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"
49 #include "adt/pqueue.h"
53 #include "irbackedge_t.h"
54 #include "opt_inline_t.h"
59 #include "analyze_irg_args.h"
60 #include "iredges_t.h"
64 #include "iropt_dbg.h"
66 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
68 /*------------------------------------------------------------------*/
69 /* Routines for dead node elimination / copying garbage collection */
71 /*------------------------------------------------------------------*/
74 * Remember the new node in the old node by using a field all nodes have.
76 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
79 * Get this new node, before the old node is forgotten.
81 #define get_new_node(oldn) get_irn_link(oldn)
84 * Check if a new node was set.
86 #define has_new_node(n) (get_new_node(n) != NULL)
89 * We use the block_visited flag to mark that we have computed the
90 * number of useful predecessors for this block.
91 * Further we encode the new arity in this flag in the old blocks.
92 * Remembering the arity is useful, as it saves a lot of pointer
93 * accesses. This function is called for all Phi and Block nodes
97 compute_new_arity(ir_node *b) {
98 int i, res, irn_arity;
101 irg_v = get_irg_block_visited(current_ir_graph);
102 block_v = get_Block_block_visited(b);
103 if (block_v >= irg_v) {
104 /* we computed the number of preds for this block and saved it in the
106 return block_v - irg_v;
108 /* compute the number of good predecessors */
109 res = irn_arity = get_irn_arity(b);
110 for (i = 0; i < irn_arity; i++)
111 if (is_Bad(get_irn_n(b, i))) res--;
112 /* save it in the flag. */
113 set_Block_block_visited(b, irg_v + res);
119 * Copies the node to the new obstack. The Ins of the new node point to
120 * the predecessors on the old obstack. For block/phi nodes not all
121 * predecessors might be copied. n->link points to the new node.
122 * For Phi and Block nodes the function allocates in-arrays with an arity
123 * only for useful predecessors. The arity is determined by counting
124 * the non-bad predecessors of the block.
126 * @param n The node to be copied
127 * @param env if non-NULL, the node number attribute will be copied to the new node
129 * Note: Also used for loop unrolling.
131 static void copy_node(ir_node *n, void *env) {
134 ir_op *op = get_irn_op(n);
138 /* node copied already */
140 } else if (op == op_Block) {
142 new_arity = compute_new_arity(n);
143 n->attr.block.graph_arr = NULL;
145 block = get_nodes_block(n);
147 new_arity = compute_new_arity(block);
149 new_arity = get_irn_arity(n);
152 nn = new_ir_node(get_irn_dbg_info(n),
159 /* Copy the attributes. These might point to additional data. If this
160 was allocated on the old obstack the pointers now are dangling. This
161 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
162 if (op == op_Block) {
163 /* we cannot allow blocks WITHOUT macroblock input */
164 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
166 copy_node_attr(n, nn);
170 int copy_node_nr = env != NULL;
172 /* for easier debugging, we want to copy the node numbers too */
173 nn->node_nr = n->node_nr;
179 hook_dead_node_elim_subst(current_ir_graph, n, nn);
183 * Copies new predecessors of old node to new node remembered in link.
184 * Spare the Bad predecessors of Phi and Block nodes.
186 static void copy_preds(ir_node *n, void *env) {
191 nn = get_new_node(n);
194 /* copy the macro block header */
195 ir_node *mbh = get_Block_MacroBlock(n);
198 /* this block is a macroblock header */
199 set_Block_MacroBlock(nn, nn);
201 /* get the macro block header */
202 ir_node *nmbh = get_new_node(mbh);
203 assert(nmbh != NULL);
204 set_Block_MacroBlock(nn, nmbh);
207 /* Don't copy Bad nodes. */
209 irn_arity = get_irn_arity(n);
210 for (i = 0; i < irn_arity; i++) {
211 if (! is_Bad(get_irn_n(n, i))) {
212 ir_node *pred = get_irn_n(n, i);
213 set_irn_n(nn, j, get_new_node(pred));
217 /* repair the block visited flag from above misuse. Repair it in both
218 graphs so that the old one can still be used. */
219 set_Block_block_visited(nn, 0);
220 set_Block_block_visited(n, 0);
221 /* Local optimization could not merge two subsequent blocks if
222 in array contained Bads. Now it's possible.
223 We don't call optimize_in_place as it requires
224 that the fields in ir_graph are set properly. */
225 if ((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(sizeof(*rebirth_obst));
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(sizeof(res[0]));
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 #ifndef FIRM_ENABLE_HOOKS
679 assert(0 && "need hooks enabled");
682 register_hook(hook_dead_node_elim, &res->dead_node_elim);
683 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
688 * Free a Survive DCE environment.
690 void free_survive_dce(survive_dce_t *sd) {
691 obstack_free(&sd->obst, NULL);
692 pmap_destroy(sd->places);
693 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
694 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
699 * Register a node pointer to be patched upon DCE.
700 * When DCE occurs, the node pointer specified by @p place will be
701 * patched to the new address of the node it is pointing to.
703 * @param sd The Survive DCE environment.
704 * @param place The address of the node pointer.
706 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
707 if (*place != NULL) {
708 ir_node *irn = *place;
709 survive_dce_list_t *curr = pmap_get(sd->places, irn);
710 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
715 pmap_insert(sd->places, irn, nw);
719 /*--------------------------------------------------------------------*/
720 /* Functionality for inlining */
721 /*--------------------------------------------------------------------*/
724 * Copy node for inlineing. Updates attributes that change when
725 * inlineing but not for dead node elimination.
727 * Copies the node by calling copy_node() and then updates the entity if
728 * it's a local one. env must be a pointer of the frame type of the
729 * inlined procedure. The new entities must be in the link field of
732 static void copy_node_inline(ir_node *n, void *env) {
734 ir_type *frame_tp = (ir_type *)env;
738 nn = get_new_node (n);
740 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
741 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
743 } else if (is_Block(n)) {
744 nn = get_new_node (n);
745 nn->attr.block.irg = current_ir_graph;
750 * Copies new predecessors of old node and move constants to
753 static void copy_preds_inline(ir_node *n, void *env) {
757 nn = skip_Id(get_new_node(n));
758 if (is_irn_constlike(nn)) {
759 /* move Constants into the start block */
760 set_nodes_block(nn, get_irg_start_block(current_ir_graph));
762 n = identify_remember(current_ir_graph->value_table, nn);
771 * Walker: checks if P_value_arg_base is used.
773 static void find_addr(ir_node *node, void *env) {
774 int *allow_inline = env;
776 is_Start(get_Proj_pred(node)) &&
777 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
779 } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
781 * Refuse to inline alloca call unless user explicitly forced so as this
782 * may change program's memory overhead drastically when the function
783 * using alloca is called in loop. In GCC present in SPEC2000 inlining
784 * into schedule_block cause it to require 2GB of ram instead of 256MB.
786 * Sorryly this is true with our implementation also.
787 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
788 * disables inlining of functions using VLA (with are completely save).
791 * - add a flag to the Alloc node for "real" alloca() calls
792 * - add a new Stack-Restore node at the end of a function using alloca()
799 * Check if we can inline a given call.
800 * Currently, we cannot inline two cases:
801 * - call with compound arguments
802 * - graphs that take the address of a parameter
804 * check these conditions here
806 static int can_inline(ir_node *call, ir_graph *called_graph) {
807 ir_type *call_type = get_Call_type(call);
808 int params, ress, i, res;
809 assert(is_Method_type(call_type));
811 params = get_method_n_params(call_type);
812 ress = get_method_n_ress(call_type);
814 /* check parameters for compound arguments */
815 for (i = 0; i < params; ++i) {
816 ir_type *p_type = get_method_param_type(call_type, i);
818 if (is_compound_type(p_type))
822 /* check results for compound arguments */
823 for (i = 0; i < ress; ++i) {
824 ir_type *r_type = get_method_res_type(call_type, i);
826 if (is_compound_type(r_type))
831 irg_walk_graph(called_graph, find_addr, NULL, &res);
837 exc_handler = 0, /**< There is a handler. */
838 exc_to_end = 1, /**< Branches to End. */
839 exc_no_handler = 2 /**< Exception handling not represented. */
842 /* Inlines a method at the given call site. */
843 int inline_method(ir_node *call, ir_graph *called_graph) {
845 ir_node *post_call, *post_bl;
846 ir_node *in[pn_Start_max];
847 ir_node *end, *end_bl, *block;
852 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
853 enum exc_mode exc_handling;
854 ir_type *called_frame, *curr_frame, *mtp, *ctp;
857 irg_inline_property prop = get_irg_inline_property(called_graph);
858 unsigned long visited;
860 if (prop == irg_inline_forbidden)
863 ent = get_irg_entity(called_graph);
865 mtp = get_entity_type(ent);
866 ctp = get_Call_type(call);
867 if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
868 /* this is a bad feature of C: without a prototype, we can can call a function with less
869 parameters than needed. Currently we don't support this, although it would be
870 to use Unknown than. */
874 /* Argh, compiling C has some bad consequences:
875 the call type AND the method type might be different.
876 It is implementation defendant what happens in that case.
877 We support inlining, if the bitsize of the types matches AND
878 the same arithmetic is used. */
879 n_params = get_method_n_params(mtp);
880 for (i = n_params - 1; i >= 0; --i) {
881 ir_type *param_tp = get_method_param_type(mtp, i);
882 ir_type *arg_tp = get_method_param_type(ctp, i);
884 if (param_tp != arg_tp) {
885 ir_mode *pmode = get_type_mode(param_tp);
886 ir_mode *amode = get_type_mode(arg_tp);
888 if (pmode == NULL || amode == NULL)
890 if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
892 if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
894 /* otherwise we can simply "reinterpret" the bits */
898 irg = get_irn_irg(call);
901 * We cannot inline a recursive call. The graph must be copied before
902 * the call the inline_method() using create_irg_copy().
904 if (called_graph == irg)
908 * currently, we cannot inline two cases:
909 * - call with compound arguments
910 * - graphs that take the address of a parameter
912 if (! can_inline(call, called_graph))
915 rem = current_ir_graph;
916 current_ir_graph = irg;
918 DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
920 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
921 rem_opt = get_opt_optimize();
924 /* Handle graph state */
925 assert(get_irg_phase_state(irg) != phase_building);
926 assert(get_irg_pinned(irg) == op_pin_state_pinned);
927 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
928 set_irg_outs_inconsistent(irg);
929 set_irg_extblk_inconsistent(irg);
930 set_irg_doms_inconsistent(irg);
931 set_irg_loopinfo_inconsistent(irg);
932 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
934 /* -- Check preconditions -- */
935 assert(is_Call(call));
937 /* here we know we WILL inline, so inform the statistics */
938 hook_inline(call, called_graph);
940 /* -- Decide how to handle exception control flow: Is there a handler
941 for the Call node, or do we branch directly to End on an exception?
943 0 There is a handler.
945 2 Exception handling not represented in Firm. -- */
947 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
948 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
949 long proj_nr = get_Proj_proj(proj);
950 if (proj_nr == pn_Call_X_except) Xproj = proj;
951 if (proj_nr == pn_Call_M_except) Mproj = proj;
953 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
954 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
955 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
958 /* create the argument tuple */
959 NEW_ARR_A(ir_type *, args_in, n_params);
961 block = get_nodes_block(call);
962 for (i = n_params - 1; i >= 0; --i) {
963 ir_node *arg = get_Call_param(call, i);
964 ir_type *param_tp = get_method_param_type(mtp, i);
965 ir_mode *mode = get_type_mode(param_tp);
967 if (mode != get_irn_mode(arg)) {
968 arg = new_r_Conv(irg, block, arg, mode);
974 the procedure and later replaces the Start node of the called graph.
975 Post_call is the old Call node and collects the results of the called
976 graph. Both will end up being a tuple. -- */
977 post_bl = get_nodes_block(call);
978 set_irg_current_block(irg, post_bl);
979 /* XxMxPxPxPxT of Start + parameter of Call */
980 in[pn_Start_X_initial_exec] = new_Jmp();
981 in[pn_Start_M] = get_Call_mem(call);
982 in[pn_Start_P_frame_base] = get_irg_frame(irg);
983 in[pn_Start_P_tls] = get_irg_tls(irg);
984 in[pn_Start_T_args] = new_Tuple(n_params, args_in);
985 /* in[pn_Start_P_value_arg_base] = ??? */
986 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
987 pre_call = new_Tuple(pn_Start_max - 1, 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 called_frame = get_irg_frame_type(called_graph);
1026 curr_frame = get_irg_frame_type(irg);
1027 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
1028 ir_entity *new_ent, *old_ent;
1029 old_ent = get_class_member(called_frame, i);
1030 new_ent = copy_entity_own(old_ent, curr_frame);
1031 set_entity_link(old_ent, new_ent);
1034 /* visited is > than that of called graph. With this trick visited will
1035 remain unchanged so that an outer walker, e.g., searching the call nodes
1036 to inline, calling this inline will not visit the inlined nodes. */
1037 set_irg_visited(irg, get_irg_visited(irg)-1);
1039 /* -- Performing dead node elimination inlines the graph -- */
1040 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1042 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
1043 get_irg_frame_type(called_graph));
1045 /* Repair called_graph */
1046 set_irg_visited(called_graph, get_irg_visited(irg));
1047 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
1048 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1050 /* -- Merge the end of the inlined procedure with the call site -- */
1051 /* We will turn the old Call node into a Tuple with the following
1054 0: Phi of all Memories of Return statements.
1055 1: Jmp from new Block that merges the control flow from all exception
1056 predecessors of the old end block.
1057 2: Tuple of all arguments.
1058 3: Phi of Exception memories.
1059 In case the old Call directly branches to End on an exception we don't
1060 need the block merging all exceptions nor the Phi of the exception
1064 /* -- Precompute some values -- */
1065 end_bl = get_new_node(get_irg_end_block(called_graph));
1066 end = get_new_node(get_irg_end(called_graph));
1067 arity = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret */
1068 n_res = get_method_n_ress(get_Call_type(call));
1070 res_pred = xmalloc(n_res * sizeof(*res_pred));
1071 cf_pred = xmalloc(arity * sizeof(*res_pred));
1073 set_irg_current_block(irg, post_bl); /* just to make sure */
1075 /* -- archive keepalives -- */
1076 irn_arity = get_irn_arity(end);
1077 for (i = 0; i < irn_arity; i++) {
1078 ir_node *ka = get_End_keepalive(end, i);
1080 add_End_keepalive(get_irg_end(irg), ka);
1083 /* The new end node will die. We need not free as the in array is on the obstack:
1084 copy_node() only generated 'D' arrays. */
1086 /* -- Replace Return nodes by Jump nodes. -- */
1088 for (i = 0; i < arity; i++) {
1090 ret = get_Block_cfgpred(end_bl, i);
1091 if (is_Return(ret)) {
1092 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1096 set_irn_in(post_bl, n_ret, cf_pred);
1098 /* -- Build a Tuple for all results of the method.
1099 Add Phi node if there was more than one Return. -- */
1100 turn_into_tuple(post_call, pn_Call_max);
1101 /* First the Memory-Phi */
1103 for (i = 0; i < arity; i++) {
1104 ret = get_Block_cfgpred(end_bl, i);
1105 if (is_Return(ret)) {
1106 cf_pred[n_ret] = get_Return_mem(ret);
1110 phi = new_Phi(n_ret, cf_pred, mode_M);
1111 set_Tuple_pred(call, pn_Call_M_regular, phi);
1112 /* Conserve Phi-list for further inlinings -- but might be optimized */
1113 if (get_nodes_block(phi) == post_bl) {
1114 set_irn_link(phi, get_irn_link(post_bl));
1115 set_irn_link(post_bl, phi);
1117 /* Now the real results */
1119 for (j = 0; j < n_res; j++) {
1121 for (i = 0; i < arity; i++) {
1122 ret = get_Block_cfgpred(end_bl, i);
1123 if (is_Return(ret)) {
1124 cf_pred[n_ret] = get_Return_res(ret, j);
1129 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1133 /* Conserve Phi-list for further inlinings -- but might be optimized */
1134 if (get_nodes_block(phi) == post_bl) {
1135 set_Phi_next(phi, get_Block_phis(post_bl));
1136 set_Block_phis(post_bl, phi);
1139 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1141 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1143 /* handle the regular call */
1144 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1146 /* For now, we cannot inline calls with value_base */
1147 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1149 /* Finally the exception control flow.
1150 We have two (three) possible situations:
1151 First if the Call branches to an exception handler: We need to add a Phi node to
1152 collect the memory containing the exception objects. Further we need
1153 to add another block to get a correct representation of this Phi. To
1154 this block we add a Jmp that resolves into the X output of the Call
1155 when the Call is turned into a tuple.
1156 Second the Call branches to End, the exception is not handled. Just
1157 add all inlined exception branches to the End node.
1158 Third: there is no Exception edge at all. Handle as case two. */
1159 if (exc_handling == exc_handler) {
1161 for (i = 0; i < arity; i++) {
1163 ret = get_Block_cfgpred(end_bl, i);
1164 irn = skip_Proj(ret);
1165 if (is_fragile_op(irn) || is_Raise(irn)) {
1166 cf_pred[n_exc] = ret;
1171 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1172 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1173 /* The Phi for the memories with the exception objects */
1175 for (i = 0; i < arity; i++) {
1177 ret = skip_Proj(get_Block_cfgpred(end_bl, i));
1179 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1181 } else if (is_fragile_op(ret)) {
1182 /* We rely that all cfops have the memory output at the same position. */
1183 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1185 } else if (is_Raise(ret)) {
1186 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1190 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1192 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1193 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1196 ir_node *main_end_bl;
1197 int main_end_bl_arity;
1198 ir_node **end_preds;
1200 /* assert(exc_handling == 1 || no exceptions. ) */
1202 for (i = 0; i < arity; i++) {
1203 ir_node *ret = get_Block_cfgpred(end_bl, i);
1204 ir_node *irn = skip_Proj(ret);
1206 if (is_fragile_op(irn) || is_Raise(irn)) {
1207 cf_pred[n_exc] = ret;
1211 main_end_bl = get_irg_end_block(irg);
1212 main_end_bl_arity = get_irn_arity(main_end_bl);
1213 end_preds = xmalloc((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1215 for (i = 0; i < main_end_bl_arity; ++i)
1216 end_preds[i] = get_irn_n(main_end_bl, i);
1217 for (i = 0; i < n_exc; ++i)
1218 end_preds[main_end_bl_arity + i] = cf_pred[i];
1219 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1220 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1221 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1227 /* -- Turn CSE back on. -- */
1228 set_optimize(rem_opt);
1229 current_ir_graph = rem;
1234 /********************************************************************/
1235 /* Apply inlineing to small methods. */
1236 /********************************************************************/
1238 static struct obstack temp_obst;
1240 /** Represents a possible inlinable call in a graph. */
1241 typedef struct _call_entry call_entry;
1242 struct _call_entry {
1243 ir_node *call; /**< the Call node */
1244 ir_graph *callee; /**< the callee IR-graph called here */
1245 call_entry *next; /**< for linking the next one */
1246 int loop_depth; /**< the loop depth of this call */
1247 unsigned local_adr : 1;
1251 * environment for inlining small irgs
1253 typedef struct _inline_env_t {
1254 struct obstack obst; /**< an obstack where call_entries are allocated on. */
1255 call_entry *head; /**< the head of the call entry list */
1256 call_entry *tail; /**< the tail of the call entry list */
1260 * Returns the irg called from a Call node. If the irg is not
1261 * known, NULL is returned.
1263 * @param call the call node
1265 static ir_graph *get_call_called_irg(ir_node *call) {
1268 addr = get_Call_ptr(call);
1269 if (is_Global(addr)) {
1270 ir_entity *ent = get_Global_entity(addr);
1271 return get_entity_irg(ent);
1278 * Walker: Collect all calls to known graphs inside a graph.
1280 static void collect_calls(ir_node *call, void *env) {
1281 if (is_Call(call)) {
1282 ir_graph *called_irg = get_call_called_irg(call);
1284 if (called_irg != NULL) {
1285 /* The Call node calls a locally defined method. Remember to inline. */
1286 inline_env_t *ienv = env;
1287 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1289 entry->callee = called_irg;
1291 entry->loop_depth = 0;
1293 if (ienv->tail == NULL)
1296 ienv->tail->next = entry;
1303 * Inlines all small methods at call sites where the called address comes
1304 * from a Const node that references the entity representing the called
1306 * The size argument is a rough measure for the code size of the method:
1307 * Methods where the obstack containing the firm graph is smaller than
1310 void inline_small_irgs(ir_graph *irg, int size) {
1311 ir_graph *rem = current_ir_graph;
1315 current_ir_graph = irg;
1316 /* Handle graph state */
1317 assert(get_irg_phase_state(irg) != phase_building);
1318 free_callee_info(irg);
1320 /* Find Call nodes to inline.
1321 (We can not inline during a walk of the graph, as inlineing the same
1322 method several times changes the visited flag of the walked graph:
1323 after the first inlineing visited of the callee equals visited of
1324 the caller. With the next inlineing both are increased.) */
1325 obstack_init(&env.obst);
1326 env.head = env.tail = NULL;
1327 irg_walk_graph(irg, NULL, collect_calls, &env);
1329 if (env.head != NULL) {
1330 /* There are calls to inline */
1331 collect_phiprojs(irg);
1333 for (entry = env.head; entry != NULL; entry = entry->next) {
1334 ir_graph *callee = entry->callee;
1335 irg_inline_property prop = get_irg_inline_property(callee);
1337 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1338 /* do not inline forbidden / weak graphs */
1342 if (prop >= irg_inline_forced ||
1343 _obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst) < size) {
1344 inline_method(entry->call, callee);
1348 obstack_free(&env.obst, NULL);
1349 current_ir_graph = rem;
1353 * Environment for inlining irgs.
1356 unsigned n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1357 unsigned n_blocks; /**< Number of Blocks in graph without Start and End block. */
1358 unsigned n_nodes_orig; /**< for statistics */
1359 unsigned n_call_nodes; /**< Number of Call nodes in the graph. */
1360 unsigned n_call_nodes_orig; /**< for statistics */
1361 unsigned n_callers; /**< Number of known graphs that call this graphs. */
1362 unsigned n_callers_orig; /**< for statistics */
1363 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1364 unsigned local_vars:1; /**< Set, if a inlined function gets the address of an inlined variable. */
1365 unsigned recursive:1; /**< Set, if this function is self recursive. */
1366 call_entry *call_head; /**< The head of the list of all call nodes in this graph. */
1367 call_entry *call_tail; /**< The tail of the list of all call nodes in this graph .*/
1368 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1372 * Allocate a new environment for inlining.
1374 static inline_irg_env *alloc_inline_irg_env(void) {
1375 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1376 env->n_nodes = -2; /* do not count count Start, End */
1377 env->n_blocks = -2; /* do not count count Start, End Block */
1378 env->n_nodes_orig = -2; /* do not count Start, End */
1379 env->call_head = NULL;
1380 env->call_tail = NULL;
1381 env->n_call_nodes = 0;
1382 env->n_call_nodes_orig = 0;
1384 env->n_callers_orig = 0;
1385 env->got_inline = 0;
1386 env->local_vars = 0;
1388 env->local_weights = NULL;
1392 typedef struct walker_env {
1393 inline_irg_env *x; /**< the inline environment */
1394 char ignore_runtime; /**< the ignore runtime flag */
1395 char ignore_callers; /**< if set, do change callers data */
1399 * post-walker: collect all calls in the inline-environment
1400 * of a graph and sum some statistics.
1402 static void collect_calls2(ir_node *call, void *ctx) {
1404 inline_irg_env *x = env->x;
1405 ir_opcode code = get_irn_opcode(call);
1409 /* count meaningful nodes in irg */
1410 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1411 if (code != iro_Block) {
1419 if (code != iro_Call) return;
1421 /* check, if it's a runtime call */
1422 if (env->ignore_runtime) {
1423 ir_node *symc = get_Call_ptr(call);
1425 if (is_Global(symc)) {
1426 ir_entity *ent = get_Global_entity(symc);
1428 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1433 /* collect all call nodes */
1435 ++x->n_call_nodes_orig;
1437 callee = get_call_called_irg(call);
1438 if (callee != NULL) {
1439 if (! env->ignore_callers) {
1440 inline_irg_env *callee_env = get_irg_link(callee);
1441 /* count all static callers */
1442 ++callee_env->n_callers;
1443 ++callee_env->n_callers_orig;
1445 if (callee == current_ir_graph)
1448 /* link it in the list of possible inlinable entries */
1449 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1451 entry->callee = callee;
1453 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1455 entry->next = x->call_head;
1456 x->call_head = entry;
1458 if (entry->next == NULL) {
1459 /* keep tail up to date */
1460 x->call_tail = entry;
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 * Append the nodes of the list src to the nodes of the list in environment dst.
1486 static void append_call_list(inline_irg_env *dst, call_entry *src) {
1487 call_entry *entry, *nentry;
1489 /* Note that the src list points to Call nodes in the inlined graph, but
1490 we need Call nodes in our graph. Luckily the inliner leaves this information
1491 in the link field. */
1492 for (entry = src; entry != NULL; entry = entry->next) {
1493 nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1494 nentry->call = get_irn_link(entry->call);
1495 nentry->callee = entry->callee;
1496 nentry->next = NULL;
1497 nentry->loop_depth = entry->loop_depth;
1498 dst->call_tail->next = nentry;
1499 dst->call_tail = nentry;
1503 static call_entry *duplicate_call_entry(const call_entry *entry,
1504 int loop_depth_delta, ir_node *new_call)
1506 call_entry *nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1507 nentry->call = new_call;
1508 nentry->callee = entry->callee;
1509 nentry->next = NULL;
1510 nentry->loop_depth = entry->loop_depth + loop_depth_delta;
1516 * Inlines small leave methods at call sites where the called address comes
1517 * from a Const node that references the entity representing the called
1519 * The size argument is a rough measure for the code size of the method:
1520 * Methods where the obstack containing the firm graph is smaller than
1523 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
1524 unsigned size, int ignore_runtime)
1526 inline_irg_env *env;
1532 call_entry *entry, *tail;
1533 const call_entry *centry;
1534 pmap *copied_graphs;
1535 pmap_entry *pm_entry;
1537 rem = current_ir_graph;
1538 obstack_init(&temp_obst);
1540 /* a map for the copied graphs, used to inline recursive calls */
1541 copied_graphs = pmap_create();
1543 /* extend all irgs by a temporary data structure for inlining. */
1544 n_irgs = get_irp_n_irgs();
1545 for (i = 0; i < n_irgs; ++i)
1546 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1548 /* Precompute information in temporary data structure. */
1549 wenv.ignore_runtime = ignore_runtime;
1550 wenv.ignore_callers = 0;
1551 for (i = 0; i < n_irgs; ++i) {
1552 ir_graph *irg = get_irp_irg(i);
1554 assert(get_irg_phase_state(irg) != phase_building);
1555 free_callee_info(irg);
1557 assure_cf_loop(irg);
1558 wenv.x = get_irg_link(irg);
1559 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1562 /* -- and now inline. -- */
1564 /* Inline leaves recursively -- we might construct new leaves. */
1568 for (i = 0; i < n_irgs; ++i) {
1570 int phiproj_computed = 0;
1572 current_ir_graph = get_irp_irg(i);
1573 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1576 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1578 irg_inline_property prop;
1580 if (env->n_nodes > maxsize)
1584 callee = entry->callee;
1586 prop = get_irg_inline_property(callee);
1587 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1588 /* do not inline forbidden / weak graphs */
1592 if (is_leave(callee) && (
1593 is_smaller(callee, leavesize) || prop >= irg_inline_forced)) {
1594 if (!phiproj_computed) {
1595 phiproj_computed = 1;
1596 collect_phiprojs(current_ir_graph);
1598 did_inline = inline_method(call, callee);
1601 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1603 /* was inlined, must be recomputed */
1604 phiproj_computed = 0;
1606 /* Do some statistics */
1607 env->got_inline = 1;
1608 --env->n_call_nodes;
1609 env->n_nodes += callee_env->n_nodes;
1610 --callee_env->n_callers;
1612 /* remove this call from the list */
1614 tail->next = entry->next;
1616 env->call_head = entry->next;
1622 env->call_tail = tail;
1624 } while (did_inline);
1626 /* inline other small functions. */
1627 for (i = 0; i < n_irgs; ++i) {
1629 int phiproj_computed = 0;
1631 current_ir_graph = get_irp_irg(i);
1632 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1634 /* note that the list of possible calls is updated during the process */
1636 for (entry = env->call_head; entry != NULL; entry = entry->next) {
1637 irg_inline_property prop;
1642 callee = entry->callee;
1644 prop = get_irg_inline_property(callee);
1645 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1646 /* do not inline forbidden / weak graphs */
1650 e = pmap_find(copied_graphs, callee);
1653 * Remap callee if we have a copy.
1654 * FIXME: Should we do this only for recursive Calls ?
1659 if (prop >= irg_inline_forced ||
1660 (is_smaller(callee, size) && env->n_nodes < maxsize) /* small function */) {
1661 if (current_ir_graph == callee) {
1663 * Recursive call: we cannot directly inline because we cannot walk
1664 * the graph and change it. So we have to make a copy of the graph
1668 inline_irg_env *callee_env;
1672 * No copy yet, create one.
1673 * Note that recursive methods are never leaves, so it is sufficient
1674 * to test this condition here.
1676 copy = create_irg_copy(callee);
1678 /* create_irg_copy() destroys the Proj links, recompute them */
1679 phiproj_computed = 0;
1681 /* allocate new environment */
1682 callee_env = alloc_inline_irg_env();
1683 set_irg_link(copy, callee_env);
1685 assure_cf_loop(copy);
1686 wenv.x = callee_env;
1687 wenv.ignore_callers = 1;
1688 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1691 * Enter the entity of the original graph. This is needed
1692 * for inline_method(). However, note that ent->irg still points
1693 * to callee, NOT to copy.
1695 set_irg_entity(copy, get_irg_entity(callee));
1697 pmap_insert(copied_graphs, callee, copy);
1700 /* we have only one caller: the original graph */
1701 callee_env->n_callers = 1;
1702 callee_env->n_callers_orig = 1;
1704 if (! phiproj_computed) {
1705 phiproj_computed = 1;
1706 collect_phiprojs(current_ir_graph);
1708 did_inline = inline_method(call, callee);
1710 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1712 /* was inlined, must be recomputed */
1713 phiproj_computed = 0;
1715 /* callee was inline. Append it's call list. */
1716 env->got_inline = 1;
1717 --env->n_call_nodes;
1718 append_call_list(env, callee_env->call_head);
1719 env->n_call_nodes += callee_env->n_call_nodes;
1720 env->n_nodes += callee_env->n_nodes;
1721 --callee_env->n_callers;
1723 /* after we have inlined callee, all called methods inside callee
1724 are now called once more */
1725 for (centry = callee_env->call_head; centry != NULL; centry = centry->next) {
1726 inline_irg_env *penv = get_irg_link(centry->callee);
1730 /* remove this call from the list */
1732 tail->next = entry->next;
1734 env->call_head = entry->next;
1740 env->call_tail = tail;
1743 for (i = 0; i < n_irgs; ++i) {
1744 irg = get_irp_irg(i);
1745 env = (inline_irg_env *)get_irg_link(irg);
1747 if (env->got_inline) {
1748 optimize_graph_df(irg);
1751 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1752 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1753 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1754 env->n_callers_orig, env->n_callers,
1755 get_entity_name(get_irg_entity(irg))));
1759 /* kill the copied graphs: we don't need them anymore */
1760 foreach_pmap(copied_graphs, pm_entry) {
1761 ir_graph *copy = pm_entry->value;
1763 /* reset the entity, otherwise it will be deleted in the next step ... */
1764 set_irg_entity(copy, NULL);
1765 free_ir_graph(copy);
1767 pmap_destroy(copied_graphs);
1769 obstack_free(&temp_obst, NULL);
1770 current_ir_graph = rem;
1774 * Calculate the parameter weights for transmitting the address of a local variable.
1776 static unsigned calc_method_local_weight(ir_node *arg) {
1778 unsigned v, weight = 0;
1780 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1781 ir_node *succ = get_irn_out(arg, i);
1783 switch (get_irn_opcode(succ)) {
1786 /* Loads and Store can be removed */
1790 /* check if all args are constant */
1791 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1792 ir_node *idx = get_Sel_index(succ, j);
1793 if (! is_Const(idx))
1796 /* Check users on this Sel. Note: if a 0 is returned here, there was
1797 some unsupported node. */
1798 v = calc_method_local_weight(succ);
1801 /* we can kill one Sel with constant indexes, this is cheap */
1805 /* when looking backward we might find Id nodes */
1806 weight += calc_method_local_weight(succ);
1809 /* unoptimized tuple */
1810 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1811 ir_node *pred = get_Tuple_pred(succ, j);
1813 /* look for Proj(j) */
1814 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1815 ir_node *succ_succ = get_irn_out(succ, k);
1816 if (is_Proj(succ_succ)) {
1817 if (get_Proj_proj(succ_succ) == j) {
1819 weight += calc_method_local_weight(succ_succ);
1822 /* this should NOT happen */
1830 /* any other node: unsupported yet or bad. */
1838 * Calculate the parameter weights for transmitting the address of a local variable.
1840 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1841 ir_entity *ent = get_irg_entity(irg);
1843 int nparams, i, proj_nr;
1844 ir_node *irg_args, *arg;
1846 mtp = get_entity_type(ent);
1847 nparams = get_method_n_params(mtp);
1849 /* allocate a new array. currently used as 'analysed' flag */
1850 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1852 /* If the method haven't parameters we have nothing to do. */
1856 assure_irg_outs(irg);
1857 irg_args = get_irg_args(irg);
1858 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1859 arg = get_irn_out(irg_args, i);
1860 proj_nr = get_Proj_proj(arg);
1861 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1866 * Calculate the benefice for transmitting an local variable address.
1867 * After inlining, the local variable might be transformed into a
1868 * SSA variable by scalar_replacement().
1870 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1871 inline_irg_env *env = get_irg_link(callee);
1873 if (env->local_weights != NULL) {
1874 if (pos < ARR_LEN(env->local_weights))
1875 return env->local_weights[pos];
1879 analyze_irg_local_weights(env, callee);
1881 if (pos < ARR_LEN(env->local_weights))
1882 return env->local_weights[pos];
1887 * Calculate a benefice value for inlining the given call.
1889 * @param call the call node we have to inspect
1890 * @param callee the called graph
1891 * @param local_adr set after return if an address of a local variable is
1892 * transmitted as a parameter
1894 static int calc_inline_benefice(const call_entry *entry, ir_graph *callee,
1895 unsigned *local_adr)
1897 ir_node *call = entry->call;
1898 ir_entity *ent = get_irg_entity(callee);
1902 int i, n_params, all_const;
1904 irg_inline_property prop;
1906 inline_irg_env *callee_env;
1908 prop = get_irg_inline_property(callee);
1909 if (prop == irg_inline_forbidden) {
1910 DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden\n",
1915 if ( (get_irg_additional_properties(callee)
1916 | get_entity_additional_properties(ent))
1917 & (mtp_property_noreturn | mtp_property_weak)) {
1918 DB((dbg, LEVEL_2, "In %+F Call to %+F: not inlining noreturn or weak\n",
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
1955 * inlining, scalar_replacement might be able to remove the
1956 * local variable, so honor this.
1958 v = get_method_local_adress_weight(callee, i);
1966 callee_env = get_irg_link(callee);
1967 if (callee_env->n_callers == 1
1968 && callee != current_ir_graph
1969 && get_entity_visibility(ent) == visibility_local) {
1973 /* give a bonus for functions with one block */
1974 if (callee_env->n_blocks == 1)
1975 weight = weight * 3 / 2;
1977 /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1978 if (callee_env->n_nodes < 30 && !callee_env->recursive)
1981 /* and finally for leaves: they do not increase the register pressure
1982 because of callee safe registers */
1983 if (callee_env->n_call_nodes == 0)
1987 * Reduce the weight for recursive function IFF not all arguments are const.
1988 * inlining recursive functions is rarely good.
1990 if (callee_env->recursive && !all_const)
1993 /** it's important to inline inner loops first */
1994 if (entry->loop_depth > 30)
1995 weight += 30 * 1024;
1997 weight += entry->loop_depth * 1024;
2000 * All arguments constant is probably a good sign, give an extra bonus
2008 static ir_graph **irgs;
2009 static int last_irg;
2011 static void callgraph_walker(ir_graph *irg, void *data)
2014 irgs[last_irg++] = irg;
2017 static ir_graph **create_irg_list(void)
2019 ir_entity **free_methods;
2021 int n_irgs = get_irp_n_irgs();
2023 cgana(&arr_len, &free_methods);
2024 xfree(free_methods);
2026 compute_callgraph();
2029 irgs = xmalloc(n_irgs * sizeof(*irgs));
2030 memset(irgs, 0, sizeof(n_irgs * sizeof(*irgs)));
2032 callgraph_walk(NULL, callgraph_walker, NULL);
2033 assert(n_irgs == last_irg);
2038 static void maybe_push_call(pqueue_t *pqueue, call_entry *call,
2039 int inline_threshold)
2041 ir_graph *callee = call->callee;
2042 irg_inline_property prop = get_irg_inline_property(callee);
2046 benefice = calc_inline_benefice(call, callee, &local_adr);
2047 call->local_adr = local_adr;
2049 DB((dbg, LEVEL_2, "In %+F Call %+F to %+F has benefice %d\n",
2050 get_irn_irg(call->call), call->call, callee, benefice));
2052 if (benefice < inline_threshold && !(prop & irg_inline_forced))
2055 pqueue_put(pqueue, call, benefice);
2058 static void inline_into(ir_graph *irg, unsigned maxsize,
2059 int inline_threshold, pmap *copied_graphs)
2061 int phiproj_computed = 0;
2062 inline_irg_env *env = get_irg_link(irg);
2063 call_entry *curr_call;
2066 if (env->n_nodes > maxsize) {
2067 DB((dbg, LEVEL_2, "%+F: too big (%d)\n", irg, env->n_nodes));
2071 current_ir_graph = irg;
2073 /* put irgs into the pqueue */
2074 pqueue_t *pqueue = new_pqueue();
2076 for (curr_call = env->call_head; curr_call != NULL;
2077 curr_call = curr_call->next) {
2079 if (is_Tuple(curr_call->call))
2081 assert(is_Call(curr_call->call));
2083 maybe_push_call(pqueue, curr_call, inline_threshold);
2086 /* note that the list of possible calls is updated during the process */
2087 while (!pqueue_empty(pqueue)) {
2090 call_entry *curr_call = pqueue_pop_front(pqueue);
2091 ir_graph *callee = curr_call->callee;
2092 ir_node *call_node = curr_call->call;
2093 irg_inline_property prop = get_irg_inline_property(callee);
2094 const call_entry *centry;
2095 inline_irg_env *callee_env = get_irg_link(callee);
2096 int depth = curr_call->loop_depth;
2098 if (! (prop & irg_inline_forced)
2099 && env->n_nodes + callee_env->n_nodes > maxsize) {
2100 DB((dbg, LEVEL_2, "%+F: too big (%d) + %+F (%d)\n", irg,
2101 env->n_nodes, callee, callee_env->n_nodes));
2105 e = pmap_find(copied_graphs, callee);
2108 * Remap callee if we have a copy.
2109 * FIXME: Should we do this only for recursive Calls ?
2114 if (current_ir_graph == callee) {
2116 * Recursive call: we cannot directly inline because we cannot
2117 * walk the graph and change it. So we have to make a copy of
2120 inline_irg_env *callee_env;
2124 * No copy yet, create one.
2125 * Note that recursive methods are never leaves, so it is
2126 * sufficient to test this condition here.
2128 copy = create_irg_copy(callee);
2130 /* create_irg_copy() destroys the Proj links, recompute them */
2131 phiproj_computed = 0;
2133 /* allocate new environment */
2134 callee_env = alloc_inline_irg_env();
2135 set_irg_link(copy, callee_env);
2137 assure_cf_loop(copy);
2138 wenv.x = callee_env;
2139 wenv.ignore_callers = 1;
2140 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2143 * Enter the entity of the original graph. This is needed
2144 * for inline_method(). However, note that ent->irg still points
2145 * to callee, NOT to copy.
2147 set_irg_entity(copy, get_irg_entity(callee));
2149 pmap_insert(copied_graphs, callee, copy);
2152 /* we have only one caller: the original graph */
2153 callee_env->n_callers = 1;
2154 callee_env->n_callers_orig = 1;
2156 if (! phiproj_computed) {
2157 phiproj_computed = 1;
2158 collect_phiprojs(current_ir_graph);
2160 did_inline = inline_method(call_node, callee);
2164 /* was inlined, must be recomputed */
2165 phiproj_computed = 0;
2167 /* after we have inlined callee, all called methods inside
2168 * callee are now called once more */
2169 for (centry = callee_env->call_head; centry != NULL;
2170 centry = centry->next) {
2171 inline_irg_env *penv = get_irg_link(centry->callee);
2175 /* callee was inline. Append it's call list. */
2176 env->got_inline = 1;
2177 if (curr_call->local_adr)
2178 env->local_vars = 1;
2179 --env->n_call_nodes;
2181 /* we just generate a bunch of new calls */
2182 for (centry = callee_env->call_head; centry != NULL;
2183 centry = centry->next) {
2184 /* Note that the src list points to Call nodes in the inlined graph,
2185 * but we need Call nodes in our graph. Luckily the inliner leaves
2186 * this information in the link field. */
2188 ir_node *new_call = get_irn_link(centry->call);
2189 call_entry *new_entry;
2191 if (!is_Call(new_call))
2194 new_entry = duplicate_call_entry(centry, depth, new_call);
2195 maybe_push_call(pqueue, new_entry, inline_threshold);
2198 env->n_call_nodes += callee_env->n_call_nodes;
2199 env->n_nodes += callee_env->n_nodes;
2200 --callee_env->n_callers;
2207 * Heuristic inliner. Calculates a benefice value for every call and inlines
2208 * those calls with a value higher than the threshold.
2210 void inline_functions(unsigned maxsize, int inline_threshold) {
2211 inline_irg_env *env;
2215 pmap *copied_graphs;
2216 pmap_entry *pm_entry;
2219 rem = current_ir_graph;
2220 obstack_init(&temp_obst);
2222 irgs = create_irg_list();
2224 /* a map for the copied graphs, used to inline recursive calls */
2225 copied_graphs = pmap_create();
2227 /* extend all irgs by a temporary data structure for inlining. */
2228 n_irgs = get_irp_n_irgs();
2229 for (i = 0; i < n_irgs; ++i)
2230 set_irg_link(irgs[i], alloc_inline_irg_env());
2232 /* Precompute information in temporary data structure. */
2233 wenv.ignore_runtime = 0;
2234 wenv.ignore_callers = 0;
2235 for (i = 0; i < n_irgs; ++i) {
2236 ir_graph *irg = irgs[i];
2238 free_callee_info(irg);
2240 wenv.x = get_irg_link(irg);
2241 assure_cf_loop(irg);
2242 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
2245 /* -- and now inline. -- */
2246 for (i = 0; i < n_irgs; ++i) {
2247 ir_graph *irg = irgs[i];
2249 inline_into(irg, maxsize, inline_threshold, copied_graphs);
2252 for (i = 0; i < n_irgs; ++i) {
2253 ir_graph *irg = irgs[i];
2255 env = get_irg_link(irg);
2256 if (env->got_inline) {
2257 /* this irg got calls inlined: optimize it */
2260 /* scalar replacement does not work well with Tuple nodes, so optimize them away */
2261 optimize_graph_df(irg);
2263 if (env->local_vars) {
2264 if (scalar_replacement_opt(irg)) {
2265 optimize_graph_df(irg);
2270 if (env->local_vars) {
2271 scalar_replacement_opt(irg);
2276 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2277 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2278 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2279 env->n_callers_orig, env->n_callers,
2280 get_entity_name(get_irg_entity(irg))));
2284 /* kill the copied graphs: we don't need them anymore */
2285 foreach_pmap(copied_graphs, pm_entry) {
2286 ir_graph *copy = pm_entry->value;
2288 /* reset the entity, otherwise it will be deleted in the next step ... */
2289 set_irg_entity(copy, NULL);
2290 free_ir_graph(copy);
2292 pmap_destroy(copied_graphs);
2296 obstack_free(&temp_obst, NULL);
2297 current_ir_graph = rem;
2300 void firm_init_inline(void) {
2301 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");