2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Dead node elimination and Procedure Inlining.
23 * @author Michael Beck, Goetz Lindenmaier
34 #include "irgraph_t.h"
37 #include "iroptimize.h"
44 #include "adt/array.h"
48 #include "adt/xmalloc.h"
52 #include "irbackedge_t.h"
57 #include "analyze_irg_args.h"
58 #include "iredges_t.h"
64 /*------------------------------------------------------------------*/
65 /* Routines for dead node elimination / copying garbage collection */
67 /*------------------------------------------------------------------*/
70 * Remember the new node in the old node by using a field all nodes have.
72 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
75 * Get this new node, before the old node is forgotten.
77 #define get_new_node(oldn) get_irn_link(oldn)
80 * Check if a new node was set.
82 #define has_new_node(n) (get_new_node(n) != NULL)
85 * We use the block_visited flag to mark that we have computed the
86 * number of useful predecessors for this block.
87 * Further we encode the new arity in this flag in the old blocks.
88 * Remembering the arity is useful, as it saves a lot of pointer
89 * accesses. This function is called for all Phi and Block nodes
93 compute_new_arity(ir_node *b) {
94 int i, res, irn_arity;
97 irg_v = get_irg_block_visited(current_ir_graph);
98 block_v = get_Block_block_visited(b);
99 if (block_v >= irg_v) {
100 /* we computed the number of preds for this block and saved it in the
102 return block_v - irg_v;
104 /* compute the number of good predecessors */
105 res = irn_arity = get_irn_arity(b);
106 for (i = 0; i < irn_arity; i++)
107 if (is_Bad(get_irn_n(b, i))) res--;
108 /* save it in the flag. */
109 set_Block_block_visited(b, irg_v + res);
115 * Copies the node to the new obstack. The Ins of the new node point to
116 * the predecessors on the old obstack. For block/phi nodes not all
117 * predecessors might be copied. n->link points to the new node.
118 * For Phi and Block nodes the function allocates in-arrays with an arity
119 * only for useful predecessors. The arity is determined by counting
120 * the non-bad predecessors of the block.
122 * @param n The node to be copied
123 * @param env if non-NULL, the node number attribute will be copied to the new node
125 * Note: Also used for loop unrolling.
127 static void copy_node(ir_node *n, void *env) {
130 ir_op *op = get_irn_op(n);
134 /* node copied already */
136 } else if (op == op_Block) {
138 new_arity = compute_new_arity(n);
139 n->attr.block.graph_arr = NULL;
141 block = get_nodes_block(n);
143 new_arity = compute_new_arity(block);
145 new_arity = get_irn_arity(n);
148 nn = new_ir_node(get_irn_dbg_info(n),
155 /* Copy the attributes. These might point to additional data. If this
156 was allocated on the old obstack the pointers now are dangling. This
157 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
158 if (op == op_Block) {
159 /* we cannot allow blocks WITHOUT macroblock input */
160 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
162 copy_node_attr(n, nn);
166 int copy_node_nr = env != NULL;
168 /* for easier debugging, we want to copy the node numbers too */
169 nn->node_nr = n->node_nr;
175 hook_dead_node_elim_subst(current_ir_graph, n, nn);
179 * Copies new predecessors of old node to new node remembered in link.
180 * Spare the Bad predecessors of Phi and Block nodes.
182 static void copy_preds(ir_node *n, void *env) {
187 nn = get_new_node(n);
190 /* copy the macro block header */
191 ir_node *mbh = get_Block_MacroBlock(n);
194 /* this block is a macroblock header */
195 set_Block_MacroBlock(nn, nn);
197 /* get the macro block header */
198 ir_node *nmbh = get_new_node(mbh);
199 assert(nmbh != NULL);
200 set_Block_MacroBlock(nn, nmbh);
203 /* Don't copy Bad nodes. */
205 irn_arity = get_irn_arity(n);
206 for (i = 0; i < irn_arity; i++) {
207 if (! is_Bad(get_irn_n(n, i))) {
208 ir_node *pred = get_irn_n(n, i);
209 set_irn_n(nn, j, get_new_node(pred));
213 /* repair the block visited flag from above misuse. Repair it in both
214 graphs so that the old one can still be used. */
215 set_Block_block_visited(nn, 0);
216 set_Block_block_visited(n, 0);
217 /* Local optimization could not merge two subsequent blocks if
218 in array contained Bads. Now it's possible.
219 We don't call optimize_in_place as it requires
220 that the fields in ir_graph are set properly. */
221 if ((get_opt_control_flow_straightening()) &&
222 (get_Block_n_cfgpreds(nn) == 1) &&
223 is_Jmp(get_Block_cfgpred(nn, 0))) {
224 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
226 /* Jmp jumps into the block it is in -- deal self cycle. */
227 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
228 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
233 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
234 /* Don't copy node if corresponding predecessor in block is Bad.
235 The Block itself should not be Bad. */
236 block = get_nodes_block(n);
237 set_nodes_block(nn, get_new_node(block));
239 irn_arity = get_irn_arity(n);
240 for (i = 0; i < irn_arity; i++) {
241 if (! is_Bad(get_irn_n(block, i))) {
242 ir_node *pred = get_irn_n(n, i);
243 set_irn_n(nn, j, get_new_node(pred));
244 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
248 /* If the pre walker reached this Phi after the post walker visited the
249 block block_visited is > 0. */
250 set_Block_block_visited(get_nodes_block(n), 0);
251 /* Compacting the Phi's ins might generate Phis with only one
253 if (get_irn_arity(nn) == 1)
254 exchange(nn, get_irn_n(nn, 0));
256 irn_arity = get_irn_arity(n);
257 for (i = -1; i < irn_arity; i++)
258 set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
260 /* Now the new node is complete. We can add it to the hash table for CSE.
261 @@@ inlining aborts if we identify End. Why? */
263 add_identities(current_ir_graph->value_table, nn);
267 * Copies the graph recursively, compacts the keep-alives of the end node.
269 * @param irg the graph to be copied
270 * @param copy_node_nr If non-zero, the node number will be copied
272 static void copy_graph(ir_graph *irg, int copy_node_nr) {
273 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
274 ir_node *ka; /* keep alive */
278 /* Some nodes must be copied by hand, sigh */
279 vfl = get_irg_visited(irg);
280 set_irg_visited(irg, vfl + 1);
282 oe = get_irg_end(irg);
283 mark_irn_visited(oe);
284 /* copy the end node by hand, allocate dynamic in array! */
285 ne = new_ir_node(get_irn_dbg_info(oe),
292 /* Copy the attributes. Well, there might be some in the future... */
293 copy_node_attr(oe, ne);
294 set_new_node(oe, ne);
296 /* copy the Bad node */
297 ob = get_irg_bad(irg);
298 mark_irn_visited(ob);
299 nb = new_ir_node(get_irn_dbg_info(ob),
306 copy_node_attr(ob, nb);
307 set_new_node(ob, nb);
309 /* copy the NoMem node */
310 om = get_irg_no_mem(irg);
311 mark_irn_visited(om);
312 nm = new_ir_node(get_irn_dbg_info(om),
319 copy_node_attr(om, nm);
320 set_new_node(om, nm);
322 /* copy the live nodes */
323 set_irg_visited(irg, vfl);
324 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
326 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
328 /* visit the anchors as well */
329 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
330 ir_node *n = get_irg_anchor(irg, i);
332 if (n && (get_irn_visited(n) <= vfl)) {
333 set_irg_visited(irg, vfl);
334 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
338 /* copy_preds for the end node ... */
339 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
341 /*- ... and now the keep alives. -*/
342 /* First pick the not marked block nodes and walk them. We must pick these
343 first as else we will oversee blocks reachable from Phis. */
344 irn_arity = get_End_n_keepalives(oe);
345 for (i = 0; i < irn_arity; i++) {
346 ka = get_End_keepalive(oe, i);
348 if (get_irn_visited(ka) <= vfl) {
349 /* We must keep the block alive and copy everything reachable */
350 set_irg_visited(irg, vfl);
351 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
353 add_End_keepalive(ne, get_new_node(ka));
357 /* Now pick other nodes. Here we will keep all! */
358 irn_arity = get_End_n_keepalives(oe);
359 for (i = 0; i < irn_arity; i++) {
360 ka = get_End_keepalive(oe, i);
362 if (get_irn_visited(ka) <= vfl) {
363 /* We didn't copy the node yet. */
364 set_irg_visited(irg, vfl);
365 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
367 add_End_keepalive(ne, get_new_node(ka));
371 /* start block sometimes only reached after keep alives */
372 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
373 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
377 * Copies the graph reachable from current_ir_graph->end to the obstack
378 * in current_ir_graph and fixes the environment.
379 * Then fixes the fields in current_ir_graph containing nodes of the
382 * @param copy_node_nr If non-zero, the node number will be copied
385 copy_graph_env(int copy_node_nr) {
386 ir_graph *irg = current_ir_graph;
387 ir_node *old_end, *new_anchor;
390 /* remove end_except and end_reg nodes */
391 old_end = get_irg_end(irg);
392 set_irg_end_except (irg, old_end);
393 set_irg_end_reg (irg, old_end);
395 /* Not all nodes remembered in irg might be reachable
396 from the end node. Assure their link is set to NULL, so that
397 we can test whether new nodes have been computed. */
398 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
399 ir_node *n = get_irg_anchor(irg, i);
401 set_new_node(n, NULL);
403 /* we use the block walk flag for removing Bads from Blocks ins. */
404 inc_irg_block_visited(irg);
407 copy_graph(irg, copy_node_nr);
410 old_end = get_irg_end(irg);
411 new_anchor = new_Anchor(irg);
413 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
414 ir_node *n = get_irg_anchor(irg, i);
416 set_irn_n(new_anchor, i, get_new_node(n));
419 irg->anchor = new_anchor;
421 /* ensure the new anchor is placed in the endblock */
422 set_nodes_block(new_anchor, get_irg_end_block(irg));
426 * Copies all reachable nodes to a new obstack. Removes bad inputs
427 * from block nodes and the corresponding inputs from Phi nodes.
428 * Merges single exit blocks with single entry blocks and removes
430 * Adds all new nodes to a new hash table for CSE. Does not
431 * perform CSE, so the hash table might contain common subexpressions.
433 void dead_node_elimination(ir_graph *irg) {
435 #ifdef INTERPROCEDURAL_VIEW
436 int rem_ipview = get_interprocedural_view();
438 struct obstack *graveyard_obst = NULL;
439 struct obstack *rebirth_obst = NULL;
440 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
442 /* inform statistics that we started a dead-node elimination run */
443 hook_dead_node_elim(irg, 1);
445 /* Remember external state of current_ir_graph. */
446 rem = current_ir_graph;
447 current_ir_graph = irg;
448 #ifdef INTERPROCEDURAL_VIEW
449 set_interprocedural_view(0);
452 assert(get_irg_phase_state(irg) != phase_building);
454 /* Handle graph state */
455 free_callee_info(irg);
459 /* @@@ so far we loose loops when copying */
460 free_loop_information(irg);
462 set_irg_doms_inconsistent(irg);
464 /* A quiet place, where the old obstack can rest in peace,
465 until it will be cremated. */
466 graveyard_obst = irg->obst;
468 /* A new obstack, where the reachable nodes will be copied to. */
469 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
470 irg->obst = rebirth_obst;
471 obstack_init(irg->obst);
472 irg->last_node_idx = 0;
474 /* We also need a new value table for CSE */
475 del_identities(irg->value_table);
476 irg->value_table = new_identities();
478 /* Copy the graph from the old to the new obstack */
479 copy_graph_env(/*copy_node_nr=*/1);
481 /* Free memory from old unoptimized obstack */
482 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
483 xfree(graveyard_obst); /* ... then free it. */
485 /* inform statistics that the run is over */
486 hook_dead_node_elim(irg, 0);
488 current_ir_graph = rem;
489 #ifdef INTERPROCEDURAL_VIEW
490 set_interprocedural_view(rem_ipview);
495 * Relink bad predecessors of a block and store the old in array to the
496 * link field. This function is called by relink_bad_predecessors().
497 * The array of link field starts with the block operand at position 0.
498 * If block has bad predecessors, create a new in array without bad preds.
499 * Otherwise let in array untouched.
501 static void relink_bad_block_predecessors(ir_node *n, void *env) {
502 ir_node **new_in, *irn;
503 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
506 /* if link field of block is NULL, look for bad predecessors otherwise
507 this is already done */
508 if (is_Block(n) && get_irn_link(n) == NULL) {
509 /* save old predecessors in link field (position 0 is the block operand)*/
510 set_irn_link(n, get_irn_in(n));
512 /* count predecessors without bad nodes */
513 old_irn_arity = get_irn_arity(n);
514 for (i = 0; i < old_irn_arity; i++)
515 if (!is_Bad(get_irn_n(n, i)))
518 /* arity changing: set new predecessors without bad nodes */
519 if (new_irn_arity < old_irn_arity) {
520 /* Get new predecessor array. We do not resize the array, as we must
521 keep the old one to update Phis. */
522 new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
524 /* set new predecessors in array */
527 for (i = 0; i < old_irn_arity; i++) {
528 irn = get_irn_n(n, i);
530 new_in[new_irn_n] = irn;
531 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
535 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
536 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
538 } /* ir node has bad predecessors */
539 } /* Block is not relinked */
543 * Relinks Bad predecessors from Blocks and Phis called by walker
544 * remove_bad_predecesors(). If n is a Block, call
545 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
546 * function of Phi's Block. If this block has bad predecessors, relink preds
549 static void relink_bad_predecessors(ir_node *n, void *env) {
550 ir_node *block, **old_in;
551 int i, old_irn_arity, new_irn_arity;
553 /* relink bad predecessors of a block */
555 relink_bad_block_predecessors(n, env);
557 /* If Phi node relink its block and its predecessors */
559 /* Relink predecessors of phi's block */
560 block = get_nodes_block(n);
561 if (get_irn_link(block) == NULL)
562 relink_bad_block_predecessors(block, env);
564 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
565 old_irn_arity = ARR_LEN(old_in);
567 /* Relink Phi predecessors if count of predecessors changed */
568 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
569 /* set new predecessors in array
570 n->in[0] remains the same block */
572 for(i = 1; i < old_irn_arity; i++)
573 if (!is_Bad(old_in[i])) {
574 n->in[new_irn_arity] = n->in[i];
575 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
579 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
580 ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
582 } /* n is a Phi node */
586 * Removes Bad Bad predecessors from Blocks and the corresponding
587 * inputs to Phi nodes as in dead_node_elimination but without
589 * On walking up set the link field to NULL, on walking down call
590 * relink_bad_predecessors() (This function stores the old in array
591 * to the link field and sets a new in array if arity of predecessors
594 void remove_bad_predecessors(ir_graph *irg) {
595 panic("Fix backedge handling first");
596 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
603 __)|_| | \_/ | \_/(/_ |_/\__|__
605 The following stuff implements a facility that automatically patches
606 registered ir_node pointers to the new node when a dead node elimination occurs.
609 struct _survive_dce_t {
613 hook_entry_t dead_node_elim;
614 hook_entry_t dead_node_elim_subst;
617 typedef struct _survive_dce_list_t {
618 struct _survive_dce_list_t *next;
620 } survive_dce_list_t;
622 static void dead_node_hook(void *context, ir_graph *irg, int start) {
623 survive_dce_t *sd = context;
626 /* Create a new map before the dead node elimination is performed. */
628 sd->new_places = pmap_create_ex(pmap_count(sd->places));
630 /* Patch back all nodes if dead node elimination is over and something is to be done. */
631 pmap_destroy(sd->places);
632 sd->places = sd->new_places;
633 sd->new_places = NULL;
638 * Hook called when dead node elimination replaces old by nw.
640 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
641 survive_dce_t *sd = context;
642 survive_dce_list_t *list = pmap_get(sd->places, old);
645 /* If the node is to be patched back, write the new address to all registered locations. */
647 survive_dce_list_t *p;
649 for (p = list; p; p = p->next)
652 pmap_insert(sd->new_places, nw, list);
657 * Make a new Survive DCE environment.
659 survive_dce_t *new_survive_dce(void) {
660 survive_dce_t *res = xmalloc(sizeof(res[0]));
661 obstack_init(&res->obst);
662 res->places = pmap_create();
663 res->new_places = NULL;
665 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
666 res->dead_node_elim.context = res;
667 res->dead_node_elim.next = NULL;
669 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
670 res->dead_node_elim_subst.context = res;
671 res->dead_node_elim_subst.next = NULL;
673 #ifndef FIRM_ENABLE_HOOKS
674 assert(0 && "need hooks enabled");
677 register_hook(hook_dead_node_elim, &res->dead_node_elim);
678 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
683 * Free a Survive DCE environment.
685 void free_survive_dce(survive_dce_t *sd) {
686 obstack_free(&sd->obst, NULL);
687 pmap_destroy(sd->places);
688 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
689 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
694 * Register a node pointer to be patched upon DCE.
695 * When DCE occurs, the node pointer specified by @p place will be
696 * patched to the new address of the node it is pointing to.
698 * @param sd The Survive DCE environment.
699 * @param place The address of the node pointer.
701 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
702 if (*place != NULL) {
703 ir_node *irn = *place;
704 survive_dce_list_t *curr = pmap_get(sd->places, irn);
705 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
710 pmap_insert(sd->places, irn, nw);
714 /*--------------------------------------------------------------------*/
715 /* Functionality for inlining */
716 /*--------------------------------------------------------------------*/
719 * Copy node for inlineing. Updates attributes that change when
720 * inlineing but not for dead node elimination.
722 * Copies the node by calling copy_node() and then updates the entity if
723 * it's a local one. env must be a pointer of the frame type of the
724 * inlined procedure. The new entities must be in the link field of
728 copy_node_inline(ir_node *n, void *env) {
730 ir_type *frame_tp = (ir_type *)env;
734 nn = get_new_node (n);
736 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
737 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
739 } else if (is_Block(n)) {
740 nn = get_new_node (n);
741 nn->attr.block.irg = current_ir_graph;
746 * Walker: checks if P_value_arg_base is used.
748 static void find_addr(ir_node *node, void *env) {
749 int *allow_inline = env;
751 is_Start(get_Proj_pred(node)) &&
752 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
758 * Check if we can inline a given call.
759 * Currently, we cannot inline two cases:
760 * - call with compound arguments
761 * - graphs that take the address of a parameter
763 * check these conditions here
765 static int can_inline(ir_node *call, ir_graph *called_graph) {
766 ir_type *call_type = get_Call_type(call);
767 int params, ress, i, res;
768 assert(is_Method_type(call_type));
770 params = get_method_n_params(call_type);
771 ress = get_method_n_ress(call_type);
773 /* check parameters for compound arguments */
774 for (i = 0; i < params; ++i) {
775 ir_type *p_type = get_method_param_type(call_type, i);
777 if (is_compound_type(p_type))
781 /* check results for compound arguments */
782 for (i = 0; i < ress; ++i) {
783 ir_type *r_type = get_method_res_type(call_type, i);
785 if (is_compound_type(r_type))
790 irg_walk_graph(called_graph, find_addr, NULL, &res);
796 exc_handler = 0, /**< There is a handler. */
797 exc_to_end = 1, /**< Branches to End. */
798 exc_no_handler = 2 /**< Exception handling not represented. */
801 /* Inlines a method at the given call site. */
802 int inline_method(ir_node *call, ir_graph *called_graph) {
804 ir_node *post_call, *post_bl;
805 ir_node *in[pn_Start_max];
806 ir_node *end, *end_bl;
810 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity;
811 enum exc_mode exc_handling;
812 ir_type *called_frame, *curr_frame;
813 irg_inline_property prop = get_irg_inline_property(called_graph);
816 if (prop == irg_inline_forbidden)
819 ent = get_irg_entity(called_graph);
821 /* Do not inline variadic functions. */
822 if (get_method_variadicity(get_entity_type(ent)) == variadicity_variadic)
825 assert(get_method_n_params(get_entity_type(ent)) ==
826 get_method_n_params(get_Call_type(call)));
829 * We cannot inline a recursive call. The graph must be copied before
830 * the call the inline_method() using create_irg_copy().
832 if (called_graph == current_ir_graph)
836 * currently, we cannot inline two cases:
837 * - call with compound arguments
838 * - graphs that take the address of a parameter
840 if (! can_inline(call, called_graph))
843 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
844 rem_opt = get_opt_optimize();
847 /* Handle graph state */
848 assert(get_irg_phase_state(current_ir_graph) != phase_building);
849 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
850 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
851 set_irg_outs_inconsistent(current_ir_graph);
852 set_irg_extblk_inconsistent(current_ir_graph);
853 set_irg_doms_inconsistent(current_ir_graph);
854 set_irg_loopinfo_inconsistent(current_ir_graph);
855 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
857 /* -- Check preconditions -- */
858 assert(is_Call(call));
860 /* here we know we WILL inline, so inform the statistics */
861 hook_inline(call, called_graph);
863 /* -- Decide how to handle exception control flow: Is there a handler
864 for the Call node, or do we branch directly to End on an exception?
866 0 There is a handler.
868 2 Exception handling not represented in Firm. -- */
870 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
871 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
872 long proj_nr = get_Proj_proj(proj);
873 if (proj_nr == pn_Call_X_except) Xproj = proj;
874 if (proj_nr == pn_Call_M_except) Mproj = proj;
876 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
877 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
878 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
882 the procedure and later replaces the Start node of the called graph.
883 Post_call is the old Call node and collects the results of the called
884 graph. Both will end up being a tuple. -- */
885 post_bl = get_nodes_block(call);
886 set_irg_current_block(current_ir_graph, post_bl);
887 /* XxMxPxPxPxT of Start + parameter of Call */
888 in[pn_Start_X_initial_exec] = new_Jmp();
889 in[pn_Start_M] = get_Call_mem(call);
890 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
891 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
892 in[pn_Start_P_tls] = get_irg_tls(current_ir_graph);
893 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
894 /* in[pn_Start_P_value_arg_base] = ??? */
895 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
896 pre_call = new_Tuple(pn_Start_max - 1, in);
900 The new block gets the ins of the old block, pre_call and all its
901 predecessors and all Phi nodes. -- */
902 part_block(pre_call);
904 /* -- Prepare state for dead node elimination -- */
905 /* Visited flags in calling irg must be >= flag in called irg.
906 Else walker and arity computation will not work. */
907 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
908 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
909 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
910 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
911 /* Set pre_call as new Start node in link field of the start node of
912 calling graph and pre_calls block as new block for the start block
914 Further mark these nodes so that they are not visited by the
916 set_irn_link(get_irg_start(called_graph), pre_call);
917 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
918 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
919 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
920 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
921 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
923 /* Initialize for compaction of in arrays */
924 inc_irg_block_visited(current_ir_graph);
926 /* -- Replicate local entities of the called_graph -- */
927 /* copy the entities. */
928 called_frame = get_irg_frame_type(called_graph);
929 curr_frame = get_irg_frame_type(current_ir_graph);
930 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
931 ir_entity *new_ent, *old_ent;
932 old_ent = get_class_member(called_frame, i);
933 new_ent = copy_entity_own(old_ent, curr_frame);
934 set_entity_link(old_ent, new_ent);
937 /* visited is > than that of called graph. With this trick visited will
938 remain unchanged so that an outer walker, e.g., searching the call nodes
939 to inline, calling this inline will not visit the inlined nodes. */
940 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
942 /* -- Performing dead node elimination inlines the graph -- */
943 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
945 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
946 get_irg_frame_type(called_graph));
948 /* Repair called_graph */
949 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
950 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
951 set_Block_block_visited(get_irg_start_block(called_graph), 0);
953 /* -- Merge the end of the inlined procedure with the call site -- */
954 /* We will turn the old Call node into a Tuple with the following
957 0: Phi of all Memories of Return statements.
958 1: Jmp from new Block that merges the control flow from all exception
959 predecessors of the old end block.
960 2: Tuple of all arguments.
961 3: Phi of Exception memories.
962 In case the old Call directly branches to End on an exception we don't
963 need the block merging all exceptions nor the Phi of the exception
967 /* -- Precompute some values -- */
968 end_bl = get_new_node(get_irg_end_block(called_graph));
969 end = get_new_node(get_irg_end(called_graph));
970 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
971 n_res = get_method_n_ress(get_Call_type(call));
973 res_pred = xmalloc(n_res * sizeof(*res_pred));
974 cf_pred = xmalloc(arity * sizeof(*res_pred));
976 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
978 /* -- archive keepalives -- */
979 irn_arity = get_irn_arity(end);
980 for (i = 0; i < irn_arity; i++) {
981 ir_node *ka = get_End_keepalive(end, i);
983 add_End_keepalive(get_irg_end(current_ir_graph), ka);
986 /* The new end node will die. We need not free as the in array is on the obstack:
987 copy_node() only generated 'D' arrays. */
989 /* -- Replace Return nodes by Jump nodes. -- */
991 for (i = 0; i < arity; i++) {
993 ret = get_irn_n(end_bl, i);
994 if (is_Return(ret)) {
995 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
999 set_irn_in(post_bl, n_ret, cf_pred);
1001 /* -- Build a Tuple for all results of the method.
1002 Add Phi node if there was more than one Return. -- */
1003 turn_into_tuple(post_call, pn_Call_max);
1004 /* First the Memory-Phi */
1006 for (i = 0; i < arity; i++) {
1007 ret = get_irn_n(end_bl, i);
1008 if (is_Return(ret)) {
1009 cf_pred[n_ret] = get_Return_mem(ret);
1013 phi = new_Phi(n_ret, cf_pred, mode_M);
1014 set_Tuple_pred(call, pn_Call_M_regular, phi);
1015 /* Conserve Phi-list for further inlinings -- but might be optimized */
1016 if (get_nodes_block(phi) == post_bl) {
1017 set_irn_link(phi, get_irn_link(post_bl));
1018 set_irn_link(post_bl, phi);
1020 /* Now the real results */
1022 for (j = 0; j < n_res; j++) {
1024 for (i = 0; i < arity; i++) {
1025 ret = get_irn_n(end_bl, i);
1026 if (is_Return(ret)) {
1027 cf_pred[n_ret] = get_Return_res(ret, j);
1032 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1036 /* Conserve Phi-list for further inlinings -- but might be optimized */
1037 if (get_nodes_block(phi) == post_bl) {
1038 set_Phi_next(phi, get_Block_phis(post_bl));
1039 set_Block_phis(post_bl, phi);
1042 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1044 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1046 /* handle the regular call */
1047 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1049 /* For now, we cannot inline calls with value_base */
1050 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1052 /* Finally the exception control flow.
1053 We have two (three) possible situations:
1054 First if the Call branches to an exception handler: We need to add a Phi node to
1055 collect the memory containing the exception objects. Further we need
1056 to add another block to get a correct representation of this Phi. To
1057 this block we add a Jmp that resolves into the X output of the Call
1058 when the Call is turned into a tuple.
1059 Second the Call branches to End, the exception is not handled. Just
1060 add all inlined exception branches to the End node.
1061 Third: there is no Exception edge at all. Handle as case two. */
1062 if (exc_handling == exc_handler) {
1064 for (i = 0; i < arity; i++) {
1066 ret = get_irn_n(end_bl, i);
1067 irn = skip_Proj(ret);
1068 if (is_fragile_op(irn) || is_Raise(irn)) {
1069 cf_pred[n_exc] = ret;
1074 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1075 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1076 /* The Phi for the memories with the exception objects */
1078 for (i = 0; i < arity; i++) {
1080 ret = skip_Proj(get_irn_n(end_bl, i));
1082 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1084 } else if (is_fragile_op(ret)) {
1085 /* We rely that all cfops have the memory output at the same position. */
1086 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1088 } else if (is_Raise(ret)) {
1089 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1093 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1095 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1096 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1099 ir_node *main_end_bl;
1100 int main_end_bl_arity;
1101 ir_node **end_preds;
1103 /* assert(exc_handling == 1 || no exceptions. ) */
1105 for (i = 0; i < arity; i++) {
1106 ir_node *ret = get_irn_n(end_bl, i);
1107 ir_node *irn = skip_Proj(ret);
1109 if (is_fragile_op(irn) || is_Raise(irn)) {
1110 cf_pred[n_exc] = ret;
1114 main_end_bl = get_irg_end_block(current_ir_graph);
1115 main_end_bl_arity = get_irn_arity(main_end_bl);
1116 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1118 for (i = 0; i < main_end_bl_arity; ++i)
1119 end_preds[i] = get_irn_n(main_end_bl, i);
1120 for (i = 0; i < n_exc; ++i)
1121 end_preds[main_end_bl_arity + i] = cf_pred[i];
1122 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1123 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1124 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1130 /* -- Turn CSE back on. -- */
1131 set_optimize(rem_opt);
1136 /********************************************************************/
1137 /* Apply inlineing to small methods. */
1138 /********************************************************************/
1140 static struct obstack temp_obst;
1142 /** Represents a possible inlinable call in a graph. */
1143 typedef struct _call_entry call_entry;
1144 struct _call_entry {
1145 ir_node *call; /**< the Call */
1146 ir_graph *callee; /**< the callee called here */
1147 call_entry *next; /**< for linking the next one */
1148 unsigned weight; /**< the weight of the call */
1152 * environment for inlining small irgs
1154 typedef struct _inline_env_t {
1155 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1156 call_entry *head; /**< the head of the call entry list */
1157 call_entry *tail; /**< the tail of the call entry list */
1161 * Returns the irg called from a Call node. If the irg is not
1162 * known, NULL is returned.
1164 * @param call the call node
1166 static ir_graph *get_call_called_irg(ir_node *call) {
1169 addr = get_Call_ptr(call);
1170 if (is_Global(addr)) {
1171 ir_entity *ent = get_Global_entity(addr);
1172 return get_entity_irg(ent);
1179 * Walker: Collect all calls to known graphs inside a graph.
1181 static void collect_calls(ir_node *call, void *env) {
1182 if (is_Call(call)) {
1183 ir_graph *called_irg = get_call_called_irg(call);
1185 if (called_irg != NULL) {
1186 /* The Call node calls a locally defined method. Remember to inline. */
1187 inline_env_t *ienv = env;
1188 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1190 entry->callee = called_irg;
1194 if (ienv->tail == NULL)
1197 ienv->tail->next = entry;
1204 * Inlines all small methods at call sites where the called address comes
1205 * from a Const node that references the entity representing the called
1207 * The size argument is a rough measure for the code size of the method:
1208 * Methods where the obstack containing the firm graph is smaller than
1211 void inline_small_irgs(ir_graph *irg, int size) {
1212 ir_graph *rem = current_ir_graph;
1215 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1217 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1219 current_ir_graph = irg;
1220 /* Handle graph state */
1221 assert(get_irg_phase_state(irg) != phase_building);
1222 free_callee_info(irg);
1224 /* Find Call nodes to inline.
1225 (We can not inline during a walk of the graph, as inlineing the same
1226 method several times changes the visited flag of the walked graph:
1227 after the first inlineing visited of the callee equals visited of
1228 the caller. With the next inlineing both are increased.) */
1229 obstack_init(&env.obst);
1230 env.head = env.tail = NULL;
1231 irg_walk_graph(irg, NULL, collect_calls, &env);
1233 if (env.head != NULL) {
1234 /* There are calls to inline */
1235 collect_phiprojs(irg);
1236 for (entry = env.head; entry != NULL; entry = entry->next) {
1237 ir_graph *callee = entry->callee;
1238 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1239 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1240 inline_method(entry->call, callee);
1244 obstack_free(&env.obst, NULL);
1245 current_ir_graph = rem;
1249 * Environment for inlining irgs.
1252 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1253 int n_blocks; /**< Number of Blocks in graph without Start and End block. */
1254 int n_nodes_orig; /**< for statistics */
1255 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1256 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1257 int n_call_nodes; /**< Number of Call nodes in the graph. */
1258 int n_call_nodes_orig; /**< for statistics */
1259 int n_callers; /**< Number of known graphs that call this graphs. */
1260 int n_callers_orig; /**< for statistics */
1261 int got_inline; /**< Set, if at least one call inside this graph was inlined. */
1262 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1266 * Allocate a new environment for inlining.
1268 static inline_irg_env *alloc_inline_irg_env(void) {
1269 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1270 env->n_nodes = -2; /* do not count count Start, End */
1271 env->n_blocks = -2; /* do not count count Start, End Block */
1272 env->n_nodes_orig = -2; /* do not count Start, End */
1273 env->call_head = NULL;
1274 env->call_tail = NULL;
1275 env->n_call_nodes = 0;
1276 env->n_call_nodes_orig = 0;
1278 env->n_callers_orig = 0;
1279 env->got_inline = 0;
1280 env->local_weights = NULL;
1284 typedef struct walker_env {
1285 inline_irg_env *x; /**< the inline environment */
1286 char ignore_runtime; /**< the ignore runtime flag */
1287 char ignore_callers; /**< if set, do change callers data */
1291 * post-walker: collect all calls in the inline-environment
1292 * of a graph and sum some statistics.
1294 static void collect_calls2(ir_node *call, void *ctx) {
1296 inline_irg_env *x = env->x;
1297 ir_opcode code = get_irn_opcode(call);
1301 /* count meaningful nodes in irg */
1302 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1303 if (code != iro_Block) {
1311 if (code != iro_Call) return;
1313 /* check, if it's a runtime call */
1314 if (env->ignore_runtime) {
1315 ir_node *symc = get_Call_ptr(call);
1317 if (is_Global(symc)) {
1318 ir_entity *ent = get_Global_entity(symc);
1320 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1325 /* collect all call nodes */
1327 ++x->n_call_nodes_orig;
1329 callee = get_call_called_irg(call);
1330 if (callee != NULL) {
1331 if (! env->ignore_callers) {
1332 inline_irg_env *callee_env = get_irg_link(callee);
1333 /* count all static callers */
1334 ++callee_env->n_callers;
1335 ++callee_env->n_callers_orig;
1338 /* link it in the list of possible inlinable entries */
1339 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1341 entry->callee = callee;
1343 if (x->call_tail == NULL)
1344 x->call_head = entry;
1346 x->call_tail->next = entry;
1347 x->call_tail = entry;
1352 * Returns TRUE if the number of callers is 0 in the irg's environment,
1353 * hence this irg is a leave.
1355 INLINE static int is_leave(ir_graph *irg) {
1356 inline_irg_env *env = get_irg_link(irg);
1357 return env->n_call_nodes == 0;
1361 * Returns TRUE if the number of nodes in the callee is
1362 * smaller then size in the irg's environment.
1364 INLINE static int is_smaller(ir_graph *callee, int size) {
1365 inline_irg_env *env = get_irg_link(callee);
1366 return env->n_nodes < size;
1370 * Append the nodes of the list src to the nodes of the list in environment dst.
1372 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1373 call_entry *entry, *nentry;
1375 /* Note that the src list points to Call nodes in the inlined graph, but
1376 we need Call nodes in our graph. Luckily the inliner leaves this information
1377 in the link field. */
1378 for (entry = src; entry != NULL; entry = entry->next) {
1379 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1380 nentry->call = get_irn_link(entry->call);
1381 nentry->callee = entry->callee;
1382 nentry->next = NULL;
1383 dst->call_tail->next = nentry;
1384 dst->call_tail = nentry;
1389 * Inlines small leave methods at call sites where the called address comes
1390 * from a Const node that references the entity representing the called
1392 * The size argument is a rough measure for the code size of the method:
1393 * Methods where the obstack containing the firm graph is smaller than
1396 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1397 inline_irg_env *env;
1403 call_entry *entry, *tail;
1404 const call_entry *centry;
1405 pmap *copied_graphs;
1406 pmap_entry *pm_entry;
1407 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1409 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1410 rem = current_ir_graph;
1411 obstack_init(&temp_obst);
1413 /* a map for the copied graphs, used to inline recursive calls */
1414 copied_graphs = pmap_create();
1416 /* extend all irgs by a temporary data structure for inlining. */
1417 n_irgs = get_irp_n_irgs();
1418 for (i = 0; i < n_irgs; ++i)
1419 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1421 /* Precompute information in temporary data structure. */
1422 wenv.ignore_runtime = ignore_runtime;
1423 wenv.ignore_callers = 0;
1424 for (i = 0; i < n_irgs; ++i) {
1425 ir_graph *irg = get_irp_irg(i);
1427 assert(get_irg_phase_state(irg) != phase_building);
1428 free_callee_info(irg);
1430 wenv.x = get_irg_link(irg);
1431 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1434 /* -- and now inline. -- */
1436 /* Inline leaves recursively -- we might construct new leaves. */
1440 for (i = 0; i < n_irgs; ++i) {
1442 int phiproj_computed = 0;
1444 current_ir_graph = get_irp_irg(i);
1445 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1448 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1451 if (env->n_nodes > maxsize) break;
1454 callee = entry->callee;
1456 if (is_leave(callee) && (
1457 is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1458 if (!phiproj_computed) {
1459 phiproj_computed = 1;
1460 collect_phiprojs(current_ir_graph);
1462 did_inline = inline_method(call, callee);
1465 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1467 /* was inlined, must be recomputed */
1468 phiproj_computed = 0;
1470 /* Do some statistics */
1471 env->got_inline = 1;
1472 --env->n_call_nodes;
1473 env->n_nodes += callee_env->n_nodes;
1474 --callee_env->n_callers;
1476 /* remove this call from the list */
1478 tail->next = entry->next;
1480 env->call_head = entry->next;
1486 env->call_tail = tail;
1488 } while (did_inline);
1490 /* inline other small functions. */
1491 for (i = 0; i < n_irgs; ++i) {
1493 int phiproj_computed = 0;
1495 current_ir_graph = get_irp_irg(i);
1496 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1498 /* note that the list of possible calls is updated during the process */
1500 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1505 callee = entry->callee;
1507 e = pmap_find(copied_graphs, callee);
1510 * Remap callee if we have a copy.
1511 * FIXME: Should we do this only for recursive Calls ?
1516 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1517 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1518 if (current_ir_graph == callee) {
1520 * Recursive call: we cannot directly inline because we cannot walk
1521 * the graph and change it. So we have to make a copy of the graph
1525 inline_irg_env *callee_env;
1529 * No copy yet, create one.
1530 * Note that recursive methods are never leaves, so it is sufficient
1531 * to test this condition here.
1533 copy = create_irg_copy(callee);
1535 /* create_irg_copy() destroys the Proj links, recompute them */
1536 phiproj_computed = 0;
1538 /* allocate new environment */
1539 callee_env = alloc_inline_irg_env();
1540 set_irg_link(copy, callee_env);
1542 wenv.x = callee_env;
1543 wenv.ignore_callers = 1;
1544 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1547 * Enter the entity of the original graph. This is needed
1548 * for inline_method(). However, note that ent->irg still points
1549 * to callee, NOT to copy.
1551 set_irg_entity(copy, get_irg_entity(callee));
1553 pmap_insert(copied_graphs, callee, copy);
1556 /* we have only one caller: the original graph */
1557 callee_env->n_callers = 1;
1558 callee_env->n_callers_orig = 1;
1560 if (! phiproj_computed) {
1561 phiproj_computed = 1;
1562 collect_phiprojs(current_ir_graph);
1564 did_inline = inline_method(call, callee);
1566 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1568 /* was inlined, must be recomputed */
1569 phiproj_computed = 0;
1571 /* callee was inline. Append it's call list. */
1572 env->got_inline = 1;
1573 --env->n_call_nodes;
1574 append_call_list(env, callee_env->call_head);
1575 env->n_call_nodes += callee_env->n_call_nodes;
1576 env->n_nodes += callee_env->n_nodes;
1577 --callee_env->n_callers;
1579 /* after we have inlined callee, all called methods inside callee
1580 are now called once more */
1581 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1582 inline_irg_env *penv = get_irg_link(centry->callee);
1586 /* remove this call from the list */
1588 tail->next = entry->next;
1590 env->call_head = entry->next;
1596 env->call_tail = tail;
1599 for (i = 0; i < n_irgs; ++i) {
1600 irg = get_irp_irg(i);
1601 env = (inline_irg_env *)get_irg_link(irg);
1603 if (env->got_inline) {
1604 /* this irg got calls inlined */
1605 set_irg_outs_inconsistent(irg);
1606 set_irg_doms_inconsistent(irg);
1608 optimize_graph_df(irg);
1611 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1612 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1613 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1614 env->n_callers_orig, env->n_callers,
1615 get_entity_name(get_irg_entity(irg))));
1619 /* kill the copied graphs: we don't need them anymore */
1620 foreach_pmap(copied_graphs, pm_entry) {
1621 ir_graph *copy = pm_entry->value;
1623 /* reset the entity, otherwise it will be deleted in the next step ... */
1624 set_irg_entity(copy, NULL);
1625 free_ir_graph(copy);
1627 pmap_destroy(copied_graphs);
1629 obstack_free(&temp_obst, NULL);
1630 current_ir_graph = rem;
1634 static void *VISITED = &v;
1637 * Calculate the parameter weights for transmitting the address of a local variable.
1639 static unsigned calc_method_local_weight(ir_node *arg) {
1641 unsigned v, weight = 0;
1643 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1644 ir_node *succ = get_irn_out(arg, i);
1646 switch (get_irn_opcode(succ)) {
1649 /* Loads and Store can be removed */
1653 /* check if all args are constant */
1654 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1655 ir_node *idx = get_Sel_index(succ, j);
1656 if (! is_Const(idx))
1659 /* Check users on this Sel. Note: if a 0 is returned here, there was
1660 some unsupported node. */
1661 v = calc_method_local_weight(succ);
1664 /* we can kill one Sel with constant indexes, this is cheap */
1668 /* any other node: unsupported yet or bad. */
1676 * Calculate the parameter weights for transmitting the address of a local variable.
1678 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1679 ir_entity *ent = get_irg_entity(irg);
1681 int nparams, i, proj_nr;
1682 ir_node *irg_args, *arg;
1684 mtp = get_entity_type(ent);
1685 nparams = get_method_n_params(mtp);
1687 /* allocate a new array. currently used as 'analysed' flag */
1688 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1690 /* If the method haven't parameters we have nothing to do. */
1694 assure_irg_outs(irg);
1695 irg_args = get_irg_args(irg);
1696 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1697 arg = get_irn_out(irg_args, i);
1698 proj_nr = get_Proj_proj(arg);
1699 ent->attr.mtd_attr.param_weight[proj_nr] = calc_method_local_weight(arg);
1704 * Calculate the benefice for transmitting an local variable address.
1705 * After inlining, the local variable might be transformed into a
1706 * SSA variable by scalar_replacement().
1708 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1709 inline_irg_env *env = get_irg_link(callee);
1711 if (env->local_weights != NULL) {
1712 if (pos < ARR_LEN(env->local_weights))
1713 return env->local_weights[pos];
1717 analyze_irg_local_weights(env, callee);
1719 if (pos < ARR_LEN(env->local_weights))
1720 return env->local_weights[pos];
1725 * calculate a benefice value for inlining the given call.
1727 static int calc_inline_benefice(ir_node *call, ir_graph *callee) {
1728 ir_entity *ent = get_irg_entity(callee);
1735 inline_irg_env *curr_env, *callee_env;
1737 if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1738 /* do NOT inline noreturn calls */
1742 /* costs for every passed parameter */
1743 n_params = get_Call_n_params(call);
1744 mtp = get_entity_type(ent);
1745 cc = get_method_calling_convention(mtp);
1746 if (cc & cc_reg_param) {
1747 /* register parameter, smaller costs for register parameters */
1748 int max_regs = cc & ~cc_bits;
1750 if (max_regs < n_params)
1751 weight += max_regs * 2 + (n_params - max_regs) * 5;
1753 weight += n_params * 2;
1755 /* parameters are passed an stack */
1756 weight += 5 * n_params;
1759 /* constant parameters improve the benefice */
1760 frame_ptr = get_irg_frame(current_ir_graph);
1761 for (i = 0; i < n_params; ++i) {
1762 ir_node *param = get_Call_param(call, i);
1764 if (is_Const(param) || is_SymConst(param))
1765 weight += get_method_param_weight(ent, i);
1766 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1768 * An address of a local variable is transmitted. After inlining,
1769 * scalar_replacement might be able to remove the local variable,
1772 weight += get_method_local_adress_weight(callee, i);
1776 callee_env = get_irg_link(callee);
1777 if (get_entity_visibility(ent) == visibility_local &&
1778 callee_env->n_callers_orig == 1 &&
1779 callee != current_ir_graph) {
1780 /* we are the only caller, give big bonus */
1784 /* do not inline big functions */
1785 weight -= callee_env->n_nodes;
1787 /* reduce the benefice if the current function is already big */
1788 curr_env = get_irg_link(current_ir_graph);
1789 weight -= curr_env->n_nodes / 100;
1791 /* give a bonus for functions with one block */
1792 if (callee_env->n_blocks == 1)
1793 weight = weight * 3 / 2;
1799 * Heuristic inliner. Calculates a benifice value for every call and inlines
1800 * those calls with a value higher than the threshold.
1802 void inline_functions(int inline_threshold) {
1803 inline_irg_env *env;
1809 call_entry *entry, *tail;
1810 const call_entry *centry;
1811 struct obstack obst;
1812 pmap *copied_graphs;
1813 pmap_entry *pm_entry;
1814 DEBUG_ONLY(firm_dbg_module_t *dbg;)
1816 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");
1817 rem = current_ir_graph;
1818 obstack_init(&obst);
1820 /* a map for the copied graphs, used to inline recursive calls */
1821 copied_graphs = pmap_create();
1823 /* extend all irgs by a temporary data structure for inlining. */
1824 n_irgs = get_irp_n_irgs();
1825 for (i = 0; i < n_irgs; ++i)
1826 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1828 /* Precompute information in temporary data structure. */
1829 wenv.ignore_runtime = 0;
1830 wenv.ignore_callers = 0;
1831 for (i = 0; i < n_irgs; ++i) {
1832 ir_graph *irg = get_irp_irg(i);
1834 assert(get_irg_phase_state(irg) != phase_building);
1835 free_callee_info(irg);
1837 wenv.x = get_irg_link(irg);
1838 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1841 /* -- and now inline. -- */
1842 for (i = 0; i < n_irgs; ++i) {
1844 int phiproj_computed = 0;
1846 current_ir_graph = get_irp_irg(i);
1847 env = get_irg_link(current_ir_graph);
1849 /* note that the list of possible calls is updated during the process */
1851 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1857 callee = entry->callee;
1859 /* calculate the benifice on the original call to prevent excessive inlining */
1860 benefice = calc_inline_benefice(call, callee);
1861 DB((dbg, SET_LEVEL_2, "In %+F Call %+F has benefice %d\n", current_ir_graph, callee, benefice));
1863 e = pmap_find(copied_graphs, callee);
1866 * Remap callee if we have a copy.
1867 * FIXME: Should we do this only for recursive Calls ?
1872 if (benefice > -inline_threshold ||
1873 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1874 if (current_ir_graph == callee) {
1876 * Recursive call: we cannot directly inline because we cannot walk
1877 * the graph and change it. So we have to make a copy of the graph
1881 inline_irg_env *callee_env;
1885 * No copy yet, create one.
1886 * Note that recursive methods are never leaves, so it is sufficient
1887 * to test this condition here.
1889 copy = create_irg_copy(callee);
1891 /* create_irg_copy() destroys the Proj links, recompute them */
1892 phiproj_computed = 0;
1894 /* allocate new environment */
1895 callee_env = alloc_inline_irg_env();
1896 set_irg_link(copy, callee_env);
1898 wenv.x = callee_env;
1899 wenv.ignore_callers = 1;
1900 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1903 * Enter the entity of the original graph. This is needed
1904 * for inline_method(). However, note that ent->irg still points
1905 * to callee, NOT to copy.
1907 set_irg_entity(copy, get_irg_entity(callee));
1909 pmap_insert(copied_graphs, callee, copy);
1912 /* we have only one caller: the original graph */
1913 callee_env->n_callers = 1;
1914 callee_env->n_callers_orig = 1;
1916 if (! phiproj_computed) {
1917 phiproj_computed = 1;
1918 collect_phiprojs(current_ir_graph);
1920 did_inline = inline_method(call, callee);
1922 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1924 /* was inlined, must be recomputed */
1925 phiproj_computed = 0;
1927 /* callee was inline. Append it's call list. */
1928 env->got_inline = 1;
1929 --env->n_call_nodes;
1930 append_call_list(env, callee_env->call_head);
1931 env->n_call_nodes += callee_env->n_call_nodes;
1932 env->n_nodes += callee_env->n_nodes;
1933 --callee_env->n_callers;
1935 /* after we have inlined callee, all called methods inside callee
1936 are now called once more */
1937 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1938 inline_irg_env *penv = get_irg_link(centry->callee);
1942 /* remove this call from the list */
1944 tail->next = entry->next;
1946 env->call_head = entry->next;
1952 env->call_tail = tail;
1955 for (i = 0; i < n_irgs; ++i) {
1956 irg = get_irp_irg(i);
1957 env = (inline_irg_env *)get_irg_link(irg);
1959 if (env->got_inline) {
1960 /* this irg got calls inlined */
1961 set_irg_outs_inconsistent(irg);
1962 set_irg_doms_inconsistent(irg);
1964 optimize_graph_df(irg);
1967 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1968 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1969 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1970 env->n_callers_orig, env->n_callers,
1971 get_entity_name(get_irg_entity(irg))));
1975 /* kill the copied graphs: we don't need them anymore */
1976 foreach_pmap(copied_graphs, pm_entry) {
1977 ir_graph *copy = pm_entry->value;
1979 /* reset the entity, otherwise it will be deleted in the next step ... */
1980 set_irg_entity(copy, NULL);
1981 free_ir_graph(copy);
1983 pmap_destroy(copied_graphs);
1985 obstack_free(&obst, NULL);
1986 current_ir_graph = rem;