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_globals] = get_irg_globals(irg);
940 in[pn_Start_P_tls] = get_irg_tls(irg);
941 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
942 /* in[pn_Start_P_value_arg_base] = ??? */
943 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
944 pre_call = new_Tuple(pn_Start_max - 1, in);
948 The new block gets the ins of the old block, pre_call and all its
949 predecessors and all Phi nodes. -- */
950 part_block(pre_call);
952 /* -- Prepare state for dead node elimination -- */
953 /* Visited flags in calling irg must be >= flag in called irg.
954 Else walker and arity computation will not work. */
955 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
956 set_irg_visited(irg, get_irg_visited(called_graph)+1);
957 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
958 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
959 /* Set pre_call as new Start node in link field of the start node of
960 calling graph and pre_calls block as new block for the start block
962 Further mark these nodes so that they are not visited by the
964 set_irn_link(get_irg_start(called_graph), pre_call);
965 set_irn_visited(get_irg_start(called_graph), get_irg_visited(irg));
966 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
967 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(irg));
968 set_irn_link(get_irg_bad(called_graph), get_irg_bad(irg));
969 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(irg));
971 /* Initialize for compaction of in arrays */
972 inc_irg_block_visited(irg);
974 /* -- Replicate local entities of the called_graph -- */
975 /* copy the entities. */
976 called_frame = get_irg_frame_type(called_graph);
977 curr_frame = get_irg_frame_type(irg);
978 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
979 ir_entity *new_ent, *old_ent;
980 old_ent = get_class_member(called_frame, i);
981 new_ent = copy_entity_own(old_ent, curr_frame);
982 set_entity_link(old_ent, new_ent);
985 /* visited is > than that of called graph. With this trick visited will
986 remain unchanged so that an outer walker, e.g., searching the call nodes
987 to inline, calling this inline will not visit the inlined nodes. */
988 set_irg_visited(irg, get_irg_visited(irg)-1);
990 /* -- Performing dead node elimination inlines the graph -- */
991 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
993 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
994 get_irg_frame_type(called_graph));
996 /* Repair called_graph */
997 set_irg_visited(called_graph, get_irg_visited(irg));
998 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
999 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1001 /* -- Merge the end of the inlined procedure with the call site -- */
1002 /* We will turn the old Call node into a Tuple with the following
1005 0: Phi of all Memories of Return statements.
1006 1: Jmp from new Block that merges the control flow from all exception
1007 predecessors of the old end block.
1008 2: Tuple of all arguments.
1009 3: Phi of Exception memories.
1010 In case the old Call directly branches to End on an exception we don't
1011 need the block merging all exceptions nor the Phi of the exception
1015 /* -- Precompute some values -- */
1016 end_bl = get_new_node(get_irg_end_block(called_graph));
1017 end = get_new_node(get_irg_end(called_graph));
1018 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
1019 n_res = get_method_n_ress(get_Call_type(call));
1021 res_pred = xmalloc(n_res * sizeof(*res_pred));
1022 cf_pred = xmalloc(arity * sizeof(*res_pred));
1024 set_irg_current_block(irg, post_bl); /* just to make sure */
1026 /* -- archive keepalives -- */
1027 irn_arity = get_irn_arity(end);
1028 for (i = 0; i < irn_arity; i++) {
1029 ir_node *ka = get_End_keepalive(end, i);
1031 add_End_keepalive(get_irg_end(irg), ka);
1034 /* The new end node will die. We need not free as the in array is on the obstack:
1035 copy_node() only generated 'D' arrays. */
1037 /* -- Replace Return nodes by Jump nodes. -- */
1039 for (i = 0; i < arity; i++) {
1041 ret = get_irn_n(end_bl, i);
1042 if (is_Return(ret)) {
1043 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1047 set_irn_in(post_bl, n_ret, cf_pred);
1049 /* -- Build a Tuple for all results of the method.
1050 Add Phi node if there was more than one Return. -- */
1051 turn_into_tuple(post_call, pn_Call_max);
1052 /* First the Memory-Phi */
1054 for (i = 0; i < arity; i++) {
1055 ret = get_irn_n(end_bl, i);
1056 if (is_Return(ret)) {
1057 cf_pred[n_ret] = get_Return_mem(ret);
1061 phi = new_Phi(n_ret, cf_pred, mode_M);
1062 set_Tuple_pred(call, pn_Call_M_regular, phi);
1063 /* Conserve Phi-list for further inlinings -- but might be optimized */
1064 if (get_nodes_block(phi) == post_bl) {
1065 set_irn_link(phi, get_irn_link(post_bl));
1066 set_irn_link(post_bl, phi);
1068 /* Now the real results */
1070 for (j = 0; j < n_res; j++) {
1072 for (i = 0; i < arity; i++) {
1073 ret = get_irn_n(end_bl, i);
1074 if (is_Return(ret)) {
1075 cf_pred[n_ret] = get_Return_res(ret, j);
1080 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1084 /* Conserve Phi-list for further inlinings -- but might be optimized */
1085 if (get_nodes_block(phi) == post_bl) {
1086 set_Phi_next(phi, get_Block_phis(post_bl));
1087 set_Block_phis(post_bl, phi);
1090 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1092 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1094 /* handle the regular call */
1095 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1097 /* For now, we cannot inline calls with value_base */
1098 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1100 /* Finally the exception control flow.
1101 We have two (three) possible situations:
1102 First if the Call branches to an exception handler: We need to add a Phi node to
1103 collect the memory containing the exception objects. Further we need
1104 to add another block to get a correct representation of this Phi. To
1105 this block we add a Jmp that resolves into the X output of the Call
1106 when the Call is turned into a tuple.
1107 Second the Call branches to End, the exception is not handled. Just
1108 add all inlined exception branches to the End node.
1109 Third: there is no Exception edge at all. Handle as case two. */
1110 if (exc_handling == exc_handler) {
1112 for (i = 0; i < arity; i++) {
1114 ret = get_irn_n(end_bl, i);
1115 irn = skip_Proj(ret);
1116 if (is_fragile_op(irn) || is_Raise(irn)) {
1117 cf_pred[n_exc] = ret;
1122 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1123 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1124 /* The Phi for the memories with the exception objects */
1126 for (i = 0; i < arity; i++) {
1128 ret = skip_Proj(get_irn_n(end_bl, i));
1130 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1132 } else if (is_fragile_op(ret)) {
1133 /* We rely that all cfops have the memory output at the same position. */
1134 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1136 } else if (is_Raise(ret)) {
1137 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1141 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1143 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1144 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1147 ir_node *main_end_bl;
1148 int main_end_bl_arity;
1149 ir_node **end_preds;
1151 /* assert(exc_handling == 1 || no exceptions. ) */
1153 for (i = 0; i < arity; i++) {
1154 ir_node *ret = get_irn_n(end_bl, i);
1155 ir_node *irn = skip_Proj(ret);
1157 if (is_fragile_op(irn) || is_Raise(irn)) {
1158 cf_pred[n_exc] = ret;
1162 main_end_bl = get_irg_end_block(irg);
1163 main_end_bl_arity = get_irn_arity(main_end_bl);
1164 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1166 for (i = 0; i < main_end_bl_arity; ++i)
1167 end_preds[i] = get_irn_n(main_end_bl, i);
1168 for (i = 0; i < n_exc; ++i)
1169 end_preds[main_end_bl_arity + i] = cf_pred[i];
1170 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1171 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1172 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1178 /* -- Turn CSE back on. -- */
1179 set_optimize(rem_opt);
1180 current_ir_graph = rem;
1185 /********************************************************************/
1186 /* Apply inlineing to small methods. */
1187 /********************************************************************/
1189 static struct obstack temp_obst;
1191 /** Represents a possible inlinable call in a graph. */
1192 typedef struct _call_entry call_entry;
1193 struct _call_entry {
1194 ir_node *call; /**< the Call */
1195 ir_graph *callee; /**< the callee called here */
1196 call_entry *next; /**< for linking the next one */
1197 unsigned weight; /**< the weight of the call */
1201 * environment for inlining small irgs
1203 typedef struct _inline_env_t {
1204 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1205 call_entry *head; /**< the head of the call entry list */
1206 call_entry *tail; /**< the tail of the call entry list */
1210 * Returns the irg called from a Call node. If the irg is not
1211 * known, NULL is returned.
1213 * @param call the call node
1215 static ir_graph *get_call_called_irg(ir_node *call) {
1218 addr = get_Call_ptr(call);
1219 if (is_Global(addr)) {
1220 ir_entity *ent = get_Global_entity(addr);
1221 return get_entity_irg(ent);
1228 * Walker: Collect all calls to known graphs inside a graph.
1230 static void collect_calls(ir_node *call, void *env) {
1231 if (is_Call(call)) {
1232 ir_graph *called_irg = get_call_called_irg(call);
1234 if (called_irg != NULL) {
1235 /* The Call node calls a locally defined method. Remember to inline. */
1236 inline_env_t *ienv = env;
1237 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1239 entry->callee = called_irg;
1243 if (ienv->tail == NULL)
1246 ienv->tail->next = entry;
1253 * Inlines all small methods at call sites where the called address comes
1254 * from a Const node that references the entity representing the called
1256 * The size argument is a rough measure for the code size of the method:
1257 * Methods where the obstack containing the firm graph is smaller than
1260 void inline_small_irgs(ir_graph *irg, int size) {
1261 ir_graph *rem = current_ir_graph;
1265 current_ir_graph = irg;
1266 /* Handle graph state */
1267 assert(get_irg_phase_state(irg) != phase_building);
1268 free_callee_info(irg);
1270 /* Find Call nodes to inline.
1271 (We can not inline during a walk of the graph, as inlineing the same
1272 method several times changes the visited flag of the walked graph:
1273 after the first inlineing visited of the callee equals visited of
1274 the caller. With the next inlineing both are increased.) */
1275 obstack_init(&env.obst);
1276 env.head = env.tail = NULL;
1277 irg_walk_graph(irg, NULL, collect_calls, &env);
1279 if (env.head != NULL) {
1280 /* There are calls to inline */
1281 collect_phiprojs(irg);
1282 for (entry = env.head; entry != NULL; entry = entry->next) {
1283 ir_graph *callee = entry->callee;
1284 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1285 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1286 inline_method(entry->call, callee);
1290 obstack_free(&env.obst, NULL);
1291 current_ir_graph = rem;
1295 * Environment for inlining irgs.
1298 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1299 int n_blocks; /**< Number of Blocks in graph without Start and End block. */
1300 int n_nodes_orig; /**< for statistics */
1301 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1302 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1303 int n_call_nodes; /**< Number of Call nodes in the graph. */
1304 int n_call_nodes_orig; /**< for statistics */
1305 int n_callers; /**< Number of known graphs that call this graphs. */
1306 int n_callers_orig; /**< for statistics */
1307 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1308 unsigned local_vars:1; /**< Set, if a inlined function gets the address of an inlined variable. */
1309 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1313 * Allocate a new environment for inlining.
1315 static inline_irg_env *alloc_inline_irg_env(void) {
1316 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1317 env->n_nodes = -2; /* do not count count Start, End */
1318 env->n_blocks = -2; /* do not count count Start, End Block */
1319 env->n_nodes_orig = -2; /* do not count Start, End */
1320 env->call_head = NULL;
1321 env->call_tail = NULL;
1322 env->n_call_nodes = 0;
1323 env->n_call_nodes_orig = 0;
1325 env->n_callers_orig = 0;
1326 env->got_inline = 0;
1327 env->local_vars = 0;
1328 env->local_weights = NULL;
1332 typedef struct walker_env {
1333 inline_irg_env *x; /**< the inline environment */
1334 char ignore_runtime; /**< the ignore runtime flag */
1335 char ignore_callers; /**< if set, do change callers data */
1339 * post-walker: collect all calls in the inline-environment
1340 * of a graph and sum some statistics.
1342 static void collect_calls2(ir_node *call, void *ctx) {
1344 inline_irg_env *x = env->x;
1345 ir_opcode code = get_irn_opcode(call);
1349 /* count meaningful nodes in irg */
1350 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1351 if (code != iro_Block) {
1359 if (code != iro_Call) return;
1361 /* check, if it's a runtime call */
1362 if (env->ignore_runtime) {
1363 ir_node *symc = get_Call_ptr(call);
1365 if (is_Global(symc)) {
1366 ir_entity *ent = get_Global_entity(symc);
1368 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1373 /* collect all call nodes */
1375 ++x->n_call_nodes_orig;
1377 callee = get_call_called_irg(call);
1378 if (callee != NULL) {
1379 if (! env->ignore_callers) {
1380 inline_irg_env *callee_env = get_irg_link(callee);
1381 /* count all static callers */
1382 ++callee_env->n_callers;
1383 ++callee_env->n_callers_orig;
1386 /* link it in the list of possible inlinable entries */
1387 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1389 entry->callee = callee;
1391 if (x->call_tail == NULL)
1392 x->call_head = entry;
1394 x->call_tail->next = entry;
1395 x->call_tail = entry;
1400 * Returns TRUE if the number of callers is 0 in the irg's environment,
1401 * hence this irg is a leave.
1403 INLINE static int is_leave(ir_graph *irg) {
1404 inline_irg_env *env = get_irg_link(irg);
1405 return env->n_call_nodes == 0;
1409 * Returns TRUE if the number of nodes in the callee is
1410 * smaller then size in the irg's environment.
1412 INLINE static int is_smaller(ir_graph *callee, int size) {
1413 inline_irg_env *env = get_irg_link(callee);
1414 return env->n_nodes < size;
1418 * Append the nodes of the list src to the nodes of the list in environment dst.
1420 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1421 call_entry *entry, *nentry;
1423 /* Note that the src list points to Call nodes in the inlined graph, but
1424 we need Call nodes in our graph. Luckily the inliner leaves this information
1425 in the link field. */
1426 for (entry = src; entry != NULL; entry = entry->next) {
1427 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1428 nentry->call = get_irn_link(entry->call);
1429 nentry->callee = entry->callee;
1430 nentry->next = NULL;
1431 dst->call_tail->next = nentry;
1432 dst->call_tail = nentry;
1437 * Inlines small leave methods at call sites where the called address comes
1438 * from a Const node that references the entity representing the called
1440 * The size argument is a rough measure for the code size of the method:
1441 * Methods where the obstack containing the firm graph is smaller than
1444 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1445 inline_irg_env *env;
1451 call_entry *entry, *tail;
1452 const call_entry *centry;
1453 pmap *copied_graphs;
1454 pmap_entry *pm_entry;
1456 rem = current_ir_graph;
1457 obstack_init(&temp_obst);
1459 /* a map for the copied graphs, used to inline recursive calls */
1460 copied_graphs = pmap_create();
1462 /* extend all irgs by a temporary data structure for inlining. */
1463 n_irgs = get_irp_n_irgs();
1464 for (i = 0; i < n_irgs; ++i)
1465 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1467 /* Precompute information in temporary data structure. */
1468 wenv.ignore_runtime = ignore_runtime;
1469 wenv.ignore_callers = 0;
1470 for (i = 0; i < n_irgs; ++i) {
1471 ir_graph *irg = get_irp_irg(i);
1473 assert(get_irg_phase_state(irg) != phase_building);
1474 free_callee_info(irg);
1476 wenv.x = get_irg_link(irg);
1477 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1480 /* -- and now inline. -- */
1482 /* Inline leaves recursively -- we might construct new leaves. */
1486 for (i = 0; i < n_irgs; ++i) {
1488 int phiproj_computed = 0;
1490 current_ir_graph = get_irp_irg(i);
1491 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1494 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1497 if (env->n_nodes > maxsize) break;
1500 callee = entry->callee;
1502 if (is_leave(callee) && (
1503 is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1504 if (!phiproj_computed) {
1505 phiproj_computed = 1;
1506 collect_phiprojs(current_ir_graph);
1508 did_inline = inline_method(call, callee);
1511 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1513 /* was inlined, must be recomputed */
1514 phiproj_computed = 0;
1516 /* Do some statistics */
1517 env->got_inline = 1;
1518 --env->n_call_nodes;
1519 env->n_nodes += callee_env->n_nodes;
1520 --callee_env->n_callers;
1522 /* remove this call from the list */
1524 tail->next = entry->next;
1526 env->call_head = entry->next;
1532 env->call_tail = tail;
1534 } while (did_inline);
1536 /* inline other small functions. */
1537 for (i = 0; i < n_irgs; ++i) {
1539 int phiproj_computed = 0;
1541 current_ir_graph = get_irp_irg(i);
1542 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1544 /* note that the list of possible calls is updated during the process */
1546 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1551 callee = entry->callee;
1553 e = pmap_find(copied_graphs, callee);
1556 * Remap callee if we have a copy.
1557 * FIXME: Should we do this only for recursive Calls ?
1562 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1563 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1564 if (current_ir_graph == callee) {
1566 * Recursive call: we cannot directly inline because we cannot walk
1567 * the graph and change it. So we have to make a copy of the graph
1571 inline_irg_env *callee_env;
1575 * No copy yet, create one.
1576 * Note that recursive methods are never leaves, so it is sufficient
1577 * to test this condition here.
1579 copy = create_irg_copy(callee);
1581 /* create_irg_copy() destroys the Proj links, recompute them */
1582 phiproj_computed = 0;
1584 /* allocate new environment */
1585 callee_env = alloc_inline_irg_env();
1586 set_irg_link(copy, callee_env);
1588 wenv.x = callee_env;
1589 wenv.ignore_callers = 1;
1590 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1593 * Enter the entity of the original graph. This is needed
1594 * for inline_method(). However, note that ent->irg still points
1595 * to callee, NOT to copy.
1597 set_irg_entity(copy, get_irg_entity(callee));
1599 pmap_insert(copied_graphs, callee, copy);
1602 /* we have only one caller: the original graph */
1603 callee_env->n_callers = 1;
1604 callee_env->n_callers_orig = 1;
1606 if (! phiproj_computed) {
1607 phiproj_computed = 1;
1608 collect_phiprojs(current_ir_graph);
1610 did_inline = inline_method(call, callee);
1612 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1614 /* was inlined, must be recomputed */
1615 phiproj_computed = 0;
1617 /* callee was inline. Append it's call list. */
1618 env->got_inline = 1;
1619 --env->n_call_nodes;
1620 append_call_list(env, callee_env->call_head);
1621 env->n_call_nodes += callee_env->n_call_nodes;
1622 env->n_nodes += callee_env->n_nodes;
1623 --callee_env->n_callers;
1625 /* after we have inlined callee, all called methods inside callee
1626 are now called once more */
1627 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1628 inline_irg_env *penv = get_irg_link(centry->callee);
1632 /* remove this call from the list */
1634 tail->next = entry->next;
1636 env->call_head = entry->next;
1642 env->call_tail = tail;
1645 for (i = 0; i < n_irgs; ++i) {
1646 irg = get_irp_irg(i);
1647 env = (inline_irg_env *)get_irg_link(irg);
1649 if (env->got_inline) {
1650 optimize_graph_df(irg);
1653 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1654 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1655 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1656 env->n_callers_orig, env->n_callers,
1657 get_entity_name(get_irg_entity(irg))));
1661 /* kill the copied graphs: we don't need them anymore */
1662 foreach_pmap(copied_graphs, pm_entry) {
1663 ir_graph *copy = pm_entry->value;
1665 /* reset the entity, otherwise it will be deleted in the next step ... */
1666 set_irg_entity(copy, NULL);
1667 free_ir_graph(copy);
1669 pmap_destroy(copied_graphs);
1671 obstack_free(&temp_obst, NULL);
1672 current_ir_graph = rem;
1676 * Calculate the parameter weights for transmitting the address of a local variable.
1678 static unsigned calc_method_local_weight(ir_node *arg) {
1680 unsigned v, weight = 0;
1682 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1683 ir_node *succ = get_irn_out(arg, i);
1685 switch (get_irn_opcode(succ)) {
1688 /* Loads and Store can be removed */
1692 /* check if all args are constant */
1693 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1694 ir_node *idx = get_Sel_index(succ, j);
1695 if (! is_Const(idx))
1698 /* Check users on this Sel. Note: if a 0 is returned here, there was
1699 some unsupported node. */
1700 v = calc_method_local_weight(succ);
1703 /* we can kill one Sel with constant indexes, this is cheap */
1707 /* when looking backward we might find Id nodes */
1708 weight += calc_method_local_weight(succ);
1711 /* unoptimized tuple */
1712 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1713 ir_node *pred = get_Tuple_pred(succ, j);
1715 /* look for Proj(j) */
1716 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1717 ir_node *succ_succ = get_irn_out(succ, k);
1718 if (is_Proj(succ_succ)) {
1719 if (get_Proj_proj(succ_succ) == j) {
1721 weight += calc_method_local_weight(succ_succ);
1724 /* this should NOT happen */
1731 /* any other node: unsupported yet or bad. */
1739 * Calculate the parameter weights for transmitting the address of a local variable.
1741 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1742 ir_entity *ent = get_irg_entity(irg);
1744 int nparams, i, proj_nr;
1745 ir_node *irg_args, *arg;
1747 mtp = get_entity_type(ent);
1748 nparams = get_method_n_params(mtp);
1750 /* allocate a new array. currently used as 'analysed' flag */
1751 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1753 /* If the method haven't parameters we have nothing to do. */
1757 assure_irg_outs(irg);
1758 irg_args = get_irg_args(irg);
1759 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1760 arg = get_irn_out(irg_args, i);
1761 proj_nr = get_Proj_proj(arg);
1762 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1767 * Calculate the benefice for transmitting an local variable address.
1768 * After inlining, the local variable might be transformed into a
1769 * SSA variable by scalar_replacement().
1771 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1772 inline_irg_env *env = get_irg_link(callee);
1774 if (env->local_weights != NULL) {
1775 if (pos < ARR_LEN(env->local_weights))
1776 return env->local_weights[pos];
1780 analyze_irg_local_weights(env, callee);
1782 if (pos < ARR_LEN(env->local_weights))
1783 return env->local_weights[pos];
1788 * calculate a benefice value for inlining the given call.
1790 static int calc_inline_benefice(ir_node *call, ir_graph *callee, unsigned *local_adr) {
1791 ir_entity *ent = get_irg_entity(callee);
1798 inline_irg_env *curr_env, *callee_env;
1800 if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1801 /* do NOT inline noreturn calls */
1805 /* costs for every passed parameter */
1806 n_params = get_Call_n_params(call);
1807 mtp = get_entity_type(ent);
1808 cc = get_method_calling_convention(mtp);
1809 if (cc & cc_reg_param) {
1810 /* register parameter, smaller costs for register parameters */
1811 int max_regs = cc & ~cc_bits;
1813 if (max_regs < n_params)
1814 weight += max_regs * 2 + (n_params - max_regs) * 5;
1816 weight += n_params * 2;
1818 /* parameters are passed an stack */
1819 weight += 5 * n_params;
1822 /* constant parameters improve the benefice */
1823 frame_ptr = get_irg_frame(current_ir_graph);
1824 for (i = 0; i < n_params; ++i) {
1825 ir_node *param = get_Call_param(call, i);
1827 if (is_Const(param) || is_SymConst(param))
1828 weight += get_method_param_weight(ent, i);
1829 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1831 * An address of a local variable is transmitted. After inlining,
1832 * scalar_replacement might be able to remove the local variable,
1835 v = get_method_local_adress_weight(callee, i);
1842 callee_env = get_irg_link(callee);
1843 if (get_entity_visibility(ent) == visibility_local &&
1844 callee_env->n_callers_orig == 1 &&
1845 callee != current_ir_graph) {
1846 /* we are the only caller, give big bonus */
1850 /* do not inline big functions */
1851 weight -= callee_env->n_nodes;
1853 /* reduce the benefice if the current function is already big */
1854 curr_env = get_irg_link(current_ir_graph);
1855 weight -= curr_env->n_nodes / 100;
1857 /* give a bonus for functions with one block */
1858 if (callee_env->n_blocks == 1)
1859 weight = weight * 3 / 2;
1865 * Heuristic inliner. Calculates a benifice value for every call and inlines
1866 * those calls with a value higher than the threshold.
1868 void inline_functions(int inline_threshold) {
1869 inline_irg_env *env;
1874 call_entry *entry, *tail;
1875 const call_entry *centry;
1876 pmap *copied_graphs;
1877 pmap_entry *pm_entry;
1879 rem = current_ir_graph;
1880 obstack_init(&temp_obst);
1882 /* a map for the copied graphs, used to inline recursive calls */
1883 copied_graphs = pmap_create();
1885 /* extend all irgs by a temporary data structure for inlining. */
1886 n_irgs = get_irp_n_irgs();
1887 for (i = 0; i < n_irgs; ++i)
1888 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1890 /* Precompute information in temporary data structure. */
1891 wenv.ignore_runtime = 0;
1892 wenv.ignore_callers = 0;
1893 for (i = 0; i < n_irgs; ++i) {
1894 ir_graph *irg = get_irp_irg(i);
1896 assert(get_irg_phase_state(irg) != phase_building);
1897 free_callee_info(irg);
1899 wenv.x = get_irg_link(irg);
1900 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1903 /* -- and now inline. -- */
1904 for (i = 0; i < n_irgs; ++i) {
1905 int phiproj_computed = 0;
1907 ir_graph *irg = get_irp_irg(i);
1909 current_ir_graph = irg;
1910 env = get_irg_link(irg);
1912 /* note that the list of possible calls is updated during the process */
1914 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1921 callee = entry->callee;
1923 /* calculate the benifice on the original call to prevent excessive inlining */
1925 benefice = calc_inline_benefice(call, callee, &local_adr);
1926 DB((dbg, LEVEL_2, "In %+F Call %+F has benefice %d\n", irg, callee, benefice));
1928 e = pmap_find(copied_graphs, callee);
1931 * Remap callee if we have a copy.
1932 * FIXME: Should we do this only for recursive Calls ?
1937 if (benefice > -inline_threshold ||
1938 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1939 if (current_ir_graph == callee) {
1941 * Recursive call: we cannot directly inline because we cannot walk
1942 * the graph and change it. So we have to make a copy of the graph
1946 inline_irg_env *callee_env;
1950 * No copy yet, create one.
1951 * Note that recursive methods are never leaves, so it is sufficient
1952 * to test this condition here.
1954 copy = create_irg_copy(callee);
1956 /* create_irg_copy() destroys the Proj links, recompute them */
1957 phiproj_computed = 0;
1959 /* allocate new environment */
1960 callee_env = alloc_inline_irg_env();
1961 set_irg_link(copy, callee_env);
1963 wenv.x = callee_env;
1964 wenv.ignore_callers = 1;
1965 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1968 * Enter the entity of the original graph. This is needed
1969 * for inline_method(). However, note that ent->irg still points
1970 * to callee, NOT to copy.
1972 set_irg_entity(copy, get_irg_entity(callee));
1974 pmap_insert(copied_graphs, callee, copy);
1977 /* we have only one caller: the original graph */
1978 callee_env->n_callers = 1;
1979 callee_env->n_callers_orig = 1;
1981 if (! phiproj_computed) {
1982 phiproj_computed = 1;
1983 collect_phiprojs(current_ir_graph);
1985 did_inline = inline_method(call, callee);
1987 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1989 /* was inlined, must be recomputed */
1990 phiproj_computed = 0;
1992 /* callee was inline. Append it's call list. */
1993 env->got_inline = 1;
1995 env->local_vars = 1;
1996 --env->n_call_nodes;
1997 append_call_list(env, callee_env->call_head);
1998 env->n_call_nodes += callee_env->n_call_nodes;
1999 env->n_nodes += callee_env->n_nodes;
2000 --callee_env->n_callers;
2002 /* after we have inlined callee, all called methods inside callee
2003 are now called once more */
2004 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
2005 inline_irg_env *penv = get_irg_link(centry->callee);
2009 /* remove this call from the list */
2011 tail->next = entry->next;
2013 env->call_head = entry->next;
2019 env->call_tail = tail;
2021 if (env->got_inline) {
2022 /* this irg got calls inlined: optimize it */
2024 /* scalar replacement does not work well with Tuple nodes, so optimize them away */
2025 optimize_graph_df(irg);
2027 if (env->local_vars) {
2028 if (scalar_replacement_opt(irg)) {
2029 optimize_graph_df(irg);
2035 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2036 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2037 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2038 env->n_callers_orig, env->n_callers,
2039 get_entity_name(get_irg_entity(irg))));
2043 /* kill the copied graphs: we don't need them anymore */
2044 foreach_pmap(copied_graphs, pm_entry) {
2045 ir_graph *copy = pm_entry->value;
2047 /* reset the entity, otherwise it will be deleted in the next step ... */
2048 set_irg_entity(copy, NULL);
2049 free_ir_graph(copy);
2051 pmap_destroy(copied_graphs);
2053 obstack_free(&temp_obst, NULL);
2054 current_ir_graph = rem;
2057 void firm_init_inline(void) {
2058 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");