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) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
19 #include "irgraph_t.h"
32 #include "pdeq.h" /* Fuer code placement */
37 #include "irbackedge_t.h"
44 #include "iredges_t.h"
47 /*------------------------------------------------------------------*/
48 /* apply optimizations of iropt to all nodes. */
49 /*------------------------------------------------------------------*/
52 * A wrapper around optimize_inplace_2() to be called from a walker.
54 static void optimize_in_place_wrapper (ir_node *n, void *env) {
55 ir_node *optimized = optimize_in_place_2(n);
56 if (optimized != n) exchange (n, optimized);
60 * Do local optimizations for a node.
62 * @param n the IR-node where to start. Typically the End node
65 * @note current_ir_graph must be set
67 static INLINE void do_local_optimize(ir_node *n) {
68 /* Handle graph state */
69 assert(get_irg_phase_state(current_ir_graph) != phase_building);
71 if (get_opt_global_cse())
72 set_irg_pinned(current_ir_graph, op_pin_state_floats);
73 set_irg_outs_inconsistent(current_ir_graph);
74 set_irg_doms_inconsistent(current_ir_graph);
75 set_irg_loopinfo_inconsistent(current_ir_graph);
77 /* Clean the value_table in irg for the CSE. */
78 del_identities(current_ir_graph->value_table);
79 current_ir_graph->value_table = new_identities();
81 /* walk over the graph */
82 irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
85 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
86 void local_optimize_node(ir_node *n) {
87 ir_graph *rem = current_ir_graph;
88 current_ir_graph = get_irn_irg(n);
92 current_ir_graph = rem;
96 * Block-Walker: uses dominance depth to mark dead blocks.
98 static void kill_dead_blocks(ir_node *block, void *env)
100 if (get_Block_dom_depth(block) < 0) {
102 * Note that the new dominance code correctly handles
103 * the End block, i.e. it is always reachable from Start
105 set_Block_dead(block);
110 local_optimize_graph(ir_graph *irg) {
111 ir_graph *rem = current_ir_graph;
112 current_ir_graph = irg;
114 if (get_irg_dom_state(irg) == dom_consistent)
115 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
117 do_local_optimize(get_irg_end(irg));
119 current_ir_graph = rem;
123 * Enqueue all users of a node to a wait queue.
124 * Handles mode_T nodes.
126 static void enqueue_users(ir_node *n, pdeq *waitq) {
127 const ir_edge_t *edge;
129 foreach_out_edge(n, edge) {
130 ir_node *succ = get_edge_src_irn(edge);
132 if (get_irn_link(succ) != waitq) {
133 pdeq_putr(waitq, succ);
134 set_irn_link(succ, waitq);
136 if (get_irn_mode(succ) == mode_T) {
137 /* A mode_T node has Proj's. Because most optimizations
138 run on the Proj's we have to enqueue them also. */
139 enqueue_users(succ, waitq);
145 * Data flow optimization walker.
146 * Optimizes all nodes and enqueue it's users
149 static void opt_walker(ir_node *n, void *env) {
153 optimized = optimize_in_place_2(n);
154 set_irn_link(optimized, NULL);
156 if (optimized != n) {
157 enqueue_users(n, waitq);
158 exchange(n, optimized);
162 /* Applies local optimizations to all nodes in the graph until fixpoint. */
163 void optimize_graph_df(ir_graph *irg) {
164 pdeq *waitq = new_pdeq();
165 int state = edges_activated(irg);
166 ir_graph *rem = current_ir_graph;
168 current_ir_graph = irg;
173 if (get_opt_global_cse())
174 set_irg_pinned(current_ir_graph, op_pin_state_floats);
176 /* Clean the value_table in irg for the CSE. */
177 del_identities(irg->value_table);
178 irg->value_table = new_identities();
180 if (get_irg_dom_state(irg) == dom_consistent)
181 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
183 /* invalidate info */
184 set_irg_outs_inconsistent(irg);
185 set_irg_doms_inconsistent(irg);
186 set_irg_loopinfo_inconsistent(irg);
188 /* walk over the graph */
189 irg_walk_graph(irg, NULL, opt_walker, waitq);
191 /* finish the wait queue */
192 while (! pdeq_empty(waitq)) {
193 ir_node *n = pdeq_getl(waitq);
195 opt_walker(n, waitq);
201 edges_deactivate(irg);
203 current_ir_graph = rem;
207 /*------------------------------------------------------------------*/
208 /* Routines for dead node elimination / copying garbage collection */
209 /* of the obstack. */
210 /*------------------------------------------------------------------*/
213 * Remember the new node in the old node by using a field all nodes have.
215 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
218 * Get this new node, before the old node is forgotten.
220 #define get_new_node(oldn) get_irn_link(oldn)
223 * Check if a new node was set.
225 #define has_new_node(n) (get_new_node(n) != NULL)
228 * We use the block_visited flag to mark that we have computed the
229 * number of useful predecessors for this block.
230 * Further we encode the new arity in this flag in the old blocks.
231 * Remembering the arity is useful, as it saves a lot of pointer
232 * accesses. This function is called for all Phi and Block nodes
236 compute_new_arity(ir_node *b) {
237 int i, res, irn_arity;
240 irg_v = get_irg_block_visited(current_ir_graph);
241 block_v = get_Block_block_visited(b);
242 if (block_v >= irg_v) {
243 /* we computed the number of preds for this block and saved it in the
245 return block_v - irg_v;
247 /* compute the number of good predecessors */
248 res = irn_arity = get_irn_arity(b);
249 for (i = 0; i < irn_arity; i++)
250 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
251 /* save it in the flag. */
252 set_Block_block_visited(b, irg_v + res);
258 * Copies the node to the new obstack. The Ins of the new node point to
259 * the predecessors on the old obstack. For block/phi nodes not all
260 * predecessors might be copied. n->link points to the new node.
261 * For Phi and Block nodes the function allocates in-arrays with an arity
262 * only for useful predecessors. The arity is determined by counting
263 * the non-bad predecessors of the block.
265 * @param n The node to be copied
266 * @param env if non-NULL, the node number attribute will be copied to the new node
268 * Note: Also used for loop unrolling.
270 static void copy_node(ir_node *n, void *env) {
273 ir_op *op = get_irn_op(n);
274 int copy_node_nr = env != NULL;
276 /* The end node looses it's flexible in array. This doesn't matter,
277 as dead node elimination builds End by hand, inlineing doesn't use
279 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
282 /* node copied already */
284 } else if (op == op_Block) {
286 new_arity = compute_new_arity(n);
287 n->attr.block.graph_arr = NULL;
289 block = get_nodes_block(n);
291 new_arity = compute_new_arity(block);
293 new_arity = get_irn_arity(n);
296 nn = new_ir_node(get_irn_dbg_info(n),
303 /* Copy the attributes. These might point to additional data. If this
304 was allocated on the old obstack the pointers now are dangling. This
305 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
306 copy_node_attr(n, nn);
307 new_backedge_info(nn);
311 /* for easier debugging, we want to copy the node numbers too */
312 nn->node_nr = n->node_nr;
317 hook_dead_node_elim_subst(current_ir_graph, n, nn);
321 * Copies new predecessors of old node to new node remembered in link.
322 * Spare the Bad predecessors of Phi and Block nodes.
325 copy_preds(ir_node *n, void *env) {
329 nn = get_new_node(n);
331 /* printf("\n old node: "); DDMSG2(n);
332 printf(" new node: "); DDMSG2(nn);
333 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
336 /* Don't copy Bad nodes. */
338 irn_arity = get_irn_arity(n);
339 for (i = 0; i < irn_arity; i++)
340 if (! is_Bad(get_irn_n(n, i))) {
341 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
342 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
345 /* repair the block visited flag from above misuse. Repair it in both
346 graphs so that the old one can still be used. */
347 set_Block_block_visited(nn, 0);
348 set_Block_block_visited(n, 0);
349 /* Local optimization could not merge two subsequent blocks if
350 in array contained Bads. Now it's possible.
351 We don't call optimize_in_place as it requires
352 that the fields in ir_graph are set properly. */
353 if ((get_opt_control_flow_straightening()) &&
354 (get_Block_n_cfgpreds(nn) == 1) &&
355 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
356 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
358 /* Jmp jumps into the block it is in -- deal self cycle. */
359 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
360 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
365 } else if (get_irn_op(n) == op_Phi) {
366 /* Don't copy node if corresponding predecessor in block is Bad.
367 The Block itself should not be Bad. */
368 block = get_nodes_block(n);
369 set_irn_n(nn, -1, get_new_node(block));
371 irn_arity = get_irn_arity(n);
372 for (i = 0; i < irn_arity; i++)
373 if (! is_Bad(get_irn_n(block, i))) {
374 set_irn_n(nn, j, get_new_node(get_irn_n(n, i)));
375 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
378 /* If the pre walker reached this Phi after the post walker visited the
379 block block_visited is > 0. */
380 set_Block_block_visited(get_nodes_block(n), 0);
381 /* Compacting the Phi's ins might generate Phis with only one
383 if (get_irn_arity(nn) == 1)
384 exchange(nn, get_irn_n(nn, 0));
386 irn_arity = get_irn_arity(n);
387 for (i = -1; i < irn_arity; i++)
388 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
390 /* Now the new node is complete. We can add it to the hash table for CSE.
391 @@@ inlining aborts if we identify End. Why? */
392 if (get_irn_op(nn) != op_End)
393 add_identities(current_ir_graph->value_table, nn);
397 * Copies the graph recursively, compacts the keep-alives of the end node.
399 * @param irg the graph to be copied
400 * @param copy_node_nr If non-zero, the node number will be copied
402 static void copy_graph(ir_graph *irg, int copy_node_nr) {
403 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
404 ir_node *ka; /* keep alive */
408 /* Some nodes must be copied by hand, sigh */
409 vfl = get_irg_visited(irg);
410 set_irg_visited(irg, vfl + 1);
412 oe = get_irg_end(irg);
413 mark_irn_visited(oe);
414 /* copy the end node by hand, allocate dynamic in array! */
415 ne = new_ir_node(get_irn_dbg_info(oe),
422 /* Copy the attributes. Well, there might be some in the future... */
423 copy_node_attr(oe, ne);
424 set_new_node(oe, ne);
426 /* copy the Bad node */
427 ob = get_irg_bad(irg);
428 mark_irn_visited(ob);
429 nb = new_ir_node(get_irn_dbg_info(ob),
436 copy_node_attr(ob, nb);
437 set_new_node(ob, nb);
439 /* copy the NoMem node */
440 om = get_irg_no_mem(irg);
441 mark_irn_visited(om);
442 nm = new_ir_node(get_irn_dbg_info(om),
449 copy_node_attr(om, nm);
450 set_new_node(om, nm);
452 /* copy the live nodes */
453 set_irg_visited(irg, vfl);
454 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
456 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
458 /* visit the anchors as well */
459 for (i = anchor_max - 1; i >= 0; --i) {
460 ir_node *n = irg->anchors[i];
462 if (n && (get_irn_visited(n) <= vfl)) {
463 set_irg_visited(irg, vfl);
464 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
468 /* copy_preds for the end node ... */
469 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
471 /*- ... and now the keep alives. -*/
472 /* First pick the not marked block nodes and walk them. We must pick these
473 first as else we will oversee blocks reachable from Phis. */
474 irn_arity = get_End_n_keepalives(oe);
475 for (i = 0; i < irn_arity; i++) {
476 ka = get_End_keepalive(oe, i);
478 if (get_irn_visited(ka) <= vfl) {
479 /* We must keep the block alive and copy everything reachable */
480 set_irg_visited(irg, vfl);
481 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
483 add_End_keepalive(ne, get_new_node(ka));
487 /* Now pick other nodes. Here we will keep all! */
488 irn_arity = get_End_n_keepalives(oe);
489 for (i = 0; i < irn_arity; i++) {
490 ka = get_End_keepalive(oe, i);
492 if (get_irn_visited(ka) <= vfl) {
493 /* We didn't copy the node yet. */
494 set_irg_visited(irg, vfl);
495 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
497 add_End_keepalive(ne, get_new_node(ka));
501 /* start block sometimes only reached after keep alives */
502 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
503 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
507 * Copies the graph reachable from current_ir_graph->end to the obstack
508 * in current_ir_graph and fixes the environment.
509 * Then fixes the fields in current_ir_graph containing nodes of the
512 * @param copy_node_nr If non-zero, the node number will be copied
515 copy_graph_env(int copy_node_nr) {
516 ir_graph *irg = current_ir_graph;
517 ir_node *old_end, *n;
520 /* remove end_except and end_reg nodes */
521 old_end = get_irg_end(irg);
522 set_irg_end_except (irg, old_end);
523 set_irg_end_reg (irg, old_end);
525 /* Not all nodes remembered in irg might be reachable
526 from the end node. Assure their link is set to NULL, so that
527 we can test whether new nodes have been computed. */
528 for (i = anchor_max - 1; i >= 0; --i) {
530 set_new_node(irg->anchors[i], NULL);
532 /* we use the block walk flag for removing Bads from Blocks ins. */
533 inc_irg_block_visited(irg);
536 copy_graph(irg, copy_node_nr);
538 /* fix the fields in irg */
539 old_end = get_irg_end(irg);
540 for (i = anchor_max - 1; i >= 0; --i) {
543 irg->anchors[i] = get_new_node(n);
549 * Copies all reachable nodes to a new obstack. Removes bad inputs
550 * from block nodes and the corresponding inputs from Phi nodes.
551 * Merges single exit blocks with single entry blocks and removes
553 * Adds all new nodes to a new hash table for CSE. Does not
554 * perform CSE, so the hash table might contain common subexpressions.
557 dead_node_elimination(ir_graph *irg) {
558 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
560 int rem_ipview = get_interprocedural_view();
561 struct obstack *graveyard_obst = NULL;
562 struct obstack *rebirth_obst = NULL;
563 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
565 /* inform statistics that we started a dead-node elimination run */
566 hook_dead_node_elim(irg, 1);
568 /* Remember external state of current_ir_graph. */
569 rem = current_ir_graph;
570 current_ir_graph = irg;
571 set_interprocedural_view(0);
573 assert(get_irg_phase_state(irg) != phase_building);
575 /* Handle graph state */
576 free_callee_info(irg);
580 /* @@@ so far we loose loops when copying */
581 free_loop_information(irg);
583 set_irg_doms_inconsistent(irg);
585 /* A quiet place, where the old obstack can rest in peace,
586 until it will be cremated. */
587 graveyard_obst = irg->obst;
589 /* A new obstack, where the reachable nodes will be copied to. */
590 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
591 irg->obst = rebirth_obst;
592 obstack_init(irg->obst);
593 irg->last_node_idx = 0;
595 /* We also need a new value table for CSE */
596 del_identities(irg->value_table);
597 irg->value_table = new_identities();
599 /* Copy the graph from the old to the new obstack */
600 copy_graph_env(/*copy_node_nr=*/1);
602 /* Free memory from old unoptimized obstack */
603 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
604 xfree (graveyard_obst); /* ... then free it. */
606 /* inform statistics that the run is over */
607 hook_dead_node_elim(irg, 0);
609 current_ir_graph = rem;
610 set_interprocedural_view(rem_ipview);
615 * Relink bad predecessors of a block and store the old in array to the
616 * link field. This function is called by relink_bad_predecessors().
617 * The array of link field starts with the block operand at position 0.
618 * If block has bad predecessors, create a new in array without bad preds.
619 * Otherwise let in array untouched.
621 static void relink_bad_block_predecessors(ir_node *n, void *env) {
622 ir_node **new_in, *irn;
623 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
625 /* if link field of block is NULL, look for bad predecessors otherwise
626 this is already done */
627 if (get_irn_op(n) == op_Block &&
628 get_irn_link(n) == NULL) {
630 /* save old predecessors in link field (position 0 is the block operand)*/
631 set_irn_link(n, get_irn_in(n));
633 /* count predecessors without bad nodes */
634 old_irn_arity = get_irn_arity(n);
635 for (i = 0; i < old_irn_arity; i++)
636 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
638 /* arity changing: set new predecessors without bad nodes */
639 if (new_irn_arity < old_irn_arity) {
640 /* Get new predecessor array. We do not resize the array, as we must
641 keep the old one to update Phis. */
642 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
644 /* set new predecessors in array */
647 for (i = 0; i < old_irn_arity; i++) {
648 irn = get_irn_n(n, i);
650 new_in[new_irn_n] = irn;
651 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
655 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
656 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
659 } /* ir node has bad predecessors */
661 } /* Block is not relinked */
665 * Relinks Bad predecessors from Blocks and Phis called by walker
666 * remove_bad_predecesors(). If n is a Block, call
667 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
668 * function of Phi's Block. If this block has bad predecessors, relink preds
671 static void relink_bad_predecessors(ir_node *n, void *env) {
672 ir_node *block, **old_in;
673 int i, old_irn_arity, new_irn_arity;
675 /* relink bad predecessors of a block */
676 if (get_irn_op(n) == op_Block)
677 relink_bad_block_predecessors(n, env);
679 /* If Phi node relink its block and its predecessors */
680 if (get_irn_op(n) == op_Phi) {
682 /* Relink predecessors of phi's block */
683 block = get_nodes_block(n);
684 if (get_irn_link(block) == NULL)
685 relink_bad_block_predecessors(block, env);
687 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
688 old_irn_arity = ARR_LEN(old_in);
690 /* Relink Phi predecessors if count of predecessors changed */
691 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
692 /* set new predecessors in array
693 n->in[0] remains the same block */
695 for(i = 1; i < old_irn_arity; i++)
696 if (!is_Bad((ir_node *)old_in[i])) {
697 n->in[new_irn_arity] = n->in[i];
698 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
702 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
703 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
706 } /* n is a Phi node */
710 * Removes Bad Bad predecessors from Blocks and the corresponding
711 * inputs to Phi nodes as in dead_node_elimination but without
713 * On walking up set the link field to NULL, on walking down call
714 * relink_bad_predecessors() (This function stores the old in array
715 * to the link field and sets a new in array if arity of predecessors
718 void remove_bad_predecessors(ir_graph *irg) {
719 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
726 __)|_| | \_/ | \_/(/_ |_/\__|__
728 The following stuff implements a facility that automatically patches
729 registered ir_node pointers to the new node when a dead node elimination occurs.
732 struct _survive_dce_t {
736 hook_entry_t dead_node_elim;
737 hook_entry_t dead_node_elim_subst;
740 typedef struct _survive_dce_list_t {
741 struct _survive_dce_list_t *next;
743 } survive_dce_list_t;
745 static void dead_node_hook(void *context, ir_graph *irg, int start)
747 survive_dce_t *sd = context;
749 /* Create a new map before the dead node elimination is performed. */
751 sd->new_places = pmap_create_ex(pmap_count(sd->places));
754 /* Patch back all nodes if dead node elimination is over and something is to be done. */
756 pmap_destroy(sd->places);
757 sd->places = sd->new_places;
758 sd->new_places = NULL;
763 * Hook called when dead node elimination replaces old by nw.
765 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw)
767 survive_dce_t *sd = context;
768 survive_dce_list_t *list = pmap_get(sd->places, old);
770 /* If the node is to be patched back, write the new address to all registered locations. */
772 survive_dce_list_t *p;
774 for(p = list; p; p = p->next)
777 pmap_insert(sd->new_places, nw, list);
782 * Make a new Survive DCE environment.
784 survive_dce_t *new_survive_dce(void)
786 survive_dce_t *res = xmalloc(sizeof(res[0]));
787 obstack_init(&res->obst);
788 res->places = pmap_create();
789 res->new_places = NULL;
791 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
792 res->dead_node_elim.context = res;
793 res->dead_node_elim.next = NULL;
795 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
796 res->dead_node_elim_subst.context = res;
797 res->dead_node_elim_subst.next = NULL;
799 register_hook(hook_dead_node_elim, &res->dead_node_elim);
800 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
805 * Free a Survive DCE environment.
807 void free_survive_dce(survive_dce_t *sd)
809 obstack_free(&sd->obst, NULL);
810 pmap_destroy(sd->places);
811 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
812 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
817 * Register a node pointer to be patched upon DCE.
818 * When DCE occurs, the node pointer specified by @p place will be
819 * patched to the new address of the node it is pointing to.
821 * @param sd The Survive DCE environment.
822 * @param place The address of the node pointer.
824 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place)
827 ir_node *irn = *place;
828 survive_dce_list_t *curr = pmap_get(sd->places, irn);
829 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw));
834 pmap_insert(sd->places, irn, nw);
838 /*--------------------------------------------------------------------*/
839 /* Functionality for inlining */
840 /*--------------------------------------------------------------------*/
843 * Copy node for inlineing. Updates attributes that change when
844 * inlineing but not for dead node elimination.
846 * Copies the node by calling copy_node() and then updates the entity if
847 * it's a local one. env must be a pointer of the frame type of the
848 * inlined procedure. The new entities must be in the link field of
852 copy_node_inline (ir_node *n, void *env) {
854 ir_type *frame_tp = (ir_type *)env;
857 if (get_irn_op(n) == op_Sel) {
858 nn = get_new_node (n);
860 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
861 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
863 } else if (get_irn_op(n) == op_Block) {
864 nn = get_new_node (n);
865 nn->attr.block.irg = current_ir_graph;
870 * Check if the varlue_arg_base is used.
872 static void find_addr(ir_node *node, void *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);
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))
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 int inline_method(ir_node *call, ir_graph *called_graph) {
919 ir_node *post_call, *post_bl;
920 ir_node *in[pn_Start_max];
921 ir_node *end, *end_bl;
925 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
927 ir_type *called_frame;
928 irg_inline_property prop = get_irg_inline_property(called_graph);
930 if ( (prop < irg_inline_forced) &&
931 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
933 /* Do not inline variadic functions. */
934 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
937 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
938 get_method_n_params(get_Call_type(call)));
941 * currently, we cannot inline two cases:
942 * - call with compound arguments
943 * - graphs that take the address of a parameter
945 if (! can_inline(call, called_graph))
948 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
949 rem_opt = get_opt_optimize();
952 /* Handle graph state */
953 assert(get_irg_phase_state(current_ir_graph) != phase_building);
954 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
955 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
956 set_irg_outs_inconsistent(current_ir_graph);
957 set_irg_extblk_inconsistent(current_ir_graph);
958 set_irg_doms_inconsistent(current_ir_graph);
959 set_irg_loopinfo_inconsistent(current_ir_graph);
960 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
962 /* -- Check preconditions -- */
963 assert(is_Call(call));
964 /* @@@ does not work for InterfaceIII.java after cgana
965 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
966 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
967 get_Call_type(call)));
969 if (called_graph == current_ir_graph) {
970 set_optimize(rem_opt);
974 /* here we know we WILL inline, so inform the statistics */
975 hook_inline(call, called_graph);
977 /* -- Decide how to handle exception control flow: Is there a handler
978 for the Call node, or do we branch directly to End on an exception?
980 0 There is a handler.
982 2 Exception handling not represented in Firm. -- */
984 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
985 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
986 assert(is_Proj(proj));
987 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
988 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
990 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
991 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
992 else { exc_handling = 2; } /* !Mproj && !Xproj */
997 the procedure and later replaces the Start node of the called graph.
998 Post_call is the old Call node and collects the results of the called
999 graph. Both will end up being a tuple. -- */
1000 post_bl = get_nodes_block(call);
1001 set_irg_current_block(current_ir_graph, post_bl);
1002 /* XxMxPxPxPxT of Start + parameter of Call */
1003 in[pn_Start_X_initial_exec] = new_Jmp();
1004 in[pn_Start_M] = get_Call_mem(call);
1005 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
1006 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
1007 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
1008 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
1009 /* in[pn_Start_P_value_arg_base] = ??? */
1010 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
1011 pre_call = new_Tuple(pn_Start_max - 1, in);
1015 The new block gets the ins of the old block, pre_call and all its
1016 predecessors and all Phi nodes. -- */
1017 part_block(pre_call);
1019 /* -- Prepare state for dead node elimination -- */
1020 /* Visited flags in calling irg must be >= flag in called irg.
1021 Else walker and arity computation will not work. */
1022 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1023 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1024 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1025 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1026 /* Set pre_call as new Start node in link field of the start node of
1027 calling graph and pre_calls block as new block for the start block
1029 Further mark these nodes so that they are not visited by the
1031 set_irn_link(get_irg_start(called_graph), pre_call);
1032 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1033 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1034 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1035 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1036 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1038 /* Initialize for compaction of in arrays */
1039 inc_irg_block_visited(current_ir_graph);
1041 /* -- Replicate local entities of the called_graph -- */
1042 /* copy the entities. */
1043 called_frame = get_irg_frame_type(called_graph);
1044 for (i = 0; i < get_class_n_members(called_frame); i++) {
1045 entity *new_ent, *old_ent;
1046 old_ent = get_class_member(called_frame, i);
1047 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1048 set_entity_link(old_ent, new_ent);
1051 /* visited is > than that of called graph. With this trick visited will
1052 remain unchanged so that an outer walker, e.g., searching the call nodes
1053 to inline, calling this inline will not visit the inlined nodes. */
1054 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1056 /* -- Performing dead node elimination inlines the graph -- */
1057 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1059 /* @@@ endless loops are not copied!! -- they should be, I think... */
1060 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1061 get_irg_frame_type(called_graph));
1063 /* Repair called_graph */
1064 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1065 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1066 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1068 /* -- Merge the end of the inlined procedure with the call site -- */
1069 /* We will turn the old Call node into a Tuple with the following
1072 0: Phi of all Memories of Return statements.
1073 1: Jmp from new Block that merges the control flow from all exception
1074 predecessors of the old end block.
1075 2: Tuple of all arguments.
1076 3: Phi of Exception memories.
1077 In case the old Call directly branches to End on an exception we don't
1078 need the block merging all exceptions nor the Phi of the exception
1082 /* -- Precompute some values -- */
1083 end_bl = get_new_node(get_irg_end_block(called_graph));
1084 end = get_new_node(get_irg_end(called_graph));
1085 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1086 n_res = get_method_n_ress(get_Call_type(call));
1088 res_pred = xmalloc (n_res * sizeof(*res_pred));
1089 cf_pred = xmalloc (arity * sizeof(*res_pred));
1091 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1093 /* -- archive keepalives -- */
1094 irn_arity = get_irn_arity(end);
1095 for (i = 0; i < irn_arity; i++)
1096 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
1098 /* The new end node will die. We need not free as the in array is on the obstack:
1099 copy_node() only generated 'D' arrays. */
1101 /* -- Replace Return nodes by Jump nodes. -- */
1103 for (i = 0; i < arity; i++) {
1105 ret = get_irn_n(end_bl, i);
1106 if (is_Return(ret)) {
1107 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1111 set_irn_in(post_bl, n_ret, cf_pred);
1113 /* -- Build a Tuple for all results of the method.
1114 Add Phi node if there was more than one Return. -- */
1115 turn_into_tuple(post_call, 4);
1116 /* First the Memory-Phi */
1118 for (i = 0; i < arity; i++) {
1119 ret = get_irn_n(end_bl, i);
1120 if (is_Return(ret)) {
1121 cf_pred[n_ret] = get_Return_mem(ret);
1125 phi = new_Phi(n_ret, cf_pred, mode_M);
1126 set_Tuple_pred(call, pn_Call_M_regular, phi);
1127 /* Conserve Phi-list for further inlinings -- but might be optimized */
1128 if (get_nodes_block(phi) == post_bl) {
1129 set_irn_link(phi, get_irn_link(post_bl));
1130 set_irn_link(post_bl, phi);
1132 /* Now the real results */
1134 for (j = 0; j < n_res; j++) {
1136 for (i = 0; i < arity; i++) {
1137 ret = get_irn_n(end_bl, i);
1138 if (get_irn_op(ret) == op_Return) {
1139 cf_pred[n_ret] = get_Return_res(ret, j);
1144 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1148 /* Conserve Phi-list for further inlinings -- but might be optimized */
1149 if (get_nodes_block(phi) == post_bl) {
1150 set_irn_link(phi, get_irn_link(post_bl));
1151 set_irn_link(post_bl, phi);
1154 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1156 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1158 /* Finally the exception control flow.
1159 We have two (three) possible situations:
1160 First if the Call branches to an exception handler: We need to add a Phi node to
1161 collect the memory containing the exception objects. Further we need
1162 to add another block to get a correct representation of this Phi. To
1163 this block we add a Jmp that resolves into the X output of the Call
1164 when the Call is turned into a tuple.
1165 Second the Call branches to End, the exception is not handled. Just
1166 add all inlined exception branches to the End node.
1167 Third: there is no Exception edge at all. Handle as case two. */
1168 if (exc_handling == 0) {
1170 for (i = 0; i < arity; i++) {
1172 ret = get_irn_n(end_bl, i);
1173 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1174 cf_pred[n_exc] = ret;
1179 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1180 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1181 /* The Phi for the memories with the exception objects */
1183 for (i = 0; i < arity; i++) {
1185 ret = skip_Proj(get_irn_n(end_bl, i));
1187 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1189 } else if (is_fragile_op(ret)) {
1190 /* We rely that all cfops have the memory output at the same position. */
1191 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1193 } else if (get_irn_op(ret) == op_Raise) {
1194 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1198 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1200 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1201 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1204 ir_node *main_end_bl;
1205 int main_end_bl_arity;
1206 ir_node **end_preds;
1208 /* assert(exc_handling == 1 || no exceptions. ) */
1210 for (i = 0; i < arity; i++) {
1211 ir_node *ret = get_irn_n(end_bl, i);
1213 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1214 cf_pred[n_exc] = ret;
1218 main_end_bl = get_irg_end_block(current_ir_graph);
1219 main_end_bl_arity = get_irn_arity(main_end_bl);
1220 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1222 for (i = 0; i < main_end_bl_arity; ++i)
1223 end_preds[i] = get_irn_n(main_end_bl, i);
1224 for (i = 0; i < n_exc; ++i)
1225 end_preds[main_end_bl_arity + i] = cf_pred[i];
1226 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1227 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1228 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1234 #if 0 /* old. now better, correcter, faster implementation. */
1236 /* -- If the exception control flow from the inlined Call directly
1237 branched to the end block we now have the following control
1238 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1239 remove the Jmp along with it's empty block and add Jmp's
1240 predecessors as predecessors of this end block. No problem if
1241 there is no exception, because then branches Bad to End which
1243 @@@ can't we know this beforehand: by getting the Proj(1) from
1244 the Call link list and checking whether it goes to Proj. */
1245 /* find the problematic predecessor of the end block. */
1246 end_bl = get_irg_end_block(current_ir_graph);
1247 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1248 cf_op = get_Block_cfgpred(end_bl, i);
1249 if (get_irn_op(cf_op) == op_Proj) {
1250 cf_op = get_Proj_pred(cf_op);
1251 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1252 /* There are unoptimized tuples from inlineing before when no exc */
1253 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1254 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1255 assert(get_irn_op(cf_op) == op_Jmp);
1261 if (i < get_Block_n_cfgpreds(end_bl)) {
1262 bl = get_nodes_block(cf_op);
1263 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1264 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1265 for (j = 0; j < i; j++)
1266 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1267 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1268 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1269 for (j = j; j < arity; j++)
1270 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1271 set_irn_in(end_bl, arity, cf_pred);
1273 /* Remove the exception pred from post-call Tuple. */
1274 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1279 /* -- Turn CSE back on. -- */
1280 set_optimize(rem_opt);
1285 /********************************************************************/
1286 /* Apply inlineing to small methods. */
1287 /********************************************************************/
1289 /* It makes no sense to inline too many calls in one procedure. Anyways,
1290 I didn't get a version with NEW_ARR_F to run. */
1291 #define MAX_INLINE 1024
1294 * environment for inlining small irgs
1296 typedef struct _inline_env_t {
1298 ir_node *calls[MAX_INLINE];
1302 * Returns the irg called from a Call node. If the irg is not
1303 * known, NULL is returned.
1305 static ir_graph *get_call_called_irg(ir_node *call) {
1307 ir_graph *called_irg = NULL;
1309 assert(is_Call(call));
1311 addr = get_Call_ptr(call);
1312 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1313 called_irg = get_entity_irg(get_SymConst_entity(addr));
1319 static void collect_calls(ir_node *call, void *env) {
1322 if (! is_Call(call)) return;
1324 addr = get_Call_ptr(call);
1326 if (get_irn_op(addr) == op_SymConst) {
1327 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1328 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1329 inline_env_t *ienv = (inline_env_t *)env;
1330 if (called_irg && ienv->pos < MAX_INLINE) {
1331 /* The Call node calls a locally defined method. Remember to inline. */
1332 ienv->calls[ienv->pos++] = call;
1339 * Inlines all small methods at call sites where the called address comes
1340 * from a Const node that references the entity representing the called
1342 * The size argument is a rough measure for the code size of the method:
1343 * Methods where the obstack containing the firm graph is smaller than
1346 void inline_small_irgs(ir_graph *irg, int size) {
1348 ir_graph *rem = current_ir_graph;
1349 inline_env_t env /* = {0, NULL}*/;
1351 if (!(get_opt_optimize() && get_opt_inline())) return;
1353 current_ir_graph = irg;
1354 /* Handle graph state */
1355 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1356 free_callee_info(current_ir_graph);
1358 /* Find Call nodes to inline.
1359 (We can not inline during a walk of the graph, as inlineing the same
1360 method several times changes the visited flag of the walked graph:
1361 after the first inlineing visited of the callee equals visited of
1362 the caller. With the next inlineing both are increased.) */
1364 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1366 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1367 /* There are calls to inline */
1368 collect_phiprojs(irg);
1369 for (i = 0; i < env.pos; i++) {
1371 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1372 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1373 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1374 inline_method(env.calls[i], callee);
1379 current_ir_graph = rem;
1383 * Environment for inlining irgs.
1386 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1387 int n_nodes_orig; /**< for statistics */
1388 eset *call_nodes; /**< All call nodes in this graph */
1390 int n_call_nodes_orig; /**< for statistics */
1391 int n_callers; /**< Number of known graphs that call this graphs. */
1392 int n_callers_orig; /**< for statistics */
1396 * Allocate a new environment for inlining.
1398 static inline_irg_env *new_inline_irg_env(void) {
1399 inline_irg_env *env = xmalloc(sizeof(*env));
1400 env->n_nodes = -2; /* do not count count Start, End */
1401 env->n_nodes_orig = -2; /* do not count Start, End */
1402 env->call_nodes = eset_create();
1403 env->n_call_nodes = 0;
1404 env->n_call_nodes_orig = 0;
1406 env->n_callers_orig = 0;
1411 * destroy an environment for inlining.
1413 static void free_inline_irg_env(inline_irg_env *env) {
1414 eset_destroy(env->call_nodes);
1419 * post-walker: collect all calls in the inline-environment
1420 * of a graph and sum some statistics.
1422 static void collect_calls2(ir_node *call, void *env) {
1423 inline_irg_env *x = (inline_irg_env *)env;
1424 ir_op *op = get_irn_op(call);
1427 /* count meaningful nodes in irg */
1428 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1433 if (op != op_Call) return;
1435 /* collect all call nodes */
1436 eset_insert(x->call_nodes, call);
1438 x->n_call_nodes_orig++;
1440 /* count all static callers */
1441 callee = get_call_called_irg(call);
1443 inline_irg_env *callee_env = get_irg_link(callee);
1444 callee_env->n_callers++;
1445 callee_env->n_callers_orig++;
1450 * Returns TRUE if the number of callers in 0 in the irg's environment,
1451 * hence this irg is a leave.
1453 INLINE static int is_leave(ir_graph *irg) {
1454 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1458 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1460 INLINE static int is_smaller(ir_graph *callee, int size) {
1461 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1466 * Inlines small leave methods at call sites where the called address comes
1467 * from a Const node that references the entity representing the called
1469 * The size argument is a rough measure for the code size of the method:
1470 * Methods where the obstack containing the firm graph is smaller than
1473 void inline_leave_functions(int maxsize, int leavesize, int size) {
1474 inline_irg_env *env;
1475 int i, n_irgs = get_irp_n_irgs();
1476 ir_graph *rem = current_ir_graph;
1479 if (!(get_opt_optimize() && get_opt_inline())) return;
1481 /* extend all irgs by a temporary data structure for inlining. */
1482 for (i = 0; i < n_irgs; ++i)
1483 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1485 /* Precompute information in temporary data structure. */
1486 for (i = 0; i < n_irgs; ++i) {
1487 current_ir_graph = get_irp_irg(i);
1488 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1489 free_callee_info(current_ir_graph);
1491 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1492 get_irg_link(current_ir_graph));
1495 /* -- and now inline. -- */
1497 /* Inline leaves recursively -- we might construct new leaves. */
1498 while (did_inline) {
1501 for (i = 0; i < n_irgs; ++i) {
1503 int phiproj_computed = 0;
1505 current_ir_graph = get_irp_irg(i);
1506 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1508 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1511 if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
1512 callee = get_call_called_irg(call);
1514 if (env->n_nodes > maxsize) continue; // break;
1516 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1517 if (!phiproj_computed) {
1518 phiproj_computed = 1;
1519 collect_phiprojs(current_ir_graph);
1521 did_inline = inline_method(call, callee);
1524 /* Do some statistics */
1525 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1526 env->n_call_nodes --;
1527 env->n_nodes += callee_env->n_nodes;
1528 callee_env->n_callers--;
1535 /* inline other small functions. */
1536 for (i = 0; i < n_irgs; ++i) {
1539 int phiproj_computed = 0;
1541 current_ir_graph = get_irp_irg(i);
1542 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1544 /* we can not walk and change a set, nor remove from it.
1546 walkset = env->call_nodes;
1547 env->call_nodes = eset_create();
1548 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1551 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1552 callee = get_call_called_irg(call);
1555 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1556 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1557 if (!phiproj_computed) {
1558 phiproj_computed = 1;
1559 collect_phiprojs(current_ir_graph);
1561 if (inline_method(call, callee)) {
1562 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1563 env->n_call_nodes--;
1564 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1565 env->n_call_nodes += callee_env->n_call_nodes;
1566 env->n_nodes += callee_env->n_nodes;
1567 callee_env->n_callers--;
1570 eset_insert(env->call_nodes, call);
1573 eset_destroy(walkset);
1576 for (i = 0; i < n_irgs; ++i) {
1577 current_ir_graph = get_irp_irg(i);
1579 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1580 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1581 (env->n_callers_orig != env->n_callers))
1582 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1583 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1584 env->n_callers_orig, env->n_callers,
1585 get_entity_name(get_irg_entity(current_ir_graph)));
1587 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1590 current_ir_graph = rem;
1593 /*******************************************************************/
1594 /* Code Placement. Pins all floating nodes to a block where they */
1595 /* will be executed only if needed. */
1596 /*******************************************************************/
1599 * Returns non-zero, is a block is not reachable from Start.
1601 * @param block the block to test
1604 is_Block_unreachable(ir_node *block) {
1605 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1609 * Find the earliest correct block for N. --- Place N into the
1610 * same Block as its dominance-deepest Input.
1612 * We have to avoid calls to get_nodes_block() here
1613 * because the graph is floating.
1615 * move_out_of_loops() expects that place_floats_early() have placed
1616 * all "living" nodes into a living block. That's why we must
1617 * move nodes in dead block with "live" successors into a valid
1619 * We move them just into the same block as it's successor (or
1620 * in case of a Phi into the effective use block). For Phi successors,
1621 * this may still be a dead block, but then there is no real use, as
1622 * the control flow will be dead later.
1625 place_floats_early(ir_node *n, pdeq *worklist)
1629 /* we must not run into an infinite loop */
1630 assert(irn_not_visited(n));
1631 mark_irn_visited(n);
1633 /* Place floating nodes. */
1634 if (get_irn_pinned(n) == op_pin_state_floats) {
1635 ir_node *curr_block = get_irn_n(n, -1);
1636 int in_dead_block = is_Block_unreachable(curr_block);
1638 ir_node *b = NULL; /* The block to place this node in */
1640 assert(get_irn_op(n) != op_Block);
1642 if ((get_irn_op(n) == op_Const) ||
1643 (get_irn_op(n) == op_SymConst) ||
1645 (get_irn_op(n) == op_Unknown)) {
1646 /* These nodes will not be placed by the loop below. */
1647 b = get_irg_start_block(current_ir_graph);
1651 /* find the block for this node. */
1652 irn_arity = get_irn_arity(n);
1653 for (i = 0; i < irn_arity; i++) {
1654 ir_node *pred = get_irn_n(n, i);
1655 ir_node *pred_block;
1657 if ((irn_not_visited(pred))
1658 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1661 * If the current node is NOT in a dead block, but one of its
1662 * predecessors is, we must move the predecessor to a live block.
1663 * Such thing can happen, if global CSE chose a node from a dead block.
1664 * We move it simple to our block.
1665 * Note that neither Phi nor End nodes are floating, so we don't
1666 * need to handle them here.
1668 if (! in_dead_block) {
1669 if (get_irn_pinned(pred) == op_pin_state_floats &&
1670 is_Block_unreachable(get_irn_n(pred, -1)))
1671 set_nodes_block(pred, curr_block);
1673 place_floats_early(pred, worklist);
1677 * A node in the Bad block must stay in the bad block,
1678 * so don't compute a new block for it.
1683 /* Because all loops contain at least one op_pin_state_pinned node, now all
1684 our inputs are either op_pin_state_pinned or place_early() has already
1685 been finished on them. We do not have any unfinished inputs! */
1686 pred_block = get_irn_n(pred, -1);
1687 if ((!is_Block_dead(pred_block)) &&
1688 (get_Block_dom_depth(pred_block) > depth)) {
1690 depth = get_Block_dom_depth(pred_block);
1692 /* Avoid that the node is placed in the Start block */
1693 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1694 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1695 assert(b != get_irg_start_block(current_ir_graph));
1700 set_nodes_block(n, b);
1704 * Add predecessors of non floating nodes and non-floating predecessors
1705 * of floating nodes to worklist and fix their blocks if the are in dead block.
1707 irn_arity = get_irn_arity(n);
1709 if (get_irn_op(n) == op_End) {
1711 * Simplest case: End node. Predecessors are keep-alives,
1712 * no need to move out of dead block.
1714 for (i = -1; i < irn_arity; ++i) {
1715 ir_node *pred = get_irn_n(n, i);
1716 if (irn_not_visited(pred))
1717 pdeq_putr(worklist, pred);
1720 else if (is_Block(n)) {
1722 * Blocks: Predecessors are control flow, no need to move
1723 * them out of dead block.
1725 for (i = irn_arity - 1; i >= 0; --i) {
1726 ir_node *pred = get_irn_n(n, i);
1727 if (irn_not_visited(pred))
1728 pdeq_putr(worklist, pred);
1731 else if (is_Phi(n)) {
1733 ir_node *curr_block = get_irn_n(n, -1);
1734 int in_dead_block = is_Block_unreachable(curr_block);
1737 * Phi nodes: move nodes from dead blocks into the effective use
1738 * of the Phi-input if the Phi is not in a bad block.
1740 pred = get_irn_n(n, -1);
1741 if (irn_not_visited(pred))
1742 pdeq_putr(worklist, pred);
1744 for (i = irn_arity - 1; i >= 0; --i) {
1745 ir_node *pred = get_irn_n(n, i);
1747 if (irn_not_visited(pred)) {
1748 if (! in_dead_block &&
1749 get_irn_pinned(pred) == op_pin_state_floats &&
1750 is_Block_unreachable(get_irn_n(pred, -1))) {
1751 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1753 pdeq_putr(worklist, pred);
1759 ir_node *curr_block = get_irn_n(n, -1);
1760 int in_dead_block = is_Block_unreachable(curr_block);
1763 * All other nodes: move nodes from dead blocks into the same block.
1765 pred = get_irn_n(n, -1);
1766 if (irn_not_visited(pred))
1767 pdeq_putr(worklist, pred);
1769 for (i = irn_arity - 1; i >= 0; --i) {
1770 ir_node *pred = get_irn_n(n, i);
1772 if (irn_not_visited(pred)) {
1773 if (! in_dead_block &&
1774 get_irn_pinned(pred) == op_pin_state_floats &&
1775 is_Block_unreachable(get_irn_n(pred, -1))) {
1776 set_nodes_block(pred, curr_block);
1778 pdeq_putr(worklist, pred);
1785 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1786 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1787 * places all floating nodes reachable from its argument through floating
1788 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1790 static INLINE void place_early(pdeq *worklist) {
1792 inc_irg_visited(current_ir_graph);
1794 /* this inits the worklist */
1795 place_floats_early(get_irg_end(current_ir_graph), worklist);
1797 /* Work the content of the worklist. */
1798 while (!pdeq_empty(worklist)) {
1799 ir_node *n = pdeq_getl(worklist);
1800 if (irn_not_visited(n))
1801 place_floats_early(n, worklist);
1804 set_irg_outs_inconsistent(current_ir_graph);
1805 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1809 * Compute the deepest common ancestor of block and dca.
1811 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1815 /* we do not want to place nodes in dead blocks */
1816 if (is_Block_dead(block))
1819 /* We found a first legal placement. */
1820 if (!dca) return block;
1822 /* Find a placement that is dominates both, dca and block. */
1823 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1824 block = get_Block_idom(block);
1826 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1827 dca = get_Block_idom(dca);
1830 while (block != dca)
1831 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1836 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1837 * I.e., DCA is the block where we might place PRODUCER.
1838 * A data flow edge points from producer to consumer.
1841 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1843 ir_node *block = NULL;
1845 /* Compute the latest block into which we can place a node so that it is
1847 if (get_irn_op(consumer) == op_Phi) {
1848 /* our consumer is a Phi-node, the effective use is in all those
1849 blocks through which the Phi-node reaches producer */
1851 ir_node *phi_block = get_nodes_block(consumer);
1852 irn_arity = get_irn_arity(consumer);
1854 for (i = 0; i < irn_arity; i++) {
1855 if (get_irn_n(consumer, i) == producer) {
1856 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1858 if (! is_Block_unreachable(new_block))
1859 block = calc_dca(block, new_block);
1864 block = get_irn_n(producer, -1);
1867 assert(is_no_Block(consumer));
1868 block = get_nodes_block(consumer);
1871 /* Compute the deepest common ancestor of block and dca. */
1872 return calc_dca(dca, block);
1875 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1877 static INLINE int get_irn_loop_depth(ir_node *n) {
1878 return get_loop_depth(get_irn_loop(n));
1882 * Move n to a block with less loop depth than it's current block. The
1883 * new block must be dominated by early.
1885 * @param n the node that should be moved
1886 * @param early the earliest block we can n move to
1889 move_out_of_loops (ir_node *n, ir_node *early)
1891 ir_node *best, *dca;
1895 /* Find the region deepest in the dominator tree dominating
1896 dca with the least loop nesting depth, but still dominated
1897 by our early placement. */
1898 dca = get_nodes_block(n);
1901 while (dca != early) {
1902 dca = get_Block_idom(dca);
1903 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1904 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1908 if (best != get_nodes_block(n)) {
1910 printf("Moving out of loop: "); DDMN(n);
1911 printf(" Outermost block: "); DDMN(early);
1912 printf(" Best block: "); DDMN(best);
1913 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1915 set_nodes_block(n, best);
1920 * Find the latest legal block for N and place N into the
1921 * `optimal' Block between the latest and earliest legal block.
1922 * The `optimal' block is the dominance-deepest block of those
1923 * with the least loop-nesting-depth. This places N out of as many
1924 * loops as possible and then makes it as control dependent as
1928 place_floats_late(ir_node *n, pdeq *worklist)
1933 assert(irn_not_visited(n)); /* no multiple placement */
1935 mark_irn_visited(n);
1937 /* no need to place block nodes, control nodes are already placed. */
1938 if ((get_irn_op(n) != op_Block) &&
1940 (get_irn_mode(n) != mode_X)) {
1941 /* Remember the early_blk placement of this block to move it
1942 out of loop no further than the early_blk placement. */
1943 early_blk = get_irn_n(n, -1);
1946 * BEWARE: Here we also get code, that is live, but
1947 * was in a dead block. If the node is life, but because
1948 * of CSE in a dead block, we still might need it.
1951 /* Assure that our users are all placed, except the Phi-nodes.
1952 --- Each data flow cycle contains at least one Phi-node. We
1953 have to break the `user has to be placed before the
1954 producer' dependence cycle and the Phi-nodes are the
1955 place to do so, because we need to base our placement on the
1956 final region of our users, which is OK with Phi-nodes, as they
1957 are op_pin_state_pinned, and they never have to be placed after a
1958 producer of one of their inputs in the same block anyway. */
1959 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1960 ir_node *succ = get_irn_out(n, i);
1961 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1962 place_floats_late(succ, worklist);
1965 if (! is_Block_dead(early_blk)) {
1966 /* do only move things that where not dead */
1968 /* We have to determine the final block of this node... except for
1970 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1971 (get_irn_op(n) != op_Const) &&
1972 (get_irn_op(n) != op_SymConst)) {
1973 ir_node *dca = NULL; /* deepest common ancestor in the
1974 dominator tree of all nodes'
1975 blocks depending on us; our final
1976 placement has to dominate DCA. */
1977 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1978 ir_node *succ = get_irn_out(n, i);
1981 if (get_irn_op(succ) == op_End) {
1983 * This consumer is the End node, a keep alive edge.
1984 * This is not a real consumer, so we ignore it
1989 /* ignore if succ is in dead code */
1990 succ_blk = get_irn_n(succ, -1);
1991 if (is_Block_unreachable(succ_blk))
1993 dca = consumer_dom_dca(dca, succ, n);
1996 set_nodes_block(n, dca);
1997 move_out_of_loops(n, early_blk);
2003 /* Add predecessors of all non-floating nodes on list. (Those of floating
2004 nodes are placed already and therefore are marked.) */
2005 for (i = 0; i < get_irn_n_outs(n); i++) {
2006 ir_node *succ = get_irn_out(n, i);
2007 if (irn_not_visited(get_irn_out(n, i))) {
2008 pdeq_putr(worklist, succ);
2013 static INLINE void place_late(pdeq *worklist) {
2015 inc_irg_visited(current_ir_graph);
2017 /* This fills the worklist initially. */
2018 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2020 /* And now empty the worklist again... */
2021 while (!pdeq_empty(worklist)) {
2022 ir_node *n = pdeq_getl(worklist);
2023 if (irn_not_visited(n))
2024 place_floats_late(n, worklist);
2028 void place_code(ir_graph *irg) {
2030 ir_graph *rem = current_ir_graph;
2032 current_ir_graph = irg;
2034 if (!(get_opt_optimize() && get_opt_global_cse())) return;
2036 /* Handle graph state */
2037 assert(get_irg_phase_state(irg) != phase_building);
2040 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2041 free_loop_information(irg);
2042 construct_backedges(irg);
2045 /* Place all floating nodes as early as possible. This guarantees
2046 a legal code placement. */
2047 worklist = new_pdeq();
2048 place_early(worklist);
2050 /* place_early() invalidates the outs, place_late needs them. */
2051 compute_irg_outs(irg);
2053 /* Now move the nodes down in the dominator tree. This reduces the
2054 unnecessary executions of the node. */
2055 place_late(worklist);
2057 set_irg_outs_inconsistent(current_ir_graph);
2058 set_irg_loopinfo_inconsistent(current_ir_graph);
2060 current_ir_graph = rem;
2064 * Called by walker of remove_critical_cf_edges().
2066 * Place an empty block to an edge between a blocks of multiple
2067 * predecessors and a block of multiple successors.
2070 * @param env Environment of walker. The changed field.
2072 static void walk_critical_cf_edges(ir_node *n, void *env) {
2074 ir_node *pre, *block, *jmp;
2076 ir_graph *irg = get_irn_irg(n);
2078 /* Block has multiple predecessors */
2079 arity = get_irn_arity(n);
2081 if (n == get_irg_end_block(irg))
2082 return; /* No use to add a block here. */
2084 for (i = 0; i < arity; ++i) {
2087 pre = get_irn_n(n, i);
2088 cfop = get_irn_op(skip_Proj(pre));
2089 /* Predecessor has multiple successors. Insert new control flow edge but
2090 ignore exception edges. */
2091 if (! is_op_fragile(cfop) && is_op_forking(cfop)) {
2092 /* set predecessor of new block */
2093 block = new_r_Block(irg, 1, &pre);
2094 /* insert new jmp node to new block */
2095 jmp = new_r_Jmp(irg, block);
2096 /* set successor of new block */
2097 set_irn_n(n, i, jmp);
2099 } /* predecessor has multiple successors */
2100 } /* for all predecessors */
2101 } /* n is a multi-entry block */
2104 void remove_critical_cf_edges(ir_graph *irg) {
2107 irg_block_walk_graph(irg, NULL, walk_critical_cf_edges, &changed);
2109 /* control flow changed */
2110 set_irg_outs_inconsistent(irg);
2111 set_irg_extblk_inconsistent(irg);
2112 set_irg_doms_inconsistent(irg);
2113 set_irg_loopinfo_inconsistent(irg);