2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Dead node elimination and Procedure Inlining.
23 * @author Michael Beck, Goetz Lindenmaier
32 #include "irgraph_t.h"
35 #include "iroptimize.h"
52 #include "irbackedge_t.h"
53 #include "opt_inline_t.h"
58 #include "analyze_irg_args.h"
59 #include "iredges_t.h"
63 #include "iropt_dbg.h"
65 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
67 /*------------------------------------------------------------------*/
68 /* Routines for dead node elimination / copying garbage collection */
70 /*------------------------------------------------------------------*/
73 * Remember the new node in the old node by using a field all nodes have.
75 #define set_new_node(oldn, newn) set_irn_link(oldn, newn)
78 * Get this new node, before the old node is forgotten.
80 #define get_new_node(oldn) get_irn_link(oldn)
83 * Check if a new node was set.
85 #define has_new_node(n) (get_new_node(n) != NULL)
88 * We use the block_visited flag to mark that we have computed the
89 * number of useful predecessors for this block.
90 * Further we encode the new arity in this flag in the old blocks.
91 * Remembering the arity is useful, as it saves a lot of pointer
92 * accesses. This function is called for all Phi and Block nodes
96 compute_new_arity(ir_node *b) {
97 int i, res, irn_arity;
100 irg_v = get_irg_block_visited(current_ir_graph);
101 block_v = get_Block_block_visited(b);
102 if (block_v >= irg_v) {
103 /* we computed the number of preds for this block and saved it in the
105 return block_v - irg_v;
107 /* compute the number of good predecessors */
108 res = irn_arity = get_irn_arity(b);
109 for (i = 0; i < irn_arity; i++)
110 if (is_Bad(get_irn_n(b, i))) res--;
111 /* save it in the flag. */
112 set_Block_block_visited(b, irg_v + res);
118 * Copies the node to the new obstack. The Ins of the new node point to
119 * the predecessors on the old obstack. For block/phi nodes not all
120 * predecessors might be copied. n->link points to the new node.
121 * For Phi and Block nodes the function allocates in-arrays with an arity
122 * only for useful predecessors. The arity is determined by counting
123 * the non-bad predecessors of the block.
125 * @param n The node to be copied
126 * @param env if non-NULL, the node number attribute will be copied to the new node
128 * Note: Also used for loop unrolling.
130 static void copy_node(ir_node *n, void *env) {
133 ir_op *op = get_irn_op(n);
137 /* node copied already */
139 } else if (op == op_Block) {
141 new_arity = compute_new_arity(n);
142 n->attr.block.graph_arr = NULL;
144 block = get_nodes_block(n);
146 new_arity = compute_new_arity(block);
148 new_arity = get_irn_arity(n);
151 nn = new_ir_node(get_irn_dbg_info(n),
158 /* Copy the attributes. These might point to additional data. If this
159 was allocated on the old obstack the pointers now are dangling. This
160 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
161 if (op == op_Block) {
162 /* we cannot allow blocks WITHOUT macroblock input */
163 set_Block_MacroBlock(nn, get_Block_MacroBlock(n));
165 copy_node_attr(n, nn);
169 int copy_node_nr = env != NULL;
171 /* for easier debugging, we want to copy the node numbers too */
172 nn->node_nr = n->node_nr;
178 hook_dead_node_elim_subst(current_ir_graph, n, nn);
182 * Copies new predecessors of old node to new node remembered in link.
183 * Spare the Bad predecessors of Phi and Block nodes.
185 static void copy_preds(ir_node *n, void *env) {
190 nn = get_new_node(n);
193 /* copy the macro block header */
194 ir_node *mbh = get_Block_MacroBlock(n);
197 /* this block is a macroblock header */
198 set_Block_MacroBlock(nn, nn);
200 /* get the macro block header */
201 ir_node *nmbh = get_new_node(mbh);
202 assert(nmbh != NULL);
203 set_Block_MacroBlock(nn, nmbh);
206 /* Don't copy Bad nodes. */
208 irn_arity = get_irn_arity(n);
209 for (i = 0; i < irn_arity; i++) {
210 if (! is_Bad(get_irn_n(n, i))) {
211 ir_node *pred = get_irn_n(n, i);
212 set_irn_n(nn, j, get_new_node(pred));
216 /* repair the block visited flag from above misuse. Repair it in both
217 graphs so that the old one can still be used. */
218 set_Block_block_visited(nn, 0);
219 set_Block_block_visited(n, 0);
220 /* Local optimization could not merge two subsequent blocks if
221 in array contained Bads. Now it's possible.
222 We don't call optimize_in_place as it requires
223 that the fields in ir_graph are set properly. */
224 if ((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(struct obstack);
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(survive_dce_t);
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 register_hook(hook_dead_node_elim, &res->dead_node_elim);
678 register_hook(hook_dead_node_elim_subst, &res->dead_node_elim_subst);
683 * Free a Survive DCE environment.
685 void free_survive_dce(survive_dce_t *sd) {
686 obstack_free(&sd->obst, NULL);
687 pmap_destroy(sd->places);
688 unregister_hook(hook_dead_node_elim, &sd->dead_node_elim);
689 unregister_hook(hook_dead_node_elim_subst, &sd->dead_node_elim_subst);
694 * Register a node pointer to be patched upon DCE.
695 * When DCE occurs, the node pointer specified by @p place will be
696 * patched to the new address of the node it is pointing to.
698 * @param sd The Survive DCE environment.
699 * @param place The address of the node pointer.
701 void survive_dce_register_irn(survive_dce_t *sd, ir_node **place) {
702 if (*place != NULL) {
703 ir_node *irn = *place;
704 survive_dce_list_t *curr = pmap_get(sd->places, irn);
705 survive_dce_list_t *nw = obstack_alloc(&sd->obst, sizeof(nw[0]));
710 pmap_insert(sd->places, irn, nw);
714 /*--------------------------------------------------------------------*/
715 /* Functionality for inlining */
716 /*--------------------------------------------------------------------*/
719 * Copy node for inlineing. Updates attributes that change when
720 * inlineing but not for dead node elimination.
722 * Copies the node by calling copy_node() and then updates the entity if
723 * it's a local one. env must be a pointer of the frame type of the
724 * inlined procedure. The new entities must be in the link field of
727 static void copy_node_inline(ir_node *n, void *env) {
729 ir_type *frame_tp = (ir_type *)env;
733 nn = get_new_node (n);
735 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
736 set_Sel_entity(nn, get_entity_link(get_Sel_entity(n)));
738 } else if (is_Block(n)) {
739 nn = get_new_node (n);
740 nn->attr.block.irg = current_ir_graph;
745 * Copies new predecessors of old node and move constants to
748 static void copy_preds_inline(ir_node *n, void *env) {
752 nn = skip_Id(get_new_node(n));
753 if (is_irn_constlike(nn)) {
754 /* move Constants into the start block */
755 set_nodes_block(nn, get_irg_start_block(current_ir_graph));
757 n = identify_remember(current_ir_graph->value_table, nn);
766 * Walker: checks if P_value_arg_base is used.
768 static void find_addr(ir_node *node, void *env) {
769 int *allow_inline = env;
771 is_Start(get_Proj_pred(node)) &&
772 get_Proj_proj(node) == pn_Start_P_value_arg_base) {
774 } else if (is_Alloc(node) && get_Alloc_where(node) == stack_alloc) {
776 * Refuse to inline alloca call unless user explicitly forced so as this
777 * may change program's memory overhead drastically when the function
778 * using alloca is called in loop. In GCC present in SPEC2000 inlining
779 * into schedule_block cause it to require 2GB of ram instead of 256MB.
781 * Sorrily this is true with our implementation also.
782 * Moreover, we cannot differentiate between alloca() and VLA yet, so this
783 * disables inlining of functions using VLA (with are completely save).
786 * - add a flag to the Alloc node for "real" alloca() calls
787 * - add a new Stack-Restore node at the end of a function using alloca()
794 * Check if we can inline a given call.
795 * Currently, we cannot inline two cases:
796 * - call with compound arguments
797 * - graphs that take the address of a parameter
799 * check these conditions here
801 static int can_inline(ir_node *call, ir_graph *called_graph) {
802 ir_type *call_type = get_Call_type(call);
803 int params, ress, i, res;
804 assert(is_Method_type(call_type));
806 params = get_method_n_params(call_type);
807 ress = get_method_n_ress(call_type);
809 /* check parameters for compound arguments */
810 for (i = 0; i < params; ++i) {
811 ir_type *p_type = get_method_param_type(call_type, i);
813 if (is_compound_type(p_type))
817 /* check results for compound arguments */
818 for (i = 0; i < ress; ++i) {
819 ir_type *r_type = get_method_res_type(call_type, i);
821 if (is_compound_type(r_type))
826 irg_walk_graph(called_graph, find_addr, NULL, &res);
832 exc_handler = 0, /**< There is a handler. */
833 exc_to_end = 1, /**< Branches to End. */
834 exc_no_handler = 2 /**< Exception handling not represented. */
837 /* Inlines a method at the given call site. */
838 int inline_method(ir_node *call, ir_graph *called_graph) {
840 ir_node *post_call, *post_bl;
841 ir_node *in[pn_Start_max];
842 ir_node *end, *end_bl, *block;
847 int arity, n_ret, n_exc, n_res, i, n, j, rem_opt, irn_arity, n_params;
848 enum exc_mode exc_handling;
849 ir_type *called_frame, *curr_frame, *mtp, *ctp;
852 irg_inline_property prop = get_irg_inline_property(called_graph);
853 unsigned long visited;
855 if (prop == irg_inline_forbidden)
858 ent = get_irg_entity(called_graph);
860 mtp = get_entity_type(ent);
861 ctp = get_Call_type(call);
862 if (get_method_n_params(mtp) > get_method_n_params(ctp)) {
863 /* this is a bad feature of C: without a prototype, we can can call a function with less
864 parameters than needed. Currently we don't support this, although it would be
865 to use Unknown than. */
869 /* Argh, compiling C has some bad consequences:
870 the call type AND the method type might be different.
871 It is implementation defendant what happens in that case.
872 We support inlining, if the bitsize of the types matches AND
873 the same arithmetic is used. */
874 n_params = get_method_n_params(mtp);
875 for (i = n_params - 1; i >= 0; --i) {
876 ir_type *param_tp = get_method_param_type(mtp, i);
877 ir_type *arg_tp = get_method_param_type(ctp, i);
879 if (param_tp != arg_tp) {
880 ir_mode *pmode = get_type_mode(param_tp);
881 ir_mode *amode = get_type_mode(arg_tp);
883 if (pmode == NULL || amode == NULL)
885 if (get_mode_size_bits(pmode) != get_mode_size_bits(amode))
887 if (get_mode_arithmetic(pmode) != get_mode_arithmetic(amode))
889 /* otherwise we can simply "reinterpret" the bits */
893 irg = get_irn_irg(call);
896 * We cannot inline a recursive call. The graph must be copied before
897 * the call the inline_method() using create_irg_copy().
899 if (called_graph == irg)
903 * currently, we cannot inline two cases:
904 * - call with compound arguments
905 * - graphs that take the address of a parameter
907 if (! can_inline(call, called_graph))
910 rem = current_ir_graph;
911 current_ir_graph = irg;
913 DB((dbg, LEVEL_1, "Inlining %+F(%+F) into %+F\n", call, called_graph, irg));
915 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
916 rem_opt = get_opt_optimize();
919 /* Handle graph state */
920 assert(get_irg_phase_state(irg) != phase_building);
921 assert(get_irg_pinned(irg) == op_pin_state_pinned);
922 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
923 set_irg_outs_inconsistent(irg);
924 set_irg_extblk_inconsistent(irg);
925 set_irg_doms_inconsistent(irg);
926 set_irg_loopinfo_inconsistent(irg);
927 set_irg_callee_info_state(irg, irg_callee_info_inconsistent);
929 /* -- Check preconditions -- */
930 assert(is_Call(call));
932 /* here we know we WILL inline, so inform the statistics */
933 hook_inline(call, called_graph);
935 /* -- Decide how to handle exception control flow: Is there a handler
936 for the Call node, or do we branch directly to End on an exception?
938 0 There is a handler.
940 2 Exception handling not represented in Firm. -- */
942 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
943 for (proj = get_irn_link(call); proj; proj = get_irn_link(proj)) {
944 long proj_nr = get_Proj_proj(proj);
945 if (proj_nr == pn_Call_X_except) Xproj = proj;
946 if (proj_nr == pn_Call_M_except) Mproj = proj;
948 if (Mproj) { assert(Xproj); exc_handling = exc_handler; } /* Mproj */
949 else if (Xproj) { exc_handling = exc_to_end; } /* !Mproj && Xproj */
950 else { exc_handling = exc_no_handler; } /* !Mproj && !Xproj */
953 /* create the argument tuple */
954 NEW_ARR_A(ir_type *, args_in, n_params);
956 block = get_nodes_block(call);
957 for (i = n_params - 1; i >= 0; --i) {
958 ir_node *arg = get_Call_param(call, i);
959 ir_type *param_tp = get_method_param_type(mtp, i);
960 ir_mode *mode = get_type_mode(param_tp);
962 if (mode != get_irn_mode(arg)) {
963 arg = new_r_Conv(irg, block, arg, mode);
969 the procedure and later replaces the Start node of the called graph.
970 Post_call is the old Call node and collects the results of the called
971 graph. Both will end up being a tuple. -- */
972 post_bl = get_nodes_block(call);
973 set_irg_current_block(irg, post_bl);
974 /* XxMxPxPxPxT of Start + parameter of Call */
975 in[pn_Start_X_initial_exec] = new_Jmp();
976 in[pn_Start_M] = get_Call_mem(call);
977 in[pn_Start_P_frame_base] = get_irg_frame(irg);
978 in[pn_Start_P_tls] = get_irg_tls(irg);
979 in[pn_Start_T_args] = new_Tuple(n_params, args_in);
980 /* in[pn_Start_P_value_arg_base] = ??? */
981 assert(pn_Start_P_value_arg_base == pn_Start_max - 1 && "pn_Start_P_value_arg_base not supported, fix");
982 pre_call = new_Tuple(pn_Start_max - 1, in);
986 The new block gets the ins of the old block, pre_call and all its
987 predecessors and all Phi nodes. -- */
988 part_block(pre_call);
990 /* -- Prepare state for dead node elimination -- */
991 /* Visited flags in calling irg must be >= flag in called irg.
992 Else walker and arity computation will not work. */
993 if (get_irg_visited(irg) <= get_irg_visited(called_graph))
994 set_irg_visited(irg, get_irg_visited(called_graph) + 1);
995 if (get_irg_block_visited(irg) < get_irg_block_visited(called_graph))
996 set_irg_block_visited(irg, get_irg_block_visited(called_graph));
997 visited = get_irg_visited(irg);
999 /* Set pre_call as new Start node in link field of the start node of
1000 calling graph and pre_calls block as new block for the start block
1002 Further mark these nodes so that they are not visited by the
1004 set_irn_link(get_irg_start(called_graph), pre_call);
1005 set_irn_visited(get_irg_start(called_graph), visited);
1006 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
1007 set_irn_visited(get_irg_start_block(called_graph), visited);
1009 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
1010 set_irn_visited(get_irg_bad(called_graph), visited);
1012 set_irn_link(get_irg_no_mem(called_graph), get_irg_no_mem(current_ir_graph));
1013 set_irn_visited(get_irg_no_mem(called_graph), visited);
1015 /* Initialize for compaction of in arrays */
1016 inc_irg_block_visited(irg);
1018 /* -- Replicate local entities of the called_graph -- */
1019 /* copy the entities. */
1020 irp_reserve_resources(irp, IR_RESOURCE_ENTITY_LINK);
1021 called_frame = get_irg_frame_type(called_graph);
1022 curr_frame = get_irg_frame_type(irg);
1023 for (i = 0, n = get_class_n_members(called_frame); i < n; ++i) {
1024 ir_entity *new_ent, *old_ent;
1025 old_ent = get_class_member(called_frame, i);
1026 new_ent = copy_entity_own(old_ent, curr_frame);
1027 set_entity_link(old_ent, new_ent);
1030 /* visited is > than that of called graph. With this trick visited will
1031 remain unchanged so that an outer walker, e.g., searching the call nodes
1032 to inline, calling this inline will not visit the inlined nodes. */
1033 set_irg_visited(irg, get_irg_visited(irg)-1);
1035 /* -- Performing dead node elimination inlines the graph -- */
1036 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
1038 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds_inline,
1039 get_irg_frame_type(called_graph));
1041 irp_free_resources(irp, IR_RESOURCE_ENTITY_LINK);
1043 /* Repair called_graph */
1044 set_irg_visited(called_graph, get_irg_visited(irg));
1045 set_irg_block_visited(called_graph, get_irg_block_visited(irg));
1046 set_Block_block_visited(get_irg_start_block(called_graph), 0);
1048 /* -- Merge the end of the inlined procedure with the call site -- */
1049 /* We will turn the old Call node into a Tuple with the following
1052 0: Phi of all Memories of Return statements.
1053 1: Jmp from new Block that merges the control flow from all exception
1054 predecessors of the old end block.
1055 2: Tuple of all arguments.
1056 3: Phi of Exception memories.
1057 In case the old Call directly branches to End on an exception we don't
1058 need the block merging all exceptions nor the Phi of the exception
1062 /* -- Precompute some values -- */
1063 end_bl = get_new_node(get_irg_end_block(called_graph));
1064 end = get_new_node(get_irg_end(called_graph));
1065 arity = get_Block_n_cfgpreds(end_bl); /* arity = n_exc + n_ret */
1066 n_res = get_method_n_ress(get_Call_type(call));
1068 res_pred = XMALLOCN(ir_node*, n_res);
1069 cf_pred = XMALLOCN(ir_node*, arity);
1071 set_irg_current_block(irg, post_bl); /* just to make sure */
1073 /* -- archive keepalives -- */
1074 irn_arity = get_irn_arity(end);
1075 for (i = 0; i < irn_arity; i++) {
1076 ir_node *ka = get_End_keepalive(end, i);
1078 add_End_keepalive(get_irg_end(irg), ka);
1081 /* The new end node will die. We need not free as the in array is on the obstack:
1082 copy_node() only generated 'D' arrays. */
1084 /* -- Replace Return nodes by Jump nodes. -- */
1086 for (i = 0; i < arity; i++) {
1088 ret = get_Block_cfgpred(end_bl, i);
1089 if (is_Return(ret)) {
1090 cf_pred[n_ret] = new_r_Jmp(irg, get_nodes_block(ret));
1094 set_irn_in(post_bl, n_ret, cf_pred);
1096 /* -- Build a Tuple for all results of the method.
1097 Add Phi node if there was more than one Return. -- */
1098 turn_into_tuple(post_call, pn_Call_max);
1099 /* First the Memory-Phi */
1101 for (i = 0; i < arity; i++) {
1102 ret = get_Block_cfgpred(end_bl, i);
1103 if (is_Return(ret)) {
1104 cf_pred[n_ret] = get_Return_mem(ret);
1108 phi = new_Phi(n_ret, cf_pred, mode_M);
1109 set_Tuple_pred(call, pn_Call_M_regular, phi);
1110 /* Conserve Phi-list for further inlinings -- but might be optimized */
1111 if (get_nodes_block(phi) == post_bl) {
1112 set_irn_link(phi, get_irn_link(post_bl));
1113 set_irn_link(post_bl, phi);
1115 /* Now the real results */
1117 for (j = 0; j < n_res; j++) {
1119 for (i = 0; i < arity; i++) {
1120 ret = get_Block_cfgpred(end_bl, i);
1121 if (is_Return(ret)) {
1122 cf_pred[n_ret] = get_Return_res(ret, j);
1127 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
1131 /* Conserve Phi-list for further inlinings -- but might be optimized */
1132 if (get_nodes_block(phi) == post_bl) {
1133 set_Phi_next(phi, get_Block_phis(post_bl));
1134 set_Block_phis(post_bl, phi);
1137 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
1139 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
1141 /* handle the regular call */
1142 set_Tuple_pred(call, pn_Call_X_regular, new_Jmp());
1144 /* For now, we cannot inline calls with value_base */
1145 set_Tuple_pred(call, pn_Call_P_value_res_base, new_Bad());
1147 /* Finally the exception control flow.
1148 We have two (three) possible situations:
1149 First if the Call branches to an exception handler: We need to add a Phi node to
1150 collect the memory containing the exception objects. Further we need
1151 to add another block to get a correct representation of this Phi. To
1152 this block we add a Jmp that resolves into the X output of the Call
1153 when the Call is turned into a tuple.
1154 Second the Call branches to End, the exception is not handled. Just
1155 add all inlined exception branches to the End node.
1156 Third: there is no Exception edge at all. Handle as case two. */
1157 if (exc_handling == exc_handler) {
1159 for (i = 0; i < arity; i++) {
1161 ret = get_Block_cfgpred(end_bl, i);
1162 irn = skip_Proj(ret);
1163 if (is_fragile_op(irn) || is_Raise(irn)) {
1164 cf_pred[n_exc] = ret;
1169 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1170 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1171 /* The Phi for the memories with the exception objects */
1173 for (i = 0; i < arity; i++) {
1175 ret = skip_Proj(get_Block_cfgpred(end_bl, i));
1177 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 3);
1179 } else if (is_fragile_op(ret)) {
1180 /* We rely that all cfops have the memory output at the same position. */
1181 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 0);
1183 } else if (is_Raise(ret)) {
1184 cf_pred[n_exc] = new_r_Proj(irg, get_nodes_block(ret), ret, mode_M, 1);
1188 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1190 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1191 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1194 ir_node *main_end_bl;
1195 int main_end_bl_arity;
1196 ir_node **end_preds;
1198 /* assert(exc_handling == 1 || no exceptions. ) */
1200 for (i = 0; i < arity; i++) {
1201 ir_node *ret = get_Block_cfgpred(end_bl, i);
1202 ir_node *irn = skip_Proj(ret);
1204 if (is_fragile_op(irn) || is_Raise(irn)) {
1205 cf_pred[n_exc] = ret;
1209 main_end_bl = get_irg_end_block(irg);
1210 main_end_bl_arity = get_irn_arity(main_end_bl);
1211 end_preds = XMALLOCN(ir_node*, n_exc + main_end_bl_arity);
1213 for (i = 0; i < main_end_bl_arity; ++i)
1214 end_preds[i] = get_irn_n(main_end_bl, i);
1215 for (i = 0; i < n_exc; ++i)
1216 end_preds[main_end_bl_arity + i] = cf_pred[i];
1217 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1218 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1219 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1225 /* -- Turn CSE back on. -- */
1226 set_optimize(rem_opt);
1227 current_ir_graph = rem;
1232 /********************************************************************/
1233 /* Apply inlining to small methods. */
1234 /********************************************************************/
1236 static struct obstack temp_obst;
1238 /** Represents a possible inlinable call in a graph. */
1239 typedef struct _call_entry {
1240 ir_node *call; /**< The Call node. */
1241 ir_graph *callee; /**< The callee IR-graph. */
1242 list_head list; /**< List head for linking the next one. */
1243 int loop_depth; /**< The loop depth of this call. */
1244 int benefice; /**< The calculated benefice of this call. */
1245 unsigned local_adr:1; /**< Set if this call gets an address of a local variable. */
1246 unsigned all_const:1; /**< Set if this call has only constant parameters. */
1250 * environment for inlining small irgs
1252 typedef struct _inline_env_t {
1253 struct obstack obst; /**< An obstack where call_entries are allocated on. */
1254 list_head calls; /**< 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) {
1280 if (is_Call(call)) {
1281 ir_graph *called_irg = get_call_called_irg(call);
1283 if (called_irg != NULL) {
1284 /* The Call node calls a locally defined method. Remember to inline. */
1285 inline_env_t *ienv = env;
1286 call_entry *entry = obstack_alloc(&ienv->obst, sizeof(*entry));
1288 entry->callee = called_irg;
1289 entry->loop_depth = 0;
1290 entry->benefice = 0;
1291 entry->local_adr = 0;
1292 entry->all_const = 0;
1294 list_add_tail(&entry->list, &ienv->calls);
1300 * Inlines all small methods at call sites where the called address comes
1301 * from a Const node that references the entity representing the called
1303 * The size argument is a rough measure for the code size of the method:
1304 * Methods where the obstack containing the firm graph is smaller than
1307 void inline_small_irgs(ir_graph *irg, int size) {
1308 ir_graph *rem = current_ir_graph;
1312 current_ir_graph = irg;
1313 /* Handle graph state */
1314 assert(get_irg_phase_state(irg) != phase_building);
1315 free_callee_info(irg);
1317 /* Find Call nodes to inline.
1318 (We can not inline during a walk of the graph, as inlining the same
1319 method several times changes the visited flag of the walked graph:
1320 after the first inlining visited of the callee equals visited of
1321 the caller. With the next inlining both are increased.) */
1322 obstack_init(&env.obst);
1323 INIT_LIST_HEAD(&env.calls);
1324 irg_walk_graph(irg, NULL, collect_calls, &env);
1326 if (! list_empty(&env.calls)) {
1327 /* There are calls to inline */
1328 collect_phiprojs(irg);
1330 list_for_each_entry(call_entry, entry, &env.calls, list) {
1331 ir_graph *callee = entry->callee;
1332 irg_inline_property prop = get_irg_inline_property(callee);
1334 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1335 /* do not inline forbidden / weak graphs */
1339 if (prop >= irg_inline_forced ||
1340 _obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst) < size) {
1341 inline_method(entry->call, callee);
1345 obstack_free(&env.obst, NULL);
1346 current_ir_graph = rem;
1350 * Environment for inlining irgs.
1353 list_head calls; /**< List of of all call nodes in this graph. */
1354 unsigned *local_weights; /**< Once allocated, the beneficial weight for transmitting local addresses. */
1355 unsigned n_nodes; /**< Number of nodes in graph except Id, Tuple, Proj, Start, End. */
1356 unsigned n_blocks; /**< Number of Blocks in graph without Start and End block. */
1357 unsigned n_nodes_orig; /**< for statistics */
1358 unsigned n_call_nodes; /**< Number of Call nodes in the graph. */
1359 unsigned n_call_nodes_orig; /**< for statistics */
1360 unsigned n_callers; /**< Number of known graphs that call this graphs. */
1361 unsigned n_callers_orig; /**< for statistics */
1362 unsigned got_inline:1; /**< Set, if at least one call inside this graph was inlined. */
1363 unsigned local_vars:1; /**< Set, if an inlined function got the address of a local variable. */
1364 unsigned recursive:1; /**< Set, if this function is self recursive. */
1368 * Allocate a new environment for inlining.
1370 static inline_irg_env *alloc_inline_irg_env(void) {
1371 inline_irg_env *env = obstack_alloc(&temp_obst, sizeof(*env));
1372 INIT_LIST_HEAD(&env->calls);
1373 env->local_weights = NULL;
1374 env->n_nodes = -2; /* do not count count Start, End */
1375 env->n_blocks = -2; /* do not count count Start, End Block */
1376 env->n_nodes_orig = -2; /* do not count Start, End */
1377 env->n_call_nodes = 0;
1378 env->n_call_nodes_orig = 0;
1380 env->n_callers_orig = 0;
1381 env->got_inline = 0;
1382 env->local_vars = 0;
1387 typedef struct walker_env {
1388 inline_irg_env *x; /**< the inline environment */
1389 char ignore_runtime; /**< the ignore runtime flag */
1390 char ignore_callers; /**< if set, do change callers data */
1394 * post-walker: collect all calls in the inline-environment
1395 * of a graph and sum some statistics.
1397 static void collect_calls2(ir_node *call, void *ctx) {
1399 inline_irg_env *x = env->x;
1400 ir_opcode code = get_irn_opcode(call);
1404 /* count meaningful nodes in irg */
1405 if (code != iro_Proj && code != iro_Tuple && code != iro_Sync) {
1406 if (code != iro_Block) {
1414 if (code != iro_Call) return;
1416 /* check, if it's a runtime call */
1417 if (env->ignore_runtime) {
1418 ir_node *symc = get_Call_ptr(call);
1420 if (is_Global(symc)) {
1421 ir_entity *ent = get_Global_entity(symc);
1423 if (get_entity_additional_properties(ent) & mtp_property_runtime)
1428 /* collect all call nodes */
1430 ++x->n_call_nodes_orig;
1432 callee = get_call_called_irg(call);
1433 if (callee != NULL) {
1434 if (! env->ignore_callers) {
1435 inline_irg_env *callee_env = get_irg_link(callee);
1436 /* count all static callers */
1437 ++callee_env->n_callers;
1438 ++callee_env->n_callers_orig;
1440 if (callee == current_ir_graph)
1443 /* link it in the list of possible inlinable entries */
1444 entry = obstack_alloc(&temp_obst, sizeof(*entry));
1446 entry->callee = callee;
1447 entry->loop_depth = get_irn_loop(get_nodes_block(call))->depth;
1448 entry->benefice = 0;
1449 entry->local_adr = 0;
1450 entry->all_const = 0;
1452 list_add_tail(&entry->list, &x->calls);
1457 * Returns TRUE if the number of callers is 0 in the irg's environment,
1458 * hence this irg is a leave.
1460 inline static int is_leave(ir_graph *irg) {
1461 inline_irg_env *env = get_irg_link(irg);
1462 return env->n_call_nodes == 0;
1466 * Returns TRUE if the number of nodes in the callee is
1467 * smaller then size in the irg's environment.
1469 inline static int is_smaller(ir_graph *callee, unsigned size) {
1470 inline_irg_env *env = get_irg_link(callee);
1471 return env->n_nodes < size;
1475 * Duplicate a call entry.
1477 * @param entry the original entry to duplicate
1478 * @param new_call the new call node
1479 * @param loop_depth_delta
1480 * delta value for the loop depth
1482 static call_entry *duplicate_call_entry(const call_entry *entry,
1483 ir_node *new_call, int loop_depth_delta) {
1484 call_entry *nentry = obstack_alloc(&temp_obst, sizeof(*nentry));
1485 nentry->call = new_call;
1486 nentry->callee = entry->callee;
1487 nentry->benefice = entry->benefice;
1488 nentry->loop_depth = entry->loop_depth + loop_depth_delta;
1489 nentry->local_adr = entry->local_adr;
1490 nentry->all_const = entry->all_const;
1496 * Append all call nodes of the source environment to the nodes of in the destination
1499 * @param dst destination environment
1500 * @param src source environment
1501 * @param loop_depth the loop depth of the call that is replaced by the src list
1503 static void append_call_list(inline_irg_env *dst, inline_irg_env *src, int loop_depth) {
1504 call_entry *entry, *nentry;
1506 /* Note that the src list points to Call nodes in the inlined graph, but
1507 we need Call nodes in our graph. Luckily the inliner leaves this information
1508 in the link field. */
1509 list_for_each_entry(call_entry, entry, &src->calls, list) {
1510 nentry = duplicate_call_entry(entry, get_irn_link(entry->call), loop_depth);
1511 list_add_tail(&nentry->list, &dst->calls);
1513 dst->n_call_nodes += src->n_call_nodes;
1514 dst->n_nodes += src->n_nodes;
1518 * Inlines small leave methods at call sites where the called address comes
1519 * from a Const node that references the entity representing the called
1521 * The size argument is a rough measure for the code size of the method:
1522 * Methods where the obstack containing the firm graph is smaller than
1525 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
1526 unsigned size, int ignore_runtime)
1528 inline_irg_env *env;
1534 call_entry *entry, *next;
1535 const call_entry *centry;
1536 pmap *copied_graphs;
1537 pmap_entry *pm_entry;
1539 rem = current_ir_graph;
1540 obstack_init(&temp_obst);
1542 /* a map for the copied graphs, used to inline recursive calls */
1543 copied_graphs = pmap_create();
1545 /* extend all irgs by a temporary data structure for inlining. */
1546 n_irgs = get_irp_n_irgs();
1547 for (i = 0; i < n_irgs; ++i)
1548 set_irg_link(get_irp_irg(i), alloc_inline_irg_env());
1550 /* Pre-compute information in temporary data structure. */
1551 wenv.ignore_runtime = ignore_runtime;
1552 wenv.ignore_callers = 0;
1553 for (i = 0; i < n_irgs; ++i) {
1554 ir_graph *irg = get_irp_irg(i);
1556 assert(get_irg_phase_state(irg) != phase_building);
1557 free_callee_info(irg);
1559 assure_cf_loop(irg);
1560 wenv.x = get_irg_link(irg);
1561 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
1564 /* -- and now inline. -- */
1566 /* Inline leaves recursively -- we might construct new leaves. */
1570 for (i = 0; i < n_irgs; ++i) {
1572 int phiproj_computed = 0;
1574 current_ir_graph = get_irp_irg(i);
1575 env = get_irg_link(current_ir_graph);
1577 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1579 irg_inline_property prop;
1581 if (env->n_nodes > maxsize)
1585 callee = entry->callee;
1587 prop = get_irg_inline_property(callee);
1588 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1589 /* do not inline forbidden / weak graphs */
1593 if (is_leave(callee) && (
1594 is_smaller(callee, leavesize) || prop >= irg_inline_forced)) {
1595 if (!phiproj_computed) {
1596 phiproj_computed = 1;
1597 collect_phiprojs(current_ir_graph);
1599 did_inline = inline_method(call, callee);
1602 inline_irg_env *callee_env = get_irg_link(callee);
1604 /* call was inlined, Phi/Projs for current graph must be recomputed */
1605 phiproj_computed = 0;
1607 /* Do some statistics */
1608 env->got_inline = 1;
1609 --env->n_call_nodes;
1610 env->n_nodes += callee_env->n_nodes;
1611 --callee_env->n_callers;
1613 /* remove this call from the list */
1614 list_del(&entry->list);
1620 } while (did_inline);
1622 /* inline other small functions. */
1623 for (i = 0; i < n_irgs; ++i) {
1625 int phiproj_computed = 0;
1627 current_ir_graph = get_irp_irg(i);
1628 env = get_irg_link(current_ir_graph);
1630 /* note that the list of possible calls is updated during the process */
1631 list_for_each_entry_safe(call_entry, entry, next, &env->calls, list) {
1632 irg_inline_property prop;
1637 callee = entry->callee;
1639 prop = get_irg_inline_property(callee);
1640 if (prop == irg_inline_forbidden || get_irg_additional_properties(callee) & mtp_property_weak) {
1641 /* do not inline forbidden / weak graphs */
1645 e = pmap_find(copied_graphs, callee);
1648 * Remap callee if we have a copy.
1649 * FIXME: Should we do this only for recursive Calls ?
1654 if (prop >= irg_inline_forced ||
1655 (is_smaller(callee, size) && env->n_nodes < maxsize) /* small function */) {
1656 if (current_ir_graph == callee) {
1658 * Recursive call: we cannot directly inline because we cannot walk
1659 * the graph and change it. So we have to make a copy of the graph
1663 inline_irg_env *callee_env;
1667 * No copy yet, create one.
1668 * Note that recursive methods are never leaves, so it is sufficient
1669 * to test this condition here.
1671 copy = create_irg_copy(callee);
1673 /* create_irg_copy() destroys the Proj links, recompute them */
1674 phiproj_computed = 0;
1676 /* allocate new environment */
1677 callee_env = alloc_inline_irg_env();
1678 set_irg_link(copy, callee_env);
1680 assure_cf_loop(copy);
1681 wenv.x = callee_env;
1682 wenv.ignore_callers = 1;
1683 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
1686 * Enter the entity of the original graph. This is needed
1687 * for inline_method(). However, note that ent->irg still points
1688 * to callee, NOT to copy.
1690 set_irg_entity(copy, get_irg_entity(callee));
1692 pmap_insert(copied_graphs, callee, copy);
1695 /* we have only one caller: the original graph */
1696 callee_env->n_callers = 1;
1697 callee_env->n_callers_orig = 1;
1699 if (! phiproj_computed) {
1700 phiproj_computed = 1;
1701 collect_phiprojs(current_ir_graph);
1703 did_inline = inline_method(call, callee);
1705 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1707 /* call was inlined, Phi/Projs for current graph must be recomputed */
1708 phiproj_computed = 0;
1710 /* callee was inline. Append it's call list. */
1711 env->got_inline = 1;
1712 --env->n_call_nodes;
1713 append_call_list(env, callee_env, entry->loop_depth);
1714 --callee_env->n_callers;
1716 /* after we have inlined callee, all called methods inside callee
1717 are now called once more */
1718 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
1719 inline_irg_env *penv = get_irg_link(centry->callee);
1723 /* remove this call from the list */
1724 list_del(&entry->list);
1731 for (i = 0; i < n_irgs; ++i) {
1732 irg = get_irp_irg(i);
1733 env = get_irg_link(irg);
1735 if (env->got_inline) {
1736 optimize_graph_df(irg);
1739 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
1740 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1741 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1742 env->n_callers_orig, env->n_callers,
1743 get_entity_name(get_irg_entity(irg))));
1747 /* kill the copied graphs: we don't need them anymore */
1748 foreach_pmap(copied_graphs, pm_entry) {
1749 ir_graph *copy = pm_entry->value;
1751 /* reset the entity, otherwise it will be deleted in the next step ... */
1752 set_irg_entity(copy, NULL);
1753 free_ir_graph(copy);
1755 pmap_destroy(copied_graphs);
1757 obstack_free(&temp_obst, NULL);
1758 current_ir_graph = rem;
1762 * Calculate the parameter weights for transmitting the address of a local variable.
1764 static unsigned calc_method_local_weight(ir_node *arg) {
1766 unsigned v, weight = 0;
1768 for (i = get_irn_n_outs(arg) - 1; i >= 0; --i) {
1769 ir_node *succ = get_irn_out(arg, i);
1771 switch (get_irn_opcode(succ)) {
1774 /* Loads and Store can be removed */
1778 /* check if all args are constant */
1779 for (j = get_Sel_n_indexs(succ) - 1; j >= 0; --j) {
1780 ir_node *idx = get_Sel_index(succ, j);
1781 if (! is_Const(idx))
1784 /* Check users on this Sel. Note: if a 0 is returned here, there was
1785 some unsupported node. */
1786 v = calc_method_local_weight(succ);
1789 /* we can kill one Sel with constant indexes, this is cheap */
1793 /* when looking backward we might find Id nodes */
1794 weight += calc_method_local_weight(succ);
1797 /* unoptimized tuple */
1798 for (j = get_Tuple_n_preds(succ) - 1; j >= 0; --j) {
1799 ir_node *pred = get_Tuple_pred(succ, j);
1801 /* look for Proj(j) */
1802 for (k = get_irn_n_outs(succ) - 1; k >= 0; --k) {
1803 ir_node *succ_succ = get_irn_out(succ, k);
1804 if (is_Proj(succ_succ)) {
1805 if (get_Proj_proj(succ_succ) == j) {
1807 weight += calc_method_local_weight(succ_succ);
1810 /* this should NOT happen */
1818 /* any other node: unsupported yet or bad. */
1826 * Calculate the parameter weights for transmitting the address of a local variable.
1828 static void analyze_irg_local_weights(inline_irg_env *env, ir_graph *irg) {
1829 ir_entity *ent = get_irg_entity(irg);
1831 int nparams, i, proj_nr;
1832 ir_node *irg_args, *arg;
1834 mtp = get_entity_type(ent);
1835 nparams = get_method_n_params(mtp);
1837 /* allocate a new array. currently used as 'analysed' flag */
1838 env->local_weights = NEW_ARR_D(unsigned, &temp_obst, nparams);
1840 /* If the method haven't parameters we have nothing to do. */
1844 assure_irg_outs(irg);
1845 irg_args = get_irg_args(irg);
1846 for (i = get_irn_n_outs(irg_args) - 1; i >= 0; --i) {
1847 arg = get_irn_out(irg_args, i);
1848 proj_nr = get_Proj_proj(arg);
1849 env->local_weights[proj_nr] = calc_method_local_weight(arg);
1854 * Calculate the benefice for transmitting an local variable address.
1855 * After inlining, the local variable might be transformed into a
1856 * SSA variable by scalar_replacement().
1858 static unsigned get_method_local_adress_weight(ir_graph *callee, int pos) {
1859 inline_irg_env *env = get_irg_link(callee);
1861 if (env->local_weights != NULL) {
1862 if (pos < ARR_LEN(env->local_weights))
1863 return env->local_weights[pos];
1867 analyze_irg_local_weights(env, callee);
1869 if (pos < ARR_LEN(env->local_weights))
1870 return env->local_weights[pos];
1875 * Calculate a benefice value for inlining the given call.
1877 * @param call the call node we have to inspect
1878 * @param callee the called graph
1880 static int calc_inline_benefice(call_entry *entry, ir_graph *callee)
1882 ir_node *call = entry->call;
1883 ir_entity *ent = get_irg_entity(callee);
1887 int i, n_params, all_const;
1889 irg_inline_property prop;
1891 inline_irg_env *callee_env;
1893 prop = get_irg_inline_property(callee);
1894 if (prop == irg_inline_forbidden) {
1895 DB((dbg, LEVEL_2, "In %+F Call to %+F: inlining forbidden\n",
1897 return entry->benefice = INT_MIN;
1900 if (get_irg_additional_properties(callee) & (mtp_property_noreturn | mtp_property_weak)) {
1901 DB((dbg, LEVEL_2, "In %+F Call to %+F: not inlining noreturn or weak\n",
1903 return entry->benefice = INT_MIN;
1906 /* costs for every passed parameter */
1907 n_params = get_Call_n_params(call);
1908 mtp = get_entity_type(ent);
1909 cc = get_method_calling_convention(mtp);
1910 if (cc & cc_reg_param) {
1911 /* register parameter, smaller costs for register parameters */
1912 int max_regs = cc & ~cc_bits;
1914 if (max_regs < n_params)
1915 weight += max_regs * 2 + (n_params - max_regs) * 5;
1917 weight += n_params * 2;
1919 /* parameters are passed an stack */
1920 weight += 5 * n_params;
1923 /* constant parameters improve the benefice */
1924 frame_ptr = get_irg_frame(current_ir_graph);
1926 for (i = 0; i < n_params; ++i) {
1927 ir_node *param = get_Call_param(call, i);
1929 if (is_Const(param)) {
1930 weight += get_method_param_weight(ent, i);
1933 if (is_SymConst(param))
1934 weight += get_method_param_weight(ent, i);
1935 else if (is_Sel(param) && get_Sel_ptr(param) == frame_ptr) {
1937 * An address of a local variable is transmitted. After
1938 * inlining, scalar_replacement might be able to remove the
1939 * local variable, so honor this.
1941 v = get_method_local_adress_weight(callee, i);
1944 entry->local_adr = 1;
1948 entry->all_const = all_const;
1950 callee_env = get_irg_link(callee);
1951 if (callee_env->n_callers == 1 &&
1952 callee != current_ir_graph &&
1953 get_entity_visibility(ent) == visibility_local) {
1957 /* give a bonus for functions with one block */
1958 if (callee_env->n_blocks == 1)
1959 weight = weight * 3 / 2;
1961 /* and one for small non-recursive functions: we want them to be inlined in mostly every case */
1962 if (callee_env->n_nodes < 30 && !callee_env->recursive)
1965 /* and finally for leaves: they do not increase the register pressure
1966 because of callee safe registers */
1967 if (callee_env->n_call_nodes == 0)
1970 /** it's important to inline inner loops first */
1971 if (entry->loop_depth > 30)
1972 weight += 30 * 1024;
1974 weight += entry->loop_depth * 1024;
1977 * All arguments constant is probably a good sign, give an extra bonus
1982 return entry->benefice = weight;
1985 static ir_graph **irgs;
1986 static int last_irg;
1989 * Callgraph walker, collect all visited graphs.
1991 static void callgraph_walker(ir_graph *irg, void *data) {
1993 irgs[last_irg++] = irg;
1997 * Creates an inline order for all graphs.
1999 * @return the list of graphs.
2001 static ir_graph **create_irg_list(void) {
2002 ir_entity **free_methods;
2004 int n_irgs = get_irp_n_irgs();
2006 cgana(&arr_len, &free_methods);
2007 xfree(free_methods);
2009 compute_callgraph();
2012 irgs = XMALLOCNZ(ir_graph*, n_irgs);
2014 callgraph_walk(NULL, callgraph_walker, NULL);
2015 assert(n_irgs == last_irg);
2021 * Push a call onto the priority list if its benefice is big enough.
2023 * @param pqueue the priority queue of calls
2024 * @param call the call entry
2025 * @param inlien_threshold
2026 * the threshold value
2028 static void maybe_push_call(pqueue_t *pqueue, call_entry *call,
2029 int inline_threshold)
2031 ir_graph *callee = call->callee;
2032 irg_inline_property prop = get_irg_inline_property(callee);
2033 int benefice = calc_inline_benefice(call, callee);
2035 DB((dbg, LEVEL_2, "In %+F Call %+F to %+F has benefice %d\n",
2036 get_irn_irg(call->call), call->call, callee, benefice));
2038 if (prop < irg_inline_forced && benefice < inline_threshold) {
2042 pqueue_put(pqueue, call, benefice);
2046 * Try to inline calls into a graph.
2048 * @param irg the graph into which we inline
2049 * @param maxsize do NOT inline if the size of irg gets
2050 * bigger than this amount
2051 * @param inline_threshold
2052 * threshold value for inline decision
2053 * @param copied_graphs
2054 * map containing copied of recursive graphs
2056 static void inline_into(ir_graph *irg, unsigned maxsize,
2057 int inline_threshold, pmap *copied_graphs)
2059 int phiproj_computed = 0;
2060 inline_irg_env *env = get_irg_link(irg);
2061 call_entry *curr_call;
2065 if (env->n_call_nodes == 0)
2068 if (env->n_nodes > maxsize) {
2069 DB((dbg, LEVEL_2, "%+F: too big (%d)\n", irg, env->n_nodes));
2073 current_ir_graph = irg;
2075 /* put irgs into the pqueue */
2076 pqueue = new_pqueue();
2078 list_for_each_entry(call_entry, curr_call, &env->calls, list) {
2079 assert(is_Call(curr_call->call));
2080 maybe_push_call(pqueue, curr_call, inline_threshold);
2083 /* note that the list of possible calls is updated during the process */
2084 while (!pqueue_empty(pqueue)) {
2086 call_entry *curr_call = pqueue_pop_front(pqueue);
2087 ir_graph *callee = curr_call->callee;
2088 ir_node *call_node = curr_call->call;
2089 inline_irg_env *callee_env = get_irg_link(callee);
2090 irg_inline_property prop = get_irg_inline_property(callee);
2092 const call_entry *centry;
2095 if ((prop < irg_inline_forced) && env->n_nodes + callee_env->n_nodes > maxsize) {
2096 DB((dbg, LEVEL_2, "%+F: too big (%d) + %+F (%d)\n", irg,
2097 env->n_nodes, callee, callee_env->n_nodes));
2101 e = pmap_find(copied_graphs, callee);
2103 int benefice = curr_call->benefice;
2105 * Reduce the weight for recursive function IFF not all arguments are const.
2106 * inlining recursive functions is rarely good.
2108 if (!curr_call->all_const)
2110 if (benefice < inline_threshold)
2114 * Remap callee if we have a copy.
2117 callee_env = get_irg_link(callee);
2120 if (current_ir_graph == callee) {
2122 * Recursive call: we cannot directly inline because we cannot
2123 * walk the graph and change it. So we have to make a copy of
2126 int benefice = curr_call->benefice;
2130 * Reduce the weight for recursive function IFF not all arguments are const.
2131 * inlining recursive functions is rarely good.
2133 if (!curr_call->all_const)
2135 if (benefice < inline_threshold)
2139 * No copy yet, create one.
2140 * Note that recursive methods are never leaves, so it is
2141 * sufficient to test this condition here.
2143 copy = create_irg_copy(callee);
2145 /* create_irg_copy() destroys the Proj links, recompute them */
2146 phiproj_computed = 0;
2148 /* allocate a new environment */
2149 callee_env = alloc_inline_irg_env();
2150 set_irg_link(copy, callee_env);
2152 assure_cf_loop(copy);
2153 wenv.x = callee_env;
2154 wenv.ignore_callers = 1;
2155 irg_walk_graph(copy, NULL, collect_calls2, &wenv);
2158 * Enter the entity of the original graph. This is needed
2159 * for inline_method(). However, note that ent->irg still points
2160 * to callee, NOT to copy.
2162 set_irg_entity(copy, get_irg_entity(callee));
2164 pmap_insert(copied_graphs, callee, copy);
2167 /* we have only one caller: the original graph */
2168 callee_env->n_callers = 1;
2169 callee_env->n_callers_orig = 1;
2171 if (! phiproj_computed) {
2172 phiproj_computed = 1;
2173 collect_phiprojs(current_ir_graph);
2175 did_inline = inline_method(call_node, callee);
2179 /* call was inlined, Phi/Projs for current graph must be recomputed */
2180 phiproj_computed = 0;
2182 /* remove it from the caller list */
2183 list_del(&curr_call->list);
2185 /* callee was inline. Append it's call list. */
2186 env->got_inline = 1;
2187 if (curr_call->local_adr)
2188 env->local_vars = 1;
2189 --env->n_call_nodes;
2191 /* we just generate a bunch of new calls */
2192 loop_depth = curr_call->loop_depth;
2193 list_for_each_entry(call_entry, centry, &callee_env->calls, list) {
2194 inline_irg_env *penv = get_irg_link(centry->callee);
2196 call_entry *new_entry;
2198 /* after we have inlined callee, all called methods inside
2199 * callee are now called once more */
2202 /* Note that the src list points to Call nodes in the inlined graph,
2203 * but we need Call nodes in our graph. Luckily the inliner leaves
2204 * this information in the link field. */
2205 new_call = get_irn_link(centry->call);
2206 assert(is_Call(new_call));
2208 new_entry = duplicate_call_entry(centry, new_call, loop_depth);
2209 list_add_tail(&new_entry->list, &env->calls);
2210 maybe_push_call(pqueue, new_entry, inline_threshold);
2213 env->n_call_nodes += callee_env->n_call_nodes;
2214 env->n_nodes += callee_env->n_nodes;
2215 --callee_env->n_callers;
2222 * Heuristic inliner. Calculates a benefice value for every call and inlines
2223 * those calls with a value higher than the threshold.
2225 void inline_functions(unsigned maxsize, int inline_threshold) {
2226 inline_irg_env *env;
2230 pmap *copied_graphs;
2231 pmap_entry *pm_entry;
2234 rem = current_ir_graph;
2235 obstack_init(&temp_obst);
2237 irgs = create_irg_list();
2239 /* a map for the copied graphs, used to inline recursive calls */
2240 copied_graphs = pmap_create();
2242 /* extend all irgs by a temporary data structure for inlining. */
2243 n_irgs = get_irp_n_irgs();
2244 for (i = 0; i < n_irgs; ++i)
2245 set_irg_link(irgs[i], alloc_inline_irg_env());
2247 /* Pre-compute information in temporary data structure. */
2248 wenv.ignore_runtime = 0;
2249 wenv.ignore_callers = 0;
2250 for (i = 0; i < n_irgs; ++i) {
2251 ir_graph *irg = irgs[i];
2253 free_callee_info(irg);
2255 wenv.x = get_irg_link(irg);
2256 assure_cf_loop(irg);
2257 irg_walk_graph(irg, NULL, collect_calls2, &wenv);
2260 /* -- and now inline. -- */
2261 for (i = 0; i < n_irgs; ++i) {
2262 ir_graph *irg = irgs[i];
2264 inline_into(irg, maxsize, inline_threshold, copied_graphs);
2267 for (i = 0; i < n_irgs; ++i) {
2268 ir_graph *irg = irgs[i];
2270 env = get_irg_link(irg);
2271 if (env->got_inline) {
2272 /* this irg got calls inlined: optimize it */
2273 if (get_opt_combo()) {
2274 if (env->local_vars) {
2275 scalar_replacement_opt(irg);
2279 if (env->local_vars) {
2280 if (scalar_replacement_opt(irg)) {
2281 optimize_graph_df(irg);
2287 if (env->got_inline || (env->n_callers_orig != env->n_callers)) {
2288 DB((dbg, LEVEL_1, "Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
2289 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
2290 env->n_callers_orig, env->n_callers,
2291 get_entity_name(get_irg_entity(irg))));
2295 /* kill the copied graphs: we don't need them anymore */
2296 foreach_pmap(copied_graphs, pm_entry) {
2297 ir_graph *copy = pm_entry->value;
2299 /* reset the entity, otherwise it will be deleted in the next step ... */
2300 set_irg_entity(copy, NULL);
2301 free_ir_graph(copy);
2303 pmap_destroy(copied_graphs);
2307 obstack_free(&temp_obst, NULL);
2308 current_ir_graph = rem;
2311 void firm_init_inline(void) {
2312 FIRM_DBG_REGISTER(dbg, "firm.opt.inline");