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
32 #include "irgraph_t.h"
35 #include "iroptimize.h"
52 #include "irbackedge_t.h"
53 #include "opt_inline_t.h"
58 #include "analyze_irg_args.h"
59 #include "iredges_t.h"
63 #include "iropt_dbg.h"
65 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
67 /*------------------------------------------------------------------*/
68 /* Routines for dead node elimination / copying garbage collection */
70 /*------------------------------------------------------------------*/
73 * Remember the new node in the old node by using a field all nodes have.
75 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
78 * Get this new node, before the old node is forgotten.
80 #define get_new_node(oldn) get_irn_link(oldn)
83 * Check if a new node was set.
85 #define has_new_node(n) (get_new_node(n) != NULL)
88 * We use the block_visited flag to mark that we have computed the
89 * number of useful predecessors for this block.
90 * Further we encode the new arity in this flag in the old blocks.
91 * Remembering the arity is useful, as it saves a lot of pointer
92 * accesses. This function is called for all Phi and Block nodes
96 compute_new_arity(ir_node *b) {
97 int i, res, irn_arity;
100 irg_v = get_irg_block_visited(current_ir_graph);
101 block_v = get_Block_block_visited(b);
102 if (block_v >= irg_v) {
103 /* we computed the number of preds for this block and saved it in the
105 return block_v - irg_v;
107 /* compute the number of good predecessors */
108 res = irn_arity = get_irn_arity(b);
109 for (i = 0; i < irn_arity; i++)
110 if (is_Bad(get_irn_n(b, i))) res--;
111 /* save it in the flag. */
112 set_Block_block_visited(b, irg_v + res);
118 * Copies the node to the new obstack. The Ins of the new node point to
119 * the predecessors on the old obstack. For block/phi nodes not all
120 * predecessors might be copied. n->link points to the new node.
121 * For Phi and Block nodes the function allocates in-arrays with an arity
122 * only for useful predecessors. The arity is determined by counting
123 * the non-bad predecessors of the block.
125 * @param n The node to be copied
126 * @param env if non-NULL, the node number attribute will be copied to the new node
128 * Note: Also used for loop unrolling.
130 static void copy_node(ir_node *n, void *env) {
133 ir_op *op = get_irn_op(n);
137 /* node copied already */
139 } else if (op == op_Block) {
141 new_arity = compute_new_arity(n);
142 n->attr.block.graph_arr = NULL;
144 block = get_nodes_block(n);
146 new_arity = compute_new_arity(block);
148 new_arity = get_irn_arity(n);
151 nn = new_ir_node(get_irn_dbg_info(n),
158 /* Copy the attributes. These might point to additional data. If this
159 was allocated on the old obstack the pointers now are dangling. This
160 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
161 if (op == op_Block) {
162 /* we cannot allow blocks WITHOUT macroblock input */
163 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
165 copy_node_attr(n, nn);
169 int copy_node_nr = env != NULL;
171 /* for easier debugging, we want to copy the node numbers too */
172 nn->node_nr = n->node_nr;
178 hook_dead_node_elim_subst(current_ir_graph, n, nn);
182 * Copies new predecessors of old node to new node remembered in link.
183 * Spare the Bad predecessors of Phi and Block nodes.
185 static void copy_preds(ir_node *n, void *env) {
190 nn = get_new_node(n);
193 /* copy the macro block header */
194 ir_node *mbh = get_Block_MacroBlock(n);
197 /* this block is a macroblock header */
198 set_Block_MacroBlock(nn, nn);
200 /* get the macro block header */
201 ir_node *nmbh = get_new_node(mbh);
202 assert(nmbh != NULL);
203 set_Block_MacroBlock(nn, nmbh);
206 /* Don't copy Bad nodes. */
208 irn_arity = get_irn_arity(n);
209 for (i = 0; i < irn_arity; i++) {
210 if (! is_Bad(get_irn_n(n, i))) {
211 ir_node *pred = get_irn_n(n, i);
212 set_irn_n(nn, j, get_new_node(pred));
216 /* repair the block visited flag from above misuse. Repair it in both
217 graphs so that the old one can still be used. */
218 set_Block_block_visited(nn, 0);
219 set_Block_block_visited(n, 0);
220 /* Local optimization could not merge two subsequent blocks if
221 in array contained Bads. Now it's possible.
222 We don't call optimize_in_place as it requires
223 that the fields in ir_graph are set properly. */
224 if (!has_Block_label(nn) &&
225 get_opt_control_flow_straightening() &&
226 get_Block_n_cfgpreds(nn) == 1 &&
227 is_Jmp(get_Block_cfgpred(nn, 0))) {
228 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
230 /* Jmp jumps into the block it is in -- deal self cycle. */
231 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
232 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
237 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
238 /* Don't copy node if corresponding predecessor in block is Bad.
239 The Block itself should not be Bad. */
240 block = get_nodes_block(n);
241 set_nodes_block(nn, get_new_node(block));
243 irn_arity = get_irn_arity(n);
244 for (i = 0; i < irn_arity; i++) {
245 if (! is_Bad(get_irn_n(block, i))) {
246 ir_node *pred = get_irn_n(n, i);
247 set_irn_n(nn, j, get_new_node(pred));
248 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
252 /* If the pre walker reached this Phi after the post walker visited the
253 block block_visited is > 0. */
254 set_Block_block_visited(get_nodes_block(n), 0);
255 /* Compacting the Phi's ins might generate Phis with only one
257 if (get_irn_arity(nn) == 1)
258 exchange(nn, get_irn_n(nn, 0));
260 irn_arity = get_irn_arity(n);
261 for (i = -1; i < irn_arity; i++)
262 set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
264 /* Now the new node is complete. We can add it to the hash table for CSE.
265 @@@ inlining aborts if we identify End. Why? */
267 add_identities(current_ir_graph->value_table, nn);
271 * Copies the graph recursively, compacts the keep-alives of the end node.
273 * @param irg the graph to be copied
274 * @param copy_node_nr If non-zero, the node number will be copied
276 static void copy_graph(ir_graph *irg, int copy_node_nr) {
277 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
278 ir_node *ka; /* keep alive */
282 /* Some nodes must be copied by hand, sigh */
283 vfl = get_irg_visited(irg);
284 set_irg_visited(irg, vfl + 1);
286 oe = get_irg_end(irg);
287 mark_irn_visited(oe);
288 /* copy the end node by hand, allocate dynamic in array! */
289 ne = new_ir_node(get_irn_dbg_info(oe),
296 /* Copy the attributes. Well, there might be some in the future... */
297 copy_node_attr(oe, ne);
298 set_new_node(oe, ne);
300 /* copy the Bad node */
301 ob = get_irg_bad(irg);
302 mark_irn_visited(ob);
303 nb = new_ir_node(get_irn_dbg_info(ob),
310 copy_node_attr(ob, nb);
311 set_new_node(ob, nb);
313 /* copy the NoMem node */
314 om = get_irg_no_mem(irg);
315 mark_irn_visited(om);
316 nm = new_ir_node(get_irn_dbg_info(om),
323 copy_node_attr(om, nm);
324 set_new_node(om, nm);
326 /* copy the live nodes */
327 set_irg_visited(irg, vfl);
328 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
330 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
332 /* visit the anchors as well */
333 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
334 ir_node *n = get_irg_anchor(irg, i);
336 if (n && (get_irn_visited(n) <= vfl)) {
337 set_irg_visited(irg, vfl);
338 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
342 /* copy_preds for the end node ... */
343 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
345 /*- ... and now the keep alives. -*/
346 /* First pick the not marked block nodes and walk them. We must pick these
347 first as else we will oversee blocks reachable from Phis. */
348 irn_arity = get_End_n_keepalives(oe);
349 for (i = 0; i < irn_arity; i++) {
350 ka = get_End_keepalive(oe, i);
352 if (get_irn_visited(ka) <= vfl) {
353 /* We must keep the block alive and copy everything reachable */
354 set_irg_visited(irg, vfl);
355 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
357 add_End_keepalive(ne, get_new_node(ka));
361 /* Now pick other nodes. Here we will keep all! */
362 irn_arity = get_End_n_keepalives(oe);
363 for (i = 0; i < irn_arity; i++) {
364 ka = get_End_keepalive(oe, i);
366 if (get_irn_visited(ka) <= vfl) {
367 /* We didn't copy the node yet. */
368 set_irg_visited(irg, vfl);
369 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
371 add_End_keepalive(ne, get_new_node(ka));
375 /* start block sometimes only reached after keep alives */
376 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
377 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
381 * Copies the graph reachable from current_ir_graph->end to the obstack
382 * in current_ir_graph and fixes the environment.
383 * Then fixes the fields in current_ir_graph containing nodes of the
386 * @param copy_node_nr If non-zero, the node number will be copied
389 copy_graph_env(int copy_node_nr) {
390 ir_graph *irg = current_ir_graph;
391 ir_node *old_end, *new_anchor;
394 /* remove end_except and end_reg nodes */
395 old_end = get_irg_end(irg);
396 set_irg_end_except (irg, old_end);
397 set_irg_end_reg (irg, old_end);
399 /* Not all nodes remembered in irg might be reachable
400 from the end node. Assure their link is set to NULL, so that
401 we can test whether new nodes have been computed. */
402 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
403 ir_node *n = get_irg_anchor(irg, i);
405 set_new_node(n, NULL);
407 /* we use the block walk flag for removing Bads from Blocks ins. */
408 inc_irg_block_visited(irg);
411 copy_graph(irg, copy_node_nr);
414 old_end = get_irg_end(irg);
415 new_anchor = new_Anchor(irg);
417 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
418 ir_node *n = get_irg_anchor(irg, i);
420 set_irn_n(new_anchor, i, get_new_node(n));
423 irg->anchor = new_anchor;
425 /* ensure the new anchor is placed in the endblock */
426 set_nodes_block(new_anchor, get_irg_end_block(irg));
430 * Copies all reachable nodes to a new obstack. Removes bad inputs
431 * from block nodes and the corresponding inputs from Phi nodes.
432 * Merges single exit blocks with single entry blocks and removes
434 * Adds all new nodes to a new hash table for CSE. Does not
435 * perform CSE, so the hash table might contain common subexpressions.
437 void dead_node_elimination(ir_graph *irg) {
439 #ifdef INTERPROCEDURAL_VIEW
440 int rem_ipview = get_interprocedural_view();
442 struct obstack *graveyard_obst = NULL;
443 struct obstack *rebirth_obst = NULL;
445 edges_deactivate(irg);
447 /* inform statistics that we started a dead-node elimination run */
448 hook_dead_node_elim(irg, 1);
450 /* Remember external state of current_ir_graph. */
451 rem = current_ir_graph;
452 current_ir_graph = irg;
453 #ifdef INTERPROCEDURAL_VIEW
454 set_interprocedural_view(0);
457 assert(get_irg_phase_state(irg) != phase_building);
459 /* Handle graph state */
460 free_callee_info(irg);
464 /* @@@ so far we loose loops when copying */
465 free_loop_information(irg);
467 set_irg_doms_inconsistent(irg);
469 /* A quiet place, where the old obstack can rest in peace,
470 until it will be cremated. */
471 graveyard_obst = irg->obst;
473 /* A new obstack, where the reachable nodes will be copied to. */
474 rebirth_obst = XMALLOC(struct obstack);
475 irg->obst = rebirth_obst;
476 obstack_init(irg->obst);
477 irg->last_node_idx = 0;
479 /* We also need a new value table for CSE */
480 del_identities(irg->value_table);
481 irg->value_table = new_identities();
483 /* Copy the graph from the old to the new obstack */
484 copy_graph_env(/*copy_node_nr=*/1);
486 /* Free memory from old unoptimized obstack */
487 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
488 xfree(graveyard_obst); /* ... then free it. */
490 /* inform statistics that the run is over */
491 hook_dead_node_elim(irg, 0);
493 current_ir_graph = rem;
494 #ifdef INTERPROCEDURAL_VIEW
495 set_interprocedural_view(rem_ipview);
500 * Relink bad predecessors of a block and store the old in array to the
501 * link field. This function is called by relink_bad_predecessors().
502 * The array of link field starts with the block operand at position 0.
503 * If block has bad predecessors, create a new in array without bad preds.
504 * Otherwise let in array untouched.
506 static void relink_bad_block_predecessors(ir_node *n, void *env) {
507 ir_node **new_in, *irn;
508 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
511 /* if link field of block is NULL, look for bad predecessors otherwise
512 this is already done */
513 if (is_Block(n) && get_irn_link(n) == NULL) {
514 /* save old predecessors in link field (position 0 is the block operand)*/
515 set_irn_link(n, get_irn_in(n));
517 /* count predecessors without bad nodes */
518 old_irn_arity = get_irn_arity(n);
519 for (i = 0; i < old_irn_arity; i++)
520 if (!is_Bad(get_irn_n(n, i)))
523 /* arity changing: set new predecessors without bad nodes */
524 if (new_irn_arity < old_irn_arity) {
525 /* Get new predecessor array. We do not resize the array, as we must
526 keep the old one to update Phis. */
527 new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
529 /* set new predecessors in array */
532 for (i = 0; i < old_irn_arity; i++) {
533 irn = get_irn_n(n, i);
535 new_in[new_irn_n] = irn;
536 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
540 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
541 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
543 } /* ir node has bad predecessors */
544 } /* Block is not relinked */
548 * Relinks Bad predecessors from Blocks and Phis called by walker
549 * remove_bad_predecesors(). If n is a Block, call
550 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
551 * function of Phi's Block. If this block has bad predecessors, relink preds
554 static void relink_bad_predecessors(ir_node *n, void *env) {
555 ir_node *block, **old_in;
556 int i, old_irn_arity, new_irn_arity;
558 /* relink bad predecessors of a block */
560 relink_bad_block_predecessors(n, env);
562 /* If Phi node relink its block and its predecessors */
564 /* Relink predecessors of phi's block */
565 block = get_nodes_block(n);
566 if (get_irn_link(block) == NULL)
567 relink_bad_block_predecessors(block, env);
569 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
570 old_irn_arity = ARR_LEN(old_in);
572 /* Relink Phi predecessors if count of predecessors changed */
573 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
574 /* set new predecessors in array
575 n->in[0] remains the same block */
577 for(i = 1; i < old_irn_arity; i++)
578 if (!is_Bad(old_in[i])) {
579 n->in[new_irn_arity] = n->in[i];
580 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
584 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
585 ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
587 } /* n is a Phi node */
591 * Removes Bad Bad predecessors from Blocks and the corresponding
592 * inputs to Phi nodes as in dead_node_elimination but without
594 * On walking up set the link field to NULL, on walking down call
595 * relink_bad_predecessors() (This function stores the old in array
596 * to the link field and sets a new in array if arity of predecessors
599 void remove_bad_predecessors(ir_graph *irg) {
600 panic("Fix backedge handling first");
601 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
608 __)|_| | \_/ | \_/(/_ |_/\__|__
610 The following stuff implements a facility that automatically patches
611 registered ir_node pointers to the new node when a dead node elimination occurs.
614 struct _survive_dce_t {
618 hook_entry_t dead_node_elim;
619 hook_entry_t dead_node_elim_subst;
622 typedef struct _survive_dce_list_t {
623 struct _survive_dce_list_t *next;
625 } survive_dce_list_t;
627 static void dead_node_hook(void *context, ir_graph *irg, int start) {
628 survive_dce_t *sd = context;
631 /* Create a new map before the dead node elimination is performed. */
633 sd->new_places = pmap_create_ex(pmap_count(sd->places));
635 /* Patch back all nodes if dead node elimination is over and something is to be done. */
636 pmap_destroy(sd->places);
637 sd->places = sd->new_places;
638 sd->new_places = NULL;
643 * Hook called when dead node elimination replaces old by nw.
645 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
646 survive_dce_t *sd = context;
647 survive_dce_list_t *list = pmap_get(sd->places, old);
650 /* If the node is to be patched back, write the new address to all registered locations. */
652 survive_dce_list_t *p;
654 for (p = list; p; p = p->next)
657 pmap_insert(sd->new_places, nw, list);
662 * Make a new Survive DCE environment.
664 survive_dce_t *new_survive_dce(void) {
665 survive_dce_t *res = XMALLOC(survive_dce_t);
666 obstack_init(&res->obst);
667 res->places = pmap_create();
668 res->new_places = NULL;
670 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
671 res->dead_node_elim.context = res;
672 res->dead_node_elim.next = NULL;
674 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
675 res->dead_node_elim_subst.context = res;
676 res->dead_node_elim_subst.next = NULL;
678 register_hook(hook_dead_node_elim, &res->dead_node_elim);
679 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
684 * Free a Survive DCE environment.
686 void free_survive_dce(survive_dce_t *sd) {
687 obstack_free(&sd->obst, NULL);
688 pmap_destroy(sd->places);
689 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
690 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
695 * Register a node pointer to be patched upon DCE.
696 * When DCE occurs, the node pointer specified by @p place will be
697 * patched to the new address of the node it is pointing to.
699 * @param sd The Survive DCE environment.
700 * @param place The address of the node pointer.
702 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
703 if (*place != NULL) {
704 ir_node *irn = *place;
705 survive_dce_list_t *curr = pmap_get(sd->places, irn);
706 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
711 pmap_insert(sd->places, irn, nw);
715 /*--------------------------------------------------------------------*/
716 /* Functionality for inlining */
717 /*--------------------------------------------------------------------*/
720 * Copy node for inlineing. Updates attributes that change when
721 * inlineing but not for dead node elimination.
723 * Copies the node by calling copy_node() and then updates the entity if
724 * it's a local one. env must be a pointer of the frame type of the
725 * inlined procedure. The new entities must be in the link field of
728 static void copy_node_inline(ir_node *n, void *env) {
730 ir_type *frame_tp = (ir_type *)env;
734 nn = get_new_node (n);
736 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
737 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
739 } else if (is_Block(n)) {
740 nn = get_new_node (n);
741 nn->attr.block.irg = current_ir_graph;
746 * Copies new predecessors of old node and move constants to
749 static void copy_preds_inline(ir_node *n, void *env) {
753 nn = skip_Id(get_new_node(n));
754 if (is_irn_constlike(nn)) {
755 /* move Constants into the start block */
756 set_nodes_block(nn, get_irg_start_block(current_ir_graph));
758 n = identify_remember(current_ir_graph->value_table, nn);
767 * Walker: checks if P_value_arg_base is used.
769 static void find_addr(ir_node *node, void *env) {
770 int *allow_inline = env;
772 is_Start(get_Proj_pred(node)) &&
773 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
775 } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
777 * Refuse to inline alloca call unless user explicitly forced so as this
778 * may change program's memory overhead drastically when the function
779 * using alloca is called in loop. In GCC present in SPEC2000 inlining
780 * into schedule_block cause it to require 2GB of ram instead of 256MB.
782 * Sorrily this is true with our implementation also.
783 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
784 * disables inlining of functions using VLA (with are completely save).
787 * - add a flag to the Alloc node for "real" alloca() calls
788 * - add a new Stack-Restore node at the end of a function using alloca()
795 * Check if we can inline a given call.
796 * Currently, we cannot inline two cases:
797 * - call with compound arguments
798 * - graphs that take the address of a parameter
800 * check these conditions here
802 static int can_inline(ir_node *call, ir_graph *called_graph) {
803 ir_type *call_type = get_Call_type(call);
804 int params, ress, i, res;
805 assert(is_Method_type(call_type));
807 params = get_method_n_params(call_type);
808 ress = get_method_n_ress(call_type);
810 /* check parameters for compound arguments */
811 for (i = 0; i < params; ++i) {
812 ir_type *p_type = get_method_param_type(call_type, i);
814 if (is_compound_type(p_type))
818 /* check results for compound arguments */
819 for (i = 0; i < ress; ++i) {
820 ir_type *r_type = get_method_res_type(call_type, i);
822 if (is_compound_type(r_type))
827 irg_walk_graph(called_graph, find_addr, NULL, &res);
833 exc_handler = 0, /**< There is a handler. */
834 exc_to_end = 1, /**< Branches to End. */
835 exc_no_handler = 2 /**< Exception handling not represented. */
838 /* Inlines a method at the given call site. */
839 int inline_method(ir_node *call, ir_graph *called_graph) {
841 ir_node *post_call, *post_bl;
842 ir_node *in[pn_Start_max];
843 ir_node *end, *end_bl, *block;
848 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
849 enum exc_mode exc_handling;
850 ir_type *called_frame, *curr_frame, *mtp, *ctp;
853 irg_inline_property prop = get_irg_inline_property(called_graph);
854 unsigned long visited;
856 if (prop == irg_inline_forbidden)
859 ent = get_irg_entity(called_graph);
861 mtp = get_entity_type(ent);
862 ctp = get_Call_type(call);
863 if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
864 /* this is a bad feature of C: without a prototype, we can can call a function with less
865 parameters than needed. Currently we don't support this, although it would be
866 to use Unknown than. */
870 /* Argh, compiling C has some bad consequences:
871 the call type AND the method type might be different.
872 It is implementation defendant what happens in that case.
873 We support inlining, if the bitsize of the types matches AND
874 the same arithmetic is used. */
875 n_params = get_method_n_params(mtp);
876 for (i = n_params - 1; i >= 0; --i) {
877 ir_type *param_tp = get_method_param_type(mtp, i);
878 ir_type *arg_tp = get_method_param_type(ctp, i);
880 if (param_tp != arg_tp) {
881 ir_mode *pmode = get_type_mode(param_tp);
882 ir_mode *amode = get_type_mode(arg_tp);
884 if (pmode == NULL || amode == NULL)
886 if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
888 if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
890 /* otherwise we can simply "reinterpret" the bits */
894 irg = get_irn_irg(call);
897 * We cannot inline a recursive call. The graph must be copied before
898 * the call the inline_method() using create_irg_copy().
900 if (called_graph == irg)
904 * currently, we cannot inline two cases:
905 * - call with compound arguments
906 * - graphs that take the address of a parameter
908 if (! can_inline(call, called_graph))
911 rem = current_ir_graph;
912 current_ir_graph = irg;
914 DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
916 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
917 rem_opt = get_opt_optimize();
920 /* Handle graph state */
921 assert(get_irg_phase_state(irg) != phase_building);
922 assert(get_irg_pinned(irg) == op_pin_state_pinned);
923 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
924 set_irg_outs_inconsistent(irg);
925 set_irg_extblk_inconsistent(irg);
926 set_irg_doms_inconsistent(irg);
927 set_irg_loopinfo_inconsistent(irg);
928 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
930 /* -- Check preconditions -- */
931 assert(is_Call(call));
933 /* here we know we WILL inline, so inform the statistics */
934 hook_inline(call, called_graph);
936 /* -- Decide how to handle exception control flow: Is there a handler
937 for the Call node, or do we branch directly to End on an exception?
939 0 There is a handler.
941 2 Exception handling not represented in Firm. -- */
943 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
944 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
945 long proj_nr = get_Proj_proj(proj);
946 if (proj_nr == pn_Call_X_except) Xproj = proj;
947 if (proj_nr == pn_Call_M_except) Mproj = proj;
949 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
950 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
951 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
954 /* create the argument tuple */
955 NEW_ARR_A(ir_type *, args_in, n_params);
957 block = get_nodes_block(call);
958 for (i = n_params - 1; i >= 0; --i) {
959 ir_node *arg = get_Call_param(call, i);
960 ir_type *param_tp = get_method_param_type(mtp, i);
961 ir_mode *mode = get_type_mode(param_tp);
963 if (mode != get_irn_mode(arg)) {
964 arg = new_r_Conv(irg, block, arg, mode);
970 the procedure and later replaces the Start node of the called graph.
971 Post_call is the old Call node and collects the results of the called
972 graph. Both will end up being a tuple. -- */
973 post_bl = get_nodes_block(call);
974 set_irg_current_block(irg, post_bl);
975 /* XxMxPxPxPxT of Start + parameter of Call */
976 in[pn_Start_X_initial_exec] = new_Jmp();
977 in[pn_Start_M] = get_Call_mem(call);
978 in[pn_Start_P_frame_base] = get_irg_frame(irg);
979 in[pn_Start_P_tls] = get_irg_tls(irg);
980 in[pn_Start_T_args] = new_Tuple(n_params, args_in);
981 /* in[pn_Start_P_value_arg_base] = ??? */
982 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
983 pre_call = new_Tuple(pn_Start_max - 1, in);
987 The new block gets the ins of the old block, pre_call and all its
988 predecessors and all Phi nodes. -- */
989 part_block(pre_call);
991 /* -- Prepare state for dead node elimination -- */
992 /* Visited flags in calling irg must be >= flag in called irg.
993 Else walker and arity computation will not work. */
994 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
995 set_irg_visited(irg, get_irg_visited(called_graph) + 1);
996 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
997 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
998 visited = get_irg_visited(irg);
1000 /* Set pre_call as new Start node in link field of the start node of
1001 calling graph and pre_calls block as new block for the start block
1003 Further mark these nodes so that they are not visited by the
1005 set_irn_link(get_irg_start(called_graph), pre_call);
1006 set_irn_visited(get_irg_start(called_graph), visited);
1007 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1008 set_irn_visited(get_irg_start_block(called_graph), visited);
1010 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1011 set_irn_visited(get_irg_bad(called_graph), visited);
1013 set_irn_link(get_irg_no_mem(called_graph), get_irg_no_mem(current_ir_graph));
1014 set_irn_visited(get_irg_no_mem(called_graph), visited);
1016 /* Initialize for compaction of in arrays */
1017 inc_irg_block_visited(irg);
1019 /* -- Replicate local entities of the called_graph -- */
1020 /* copy the entities. */
1021 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1022 called_frame = get_irg_frame_type(called_graph);
1023 curr_frame = get_irg_frame_type(irg);
1024 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
1025 ir_entity *new_ent, *old_ent;
1026 old_ent = get_class_member(called_frame, i);
1027 new_ent = copy_entity_own(old_ent, curr_frame);
1028 set_entity_link(old_ent, new_ent);
1031 /* visited is > than that of called graph. With this trick visited will
1032 remain unchanged so that an outer walker, e.g., searching the call nodes
1033 to inline, calling this inline will not visit the inlined nodes. */
1034 set_irg_visited(irg, get_irg_visited(irg)-1);
1036 /* -- Performing dead node elimination inlines the graph -- */
1037 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1039 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
1040 get_irg_frame_type(called_graph));
1042 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1044 /* Repair called_graph */
1045 set_irg_visited(called_graph, get_irg_visited(irg));
1046 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
1047 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1049 /* -- Merge the end of the inlined procedure with the call site -- */
1050 /* We will turn the old Call node into a Tuple with the following
1053 0: Phi of all Memories of Return statements.
1054 1: Jmp from new Block that merges the control flow from all exception
1055 predecessors of the old end block.
1056 2: Tuple of all arguments.
1057 3: Phi of Exception memories.
1058 In case the old Call directly branches to End on an exception we don't
1059 need the block merging all exceptions nor the Phi of the exception
1063 /* -- Precompute some values -- */
1064 end_bl = get_new_node(get_irg_end_block(called_graph));
1065 end = get_new_node(get_irg_end(called_graph));
1066 arity = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret */
1067 n_res = get_method_n_ress(get_Call_type(call));
1069 res_pred = XMALLOCN(ir_node*, n_res);
1070 cf_pred = XMALLOCN(ir_node*, arity);
1072 set_irg_current_block(irg, post_bl); /* just to make sure */
1074 /* -- archive keepalives -- */
1075 irn_arity = get_irn_arity(end);
1076 for (i = 0; i < irn_arity; i++) {
1077 ir_node *ka = get_End_keepalive(end, i);
1079 add_End_keepalive(get_irg_end(irg), ka);
1082 /* The new end node will die. We need not free as the in array is on the obstack:
1083 copy_node() only generated 'D' arrays. */
1085 /* -- Replace Return nodes by Jump nodes. -- */
1087 for (i = 0; i < arity; i++) {
1089 ret = get_Block_cfgpred(end_bl, i);
1090 if (is_Return(ret)) {
1091 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1095 set_irn_in(post_bl, n_ret, cf_pred);
1097 /* -- Build a Tuple for all results of the method.
1098 Add Phi node if there was more than one Return. -- */
1099 turn_into_tuple(post_call, pn_Call_max);
1100 /* First the Memory-Phi */
1102 for (i = 0; i < arity; i++) {
1103 ret = get_Block_cfgpred(end_bl, i);
1104 if (is_Return(ret)) {
1105 cf_pred[n_ret] = get_Return_mem(ret);
1109 phi = new_Phi(n_ret, cf_pred, mode_M);
1110 set_Tuple_pred(call, pn_Call_M_regular, phi);
1111 /* Conserve Phi-list for further inlinings -- but might be optimized */
1112 if (get_nodes_block(phi) == post_bl) {
1113 set_irn_link(phi, get_irn_link(post_bl));
1114 set_irn_link(post_bl, phi);
1116 /* Now the real results */
1118 for (j = 0; j < n_res; j++) {
1120 for (i = 0; i < arity; i++) {
1121 ret = get_Block_cfgpred(end_bl, i);
1122 if (is_Return(ret)) {
1123 cf_pred[n_ret] = get_Return_res(ret, j);
1128 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1132 /* Conserve Phi-list for further inlinings -- but might be optimized */
1133 if (get_nodes_block(phi) == post_bl) {
1134 set_Phi_next(phi, get_Block_phis(post_bl));
1135 set_Block_phis(post_bl, phi);
1138 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1140 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1142 /* handle the regular call */
1143 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1145 /* For now, we cannot inline calls with value_base */
1146 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1148 /* Finally the exception control flow.
1149 We have two (three) possible situations:
1150 First if the Call branches to an exception handler: We need to add a Phi node to
1151 collect the memory containing the exception objects. Further we need
1152 to add another block to get a correct representation of this Phi. To
1153 this block we add a Jmp that resolves into the X output of the Call
1154 when the Call is turned into a tuple.
1155 Second the Call branches to End, the exception is not handled. Just
1156 add all inlined exception branches to the End node.
1157 Third: there is no Exception edge at all. Handle as case two. */
1158 if (exc_handling == exc_handler) {
1160 for (i = 0; i < arity; i++) {
1162 ret = get_Block_cfgpred(end_bl, i);
1163 irn = skip_Proj(ret);
1164 if (is_fragile_op(irn) || is_Raise(irn)) {
1165 cf_pred[n_exc] = ret;
1170 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1171 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1172 /* The Phi for the memories with the exception objects */
1174 for (i = 0; i < arity; i++) {
1176 ret = skip_Proj(get_Block_cfgpred(end_bl, i));
1178 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1180 } else if (is_fragile_op(ret)) {
1181 /* We rely that all cfops have the memory output at the same position. */
1182 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1184 } else if (is_Raise(ret)) {
1185 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1189 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1191 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1192 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1195 ir_node *main_end_bl;
1196 int main_end_bl_arity;
1197 ir_node **end_preds;
1199 /* assert(exc_handling == 1 || no exceptions. ) */
1201 for (i = 0; i < arity; i++) {
1202 ir_node *ret = get_Block_cfgpred(end_bl, i);
1203 ir_node *irn = skip_Proj(ret);
1205 if (is_fragile_op(irn) || is_Raise(irn)) {
1206 cf_pred[n_exc] = ret;
1210 main_end_bl = get_irg_end_block(irg);
1211 main_end_bl_arity = get_irn_arity(main_end_bl);
1212 end_preds = XMALLOCN(ir_node*, n_exc + main_end_bl_arity);
1214 for (i = 0; i < main_end_bl_arity; ++i)
1215 end_preds[i] = get_irn_n(main_end_bl, i);
1216 for (i = 0; i < n_exc; ++i)
1217 end_preds[main_end_bl_arity + i] = cf_pred[i];
1218 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1219 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1220 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1226 /* -- Turn CSE back on. -- */
1227 set_optimize(rem_opt);
1228 current_ir_graph = rem;
1233 /********************************************************************/
1234 /* Apply inlining to small methods. */
1235 /********************************************************************/
1237 static struct obstack temp_obst;
1239 /** Represents a possible inlinable call in a graph. */
1240 typedef struct _call_entry {
1241 ir_node *call; /**< The Call node. */
1242 ir_graph *callee; /**< The callee IR-graph. */
1243 list_head list; /**< List head for linking the next one. */
1244 int loop_depth; /**< The loop depth of this call. */
1245 int benefice; /**< The calculated benefice of this call. */
1246 unsigned local_adr:1; /**< Set if this call gets an address of a local variable. */
1247 unsigned all_const:1; /**< Set if this call has only constant parameters. */
1251 * environment for inlining small irgs
1253 typedef struct _inline_env_t {
1254 struct obstack obst; /**< An obstack where call_entries are allocated on. */
1255 list_head calls; /**< The call entry list. */
1259 * Returns the irg called from a Call node. If the irg is not
1260 * known, NULL is returned.
1262 * @param call the call node
1264 static ir_graph *get_call_called_irg(ir_node *call) {
1267 addr = get_Call_ptr(call);
1268 if (is_Global(addr)) {
1269 ir_entity *ent = get_Global_entity(addr);
1270 return get_entity_irg(ent);
1277 * Walker: Collect all calls to known graphs inside a graph.
1279 static void collect_calls(ir_node *call, void *env) {
1281 if (is_Call(call)) {
1282 ir_graph *called_irg = get_call_called_irg(call);
1284 if (called_irg != NULL) {
1285 /* The Call node calls a locally defined method. Remember to inline. */
1286 inline_env_t *ienv = env;
1287 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1289 entry->callee = called_irg;
1290 entry->loop_depth = 0;
1291 entry->benefice = 0;
1292 entry->local_adr = 0;
1293 entry->all_const = 0;
1295 list_add_tail(&entry->list, &ienv->calls);
1301 * Inlines all small methods at call sites where the called address comes
1302 * from a Const node that references the entity representing the called
1304 * The size argument is a rough measure for the code size of the method:
1305 * Methods where the obstack containing the firm graph is smaller than
1308 void inline_small_irgs(ir_graph *irg, int size) {
1309 ir_graph *rem = current_ir_graph;
1313 current_ir_graph = irg;
1314 /* Handle graph state */
1315 assert(get_irg_phase_state(irg) != phase_building);
1316 free_callee_info(irg);
1318 /* Find Call nodes to inline.
1319 (We can not inline during a walk of the graph, as inlining the same
1320 method several times changes the visited flag of the walked graph:
1321 after the first inlining visited of the callee equals visited of
1322 the caller. With the next inlining both are increased.) */
1323 obstack_init(&env.obst);
1324 INIT_LIST_HEAD(&env.calls);
1325 irg_walk_graph(irg, NULL, collect_calls, &env);
1327 if (! list_empty(&env.calls)) {
1328 /* There are calls to inline */
1329 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1330 collect_phiprojs(irg);
1332 list_for_each_entry(call_entry, entry, &env.calls, list) {
1333 ir_graph *callee = entry->callee;
1334 irg_inline_property prop = get_irg_inline_property(callee);
1336 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1337 /* do not inline forbidden / weak graphs */
1341 if (prop >= irg_inline_forced ||
1342 _obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst) < size) {
1343 inline_method(entry->call, callee);
1346 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1348 obstack_free(&env.obst, NULL);
1349 current_ir_graph = rem;
1353 * Environment for inlining irgs.
1356 list_head calls; /**< List of of all call nodes in this graph. */
1357 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1358 unsigned n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1359 unsigned n_blocks; /**< Number of Blocks in graph without Start and End block. */
1360 unsigned n_nodes_orig; /**< for statistics */
1361 unsigned n_call_nodes; /**< Number of Call nodes in the graph. */
1362 unsigned n_call_nodes_orig; /**< for statistics */
1363 unsigned n_callers; /**< Number of known graphs that call this graphs. */
1364 unsigned n_callers_orig; /**< for statistics */
1365 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1366 unsigned local_vars:1; /**< Set, if an inlined function got the address of a local variable. */
1367 unsigned recursive:1; /**< Set, if this function is self recursive. */
1371 * Allocate a new environment for inlining.
1373 static inline_irg_env *alloc_inline_irg_env(void) {
1374 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1375 INIT_LIST_HEAD(&env->calls);
1376 env->local_weights = NULL;
1377 env->n_nodes = -2; /* do not count count Start, End */
1378 env->n_blocks = -2; /* do not count count Start, End Block */
1379 env->n_nodes_orig = -2; /* do not count Start, End */
1380 env->n_call_nodes = 0;
1381 env->n_call_nodes_orig = 0;
1383 env->n_callers_orig = 0;
1384 env->got_inline = 0;
1385 env->local_vars = 0;
1390 typedef struct walker_env {
1391 inline_irg_env *x; /**< the inline environment */
1392 char ignore_runtime; /**< the ignore runtime flag */
1393 char ignore_callers; /**< if set, do change callers data */
1397 * post-walker: collect all calls in the inline-environment
1398 * of a graph and sum some statistics.
1400 static void collect_calls2(ir_node *call, void *ctx) {
1402 inline_irg_env *x = env->x;
1403 ir_opcode code = get_irn_opcode(call);
1407 /* count meaningful nodes in irg */
1408 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1409 if (code != iro_Block) {
1417 if (code != iro_Call) return;
1419 /* check, if it's a runtime call */
1420 if (env->ignore_runtime) {
1421 ir_node *symc = get_Call_ptr(call);
1423 if (is_Global(symc)) {
1424 ir_entity *ent = get_Global_entity(symc);
1426 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1431 /* collect all call nodes */
1433 ++x->n_call_nodes_orig;
1435 callee = get_call_called_irg(call);
1436 if (callee != NULL) {
1437 if (! env->ignore_callers) {
1438 inline_irg_env *callee_env = get_irg_link(callee);
1439 /* count all static callers */
1440 ++callee_env->n_callers;
1441 ++callee_env->n_callers_orig;
1443 if (callee == current_ir_graph)
1446 /* link it in the list of possible inlinable entries */
1447 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1449 entry->callee = callee;
1450 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1451 entry->benefice = 0;
1452 entry->local_adr = 0;
1453 entry->all_const = 0;
1455 list_add_tail(&entry->list, &x->calls);
1460 * Returns TRUE if the number of callers is 0 in the irg's environment,
1461 * hence this irg is a leave.
1463 inline static int is_leave(ir_graph *irg) {
1464 inline_irg_env *env = get_irg_link(irg);
1465 return env->n_call_nodes == 0;
1469 * Returns TRUE if the number of nodes in the callee is
1470 * smaller then size in the irg's environment.
1472 inline static int is_smaller(ir_graph *callee, unsigned size) {
1473 inline_irg_env *env = get_irg_link(callee);
1474 return env->n_nodes < size;
1478 * Duplicate a call entry.
1480 * @param entry the original entry to duplicate
1481 * @param new_call the new call node
1482 * @param loop_depth_delta
1483 * delta value for the loop depth
1485 static call_entry *duplicate_call_entry(const call_entry *entry,
1486 ir_node *new_call, int loop_depth_delta) {
1487 call_entry *nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1488 nentry->call = new_call;
1489 nentry->callee = entry->callee;
1490 nentry->benefice = entry->benefice;
1491 nentry->loop_depth = entry->loop_depth + loop_depth_delta;
1492 nentry->local_adr = entry->local_adr;
1493 nentry->all_const = entry->all_const;
1499 * Append all call nodes of the source environment to the nodes of in the destination
1502 * @param dst destination environment
1503 * @param src source environment
1504 * @param loop_depth the loop depth of the call that is replaced by the src list
1506 static void append_call_list(inline_irg_env *dst, inline_irg_env *src, int loop_depth) {
1507 call_entry *entry, *nentry;
1509 /* Note that the src list points to Call nodes in the inlined graph, but
1510 we need Call nodes in our graph. Luckily the inliner leaves this information
1511 in the link field. */
1512 list_for_each_entry(call_entry, entry, &src->calls, list) {
1513 nentry = duplicate_call_entry(entry, get_irn_link(entry->call), loop_depth);
1514 list_add_tail(&nentry->list, &dst->calls);
1516 dst->n_call_nodes += src->n_call_nodes;
1517 dst->n_nodes += src->n_nodes;
1521 * Inlines small leave methods at call sites where the called address comes
1522 * from a Const node that references the entity representing the called
1524 * The size argument is a rough measure for the code size of the method:
1525 * Methods where the obstack containing the firm graph is smaller than
1528 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
1529 unsigned size, int ignore_runtime)
1531 inline_irg_env *env;
1537 call_entry *entry, *next;
1538 const call_entry *centry;
1539 pmap *copied_graphs;
1540 pmap_entry *pm_entry;
1542 rem = current_ir_graph;
1543 obstack_init(&temp_obst);
1545 /* a map for the copied graphs, used to inline recursive calls */
1546 copied_graphs = pmap_create();
1548 /* extend all irgs by a temporary data structure for inlining. */
1549 n_irgs = get_irp_n_irgs();
1550 for (i = 0; i < n_irgs; ++i)
1551 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1553 /* Pre-compute information in temporary data structure. */
1554 wenv.ignore_runtime = ignore_runtime;
1555 wenv.ignore_callers = 0;
1556 for (i = 0; i < n_irgs; ++i) {
1557 ir_graph *irg = get_irp_irg(i);
1559 assert(get_irg_phase_state(irg) != phase_building);
1560 free_callee_info(irg);
1562 assure_cf_loop(irg);
1563 wenv.x = get_irg_link(irg);
1564 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1567 /* -- and now inline. -- */
1569 /* Inline leaves recursively -- we might construct new leaves. */
1573 for (i = 0; i < n_irgs; ++i) {
1575 int phiproj_computed = 0;
1577 current_ir_graph = get_irp_irg(i);
1578 env = get_irg_link(current_ir_graph);
1580 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1581 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1583 irg_inline_property prop;
1585 if (env->n_nodes > maxsize)
1589 callee = entry->callee;
1591 prop = get_irg_inline_property(callee);
1592 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1593 /* do not inline forbidden / weak graphs */
1597 if (is_leave(callee) && (
1598 is_smaller(callee, leavesize) || prop >= irg_inline_forced)) {
1599 if (!phiproj_computed) {
1600 phiproj_computed = 1;
1601 collect_phiprojs(current_ir_graph);
1603 did_inline = inline_method(call, callee);
1606 inline_irg_env *callee_env = get_irg_link(callee);
1608 /* call was inlined, Phi/Projs for current graph must be recomputed */
1609 phiproj_computed = 0;
1611 /* Do some statistics */
1612 env->got_inline = 1;
1613 --env->n_call_nodes;
1614 env->n_nodes += callee_env->n_nodes;
1615 --callee_env->n_callers;
1617 /* remove this call from the list */
1618 list_del(&entry->list);
1623 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1625 } while (did_inline);
1627 /* inline other small functions. */
1628 for (i = 0; i < n_irgs; ++i) {
1630 int phiproj_computed = 0;
1632 current_ir_graph = get_irp_irg(i);
1633 env = get_irg_link(current_ir_graph);
1635 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1637 /* note that the list of possible calls is updated during the process */
1638 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1639 irg_inline_property prop;
1644 callee = entry->callee;
1646 prop = get_irg_inline_property(callee);
1647 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1648 /* do not inline forbidden / weak graphs */
1652 e = pmap_find(copied_graphs, callee);
1655 * Remap callee if we have a copy.
1656 * FIXME: Should we do this only for recursive Calls ?
1661 if (prop >= irg_inline_forced ||
1662 (is_smaller(callee, size) && env->n_nodes < maxsize) /* small function */) {
1663 if (current_ir_graph == callee) {
1665 * Recursive call: we cannot directly inline because we cannot walk
1666 * the graph and change it. So we have to make a copy of the graph
1670 inline_irg_env *callee_env;
1673 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1676 * No copy yet, create one.
1677 * Note that recursive methods are never leaves, so it is sufficient
1678 * to test this condition here.
1680 copy = create_irg_copy(callee);
1682 /* create_irg_copy() destroys the Proj links, recompute them */
1683 phiproj_computed = 0;
1685 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1687 /* allocate new environment */
1688 callee_env = alloc_inline_irg_env();
1689 set_irg_link(copy, callee_env);
1691 assure_cf_loop(copy);
1692 wenv.x = callee_env;
1693 wenv.ignore_callers = 1;
1694 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1697 * Enter the entity of the original graph. This is needed
1698 * for inline_method(). However, note that ent->irg still points
1699 * to callee, NOT to copy.
1701 set_irg_entity(copy, get_irg_entity(callee));
1703 pmap_insert(copied_graphs, callee, copy);
1706 /* we have only one caller: the original graph */
1707 callee_env->n_callers = 1;
1708 callee_env->n_callers_orig = 1;
1710 if (! phiproj_computed) {
1711 phiproj_computed = 1;
1712 collect_phiprojs(current_ir_graph);
1714 did_inline = inline_method(call, callee);
1716 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1718 /* call was inlined, Phi/Projs for current graph must be recomputed */
1719 phiproj_computed = 0;
1721 /* callee was inline. Append it's call list. */
1722 env->got_inline = 1;
1723 --env->n_call_nodes;
1724 append_call_list(env, callee_env, entry->loop_depth);
1725 --callee_env->n_callers;
1727 /* after we have inlined callee, all called methods inside callee
1728 are now called once more */
1729 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
1730 inline_irg_env *penv = get_irg_link(centry->callee);
1734 /* remove this call from the list */
1735 list_del(&entry->list);
1740 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1743 for (i = 0; i < n_irgs; ++i) {
1744 irg = get_irp_irg(i);
1745 env = get_irg_link(irg);
1747 if (env->got_inline) {
1748 optimize_graph_df(irg);
1751 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1752 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1753 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1754 env->n_callers_orig, env->n_callers,
1755 get_entity_name(get_irg_entity(irg))));
1759 /* kill the copied graphs: we don't need them anymore */
1760 foreach_pmap(copied_graphs, pm_entry) {
1761 ir_graph *copy = pm_entry->value;
1763 /* reset the entity, otherwise it will be deleted in the next step ... */
1764 set_irg_entity(copy, NULL);
1765 free_ir_graph(copy);
1767 pmap_destroy(copied_graphs);
1769 obstack_free(&temp_obst, NULL);
1770 current_ir_graph = rem;
1774 * Calculate the parameter weights for transmitting the address of a local variable.
1776 static unsigned calc_method_local_weight(ir_node *arg) {
1778 unsigned v, weight = 0;
1780 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1781 ir_node *succ = get_irn_out(arg, i);
1783 switch (get_irn_opcode(succ)) {
1786 /* Loads and Store can be removed */
1790 /* check if all args are constant */
1791 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1792 ir_node *idx = get_Sel_index(succ, j);
1793 if (! is_Const(idx))
1796 /* Check users on this Sel. Note: if a 0 is returned here, there was
1797 some unsupported node. */
1798 v = calc_method_local_weight(succ);
1801 /* we can kill one Sel with constant indexes, this is cheap */
1805 /* when looking backward we might find Id nodes */
1806 weight += calc_method_local_weight(succ);
1809 /* unoptimized tuple */
1810 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1811 ir_node *pred = get_Tuple_pred(succ, j);
1813 /* look for Proj(j) */
1814 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1815 ir_node *succ_succ = get_irn_out(succ, k);
1816 if (is_Proj(succ_succ)) {
1817 if (get_Proj_proj(succ_succ) == j) {
1819 weight += calc_method_local_weight(succ_succ);
1822 /* this should NOT happen */
1830 /* any other node: unsupported yet or bad. */
1838 * Calculate the parameter weights for transmitting the address of a local variable.
1840 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1841 ir_entity *ent = get_irg_entity(irg);
1843 int nparams, i, proj_nr;
1844 ir_node *irg_args, *arg;
1846 mtp = get_entity_type(ent);
1847 nparams = get_method_n_params(mtp);
1849 /* allocate a new array. currently used as 'analysed' flag */
1850 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1852 /* If the method haven't parameters we have nothing to do. */
1856 assure_irg_outs(irg);
1857 irg_args = get_irg_args(irg);
1858 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1859 arg = get_irn_out(irg_args, i);
1860 proj_nr = get_Proj_proj(arg);
1861 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1866 * Calculate the benefice for transmitting an local variable address.
1867 * After inlining, the local variable might be transformed into a
1868 * SSA variable by scalar_replacement().
1870 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1871 inline_irg_env *env = get_irg_link(callee);
1873 if (env->local_weights != NULL) {
1874 if (pos < ARR_LEN(env->local_weights))
1875 return env->local_weights[pos];
1879 analyze_irg_local_weights(env, callee);
1881 if (pos < ARR_LEN(env->local_weights))
1882 return env->local_weights[pos];
1887 * Calculate a benefice value for inlining the given call.
1889 * @param call the call node we have to inspect
1890 * @param callee the called graph
1892 static int calc_inline_benefice(call_entry *entry, ir_graph *callee)
1894 ir_node *call = entry->call;
1895 ir_entity *ent = get_irg_entity(callee);
1899 int i, n_params, all_const;
1901 irg_inline_property prop;
1903 inline_irg_env *callee_env;
1905 prop = get_irg_inline_property(callee);
1906 if (prop == irg_inline_forbidden) {
1907 DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden\n",
1909 return entry->benefice = INT_MIN;
1912 if (get_irg_additional_properties(callee) & (mtp_property_noreturn | mtp_property_weak)) {
1913 DB((dbg, LEVEL_2, "In %+F Call to %+F: not inlining noreturn or weak\n",
1915 return entry->benefice = INT_MIN;
1918 /* costs for every passed parameter */
1919 n_params = get_Call_n_params(call);
1920 mtp = get_entity_type(ent);
1921 cc = get_method_calling_convention(mtp);
1922 if (cc & cc_reg_param) {
1923 /* register parameter, smaller costs for register parameters */
1924 int max_regs = cc & ~cc_bits;
1926 if (max_regs < n_params)
1927 weight += max_regs * 2 + (n_params - max_regs) * 5;
1929 weight += n_params * 2;
1931 /* parameters are passed an stack */
1932 weight += 5 * n_params;
1935 /* constant parameters improve the benefice */
1936 frame_ptr = get_irg_frame(current_ir_graph);
1938 for (i = 0; i < n_params; ++i) {
1939 ir_node *param = get_Call_param(call, i);
1941 if (is_Const(param)) {
1942 weight += get_method_param_weight(ent, i);
1945 if (is_SymConst(param))
1946 weight += get_method_param_weight(ent, i);
1947 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1949 * An address of a local variable is transmitted. After
1950 * inlining, scalar_replacement might be able to remove the
1951 * local variable, so honor this.
1953 v = get_method_local_adress_weight(callee, i);
1956 entry->local_adr = 1;
1960 entry->all_const = all_const;
1962 callee_env = get_irg_link(callee);
1963 if (callee_env->n_callers == 1 &&
1964 callee != current_ir_graph &&
1965 get_entity_visibility(ent) == visibility_local) {
1969 /* give a bonus for functions with one block */
1970 if (callee_env->n_blocks == 1)
1971 weight = weight * 3 / 2;
1973 /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1974 if (callee_env->n_nodes < 30 && !callee_env->recursive)
1977 /* and finally for leaves: they do not increase the register pressure
1978 because of callee safe registers */
1979 if (callee_env->n_call_nodes == 0)
1982 /** it's important to inline inner loops first */
1983 if (entry->loop_depth > 30)
1984 weight += 30 * 1024;
1986 weight += entry->loop_depth * 1024;
1989 * All arguments constant is probably a good sign, give an extra bonus
1994 return entry->benefice = weight;
1997 static ir_graph **irgs;
1998 static int last_irg;
2001 * Callgraph walker, collect all visited graphs.
2003 static void callgraph_walker(ir_graph *irg, void *data) {
2005 irgs[last_irg++] = irg;
2009 * Creates an inline order for all graphs.
2011 * @return the list of graphs.
2013 static ir_graph **create_irg_list(void) {
2014 ir_entity **free_methods;
2016 int n_irgs = get_irp_n_irgs();
2018 cgana(&arr_len, &free_methods);
2019 xfree(free_methods);
2021 compute_callgraph();
2024 irgs = XMALLOCNZ(ir_graph*, n_irgs);
2026 callgraph_walk(NULL, callgraph_walker, NULL);
2027 assert(n_irgs == last_irg);
2033 * Push a call onto the priority list if its benefice is big enough.
2035 * @param pqueue the priority queue of calls
2036 * @param call the call entry
2037 * @param inlien_threshold
2038 * the threshold value
2040 static void maybe_push_call(pqueue_t *pqueue, call_entry *call,
2041 int inline_threshold)
2043 ir_graph *callee = call->callee;
2044 irg_inline_property prop = get_irg_inline_property(callee);
2045 int benefice = calc_inline_benefice(call, callee);
2047 DB((dbg, LEVEL_2, "In %+F Call %+F to %+F has benefice %d\n",
2048 get_irn_irg(call->call), call->call, callee, benefice));
2050 if (prop < irg_inline_forced && benefice < inline_threshold) {
2054 pqueue_put(pqueue, call, benefice);
2058 * Try to inline calls into a graph.
2060 * @param irg the graph into which we inline
2061 * @param maxsize do NOT inline if the size of irg gets
2062 * bigger than this amount
2063 * @param inline_threshold
2064 * threshold value for inline decision
2065 * @param copied_graphs
2066 * map containing copied of recursive graphs
2068 static void inline_into(ir_graph *irg, unsigned maxsize,
2069 int inline_threshold, pmap *copied_graphs)
2071 int phiproj_computed = 0;
2072 inline_irg_env *env = get_irg_link(irg);
2073 call_entry *curr_call;
2077 if (env->n_call_nodes == 0)
2080 if (env->n_nodes > maxsize) {
2081 DB((dbg, LEVEL_2, "%+F: too big (%d)\n", irg, env->n_nodes));
2085 current_ir_graph = irg;
2086 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2088 /* put irgs into the pqueue */
2089 pqueue = new_pqueue();
2091 list_for_each_entry(call_entry, curr_call, &env->calls, list) {
2092 assert(is_Call(curr_call->call));
2093 maybe_push_call(pqueue, curr_call, inline_threshold);
2096 /* note that the list of possible calls is updated during the process */
2097 while (!pqueue_empty(pqueue)) {
2099 call_entry *curr_call = pqueue_pop_front(pqueue);
2100 ir_graph *callee = curr_call->callee;
2101 ir_node *call_node = curr_call->call;
2102 inline_irg_env *callee_env = get_irg_link(callee);
2103 irg_inline_property prop = get_irg_inline_property(callee);
2105 const call_entry *centry;
2108 if ((prop < irg_inline_forced) && env->n_nodes + callee_env->n_nodes > maxsize) {
2109 DB((dbg, LEVEL_2, "%+F: too big (%d) + %+F (%d)\n", irg,
2110 env->n_nodes, callee, callee_env->n_nodes));
2114 e = pmap_find(copied_graphs, callee);
2116 int benefice = curr_call->benefice;
2118 * Reduce the weight for recursive function IFF not all arguments are const.
2119 * inlining recursive functions is rarely good.
2121 if (!curr_call->all_const)
2123 if (benefice < inline_threshold)
2127 * Remap callee if we have a copy.
2130 callee_env = get_irg_link(callee);
2133 if (current_ir_graph == callee) {
2135 * Recursive call: we cannot directly inline because we cannot
2136 * walk the graph and change it. So we have to make a copy of
2139 int benefice = curr_call->benefice;
2143 * Reduce the weight for recursive function IFF not all arguments are const.
2144 * inlining recursive functions is rarely good.
2146 if (!curr_call->all_const)
2148 if (benefice < inline_threshold)
2151 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2154 * No copy yet, create one.
2155 * Note that recursive methods are never leaves, so it is
2156 * sufficient to test this condition here.
2158 copy = create_irg_copy(callee);
2160 /* create_irg_copy() destroys the Proj links, recompute them */
2161 phiproj_computed = 0;
2163 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2165 /* allocate a new environment */
2166 callee_env = alloc_inline_irg_env();
2167 set_irg_link(copy, callee_env);
2169 assure_cf_loop(copy);
2170 wenv.x = callee_env;
2171 wenv.ignore_callers = 1;
2172 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2175 * Enter the entity of the original graph. This is needed
2176 * for inline_method(). However, note that ent->irg still points
2177 * to callee, NOT to copy.
2179 set_irg_entity(copy, get_irg_entity(callee));
2181 pmap_insert(copied_graphs, callee, copy);
2184 /* we have only one caller: the original graph */
2185 callee_env->n_callers = 1;
2186 callee_env->n_callers_orig = 1;
2188 if (! phiproj_computed) {
2189 phiproj_computed = 1;
2190 collect_phiprojs(current_ir_graph);
2192 did_inline = inline_method(call_node, callee);
2196 /* call was inlined, Phi/Projs for current graph must be recomputed */
2197 phiproj_computed = 0;
2199 /* remove it from the caller list */
2200 list_del(&curr_call->list);
2202 /* callee was inline. Append it's call list. */
2203 env->got_inline = 1;
2204 if (curr_call->local_adr)
2205 env->local_vars = 1;
2206 --env->n_call_nodes;
2208 /* we just generate a bunch of new calls */
2209 loop_depth = curr_call->loop_depth;
2210 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
2211 inline_irg_env *penv = get_irg_link(centry->callee);
2213 call_entry *new_entry;
2215 /* after we have inlined callee, all called methods inside
2216 * callee are now called once more */
2219 /* Note that the src list points to Call nodes in the inlined graph,
2220 * but we need Call nodes in our graph. Luckily the inliner leaves
2221 * this information in the link field. */
2222 new_call = get_irn_link(centry->call);
2223 assert(is_Call(new_call));
2225 new_entry = duplicate_call_entry(centry, new_call, loop_depth);
2226 list_add_tail(&new_entry->list, &env->calls);
2227 maybe_push_call(pqueue, new_entry, inline_threshold);
2230 env->n_call_nodes += callee_env->n_call_nodes;
2231 env->n_nodes += callee_env->n_nodes;
2232 --callee_env->n_callers;
2234 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2239 * Heuristic inliner. Calculates a benefice value for every call and inlines
2240 * those calls with a value higher than the threshold.
2242 void inline_functions(unsigned maxsize, int inline_threshold) {
2243 inline_irg_env *env;
2247 pmap *copied_graphs;
2248 pmap_entry *pm_entry;
2251 rem = current_ir_graph;
2252 obstack_init(&temp_obst);
2254 irgs = create_irg_list();
2256 /* a map for the copied graphs, used to inline recursive calls */
2257 copied_graphs = pmap_create();
2259 /* extend all irgs by a temporary data structure for inlining. */
2260 n_irgs = get_irp_n_irgs();
2261 for (i = 0; i < n_irgs; ++i)
2262 set_irg_link(irgs[i], alloc_inline_irg_env());
2264 /* Pre-compute information in temporary data structure. */
2265 wenv.ignore_runtime = 0;
2266 wenv.ignore_callers = 0;
2267 for (i = 0; i < n_irgs; ++i) {
2268 ir_graph *irg = irgs[i];
2270 free_callee_info(irg);
2272 wenv.x = get_irg_link(irg);
2273 assure_cf_loop(irg);
2274 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
2277 /* -- and now inline. -- */
2278 for (i = 0; i < n_irgs; ++i) {
2279 ir_graph *irg = irgs[i];
2281 inline_into(irg, maxsize, inline_threshold, copied_graphs);
2284 for (i = 0; i < n_irgs; ++i) {
2285 ir_graph *irg = irgs[i];
2287 env = get_irg_link(irg);
2288 if (env->got_inline) {
2289 /* this irg got calls inlined: optimize it */
2290 if (get_opt_combo()) {
2291 if (env->local_vars) {
2292 scalar_replacement_opt(irg);
2296 if (env->local_vars) {
2297 if (scalar_replacement_opt(irg)) {
2298 optimize_graph_df(irg);
2304 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2305 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2306 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2307 env->n_callers_orig, env->n_callers,
2308 get_entity_name(get_irg_entity(irg))));
2312 /* kill the copied graphs: we don't need them anymore */
2313 foreach_pmap(copied_graphs, pm_entry) {
2314 ir_graph *copy = pm_entry->value;
2316 /* reset the entity, otherwise it will be deleted in the next step ... */
2317 set_irg_entity(copy, NULL);
2318 free_ir_graph(copy);
2320 pmap_destroy(copied_graphs);
2324 obstack_free(&temp_obst, NULL);
2325 current_ir_graph = rem;
2328 void firm_init_inline(void) {
2329 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");