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) {
756 } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
758 * Refuse to inline alloca call unless user explicitly forced so as this
759 * may change program's memory overhead drastically when the function
760 * using alloca is called in loop. In GCC present in SPEC2000 inlining
761 * into schedule_block cause it to require 2GB of ram instead of 256MB.
763 * Sorryly this is true with our implementation also.
764 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
765 * disables inlining of functions using VLA (with are completely save).
768 * - add a flag to the Alloc node for "real" alloca() calls
769 * - add a new Stack-Restore node at the end of a function using alloca()
776 * Check if we can inline a given call.
777 * Currently, we cannot inline two cases:
778 * - call with compound arguments
779 * - graphs that take the address of a parameter
781 * check these conditions here
783 static int can_inline(ir_node *call, ir_graph *called_graph) {
784 ir_type *call_type = get_Call_type(call);
785 int params, ress, i, res;
786 assert(is_Method_type(call_type));
788 params = get_method_n_params(call_type);
789 ress = get_method_n_ress(call_type);
791 /* check parameters for compound arguments */
792 for (i = 0; i < params; ++i) {
793 ir_type *p_type = get_method_param_type(call_type, i);
795 if (is_compound_type(p_type))
799 /* check results for compound arguments */
800 for (i = 0; i < ress; ++i) {
801 ir_type *r_type = get_method_res_type(call_type, i);
803 if (is_compound_type(r_type))
808 irg_walk_graph(called_graph, find_addr, NULL, &res);
814 exc_handler = 0, /**< There is a handler. */
815 exc_to_end = 1, /**< Branches to End. */
816 exc_no_handler = 2 /**< Exception handling not represented. */
819 /* Inlines a method at the given call site. */
820 int inline_method(ir_node *call, ir_graph *called_graph) {
822 ir_node *post_call, *post_bl;
823 ir_node *in[pn_Start_max];
824 ir_node *end, *end_bl;
828 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity;
829 enum exc_mode exc_handling;
830 ir_type *called_frame, *curr_frame;
833 irg_inline_property prop = get_irg_inline_property(called_graph);
835 if (prop == irg_inline_forbidden)
838 ent = get_irg_entity(called_graph);
840 /* Do not inline variadic functions. */
841 if (get_method_variadicity(get_entity_type(ent)) == variadicity_variadic) {
842 /* Arg, KR functions are marked as variadic one's, so check further */
843 ir_type *mtp = get_entity_type(ent);
844 ir_type *ctp = get_Call_type(call);
845 int n_params = get_method_n_params(mtp);
848 /* This is too strong, but probably ok. Function calls with a wrong number of
849 parameters should not be inlined. */
850 if (n_params != get_method_n_params(ctp))
853 /* check types: for K&R calls, this was not done by the compiler. Again, this is
854 too strong, but ok for now. */
855 for (i = n_params - 1; i >= 0; --i) {
856 ir_type *param_tp = get_method_param_type(mtp, i);
857 ir_type *arg_tp = get_method_param_type(ctp, i);
859 if (param_tp != arg_tp)
862 DB((dbg, LEVEL_1, "Inlining allowed for variadic function %+F\n", called_graph));
863 /* types match, fine: when the frame is access, the inliner stops at can_inline() */
866 assert(get_method_n_params(get_entity_type(ent)) ==
867 get_method_n_params(get_Call_type(call)));
869 irg = get_irn_irg(call);
872 * We cannot inline a recursive call. The graph must be copied before
873 * the call the inline_method() using create_irg_copy().
875 if (called_graph == irg)
879 * currently, we cannot inline two cases:
880 * - call with compound arguments
881 * - graphs that take the address of a parameter
883 if (! can_inline(call, called_graph))
886 rem = current_ir_graph;
887 current_ir_graph = irg;
889 DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
891 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
892 rem_opt = get_opt_optimize();
895 /* Handle graph state */
896 assert(get_irg_phase_state(irg) != phase_building);
897 assert(get_irg_pinned(irg) == op_pin_state_pinned);
898 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
899 set_irg_outs_inconsistent(irg);
900 set_irg_extblk_inconsistent(irg);
901 set_irg_doms_inconsistent(irg);
902 set_irg_loopinfo_inconsistent(irg);
903 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
905 /* -- Check preconditions -- */
906 assert(is_Call(call));
908 /* here we know we WILL inline, so inform the statistics */
909 hook_inline(call, called_graph);
911 /* -- Decide how to handle exception control flow: Is there a handler
912 for the Call node, or do we branch directly to End on an exception?
914 0 There is a handler.
916 2 Exception handling not represented in Firm. -- */
918 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
919 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
920 long proj_nr = get_Proj_proj(proj);
921 if (proj_nr == pn_Call_X_except) Xproj = proj;
922 if (proj_nr == pn_Call_M_except) Mproj = proj;
924 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
925 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
926 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
930 the procedure and later replaces the Start node of the called graph.
931 Post_call is the old Call node and collects the results of the called
932 graph. Both will end up being a tuple. -- */
933 post_bl = get_nodes_block(call);
934 set_irg_current_block(irg, post_bl);
935 /* XxMxPxPxPxT of Start + parameter of Call */
936 in[pn_Start_X_initial_exec] = new_Jmp();
937 in[pn_Start_M] = get_Call_mem(call);
938 in[pn_Start_P_frame_base] = get_irg_frame(irg);
939 in[pn_Start_P_tls] = get_irg_tls(irg);
940 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
941 /* in[pn_Start_P_value_arg_base] = ??? */
942 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
943 pre_call = new_Tuple(pn_Start_max - 1, in);
947 The new block gets the ins of the old block, pre_call and all its
948 predecessors and all Phi nodes. -- */
949 part_block(pre_call);
951 /* -- Prepare state for dead node elimination -- */
952 /* Visited flags in calling irg must be >= flag in called irg.
953 Else walker and arity computation will not work. */
954 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
955 set_irg_visited(irg, get_irg_visited(called_graph)+1);
956 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
957 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
958 /* Set pre_call as new Start node in link field of the start node of
959 calling graph and pre_calls block as new block for the start block
961 Further mark these nodes so that they are not visited by the
963 set_irn_link(get_irg_start(called_graph), pre_call);
964 set_irn_visited(get_irg_start(called_graph), get_irg_visited(irg));
965 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
966 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(irg));
967 set_irn_link(get_irg_bad(called_graph), get_irg_bad(irg));
968 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(irg));
970 /* Initialize for compaction of in arrays */
971 inc_irg_block_visited(irg);
973 /* -- Replicate local entities of the called_graph -- */
974 /* copy the entities. */
975 called_frame = get_irg_frame_type(called_graph);
976 curr_frame = get_irg_frame_type(irg);
977 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
978 ir_entity *new_ent, *old_ent;
979 old_ent = get_class_member(called_frame, i);
980 new_ent = copy_entity_own(old_ent, curr_frame);
981 set_entity_link(old_ent, new_ent);
984 /* visited is > than that of called graph. With this trick visited will
985 remain unchanged so that an outer walker, e.g., searching the call nodes
986 to inline, calling this inline will not visit the inlined nodes. */
987 set_irg_visited(irg, get_irg_visited(irg)-1);
989 /* -- Performing dead node elimination inlines the graph -- */
990 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
992 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
993 get_irg_frame_type(called_graph));
995 /* Repair called_graph */
996 set_irg_visited(called_graph, get_irg_visited(irg));
997 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
998 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1000 /* -- Merge the end of the inlined procedure with the call site -- */
1001 /* We will turn the old Call node into a Tuple with the following
1004 0: Phi of all Memories of Return statements.
1005 1: Jmp from new Block that merges the control flow from all exception
1006 predecessors of the old end block.
1007 2: Tuple of all arguments.
1008 3: Phi of Exception memories.
1009 In case the old Call directly branches to End on an exception we don't
1010 need the block merging all exceptions nor the Phi of the exception
1014 /* -- Precompute some values -- */
1015 end_bl = get_new_node(get_irg_end_block(called_graph));
1016 end = get_new_node(get_irg_end(called_graph));
1017 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1018 n_res = get_method_n_ress(get_Call_type(call));
1020 res_pred = xmalloc(n_res * sizeof(*res_pred));
1021 cf_pred = xmalloc(arity * sizeof(*res_pred));
1023 set_irg_current_block(irg, post_bl); /* just to make sure */
1025 /* -- archive keepalives -- */
1026 irn_arity = get_irn_arity(end);
1027 for (i = 0; i < irn_arity; i++) {
1028 ir_node *ka = get_End_keepalive(end, i);
1030 add_End_keepalive(get_irg_end(irg), ka);
1033 /* The new end node will die. We need not free as the in array is on the obstack:
1034 copy_node() only generated 'D' arrays. */
1036 /* -- Replace Return nodes by Jump nodes. -- */
1038 for (i = 0; i < arity; i++) {
1040 ret = get_irn_n(end_bl, i);
1041 if (is_Return(ret)) {
1042 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1046 set_irn_in(post_bl, n_ret, cf_pred);
1048 /* -- Build a Tuple for all results of the method.
1049 Add Phi node if there was more than one Return. -- */
1050 turn_into_tuple(post_call, pn_Call_max);
1051 /* First the Memory-Phi */
1053 for (i = 0; i < arity; i++) {
1054 ret = get_irn_n(end_bl, i);
1055 if (is_Return(ret)) {
1056 cf_pred[n_ret] = get_Return_mem(ret);
1060 phi = new_Phi(n_ret, cf_pred, mode_M);
1061 set_Tuple_pred(call, pn_Call_M_regular, phi);
1062 /* Conserve Phi-list for further inlinings -- but might be optimized */
1063 if (get_nodes_block(phi) == post_bl) {
1064 set_irn_link(phi, get_irn_link(post_bl));
1065 set_irn_link(post_bl, phi);
1067 /* Now the real results */
1069 for (j = 0; j < n_res; j++) {
1071 for (i = 0; i < arity; i++) {
1072 ret = get_irn_n(end_bl, i);
1073 if (is_Return(ret)) {
1074 cf_pred[n_ret] = get_Return_res(ret, j);
1079 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1083 /* Conserve Phi-list for further inlinings -- but might be optimized */
1084 if (get_nodes_block(phi) == post_bl) {
1085 set_Phi_next(phi, get_Block_phis(post_bl));
1086 set_Block_phis(post_bl, phi);
1089 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1091 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1093 /* handle the regular call */
1094 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1096 /* For now, we cannot inline calls with value_base */
1097 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1099 /* Finally the exception control flow.
1100 We have two (three) possible situations:
1101 First if the Call branches to an exception handler: We need to add a Phi node to
1102 collect the memory containing the exception objects. Further we need
1103 to add another block to get a correct representation of this Phi. To
1104 this block we add a Jmp that resolves into the X output of the Call
1105 when the Call is turned into a tuple.
1106 Second the Call branches to End, the exception is not handled. Just
1107 add all inlined exception branches to the End node.
1108 Third: there is no Exception edge at all. Handle as case two. */
1109 if (exc_handling == exc_handler) {
1111 for (i = 0; i < arity; i++) {
1113 ret = get_irn_n(end_bl, i);
1114 irn = skip_Proj(ret);
1115 if (is_fragile_op(irn) || is_Raise(irn)) {
1116 cf_pred[n_exc] = ret;
1121 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1122 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1123 /* The Phi for the memories with the exception objects */
1125 for (i = 0; i < arity; i++) {
1127 ret = skip_Proj(get_irn_n(end_bl, i));
1129 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1131 } else if (is_fragile_op(ret)) {
1132 /* We rely that all cfops have the memory output at the same position. */
1133 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1135 } else if (is_Raise(ret)) {
1136 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1140 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1142 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1143 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1146 ir_node *main_end_bl;
1147 int main_end_bl_arity;
1148 ir_node **end_preds;
1150 /* assert(exc_handling == 1 || no exceptions. ) */
1152 for (i = 0; i < arity; i++) {
1153 ir_node *ret = get_irn_n(end_bl, i);
1154 ir_node *irn = skip_Proj(ret);
1156 if (is_fragile_op(irn) || is_Raise(irn)) {
1157 cf_pred[n_exc] = ret;
1161 main_end_bl = get_irg_end_block(irg);
1162 main_end_bl_arity = get_irn_arity(main_end_bl);
1163 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1165 for (i = 0; i < main_end_bl_arity; ++i)
1166 end_preds[i] = get_irn_n(main_end_bl, i);
1167 for (i = 0; i < n_exc; ++i)
1168 end_preds[main_end_bl_arity + i] = cf_pred[i];
1169 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1170 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1171 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1177 /* -- Turn CSE back on. -- */
1178 set_optimize(rem_opt);
1179 current_ir_graph = rem;
1184 /********************************************************************/
1185 /* Apply inlineing to small methods. */
1186 /********************************************************************/
1188 static struct obstack temp_obst;
1190 /** Represents a possible inlinable call in a graph. */
1191 typedef struct _call_entry call_entry;
1192 struct _call_entry {
1193 ir_node *call; /**< the Call */
1194 ir_graph *callee; /**< the callee called here */
1195 call_entry *next; /**< for linking the next one */
1196 unsigned weight; /**< the weight of the call */
1200 * environment for inlining small irgs
1202 typedef struct _inline_env_t {
1203 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1204 call_entry *head; /**< the head of the call entry list */
1205 call_entry *tail; /**< the tail of the call entry list */
1209 * Returns the irg called from a Call node. If the irg is not
1210 * known, NULL is returned.
1212 * @param call the call node
1214 static ir_graph *get_call_called_irg(ir_node *call) {
1217 addr = get_Call_ptr(call);
1218 if (is_Global(addr)) {
1219 ir_entity *ent = get_Global_entity(addr);
1220 return get_entity_irg(ent);
1227 * Walker: Collect all calls to known graphs inside a graph.
1229 static void collect_calls(ir_node *call, void *env) {
1230 if (is_Call(call)) {
1231 ir_graph *called_irg = get_call_called_irg(call);
1233 if (called_irg != NULL) {
1234 /* The Call node calls a locally defined method. Remember to inline. */
1235 inline_env_t *ienv = env;
1236 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1238 entry->callee = called_irg;
1242 if (ienv->tail == NULL)
1245 ienv->tail->next = entry;
1252 * Inlines all small methods at call sites where the called address comes
1253 * from a Const node that references the entity representing the called
1255 * The size argument is a rough measure for the code size of the method:
1256 * Methods where the obstack containing the firm graph is smaller than
1259 void inline_small_irgs(ir_graph *irg, int size) {
1260 ir_graph *rem = current_ir_graph;
1264 current_ir_graph = irg;
1265 /* Handle graph state */
1266 assert(get_irg_phase_state(irg) != phase_building);
1267 free_callee_info(irg);
1269 /* Find Call nodes to inline.
1270 (We can not inline during a walk of the graph, as inlineing the same
1271 method several times changes the visited flag of the walked graph:
1272 after the first inlineing visited of the callee equals visited of
1273 the caller. With the next inlineing both are increased.) */
1274 obstack_init(&env.obst);
1275 env.head = env.tail = NULL;
1276 irg_walk_graph(irg, NULL, collect_calls, &env);
1278 if (env.head != NULL) {
1279 /* There are calls to inline */
1280 collect_phiprojs(irg);
1281 for (entry = env.head; entry != NULL; entry = entry->next) {
1282 ir_graph *callee = entry->callee;
1283 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1284 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1285 inline_method(entry->call, callee);
1289 obstack_free(&env.obst, NULL);
1290 current_ir_graph = rem;
1294 * Environment for inlining irgs.
1297 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1298 int n_blocks; /**< Number of Blocks in graph without Start and End block. */
1299 int n_nodes_orig; /**< for statistics */
1300 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1301 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1302 int n_call_nodes; /**< Number of Call nodes in the graph. */
1303 int n_call_nodes_orig; /**< for statistics */
1304 int n_callers; /**< Number of known graphs that call this graphs. */
1305 int n_callers_orig; /**< for statistics */
1306 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1307 unsigned local_vars:1; /**< Set, if a inlined function gets the address of an inlined variable. */
1308 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1312 * Allocate a new environment for inlining.
1314 static inline_irg_env *alloc_inline_irg_env(void) {
1315 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1316 env->n_nodes = -2; /* do not count count Start, End */
1317 env->n_blocks = -2; /* do not count count Start, End Block */
1318 env->n_nodes_orig = -2; /* do not count Start, End */
1319 env->call_head = NULL;
1320 env->call_tail = NULL;
1321 env->n_call_nodes = 0;
1322 env->n_call_nodes_orig = 0;
1324 env->n_callers_orig = 0;
1325 env->got_inline = 0;
1326 env->local_vars = 0;
1327 env->local_weights = NULL;
1331 typedef struct walker_env {
1332 inline_irg_env *x; /**< the inline environment */
1333 char ignore_runtime; /**< the ignore runtime flag */
1334 char ignore_callers; /**< if set, do change callers data */
1338 * post-walker: collect all calls in the inline-environment
1339 * of a graph and sum some statistics.
1341 static void collect_calls2(ir_node *call, void *ctx) {
1343 inline_irg_env *x = env->x;
1344 ir_opcode code = get_irn_opcode(call);
1348 /* count meaningful nodes in irg */
1349 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1350 if (code != iro_Block) {
1358 if (code != iro_Call) return;
1360 /* check, if it's a runtime call */
1361 if (env->ignore_runtime) {
1362 ir_node *symc = get_Call_ptr(call);
1364 if (is_Global(symc)) {
1365 ir_entity *ent = get_Global_entity(symc);
1367 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1372 /* collect all call nodes */
1374 ++x->n_call_nodes_orig;
1376 callee = get_call_called_irg(call);
1377 if (callee != NULL) {
1378 if (! env->ignore_callers) {
1379 inline_irg_env *callee_env = get_irg_link(callee);
1380 /* count all static callers */
1381 ++callee_env->n_callers;
1382 ++callee_env->n_callers_orig;
1385 /* link it in the list of possible inlinable entries */
1386 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1388 entry->callee = callee;
1390 if (x->call_tail == NULL)
1391 x->call_head = entry;
1393 x->call_tail->next = entry;
1394 x->call_tail = entry;
1399 * Returns TRUE if the number of callers is 0 in the irg's environment,
1400 * hence this irg is a leave.
1402 INLINE static int is_leave(ir_graph *irg) {
1403 inline_irg_env *env = get_irg_link(irg);
1404 return env->n_call_nodes == 0;
1408 * Returns TRUE if the number of nodes in the callee is
1409 * smaller then size in the irg's environment.
1411 INLINE static int is_smaller(ir_graph *callee, int size) {
1412 inline_irg_env *env = get_irg_link(callee);
1413 return env->n_nodes < size;
1417 * Append the nodes of the list src to the nodes of the list in environment dst.
1419 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1420 call_entry *entry, *nentry;
1422 /* Note that the src list points to Call nodes in the inlined graph, but
1423 we need Call nodes in our graph. Luckily the inliner leaves this information
1424 in the link field. */
1425 for (entry = src; entry != NULL; entry = entry->next) {
1426 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1427 nentry->call = get_irn_link(entry->call);
1428 nentry->callee = entry->callee;
1429 nentry->next = NULL;
1430 dst->call_tail->next = nentry;
1431 dst->call_tail = nentry;
1436 * Inlines small leave methods at call sites where the called address comes
1437 * from a Const node that references the entity representing the called
1439 * The size argument is a rough measure for the code size of the method:
1440 * Methods where the obstack containing the firm graph is smaller than
1443 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1444 inline_irg_env *env;
1450 call_entry *entry, *tail;
1451 const call_entry *centry;
1452 pmap *copied_graphs;
1453 pmap_entry *pm_entry;
1455 rem = current_ir_graph;
1456 obstack_init(&temp_obst);
1458 /* a map for the copied graphs, used to inline recursive calls */
1459 copied_graphs = pmap_create();
1461 /* extend all irgs by a temporary data structure for inlining. */
1462 n_irgs = get_irp_n_irgs();
1463 for (i = 0; i < n_irgs; ++i)
1464 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1466 /* Precompute information in temporary data structure. */
1467 wenv.ignore_runtime = ignore_runtime;
1468 wenv.ignore_callers = 0;
1469 for (i = 0; i < n_irgs; ++i) {
1470 ir_graph *irg = get_irp_irg(i);
1472 assert(get_irg_phase_state(irg) != phase_building);
1473 free_callee_info(irg);
1475 wenv.x = get_irg_link(irg);
1476 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1479 /* -- and now inline. -- */
1481 /* Inline leaves recursively -- we might construct new leaves. */
1485 for (i = 0; i < n_irgs; ++i) {
1487 int phiproj_computed = 0;
1489 current_ir_graph = get_irp_irg(i);
1490 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1493 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1496 if (env->n_nodes > maxsize) break;
1499 callee = entry->callee;
1501 if (is_leave(callee) && (
1502 is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1503 if (!phiproj_computed) {
1504 phiproj_computed = 1;
1505 collect_phiprojs(current_ir_graph);
1507 did_inline = inline_method(call, callee);
1510 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1512 /* was inlined, must be recomputed */
1513 phiproj_computed = 0;
1515 /* Do some statistics */
1516 env->got_inline = 1;
1517 --env->n_call_nodes;
1518 env->n_nodes += callee_env->n_nodes;
1519 --callee_env->n_callers;
1521 /* remove this call from the list */
1523 tail->next = entry->next;
1525 env->call_head = entry->next;
1531 env->call_tail = tail;
1533 } while (did_inline);
1535 /* inline other small functions. */
1536 for (i = 0; i < n_irgs; ++i) {
1538 int phiproj_computed = 0;
1540 current_ir_graph = get_irp_irg(i);
1541 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1543 /* note that the list of possible calls is updated during the process */
1545 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1550 callee = entry->callee;
1552 e = pmap_find(copied_graphs, callee);
1555 * Remap callee if we have a copy.
1556 * FIXME: Should we do this only for recursive Calls ?
1561 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1562 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1563 if (current_ir_graph == callee) {
1565 * Recursive call: we cannot directly inline because we cannot walk
1566 * the graph and change it. So we have to make a copy of the graph
1570 inline_irg_env *callee_env;
1574 * No copy yet, create one.
1575 * Note that recursive methods are never leaves, so it is sufficient
1576 * to test this condition here.
1578 copy = create_irg_copy(callee);
1580 /* create_irg_copy() destroys the Proj links, recompute them */
1581 phiproj_computed = 0;
1583 /* allocate new environment */
1584 callee_env = alloc_inline_irg_env();
1585 set_irg_link(copy, callee_env);
1587 wenv.x = callee_env;
1588 wenv.ignore_callers = 1;
1589 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1592 * Enter the entity of the original graph. This is needed
1593 * for inline_method(). However, note that ent->irg still points
1594 * to callee, NOT to copy.
1596 set_irg_entity(copy, get_irg_entity(callee));
1598 pmap_insert(copied_graphs, callee, copy);
1601 /* we have only one caller: the original graph */
1602 callee_env->n_callers = 1;
1603 callee_env->n_callers_orig = 1;
1605 if (! phiproj_computed) {
1606 phiproj_computed = 1;
1607 collect_phiprojs(current_ir_graph);
1609 did_inline = inline_method(call, callee);
1611 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1613 /* was inlined, must be recomputed */
1614 phiproj_computed = 0;
1616 /* callee was inline. Append it's call list. */
1617 env->got_inline = 1;
1618 --env->n_call_nodes;
1619 append_call_list(env, callee_env->call_head);
1620 env->n_call_nodes += callee_env->n_call_nodes;
1621 env->n_nodes += callee_env->n_nodes;
1622 --callee_env->n_callers;
1624 /* after we have inlined callee, all called methods inside callee
1625 are now called once more */
1626 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1627 inline_irg_env *penv = get_irg_link(centry->callee);
1631 /* remove this call from the list */
1633 tail->next = entry->next;
1635 env->call_head = entry->next;
1641 env->call_tail = tail;
1644 for (i = 0; i < n_irgs; ++i) {
1645 irg = get_irp_irg(i);
1646 env = (inline_irg_env *)get_irg_link(irg);
1648 if (env->got_inline) {
1649 optimize_graph_df(irg);
1652 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1653 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1654 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1655 env->n_callers_orig, env->n_callers,
1656 get_entity_name(get_irg_entity(irg))));
1660 /* kill the copied graphs: we don't need them anymore */
1661 foreach_pmap(copied_graphs, pm_entry) {
1662 ir_graph *copy = pm_entry->value;
1664 /* reset the entity, otherwise it will be deleted in the next step ... */
1665 set_irg_entity(copy, NULL);
1666 free_ir_graph(copy);
1668 pmap_destroy(copied_graphs);
1670 obstack_free(&temp_obst, NULL);
1671 current_ir_graph = rem;
1675 * Calculate the parameter weights for transmitting the address of a local variable.
1677 static unsigned calc_method_local_weight(ir_node *arg) {
1679 unsigned v, weight = 0;
1681 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1682 ir_node *succ = get_irn_out(arg, i);
1684 switch (get_irn_opcode(succ)) {
1687 /* Loads and Store can be removed */
1691 /* check if all args are constant */
1692 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1693 ir_node *idx = get_Sel_index(succ, j);
1694 if (! is_Const(idx))
1697 /* Check users on this Sel. Note: if a 0 is returned here, there was
1698 some unsupported node. */
1699 v = calc_method_local_weight(succ);
1702 /* we can kill one Sel with constant indexes, this is cheap */
1706 /* when looking backward we might find Id nodes */
1707 weight += calc_method_local_weight(succ);
1710 /* unoptimized tuple */
1711 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1712 ir_node *pred = get_Tuple_pred(succ, j);
1714 /* look for Proj(j) */
1715 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1716 ir_node *succ_succ = get_irn_out(succ, k);
1717 if (is_Proj(succ_succ)) {
1718 if (get_Proj_proj(succ_succ) == j) {
1720 weight += calc_method_local_weight(succ_succ);
1723 /* this should NOT happen */
1730 /* any other node: unsupported yet or bad. */
1738 * Calculate the parameter weights for transmitting the address of a local variable.
1740 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1741 ir_entity *ent = get_irg_entity(irg);
1743 int nparams, i, proj_nr;
1744 ir_node *irg_args, *arg;
1746 mtp = get_entity_type(ent);
1747 nparams = get_method_n_params(mtp);
1749 /* allocate a new array. currently used as 'analysed' flag */
1750 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1752 /* If the method haven't parameters we have nothing to do. */
1756 assure_irg_outs(irg);
1757 irg_args = get_irg_args(irg);
1758 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1759 arg = get_irn_out(irg_args, i);
1760 proj_nr = get_Proj_proj(arg);
1761 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1766 * Calculate the benefice for transmitting an local variable address.
1767 * After inlining, the local variable might be transformed into a
1768 * SSA variable by scalar_replacement().
1770 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1771 inline_irg_env *env = get_irg_link(callee);
1773 if (env->local_weights != NULL) {
1774 if (pos < ARR_LEN(env->local_weights))
1775 return env->local_weights[pos];
1779 analyze_irg_local_weights(env, callee);
1781 if (pos < ARR_LEN(env->local_weights))
1782 return env->local_weights[pos];
1787 * calculate a benefice value for inlining the given call.
1789 static int calc_inline_benefice(ir_node *call, ir_graph *callee, unsigned *local_adr) {
1790 ir_entity *ent = get_irg_entity(callee);
1797 inline_irg_env *curr_env, *callee_env;
1799 if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1800 /* do NOT inline noreturn calls */
1804 /* costs for every passed parameter */
1805 n_params = get_Call_n_params(call);
1806 mtp = get_entity_type(ent);
1807 cc = get_method_calling_convention(mtp);
1808 if (cc & cc_reg_param) {
1809 /* register parameter, smaller costs for register parameters */
1810 int max_regs = cc & ~cc_bits;
1812 if (max_regs < n_params)
1813 weight += max_regs * 2 + (n_params - max_regs) * 5;
1815 weight += n_params * 2;
1817 /* parameters are passed an stack */
1818 weight += 5 * n_params;
1821 /* constant parameters improve the benefice */
1822 frame_ptr = get_irg_frame(current_ir_graph);
1823 for (i = 0; i < n_params; ++i) {
1824 ir_node *param = get_Call_param(call, i);
1826 if (is_Const(param) || is_SymConst(param))
1827 weight += get_method_param_weight(ent, i);
1828 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1830 * An address of a local variable is transmitted. After inlining,
1831 * scalar_replacement might be able to remove the local variable,
1834 v = get_method_local_adress_weight(callee, i);
1841 callee_env = get_irg_link(callee);
1842 if (get_entity_visibility(ent) == visibility_local &&
1843 callee_env->n_callers_orig == 1 &&
1844 callee != current_ir_graph) {
1845 /* we are the only caller, give big bonus */
1849 /* do not inline big functions */
1850 weight -= callee_env->n_nodes;
1852 /* reduce the benefice if the current function is already big */
1853 curr_env = get_irg_link(current_ir_graph);
1854 weight -= curr_env->n_nodes / 100;
1856 /* give a bonus for functions with one block */
1857 if (callee_env->n_blocks == 1)
1858 weight = weight * 3 / 2;
1864 * Heuristic inliner. Calculates a benifice value for every call and inlines
1865 * those calls with a value higher than the threshold.
1867 void inline_functions(int inline_threshold) {
1868 inline_irg_env *env;
1873 call_entry *entry, *tail;
1874 const call_entry *centry;
1875 pmap *copied_graphs;
1876 pmap_entry *pm_entry;
1878 rem = current_ir_graph;
1879 obstack_init(&temp_obst);
1881 /* a map for the copied graphs, used to inline recursive calls */
1882 copied_graphs = pmap_create();
1884 /* extend all irgs by a temporary data structure for inlining. */
1885 n_irgs = get_irp_n_irgs();
1886 for (i = 0; i < n_irgs; ++i)
1887 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1889 /* Precompute information in temporary data structure. */
1890 wenv.ignore_runtime = 0;
1891 wenv.ignore_callers = 0;
1892 for (i = 0; i < n_irgs; ++i) {
1893 ir_graph *irg = get_irp_irg(i);
1895 assert(get_irg_phase_state(irg) != phase_building);
1896 free_callee_info(irg);
1898 wenv.x = get_irg_link(irg);
1899 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1902 /* -- and now inline. -- */
1903 for (i = 0; i < n_irgs; ++i) {
1904 int phiproj_computed = 0;
1906 ir_graph *irg = get_irp_irg(i);
1908 current_ir_graph = irg;
1909 env = get_irg_link(irg);
1911 /* note that the list of possible calls is updated during the process */
1913 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1920 callee = entry->callee;
1922 /* calculate the benifice on the original call to prevent excessive inlining */
1924 benefice = calc_inline_benefice(call, callee, &local_adr);
1925 DB((dbg, LEVEL_2, "In %+F Call %+F has benefice %d\n", irg, callee, benefice));
1927 e = pmap_find(copied_graphs, callee);
1930 * Remap callee if we have a copy.
1931 * FIXME: Should we do this only for recursive Calls ?
1936 if (benefice > -inline_threshold ||
1937 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1938 if (current_ir_graph == callee) {
1940 * Recursive call: we cannot directly inline because we cannot walk
1941 * the graph and change it. So we have to make a copy of the graph
1945 inline_irg_env *callee_env;
1949 * No copy yet, create one.
1950 * Note that recursive methods are never leaves, so it is sufficient
1951 * to test this condition here.
1953 copy = create_irg_copy(callee);
1955 /* create_irg_copy() destroys the Proj links, recompute them */
1956 phiproj_computed = 0;
1958 /* allocate new environment */
1959 callee_env = alloc_inline_irg_env();
1960 set_irg_link(copy, callee_env);
1962 wenv.x = callee_env;
1963 wenv.ignore_callers = 1;
1964 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1967 * Enter the entity of the original graph. This is needed
1968 * for inline_method(). However, note that ent->irg still points
1969 * to callee, NOT to copy.
1971 set_irg_entity(copy, get_irg_entity(callee));
1973 pmap_insert(copied_graphs, callee, copy);
1976 /* we have only one caller: the original graph */
1977 callee_env->n_callers = 1;
1978 callee_env->n_callers_orig = 1;
1980 if (! phiproj_computed) {
1981 phiproj_computed = 1;
1982 collect_phiprojs(current_ir_graph);
1984 did_inline = inline_method(call, callee);
1986 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1988 /* was inlined, must be recomputed */
1989 phiproj_computed = 0;
1991 /* callee was inline. Append it's call list. */
1992 env->got_inline = 1;
1994 env->local_vars = 1;
1995 --env->n_call_nodes;
1996 append_call_list(env, callee_env->call_head);
1997 env->n_call_nodes += callee_env->n_call_nodes;
1998 env->n_nodes += callee_env->n_nodes;
1999 --callee_env->n_callers;
2001 /* after we have inlined callee, all called methods inside callee
2002 are now called once more */
2003 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
2004 inline_irg_env *penv = get_irg_link(centry->callee);
2008 /* remove this call from the list */
2010 tail->next = entry->next;
2012 env->call_head = entry->next;
2018 env->call_tail = tail;
2020 if (env->got_inline) {
2021 /* this irg got calls inlined: optimize it */
2023 /* scalar replacement does not work well with Tuple nodes, so optimize them away */
2024 optimize_graph_df(irg);
2026 if (env->local_vars) {
2027 if (scalar_replacement_opt(irg)) {
2028 optimize_graph_df(irg);
2034 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2035 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2036 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2037 env->n_callers_orig, env->n_callers,
2038 get_entity_name(get_irg_entity(irg))));
2042 /* kill the copied graphs: we don't need them anymore */
2043 foreach_pmap(copied_graphs, pm_entry) {
2044 ir_graph *copy = pm_entry->value;
2046 /* reset the entity, otherwise it will be deleted in the next step ... */
2047 set_irg_entity(copy, NULL);
2048 free_ir_graph(copy);
2050 pmap_destroy(copied_graphs);
2052 obstack_free(&temp_obst, NULL);
2053 current_ir_graph = rem;
2056 void firm_init_inline(void) {
2057 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");