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 ir_graph *irg = current_ir_graph;
773 if (get_Sel_ptr(node) == get_irg_frame(irg)) {
774 /* access to frame */
775 ir_entity *ent = get_Sel_entity(node);
776 if (get_entity_owner(ent) != get_irg_frame_type(irg)) {
777 /* access to value_type */
781 } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
783 * Refuse to inline alloca call unless user explicitly forced so as this
784 * may change program's memory overhead drastically when the function
785 * using alloca is called in loop. In GCC present in SPEC2000 inlining
786 * into schedule_block cause it to require 2GB of ram instead of 256MB.
788 * Sorrily this is true with our implementation also.
789 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
790 * disables inlining of functions using VLA (with are completely save).
793 * - add a flag to the Alloc node for "real" alloca() calls
794 * - add a new Stack-Restore node at the end of a function using alloca()
801 * Check if we can inline a given call.
802 * Currently, we cannot inline two cases:
803 * - call with compound arguments
804 * - graphs that take the address of a parameter
806 * check these conditions here
808 static int can_inline(ir_node *call, ir_graph *called_graph) {
809 ir_type *call_type = get_Call_type(call);
810 int params, ress, i, res;
811 assert(is_Method_type(call_type));
813 params = get_method_n_params(call_type);
814 ress = get_method_n_ress(call_type);
816 /* check parameters for compound arguments */
817 for (i = 0; i < params; ++i) {
818 ir_type *p_type = get_method_param_type(call_type, i);
820 if (is_compound_type(p_type))
824 /* check results for compound arguments */
825 for (i = 0; i < ress; ++i) {
826 ir_type *r_type = get_method_res_type(call_type, i);
828 if (is_compound_type(r_type))
833 irg_walk_graph(called_graph, find_addr, NULL, &res);
839 exc_handler = 0, /**< There is a handler. */
840 exc_to_end = 1, /**< Branches to End. */
841 exc_no_handler = 2 /**< Exception handling not represented. */
844 /* Inlines a method at the given call site. */
845 int inline_method(ir_node *call, ir_graph *called_graph) {
847 ir_node *post_call, *post_bl;
848 ir_node *in[pn_Start_max];
849 ir_node *end, *end_bl, *block;
854 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
855 enum exc_mode exc_handling;
856 ir_type *called_frame, *curr_frame, *mtp, *ctp;
859 irg_inline_property prop = get_irg_inline_property(called_graph);
860 unsigned long visited;
862 if (prop == irg_inline_forbidden)
865 ent = get_irg_entity(called_graph);
867 mtp = get_entity_type(ent);
868 ctp = get_Call_type(call);
869 if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
870 /* this is a bad feature of C: without a prototype, we can
871 * call a function with less parameters than needed. Currently
872 * we don't support this, although we could use Unknown than. */
876 /* Argh, compiling C has some bad consequences:
877 * It is implementation dependent what happens in that case.
878 * We support inlining, if the bitsize of the types matches AND
879 * the same arithmetic is used. */
880 n_params = get_method_n_params(mtp);
881 for (i = n_params - 1; i >= 0; --i) {
882 ir_type *param_tp = get_method_param_type(mtp, i);
883 ir_type *arg_tp = get_method_param_type(ctp, i);
885 if (param_tp != arg_tp) {
886 ir_mode *pmode = get_type_mode(param_tp);
887 ir_mode *amode = get_type_mode(arg_tp);
889 if (pmode == NULL || amode == NULL)
891 if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
893 if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
895 /* otherwise we can simply "reinterpret" the bits */
899 irg = get_irn_irg(call);
902 * We cannot inline a recursive call. The graph must be copied before
903 * the call the inline_method() using create_irg_copy().
905 if (called_graph == irg)
909 * currently, we cannot inline two cases:
910 * - call with compound arguments
911 * - graphs that take the address of a parameter
913 if (! can_inline(call, called_graph))
916 rem = current_ir_graph;
917 current_ir_graph = irg;
919 DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
921 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
922 rem_opt = get_opt_optimize();
925 /* Handle graph state */
926 assert(get_irg_phase_state(irg) != phase_building);
927 assert(get_irg_pinned(irg) == op_pin_state_pinned);
928 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
929 set_irg_outs_inconsistent(irg);
930 set_irg_extblk_inconsistent(irg);
931 set_irg_doms_inconsistent(irg);
932 set_irg_loopinfo_inconsistent(irg);
933 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
934 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
936 /* -- Check preconditions -- */
937 assert(is_Call(call));
939 /* here we know we WILL inline, so inform the statistics */
940 hook_inline(call, called_graph);
942 /* -- Decide how to handle exception control flow: Is there a handler
943 for the Call node, or do we branch directly to End on an exception?
945 0 There is a handler.
947 2 Exception handling not represented in Firm. -- */
949 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
950 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
951 long proj_nr = get_Proj_proj(proj);
952 if (proj_nr == pn_Call_X_except) Xproj = proj;
953 if (proj_nr == pn_Call_M_except) Mproj = proj;
955 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
956 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
957 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
960 /* create the argument tuple */
961 NEW_ARR_A(ir_type *, args_in, n_params);
963 block = get_nodes_block(call);
964 for (i = n_params - 1; i >= 0; --i) {
965 ir_node *arg = get_Call_param(call, i);
966 ir_type *param_tp = get_method_param_type(mtp, i);
967 ir_mode *mode = get_type_mode(param_tp);
969 if (mode != get_irn_mode(arg)) {
970 arg = new_r_Conv(irg, block, arg, mode, 0);
976 the procedure and later replaces the Start node of the called graph.
977 Post_call is the old Call node and collects the results of the called
978 graph. Both will end up being a tuple. -- */
979 post_bl = get_nodes_block(call);
980 set_irg_current_block(irg, post_bl);
981 /* XxMxPxPxPxT of Start + parameter of Call */
982 in[pn_Start_X_initial_exec] = new_Jmp();
983 in[pn_Start_M] = get_Call_mem(call);
984 in[pn_Start_P_frame_base] = get_irg_frame(irg);
985 in[pn_Start_P_tls] = get_irg_tls(irg);
986 in[pn_Start_T_args] = new_Tuple(n_params, args_in);
987 pre_call = new_Tuple(pn_Start_max, in);
991 The new block gets the ins of the old block, pre_call and all its
992 predecessors and all Phi nodes. -- */
993 part_block(pre_call);
995 /* -- Prepare state for dead node elimination -- */
996 /* Visited flags in calling irg must be >= flag in called irg.
997 Else walker and arity computation will not work. */
998 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
999 set_irg_visited(irg, get_irg_visited(called_graph) + 1);
1000 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
1001 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
1002 visited = get_irg_visited(irg);
1004 /* Set pre_call as new Start node in link field of the start node of
1005 calling graph and pre_calls block as new block for the start block
1007 Further mark these nodes so that they are not visited by the
1009 set_irn_link(get_irg_start(called_graph), pre_call);
1010 set_irn_visited(get_irg_start(called_graph), visited);
1011 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1012 set_irn_visited(get_irg_start_block(called_graph), visited);
1014 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1015 set_irn_visited(get_irg_bad(called_graph), visited);
1017 set_irn_link(get_irg_no_mem(called_graph), get_irg_no_mem(current_ir_graph));
1018 set_irn_visited(get_irg_no_mem(called_graph), visited);
1020 /* Initialize for compaction of in arrays */
1021 inc_irg_block_visited(irg);
1023 /* -- Replicate local entities of the called_graph -- */
1024 /* copy the entities. */
1025 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1026 called_frame = get_irg_frame_type(called_graph);
1027 curr_frame = get_irg_frame_type(irg);
1028 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
1029 ir_entity *new_ent, *old_ent;
1030 old_ent = get_class_member(called_frame, i);
1031 new_ent = copy_entity_own(old_ent, curr_frame);
1032 set_entity_link(old_ent, new_ent);
1035 /* visited is > than that of called graph. With this trick visited will
1036 remain unchanged so that an outer walker, e.g., searching the call nodes
1037 to inline, calling this inline will not visit the inlined nodes. */
1038 set_irg_visited(irg, get_irg_visited(irg)-1);
1040 /* -- Performing dead node elimination inlines the graph -- */
1041 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1043 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
1044 get_irg_frame_type(called_graph));
1046 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1048 /* Repair called_graph */
1049 set_irg_visited(called_graph, get_irg_visited(irg));
1050 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
1051 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1053 /* -- Merge the end of the inlined procedure with the call site -- */
1054 /* We will turn the old Call node into a Tuple with the following
1057 0: Phi of all Memories of Return statements.
1058 1: Jmp from new Block that merges the control flow from all exception
1059 predecessors of the old end block.
1060 2: Tuple of all arguments.
1061 3: Phi of Exception memories.
1062 In case the old Call directly branches to End on an exception we don't
1063 need the block merging all exceptions nor the Phi of the exception
1067 /* -- Precompute some values -- */
1068 end_bl = get_new_node(get_irg_end_block(called_graph));
1069 end = get_new_node(get_irg_end(called_graph));
1070 arity = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret */
1071 n_res = get_method_n_ress(get_Call_type(call));
1073 res_pred = XMALLOCN(ir_node*, n_res);
1074 cf_pred = XMALLOCN(ir_node*, arity);
1076 set_irg_current_block(irg, post_bl); /* just to make sure */
1078 /* -- archive keepalives -- */
1079 irn_arity = get_irn_arity(end);
1080 for (i = 0; i < irn_arity; i++) {
1081 ir_node *ka = get_End_keepalive(end, i);
1083 add_End_keepalive(get_irg_end(irg), ka);
1086 /* The new end node will die. We need not free as the in array is on the obstack:
1087 copy_node() only generated 'D' arrays. */
1089 /* -- Replace Return nodes by Jump nodes. -- */
1091 for (i = 0; i < arity; i++) {
1093 ret = get_Block_cfgpred(end_bl, i);
1094 if (is_Return(ret)) {
1095 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1099 set_irn_in(post_bl, n_ret, cf_pred);
1101 /* -- Build a Tuple for all results of the method.
1102 Add Phi node if there was more than one Return. -- */
1103 turn_into_tuple(post_call, pn_Call_max);
1104 /* First the Memory-Phi */
1106 for (i = 0; i < arity; i++) {
1107 ret = get_Block_cfgpred(end_bl, i);
1108 if (is_Return(ret)) {
1109 cf_pred[n_ret] = get_Return_mem(ret);
1113 phi = new_Phi(n_ret, cf_pred, mode_M);
1114 set_Tuple_pred(call, pn_Call_M_regular, phi);
1115 /* Conserve Phi-list for further inlinings -- but might be optimized */
1116 if (get_nodes_block(phi) == post_bl) {
1117 set_irn_link(phi, get_irn_link(post_bl));
1118 set_irn_link(post_bl, phi);
1120 /* Now the real results */
1122 for (j = 0; j < n_res; j++) {
1124 for (i = 0; i < arity; i++) {
1125 ret = get_Block_cfgpred(end_bl, i);
1126 if (is_Return(ret)) {
1127 cf_pred[n_ret] = get_Return_res(ret, j);
1132 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1136 /* Conserve Phi-list for further inlinings -- but might be optimized */
1137 if (get_nodes_block(phi) == post_bl) {
1138 set_Phi_next(phi, get_Block_phis(post_bl));
1139 set_Block_phis(post_bl, phi);
1142 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1144 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1146 /* handle the regular call */
1147 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1149 /* For now, we cannot inline calls with value_base */
1150 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1152 /* Finally the exception control flow.
1153 We have two (three) possible situations:
1154 First if the Call branches to an exception handler: We need to add a Phi node to
1155 collect the memory containing the exception objects. Further we need
1156 to add another block to get a correct representation of this Phi. To
1157 this block we add a Jmp that resolves into the X output of the Call
1158 when the Call is turned into a tuple.
1159 Second the Call branches to End, the exception is not handled. Just
1160 add all inlined exception branches to the End node.
1161 Third: there is no Exception edge at all. Handle as case two. */
1162 if (exc_handling == exc_handler) {
1164 for (i = 0; i < arity; i++) {
1166 ret = get_Block_cfgpred(end_bl, i);
1167 irn = skip_Proj(ret);
1168 if (is_fragile_op(irn) || is_Raise(irn)) {
1169 cf_pred[n_exc] = ret;
1174 ir_node *block = new_Block(n_exc, cf_pred);
1175 set_cur_block(block);
1177 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1178 /* The Phi for the memories with the exception objects */
1180 for (i = 0; i < arity; i++) {
1182 ret = skip_Proj(get_Block_cfgpred(end_bl, i));
1184 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1186 } else if (is_fragile_op(ret)) {
1187 /* We rely that all cfops have the memory output at the same position. */
1188 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1190 } else if (is_Raise(ret)) {
1191 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1195 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1197 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1198 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1201 ir_node *main_end_bl;
1202 int main_end_bl_arity;
1203 ir_node **end_preds;
1205 /* assert(exc_handling == 1 || no exceptions. ) */
1207 for (i = 0; i < arity; i++) {
1208 ir_node *ret = get_Block_cfgpred(end_bl, i);
1209 ir_node *irn = skip_Proj(ret);
1211 if (is_fragile_op(irn) || is_Raise(irn)) {
1212 cf_pred[n_exc] = ret;
1216 main_end_bl = get_irg_end_block(irg);
1217 main_end_bl_arity = get_irn_arity(main_end_bl);
1218 end_preds = XMALLOCN(ir_node*, n_exc + main_end_bl_arity);
1220 for (i = 0; i < main_end_bl_arity; ++i)
1221 end_preds[i] = get_irn_n(main_end_bl, i);
1222 for (i = 0; i < n_exc; ++i)
1223 end_preds[main_end_bl_arity + i] = cf_pred[i];
1224 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1225 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1226 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1232 /* -- Turn CSE back on. -- */
1233 set_optimize(rem_opt);
1234 current_ir_graph = rem;
1239 /********************************************************************/
1240 /* Apply inlining to small methods. */
1241 /********************************************************************/
1243 static struct obstack temp_obst;
1245 /** Represents a possible inlinable call in a graph. */
1246 typedef struct _call_entry {
1247 ir_node *call; /**< The Call node. */
1248 ir_graph *callee; /**< The callee IR-graph. */
1249 list_head list; /**< List head for linking the next one. */
1250 int loop_depth; /**< The loop depth of this call. */
1251 int benefice; /**< The calculated benefice of this call. */
1252 unsigned local_adr:1; /**< Set if this call gets an address of a local variable. */
1253 unsigned all_const:1; /**< Set if this call has only constant parameters. */
1257 * environment for inlining small irgs
1259 typedef struct _inline_env_t {
1260 struct obstack obst; /**< An obstack where call_entries are allocated on. */
1261 list_head calls; /**< The call entry list. */
1265 * Returns the irg called from a Call node. If the irg is not
1266 * known, NULL is returned.
1268 * @param call the call node
1270 static ir_graph *get_call_called_irg(ir_node *call) {
1273 addr = get_Call_ptr(call);
1274 if (is_Global(addr)) {
1275 ir_entity *ent = get_Global_entity(addr);
1276 return get_entity_irg(ent);
1283 * Walker: Collect all calls to known graphs inside a graph.
1285 static void collect_calls(ir_node *call, void *env) {
1287 if (is_Call(call)) {
1288 ir_graph *called_irg = get_call_called_irg(call);
1290 if (called_irg != NULL) {
1291 /* The Call node calls a locally defined method. Remember to inline. */
1292 inline_env_t *ienv = env;
1293 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1295 entry->callee = called_irg;
1296 entry->loop_depth = 0;
1297 entry->benefice = 0;
1298 entry->local_adr = 0;
1299 entry->all_const = 0;
1301 list_add_tail(&entry->list, &ienv->calls);
1307 * Inlines all small methods at call sites where the called address comes
1308 * from a Const node that references the entity representing the called
1310 * The size argument is a rough measure for the code size of the method:
1311 * Methods where the obstack containing the firm graph is smaller than
1314 void inline_small_irgs(ir_graph *irg, int size) {
1315 ir_graph *rem = current_ir_graph;
1319 current_ir_graph = irg;
1320 /* Handle graph state */
1321 assert(get_irg_phase_state(irg) != phase_building);
1322 free_callee_info(irg);
1324 /* Find Call nodes to inline.
1325 (We can not inline during a walk of the graph, as inlining the same
1326 method several times changes the visited flag of the walked graph:
1327 after the first inlining visited of the callee equals visited of
1328 the caller. With the next inlining both are increased.) */
1329 obstack_init(&env.obst);
1330 INIT_LIST_HEAD(&env.calls);
1331 irg_walk_graph(irg, NULL, collect_calls, &env);
1333 if (! list_empty(&env.calls)) {
1334 /* There are calls to inline */
1335 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1336 collect_phiprojs(irg);
1338 list_for_each_entry(call_entry, entry, &env.calls, list) {
1339 ir_graph *callee = entry->callee;
1340 irg_inline_property prop = get_irg_inline_property(callee);
1342 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1343 /* do not inline forbidden / weak graphs */
1347 if (prop >= irg_inline_forced ||
1348 _obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst) < size) {
1349 inline_method(entry->call, callee);
1352 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1354 obstack_free(&env.obst, NULL);
1355 current_ir_graph = rem;
1359 * Environment for inlining irgs.
1362 list_head calls; /**< List of of all call nodes in this graph. */
1363 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1364 unsigned n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1365 unsigned n_blocks; /**< Number of Blocks in graph without Start and End block. */
1366 unsigned n_nodes_orig; /**< for statistics */
1367 unsigned n_call_nodes; /**< Number of Call nodes in the graph. */
1368 unsigned n_call_nodes_orig; /**< for statistics */
1369 unsigned n_callers; /**< Number of known graphs that call this graphs. */
1370 unsigned n_callers_orig; /**< for statistics */
1371 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1372 unsigned local_vars:1; /**< Set, if an inlined function got the address of a local variable. */
1373 unsigned recursive:1; /**< Set, if this function is self recursive. */
1377 * Allocate a new environment for inlining.
1379 static inline_irg_env *alloc_inline_irg_env(void) {
1380 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1381 INIT_LIST_HEAD(&env->calls);
1382 env->local_weights = NULL;
1383 env->n_nodes = -2; /* do not count count Start, End */
1384 env->n_blocks = -2; /* do not count count Start, End Block */
1385 env->n_nodes_orig = -2; /* do not count Start, End */
1386 env->n_call_nodes = 0;
1387 env->n_call_nodes_orig = 0;
1389 env->n_callers_orig = 0;
1390 env->got_inline = 0;
1391 env->local_vars = 0;
1396 typedef struct walker_env {
1397 inline_irg_env *x; /**< the inline environment */
1398 char ignore_runtime; /**< the ignore runtime flag */
1399 char ignore_callers; /**< if set, do change callers data */
1403 * post-walker: collect all calls in the inline-environment
1404 * of a graph and sum some statistics.
1406 static void collect_calls2(ir_node *call, void *ctx) {
1408 inline_irg_env *x = env->x;
1409 ir_opcode code = get_irn_opcode(call);
1413 /* count meaningful nodes in irg */
1414 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1415 if (code != iro_Block) {
1423 if (code != iro_Call) return;
1425 /* check, if it's a runtime call */
1426 if (env->ignore_runtime) {
1427 ir_node *symc = get_Call_ptr(call);
1429 if (is_Global(symc)) {
1430 ir_entity *ent = get_Global_entity(symc);
1432 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1437 /* collect all call nodes */
1439 ++x->n_call_nodes_orig;
1441 callee = get_call_called_irg(call);
1442 if (callee != NULL) {
1443 if (! env->ignore_callers) {
1444 inline_irg_env *callee_env = get_irg_link(callee);
1445 /* count all static callers */
1446 ++callee_env->n_callers;
1447 ++callee_env->n_callers_orig;
1449 if (callee == current_ir_graph)
1452 /* link it in the list of possible inlinable entries */
1453 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1455 entry->callee = callee;
1456 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1457 entry->benefice = 0;
1458 entry->local_adr = 0;
1459 entry->all_const = 0;
1461 list_add_tail(&entry->list, &x->calls);
1466 * Returns TRUE if the number of callers is 0 in the irg's environment,
1467 * hence this irg is a leave.
1469 inline static int is_leave(ir_graph *irg) {
1470 inline_irg_env *env = get_irg_link(irg);
1471 return env->n_call_nodes == 0;
1475 * Returns TRUE if the number of nodes in the callee is
1476 * smaller then size in the irg's environment.
1478 inline static int is_smaller(ir_graph *callee, unsigned size) {
1479 inline_irg_env *env = get_irg_link(callee);
1480 return env->n_nodes < size;
1484 * Duplicate a call entry.
1486 * @param entry the original entry to duplicate
1487 * @param new_call the new call node
1488 * @param loop_depth_delta
1489 * delta value for the loop depth
1491 static call_entry *duplicate_call_entry(const call_entry *entry,
1492 ir_node *new_call, int loop_depth_delta) {
1493 call_entry *nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1494 nentry->call = new_call;
1495 nentry->callee = entry->callee;
1496 nentry->benefice = entry->benefice;
1497 nentry->loop_depth = entry->loop_depth + loop_depth_delta;
1498 nentry->local_adr = entry->local_adr;
1499 nentry->all_const = entry->all_const;
1505 * Append all call nodes of the source environment to the nodes of in the destination
1508 * @param dst destination environment
1509 * @param src source environment
1510 * @param loop_depth the loop depth of the call that is replaced by the src list
1512 static void append_call_list(inline_irg_env *dst, inline_irg_env *src, int loop_depth) {
1513 call_entry *entry, *nentry;
1515 /* Note that the src list points to Call nodes in the inlined graph, but
1516 we need Call nodes in our graph. Luckily the inliner leaves this information
1517 in the link field. */
1518 list_for_each_entry(call_entry, entry, &src->calls, list) {
1519 nentry = duplicate_call_entry(entry, get_irn_link(entry->call), loop_depth);
1520 list_add_tail(&nentry->list, &dst->calls);
1522 dst->n_call_nodes += src->n_call_nodes;
1523 dst->n_nodes += src->n_nodes;
1527 * Inlines small leave methods at call sites where the called address comes
1528 * from a Const node that references the entity representing the called
1530 * The size argument is a rough measure for the code size of the method:
1531 * Methods where the obstack containing the firm graph is smaller than
1534 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
1535 unsigned size, int ignore_runtime)
1537 inline_irg_env *env;
1543 call_entry *entry, *next;
1544 const call_entry *centry;
1545 pmap *copied_graphs;
1546 pmap_entry *pm_entry;
1548 rem = current_ir_graph;
1549 obstack_init(&temp_obst);
1551 /* a map for the copied graphs, used to inline recursive calls */
1552 copied_graphs = pmap_create();
1554 /* extend all irgs by a temporary data structure for inlining. */
1555 n_irgs = get_irp_n_irgs();
1556 for (i = 0; i < n_irgs; ++i)
1557 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1559 /* Pre-compute information in temporary data structure. */
1560 wenv.ignore_runtime = ignore_runtime;
1561 wenv.ignore_callers = 0;
1562 for (i = 0; i < n_irgs; ++i) {
1563 ir_graph *irg = get_irp_irg(i);
1565 assert(get_irg_phase_state(irg) != phase_building);
1566 free_callee_info(irg);
1568 assure_cf_loop(irg);
1569 wenv.x = get_irg_link(irg);
1570 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1573 /* -- and now inline. -- */
1575 /* Inline leaves recursively -- we might construct new leaves. */
1579 for (i = 0; i < n_irgs; ++i) {
1581 int phiproj_computed = 0;
1583 current_ir_graph = get_irp_irg(i);
1584 env = get_irg_link(current_ir_graph);
1586 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1587 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1589 irg_inline_property prop;
1591 if (env->n_nodes > maxsize)
1595 callee = entry->callee;
1597 prop = get_irg_inline_property(callee);
1598 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1599 /* do not inline forbidden / weak graphs */
1603 if (is_leave(callee) && (
1604 is_smaller(callee, leavesize) || prop >= irg_inline_forced)) {
1605 if (!phiproj_computed) {
1606 phiproj_computed = 1;
1607 collect_phiprojs(current_ir_graph);
1609 did_inline = inline_method(call, callee);
1612 inline_irg_env *callee_env = get_irg_link(callee);
1614 /* call was inlined, Phi/Projs for current graph must be recomputed */
1615 phiproj_computed = 0;
1617 /* Do some statistics */
1618 env->got_inline = 1;
1619 --env->n_call_nodes;
1620 env->n_nodes += callee_env->n_nodes;
1621 --callee_env->n_callers;
1623 /* remove this call from the list */
1624 list_del(&entry->list);
1629 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1631 } while (did_inline);
1633 /* inline other small functions. */
1634 for (i = 0; i < n_irgs; ++i) {
1636 int phiproj_computed = 0;
1638 current_ir_graph = get_irp_irg(i);
1639 env = get_irg_link(current_ir_graph);
1641 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1643 /* note that the list of possible calls is updated during the process */
1644 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1645 irg_inline_property prop;
1650 callee = entry->callee;
1652 prop = get_irg_inline_property(callee);
1653 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1654 /* do not inline forbidden / weak graphs */
1658 e = pmap_find(copied_graphs, callee);
1661 * Remap callee if we have a copy.
1662 * FIXME: Should we do this only for recursive Calls ?
1667 if (prop >= irg_inline_forced ||
1668 (is_smaller(callee, size) && env->n_nodes < maxsize) /* small function */) {
1669 if (current_ir_graph == callee) {
1671 * Recursive call: we cannot directly inline because we cannot walk
1672 * the graph and change it. So we have to make a copy of the graph
1676 inline_irg_env *callee_env;
1679 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1682 * No copy yet, create one.
1683 * Note that recursive methods are never leaves, so it is sufficient
1684 * to test this condition here.
1686 copy = create_irg_copy(callee);
1688 /* create_irg_copy() destroys the Proj links, recompute them */
1689 phiproj_computed = 0;
1691 ir_reserve_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1693 /* allocate new environment */
1694 callee_env = alloc_inline_irg_env();
1695 set_irg_link(copy, callee_env);
1697 assure_cf_loop(copy);
1698 wenv.x = callee_env;
1699 wenv.ignore_callers = 1;
1700 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1703 * Enter the entity of the original graph. This is needed
1704 * for inline_method(). However, note that ent->irg still points
1705 * to callee, NOT to copy.
1707 set_irg_entity(copy, get_irg_entity(callee));
1709 pmap_insert(copied_graphs, callee, copy);
1712 /* we have only one caller: the original graph */
1713 callee_env->n_callers = 1;
1714 callee_env->n_callers_orig = 1;
1716 if (! phiproj_computed) {
1717 phiproj_computed = 1;
1718 collect_phiprojs(current_ir_graph);
1720 did_inline = inline_method(call, callee);
1722 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1724 /* call was inlined, Phi/Projs for current graph must be recomputed */
1725 phiproj_computed = 0;
1727 /* callee was inline. Append it's call list. */
1728 env->got_inline = 1;
1729 --env->n_call_nodes;
1730 append_call_list(env, callee_env, entry->loop_depth);
1731 --callee_env->n_callers;
1733 /* after we have inlined callee, all called methods inside callee
1734 are now called once more */
1735 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
1736 inline_irg_env *penv = get_irg_link(centry->callee);
1740 /* remove this call from the list */
1741 list_del(&entry->list);
1746 ir_free_resources(current_ir_graph, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
1749 for (i = 0; i < n_irgs; ++i) {
1750 irg = get_irp_irg(i);
1751 env = get_irg_link(irg);
1753 if (env->got_inline) {
1754 optimize_graph_df(irg);
1757 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1758 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1759 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1760 env->n_callers_orig, env->n_callers,
1761 get_entity_name(get_irg_entity(irg))));
1765 /* kill the copied graphs: we don't need them anymore */
1766 foreach_pmap(copied_graphs, pm_entry) {
1767 ir_graph *copy = pm_entry->value;
1769 /* reset the entity, otherwise it will be deleted in the next step ... */
1770 set_irg_entity(copy, NULL);
1771 free_ir_graph(copy);
1773 pmap_destroy(copied_graphs);
1775 obstack_free(&temp_obst, NULL);
1776 current_ir_graph = rem;
1780 * Calculate the parameter weights for transmitting the address of a local variable.
1782 static unsigned calc_method_local_weight(ir_node *arg) {
1784 unsigned v, weight = 0;
1786 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1787 ir_node *succ = get_irn_out(arg, i);
1789 switch (get_irn_opcode(succ)) {
1792 /* Loads and Store can be removed */
1796 /* check if all args are constant */
1797 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1798 ir_node *idx = get_Sel_index(succ, j);
1799 if (! is_Const(idx))
1802 /* Check users on this Sel. Note: if a 0 is returned here, there was
1803 some unsupported node. */
1804 v = calc_method_local_weight(succ);
1807 /* we can kill one Sel with constant indexes, this is cheap */
1811 /* when looking backward we might find Id nodes */
1812 weight += calc_method_local_weight(succ);
1815 /* unoptimized tuple */
1816 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1817 ir_node *pred = get_Tuple_pred(succ, j);
1819 /* look for Proj(j) */
1820 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1821 ir_node *succ_succ = get_irn_out(succ, k);
1822 if (is_Proj(succ_succ)) {
1823 if (get_Proj_proj(succ_succ) == j) {
1825 weight += calc_method_local_weight(succ_succ);
1828 /* this should NOT happen */
1836 /* any other node: unsupported yet or bad. */
1844 * Calculate the parameter weights for transmitting the address of a local variable.
1846 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1847 ir_entity *ent = get_irg_entity(irg);
1849 int nparams, i, proj_nr;
1850 ir_node *irg_args, *arg;
1852 mtp = get_entity_type(ent);
1853 nparams = get_method_n_params(mtp);
1855 /* allocate a new array. currently used as 'analysed' flag */
1856 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1858 /* If the method haven't parameters we have nothing to do. */
1862 assure_irg_outs(irg);
1863 irg_args = get_irg_args(irg);
1864 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1865 arg = get_irn_out(irg_args, i);
1866 proj_nr = get_Proj_proj(arg);
1867 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1872 * Calculate the benefice for transmitting an local variable address.
1873 * After inlining, the local variable might be transformed into a
1874 * SSA variable by scalar_replacement().
1876 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1877 inline_irg_env *env = get_irg_link(callee);
1879 if (env->local_weights != NULL) {
1880 if (pos < ARR_LEN(env->local_weights))
1881 return env->local_weights[pos];
1885 analyze_irg_local_weights(env, callee);
1887 if (pos < ARR_LEN(env->local_weights))
1888 return env->local_weights[pos];
1893 * Calculate a benefice value for inlining the given call.
1895 * @param call the call node we have to inspect
1896 * @param callee the called graph
1898 static int calc_inline_benefice(call_entry *entry, ir_graph *callee)
1900 ir_node *call = entry->call;
1901 ir_entity *ent = get_irg_entity(callee);
1905 int i, n_params, all_const;
1907 irg_inline_property prop;
1909 inline_irg_env *callee_env;
1911 prop = get_irg_inline_property(callee);
1912 if (prop == irg_inline_forbidden) {
1913 DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden\n",
1915 return entry->benefice = INT_MIN;
1918 if (get_irg_additional_properties(callee) & (mtp_property_noreturn | mtp_property_weak)) {
1919 DB((dbg, LEVEL_2, "In %+F Call to %+F: not inlining noreturn or weak\n",
1921 return entry->benefice = INT_MIN;
1924 /* costs for every passed parameter */
1925 n_params = get_Call_n_params(call);
1926 mtp = get_entity_type(ent);
1927 cc = get_method_calling_convention(mtp);
1928 if (cc & cc_reg_param) {
1929 /* register parameter, smaller costs for register parameters */
1930 int max_regs = cc & ~cc_bits;
1932 if (max_regs < n_params)
1933 weight += max_regs * 2 + (n_params - max_regs) * 5;
1935 weight += n_params * 2;
1937 /* parameters are passed an stack */
1938 weight += 5 * n_params;
1941 /* constant parameters improve the benefice */
1942 frame_ptr = get_irg_frame(current_ir_graph);
1944 for (i = 0; i < n_params; ++i) {
1945 ir_node *param = get_Call_param(call, i);
1947 if (is_Const(param)) {
1948 weight += get_method_param_weight(ent, i);
1951 if (is_SymConst(param))
1952 weight += get_method_param_weight(ent, i);
1953 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1955 * An address of a local variable is transmitted. After
1956 * inlining, scalar_replacement might be able to remove the
1957 * local variable, so honor this.
1959 v = get_method_local_adress_weight(callee, i);
1962 entry->local_adr = 1;
1966 entry->all_const = all_const;
1968 callee_env = get_irg_link(callee);
1969 if (callee_env->n_callers == 1 &&
1970 callee != current_ir_graph &&
1971 get_entity_visibility(ent) == visibility_local) {
1975 /* give a bonus for functions with one block */
1976 if (callee_env->n_blocks == 1)
1977 weight = weight * 3 / 2;
1979 /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1980 if (callee_env->n_nodes < 30 && !callee_env->recursive)
1983 /* and finally for leaves: they do not increase the register pressure
1984 because of callee safe registers */
1985 if (callee_env->n_call_nodes == 0)
1988 /** it's important to inline inner loops first */
1989 if (entry->loop_depth > 30)
1990 weight += 30 * 1024;
1992 weight += entry->loop_depth * 1024;
1995 * All arguments constant is probably a good sign, give an extra bonus
2000 return entry->benefice = weight;
2003 static ir_graph **irgs;
2004 static int last_irg;
2007 * Callgraph walker, collect all visited graphs.
2009 static void callgraph_walker(ir_graph *irg, void *data) {
2011 irgs[last_irg++] = irg;
2015 * Creates an inline order for all graphs.
2017 * @return the list of graphs.
2019 static ir_graph **create_irg_list(void) {
2020 ir_entity **free_methods;
2022 int n_irgs = get_irp_n_irgs();
2024 cgana(&arr_len, &free_methods);
2025 xfree(free_methods);
2027 compute_callgraph();
2030 irgs = XMALLOCNZ(ir_graph*, n_irgs);
2032 callgraph_walk(NULL, callgraph_walker, NULL);
2033 assert(n_irgs == last_irg);
2039 * Push a call onto the priority list if its benefice is big enough.
2041 * @param pqueue the priority queue of calls
2042 * @param call the call entry
2043 * @param inlien_threshold
2044 * the threshold value
2046 static void maybe_push_call(pqueue_t *pqueue, call_entry *call,
2047 int inline_threshold)
2049 ir_graph *callee = call->callee;
2050 irg_inline_property prop = get_irg_inline_property(callee);
2051 int benefice = calc_inline_benefice(call, callee);
2053 DB((dbg, LEVEL_2, "In %+F Call %+F to %+F has benefice %d\n",
2054 get_irn_irg(call->call), call->call, callee, benefice));
2056 if (prop < irg_inline_forced && benefice < inline_threshold) {
2060 pqueue_put(pqueue, call, benefice);
2064 * Try to inline calls into a graph.
2066 * @param irg the graph into which we inline
2067 * @param maxsize do NOT inline if the size of irg gets
2068 * bigger than this amount
2069 * @param inline_threshold
2070 * threshold value for inline decision
2071 * @param copied_graphs
2072 * map containing copied of recursive graphs
2074 static void inline_into(ir_graph *irg, unsigned maxsize,
2075 int inline_threshold, pmap *copied_graphs)
2077 int phiproj_computed = 0;
2078 inline_irg_env *env = get_irg_link(irg);
2079 call_entry *curr_call;
2083 if (env->n_call_nodes == 0)
2086 if (env->n_nodes > maxsize) {
2087 DB((dbg, LEVEL_2, "%+F: too big (%d)\n", irg, env->n_nodes));
2091 current_ir_graph = irg;
2092 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2094 /* put irgs into the pqueue */
2095 pqueue = new_pqueue();
2097 list_for_each_entry(call_entry, curr_call, &env->calls, list) {
2098 assert(is_Call(curr_call->call));
2099 maybe_push_call(pqueue, curr_call, inline_threshold);
2102 /* note that the list of possible calls is updated during the process */
2103 while (!pqueue_empty(pqueue)) {
2105 call_entry *curr_call = pqueue_pop_front(pqueue);
2106 ir_graph *callee = curr_call->callee;
2107 ir_node *call_node = curr_call->call;
2108 inline_irg_env *callee_env = get_irg_link(callee);
2109 irg_inline_property prop = get_irg_inline_property(callee);
2111 const call_entry *centry;
2114 if ((prop < irg_inline_forced) && env->n_nodes + callee_env->n_nodes > maxsize) {
2115 DB((dbg, LEVEL_2, "%+F: too big (%d) + %+F (%d)\n", irg,
2116 env->n_nodes, callee, callee_env->n_nodes));
2120 e = pmap_find(copied_graphs, callee);
2122 int benefice = curr_call->benefice;
2124 * Reduce the weight for recursive function IFF not all arguments are const.
2125 * inlining recursive functions is rarely good.
2127 if (!curr_call->all_const)
2129 if (benefice < inline_threshold)
2133 * Remap callee if we have a copy.
2136 callee_env = get_irg_link(callee);
2139 if (current_ir_graph == callee) {
2141 * Recursive call: we cannot directly inline because we cannot
2142 * walk the graph and change it. So we have to make a copy of
2145 int benefice = curr_call->benefice;
2149 * Reduce the weight for recursive function IFF not all arguments are const.
2150 * inlining recursive functions is rarely good.
2152 if (!curr_call->all_const)
2154 if (benefice < inline_threshold)
2157 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2160 * No copy yet, create one.
2161 * Note that recursive methods are never leaves, so it is
2162 * sufficient to test this condition here.
2164 copy = create_irg_copy(callee);
2166 /* create_irg_copy() destroys the Proj links, recompute them */
2167 phiproj_computed = 0;
2169 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2171 /* allocate a new environment */
2172 callee_env = alloc_inline_irg_env();
2173 set_irg_link(copy, callee_env);
2175 assure_cf_loop(copy);
2176 wenv.x = callee_env;
2177 wenv.ignore_callers = 1;
2178 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2181 * Enter the entity of the original graph. This is needed
2182 * for inline_method(). However, note that ent->irg still points
2183 * to callee, NOT to copy.
2185 set_irg_entity(copy, get_irg_entity(callee));
2187 pmap_insert(copied_graphs, callee, copy);
2190 /* we have only one caller: the original graph */
2191 callee_env->n_callers = 1;
2192 callee_env->n_callers_orig = 1;
2194 if (! phiproj_computed) {
2195 phiproj_computed = 1;
2196 collect_phiprojs(current_ir_graph);
2198 did_inline = inline_method(call_node, callee);
2202 /* call was inlined, Phi/Projs for current graph must be recomputed */
2203 phiproj_computed = 0;
2205 /* remove it from the caller list */
2206 list_del(&curr_call->list);
2208 /* callee was inline. Append it's call list. */
2209 env->got_inline = 1;
2210 if (curr_call->local_adr)
2211 env->local_vars = 1;
2212 --env->n_call_nodes;
2214 /* we just generate a bunch of new calls */
2215 loop_depth = curr_call->loop_depth;
2216 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
2217 inline_irg_env *penv = get_irg_link(centry->callee);
2219 call_entry *new_entry;
2221 /* after we have inlined callee, all called methods inside
2222 * callee are now called once more */
2225 /* Note that the src list points to Call nodes in the inlined graph,
2226 * but we need Call nodes in our graph. Luckily the inliner leaves
2227 * this information in the link field. */
2228 new_call = get_irn_link(centry->call);
2229 assert(is_Call(new_call));
2231 new_entry = duplicate_call_entry(centry, new_call, loop_depth);
2232 list_add_tail(&new_entry->list, &env->calls);
2233 maybe_push_call(pqueue, new_entry, inline_threshold);
2236 env->n_call_nodes += callee_env->n_call_nodes;
2237 env->n_nodes += callee_env->n_nodes;
2238 --callee_env->n_callers;
2240 ir_free_resources(irg, IR_RESOURCE_IRN_LINK|IR_RESOURCE_PHI_LIST);
2245 * Heuristic inliner. Calculates a benefice value for every call and inlines
2246 * those calls with a value higher than the threshold.
2248 void inline_functions(unsigned maxsize, int inline_threshold) {
2249 inline_irg_env *env;
2253 pmap *copied_graphs;
2254 pmap_entry *pm_entry;
2257 rem = current_ir_graph;
2258 obstack_init(&temp_obst);
2260 irgs = create_irg_list();
2262 /* a map for the copied graphs, used to inline recursive calls */
2263 copied_graphs = pmap_create();
2265 /* extend all irgs by a temporary data structure for inlining. */
2266 n_irgs = get_irp_n_irgs();
2267 for (i = 0; i < n_irgs; ++i)
2268 set_irg_link(irgs[i], alloc_inline_irg_env());
2270 /* Pre-compute information in temporary data structure. */
2271 wenv.ignore_runtime = 0;
2272 wenv.ignore_callers = 0;
2273 for (i = 0; i < n_irgs; ++i) {
2274 ir_graph *irg = irgs[i];
2276 free_callee_info(irg);
2278 wenv.x = get_irg_link(irg);
2279 assure_cf_loop(irg);
2280 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
2283 /* -- and now inline. -- */
2284 for (i = 0; i < n_irgs; ++i) {
2285 ir_graph *irg = irgs[i];
2287 inline_into(irg, maxsize, inline_threshold, copied_graphs);
2290 for (i = 0; i < n_irgs; ++i) {
2291 ir_graph *irg = irgs[i];
2293 env = get_irg_link(irg);
2294 if (env->got_inline) {
2295 /* this irg got calls inlined: optimize it */
2296 if (get_opt_combo()) {
2297 if (env->local_vars) {
2298 scalar_replacement_opt(irg);
2302 if (env->local_vars) {
2303 if (scalar_replacement_opt(irg)) {
2304 optimize_graph_df(irg);
2310 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2311 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2312 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2313 env->n_callers_orig, env->n_callers,
2314 get_entity_name(get_irg_entity(irg))));
2318 /* kill the copied graphs: we don't need them anymore */
2319 foreach_pmap(copied_graphs, pm_entry) {
2320 ir_graph *copy = pm_entry->value;
2322 /* reset the entity, otherwise it will be deleted in the next step ... */
2323 set_irg_entity(copy, NULL);
2324 free_ir_graph(copy);
2326 pmap_destroy(copied_graphs);
2330 obstack_free(&temp_obst, NULL);
2331 current_ir_graph = rem;
2334 void firm_init_inline(void) {
2335 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");