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"
53 #include "opt_inline_t.h"
58 #include "analyze_irg_args.h"
59 #include "iredges_t.h"
64 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
66 /*------------------------------------------------------------------*/
67 /* Routines for dead node elimination / copying garbage collection */
69 /*------------------------------------------------------------------*/
72 * Remember the new node in the old node by using a field all nodes have.
74 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
77 * Get this new node, before the old node is forgotten.
79 #define get_new_node(oldn) get_irn_link(oldn)
82 * Check if a new node was set.
84 #define has_new_node(n) (get_new_node(n) != NULL)
87 * We use the block_visited flag to mark that we have computed the
88 * number of useful predecessors for this block.
89 * Further we encode the new arity in this flag in the old blocks.
90 * Remembering the arity is useful, as it saves a lot of pointer
91 * accesses. This function is called for all Phi and Block nodes
95 compute_new_arity(ir_node *b) {
96 int i, res, irn_arity;
99 irg_v = get_irg_block_visited(current_ir_graph);
100 block_v = get_Block_block_visited(b);
101 if (block_v >= irg_v) {
102 /* we computed the number of preds for this block and saved it in the
104 return block_v - irg_v;
106 /* compute the number of good predecessors */
107 res = irn_arity = get_irn_arity(b);
108 for (i = 0; i < irn_arity; i++)
109 if (is_Bad(get_irn_n(b, i))) res--;
110 /* save it in the flag. */
111 set_Block_block_visited(b, irg_v + res);
117 * Copies the node to the new obstack. The Ins of the new node point to
118 * the predecessors on the old obstack. For block/phi nodes not all
119 * predecessors might be copied. n->link points to the new node.
120 * For Phi and Block nodes the function allocates in-arrays with an arity
121 * only for useful predecessors. The arity is determined by counting
122 * the non-bad predecessors of the block.
124 * @param n The node to be copied
125 * @param env if non-NULL, the node number attribute will be copied to the new node
127 * Note: Also used for loop unrolling.
129 static void copy_node(ir_node *n, void *env) {
132 ir_op *op = get_irn_op(n);
136 /* node copied already */
138 } else if (op == op_Block) {
140 new_arity = compute_new_arity(n);
141 n->attr.block.graph_arr = NULL;
143 block = get_nodes_block(n);
145 new_arity = compute_new_arity(block);
147 new_arity = get_irn_arity(n);
150 nn = new_ir_node(get_irn_dbg_info(n),
157 /* Copy the attributes. These might point to additional data. If this
158 was allocated on the old obstack the pointers now are dangling. This
159 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
160 if (op == op_Block) {
161 /* we cannot allow blocks WITHOUT macroblock input */
162 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
164 copy_node_attr(n, nn);
168 int copy_node_nr = env != NULL;
170 /* for easier debugging, we want to copy the node numbers too */
171 nn->node_nr = n->node_nr;
177 hook_dead_node_elim_subst(current_ir_graph, n, nn);
181 * Copies new predecessors of old node to new node remembered in link.
182 * Spare the Bad predecessors of Phi and Block nodes.
184 static void copy_preds(ir_node *n, void *env) {
189 nn = get_new_node(n);
192 /* copy the macro block header */
193 ir_node *mbh = get_Block_MacroBlock(n);
196 /* this block is a macroblock header */
197 set_Block_MacroBlock(nn, nn);
199 /* get the macro block header */
200 ir_node *nmbh = get_new_node(mbh);
201 assert(nmbh != NULL);
202 set_Block_MacroBlock(nn, nmbh);
205 /* Don't copy Bad nodes. */
207 irn_arity = get_irn_arity(n);
208 for (i = 0; i < irn_arity; i++) {
209 if (! is_Bad(get_irn_n(n, i))) {
210 ir_node *pred = get_irn_n(n, i);
211 set_irn_n(nn, j, get_new_node(pred));
215 /* repair the block visited flag from above misuse. Repair it in both
216 graphs so that the old one can still be used. */
217 set_Block_block_visited(nn, 0);
218 set_Block_block_visited(n, 0);
219 /* Local optimization could not merge two subsequent blocks if
220 in array contained Bads. Now it's possible.
221 We don't call optimize_in_place as it requires
222 that the fields in ir_graph are set properly. */
223 if ((get_opt_control_flow_straightening()) &&
224 (get_Block_n_cfgpreds(nn) == 1) &&
225 is_Jmp(get_Block_cfgpred(nn, 0))) {
226 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
228 /* Jmp jumps into the block it is in -- deal self cycle. */
229 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
230 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
235 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
236 /* Don't copy node if corresponding predecessor in block is Bad.
237 The Block itself should not be Bad. */
238 block = get_nodes_block(n);
239 set_nodes_block(nn, get_new_node(block));
241 irn_arity = get_irn_arity(n);
242 for (i = 0; i < irn_arity; i++) {
243 if (! is_Bad(get_irn_n(block, i))) {
244 ir_node *pred = get_irn_n(n, i);
245 set_irn_n(nn, j, get_new_node(pred));
246 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
250 /* If the pre walker reached this Phi after the post walker visited the
251 block block_visited is > 0. */
252 set_Block_block_visited(get_nodes_block(n), 0);
253 /* Compacting the Phi's ins might generate Phis with only one
255 if (get_irn_arity(nn) == 1)
256 exchange(nn, get_irn_n(nn, 0));
258 irn_arity = get_irn_arity(n);
259 for (i = -1; i < irn_arity; i++)
260 set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
262 /* Now the new node is complete. We can add it to the hash table for CSE.
263 @@@ inlining aborts if we identify End. Why? */
265 add_identities(current_ir_graph->value_table, nn);
269 * Copies the graph recursively, compacts the keep-alives of the end node.
271 * @param irg the graph to be copied
272 * @param copy_node_nr If non-zero, the node number will be copied
274 static void copy_graph(ir_graph *irg, int copy_node_nr) {
275 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
276 ir_node *ka; /* keep alive */
280 /* Some nodes must be copied by hand, sigh */
281 vfl = get_irg_visited(irg);
282 set_irg_visited(irg, vfl + 1);
284 oe = get_irg_end(irg);
285 mark_irn_visited(oe);
286 /* copy the end node by hand, allocate dynamic in array! */
287 ne = new_ir_node(get_irn_dbg_info(oe),
294 /* Copy the attributes. Well, there might be some in the future... */
295 copy_node_attr(oe, ne);
296 set_new_node(oe, ne);
298 /* copy the Bad node */
299 ob = get_irg_bad(irg);
300 mark_irn_visited(ob);
301 nb = new_ir_node(get_irn_dbg_info(ob),
308 copy_node_attr(ob, nb);
309 set_new_node(ob, nb);
311 /* copy the NoMem node */
312 om = get_irg_no_mem(irg);
313 mark_irn_visited(om);
314 nm = new_ir_node(get_irn_dbg_info(om),
321 copy_node_attr(om, nm);
322 set_new_node(om, nm);
324 /* copy the live nodes */
325 set_irg_visited(irg, vfl);
326 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
328 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
330 /* visit the anchors as well */
331 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
332 ir_node *n = get_irg_anchor(irg, i);
334 if (n && (get_irn_visited(n) <= vfl)) {
335 set_irg_visited(irg, vfl);
336 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
340 /* copy_preds for the end node ... */
341 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
343 /*- ... and now the keep alives. -*/
344 /* First pick the not marked block nodes and walk them. We must pick these
345 first as else we will oversee blocks reachable from Phis. */
346 irn_arity = get_End_n_keepalives(oe);
347 for (i = 0; i < irn_arity; i++) {
348 ka = get_End_keepalive(oe, i);
350 if (get_irn_visited(ka) <= vfl) {
351 /* We must keep the block alive and copy everything reachable */
352 set_irg_visited(irg, vfl);
353 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
355 add_End_keepalive(ne, get_new_node(ka));
359 /* Now pick other nodes. Here we will keep all! */
360 irn_arity = get_End_n_keepalives(oe);
361 for (i = 0; i < irn_arity; i++) {
362 ka = get_End_keepalive(oe, i);
364 if (get_irn_visited(ka) <= vfl) {
365 /* We didn't copy the node yet. */
366 set_irg_visited(irg, vfl);
367 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
369 add_End_keepalive(ne, get_new_node(ka));
373 /* start block sometimes only reached after keep alives */
374 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
375 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
379 * Copies the graph reachable from current_ir_graph->end to the obstack
380 * in current_ir_graph and fixes the environment.
381 * Then fixes the fields in current_ir_graph containing nodes of the
384 * @param copy_node_nr If non-zero, the node number will be copied
387 copy_graph_env(int copy_node_nr) {
388 ir_graph *irg = current_ir_graph;
389 ir_node *old_end, *new_anchor;
392 /* remove end_except and end_reg nodes */
393 old_end = get_irg_end(irg);
394 set_irg_end_except (irg, old_end);
395 set_irg_end_reg (irg, old_end);
397 /* Not all nodes remembered in irg might be reachable
398 from the end node. Assure their link is set to NULL, so that
399 we can test whether new nodes have been computed. */
400 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
401 ir_node *n = get_irg_anchor(irg, i);
403 set_new_node(n, NULL);
405 /* we use the block walk flag for removing Bads from Blocks ins. */
406 inc_irg_block_visited(irg);
409 copy_graph(irg, copy_node_nr);
412 old_end = get_irg_end(irg);
413 new_anchor = new_Anchor(irg);
415 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
416 ir_node *n = get_irg_anchor(irg, i);
418 set_irn_n(new_anchor, i, get_new_node(n));
421 irg->anchor = new_anchor;
423 /* ensure the new anchor is placed in the endblock */
424 set_nodes_block(new_anchor, get_irg_end_block(irg));
428 * Copies all reachable nodes to a new obstack. Removes bad inputs
429 * from block nodes and the corresponding inputs from Phi nodes.
430 * Merges single exit blocks with single entry blocks and removes
432 * Adds all new nodes to a new hash table for CSE. Does not
433 * perform CSE, so the hash table might contain common subexpressions.
435 void dead_node_elimination(ir_graph *irg) {
437 #ifdef INTERPROCEDURAL_VIEW
438 int rem_ipview = get_interprocedural_view();
440 struct obstack *graveyard_obst = NULL;
441 struct obstack *rebirth_obst = NULL;
442 assert(! edges_activated(irg) && "dead node elimination requires disabled edges");
444 /* inform statistics that we started a dead-node elimination run */
445 hook_dead_node_elim(irg, 1);
447 /* Remember external state of current_ir_graph. */
448 rem = current_ir_graph;
449 current_ir_graph = irg;
450 #ifdef INTERPROCEDURAL_VIEW
451 set_interprocedural_view(0);
454 assert(get_irg_phase_state(irg) != phase_building);
456 /* Handle graph state */
457 free_callee_info(irg);
461 /* @@@ so far we loose loops when copying */
462 free_loop_information(irg);
464 set_irg_doms_inconsistent(irg);
466 /* A quiet place, where the old obstack can rest in peace,
467 until it will be cremated. */
468 graveyard_obst = irg->obst;
470 /* A new obstack, where the reachable nodes will be copied to. */
471 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
472 irg->obst = rebirth_obst;
473 obstack_init(irg->obst);
474 irg->last_node_idx = 0;
476 /* We also need a new value table for CSE */
477 del_identities(irg->value_table);
478 irg->value_table = new_identities();
480 /* Copy the graph from the old to the new obstack */
481 copy_graph_env(/*copy_node_nr=*/1);
483 /* Free memory from old unoptimized obstack */
484 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
485 xfree(graveyard_obst); /* ... then free it. */
487 /* inform statistics that the run is over */
488 hook_dead_node_elim(irg, 0);
490 current_ir_graph = rem;
491 #ifdef INTERPROCEDURAL_VIEW
492 set_interprocedural_view(rem_ipview);
497 * Relink bad predecessors of a block and store the old in array to the
498 * link field. This function is called by relink_bad_predecessors().
499 * The array of link field starts with the block operand at position 0.
500 * If block has bad predecessors, create a new in array without bad preds.
501 * Otherwise let in array untouched.
503 static void relink_bad_block_predecessors(ir_node *n, void *env) {
504 ir_node **new_in, *irn;
505 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
508 /* if link field of block is NULL, look for bad predecessors otherwise
509 this is already done */
510 if (is_Block(n) && get_irn_link(n) == NULL) {
511 /* save old predecessors in link field (position 0 is the block operand)*/
512 set_irn_link(n, get_irn_in(n));
514 /* count predecessors without bad nodes */
515 old_irn_arity = get_irn_arity(n);
516 for (i = 0; i < old_irn_arity; i++)
517 if (!is_Bad(get_irn_n(n, i)))
520 /* arity changing: set new predecessors without bad nodes */
521 if (new_irn_arity < old_irn_arity) {
522 /* Get new predecessor array. We do not resize the array, as we must
523 keep the old one to update Phis. */
524 new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
526 /* set new predecessors in array */
529 for (i = 0; i < old_irn_arity; i++) {
530 irn = get_irn_n(n, i);
532 new_in[new_irn_n] = irn;
533 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
537 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
538 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
540 } /* ir node has bad predecessors */
541 } /* Block is not relinked */
545 * Relinks Bad predecessors from Blocks and Phis called by walker
546 * remove_bad_predecesors(). If n is a Block, call
547 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
548 * function of Phi's Block. If this block has bad predecessors, relink preds
551 static void relink_bad_predecessors(ir_node *n, void *env) {
552 ir_node *block, **old_in;
553 int i, old_irn_arity, new_irn_arity;
555 /* relink bad predecessors of a block */
557 relink_bad_block_predecessors(n, env);
559 /* If Phi node relink its block and its predecessors */
561 /* Relink predecessors of phi's block */
562 block = get_nodes_block(n);
563 if (get_irn_link(block) == NULL)
564 relink_bad_block_predecessors(block, env);
566 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
567 old_irn_arity = ARR_LEN(old_in);
569 /* Relink Phi predecessors if count of predecessors changed */
570 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
571 /* set new predecessors in array
572 n->in[0] remains the same block */
574 for(i = 1; i < old_irn_arity; i++)
575 if (!is_Bad(old_in[i])) {
576 n->in[new_irn_arity] = n->in[i];
577 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
581 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
582 ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
584 } /* n is a Phi node */
588 * Removes Bad Bad predecessors from Blocks and the corresponding
589 * inputs to Phi nodes as in dead_node_elimination but without
591 * On walking up set the link field to NULL, on walking down call
592 * relink_bad_predecessors() (This function stores the old in array
593 * to the link field and sets a new in array if arity of predecessors
596 void remove_bad_predecessors(ir_graph *irg) {
597 panic("Fix backedge handling first");
598 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
605 __)|_| | \_/ | \_/(/_ |_/\__|__
607 The following stuff implements a facility that automatically patches
608 registered ir_node pointers to the new node when a dead node elimination occurs.
611 struct _survive_dce_t {
615 hook_entry_t dead_node_elim;
616 hook_entry_t dead_node_elim_subst;
619 typedef struct _survive_dce_list_t {
620 struct _survive_dce_list_t *next;
622 } survive_dce_list_t;
624 static void dead_node_hook(void *context, ir_graph *irg, int start) {
625 survive_dce_t *sd = context;
628 /* Create a new map before the dead node elimination is performed. */
630 sd->new_places = pmap_create_ex(pmap_count(sd->places));
632 /* Patch back all nodes if dead node elimination is over and something is to be done. */
633 pmap_destroy(sd->places);
634 sd->places = sd->new_places;
635 sd->new_places = NULL;
640 * Hook called when dead node elimination replaces old by nw.
642 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
643 survive_dce_t *sd = context;
644 survive_dce_list_t *list = pmap_get(sd->places, old);
647 /* If the node is to be patched back, write the new address to all registered locations. */
649 survive_dce_list_t *p;
651 for (p = list; p; p = p->next)
654 pmap_insert(sd->new_places, nw, list);
659 * Make a new Survive DCE environment.
661 survive_dce_t *new_survive_dce(void) {
662 survive_dce_t *res = xmalloc(sizeof(res[0]));
663 obstack_init(&res->obst);
664 res->places = pmap_create();
665 res->new_places = NULL;
667 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
668 res->dead_node_elim.context = res;
669 res->dead_node_elim.next = NULL;
671 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
672 res->dead_node_elim_subst.context = res;
673 res->dead_node_elim_subst.next = NULL;
675 #ifndef FIRM_ENABLE_HOOKS
676 assert(0 && "need hooks enabled");
679 register_hook(hook_dead_node_elim, &res->dead_node_elim);
680 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
685 * Free a Survive DCE environment.
687 void free_survive_dce(survive_dce_t *sd) {
688 obstack_free(&sd->obst, NULL);
689 pmap_destroy(sd->places);
690 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
691 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
696 * Register a node pointer to be patched upon DCE.
697 * When DCE occurs, the node pointer specified by @p place will be
698 * patched to the new address of the node it is pointing to.
700 * @param sd The Survive DCE environment.
701 * @param place The address of the node pointer.
703 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
704 if (*place != NULL) {
705 ir_node *irn = *place;
706 survive_dce_list_t *curr = pmap_get(sd->places, irn);
707 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
712 pmap_insert(sd->places, irn, nw);
716 /*--------------------------------------------------------------------*/
717 /* Functionality for inlining */
718 /*--------------------------------------------------------------------*/
721 * Copy node for inlineing. Updates attributes that change when
722 * inlineing but not for dead node elimination.
724 * Copies the node by calling copy_node() and then updates the entity if
725 * it's a local one. env must be a pointer of the frame type of the
726 * inlined procedure. The new entities must be in the link field of
730 copy_node_inline(ir_node *n, void *env) {
732 ir_type *frame_tp = (ir_type *)env;
736 nn = get_new_node (n);
738 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
739 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
741 } else if (is_Block(n)) {
742 nn = get_new_node (n);
743 nn->attr.block.irg = current_ir_graph;
748 * Walker: checks if P_value_arg_base is used.
750 static void find_addr(ir_node *node, void *env) {
751 int *allow_inline = env;
753 is_Start(get_Proj_pred(node)) &&
754 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
760 * Check if we can inline a given call.
761 * Currently, we cannot inline two cases:
762 * - call with compound arguments
763 * - graphs that take the address of a parameter
765 * check these conditions here
767 static int can_inline(ir_node *call, ir_graph *called_graph) {
768 ir_type *call_type = get_Call_type(call);
769 int params, ress, i, res;
770 assert(is_Method_type(call_type));
772 params = get_method_n_params(call_type);
773 ress = get_method_n_ress(call_type);
775 /* check parameters for compound arguments */
776 for (i = 0; i < params; ++i) {
777 ir_type *p_type = get_method_param_type(call_type, i);
779 if (is_compound_type(p_type))
783 /* check results for compound arguments */
784 for (i = 0; i < ress; ++i) {
785 ir_type *r_type = get_method_res_type(call_type, i);
787 if (is_compound_type(r_type))
792 irg_walk_graph(called_graph, find_addr, NULL, &res);
798 exc_handler = 0, /**< There is a handler. */
799 exc_to_end = 1, /**< Branches to End. */
800 exc_no_handler = 2 /**< Exception handling not represented. */
803 /* Inlines a method at the given call site. */
804 int inline_method(ir_node *call, ir_graph *called_graph) {
806 ir_node *post_call, *post_bl;
807 ir_node *in[pn_Start_max];
808 ir_node *end, *end_bl;
812 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity;
813 enum exc_mode exc_handling;
814 ir_type *called_frame, *curr_frame;
817 irg_inline_property prop = get_irg_inline_property(called_graph);
819 if (prop == irg_inline_forbidden)
822 ent = get_irg_entity(called_graph);
824 /* Do not inline variadic functions. */
825 if (get_method_variadicity(get_entity_type(ent)) == variadicity_variadic)
828 assert(get_method_n_params(get_entity_type(ent)) ==
829 get_method_n_params(get_Call_type(call)));
831 irg = get_irn_irg(call);
834 * We cannot inline a recursive call. The graph must be copied before
835 * the call the inline_method() using create_irg_copy().
837 if (called_graph == irg)
841 * currently, we cannot inline two cases:
842 * - call with compound arguments
843 * - graphs that take the address of a parameter
845 if (! can_inline(call, called_graph))
848 rem = current_ir_graph;
849 current_ir_graph = irg;
851 DB((dbg, SET_LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
853 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
854 rem_opt = get_opt_optimize();
857 /* Handle graph state */
858 assert(get_irg_phase_state(irg) != phase_building);
859 assert(get_irg_pinned(irg) == op_pin_state_pinned);
860 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
861 set_irg_outs_inconsistent(irg);
862 set_irg_extblk_inconsistent(irg);
863 set_irg_doms_inconsistent(irg);
864 set_irg_loopinfo_inconsistent(irg);
865 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
867 /* -- Check preconditions -- */
868 assert(is_Call(call));
870 /* here we know we WILL inline, so inform the statistics */
871 hook_inline(call, called_graph);
873 /* -- Decide how to handle exception control flow: Is there a handler
874 for the Call node, or do we branch directly to End on an exception?
876 0 There is a handler.
878 2 Exception handling not represented in Firm. -- */
880 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
881 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
882 long proj_nr = get_Proj_proj(proj);
883 if (proj_nr == pn_Call_X_except) Xproj = proj;
884 if (proj_nr == pn_Call_M_except) Mproj = proj;
886 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
887 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
888 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
892 the procedure and later replaces the Start node of the called graph.
893 Post_call is the old Call node and collects the results of the called
894 graph. Both will end up being a tuple. -- */
895 post_bl = get_nodes_block(call);
896 set_irg_current_block(irg, post_bl);
897 /* XxMxPxPxPxT of Start + parameter of Call */
898 in[pn_Start_X_initial_exec] = new_Jmp();
899 in[pn_Start_M] = get_Call_mem(call);
900 in[pn_Start_P_frame_base] = get_irg_frame(irg);
901 in[pn_Start_P_globals] = get_irg_globals(irg);
902 in[pn_Start_P_tls] = get_irg_tls(irg);
903 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
904 /* in[pn_Start_P_value_arg_base] = ??? */
905 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
906 pre_call = new_Tuple(pn_Start_max - 1, in);
910 The new block gets the ins of the old block, pre_call and all its
911 predecessors and all Phi nodes. -- */
912 part_block(pre_call);
914 /* -- Prepare state for dead node elimination -- */
915 /* Visited flags in calling irg must be >= flag in called irg.
916 Else walker and arity computation will not work. */
917 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
918 set_irg_visited(irg, get_irg_visited(called_graph)+1);
919 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
920 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
921 /* Set pre_call as new Start node in link field of the start node of
922 calling graph and pre_calls block as new block for the start block
924 Further mark these nodes so that they are not visited by the
926 set_irn_link(get_irg_start(called_graph), pre_call);
927 set_irn_visited(get_irg_start(called_graph), get_irg_visited(irg));
928 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
929 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(irg));
930 set_irn_link(get_irg_bad(called_graph), get_irg_bad(irg));
931 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(irg));
933 /* Initialize for compaction of in arrays */
934 inc_irg_block_visited(irg);
936 /* -- Replicate local entities of the called_graph -- */
937 /* copy the entities. */
938 called_frame = get_irg_frame_type(called_graph);
939 curr_frame = get_irg_frame_type(irg);
940 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
941 ir_entity *new_ent, *old_ent;
942 old_ent = get_class_member(called_frame, i);
943 new_ent = copy_entity_own(old_ent, curr_frame);
944 set_entity_link(old_ent, new_ent);
947 /* visited is > than that of called graph. With this trick visited will
948 remain unchanged so that an outer walker, e.g., searching the call nodes
949 to inline, calling this inline will not visit the inlined nodes. */
950 set_irg_visited(irg, get_irg_visited(irg)-1);
952 /* -- Performing dead node elimination inlines the graph -- */
953 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
955 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
956 get_irg_frame_type(called_graph));
958 /* Repair called_graph */
959 set_irg_visited(called_graph, get_irg_visited(irg));
960 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
961 set_Block_block_visited(get_irg_start_block(called_graph), 0);
963 /* -- Merge the end of the inlined procedure with the call site -- */
964 /* We will turn the old Call node into a Tuple with the following
967 0: Phi of all Memories of Return statements.
968 1: Jmp from new Block that merges the control flow from all exception
969 predecessors of the old end block.
970 2: Tuple of all arguments.
971 3: Phi of Exception memories.
972 In case the old Call directly branches to End on an exception we don't
973 need the block merging all exceptions nor the Phi of the exception
977 /* -- Precompute some values -- */
978 end_bl = get_new_node(get_irg_end_block(called_graph));
979 end = get_new_node(get_irg_end(called_graph));
980 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
981 n_res = get_method_n_ress(get_Call_type(call));
983 res_pred = xmalloc(n_res * sizeof(*res_pred));
984 cf_pred = xmalloc(arity * sizeof(*res_pred));
986 set_irg_current_block(irg, post_bl); /* just to make sure */
988 /* -- archive keepalives -- */
989 irn_arity = get_irn_arity(end);
990 for (i = 0; i < irn_arity; i++) {
991 ir_node *ka = get_End_keepalive(end, i);
993 add_End_keepalive(get_irg_end(irg), ka);
996 /* The new end node will die. We need not free as the in array is on the obstack:
997 copy_node() only generated 'D' arrays. */
999 /* -- Replace Return nodes by Jump nodes. -- */
1001 for (i = 0; i < arity; i++) {
1003 ret = get_irn_n(end_bl, i);
1004 if (is_Return(ret)) {
1005 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1009 set_irn_in(post_bl, n_ret, cf_pred);
1011 /* -- Build a Tuple for all results of the method.
1012 Add Phi node if there was more than one Return. -- */
1013 turn_into_tuple(post_call, pn_Call_max);
1014 /* First the Memory-Phi */
1016 for (i = 0; i < arity; i++) {
1017 ret = get_irn_n(end_bl, i);
1018 if (is_Return(ret)) {
1019 cf_pred[n_ret] = get_Return_mem(ret);
1023 phi = new_Phi(n_ret, cf_pred, mode_M);
1024 set_Tuple_pred(call, pn_Call_M_regular, phi);
1025 /* Conserve Phi-list for further inlinings -- but might be optimized */
1026 if (get_nodes_block(phi) == post_bl) {
1027 set_irn_link(phi, get_irn_link(post_bl));
1028 set_irn_link(post_bl, phi);
1030 /* Now the real results */
1032 for (j = 0; j < n_res; j++) {
1034 for (i = 0; i < arity; i++) {
1035 ret = get_irn_n(end_bl, i);
1036 if (is_Return(ret)) {
1037 cf_pred[n_ret] = get_Return_res(ret, j);
1042 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1046 /* Conserve Phi-list for further inlinings -- but might be optimized */
1047 if (get_nodes_block(phi) == post_bl) {
1048 set_Phi_next(phi, get_Block_phis(post_bl));
1049 set_Block_phis(post_bl, phi);
1052 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1054 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1056 /* handle the regular call */
1057 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1059 /* For now, we cannot inline calls with value_base */
1060 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1062 /* Finally the exception control flow.
1063 We have two (three) possible situations:
1064 First if the Call branches to an exception handler: We need to add a Phi node to
1065 collect the memory containing the exception objects. Further we need
1066 to add another block to get a correct representation of this Phi. To
1067 this block we add a Jmp that resolves into the X output of the Call
1068 when the Call is turned into a tuple.
1069 Second the Call branches to End, the exception is not handled. Just
1070 add all inlined exception branches to the End node.
1071 Third: there is no Exception edge at all. Handle as case two. */
1072 if (exc_handling == exc_handler) {
1074 for (i = 0; i < arity; i++) {
1076 ret = get_irn_n(end_bl, i);
1077 irn = skip_Proj(ret);
1078 if (is_fragile_op(irn) || is_Raise(irn)) {
1079 cf_pred[n_exc] = ret;
1084 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1085 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1086 /* The Phi for the memories with the exception objects */
1088 for (i = 0; i < arity; i++) {
1090 ret = skip_Proj(get_irn_n(end_bl, i));
1092 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1094 } else if (is_fragile_op(ret)) {
1095 /* We rely that all cfops have the memory output at the same position. */
1096 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1098 } else if (is_Raise(ret)) {
1099 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1103 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1105 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1106 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1109 ir_node *main_end_bl;
1110 int main_end_bl_arity;
1111 ir_node **end_preds;
1113 /* assert(exc_handling == 1 || no exceptions. ) */
1115 for (i = 0; i < arity; i++) {
1116 ir_node *ret = get_irn_n(end_bl, i);
1117 ir_node *irn = skip_Proj(ret);
1119 if (is_fragile_op(irn) || is_Raise(irn)) {
1120 cf_pred[n_exc] = ret;
1124 main_end_bl = get_irg_end_block(irg);
1125 main_end_bl_arity = get_irn_arity(main_end_bl);
1126 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1128 for (i = 0; i < main_end_bl_arity; ++i)
1129 end_preds[i] = get_irn_n(main_end_bl, i);
1130 for (i = 0; i < n_exc; ++i)
1131 end_preds[main_end_bl_arity + i] = cf_pred[i];
1132 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1133 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1134 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1140 /* -- Turn CSE back on. -- */
1141 set_optimize(rem_opt);
1142 current_ir_graph = rem;
1147 /********************************************************************/
1148 /* Apply inlineing to small methods. */
1149 /********************************************************************/
1151 static struct obstack temp_obst;
1153 /** Represents a possible inlinable call in a graph. */
1154 typedef struct _call_entry call_entry;
1155 struct _call_entry {
1156 ir_node *call; /**< the Call */
1157 ir_graph *callee; /**< the callee called here */
1158 call_entry *next; /**< for linking the next one */
1159 unsigned weight; /**< the weight of the call */
1163 * environment for inlining small irgs
1165 typedef struct _inline_env_t {
1166 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1167 call_entry *head; /**< the head of the call entry list */
1168 call_entry *tail; /**< the tail of the call entry list */
1172 * Returns the irg called from a Call node. If the irg is not
1173 * known, NULL is returned.
1175 * @param call the call node
1177 static ir_graph *get_call_called_irg(ir_node *call) {
1180 addr = get_Call_ptr(call);
1181 if (is_Global(addr)) {
1182 ir_entity *ent = get_Global_entity(addr);
1183 return get_entity_irg(ent);
1190 * Walker: Collect all calls to known graphs inside a graph.
1192 static void collect_calls(ir_node *call, void *env) {
1193 if (is_Call(call)) {
1194 ir_graph *called_irg = get_call_called_irg(call);
1196 if (called_irg != NULL) {
1197 /* The Call node calls a locally defined method. Remember to inline. */
1198 inline_env_t *ienv = env;
1199 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1201 entry->callee = called_irg;
1205 if (ienv->tail == NULL)
1208 ienv->tail->next = entry;
1215 * Inlines all small methods at call sites where the called address comes
1216 * from a Const node that references the entity representing the called
1218 * The size argument is a rough measure for the code size of the method:
1219 * Methods where the obstack containing the firm graph is smaller than
1222 void inline_small_irgs(ir_graph *irg, int size) {
1223 ir_graph *rem = current_ir_graph;
1227 current_ir_graph = irg;
1228 /* Handle graph state */
1229 assert(get_irg_phase_state(irg) != phase_building);
1230 free_callee_info(irg);
1232 /* Find Call nodes to inline.
1233 (We can not inline during a walk of the graph, as inlineing the same
1234 method several times changes the visited flag of the walked graph:
1235 after the first inlineing visited of the callee equals visited of
1236 the caller. With the next inlineing both are increased.) */
1237 obstack_init(&env.obst);
1238 env.head = env.tail = NULL;
1239 irg_walk_graph(irg, NULL, collect_calls, &env);
1241 if (env.head != NULL) {
1242 /* There are calls to inline */
1243 collect_phiprojs(irg);
1244 for (entry = env.head; entry != NULL; entry = entry->next) {
1245 ir_graph *callee = entry->callee;
1246 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1247 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1248 inline_method(entry->call, callee);
1252 obstack_free(&env.obst, NULL);
1253 current_ir_graph = rem;
1257 * Environment for inlining irgs.
1260 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1261 int n_blocks; /**< Number of Blocks in graph without Start and End block. */
1262 int n_nodes_orig; /**< for statistics */
1263 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1264 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1265 int n_call_nodes; /**< Number of Call nodes in the graph. */
1266 int n_call_nodes_orig; /**< for statistics */
1267 int n_callers; /**< Number of known graphs that call this graphs. */
1268 int n_callers_orig; /**< for statistics */
1269 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1270 unsigned local_vars:1; /**< Set, if a inlined function gets the address of an inlined variable. */
1271 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1275 * Allocate a new environment for inlining.
1277 static inline_irg_env *alloc_inline_irg_env(void) {
1278 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1279 env->n_nodes = -2; /* do not count count Start, End */
1280 env->n_blocks = -2; /* do not count count Start, End Block */
1281 env->n_nodes_orig = -2; /* do not count Start, End */
1282 env->call_head = NULL;
1283 env->call_tail = NULL;
1284 env->n_call_nodes = 0;
1285 env->n_call_nodes_orig = 0;
1287 env->n_callers_orig = 0;
1288 env->got_inline = 0;
1289 env->local_vars = 0;
1290 env->local_weights = NULL;
1294 typedef struct walker_env {
1295 inline_irg_env *x; /**< the inline environment */
1296 char ignore_runtime; /**< the ignore runtime flag */
1297 char ignore_callers; /**< if set, do change callers data */
1301 * post-walker: collect all calls in the inline-environment
1302 * of a graph and sum some statistics.
1304 static void collect_calls2(ir_node *call, void *ctx) {
1306 inline_irg_env *x = env->x;
1307 ir_opcode code = get_irn_opcode(call);
1311 /* count meaningful nodes in irg */
1312 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1313 if (code != iro_Block) {
1321 if (code != iro_Call) return;
1323 /* check, if it's a runtime call */
1324 if (env->ignore_runtime) {
1325 ir_node *symc = get_Call_ptr(call);
1327 if (is_Global(symc)) {
1328 ir_entity *ent = get_Global_entity(symc);
1330 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1335 /* collect all call nodes */
1337 ++x->n_call_nodes_orig;
1339 callee = get_call_called_irg(call);
1340 if (callee != NULL) {
1341 if (! env->ignore_callers) {
1342 inline_irg_env *callee_env = get_irg_link(callee);
1343 /* count all static callers */
1344 ++callee_env->n_callers;
1345 ++callee_env->n_callers_orig;
1348 /* link it in the list of possible inlinable entries */
1349 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1351 entry->callee = callee;
1353 if (x->call_tail == NULL)
1354 x->call_head = entry;
1356 x->call_tail->next = entry;
1357 x->call_tail = entry;
1362 * Returns TRUE if the number of callers is 0 in the irg's environment,
1363 * hence this irg is a leave.
1365 INLINE static int is_leave(ir_graph *irg) {
1366 inline_irg_env *env = get_irg_link(irg);
1367 return env->n_call_nodes == 0;
1371 * Returns TRUE if the number of nodes in the callee is
1372 * smaller then size in the irg's environment.
1374 INLINE static int is_smaller(ir_graph *callee, int size) {
1375 inline_irg_env *env = get_irg_link(callee);
1376 return env->n_nodes < size;
1380 * Append the nodes of the list src to the nodes of the list in environment dst.
1382 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1383 call_entry *entry, *nentry;
1385 /* Note that the src list points to Call nodes in the inlined graph, but
1386 we need Call nodes in our graph. Luckily the inliner leaves this information
1387 in the link field. */
1388 for (entry = src; entry != NULL; entry = entry->next) {
1389 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1390 nentry->call = get_irn_link(entry->call);
1391 nentry->callee = entry->callee;
1392 nentry->next = NULL;
1393 dst->call_tail->next = nentry;
1394 dst->call_tail = nentry;
1399 * Inlines small leave methods at call sites where the called address comes
1400 * from a Const node that references the entity representing the called
1402 * The size argument is a rough measure for the code size of the method:
1403 * Methods where the obstack containing the firm graph is smaller than
1406 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1407 inline_irg_env *env;
1413 call_entry *entry, *tail;
1414 const call_entry *centry;
1415 pmap *copied_graphs;
1416 pmap_entry *pm_entry;
1418 rem = current_ir_graph;
1419 obstack_init(&temp_obst);
1421 /* a map for the copied graphs, used to inline recursive calls */
1422 copied_graphs = pmap_create();
1424 /* extend all irgs by a temporary data structure for inlining. */
1425 n_irgs = get_irp_n_irgs();
1426 for (i = 0; i < n_irgs; ++i)
1427 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1429 /* Precompute information in temporary data structure. */
1430 wenv.ignore_runtime = ignore_runtime;
1431 wenv.ignore_callers = 0;
1432 for (i = 0; i < n_irgs; ++i) {
1433 ir_graph *irg = get_irp_irg(i);
1435 assert(get_irg_phase_state(irg) != phase_building);
1436 free_callee_info(irg);
1438 wenv.x = get_irg_link(irg);
1439 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1442 /* -- and now inline. -- */
1444 /* Inline leaves recursively -- we might construct new leaves. */
1448 for (i = 0; i < n_irgs; ++i) {
1450 int phiproj_computed = 0;
1452 current_ir_graph = get_irp_irg(i);
1453 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1456 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1459 if (env->n_nodes > maxsize) break;
1462 callee = entry->callee;
1464 if (is_leave(callee) && (
1465 is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1466 if (!phiproj_computed) {
1467 phiproj_computed = 1;
1468 collect_phiprojs(current_ir_graph);
1470 did_inline = inline_method(call, callee);
1473 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1475 /* was inlined, must be recomputed */
1476 phiproj_computed = 0;
1478 /* Do some statistics */
1479 env->got_inline = 1;
1480 --env->n_call_nodes;
1481 env->n_nodes += callee_env->n_nodes;
1482 --callee_env->n_callers;
1484 /* remove this call from the list */
1486 tail->next = entry->next;
1488 env->call_head = entry->next;
1494 env->call_tail = tail;
1496 } while (did_inline);
1498 /* inline other small functions. */
1499 for (i = 0; i < n_irgs; ++i) {
1501 int phiproj_computed = 0;
1503 current_ir_graph = get_irp_irg(i);
1504 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1506 /* note that the list of possible calls is updated during the process */
1508 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1513 callee = entry->callee;
1515 e = pmap_find(copied_graphs, callee);
1518 * Remap callee if we have a copy.
1519 * FIXME: Should we do this only for recursive Calls ?
1524 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1525 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1526 if (current_ir_graph == callee) {
1528 * Recursive call: we cannot directly inline because we cannot walk
1529 * the graph and change it. So we have to make a copy of the graph
1533 inline_irg_env *callee_env;
1537 * No copy yet, create one.
1538 * Note that recursive methods are never leaves, so it is sufficient
1539 * to test this condition here.
1541 copy = create_irg_copy(callee);
1543 /* create_irg_copy() destroys the Proj links, recompute them */
1544 phiproj_computed = 0;
1546 /* allocate new environment */
1547 callee_env = alloc_inline_irg_env();
1548 set_irg_link(copy, callee_env);
1550 wenv.x = callee_env;
1551 wenv.ignore_callers = 1;
1552 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1555 * Enter the entity of the original graph. This is needed
1556 * for inline_method(). However, note that ent->irg still points
1557 * to callee, NOT to copy.
1559 set_irg_entity(copy, get_irg_entity(callee));
1561 pmap_insert(copied_graphs, callee, copy);
1564 /* we have only one caller: the original graph */
1565 callee_env->n_callers = 1;
1566 callee_env->n_callers_orig = 1;
1568 if (! phiproj_computed) {
1569 phiproj_computed = 1;
1570 collect_phiprojs(current_ir_graph);
1572 did_inline = inline_method(call, callee);
1574 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1576 /* was inlined, must be recomputed */
1577 phiproj_computed = 0;
1579 /* callee was inline. Append it's call list. */
1580 env->got_inline = 1;
1581 --env->n_call_nodes;
1582 append_call_list(env, callee_env->call_head);
1583 env->n_call_nodes += callee_env->n_call_nodes;
1584 env->n_nodes += callee_env->n_nodes;
1585 --callee_env->n_callers;
1587 /* after we have inlined callee, all called methods inside callee
1588 are now called once more */
1589 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1590 inline_irg_env *penv = get_irg_link(centry->callee);
1594 /* remove this call from the list */
1596 tail->next = entry->next;
1598 env->call_head = entry->next;
1604 env->call_tail = tail;
1607 for (i = 0; i < n_irgs; ++i) {
1608 irg = get_irp_irg(i);
1609 env = (inline_irg_env *)get_irg_link(irg);
1611 if (env->got_inline) {
1612 optimize_graph_df(irg);
1615 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1616 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1617 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1618 env->n_callers_orig, env->n_callers,
1619 get_entity_name(get_irg_entity(irg))));
1623 /* kill the copied graphs: we don't need them anymore */
1624 foreach_pmap(copied_graphs, pm_entry) {
1625 ir_graph *copy = pm_entry->value;
1627 /* reset the entity, otherwise it will be deleted in the next step ... */
1628 set_irg_entity(copy, NULL);
1629 free_ir_graph(copy);
1631 pmap_destroy(copied_graphs);
1633 obstack_free(&temp_obst, NULL);
1634 current_ir_graph = rem;
1638 * Calculate the parameter weights for transmitting the address of a local variable.
1640 static unsigned calc_method_local_weight(ir_node *arg) {
1642 unsigned v, weight = 0;
1644 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1645 ir_node *succ = get_irn_out(arg, i);
1647 switch (get_irn_opcode(succ)) {
1650 /* Loads and Store can be removed */
1654 /* check if all args are constant */
1655 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1656 ir_node *idx = get_Sel_index(succ, j);
1657 if (! is_Const(idx))
1660 /* Check users on this Sel. Note: if a 0 is returned here, there was
1661 some unsupported node. */
1662 v = calc_method_local_weight(succ);
1665 /* we can kill one Sel with constant indexes, this is cheap */
1669 /* when looking backward we might find Id nodes */
1670 weight += calc_method_local_weight(succ);
1673 /* unoptimized tuple */
1674 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1675 ir_node *pred = get_Tuple_pred(succ, j);
1677 /* look for Proj(j) */
1678 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1679 ir_node *succ_succ = get_irn_out(succ, k);
1680 if (is_Proj(succ_succ)) {
1681 if (get_Proj_proj(succ_succ) == j) {
1683 weight += calc_method_local_weight(succ_succ);
1686 /* this should NOT happen */
1693 /* any other node: unsupported yet or bad. */
1701 * Calculate the parameter weights for transmitting the address of a local variable.
1703 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1704 ir_entity *ent = get_irg_entity(irg);
1706 int nparams, i, proj_nr;
1707 ir_node *irg_args, *arg;
1709 mtp = get_entity_type(ent);
1710 nparams = get_method_n_params(mtp);
1712 /* allocate a new array. currently used as 'analysed' flag */
1713 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1715 /* If the method haven't parameters we have nothing to do. */
1719 assure_irg_outs(irg);
1720 irg_args = get_irg_args(irg);
1721 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1722 arg = get_irn_out(irg_args, i);
1723 proj_nr = get_Proj_proj(arg);
1724 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1729 * Calculate the benefice for transmitting an local variable address.
1730 * After inlining, the local variable might be transformed into a
1731 * SSA variable by scalar_replacement().
1733 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1734 inline_irg_env *env = get_irg_link(callee);
1736 if (env->local_weights != NULL) {
1737 if (pos < ARR_LEN(env->local_weights))
1738 return env->local_weights[pos];
1742 analyze_irg_local_weights(env, callee);
1744 if (pos < ARR_LEN(env->local_weights))
1745 return env->local_weights[pos];
1750 * calculate a benefice value for inlining the given call.
1752 static int calc_inline_benefice(ir_node *call, ir_graph *callee, unsigned *local_adr) {
1753 ir_entity *ent = get_irg_entity(callee);
1760 inline_irg_env *curr_env, *callee_env;
1762 if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1763 /* do NOT inline noreturn calls */
1767 /* costs for every passed parameter */
1768 n_params = get_Call_n_params(call);
1769 mtp = get_entity_type(ent);
1770 cc = get_method_calling_convention(mtp);
1771 if (cc & cc_reg_param) {
1772 /* register parameter, smaller costs for register parameters */
1773 int max_regs = cc & ~cc_bits;
1775 if (max_regs < n_params)
1776 weight += max_regs * 2 + (n_params - max_regs) * 5;
1778 weight += n_params * 2;
1780 /* parameters are passed an stack */
1781 weight += 5 * n_params;
1784 /* constant parameters improve the benefice */
1785 frame_ptr = get_irg_frame(current_ir_graph);
1786 for (i = 0; i < n_params; ++i) {
1787 ir_node *param = get_Call_param(call, i);
1789 if (is_Const(param) || is_SymConst(param))
1790 weight += get_method_param_weight(ent, i);
1791 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1793 * An address of a local variable is transmitted. After inlining,
1794 * scalar_replacement might be able to remove the local variable,
1797 v = get_method_local_adress_weight(callee, i);
1804 callee_env = get_irg_link(callee);
1805 if (get_entity_visibility(ent) == visibility_local &&
1806 callee_env->n_callers_orig == 1 &&
1807 callee != current_ir_graph) {
1808 /* we are the only caller, give big bonus */
1812 /* do not inline big functions */
1813 weight -= callee_env->n_nodes;
1815 /* reduce the benefice if the current function is already big */
1816 curr_env = get_irg_link(current_ir_graph);
1817 weight -= curr_env->n_nodes / 100;
1819 /* give a bonus for functions with one block */
1820 if (callee_env->n_blocks == 1)
1821 weight = weight * 3 / 2;
1827 * Heuristic inliner. Calculates a benifice value for every call and inlines
1828 * those calls with a value higher than the threshold.
1830 void inline_functions(int inline_threshold) {
1831 inline_irg_env *env;
1836 call_entry *entry, *tail;
1837 const call_entry *centry;
1838 pmap *copied_graphs;
1839 pmap_entry *pm_entry;
1841 rem = current_ir_graph;
1842 obstack_init(&temp_obst);
1844 /* a map for the copied graphs, used to inline recursive calls */
1845 copied_graphs = pmap_create();
1847 /* extend all irgs by a temporary data structure for inlining. */
1848 n_irgs = get_irp_n_irgs();
1849 for (i = 0; i < n_irgs; ++i)
1850 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1852 /* Precompute information in temporary data structure. */
1853 wenv.ignore_runtime = 0;
1854 wenv.ignore_callers = 0;
1855 for (i = 0; i < n_irgs; ++i) {
1856 ir_graph *irg = get_irp_irg(i);
1858 assert(get_irg_phase_state(irg) != phase_building);
1859 free_callee_info(irg);
1861 wenv.x = get_irg_link(irg);
1862 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1865 /* -- and now inline. -- */
1866 for (i = 0; i < n_irgs; ++i) {
1867 int phiproj_computed = 0;
1869 ir_graph *irg = get_irp_irg(i);
1871 current_ir_graph = irg;
1872 env = get_irg_link(irg);
1874 /* note that the list of possible calls is updated during the process */
1876 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1883 callee = entry->callee;
1885 /* calculate the benifice on the original call to prevent excessive inlining */
1887 benefice = calc_inline_benefice(call, callee, &local_adr);
1888 DB((dbg, SET_LEVEL_2, "In %+F Call %+F has benefice %d\n", irg, callee, benefice));
1890 e = pmap_find(copied_graphs, callee);
1893 * Remap callee if we have a copy.
1894 * FIXME: Should we do this only for recursive Calls ?
1899 if (benefice > -inline_threshold ||
1900 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1901 if (current_ir_graph == callee) {
1903 * Recursive call: we cannot directly inline because we cannot walk
1904 * the graph and change it. So we have to make a copy of the graph
1908 inline_irg_env *callee_env;
1912 * No copy yet, create one.
1913 * Note that recursive methods are never leaves, so it is sufficient
1914 * to test this condition here.
1916 copy = create_irg_copy(callee);
1918 /* create_irg_copy() destroys the Proj links, recompute them */
1919 phiproj_computed = 0;
1921 /* allocate new environment */
1922 callee_env = alloc_inline_irg_env();
1923 set_irg_link(copy, callee_env);
1925 wenv.x = callee_env;
1926 wenv.ignore_callers = 1;
1927 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1930 * Enter the entity of the original graph. This is needed
1931 * for inline_method(). However, note that ent->irg still points
1932 * to callee, NOT to copy.
1934 set_irg_entity(copy, get_irg_entity(callee));
1936 pmap_insert(copied_graphs, callee, copy);
1939 /* we have only one caller: the original graph */
1940 callee_env->n_callers = 1;
1941 callee_env->n_callers_orig = 1;
1943 if (! phiproj_computed) {
1944 phiproj_computed = 1;
1945 collect_phiprojs(current_ir_graph);
1947 did_inline = inline_method(call, callee);
1949 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1951 /* was inlined, must be recomputed */
1952 phiproj_computed = 0;
1954 /* callee was inline. Append it's call list. */
1955 env->got_inline = 1;
1957 env->local_vars = 1;
1958 --env->n_call_nodes;
1959 append_call_list(env, callee_env->call_head);
1960 env->n_call_nodes += callee_env->n_call_nodes;
1961 env->n_nodes += callee_env->n_nodes;
1962 --callee_env->n_callers;
1964 /* after we have inlined callee, all called methods inside callee
1965 are now called once more */
1966 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1967 inline_irg_env *penv = get_irg_link(centry->callee);
1971 /* remove this call from the list */
1973 tail->next = entry->next;
1975 env->call_head = entry->next;
1981 env->call_tail = tail;
1983 if (env->got_inline) {
1984 /* this irg got calls inlined: optimize it */
1986 /* scalar replacement does not work well with Tuple nodes, so optimize them away */
1987 optimize_graph_df(irg);
1989 if (env->local_vars) {
1990 if (scalar_replacement_opt(irg)) {
1991 optimize_graph_df(irg);
1997 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1998 DB((dbg, SET_LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1999 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2000 env->n_callers_orig, env->n_callers,
2001 get_entity_name(get_irg_entity(irg))));
2005 /* kill the copied graphs: we don't need them anymore */
2006 foreach_pmap(copied_graphs, pm_entry) {
2007 ir_graph *copy = pm_entry->value;
2009 /* reset the entity, otherwise it will be deleted in the next step ... */
2010 set_irg_entity(copy, NULL);
2011 free_ir_graph(copy);
2013 pmap_destroy(copied_graphs);
2015 obstack_free(&temp_obst, NULL);
2016 current_ir_graph = rem;
2019 void firm_init_inline(void) {
2020 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");