2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Dead node elimination and Procedure Inlining.
23 * @author Michael Beck, Goetz Lindenmaier
34 #include "irgraph_t.h"
37 #include "iroptimize.h"
44 #include "adt/array.h"
48 #include "adt/xmalloc.h"
52 #include "irbackedge_t.h"
53 #include "opt_inline_t.h"
58 #include "analyze_irg_args.h"
59 #include "iredges_t.h"
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 ((get_opt_control_flow_straightening()) &&
225 (get_Block_n_cfgpreds(nn) == 1) &&
226 is_Jmp(get_Block_cfgpred(nn, 0))) {
227 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
229 /* Jmp jumps into the block it is in -- deal self cycle. */
230 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
231 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
236 } else if (is_Phi(n) && get_irn_arity(n) > 0) {
237 /* Don't copy node if corresponding predecessor in block is Bad.
238 The Block itself should not be Bad. */
239 block = get_nodes_block(n);
240 set_nodes_block(nn, get_new_node(block));
242 irn_arity = get_irn_arity(n);
243 for (i = 0; i < irn_arity; i++) {
244 if (! is_Bad(get_irn_n(block, i))) {
245 ir_node *pred = get_irn_n(n, i);
246 set_irn_n(nn, j, get_new_node(pred));
247 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
251 /* If the pre walker reached this Phi after the post walker visited the
252 block block_visited is > 0. */
253 set_Block_block_visited(get_nodes_block(n), 0);
254 /* Compacting the Phi's ins might generate Phis with only one
256 if (get_irn_arity(nn) == 1)
257 exchange(nn, get_irn_n(nn, 0));
259 irn_arity = get_irn_arity(n);
260 for (i = -1; i < irn_arity; i++)
261 set_irn_n(nn, i, get_new_node(get_irn_n(n, i)));
263 /* Now the new node is complete. We can add it to the hash table for CSE.
264 @@@ inlining aborts if we identify End. Why? */
266 add_identities(current_ir_graph->value_table, nn);
270 * Copies the graph recursively, compacts the keep-alives of the end node.
272 * @param irg the graph to be copied
273 * @param copy_node_nr If non-zero, the node number will be copied
275 static void copy_graph(ir_graph *irg, int copy_node_nr) {
276 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
277 ir_node *ka; /* keep alive */
281 /* Some nodes must be copied by hand, sigh */
282 vfl = get_irg_visited(irg);
283 set_irg_visited(irg, vfl + 1);
285 oe = get_irg_end(irg);
286 mark_irn_visited(oe);
287 /* copy the end node by hand, allocate dynamic in array! */
288 ne = new_ir_node(get_irn_dbg_info(oe),
295 /* Copy the attributes. Well, there might be some in the future... */
296 copy_node_attr(oe, ne);
297 set_new_node(oe, ne);
299 /* copy the Bad node */
300 ob = get_irg_bad(irg);
301 mark_irn_visited(ob);
302 nb = new_ir_node(get_irn_dbg_info(ob),
309 copy_node_attr(ob, nb);
310 set_new_node(ob, nb);
312 /* copy the NoMem node */
313 om = get_irg_no_mem(irg);
314 mark_irn_visited(om);
315 nm = new_ir_node(get_irn_dbg_info(om),
322 copy_node_attr(om, nm);
323 set_new_node(om, nm);
325 /* copy the live nodes */
326 set_irg_visited(irg, vfl);
327 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
329 /* Note: from yet, the visited flag of the graph is equal to vfl + 1 */
331 /* visit the anchors as well */
332 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
333 ir_node *n = get_irg_anchor(irg, i);
335 if (n && (get_irn_visited(n) <= vfl)) {
336 set_irg_visited(irg, vfl);
337 irg_walk(n, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
341 /* copy_preds for the end node ... */
342 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
344 /*- ... and now the keep alives. -*/
345 /* First pick the not marked block nodes and walk them. We must pick these
346 first as else we will oversee blocks reachable from Phis. */
347 irn_arity = get_End_n_keepalives(oe);
348 for (i = 0; i < irn_arity; i++) {
349 ka = get_End_keepalive(oe, i);
351 if (get_irn_visited(ka) <= vfl) {
352 /* We must keep the block alive and copy everything reachable */
353 set_irg_visited(irg, vfl);
354 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
356 add_End_keepalive(ne, get_new_node(ka));
360 /* Now pick other nodes. Here we will keep all! */
361 irn_arity = get_End_n_keepalives(oe);
362 for (i = 0; i < irn_arity; i++) {
363 ka = get_End_keepalive(oe, i);
365 if (get_irn_visited(ka) <= vfl) {
366 /* We didn't copy the node yet. */
367 set_irg_visited(irg, vfl);
368 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
370 add_End_keepalive(ne, get_new_node(ka));
374 /* start block sometimes only reached after keep alives */
375 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
376 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
380 * Copies the graph reachable from current_ir_graph->end to the obstack
381 * in current_ir_graph and fixes the environment.
382 * Then fixes the fields in current_ir_graph containing nodes of the
385 * @param copy_node_nr If non-zero, the node number will be copied
388 copy_graph_env(int copy_node_nr) {
389 ir_graph *irg = current_ir_graph;
390 ir_node *old_end, *new_anchor;
393 /* remove end_except and end_reg nodes */
394 old_end = get_irg_end(irg);
395 set_irg_end_except (irg, old_end);
396 set_irg_end_reg (irg, old_end);
398 /* Not all nodes remembered in irg might be reachable
399 from the end node. Assure their link is set to NULL, so that
400 we can test whether new nodes have been computed. */
401 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
402 ir_node *n = get_irg_anchor(irg, i);
404 set_new_node(n, NULL);
406 /* we use the block walk flag for removing Bads from Blocks ins. */
407 inc_irg_block_visited(irg);
410 copy_graph(irg, copy_node_nr);
413 old_end = get_irg_end(irg);
414 new_anchor = new_Anchor(irg);
416 for (i = get_irg_n_anchors(irg) - 1; i >= 0; --i) {
417 ir_node *n = get_irg_anchor(irg, i);
419 set_irn_n(new_anchor, i, get_new_node(n));
422 irg->anchor = new_anchor;
424 /* ensure the new anchor is placed in the endblock */
425 set_nodes_block(new_anchor, get_irg_end_block(irg));
429 * Copies all reachable nodes to a new obstack. Removes bad inputs
430 * from block nodes and the corresponding inputs from Phi nodes.
431 * Merges single exit blocks with single entry blocks and removes
433 * Adds all new nodes to a new hash table for CSE. Does not
434 * perform CSE, so the hash table might contain common subexpressions.
436 void dead_node_elimination(ir_graph *irg) {
438 #ifdef INTERPROCEDURAL_VIEW
439 int rem_ipview = get_interprocedural_view();
441 struct obstack *graveyard_obst = NULL;
442 struct obstack *rebirth_obst = NULL;
444 edges_deactivate(irg);
446 /* inform statistics that we started a dead-node elimination run */
447 hook_dead_node_elim(irg, 1);
449 /* Remember external state of current_ir_graph. */
450 rem = current_ir_graph;
451 current_ir_graph = irg;
452 #ifdef INTERPROCEDURAL_VIEW
453 set_interprocedural_view(0);
456 assert(get_irg_phase_state(irg) != phase_building);
458 /* Handle graph state */
459 free_callee_info(irg);
463 /* @@@ so far we loose loops when copying */
464 free_loop_information(irg);
466 set_irg_doms_inconsistent(irg);
468 /* A quiet place, where the old obstack can rest in peace,
469 until it will be cremated. */
470 graveyard_obst = irg->obst;
472 /* A new obstack, where the reachable nodes will be copied to. */
473 rebirth_obst = xmalloc(sizeof(*rebirth_obst));
474 irg->obst = rebirth_obst;
475 obstack_init(irg->obst);
476 irg->last_node_idx = 0;
478 /* We also need a new value table for CSE */
479 del_identities(irg->value_table);
480 irg->value_table = new_identities();
482 /* Copy the graph from the old to the new obstack */
483 copy_graph_env(/*copy_node_nr=*/1);
485 /* Free memory from old unoptimized obstack */
486 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
487 xfree(graveyard_obst); /* ... then free it. */
489 /* inform statistics that the run is over */
490 hook_dead_node_elim(irg, 0);
492 current_ir_graph = rem;
493 #ifdef INTERPROCEDURAL_VIEW
494 set_interprocedural_view(rem_ipview);
499 * Relink bad predecessors of a block and store the old in array to the
500 * link field. This function is called by relink_bad_predecessors().
501 * The array of link field starts with the block operand at position 0.
502 * If block has bad predecessors, create a new in array without bad preds.
503 * Otherwise let in array untouched.
505 static void relink_bad_block_predecessors(ir_node *n, void *env) {
506 ir_node **new_in, *irn;
507 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
510 /* if link field of block is NULL, look for bad predecessors otherwise
511 this is already done */
512 if (is_Block(n) && get_irn_link(n) == NULL) {
513 /* save old predecessors in link field (position 0 is the block operand)*/
514 set_irn_link(n, get_irn_in(n));
516 /* count predecessors without bad nodes */
517 old_irn_arity = get_irn_arity(n);
518 for (i = 0; i < old_irn_arity; i++)
519 if (!is_Bad(get_irn_n(n, i)))
522 /* arity changing: set new predecessors without bad nodes */
523 if (new_irn_arity < old_irn_arity) {
524 /* Get new predecessor array. We do not resize the array, as we must
525 keep the old one to update Phis. */
526 new_in = NEW_ARR_D(ir_node *, current_ir_graph->obst, (new_irn_arity+1));
528 /* set new predecessors in array */
531 for (i = 0; i < old_irn_arity; i++) {
532 irn = get_irn_n(n, i);
534 new_in[new_irn_n] = irn;
535 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
539 /* ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity); */
540 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
542 } /* ir node has bad predecessors */
543 } /* Block is not relinked */
547 * Relinks Bad predecessors from Blocks and Phis called by walker
548 * remove_bad_predecesors(). If n is a Block, call
549 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
550 * function of Phi's Block. If this block has bad predecessors, relink preds
553 static void relink_bad_predecessors(ir_node *n, void *env) {
554 ir_node *block, **old_in;
555 int i, old_irn_arity, new_irn_arity;
557 /* relink bad predecessors of a block */
559 relink_bad_block_predecessors(n, env);
561 /* If Phi node relink its block and its predecessors */
563 /* Relink predecessors of phi's block */
564 block = get_nodes_block(n);
565 if (get_irn_link(block) == NULL)
566 relink_bad_block_predecessors(block, env);
568 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
569 old_irn_arity = ARR_LEN(old_in);
571 /* Relink Phi predecessors if count of predecessors changed */
572 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
573 /* set new predecessors in array
574 n->in[0] remains the same block */
576 for(i = 1; i < old_irn_arity; i++)
577 if (!is_Bad(old_in[i])) {
578 n->in[new_irn_arity] = n->in[i];
579 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
583 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
584 ARR_SETLEN(int, n->attr.phi.u.backedge, new_irn_arity);
586 } /* n is a Phi node */
590 * Removes Bad Bad predecessors from Blocks and the corresponding
591 * inputs to Phi nodes as in dead_node_elimination but without
593 * On walking up set the link field to NULL, on walking down call
594 * relink_bad_predecessors() (This function stores the old in array
595 * to the link field and sets a new in array if arity of predecessors
598 void remove_bad_predecessors(ir_graph *irg) {
599 panic("Fix backedge handling first");
600 irg_walk_graph(irg, firm_clear_link, relink_bad_predecessors, NULL);
607 __)|_| | \_/ | \_/(/_ |_/\__|__
609 The following stuff implements a facility that automatically patches
610 registered ir_node pointers to the new node when a dead node elimination occurs.
613 struct _survive_dce_t {
617 hook_entry_t dead_node_elim;
618 hook_entry_t dead_node_elim_subst;
621 typedef struct _survive_dce_list_t {
622 struct _survive_dce_list_t *next;
624 } survive_dce_list_t;
626 static void dead_node_hook(void *context, ir_graph *irg, int start) {
627 survive_dce_t *sd = context;
630 /* Create a new map before the dead node elimination is performed. */
632 sd->new_places = pmap_create_ex(pmap_count(sd->places));
634 /* Patch back all nodes if dead node elimination is over and something is to be done. */
635 pmap_destroy(sd->places);
636 sd->places = sd->new_places;
637 sd->new_places = NULL;
642 * Hook called when dead node elimination replaces old by nw.
644 static void dead_node_subst_hook(void *context, ir_graph *irg, ir_node *old, ir_node *nw) {
645 survive_dce_t *sd = context;
646 survive_dce_list_t *list = pmap_get(sd->places, old);
649 /* If the node is to be patched back, write the new address to all registered locations. */
651 survive_dce_list_t *p;
653 for (p = list; p; p = p->next)
656 pmap_insert(sd->new_places, nw, list);
661 * Make a new Survive DCE environment.
663 survive_dce_t *new_survive_dce(void) {
664 survive_dce_t *res = xmalloc(sizeof(res[0]));
665 obstack_init(&res->obst);
666 res->places = pmap_create();
667 res->new_places = NULL;
669 res->dead_node_elim.hook._hook_dead_node_elim = dead_node_hook;
670 res->dead_node_elim.context = res;
671 res->dead_node_elim.next = NULL;
673 res->dead_node_elim_subst.hook._hook_dead_node_elim_subst = dead_node_subst_hook;
674 res->dead_node_elim_subst.context = res;
675 res->dead_node_elim_subst.next = NULL;
677 #ifndef FIRM_ENABLE_HOOKS
678 assert(0 && "need hooks enabled");
681 register_hook(hook_dead_node_elim, &res->dead_node_elim);
682 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
687 * Free a Survive DCE environment.
689 void free_survive_dce(survive_dce_t *sd) {
690 obstack_free(&sd->obst, NULL);
691 pmap_destroy(sd->places);
692 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
693 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
698 * Register a node pointer to be patched upon DCE.
699 * When DCE occurs, the node pointer specified by @p place will be
700 * patched to the new address of the node it is pointing to.
702 * @param sd The Survive DCE environment.
703 * @param place The address of the node pointer.
705 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
706 if (*place != NULL) {
707 ir_node *irn = *place;
708 survive_dce_list_t *curr = pmap_get(sd->places, irn);
709 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
714 pmap_insert(sd->places, irn, nw);
718 /*--------------------------------------------------------------------*/
719 /* Functionality for inlining */
720 /*--------------------------------------------------------------------*/
723 * Copy node for inlineing. Updates attributes that change when
724 * inlineing but not for dead node elimination.
726 * Copies the node by calling copy_node() and then updates the entity if
727 * it's a local one. env must be a pointer of the frame type of the
728 * inlined procedure. The new entities must be in the link field of
731 static void copy_node_inline(ir_node *n, void *env) {
733 ir_type *frame_tp = (ir_type *)env;
737 nn = get_new_node (n);
739 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
740 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
742 } else if (is_Block(n)) {
743 nn = get_new_node (n);
744 nn->attr.block.irg = current_ir_graph;
749 * Copies new predecessors of old node and move constants to
752 static void copy_preds_inline(ir_node *n, void *env) {
756 nn = skip_Id(get_new_node(n));
757 if (is_irn_constlike(nn)) {
758 /* move Constants into the start block */
759 set_nodes_block(nn, get_irg_start_block(current_ir_graph));
761 n = identify_remember(current_ir_graph->value_table, nn);
770 * Walker: checks if P_value_arg_base is used.
772 static void find_addr(ir_node *node, void *env) {
773 int *allow_inline = env;
775 is_Start(get_Proj_pred(node)) &&
776 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
778 } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
780 * Refuse to inline alloca call unless user explicitly forced so as this
781 * may change program's memory overhead drastically when the function
782 * using alloca is called in loop. In GCC present in SPEC2000 inlining
783 * into schedule_block cause it to require 2GB of ram instead of 256MB.
785 * Sorryly this is true with our implementation also.
786 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
787 * disables inlining of functions using VLA (with are completely save).
790 * - add a flag to the Alloc node for "real" alloca() calls
791 * - add a new Stack-Restore node at the end of a function using alloca()
798 * Check if we can inline a given call.
799 * Currently, we cannot inline two cases:
800 * - call with compound arguments
801 * - graphs that take the address of a parameter
803 * check these conditions here
805 static int can_inline(ir_node *call, ir_graph *called_graph) {
806 ir_type *call_type = get_Call_type(call);
807 int params, ress, i, res;
808 assert(is_Method_type(call_type));
810 params = get_method_n_params(call_type);
811 ress = get_method_n_ress(call_type);
813 /* check parameters for compound arguments */
814 for (i = 0; i < params; ++i) {
815 ir_type *p_type = get_method_param_type(call_type, i);
817 if (is_compound_type(p_type))
821 /* check results for compound arguments */
822 for (i = 0; i < ress; ++i) {
823 ir_type *r_type = get_method_res_type(call_type, i);
825 if (is_compound_type(r_type))
830 irg_walk_graph(called_graph, find_addr, NULL, &res);
836 exc_handler = 0, /**< There is a handler. */
837 exc_to_end = 1, /**< Branches to End. */
838 exc_no_handler = 2 /**< Exception handling not represented. */
841 /* Inlines a method at the given call site. */
842 int inline_method(ir_node *call, ir_graph *called_graph) {
844 ir_node *post_call, *post_bl;
845 ir_node *in[pn_Start_max];
846 ir_node *end, *end_bl, *block;
851 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
852 enum exc_mode exc_handling;
853 ir_type *called_frame, *curr_frame, *mtp, *ctp;
856 irg_inline_property prop = get_irg_inline_property(called_graph);
857 unsigned long visited;
859 if (prop == irg_inline_forbidden)
862 ent = get_irg_entity(called_graph);
864 mtp = get_entity_type(ent);
865 ctp = get_Call_type(call);
866 if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
867 /* this is a bad feature of C: without a prototype, we can can call a function with less
868 parameters than needed. Currently we don't support this, although it would be
869 to use Unknown than. */
873 /* Argh, compiling C has some bad consequences:
874 the call type AND the method type might be different.
875 It is implementation defendant what happens in that case.
876 We support inlining, if the bitsize of the types matches AND
877 the same arithmetic is used. */
878 n_params = get_method_n_params(mtp);
879 for (i = n_params - 1; i >= 0; --i) {
880 ir_type *param_tp = get_method_param_type(mtp, i);
881 ir_type *arg_tp = get_method_param_type(ctp, i);
883 if (param_tp != arg_tp) {
884 ir_mode *pmode = get_type_mode(param_tp);
885 ir_mode *amode = get_type_mode(arg_tp);
887 if (pmode == NULL || amode == NULL)
889 if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
891 if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
893 /* otherwise we can simply "reinterpret" the bits */
897 irg = get_irn_irg(call);
900 * We cannot inline a recursive call. The graph must be copied before
901 * the call the inline_method() using create_irg_copy().
903 if (called_graph == irg)
907 * currently, we cannot inline two cases:
908 * - call with compound arguments
909 * - graphs that take the address of a parameter
911 if (! can_inline(call, called_graph))
914 rem = current_ir_graph;
915 current_ir_graph = irg;
917 DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
919 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
920 rem_opt = get_opt_optimize();
923 /* Handle graph state */
924 assert(get_irg_phase_state(irg) != phase_building);
925 assert(get_irg_pinned(irg) == op_pin_state_pinned);
926 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
927 set_irg_outs_inconsistent(irg);
928 set_irg_extblk_inconsistent(irg);
929 set_irg_doms_inconsistent(irg);
930 set_irg_loopinfo_inconsistent(irg);
931 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
933 /* -- Check preconditions -- */
934 assert(is_Call(call));
936 /* here we know we WILL inline, so inform the statistics */
937 hook_inline(call, called_graph);
939 /* -- Decide how to handle exception control flow: Is there a handler
940 for the Call node, or do we branch directly to End on an exception?
942 0 There is a handler.
944 2 Exception handling not represented in Firm. -- */
946 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
947 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
948 long proj_nr = get_Proj_proj(proj);
949 if (proj_nr == pn_Call_X_except) Xproj = proj;
950 if (proj_nr == pn_Call_M_except) Mproj = proj;
952 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
953 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
954 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
957 /* create the argument tuple */
958 NEW_ARR_A(ir_type *, args_in, n_params);
960 block = get_nodes_block(call);
961 for (i = n_params - 1; i >= 0; --i) {
962 ir_node *arg = get_Call_param(call, i);
963 ir_type *param_tp = get_method_param_type(mtp, i);
964 ir_mode *mode = get_type_mode(param_tp);
966 if (mode != get_irn_mode(arg)) {
967 arg = new_r_Conv(irg, block, arg, mode);
973 the procedure and later replaces the Start node of the called graph.
974 Post_call is the old Call node and collects the results of the called
975 graph. Both will end up being a tuple. -- */
976 post_bl = get_nodes_block(call);
977 set_irg_current_block(irg, post_bl);
978 /* XxMxPxPxPxT of Start + parameter of Call */
979 in[pn_Start_X_initial_exec] = new_Jmp();
980 in[pn_Start_M] = get_Call_mem(call);
981 in[pn_Start_P_frame_base] = get_irg_frame(irg);
982 in[pn_Start_P_tls] = get_irg_tls(irg);
983 in[pn_Start_T_args] = new_Tuple(n_params, args_in);
984 /* in[pn_Start_P_value_arg_base] = ??? */
985 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
986 pre_call = new_Tuple(pn_Start_max - 1, in);
990 The new block gets the ins of the old block, pre_call and all its
991 predecessors and all Phi nodes. -- */
992 part_block(pre_call);
994 /* -- Prepare state for dead node elimination -- */
995 /* Visited flags in calling irg must be >= flag in called irg.
996 Else walker and arity computation will not work. */
997 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
998 set_irg_visited(irg, get_irg_visited(called_graph) + 1);
999 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
1000 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
1001 visited = get_irg_visited(irg);
1003 /* Set pre_call as new Start node in link field of the start node of
1004 calling graph and pre_calls block as new block for the start block
1006 Further mark these nodes so that they are not visited by the
1008 set_irn_link(get_irg_start(called_graph), pre_call);
1009 set_irn_visited(get_irg_start(called_graph), visited);
1010 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1011 set_irn_visited(get_irg_start_block(called_graph), visited);
1013 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1014 set_irn_visited(get_irg_bad(called_graph), visited);
1016 set_irn_link(get_irg_no_mem(called_graph), get_irg_no_mem(current_ir_graph));
1017 set_irn_visited(get_irg_no_mem(called_graph), visited);
1019 /* Initialize for compaction of in arrays */
1020 inc_irg_block_visited(irg);
1022 /* -- Replicate local entities of the called_graph -- */
1023 /* copy the entities. */
1024 called_frame = get_irg_frame_type(called_graph);
1025 curr_frame = get_irg_frame_type(irg);
1026 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
1027 ir_entity *new_ent, *old_ent;
1028 old_ent = get_class_member(called_frame, i);
1029 new_ent = copy_entity_own(old_ent, curr_frame);
1030 set_entity_link(old_ent, new_ent);
1033 /* visited is > than that of called graph. With this trick visited will
1034 remain unchanged so that an outer walker, e.g., searching the call nodes
1035 to inline, calling this inline will not visit the inlined nodes. */
1036 set_irg_visited(irg, get_irg_visited(irg)-1);
1038 /* -- Performing dead node elimination inlines the graph -- */
1039 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1041 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
1042 get_irg_frame_type(called_graph));
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_irn_arity(end_bl); /* arity = n_exc + n_ret */
1067 n_res = get_method_n_ress(get_Call_type(call));
1069 res_pred = xmalloc(n_res * sizeof(*res_pred));
1070 cf_pred = xmalloc(arity * sizeof(*res_pred));
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_irn_n(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_irn_n(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_irn_n(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_irn_n(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_irn_n(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_irn_n(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 = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
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 inlineing to small methods. */
1235 /********************************************************************/
1237 static struct obstack temp_obst;
1239 /** Represents a possible inlinable call in a graph. */
1240 typedef struct _call_entry call_entry;
1241 struct _call_entry {
1242 ir_node *call; /**< the Call node */
1243 ir_graph *callee; /**< the callee IR-graph called here */
1244 call_entry *next; /**< for linking the next one */
1245 int loop_depth; /**< the loop depth of this call */
1249 * environment for inlining small irgs
1251 typedef struct _inline_env_t {
1252 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1253 call_entry *head; /**< the head of the call entry list */
1254 call_entry *tail; /**< the tail of the call entry list */
1258 * Returns the irg called from a Call node. If the irg is not
1259 * known, NULL is returned.
1261 * @param call the call node
1263 static ir_graph *get_call_called_irg(ir_node *call) {
1266 addr = get_Call_ptr(call);
1267 if (is_Global(addr)) {
1268 ir_entity *ent = get_Global_entity(addr);
1269 return get_entity_irg(ent);
1276 * Walker: Collect all calls to known graphs inside a graph.
1278 static void collect_calls(ir_node *call, void *env) {
1279 if (is_Call(call)) {
1280 ir_graph *called_irg = get_call_called_irg(call);
1282 if (called_irg != NULL) {
1283 /* The Call node calls a locally defined method. Remember to inline. */
1284 inline_env_t *ienv = env;
1285 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1287 entry->callee = called_irg;
1289 entry->loop_depth = 0;
1291 if (ienv->tail == NULL)
1294 ienv->tail->next = entry;
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 inlineing the same
1320 method several times changes the visited flag of the walked graph:
1321 after the first inlineing visited of the callee equals visited of
1322 the caller. With the next inlineing both are increased.) */
1323 obstack_init(&env.obst);
1324 env.head = env.tail = NULL;
1325 irg_walk_graph(irg, NULL, collect_calls, &env);
1327 if (env.head != NULL) {
1328 /* There are calls to inline */
1329 collect_phiprojs(irg);
1330 for (entry = env.head; entry != NULL; entry = entry->next) {
1331 ir_graph *callee = entry->callee;
1332 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1333 (get_irg_inline_property(callee) >= irg_inline_forced)) {
1334 inline_method(entry->call, callee);
1338 obstack_free(&env.obst, NULL);
1339 current_ir_graph = rem;
1343 * Environment for inlining irgs.
1346 int n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1347 int n_blocks; /**< Number of Blocks in graph without Start and End block. */
1348 int n_nodes_orig; /**< for statistics */
1349 int n_call_nodes; /**< Number of Call nodes in the graph. */
1350 int n_call_nodes_orig; /**< for statistics */
1351 int n_callers; /**< Number of known graphs that call this graphs. */
1352 int n_callers_orig; /**< for statistics */
1353 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1354 unsigned local_vars:1; /**< Set, if a inlined function gets the address of an inlined variable. */
1355 unsigned recursive:1; /**< Set, if this function is self recursive. */
1356 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1357 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1358 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1362 * Allocate a new environment for inlining.
1364 static inline_irg_env *alloc_inline_irg_env(void) {
1365 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1366 env->n_nodes = -2; /* do not count count Start, End */
1367 env->n_blocks = -2; /* do not count count Start, End Block */
1368 env->n_nodes_orig = -2; /* do not count Start, End */
1369 env->call_head = NULL;
1370 env->call_tail = NULL;
1371 env->n_call_nodes = 0;
1372 env->n_call_nodes_orig = 0;
1374 env->n_callers_orig = 0;
1375 env->got_inline = 0;
1376 env->local_vars = 0;
1378 env->local_weights = NULL;
1382 typedef struct walker_env {
1383 inline_irg_env *x; /**< the inline environment */
1384 call_entry *last_call; /**< points to the last inserted call */
1385 char ignore_runtime; /**< the ignore runtime flag */
1386 char ignore_callers; /**< if set, do change callers data */
1390 * post-walker: collect all calls in the inline-environment
1391 * of a graph and sum some statistics.
1393 static void collect_calls2(ir_node *call, void *ctx) {
1395 inline_irg_env *x = env->x;
1396 ir_opcode code = get_irn_opcode(call);
1400 /* count meaningful nodes in irg */
1401 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1402 if (code != iro_Block) {
1410 if (code != iro_Call) return;
1412 /* check, if it's a runtime call */
1413 if (env->ignore_runtime) {
1414 ir_node *symc = get_Call_ptr(call);
1416 if (is_Global(symc)) {
1417 ir_entity *ent = get_Global_entity(symc);
1419 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1424 /* collect all call nodes */
1426 ++x->n_call_nodes_orig;
1428 callee = get_call_called_irg(call);
1429 if (callee != NULL) {
1430 if (! env->ignore_callers) {
1431 inline_irg_env *callee_env = get_irg_link(callee);
1432 /* count all static callers */
1433 ++callee_env->n_callers;
1434 ++callee_env->n_callers_orig;
1436 if (callee == current_ir_graph)
1439 /* link it in the list of possible inlinable entries */
1440 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1442 entry->callee = callee;
1444 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1446 /* note: we use call_tail here as a pointer to the last inserted */
1447 if (x->call_head == NULL) {
1448 x->call_head = entry;
1450 if (entry->loop_depth == env->last_call->loop_depth) {
1451 /* same depth as the last one, enqueue after it */
1452 entry->next = env->last_call->next;
1453 env->last_call->next = entry;
1454 } else if (entry->loop_depth > x->call_head->loop_depth) {
1456 entry->next = x->call_head;
1457 x->call_head = entry;
1459 /* search the insertion point */
1462 for (p = x->call_head; p->next != NULL; p = p->next)
1463 if (entry->loop_depth > p->next->loop_depth)
1465 entry->next = p->next;
1469 env->last_call = entry;
1470 if (entry->next == NULL) {
1471 /* keep tail up to date */
1472 x->call_tail = entry;
1478 * Returns TRUE if the number of callers is 0 in the irg's environment,
1479 * hence this irg is a leave.
1481 INLINE static int is_leave(ir_graph *irg) {
1482 inline_irg_env *env = get_irg_link(irg);
1483 return env->n_call_nodes == 0;
1487 * Returns TRUE if the number of nodes in the callee is
1488 * smaller then size in the irg's environment.
1490 INLINE static int is_smaller(ir_graph *callee, int size) {
1491 inline_irg_env *env = get_irg_link(callee);
1492 return env->n_nodes < size;
1496 * Append the nodes of the list src to the nodes of the list in environment dst.
1498 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1499 call_entry *entry, *nentry;
1501 /* Note that the src list points to Call nodes in the inlined graph, but
1502 we need Call nodes in our graph. Luckily the inliner leaves this information
1503 in the link field. */
1504 for (entry = src; entry != NULL; entry = entry->next) {
1505 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1506 nentry->call = get_irn_link(entry->call);
1507 nentry->callee = entry->callee;
1508 nentry->next = NULL;
1509 nentry->loop_depth = entry->loop_depth;
1510 dst->call_tail->next = nentry;
1511 dst->call_tail = nentry;
1516 * Add the nodes of the list src in front to the nodes of the list dst.
1518 static call_entry *replace_entry_by_call_list(call_entry *dst, call_entry *src) {
1519 call_entry *entry, *nentry, *head, *tail;
1521 /* Note that the src list points to Call nodes in the inlined graph, but
1522 we need Call nodes in our graph. Luckily the inliner leaves this information
1523 in the link field. */
1525 for (entry = src; entry != NULL; entry = entry->next) {
1526 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1527 nentry->call = get_irn_link(entry->call);
1528 nentry->callee = entry->callee;
1529 nentry->next = NULL;
1530 nentry->loop_depth = entry->loop_depth + dst->loop_depth;
1534 tail->next = nentry;
1537 /* skip the head of dst */
1539 tail->next = dst->next;
1547 * Inlines small leave methods at call sites where the called address comes
1548 * from a Const node that references the entity representing the called
1550 * The size argument is a rough measure for the code size of the method:
1551 * Methods where the obstack containing the firm graph is smaller than
1554 void inline_leave_functions(int maxsize, int leavesize, int size, int ignore_runtime) {
1555 inline_irg_env *env;
1561 call_entry *entry, *tail;
1562 const call_entry *centry;
1563 pmap *copied_graphs;
1564 pmap_entry *pm_entry;
1566 rem = current_ir_graph;
1567 obstack_init(&temp_obst);
1569 /* a map for the copied graphs, used to inline recursive calls */
1570 copied_graphs = pmap_create();
1572 /* extend all irgs by a temporary data structure for inlining. */
1573 n_irgs = get_irp_n_irgs();
1574 for (i = 0; i < n_irgs; ++i)
1575 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1577 /* Precompute information in temporary data structure. */
1578 wenv.ignore_runtime = ignore_runtime;
1579 wenv.ignore_callers = 0;
1580 for (i = 0; i < n_irgs; ++i) {
1581 ir_graph *irg = get_irp_irg(i);
1583 assert(get_irg_phase_state(irg) != phase_building);
1584 free_callee_info(irg);
1586 assure_cf_loop(irg);
1587 wenv.x = get_irg_link(irg);
1588 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1591 /* -- and now inline. -- */
1593 /* Inline leaves recursively -- we might construct new leaves. */
1597 for (i = 0; i < n_irgs; ++i) {
1599 int phiproj_computed = 0;
1601 current_ir_graph = get_irp_irg(i);
1602 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1605 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1608 if (env->n_nodes > maxsize) break;
1611 callee = entry->callee;
1613 if (is_leave(callee) && (
1614 is_smaller(callee, leavesize) || (get_irg_inline_property(callee) >= irg_inline_forced))) {
1615 if (!phiproj_computed) {
1616 phiproj_computed = 1;
1617 collect_phiprojs(current_ir_graph);
1619 did_inline = inline_method(call, callee);
1622 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1624 /* was inlined, must be recomputed */
1625 phiproj_computed = 0;
1627 /* Do some statistics */
1628 env->got_inline = 1;
1629 --env->n_call_nodes;
1630 env->n_nodes += callee_env->n_nodes;
1631 --callee_env->n_callers;
1633 /* remove this call from the list */
1635 tail->next = entry->next;
1637 env->call_head = entry->next;
1643 env->call_tail = tail;
1645 } while (did_inline);
1647 /* inline other small functions. */
1648 for (i = 0; i < n_irgs; ++i) {
1650 int phiproj_computed = 0;
1652 current_ir_graph = get_irp_irg(i);
1653 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1655 /* note that the list of possible calls is updated during the process */
1657 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1662 callee = entry->callee;
1664 e = pmap_find(copied_graphs, callee);
1667 * Remap callee if we have a copy.
1668 * FIXME: Should we do this only for recursive Calls ?
1673 if (((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1674 (get_irg_inline_property(callee) >= irg_inline_forced))) {
1675 if (current_ir_graph == callee) {
1677 * Recursive call: we cannot directly inline because we cannot walk
1678 * the graph and change it. So we have to make a copy of the graph
1682 inline_irg_env *callee_env;
1686 * No copy yet, create one.
1687 * Note that recursive methods are never leaves, so it is sufficient
1688 * to test this condition here.
1690 copy = create_irg_copy(callee);
1692 /* create_irg_copy() destroys the Proj links, recompute them */
1693 phiproj_computed = 0;
1695 /* allocate new environment */
1696 callee_env = alloc_inline_irg_env();
1697 set_irg_link(copy, callee_env);
1699 assure_cf_loop(copy);
1700 wenv.x = callee_env;
1701 wenv.ignore_callers = 1;
1702 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1705 * Enter the entity of the original graph. This is needed
1706 * for inline_method(). However, note that ent->irg still points
1707 * to callee, NOT to copy.
1709 set_irg_entity(copy, get_irg_entity(callee));
1711 pmap_insert(copied_graphs, callee, copy);
1714 /* we have only one caller: the original graph */
1715 callee_env->n_callers = 1;
1716 callee_env->n_callers_orig = 1;
1718 if (! phiproj_computed) {
1719 phiproj_computed = 1;
1720 collect_phiprojs(current_ir_graph);
1722 did_inline = inline_method(call, callee);
1724 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1726 /* was inlined, must be recomputed */
1727 phiproj_computed = 0;
1729 /* callee was inline. Append it's call list. */
1730 env->got_inline = 1;
1731 --env->n_call_nodes;
1732 append_call_list(env, callee_env->call_head);
1733 env->n_call_nodes += callee_env->n_call_nodes;
1734 env->n_nodes += callee_env->n_nodes;
1735 --callee_env->n_callers;
1737 /* after we have inlined callee, all called methods inside callee
1738 are now called once more */
1739 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1740 inline_irg_env *penv = get_irg_link(centry->callee);
1744 /* remove this call from the list */
1746 tail->next = entry->next;
1748 env->call_head = entry->next;
1754 env->call_tail = tail;
1757 for (i = 0; i < n_irgs; ++i) {
1758 irg = get_irp_irg(i);
1759 env = (inline_irg_env *)get_irg_link(irg);
1761 if (env->got_inline) {
1762 optimize_graph_df(irg);
1765 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1766 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1767 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1768 env->n_callers_orig, env->n_callers,
1769 get_entity_name(get_irg_entity(irg))));
1773 /* kill the copied graphs: we don't need them anymore */
1774 foreach_pmap(copied_graphs, pm_entry) {
1775 ir_graph *copy = pm_entry->value;
1777 /* reset the entity, otherwise it will be deleted in the next step ... */
1778 set_irg_entity(copy, NULL);
1779 free_ir_graph(copy);
1781 pmap_destroy(copied_graphs);
1783 obstack_free(&temp_obst, NULL);
1784 current_ir_graph = rem;
1788 * Calculate the parameter weights for transmitting the address of a local variable.
1790 static unsigned calc_method_local_weight(ir_node *arg) {
1792 unsigned v, weight = 0;
1794 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1795 ir_node *succ = get_irn_out(arg, i);
1797 switch (get_irn_opcode(succ)) {
1800 /* Loads and Store can be removed */
1804 /* check if all args are constant */
1805 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1806 ir_node *idx = get_Sel_index(succ, j);
1807 if (! is_Const(idx))
1810 /* Check users on this Sel. Note: if a 0 is returned here, there was
1811 some unsupported node. */
1812 v = calc_method_local_weight(succ);
1815 /* we can kill one Sel with constant indexes, this is cheap */
1819 /* when looking backward we might find Id nodes */
1820 weight += calc_method_local_weight(succ);
1823 /* unoptimized tuple */
1824 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1825 ir_node *pred = get_Tuple_pred(succ, j);
1827 /* look for Proj(j) */
1828 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1829 ir_node *succ_succ = get_irn_out(succ, k);
1830 if (is_Proj(succ_succ)) {
1831 if (get_Proj_proj(succ_succ) == j) {
1833 weight += calc_method_local_weight(succ_succ);
1836 /* this should NOT happen */
1844 /* any other node: unsupported yet or bad. */
1852 * Calculate the parameter weights for transmitting the address of a local variable.
1854 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1855 ir_entity *ent = get_irg_entity(irg);
1857 int nparams, i, proj_nr;
1858 ir_node *irg_args, *arg;
1860 mtp = get_entity_type(ent);
1861 nparams = get_method_n_params(mtp);
1863 /* allocate a new array. currently used as 'analysed' flag */
1864 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1866 /* If the method haven't parameters we have nothing to do. */
1870 assure_irg_outs(irg);
1871 irg_args = get_irg_args(irg);
1872 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1873 arg = get_irn_out(irg_args, i);
1874 proj_nr = get_Proj_proj(arg);
1875 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1880 * Calculate the benefice for transmitting an local variable address.
1881 * After inlining, the local variable might be transformed into a
1882 * SSA variable by scalar_replacement().
1884 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1885 inline_irg_env *env = get_irg_link(callee);
1887 if (env->local_weights != NULL) {
1888 if (pos < ARR_LEN(env->local_weights))
1889 return env->local_weights[pos];
1893 analyze_irg_local_weights(env, callee);
1895 if (pos < ARR_LEN(env->local_weights))
1896 return env->local_weights[pos];
1901 * Calculate a benefice value for inlining the given call.
1903 * @param call the call node we have to inspect
1904 * @param callee the called graph
1905 * @param local_adr set after return if an address of a local variable is
1906 * transmitted as a parameter
1908 static int calc_inline_benefice(ir_node *call, ir_graph *callee, unsigned *local_adr) {
1909 ir_entity *ent = get_irg_entity(callee);
1913 int i, n_params, all_const;
1916 inline_irg_env *curr_env, *callee_env;
1918 if (get_entity_additional_properties(ent) & mtp_property_noreturn) {
1919 /* do NOT inline noreturn calls */
1923 /* costs for every passed parameter */
1924 n_params = get_Call_n_params(call);
1925 mtp = get_entity_type(ent);
1926 cc = get_method_calling_convention(mtp);
1927 if (cc & cc_reg_param) {
1928 /* register parameter, smaller costs for register parameters */
1929 int max_regs = cc & ~cc_bits;
1931 if (max_regs < n_params)
1932 weight += max_regs * 2 + (n_params - max_regs) * 5;
1934 weight += n_params * 2;
1936 /* parameters are passed an stack */
1937 weight += 5 * n_params;
1940 /* constant parameters improve the benefice */
1941 frame_ptr = get_irg_frame(current_ir_graph);
1943 for (i = 0; i < n_params; ++i) {
1944 ir_node *param = get_Call_param(call, i);
1946 if (is_Const(param)) {
1947 weight += get_method_param_weight(ent, i);
1950 if (is_SymConst(param))
1951 weight += get_method_param_weight(ent, i);
1952 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1954 * An address of a local variable is transmitted. After inlining,
1955 * scalar_replacement might be able to remove the local variable,
1958 v = get_method_local_adress_weight(callee, i);
1966 callee_env = get_irg_link(callee);
1967 if (get_entity_visibility(ent) == visibility_local &&
1968 callee_env->n_callers_orig == 1 &&
1969 callee != current_ir_graph) {
1970 /* we are the only caller, give big bonus */
1974 /* do not inline big functions */
1975 weight -= callee_env->n_nodes;
1977 /* reduce the benefice if the current function is already big */
1978 curr_env = get_irg_link(current_ir_graph);
1979 weight -= curr_env->n_nodes / 50;
1981 /* give a bonus for functions with one block */
1982 if (callee_env->n_blocks == 1)
1983 weight = weight * 3 / 2;
1985 /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1986 else if (callee_env->n_nodes < 20 && !callee_env->recursive)
1989 /* and finally for leaves: they do not increase the register pressure
1990 because of callee safe registers */
1991 else if (callee_env->n_call_nodes == 0)
1995 * Reduce the weight for recursive function IFF not all arguments are const.
1996 * inlining recursive functions is rarely good.
1998 if (callee_env->recursive && !all_const)
2002 * All arguments constant is probably a good sign, give an extra bonus
2011 * Heuristic inliner. Calculates a benefice value for every call and inlines
2012 * those calls with a value higher than the threshold.
2014 void inline_functions(int maxsize, int inline_threshold) {
2015 inline_irg_env *env;
2020 call_entry *curr_call, **last_call;
2021 const call_entry *centry;
2022 pmap *copied_graphs;
2023 pmap_entry *pm_entry;
2025 rem = current_ir_graph;
2026 obstack_init(&temp_obst);
2028 /* a map for the copied graphs, used to inline recursive calls */
2029 copied_graphs = pmap_create();
2031 /* extend all irgs by a temporary data structure for inlining. */
2032 n_irgs = get_irp_n_irgs();
2033 for (i = 0; i < n_irgs; ++i)
2034 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
2036 /* Precompute information in temporary data structure. */
2037 wenv.ignore_runtime = 0;
2038 wenv.ignore_callers = 0;
2039 for (i = 0; i < n_irgs; ++i) {
2040 ir_graph *irg = get_irp_irg(i);
2042 assert(get_irg_phase_state(irg) != phase_building);
2043 free_callee_info(irg);
2045 wenv.x = get_irg_link(irg);
2046 wenv.last_call = NULL;
2047 assure_cf_loop(irg);
2048 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
2051 /* -- and now inline. -- */
2052 for (i = 0; i < n_irgs; ++i) {
2053 int phiproj_computed = 0;
2055 ir_graph *irg = get_irp_irg(i);
2057 current_ir_graph = irg;
2058 env = get_irg_link(irg);
2060 /* note that the list of possible calls is updated during the process */
2061 last_call = &env->call_head;
2062 for (curr_call = env->call_head; curr_call != NULL;) {
2068 if (env->n_nodes > maxsize) break;
2070 call = curr_call->call;
2071 callee = curr_call->callee;
2073 e = pmap_find(copied_graphs, callee);
2076 * Remap callee if we have a copy.
2077 * FIXME: Should we do this only for recursive Calls ?
2082 /* calculate the benefice on the original call to prevent excessive inlining */
2084 benefice = calc_inline_benefice(call, callee, &local_adr);
2085 DB((dbg, LEVEL_2, "In %+F Call %+F has benefice %d\n", irg, callee, benefice));
2087 if (benefice > -inline_threshold ||
2088 (get_irg_inline_property(callee) >= irg_inline_forced)) {
2089 if (current_ir_graph == callee) {
2091 * Recursive call: we cannot directly inline because we cannot walk
2092 * the graph and change it. So we have to make a copy of the graph
2096 inline_irg_env *callee_env;
2100 * No copy yet, create one.
2101 * Note that recursive methods are never leaves, so it is sufficient
2102 * to test this condition here.
2104 copy = create_irg_copy(callee);
2106 /* create_irg_copy() destroys the Proj links, recompute them */
2107 phiproj_computed = 0;
2109 /* allocate new environment */
2110 callee_env = alloc_inline_irg_env();
2111 set_irg_link(copy, callee_env);
2113 assure_cf_loop(copy);
2114 wenv.x = callee_env;
2115 wenv.ignore_callers = 1;
2116 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2119 * Enter the entity of the original graph. This is needed
2120 * for inline_method(). However, note that ent->irg still points
2121 * to callee, NOT to copy.
2123 set_irg_entity(copy, get_irg_entity(callee));
2125 pmap_insert(copied_graphs, callee, copy);
2128 /* we have only one caller: the original graph */
2129 callee_env->n_callers = 1;
2130 callee_env->n_callers_orig = 1;
2132 if (! phiproj_computed) {
2133 phiproj_computed = 1;
2134 collect_phiprojs(current_ir_graph);
2136 did_inline = inline_method(call, callee);
2138 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
2140 /* was inlined, must be recomputed */
2141 phiproj_computed = 0;
2143 /* after we have inlined callee, all called methods inside callee
2144 are now called once more */
2145 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
2146 inline_irg_env *penv = get_irg_link(centry->callee);
2150 /* callee was inline. Append it's call list. */
2151 env->got_inline = 1;
2153 env->local_vars = 1;
2154 --env->n_call_nodes;
2155 curr_call = replace_entry_by_call_list(curr_call, callee_env->call_head);
2156 env->n_call_nodes += callee_env->n_call_nodes;
2157 env->n_nodes += callee_env->n_nodes;
2158 --callee_env->n_callers;
2160 /* remove the current call entry from the list */
2161 *last_call = curr_call;
2165 last_call = &curr_call->next;
2166 curr_call = curr_call->next;
2171 for (i = 0; i < n_irgs; ++i) {
2172 ir_graph *irg = get_irp_irg(i);
2174 env = get_irg_link(irg);
2175 if (env->got_inline) {
2176 /* this irg got calls inlined: optimize it */
2178 /* scalar replacement does not work well with Tuple nodes, so optimize them away */
2179 optimize_graph_df(irg);
2181 if (env->local_vars) {
2182 if (scalar_replacement_opt(irg)) {
2183 optimize_graph_df(irg);
2188 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2189 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2190 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2191 env->n_callers_orig, env->n_callers,
2192 get_entity_name(get_irg_entity(irg))));
2196 /* kill the copied graphs: we don't need them anymore */
2197 foreach_pmap(copied_graphs, pm_entry) {
2198 ir_graph *copy = pm_entry->value;
2200 /* reset the entity, otherwise it will be deleted in the next step ... */
2201 set_irg_entity(copy, NULL);
2202 free_ir_graph(copy);
2204 pmap_destroy(copied_graphs);
2206 obstack_free(&temp_obst, NULL);
2207 current_ir_graph = rem;
2210 void firm_init_inline(void) {
2211 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");