *dca;
1832 /* Find the region deepest in the dominator tree dominating
1833 dca with the least loop nesting depth, but still dominated
1834 by our early placement. */
1835 dca = get_nodes_block(n);
1838 while (dca != early) {
1839 dca = get_Block_idom(dca);
1840 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1841 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1845 if (best != get_nodes_block(n)) {
1847 printf("Moving out of loop: "); DDMN(n);
1848 printf(" Outermost block: "); DDMN(early);
1849 printf(" Best block: "); DDMN(best);
1850 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1852 set_nodes_block(n, best);
1857 * Find the latest legal block for N and place N into the
1858 * `optimal' Block between the latest and earliest legal block.
1859 * The `optimal' block is the dominance-deepest block of those
1860 * with the least loop-nesting-depth. This places N out of as many
1861 * loops as possible and then makes it as control dependent as
1865 place_floats_late(ir_node *n, pdeq *worklist)
1870 assert(irn_not_visited(n)); /* no multiple placement */
1872 mark_irn_visited(n);
1874 /* no need to place block nodes, control nodes are already placed. */
1875 if ((get_irn_op(n) != op_Block) &&
1877 (get_irn_mode(n) != mode_X)) {
1878 /* Remember the early_blk placement of this block to move it
1879 out of loop no further than the early_blk placement. */
1880 early_blk = get_irn_n(n, -1);
1883 * BEWARE: Here we also get code, that is live, but
1884 * was in a dead block. If the node is life, but because
1885 * of CSE in a dead block, we still might need it.
1888 /* Assure that our users are all placed, except the Phi-nodes.
1889 --- Each data flow cycle contains at least one Phi-node. We
1890 have to break the `user has to be placed before the
1891 producer' dependence cycle and the Phi-nodes are the
1892 place to do so, because we need to base our placement on the
1893 final region of our users, which is OK with Phi-nodes, as they
1894 are op_pin_state_pinned, and they never have to be placed after a
1895 producer of one of their inputs in the same block anyway. */
1896 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1897 ir_node *succ = get_irn_out(n, i);
1898 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1899 place_floats_late(succ, worklist);
1902 if (! is_Block_dead(early_blk)) {
1903 /* do only move things that where not dead */
1905 /* We have to determine the final block of this node... except for
1907 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1908 (get_irn_op(n) != op_Const) &&
1909 (get_irn_op(n) != op_SymConst)) {
1910 ir_node *dca = NULL; /* deepest common ancestor in the
1911 dominator tree of all nodes'
1912 blocks depending on us; our final
1913 placement has to dominate DCA. */
1914 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1915 ir_node *succ = get_irn_out(n, i);
1918 if (get_irn_op(succ) == op_End) {
1920 * This consumer is the End node, a keep alive edge.
1921 * This is not a real consumer, so we ignore it
1926 /* ignore if succ is in dead code */
1927 succ_blk = get_irn_n(succ, -1);
1928 if (is_Block_unreachable(succ_blk))
1930 dca = consumer_dom_dca(dca, succ, n);
1933 set_nodes_block(n, dca);
1934 move_out_of_loops(n, early_blk);
1940 /* Add predecessors of all non-floating nodes on list. (Those of floating
1941 nodes are placed already and therefore are marked.) */
1942 for (i = 0; i < get_irn_n_outs(n); i++) {
1943 ir_node *succ = get_irn_out(n, i);
1944 if (irn_not_visited(get_irn_out(n, i))) {
1945 pdeq_putr(worklist, succ);
1950 static INLINE void place_late(pdeq *worklist) {
1952 inc_irg_visited(current_ir_graph);
1954 /* This fills the worklist initially. */
1955 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1957 /* And now empty the worklist again... */
1958 while (!pdeq_empty(worklist)) {
1959 ir_node *n = pdeq_getl(worklist);
1960 if (irn_not_visited(n))
1961 place_floats_late(n, worklist);
1965 void place_code(ir_graph *irg) {
1967 ir_graph *rem = current_ir_graph;
1969 current_ir_graph = irg;
1971 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1973 /* Handle graph state */
1974 assert(get_irg_phase_state(irg) != phase_building);
1975 if (get_irg_dom_state(irg) != dom_consistent)
1978 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1979 free_loop_information(irg);
1980 construct_backedges(irg);
1983 /* Place all floating nodes as early as possible. This guarantees
1984 a legal code placement. */
1985 worklist = new_pdeq();
1986 place_early(worklist);
1988 /* place_early() invalidates the outs, place_late needs them. */
1989 compute_irg_outs(irg);
1991 /* Now move the nodes down in the dominator tree. This reduces the
1992 unnecessary executions of the node. */
1993 place_late(worklist);
1995 set_irg_outs_inconsistent(current_ir_graph);
1996 set_irg_loopinfo_inconsistent(current_ir_graph);
1998 current_ir_graph = rem;
2002 * Called by walker of remove_critical_cf_edges().
2004 * Place an empty block to an edge between a blocks of multiple
2005 * predecessors and a block of multiple successors.
2008 * @param env Environment of walker. The changed field.
2010 static void walk_critical_cf_edges(ir_node *n, void *env) {
2012 ir_node *pre, *block, *jmp;
2015 /* Block has multiple predecessors */
2016 if (is_Block(n) && (get_irn_arity(n) > 1)) {
2017 if (n == get_irg_end_block(current_ir_graph))
2018 return; /* No use to add a block here. */
2020 arity = get_irn_arity(n);
2021 for (i=0; i<arity; i++) {
2022 pre = get_irn_n(n, i);
2023 /* Predecessor has multiple successors. Insert new control flow edge. */
2024 if (op_Raise != get_irn_op(skip_Proj(pre))) {
2025 /* set predecessor of new block */
2026 block = new_Block(1, &pre);
2027 /* insert new jmp node to new block */
2028 set_cur_block(block);
2031 /* set successor of new block */
2032 set_irn_n(n, i, jmp);
2034 } /* predecessor has multiple successors */
2035 } /* for all predecessors */
2036 } /* n is a block */
2039 void remove_critical_cf_edges(ir_graph *irg) {
2041 irg_walk_graph(irg, NULL, walk_critical_cf_edges, &changed);
2044 /* control flow changed */
2045 set_irg_outs_inconsistent(irg);
2046 set_irg_extblk_inconsistent(irg);
2047 set_irg_doms_inconsistent(current_ir_graph);
2048 set_irg_loopinfo_inconsistent(current_ir_graph);