3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
21 #include "irgraph_t.h"
34 #include "pdeq.h" /* Fuer code placement */
39 #include "irbackedge_t.h"
46 #include "iredges_t.h"
49 /*------------------------------------------------------------------*/
50 /* apply optimizations of iropt to all nodes. */
51 /*------------------------------------------------------------------*/
54 * A wrapper around optimize_inplace_2() to be called from a walker.
56 static void optimize_in_place_wrapper (ir_node *n, void *env) {
57 ir_node *optimized = optimize_in_place_2(n);
58 if (optimized != n) exchange (n, optimized);
62 * Do local optimizations for a node.
64 * @param n the IR-node where to start. Typically the End node
67 * @note current_ir_graph must be set
69 static INLINE void do_local_optimize(ir_node *n) {
70 /* Handle graph state */
71 assert(get_irg_phase_state(current_ir_graph) != phase_building);
73 if (get_opt_global_cse())
74 set_irg_pinned(current_ir_graph, op_pin_state_floats);
75 set_irg_outs_inconsistent(current_ir_graph);
76 set_irg_doms_inconsistent(current_ir_graph);
77 set_irg_loopinfo_inconsistent(current_ir_graph);
79 /* Clean the value_table in irg for the CSE. */
80 del_identities(current_ir_graph->value_table);
81 current_ir_graph->value_table = new_identities();
83 /* walk over the graph */
84 irg_walk(n, firm_clear_link, optimize_in_place_wrapper, NULL);
87 /* Applies local optimizations (see iropt.h) to all nodes reachable from node n */
88 void local_optimize_node(ir_node *n) {
89 ir_graph *rem = current_ir_graph;
90 current_ir_graph = get_irn_irg(n);
94 current_ir_graph = rem;
98 * Block-Walker: uses dominance depth to mark dead blocks.
100 static void kill_dead_blocks(ir_node *block, void *env)
102 if (get_Block_dom_depth(block) < 0) {
104 * Note that the new dominance code correctly handles
105 * the End block, i.e. it is always reachable from Start
107 set_Block_dead(block);
112 local_optimize_graph (ir_graph *irg) {
113 ir_graph *rem = current_ir_graph;
114 current_ir_graph = irg;
116 if (get_irg_dom_state(irg) == dom_consistent)
117 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
119 do_local_optimize(get_irg_end(irg));
121 current_ir_graph = rem;
125 * Data flow optimization walker.
127 static void opt_walker(ir_node *n, void *env) {
131 optimized = optimize_in_place_2(n);
132 set_irn_link(optimized, NULL);
134 if (optimized != n) {
135 const ir_edge_t *edge;
137 foreach_out_edge(n, edge) {
138 ir_node *succ = get_edge_src_irn(edge);
140 if (get_irn_link(succ) != waitq) {
141 pdeq_putr(waitq, succ);
142 set_irn_link(succ, waitq);
145 exchange(n, optimized);
149 void optimize_graph_df(ir_graph *irg) {
150 pdeq *waitq = new_pdeq();
151 int state = edges_activated(irg);
152 ir_graph *rem = current_ir_graph;
154 current_ir_graph = irg;
159 if (get_opt_global_cse())
160 set_irg_pinned(current_ir_graph, op_pin_state_floats);
162 /* Clean the value_table in irg for the CSE. */
163 del_identities(irg->value_table);
164 irg->value_table = new_identities();
166 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
167 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
169 /* invalidate info */
170 set_irg_outs_inconsistent(irg);
171 set_irg_doms_inconsistent(irg);
172 set_irg_loopinfo_inconsistent(irg);
174 /* walk over the graph */
175 irg_walk_graph(irg, NULL, opt_walker, waitq);
177 /* finish the wait queue */
178 while (! pdeq_empty(waitq)) {
179 ir_node *n = pdeq_getl(waitq);
180 opt_walker(n, waitq);
186 edges_deactivate(irg);
188 current_ir_graph = rem;
192 /*------------------------------------------------------------------*/
193 /* Routines for dead node elimination / copying garbage collection */
194 /* of the obstack. */
195 /*------------------------------------------------------------------*/
198 * Remember the new node in the old node by using a field all nodes have.
200 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
203 * Get this new node, before the old node is forgotten.
205 #define get_new_node(oldn) get_irn_link(oldn)
208 * Check if a new node was set.
210 #define has_new_node(n) (get_new_node(n) != NULL)
213 * We use the block_visited flag to mark that we have computed the
214 * number of useful predecessors for this block.
215 * Further we encode the new arity in this flag in the old blocks.
216 * Remembering the arity is useful, as it saves a lot of pointer
217 * accesses. This function is called for all Phi and Block nodes
221 compute_new_arity(ir_node *b) {
222 int i, res, irn_arity;
225 irg_v = get_irg_block_visited(current_ir_graph);
226 block_v = get_Block_block_visited(b);
227 if (block_v >= irg_v) {
228 /* we computed the number of preds for this block and saved it in the
230 return block_v - irg_v;
232 /* compute the number of good predecessors */
233 res = irn_arity = get_irn_arity(b);
234 for (i = 0; i < irn_arity; i++)
235 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
236 /* save it in the flag. */
237 set_Block_block_visited(b, irg_v + res);
243 * Copies the node to the new obstack. The Ins of the new node point to
244 * the predecessors on the old obstack. For block/phi nodes not all
245 * predecessors might be copied. n->link points to the new node.
246 * For Phi and Block nodes the function allocates in-arrays with an arity
247 * only for useful predecessors. The arity is determined by counting
248 * the non-bad predecessors of the block.
250 * @param n The node to be copied
251 * @param env if non-NULL, the node number attribute will be copied to the new node
253 * Note: Also used for loop unrolling.
255 static void copy_node(ir_node *n, void *env) {
258 ir_op *op = get_irn_op(n);
259 int copy_node_nr = env != NULL;
261 /* The end node looses it's flexible in array. This doesn't matter,
262 as dead node elimination builds End by hand, inlineing doesn't use
264 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
267 /* node copied already */
269 } else if (op == op_Block) {
271 new_arity = compute_new_arity(n);
272 n->attr.block.graph_arr = NULL;
274 block = get_nodes_block(n);
276 new_arity = compute_new_arity(block);
278 new_arity = get_irn_arity(n);
281 nn = new_ir_node(get_irn_dbg_info(n),
288 /* Copy the attributes. These might point to additional data. If this
289 was allocated on the old obstack the pointers now are dangling. This
290 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
291 copy_node_attr(n, nn);
292 new_backedge_info(nn);
296 /* for easier debugging, we want to copy the node numbers too */
297 nn->node_nr = n->node_nr;
302 hook_dead_node_elim_subst(current_ir_graph, n, nn);
306 * Copies new predecessors of old node to new node remembered in link.
307 * Spare the Bad predecessors of Phi and Block nodes.
310 copy_preds (ir_node *n, void *env) {
314 nn = get_new_node(n);
316 /* printf("\n old node: "); DDMSG2(n);
317 printf(" new node: "); DDMSG2(nn);
318 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
321 /* Don't copy Bad nodes. */
323 irn_arity = get_irn_arity(n);
324 for (i = 0; i < irn_arity; i++)
325 if (! is_Bad(get_irn_n(n, i))) {
326 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
327 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
330 /* repair the block visited flag from above misuse. Repair it in both
331 graphs so that the old one can still be used. */
332 set_Block_block_visited(nn, 0);
333 set_Block_block_visited(n, 0);
334 /* Local optimization could not merge two subsequent blocks if
335 in array contained Bads. Now it's possible.
336 We don't call optimize_in_place as it requires
337 that the fields in ir_graph are set properly. */
338 if ((get_opt_control_flow_straightening()) &&
339 (get_Block_n_cfgpreds(nn) == 1) &&
340 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
341 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
343 /* Jmp jumps into the block it is in -- deal self cycle. */
344 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
345 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
350 } else if (get_irn_op(n) == op_Phi) {
351 /* Don't copy node if corresponding predecessor in block is Bad.
352 The Block itself should not be Bad. */
353 block = get_nodes_block(n);
354 set_irn_n (nn, -1, get_new_node(block));
356 irn_arity = get_irn_arity(n);
357 for (i = 0; i < irn_arity; i++)
358 if (! is_Bad(get_irn_n(block, i))) {
359 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
360 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
363 /* If the pre walker reached this Phi after the post walker visited the
364 block block_visited is > 0. */
365 set_Block_block_visited(get_nodes_block(n), 0);
366 /* Compacting the Phi's ins might generate Phis with only one
368 if (get_irn_arity(nn) == 1)
369 exchange(nn, get_irn_n(nn, 0));
371 irn_arity = get_irn_arity(n);
372 for (i = -1; i < irn_arity; i++)
373 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
375 /* Now the new node is complete. We can add it to the hash table for CSE.
376 @@@ inlining aborts if we identify End. Why? */
377 if (get_irn_op(nn) != op_End)
378 add_identities (current_ir_graph->value_table, nn);
382 * Copies the graph recursively, compacts the keep-alives of the end node.
384 * @param irg the graph to be copied
385 * @param copy_node_nr If non-zero, the node number will be copied
387 static void copy_graph(ir_graph *irg, int copy_node_nr) {
388 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
389 ir_node *ka; /* keep alive */
393 /* Some nodes must be copied by hand, sigh */
394 vfl = get_irg_visited(irg);
395 set_irg_visited(irg, vfl + 1);
397 oe = get_irg_end(irg);
398 mark_irn_visited(oe);
399 /* copy the end node by hand, allocate dynamic in array! */
400 ne = new_ir_node(get_irn_dbg_info(oe),
407 /* Copy the attributes. Well, there might be some in the future... */
408 copy_node_attr(oe, ne);
409 set_new_node(oe, ne);
411 /* copy the Bad node */
412 ob = get_irg_bad(irg);
413 mark_irn_visited(ob);
414 nb = new_ir_node(get_irn_dbg_info(ob),
421 copy_node_attr(ob, nb);
422 set_new_node(ob, nb);
424 /* copy the NoMem node */
425 om = get_irg_no_mem(irg);
426 mark_irn_visited(om);
427 nm = new_ir_node(get_irn_dbg_info(om),
434 copy_node_attr(om, nm);
435 set_new_node(om, nm);
437 /* copy the live nodes */
438 set_irg_visited(irg, vfl);
439 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
441 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
443 /* visit the anchors as well */
444 for (i = anchor_max - 1; i >= 0; --i) {
445 ir_node *n = irg->anchors[i];
447 if (n && (get_irn_visited(n) <= vfl)) {
448 set_irg_visited(irg, vfl);
449 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
453 /* copy_preds for the end node ... */
454 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
456 /*- ... and now the keep alives. -*/
457 /* First pick the not marked block nodes and walk them. We must pick these
458 first as else we will oversee blocks reachable from Phis. */
459 irn_arity = get_irn_arity(oe);
460 for (i = 0; i < irn_arity; i++) {
461 ka = get_irn_intra_n(oe, i);
463 (get_irn_visited(ka) <= vfl)) {
464 /* We must keep the block alive and copy everything reachable */
465 set_irg_visited(irg, vfl);
466 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
467 add_End_keepalive(ne, get_new_node(ka));
471 /* Now pick other nodes. Here we will keep all! */
472 irn_arity = get_irn_arity(oe);
473 for (i = 0; i < irn_arity; i++) {
474 ka = get_irn_intra_n(oe, i);
476 if (get_irn_visited(ka) <= vfl) {
477 /* We didn't copy the node yet. */
478 set_irg_visited(irg, vfl);
479 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
481 add_End_keepalive(ne, get_new_node(ka));
485 /* start block sometimes only reached after keep alives */
486 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
487 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
491 * Copies the graph reachable from current_ir_graph->end to the obstack
492 * in current_ir_graph and fixes the environment.
493 * Then fixes the fields in current_ir_graph containing nodes of the
496 * @param copy_node_nr If non-zero, the node number will be copied
499 copy_graph_env (int copy_node_nr) {
500 ir_graph *irg = current_ir_graph;
501 ir_node *old_end, *n;
504 /* remove end_except and end_reg nodes */
505 old_end = get_irg_end(irg);
506 set_irg_end_except (irg, old_end);
507 set_irg_end_reg (irg, old_end);
509 /* Not all nodes remembered in irg might be reachable
510 from the end node. Assure their link is set to NULL, so that
511 we can test whether new nodes have been computed. */
512 for (i = anchor_max - 1; i >= 0; --i)
514 set_new_node(irg->anchors[i], NULL);
516 /* we use the block walk flag for removing Bads from Blocks ins. */
517 inc_irg_block_visited(irg);
520 copy_graph(irg, copy_node_nr);
522 /* fix the fields in irg */
523 old_end = get_irg_end(irg);
524 for (i = anchor_max - 1; i >= 0; --i) {
527 irg->anchors[i] = get_new_node(n);
533 * Copies all reachable nodes to a new obstack. Removes bad inputs
534 * from block nodes and the corresponding inputs from Phi nodes.
535 * Merges single exit blocks with single entry blocks and removes
537 * Adds all new nodes to a new hash table for CSE. Does not
538 * perform CSE, so the hash table might contain common subexpressions.
541 dead_node_elimination(ir_graph *irg) {
543 int rem_ipview = get_interprocedural_view();
544 struct obstack *graveyard_obst = NULL;
545 struct obstack *rebirth_obst = NULL;
547 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
548 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
550 /* inform statistics that we started a dead-node elimination run */
551 hook_dead_node_elim(irg, 1);
553 /* Remember external state of current_ir_graph. */
554 rem = current_ir_graph;
555 current_ir_graph = irg;
556 set_interprocedural_view(0);
558 assert(get_irg_phase_state(current_ir_graph) != phase_building);
560 /* Handle graph state */
561 free_callee_info(current_ir_graph);
562 free_irg_outs(current_ir_graph);
565 /* @@@ so far we loose loops when copying */
566 free_loop_information(current_ir_graph);
568 set_irg_doms_inconsistent(irg);
570 /* A quiet place, where the old obstack can rest in peace,
571 until it will be cremated. */
572 graveyard_obst = irg->obst;
574 /* A new obstack, where the reachable nodes will be copied to. */
575 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
576 current_ir_graph->obst = rebirth_obst;
577 obstack_init (current_ir_graph->obst);
578 current_ir_graph->last_node_idx = 0;
580 /* We also need a new value table for CSE */
581 del_identities(irg->value_table);
582 irg->value_table = new_identities();
584 /* Copy the graph from the old to the new obstack */
587 /* Free memory from old unoptimized obstack */
588 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
589 xfree (graveyard_obst); /* ... then free it. */
591 /* inform statistics that the run is over */
592 hook_dead_node_elim(irg, 0);
594 current_ir_graph = rem;
595 set_interprocedural_view(rem_ipview);
600 * Relink bad predecessors of a block and store the old in array to the
601 * link field. This function is called by relink_bad_predecessors().
602 * The array of link field starts with the block operand at position 0.
603 * If block has bad predecessors, create a new in array without bad preds.
604 * Otherwise let in array untouched.
606 static void relink_bad_block_predecessors(ir_node *n, void *env) {
607 ir_node **new_in, *irn;
608 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
610 /* if link field of block is NULL, look for bad predecessors otherwise
611 this is already done */
612 if (get_irn_op(n) == op_Block &&
613 get_irn_link(n) == NULL) {
615 /* save old predecessors in link field (position 0 is the block operand)*/
616 set_irn_link(n, get_irn_in(n));
618 /* count predecessors without bad nodes */
619 old_irn_arity = get_irn_arity(n);
620 for (i = 0; i < old_irn_arity; i++)
621 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
623 /* arity changing: set new predecessors without bad nodes */
624 if (new_irn_arity < old_irn_arity) {
625 /* Get new predecessor array. We do not resize the array, as we must
626 keep the old one to update Phis. */
627 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
629 /* set new predecessors in array */
632 for (i = 0; i < old_irn_arity; i++) {
633 irn = get_irn_n(n, i);
635 new_in[new_irn_n] = irn;
636 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
640 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
641 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
644 } /* ir node has bad predecessors */
646 } /* Block is not relinked */
650 * Relinks Bad predecessors from Blocks and Phis called by walker
651 * remove_bad_predecesors(). If n is a Block, call
652 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
653 * function of Phi's Block. If this block has bad predecessors, relink preds
656 static void relink_bad_predecessors(ir_node *n, void *env) {
657 ir_node *block, **old_in;
658 int i, old_irn_arity, new_irn_arity;
660 /* relink bad predecessors of a block */
661 if (get_irn_op(n) == op_Block)
662 relink_bad_block_predecessors(n, env);
664 /* If Phi node relink its block and its predecessors */
665 if (get_irn_op(n) == op_Phi) {
667 /* Relink predecessors of phi's block */
668 block = get_nodes_block(n);
669 if (get_irn_link(block) == NULL)
670 relink_bad_block_predecessors(block, env);
672 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
673 old_irn_arity = ARR_LEN(old_in);
675 /* Relink Phi predecessors if count of predecessors changed */
676 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
677 /* set new predecessors in array
678 n->in[0] remains the same block */
680 for(i = 1; i < old_irn_arity; i++)
681 if (!is_Bad((ir_node *)old_in[i])) {
682 n->in[new_irn_arity] = n->in[i];
683 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
687 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
688 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
691 } /* n is a Phi node */
695 * Removes Bad Bad predecessors from Blocks and the corresponding
696 * inputs to Phi nodes as in dead_node_elimination but without
698 * On walking up set the link field to NULL, on walking down call
699 * relink_bad_predecessors() (This function stores the old in array
700 * to the link field and sets a new in array if arity of predecessors
703 void remove_bad_predecessors(ir_graph *irg) {
704 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
711 __)|_| | \_/ | \_/(/_ |_/\__|__
713 The following stuff implements a facility that automatically patches
714 registered ir_node pointers to the new node when a dead node elimination occurs.
717 struct _survive_dce_t {
721 hook_entry_t dead_node_elim;
722 hook_entry_t dead_node_elim_subst;
725 typedef struct _survive_dce_list_t {
726 struct _survive_dce_list_t *next;
728 } survive_dce_list_t;
730 static void dead_node_hook(void *context, ir_graph *irg, int start)
732 survive_dce_t *sd = context;
734 /* Create a new map before the dead node elimination is performed. */
736 sd->new_places = pmap_create_ex(pmap_count(sd->places));
739 /* Patch back all nodes if dead node elimination is over and something is to be done. */
741 pmap_destroy(sd->places);
742 sd->places = sd->new_places;
743 sd->new_places = NULL;
748 * Hook called when dead node elimination replaces old by nw.
750 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw)
752 survive_dce_t *sd = context;
753 survive_dce_list_t *list = pmap_get(sd->places, old);
755 /* If the node is to be patched back, write the new address to all registered locations. */
757 survive_dce_list_t *p;
759 for(p = list; p; p = p->next)
762 pmap_insert(sd->new_places, nw, list);
767 * Make a new Survive DCE environment.
769 survive_dce_t *new_survive_dce(void)
771 survive_dce_t *res = xmalloc(sizeof(res[0]));
772 obstack_init(&res->obst);
773 res->places = pmap_create();
774 res->new_places = NULL;
776 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
777 res->dead_node_elim.context = res;
778 res->dead_node_elim.next = NULL;
780 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
781 res->dead_node_elim_subst.context = res;
782 res->dead_node_elim_subst.next = NULL;
784 register_hook(hook_dead_node_elim, &res->dead_node_elim);
785 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
790 * Free a Survive DCE environment.
792 void free_survive_dce(survive_dce_t *sd)
794 obstack_free(&sd->obst, NULL);
795 pmap_destroy(sd->places);
796 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
797 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
802 * Register a node pointer to be patched upon DCE.
803 * When DCE occurs, the node pointer specified by @p place will be
804 * patched to the new address of the node it is pointing to.
806 * @param sd The Survive DCE environment.
807 * @param place The address of the node pointer.
809 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place)
812 ir_node *irn = *place;
813 survive_dce_list_t *curr = pmap_get(sd->places, irn);
814 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw));
819 pmap_insert(sd->places, irn, nw);
823 /*--------------------------------------------------------------------*/
824 /* Functionality for inlining */
825 /*--------------------------------------------------------------------*/
828 * Copy node for inlineing. Updates attributes that change when
829 * inlineing but not for dead node elimination.
831 * Copies the node by calling copy_node() and then updates the entity if
832 * it's a local one. env must be a pointer of the frame type of the
833 * inlined procedure. The new entities must be in the link field of
837 copy_node_inline (ir_node *n, void *env) {
839 ir_type *frame_tp = (ir_type *)env;
842 if (get_irn_op(n) == op_Sel) {
843 nn = get_new_node (n);
845 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
846 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
848 } else if (get_irn_op(n) == op_Block) {
849 nn = get_new_node (n);
850 nn->attr.block.irg = current_ir_graph;
854 static void find_addr(ir_node *node, void *env)
856 if (get_irn_opcode(node) == iro_Proj) {
857 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
863 * currently, we cannot inline two cases:
864 * - call with compound arguments
865 * - graphs that take the address of a parameter
867 * check these conditions here
869 static int can_inline(ir_node *call, ir_graph *called_graph)
871 ir_type *call_type = get_Call_type(call);
872 int params, ress, i, res;
873 assert(is_Method_type(call_type));
875 params = get_method_n_params(call_type);
876 ress = get_method_n_ress(call_type);
879 for (i = 0; i < params; ++i) {
880 ir_type *p_type = get_method_param_type(call_type, i);
882 if (is_compound_type(p_type))
887 for (i = 0; i < ress; ++i) {
888 ir_type *r_type = get_method_res_type(call_type, i);
890 if (is_compound_type(r_type))
895 irg_walk_graph(called_graph, find_addr, NULL, &res);
900 int inline_method(ir_node *call, ir_graph *called_graph) {
902 ir_node *post_call, *post_bl;
903 ir_node *in[pn_Start_max];
904 ir_node *end, *end_bl;
908 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
910 ir_type *called_frame;
911 irg_inline_property prop = get_irg_inline_property(called_graph);
913 if ( (prop != irg_inline_forced) &&
914 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
916 /* Do not inline variadic functions. */
917 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
920 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
921 get_method_n_params(get_Call_type(call)));
924 * currently, we cannot inline two cases:
925 * - call with compound arguments
926 * - graphs that take the address of a parameter
928 if (! can_inline(call, called_graph))
931 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
932 rem_opt = get_opt_optimize();
935 /* Handle graph state */
936 assert(get_irg_phase_state(current_ir_graph) != phase_building);
937 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
938 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
939 set_irg_outs_inconsistent(current_ir_graph);
940 set_irg_extblk_inconsistent(current_ir_graph);
941 set_irg_doms_inconsistent(current_ir_graph);
942 set_irg_loopinfo_inconsistent(current_ir_graph);
943 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
945 /* -- Check preconditions -- */
946 assert(is_Call(call));
947 /* @@@ does not work for InterfaceIII.java after cgana
948 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
949 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
950 get_Call_type(call)));
952 assert(get_type_tpop(get_Call_type(call)) == type_method);
953 if (called_graph == current_ir_graph) {
954 set_optimize(rem_opt);
958 /* here we know we WILL inline, so inform the statistics */
959 hook_inline(call, called_graph);
961 /* -- Decide how to handle exception control flow: Is there a handler
962 for the Call node, or do we branch directly to End on an exception?
964 0 There is a handler.
966 2 Exception handling not represented in Firm. -- */
968 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
969 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
970 assert(get_irn_op(proj) == op_Proj);
971 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
972 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
974 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
975 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
976 else { exc_handling = 2; } /* !Mproj && !Xproj */
981 the procedure and later replaces the Start node of the called graph.
982 Post_call is the old Call node and collects the results of the called
983 graph. Both will end up being a tuple. -- */
984 post_bl = get_nodes_block(call);
985 set_irg_current_block(current_ir_graph, post_bl);
986 /* XxMxPxPxPxT of Start + parameter of Call */
987 in[pn_Start_X_initial_exec] = new_Jmp();
988 in[pn_Start_M] = get_Call_mem(call);
989 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
990 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
991 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
992 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
993 /* in[pn_Start_P_value_arg_base] = ??? */
994 pre_call = new_Tuple(pn_Start_max, in);
998 The new block gets the ins of the old block, pre_call and all its
999 predecessors and all Phi nodes. -- */
1000 part_block(pre_call);
1002 /* -- Prepare state for dead node elimination -- */
1003 /* Visited flags in calling irg must be >= flag in called irg.
1004 Else walker and arity computation will not work. */
1005 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
1006 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
1007 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
1008 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
1009 /* Set pre_call as new Start node in link field of the start node of
1010 calling graph and pre_calls block as new block for the start block
1012 Further mark these nodes so that they are not visited by the
1014 set_irn_link(get_irg_start(called_graph), pre_call);
1015 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
1016 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1017 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
1018 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1019 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
1021 /* Initialize for compaction of in arrays */
1022 inc_irg_block_visited(current_ir_graph);
1024 /* -- Replicate local entities of the called_graph -- */
1025 /* copy the entities. */
1026 called_frame = get_irg_frame_type(called_graph);
1027 for (i = 0; i < get_class_n_members(called_frame); i++) {
1028 entity *new_ent, *old_ent;
1029 old_ent = get_class_member(called_frame, i);
1030 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
1031 set_entity_link(old_ent, new_ent);
1034 /* visited is > than that of called graph. With this trick visited will
1035 remain unchanged so that an outer walker, e.g., searching the call nodes
1036 to inline, calling this inline will not visit the inlined nodes. */
1037 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
1039 /* -- Performing dead node elimination inlines the graph -- */
1040 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1042 /* @@@ endless loops are not copied!! -- they should be, I think... */
1043 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
1044 get_irg_frame_type(called_graph));
1046 /* Repair called_graph */
1047 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
1048 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
1049 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1051 /* -- Merge the end of the inlined procedure with the call site -- */
1052 /* We will turn the old Call node into a Tuple with the following
1055 0: Phi of all Memories of Return statements.
1056 1: Jmp from new Block that merges the control flow from all exception
1057 predecessors of the old end block.
1058 2: Tuple of all arguments.
1059 3: Phi of Exception memories.
1060 In case the old Call directly branches to End on an exception we don't
1061 need the block merging all exceptions nor the Phi of the exception
1065 /* -- Precompute some values -- */
1066 end_bl = get_new_node(get_irg_end_block(called_graph));
1067 end = get_new_node(get_irg_end(called_graph));
1068 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1069 n_res = get_method_n_ress(get_Call_type(call));
1071 res_pred = xmalloc (n_res * sizeof(*res_pred));
1072 cf_pred = xmalloc (arity * sizeof(*res_pred));
1074 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
1076 /* -- archive keepalives -- */
1077 irn_arity = get_irn_arity(end);
1078 for (i = 0; i < irn_arity; i++)
1079 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
1081 /* The new end node will die. We need not free as the in array is on the obstack:
1082 copy_node() only generated 'D' arrays. */
1084 /* -- Replace Return nodes by Jump nodes. -- */
1086 for (i = 0; i < arity; i++) {
1088 ret = get_irn_n(end_bl, i);
1089 if (is_Return(ret)) {
1090 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
1094 set_irn_in(post_bl, n_ret, cf_pred);
1096 /* -- Build a Tuple for all results of the method.
1097 Add Phi node if there was more than one Return. -- */
1098 turn_into_tuple(post_call, 4);
1099 /* First the Memory-Phi */
1101 for (i = 0; i < arity; i++) {
1102 ret = get_irn_n(end_bl, i);
1103 if (is_Return(ret)) {
1104 cf_pred[n_ret] = get_Return_mem(ret);
1108 phi = new_Phi(n_ret, cf_pred, mode_M);
1109 set_Tuple_pred(call, pn_Call_M_regular, phi);
1110 /* Conserve Phi-list for further inlinings -- but might be optimized */
1111 if (get_nodes_block(phi) == post_bl) {
1112 set_irn_link(phi, get_irn_link(post_bl));
1113 set_irn_link(post_bl, phi);
1115 /* Now the real results */
1117 for (j = 0; j < n_res; j++) {
1119 for (i = 0; i < arity; i++) {
1120 ret = get_irn_n(end_bl, i);
1121 if (get_irn_op(ret) == op_Return) {
1122 cf_pred[n_ret] = get_Return_res(ret, j);
1127 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1131 /* Conserve Phi-list for further inlinings -- but might be optimized */
1132 if (get_nodes_block(phi) == post_bl) {
1133 set_irn_link(phi, get_irn_link(post_bl));
1134 set_irn_link(post_bl, phi);
1137 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1139 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1141 /* Finally the exception control flow.
1142 We have two (three) possible situations:
1143 First if the Call branches to an exception handler: We need to add a Phi node to
1144 collect the memory containing the exception objects. Further we need
1145 to add another block to get a correct representation of this Phi. To
1146 this block we add a Jmp that resolves into the X output of the Call
1147 when the Call is turned into a tuple.
1148 Second the Call branches to End, the exception is not handled. Just
1149 add all inlined exception branches to the End node.
1150 Third: there is no Exception edge at all. Handle as case two. */
1151 if (exc_handling == 0) {
1153 for (i = 0; i < arity; i++) {
1155 ret = get_irn_n(end_bl, i);
1156 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1157 cf_pred[n_exc] = ret;
1162 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1163 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1164 /* The Phi for the memories with the exception objects */
1166 for (i = 0; i < arity; i++) {
1168 ret = skip_Proj(get_irn_n(end_bl, i));
1170 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1172 } else if (is_fragile_op(ret)) {
1173 /* We rely that all cfops have the memory output at the same position. */
1174 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1176 } else if (get_irn_op(ret) == op_Raise) {
1177 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1181 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1183 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1184 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1187 ir_node *main_end_bl;
1188 int main_end_bl_arity;
1189 ir_node **end_preds;
1191 /* assert(exc_handling == 1 || no exceptions. ) */
1193 for (i = 0; i < arity; i++) {
1194 ir_node *ret = get_irn_n(end_bl, i);
1196 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1197 cf_pred[n_exc] = ret;
1201 main_end_bl = get_irg_end_block(current_ir_graph);
1202 main_end_bl_arity = get_irn_arity(main_end_bl);
1203 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1205 for (i = 0; i < main_end_bl_arity; ++i)
1206 end_preds[i] = get_irn_n(main_end_bl, i);
1207 for (i = 0; i < n_exc; ++i)
1208 end_preds[main_end_bl_arity + i] = cf_pred[i];
1209 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1210 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1211 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1217 #if 0 /* old. now better, correcter, faster implementation. */
1219 /* -- If the exception control flow from the inlined Call directly
1220 branched to the end block we now have the following control
1221 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1222 remove the Jmp along with it's empty block and add Jmp's
1223 predecessors as predecessors of this end block. No problem if
1224 there is no exception, because then branches Bad to End which
1226 @@@ can't we know this beforehand: by getting the Proj(1) from
1227 the Call link list and checking whether it goes to Proj. */
1228 /* find the problematic predecessor of the end block. */
1229 end_bl = get_irg_end_block(current_ir_graph);
1230 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1231 cf_op = get_Block_cfgpred(end_bl, i);
1232 if (get_irn_op(cf_op) == op_Proj) {
1233 cf_op = get_Proj_pred(cf_op);
1234 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1235 /* There are unoptimized tuples from inlineing before when no exc */
1236 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1237 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1238 assert(get_irn_op(cf_op) == op_Jmp);
1244 if (i < get_Block_n_cfgpreds(end_bl)) {
1245 bl = get_nodes_block(cf_op);
1246 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1247 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1248 for (j = 0; j < i; j++)
1249 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1250 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1251 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1252 for (j = j; j < arity; j++)
1253 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1254 set_irn_in(end_bl, arity, cf_pred);
1256 /* Remove the exception pred from post-call Tuple. */
1257 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1262 /* -- Turn CSE back on. -- */
1263 set_optimize(rem_opt);
1268 /********************************************************************/
1269 /* Apply inlineing to small methods. */
1270 /********************************************************************/
1272 /* It makes no sense to inline too many calls in one procedure. Anyways,
1273 I didn't get a version with NEW_ARR_F to run. */
1274 #define MAX_INLINE 1024
1277 * environment for inlining small irgs
1279 typedef struct _inline_env_t {
1281 ir_node *calls[MAX_INLINE];
1285 * Returns the irg called from a Call node. If the irg is not
1286 * known, NULL is returned.
1288 static ir_graph *get_call_called_irg(ir_node *call) {
1290 ir_graph *called_irg = NULL;
1292 assert(is_Call(call));
1294 addr = get_Call_ptr(call);
1295 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1296 called_irg = get_entity_irg(get_SymConst_entity(addr));
1302 static void collect_calls(ir_node *call, void *env) {
1305 if (! is_Call(call)) return;
1307 addr = get_Call_ptr(call);
1309 if (get_irn_op(addr) == op_SymConst) {
1310 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1311 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1312 inline_env_t *ienv = (inline_env_t *)env;
1313 if (called_irg && ienv->pos < MAX_INLINE) {
1314 /* The Call node calls a locally defined method. Remember to inline. */
1315 ienv->calls[ienv->pos++] = call;
1322 * Inlines all small methods at call sites where the called address comes
1323 * from a Const node that references the entity representing the called
1325 * The size argument is a rough measure for the code size of the method:
1326 * Methods where the obstack containing the firm graph is smaller than
1329 void inline_small_irgs(ir_graph *irg, int size) {
1331 ir_graph *rem = current_ir_graph;
1332 inline_env_t env /* = {0, NULL}*/;
1334 if (!(get_opt_optimize() && get_opt_inline())) return;
1336 current_ir_graph = irg;
1337 /* Handle graph state */
1338 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1339 free_callee_info(current_ir_graph);
1341 /* Find Call nodes to inline.
1342 (We can not inline during a walk of the graph, as inlineing the same
1343 method several times changes the visited flag of the walked graph:
1344 after the first inlineing visited of the callee equals visited of
1345 the caller. With the next inlineing both are increased.) */
1347 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1349 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1350 /* There are calls to inline */
1351 collect_phiprojs(irg);
1352 for (i = 0; i < env.pos; i++) {
1354 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1355 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1356 (get_irg_inline_property(callee) == irg_inline_forced)) {
1357 inline_method(env.calls[i], callee);
1362 current_ir_graph = rem;
1366 * Environment for inlining irgs.
1369 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1370 int n_nodes_orig; /**< for statistics */
1371 eset *call_nodes; /**< All call nodes in this graph */
1373 int n_call_nodes_orig; /**< for statistics */
1374 int n_callers; /**< Number of known graphs that call this graphs. */
1375 int n_callers_orig; /**< for statistics */
1379 * Allocate a new environment for inlining.
1381 static inline_irg_env *new_inline_irg_env(void) {
1382 inline_irg_env *env = xmalloc(sizeof(*env));
1383 env->n_nodes = -2; /* do not count count Start, End */
1384 env->n_nodes_orig = -2; /* do not count Start, End */
1385 env->call_nodes = eset_create();
1386 env->n_call_nodes = 0;
1387 env->n_call_nodes_orig = 0;
1389 env->n_callers_orig = 0;
1394 * destroy an environment for inlining.
1396 static void free_inline_irg_env(inline_irg_env *env) {
1397 eset_destroy(env->call_nodes);
1402 * post-walker: collect all calls in the inline-environment
1403 * of a graph and sum some statistics.
1405 static void collect_calls2(ir_node *call, void *env) {
1406 inline_irg_env *x = (inline_irg_env *)env;
1407 ir_op *op = get_irn_op(call);
1410 /* count meaningful nodes in irg */
1411 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1416 if (op != op_Call) return;
1418 /* collect all call nodes */
1419 eset_insert(x->call_nodes, call);
1421 x->n_call_nodes_orig++;
1423 /* count all static callers */
1424 callee = get_call_called_irg(call);
1426 inline_irg_env *callee_env = get_irg_link(callee);
1427 callee_env->n_callers++;
1428 callee_env->n_callers_orig++;
1433 * Returns TRUE if the number of callers in 0 in the irg's environment,
1434 * hence this irg is a leave.
1436 INLINE static int is_leave(ir_graph *irg) {
1437 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1441 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1443 INLINE static int is_smaller(ir_graph *callee, int size) {
1444 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1449 * Inlines small leave methods at call sites where the called address comes
1450 * from a Const node that references the entity representing the called
1452 * The size argument is a rough measure for the code size of the method:
1453 * Methods where the obstack containing the firm graph is smaller than
1456 void inline_leave_functions(int maxsize, int leavesize, int size) {
1457 inline_irg_env *env;
1458 int i, n_irgs = get_irp_n_irgs();
1459 ir_graph *rem = current_ir_graph;
1462 if (!(get_opt_optimize() && get_opt_inline())) return;
1464 /* extend all irgs by a temporary data structure for inlining. */
1465 for (i = 0; i < n_irgs; ++i)
1466 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1468 /* Precompute information in temporary data structure. */
1469 for (i = 0; i < n_irgs; ++i) {
1470 current_ir_graph = get_irp_irg(i);
1471 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1472 free_callee_info(current_ir_graph);
1474 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1475 get_irg_link(current_ir_graph));
1478 /* -- and now inline. -- */
1480 /* Inline leaves recursively -- we might construct new leaves. */
1481 while (did_inline) {
1484 for (i = 0; i < n_irgs; ++i) {
1486 int phiproj_computed = 0;
1488 current_ir_graph = get_irp_irg(i);
1489 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1491 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1494 if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
1495 callee = get_call_called_irg(call);
1497 if (env->n_nodes > maxsize) continue; // break;
1499 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1500 if (!phiproj_computed) {
1501 phiproj_computed = 1;
1502 collect_phiprojs(current_ir_graph);
1504 did_inline = inline_method(call, callee);
1507 /* Do some statistics */
1508 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1509 env->n_call_nodes --;
1510 env->n_nodes += callee_env->n_nodes;
1511 callee_env->n_callers--;
1518 /* inline other small functions. */
1519 for (i = 0; i < n_irgs; ++i) {
1522 int phiproj_computed = 0;
1524 current_ir_graph = get_irp_irg(i);
1525 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1527 /* we can not walk and change a set, nor remove from it.
1529 walkset = env->call_nodes;
1530 env->call_nodes = eset_create();
1531 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1534 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1535 callee = get_call_called_irg(call);
1538 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1539 (get_irg_inline_property(callee) == irg_inline_forced))) {
1540 if (!phiproj_computed) {
1541 phiproj_computed = 1;
1542 collect_phiprojs(current_ir_graph);
1544 if (inline_method(call, callee)) {
1545 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1546 env->n_call_nodes--;
1547 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1548 env->n_call_nodes += callee_env->n_call_nodes;
1549 env->n_nodes += callee_env->n_nodes;
1550 callee_env->n_callers--;
1553 eset_insert(env->call_nodes, call);
1556 eset_destroy(walkset);
1559 for (i = 0; i < n_irgs; ++i) {
1560 current_ir_graph = get_irp_irg(i);
1562 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1563 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1564 (env->n_callers_orig != env->n_callers))
1565 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1566 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1567 env->n_callers_orig, env->n_callers,
1568 get_entity_name(get_irg_entity(current_ir_graph)));
1570 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1573 current_ir_graph = rem;
1576 /*******************************************************************/
1577 /* Code Placement. Pins all floating nodes to a block where they */
1578 /* will be executed only if needed. */
1579 /*******************************************************************/
1582 * Returns non-zero, is a block is not reachable from Start.
1584 * @param block the block to test
1587 is_Block_unreachable(ir_node *block) {
1588 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1592 * Find the earliest correct block for N. --- Place N into the
1593 * same Block as its dominance-deepest Input.
1595 * We have to avoid calls to get_nodes_block() here
1596 * because the graph is floating.
1598 * move_out_of_loops() expects that place_floats_early() have placed
1599 * all "living" nodes into a living block. That's why we must
1600 * move nodes in dead block with "live" successors into a valid
1602 * We move them just into the same block as it's successor (or
1603 * in case of a Phi into the effective use block). For Phi successors,
1604 * this may still be a dead block, but then there is no real use, as
1605 * the control flow will be dead later.
1608 place_floats_early(ir_node *n, pdeq *worklist)
1612 /* we must not run into an infinite loop */
1613 assert(irn_not_visited(n));
1614 mark_irn_visited(n);
1616 /* Place floating nodes. */
1617 if (get_irn_pinned(n) == op_pin_state_floats) {
1618 ir_node *curr_block = get_irn_n(n, -1);
1619 int in_dead_block = is_Block_unreachable(curr_block);
1621 ir_node *b = NULL; /* The block to place this node in */
1623 assert(get_irn_op(n) != op_Block);
1625 if ((get_irn_op(n) == op_Const) ||
1626 (get_irn_op(n) == op_SymConst) ||
1628 (get_irn_op(n) == op_Unknown)) {
1629 /* These nodes will not be placed by the loop below. */
1630 b = get_irg_start_block(current_ir_graph);
1634 /* find the block for this node. */
1635 irn_arity = get_irn_arity(n);
1636 for (i = 0; i < irn_arity; i++) {
1637 ir_node *pred = get_irn_n(n, i);
1638 ir_node *pred_block;
1640 if ((irn_not_visited(pred))
1641 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1644 * If the current node is NOT in a dead block, but one of its
1645 * predecessors is, we must move the predecessor to a live block.
1646 * Such thing can happen, if global CSE chose a node from a dead block.
1647 * We move it simple to our block.
1648 * Note that neither Phi nor End nodes are floating, so we don't
1649 * need to handle them here.
1651 if (! in_dead_block) {
1652 if (get_irn_pinned(pred) == op_pin_state_floats &&
1653 is_Block_unreachable(get_irn_n(pred, -1)))
1654 set_nodes_block(pred, curr_block);
1656 place_floats_early(pred, worklist);
1660 * A node in the Bad block must stay in the bad block,
1661 * so don't compute a new block for it.
1666 /* Because all loops contain at least one op_pin_state_pinned node, now all
1667 our inputs are either op_pin_state_pinned or place_early() has already
1668 been finished on them. We do not have any unfinished inputs! */
1669 pred_block = get_irn_n(pred, -1);
1670 if ((!is_Block_dead(pred_block)) &&
1671 (get_Block_dom_depth(pred_block) > depth)) {
1673 depth = get_Block_dom_depth(pred_block);
1675 /* Avoid that the node is placed in the Start block */
1676 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1677 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1678 assert(b != get_irg_start_block(current_ir_graph));
1683 set_nodes_block(n, b);
1687 * Add predecessors of non floating nodes and non-floating predecessors
1688 * of floating nodes to worklist and fix their blocks if the are in dead block.
1690 irn_arity = get_irn_arity(n);
1692 if (get_irn_op(n) == op_End) {
1694 * Simplest case: End node. Predecessors are keep-alives,
1695 * no need to move out of dead block.
1697 for (i = -1; i < irn_arity; ++i) {
1698 ir_node *pred = get_irn_n(n, i);
1699 if (irn_not_visited(pred))
1700 pdeq_putr(worklist, pred);
1703 else if (is_Block(n)) {
1705 * Blocks: Predecessors are control flow, no need to move
1706 * them out of dead block.
1708 for (i = irn_arity - 1; i >= 0; --i) {
1709 ir_node *pred = get_irn_n(n, i);
1710 if (irn_not_visited(pred))
1711 pdeq_putr(worklist, pred);
1714 else if (is_Phi(n)) {
1716 ir_node *curr_block = get_irn_n(n, -1);
1717 int in_dead_block = is_Block_unreachable(curr_block);
1720 * Phi nodes: move nodes from dead blocks into the effective use
1721 * of the Phi-input if the Phi is not in a bad block.
1723 pred = get_irn_n(n, -1);
1724 if (irn_not_visited(pred))
1725 pdeq_putr(worklist, pred);
1727 for (i = irn_arity - 1; i >= 0; --i) {
1728 ir_node *pred = get_irn_n(n, i);
1730 if (irn_not_visited(pred)) {
1731 if (! in_dead_block &&
1732 get_irn_pinned(pred) == op_pin_state_floats &&
1733 is_Block_unreachable(get_irn_n(pred, -1))) {
1734 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1736 pdeq_putr(worklist, pred);
1742 ir_node *curr_block = get_irn_n(n, -1);
1743 int in_dead_block = is_Block_unreachable(curr_block);
1746 * All other nodes: move nodes from dead blocks into the same block.
1748 pred = get_irn_n(n, -1);
1749 if (irn_not_visited(pred))
1750 pdeq_putr(worklist, pred);
1752 for (i = irn_arity - 1; i >= 0; --i) {
1753 ir_node *pred = get_irn_n(n, i);
1755 if (irn_not_visited(pred)) {
1756 if (! in_dead_block &&
1757 get_irn_pinned(pred) == op_pin_state_floats &&
1758 is_Block_unreachable(get_irn_n(pred, -1))) {
1759 set_nodes_block(pred, curr_block);
1761 pdeq_putr(worklist, pred);
1768 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1769 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1770 * places all floating nodes reachable from its argument through floating
1771 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1773 static INLINE void place_early(pdeq *worklist) {
1775 inc_irg_visited(current_ir_graph);
1777 /* this inits the worklist */
1778 place_floats_early(get_irg_end(current_ir_graph), worklist);
1780 /* Work the content of the worklist. */
1781 while (!pdeq_empty(worklist)) {
1782 ir_node *n = pdeq_getl(worklist);
1783 if (irn_not_visited(n))
1784 place_floats_early(n, worklist);
1787 set_irg_outs_inconsistent(current_ir_graph);
1788 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1792 * Compute the deepest common ancestor of block and dca.
1794 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1798 /* we do not want to place nodes in dead blocks */
1799 if (is_Block_dead(block))
1802 /* We found a first legal placement. */
1803 if (!dca) return block;
1805 /* Find a placement that is dominates both, dca and block. */
1806 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1807 block = get_Block_idom(block);
1809 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1810 dca = get_Block_idom(dca);
1813 while (block != dca)
1814 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1819 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1820 * I.e., DCA is the block where we might place PRODUCER.
1821 * A data flow edge points from producer to consumer.
1824 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1826 ir_node *block = NULL;
1828 /* Compute the latest block into which we can place a node so that it is
1830 if (get_irn_op(consumer) == op_Phi) {
1831 /* our consumer is a Phi-node, the effective use is in all those
1832 blocks through which the Phi-node reaches producer */
1834 ir_node *phi_block = get_nodes_block(consumer);
1835 irn_arity = get_irn_arity(consumer);
1837 for (i = 0; i < irn_arity; i++) {
1838 if (get_irn_n(consumer, i) == producer) {
1839 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1841 if (! is_Block_unreachable(new_block))
1842 block = calc_dca(block, new_block);
1847 block = get_irn_n(producer, -1);
1850 assert(is_no_Block(consumer));
1851 block = get_nodes_block(consumer);
1854 /* Compute the deepest common ancestor of block and dca. */
1855 return calc_dca(dca, block);
1858 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1860 static INLINE int get_irn_loop_depth(ir_node *n) {
1861 return get_loop_depth(get_irn_loop(n));
1865 * Move n to a block with less loop depth than it's current block. The
1866 * new block must be dominated by early.
1868 * @param n the node that should be moved
1869 * @param early the earliest block we can n move to
1872 move_out_of_loops (ir_node *n, ir_node *early)
1874 ir_node *best, *dca;
1878 /* Find the region deepest in the dominator tree dominating
1879 dca with the least loop nesting depth, but still dominated
1880 by our early placement. */
1881 dca = get_nodes_block(n);
1884 while (dca != early) {
1885 dca = get_Block_idom(dca);
1886 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1887 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1891 if (best != get_nodes_block(n)) {
1893 printf("Moving out of loop: "); DDMN(n);
1894 printf(" Outermost block: "); DDMN(early);
1895 printf(" Best block: "); DDMN(best);
1896 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1898 set_nodes_block(n, best);
1903 * Find the latest legal block for N and place N into the
1904 * `optimal' Block between the latest and earliest legal block.
1905 * The `optimal' block is the dominance-deepest block of those
1906 * with the least loop-nesting-depth. This places N out of as many
1907 * loops as possible and then makes it as control dependent as
1911 place_floats_late(ir_node *n, pdeq *worklist)
1916 assert(irn_not_visited(n)); /* no multiple placement */
1918 mark_irn_visited(n);
1920 /* no need to place block nodes, control nodes are already placed. */
1921 if ((get_irn_op(n) != op_Block) &&
1923 (get_irn_mode(n) != mode_X)) {
1924 /* Remember the early_blk placement of this block to move it
1925 out of loop no further than the early_blk placement. */
1926 early_blk = get_irn_n(n, -1);
1929 * BEWARE: Here we also get code, that is live, but
1930 * was in a dead block. If the node is life, but because
1931 * of CSE in a dead block, we still might need it.
1934 /* Assure that our users are all placed, except the Phi-nodes.
1935 --- Each data flow cycle contains at least one Phi-node. We
1936 have to break the `user has to be placed before the
1937 producer' dependence cycle and the Phi-nodes are the
1938 place to do so, because we need to base our placement on the
1939 final region of our users, which is OK with Phi-nodes, as they
1940 are op_pin_state_pinned, and they never have to be placed after a
1941 producer of one of their inputs in the same block anyway. */
1942 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1943 ir_node *succ = get_irn_out(n, i);
1944 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1945 place_floats_late(succ, worklist);
1948 if (! is_Block_dead(early_blk)) {
1949 /* do only move things that where not dead */
1951 /* We have to determine the final block of this node... except for
1953 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1954 (get_irn_op(n) != op_Const) &&
1955 (get_irn_op(n) != op_SymConst)) {
1956 ir_node *dca = NULL; /* deepest common ancestor in the
1957 dominator tree of all nodes'
1958 blocks depending on us; our final
1959 placement has to dominate DCA. */
1960 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1961 ir_node *succ = get_irn_out(n, i);
1964 if (get_irn_op(succ) == op_End) {
1966 * This consumer is the End node, a keep alive edge.
1967 * This is not a real consumer, so we ignore it
1972 /* ignore if succ is in dead code */
1973 succ_blk = get_irn_n(succ, -1);
1974 if (is_Block_unreachable(succ_blk))
1976 dca = consumer_dom_dca(dca, succ, n);
1979 set_nodes_block(n, dca);
1980 move_out_of_loops(n, early_blk);
1986 /* Add predecessors of all non-floating nodes on list. (Those of floating
1987 nodes are placed already and therefore are marked.) */
1988 for (i = 0; i < get_irn_n_outs(n); i++) {
1989 ir_node *succ = get_irn_out(n, i);
1990 if (irn_not_visited(get_irn_out(n, i))) {
1991 pdeq_putr(worklist, succ);
1996 static INLINE void place_late(pdeq *worklist) {
1998 inc_irg_visited(current_ir_graph);
2000 /* This fills the worklist initially. */
2001 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
2003 /* And now empty the worklist again... */
2004 while (!pdeq_empty(worklist)) {
2005 ir_node *n = pdeq_getl(worklist);
2006 if (irn_not_visited(n))
2007 place_floats_late(n, worklist);
2011 void place_code(ir_graph *irg) {
2013 ir_graph *rem = current_ir_graph;
2015 current_ir_graph = irg;
2017 if (!(get_opt_optimize() && get_opt_global_cse())) return;
2019 /* Handle graph state */
2020 assert(get_irg_phase_state(irg) != phase_building);
2023 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
2024 free_loop_information(irg);
2025 construct_backedges(irg);
2028 /* Place all floating nodes as early as possible. This guarantees
2029 a legal code placement. */
2030 worklist = new_pdeq();
2031 place_early(worklist);
2033 /* place_early() invalidates the outs, place_late needs them. */
2034 compute_irg_outs(irg);
2036 /* Now move the nodes down in the dominator tree. This reduces the
2037 unnecessary executions of the node. */
2038 place_late(worklist);
2040 set_irg_outs_inconsistent(current_ir_graph);
2041 set_irg_loopinfo_inconsistent(current_ir_graph);
2043 current_ir_graph = rem;
2047 * Called by walker of remove_critical_cf_edges().
2049 * Place an empty block to an edge between a blocks of multiple
2050 * predecessors and a block of multiple successors.
2053 * @param env Environment of walker. The changed field.
2055 static void walk_critical_cf_edges(ir_node *n, void *env) {
2057 ir_node *pre, *block, *jmp;
2060 /* Block has multiple predecessors */
2061 if (is_Block(n) && (get_irn_arity(n) > 1)) {
2062 if (n == get_irg_end_block(current_ir_graph))
2063 return; /* No use to add a block here. */
2065 arity = get_irn_arity(n);
2066 for (i=0; i<arity; i++) {
2067 pre = get_irn_n(n, i);
2068 /* Predecessor has multiple successors. Insert new control flow edge. */
2069 if (op_Raise != get_irn_op(skip_Proj(pre))) {
2070 /* set predecessor of new block */
2071 block = new_Block(1, &pre);
2072 /* insert new jmp node to new block */
2073 set_cur_block(block);
2076 /* set successor of new block */
2077 set_irn_n(n, i, jmp);
2079 } /* predecessor has multiple successors */
2080 } /* for all predecessors */
2081 } /* n is a block */
2084 void remove_critical_cf_edges(ir_graph *irg) {
2086 irg_walk_graph(irg, NULL, walk_critical_cf_edges, &changed);
2089 /* control flow changed */
2090 set_irg_outs_inconsistent(irg);
2091 set_irg_extblk_inconsistent(irg);
2092 set_irg_doms_inconsistent(current_ir_graph);
2093 set_irg_loopinfo_inconsistent(current_ir_graph);