3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
21 #include "irgraph_t.h"
33 #include "pdeq.h" /* Fuer code placement */
38 #include "irbackedge_t.h"
44 #include "iredges_t.h"
47 /* Defined in iropt.c */
48 pset *new_identities (void);
49 void del_identities (pset *value_table);
50 void add_identities (pset *value_table, ir_node *node);
52 /*------------------------------------------------------------------*/
53 /* apply optimizations of iropt to all nodes. */
54 /*------------------------------------------------------------------*/
56 static void init_link (ir_node *n, void *env) {
57 set_irn_link(n, NULL);
60 #if 0 /* Old version. Avoids Ids.
61 This is not necessary: we do a post walk, and get_irn_n
62 removes ids anyways. So it's much cheaper to call the
63 optimization less often and use the exchange() algorithm. */
65 optimize_in_place_wrapper (ir_node *n, void *env) {
67 ir_node *optimized, *old;
69 irn_arity = get_irn_arity(n);
70 for (i = 0; i < irn_arity; i++) {
71 /* get_irn_n skips Id nodes, so comparison old != optimized does not
72 show all optimizations. Therefore always set new predecessor. */
73 old = get_irn_intra_n(n, i);
74 optimized = optimize_in_place_2(old);
75 set_irn_n(n, i, optimized);
78 if (get_irn_op(n) == op_Block) {
79 optimized = optimize_in_place_2(n);
80 if (optimized != n) exchange (n, optimized);
85 optimize_in_place_wrapper (ir_node *n, void *env) {
86 ir_node *optimized = optimize_in_place_2(n);
87 if (optimized != n) exchange (n, optimized);
92 static INLINE void do_local_optimize(ir_node *n) {
93 /* Handle graph state */
94 assert(get_irg_phase_state(current_ir_graph) != phase_building);
96 if (get_opt_global_cse())
97 set_irg_pinned(current_ir_graph, op_pin_state_floats);
98 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
99 set_irg_outs_inconsistent(current_ir_graph);
100 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
101 set_irg_dom_inconsistent(current_ir_graph);
102 set_irg_loopinfo_inconsistent(current_ir_graph);
104 /* Clean the value_table in irg for the CSE. */
105 del_identities(current_ir_graph->value_table);
106 current_ir_graph->value_table = new_identities();
108 /* walk over the graph */
109 irg_walk(n, init_link, optimize_in_place_wrapper, NULL);
112 void local_optimize_node(ir_node *n) {
113 ir_graph *rem = current_ir_graph;
114 current_ir_graph = get_irn_irg(n);
116 do_local_optimize(n);
118 current_ir_graph = rem;
122 * Block-Walker: uses dominance depth to mark dead blocks.
124 static void kill_dead_blocks(ir_node *block, void *env)
126 if (get_Block_dom_depth(block) < 0)
127 set_Block_dead(block);
131 local_optimize_graph (ir_graph *irg) {
132 ir_graph *rem = current_ir_graph;
133 current_ir_graph = irg;
135 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
136 irg_block_walk_graph(irg, NULL, kill_dead_blocks, NULL);
138 do_local_optimize(irg->end);
140 current_ir_graph = rem;
144 /*------------------------------------------------------------------*/
145 /* Routines for dead node elimination / copying garbage collection */
146 /* of the obstack. */
147 /*------------------------------------------------------------------*/
150 * Remember the new node in the old node by using a field all nodes have.
153 set_new_node (ir_node *old, ir_node *new)
159 * Get this new node, before the old node is forgotten.
161 static INLINE ir_node *
162 get_new_node (ir_node * n) {
167 * We use the block_visited flag to mark that we have computed the
168 * number of useful predecessors for this block.
169 * Further we encode the new arity in this flag in the old blocks.
170 * Remembering the arity is useful, as it saves a lot of pointer
171 * accesses. This function is called for all Phi and Block nodes
175 compute_new_arity(ir_node *b) {
176 int i, res, irn_arity;
179 irg_v = get_irg_block_visited(current_ir_graph);
180 block_v = get_Block_block_visited(b);
181 if (block_v >= irg_v) {
182 /* we computed the number of preds for this block and saved it in the
184 return block_v - irg_v;
186 /* compute the number of good predecessors */
187 res = irn_arity = get_irn_arity(b);
188 for (i = 0; i < irn_arity; i++)
189 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
190 /* save it in the flag. */
191 set_Block_block_visited(b, irg_v + res);
196 /* TODO: add an ir_op operation */
197 static INLINE void new_backedge_info(ir_node *n) {
198 switch(get_irn_opcode(n)) {
200 n->attr.block.cg_backedge = NULL;
201 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
204 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
207 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
214 * Copies the node to the new obstack. The Ins of the new node point to
215 * the predecessors on the old obstack. For block/phi nodes not all
216 * predecessors might be copied. n->link points to the new node.
217 * For Phi and Block nodes the function allocates in-arrays with an arity
218 * only for useful predecessors. The arity is determined by counting
219 * the non-bad predecessors of the block.
221 * @param n The node to be copied
222 * @param env if non-NULL, the node number attribute will be copied to the new node
224 * Note: Also used for loop unrolling.
226 static void copy_node(ir_node *n, void *env) {
229 ir_op *op = get_irn_op(n);
230 int copy_node_nr = env != NULL;
232 /* The end node looses it's flexible in array. This doesn't matter,
233 as dead node elimination builds End by hand, inlineing doesn't use
235 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
238 /* node copied already */
240 } else if (op == op_Block) {
242 new_arity = compute_new_arity(n);
243 n->attr.block.graph_arr = NULL;
245 block = get_nodes_block(n);
247 new_arity = compute_new_arity(block);
249 new_arity = get_irn_arity(n);
252 nn = new_ir_node(get_irn_dbg_info(n),
259 /* Copy the attributes. These might point to additional data. If this
260 was allocated on the old obstack the pointers now are dangling. This
261 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
262 copy_node_attr(n, nn);
263 new_backedge_info(nn);
267 /* for easier debugging, we want to copy the node numbers too */
268 nn->node_nr = n->node_nr;
276 * Copies new predecessors of old node to new node remembered in link.
277 * Spare the Bad predecessors of Phi and Block nodes.
280 copy_preds (ir_node *n, void *env) {
284 nn = get_new_node(n);
286 /* printf("\n old node: "); DDMSG2(n);
287 printf(" new node: "); DDMSG2(nn);
288 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
291 /* Don't copy Bad nodes. */
293 irn_arity = get_irn_arity(n);
294 for (i = 0; i < irn_arity; i++)
295 if (! is_Bad(get_irn_n(n, i))) {
296 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
297 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
300 /* repair the block visited flag from above misuse. Repair it in both
301 graphs so that the old one can still be used. */
302 set_Block_block_visited(nn, 0);
303 set_Block_block_visited(n, 0);
304 /* Local optimization could not merge two subsequent blocks if
305 in array contained Bads. Now it's possible.
306 We don't call optimize_in_place as it requires
307 that the fields in ir_graph are set properly. */
308 if ((get_opt_control_flow_straightening()) &&
309 (get_Block_n_cfgpreds(nn) == 1) &&
310 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
311 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
313 /* Jmp jumps into the block it is in -- deal self cycle. */
314 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
315 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
320 } else if (get_irn_op(n) == op_Phi) {
321 /* Don't copy node if corresponding predecessor in block is Bad.
322 The Block itself should not be Bad. */
323 block = get_nodes_block(n);
324 set_irn_n (nn, -1, get_new_node(block));
326 irn_arity = get_irn_arity(n);
327 for (i = 0; i < irn_arity; i++)
328 if (! is_Bad(get_irn_n(block, i))) {
329 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
330 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
333 /* If the pre walker reached this Phi after the post walker visited the
334 block block_visited is > 0. */
335 set_Block_block_visited(get_nodes_block(n), 0);
336 /* Compacting the Phi's ins might generate Phis with only one
338 if (get_irn_arity(nn) == 1)
339 exchange(nn, get_irn_n(nn, 0));
341 irn_arity = get_irn_arity(n);
342 for (i = -1; i < irn_arity; i++)
343 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
345 /* Now the new node is complete. We can add it to the hash table for CSE.
346 @@@ inlinening aborts if we identify End. Why? */
347 if (get_irn_op(nn) != op_End)
348 add_identities (current_ir_graph->value_table, nn);
352 * Copies the graph recursively, compacts the keepalive of the end node.
354 * @param copy_node_nr If non-zero, the node number will be copied
357 copy_graph (int copy_node_nr) {
358 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
359 ir_node *ka; /* keep alive */
362 oe = get_irg_end(current_ir_graph);
363 /* copy the end node by hand, allocate dynamic in array! */
364 ne = new_ir_node(get_irn_dbg_info(oe),
371 /* Copy the attributes. Well, there might be some in the future... */
372 copy_node_attr(oe, ne);
373 set_new_node(oe, ne);
375 /* copy the Bad node */
376 ob = get_irg_bad(current_ir_graph);
377 nb = new_ir_node(get_irn_dbg_info(ob),
384 set_new_node(ob, nb);
386 /* copy the NoMem node */
387 om = get_irg_no_mem(current_ir_graph);
388 nm = new_ir_node(get_irn_dbg_info(om),
395 set_new_node(om, nm);
397 /* copy the live nodes */
398 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
399 /* copy_preds for the end node ... */
400 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
402 /*- ... and now the keep alives. -*/
403 /* First pick the not marked block nodes and walk them. We must pick these
404 first as else we will oversee blocks reachable from Phis. */
405 irn_arity = get_irn_arity(oe);
406 for (i = 0; i < irn_arity; i++) {
407 ka = get_irn_intra_n(oe, i);
408 if ((get_irn_op(ka) == op_Block) &&
409 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
410 /* We must keep the block alive and copy everything reachable */
411 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
412 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
413 add_End_keepalive(ne, get_new_node(ka));
417 /* Now pick the Phis. Here we will keep all! */
418 irn_arity = get_irn_arity(oe);
419 for (i = 0; i < irn_arity; i++) {
420 ka = get_irn_intra_n(oe, i);
421 if ((get_irn_op(ka) == op_Phi)) {
422 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
423 /* We didn't copy the Phi yet. */
424 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
425 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
427 add_End_keepalive(ne, get_new_node(ka));
431 /* start block sometimes only reached after keep alives */
432 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
433 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
437 * Copies the graph reachable from current_ir_graph->end to the obstack
438 * in current_ir_graph and fixes the environment.
439 * Then fixes the fields in current_ir_graph containing nodes of the
442 * @param copy_node_nr If non-zero, the node number will be copied
445 copy_graph_env (int copy_node_nr) {
447 /* Not all nodes remembered in current_ir_graph might be reachable
448 from the end node. Assure their link is set to NULL, so that
449 we can test whether new nodes have been computed. */
450 set_irn_link(get_irg_frame (current_ir_graph), NULL);
451 set_irn_link(get_irg_globals (current_ir_graph), NULL);
452 set_irn_link(get_irg_args (current_ir_graph), NULL);
453 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
454 set_irn_link(get_irg_bad (current_ir_graph), NULL);
455 set_irn_link(get_irg_no_mem (current_ir_graph), NULL);
457 /* we use the block walk flag for removing Bads from Blocks ins. */
458 inc_irg_block_visited(current_ir_graph);
461 copy_graph(copy_node_nr);
463 /* fix the fields in current_ir_graph */
464 old_end = get_irg_end(current_ir_graph);
465 set_irg_end (current_ir_graph, get_new_node(old_end));
466 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
467 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
469 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
471 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
472 copy_node (get_irg_frame(current_ir_graph), INT_TO_PTR(copy_node_nr));
473 copy_preds(get_irg_frame(current_ir_graph), NULL);
475 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
476 copy_node (get_irg_globals(current_ir_graph), INT_TO_PTR(copy_node_nr));
477 copy_preds(get_irg_globals(current_ir_graph), NULL);
479 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
480 copy_node (get_irg_initial_mem(current_ir_graph), INT_TO_PTR(copy_node_nr));
481 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
483 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
484 copy_node (get_irg_args(current_ir_graph), INT_TO_PTR(copy_node_nr));
485 copy_preds(get_irg_args(current_ir_graph), NULL);
487 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
488 copy_node(get_irg_bad(current_ir_graph), INT_TO_PTR(copy_node_nr));
489 copy_preds(get_irg_bad(current_ir_graph), NULL);
491 if (get_irn_link(get_irg_no_mem(current_ir_graph)) == NULL) {
492 copy_node(get_irg_no_mem(current_ir_graph), INT_TO_PTR(copy_node_nr));
493 copy_preds(get_irg_no_mem(current_ir_graph), NULL);
495 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
496 set_irg_start_block(current_ir_graph, get_new_node(get_irg_start_block(current_ir_graph)));
497 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
498 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
499 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
500 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
501 set_irg_bad (current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
502 set_irg_no_mem (current_ir_graph, get_new_node(get_irg_no_mem(current_ir_graph)));
506 * Copies all reachable nodes to a new obstack. Removes bad inputs
507 * from block nodes and the corresponding inputs from Phi nodes.
508 * Merges single exit blocks with single entry blocks and removes
510 * Adds all new nodes to a new hash table for CSE. Does not
511 * perform CSE, so the hash table might contain common subexpressions.
514 dead_node_elimination(ir_graph *irg) {
516 int rem_ipview = get_interprocedural_view();
517 struct obstack *graveyard_obst = NULL;
518 struct obstack *rebirth_obst = NULL;
520 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
521 assert(! edges_activated(irg) && "dead node elimination requieres disabled edges");
523 /* inform statistics that we started a dead-node elimination run */
524 hook_dead_node_elim(irg, 1);
526 /* Remember external state of current_ir_graph. */
527 rem = current_ir_graph;
528 current_ir_graph = irg;
529 set_interprocedural_view(0);
531 assert(get_irg_phase_state(current_ir_graph) != phase_building);
533 /* Handle graph state */
534 free_callee_info(current_ir_graph);
535 free_irg_outs(current_ir_graph);
538 /* @@@ so far we loose loops when copying */
539 free_loop_information(current_ir_graph);
541 if (get_irg_dom_state(irg) != dom_none)
542 set_irg_dom_inconsistent(irg);
544 /* A quiet place, where the old obstack can rest in peace,
545 until it will be cremated. */
546 graveyard_obst = irg->obst;
548 /* A new obstack, where the reachable nodes will be copied to. */
549 rebirth_obst = xmalloc (sizeof(*rebirth_obst));
550 current_ir_graph->obst = rebirth_obst;
551 obstack_init (current_ir_graph->obst);
553 /* We also need a new hash table for cse */
554 del_identities (irg->value_table);
555 irg->value_table = new_identities ();
557 /* Copy the graph from the old to the new obstack */
560 /* Free memory from old unoptimized obstack */
561 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
562 xfree (graveyard_obst); /* ... then free it. */
564 /* inform statistics that the run is over */
565 hook_dead_node_elim(irg, 0);
567 current_ir_graph = rem;
568 set_interprocedural_view(rem_ipview);
573 * Relink bad predecessors of a block and store the old in array to the
574 * link field. This function is called by relink_bad_predecessors().
575 * The array of link field starts with the block operand at position 0.
576 * If block has bad predecessors, create a new in array without bad preds.
577 * Otherwise let in array untouched.
579 static void relink_bad_block_predecessors(ir_node *n, void *env) {
580 ir_node **new_in, *irn;
581 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
583 /* if link field of block is NULL, look for bad predecessors otherwise
584 this is already done */
585 if (get_irn_op(n) == op_Block &&
586 get_irn_link(n) == NULL) {
588 /* save old predecessors in link field (position 0 is the block operand)*/
589 set_irn_link(n, get_irn_in(n));
591 /* count predecessors without bad nodes */
592 old_irn_arity = get_irn_arity(n);
593 for (i = 0; i < old_irn_arity; i++)
594 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
596 /* arity changing: set new predecessors without bad nodes */
597 if (new_irn_arity < old_irn_arity) {
598 /* Get new predecessor array. We do not resize the array, as we must
599 keep the old one to update Phis. */
600 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
602 /* set new predecessors in array */
605 for (i = 0; i < old_irn_arity; i++) {
606 irn = get_irn_n(n, i);
608 new_in[new_irn_n] = irn;
609 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
613 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
614 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
617 } /* ir node has bad predecessors */
619 } /* Block is not relinked */
623 * Relinks Bad predecessors from Blocks and Phis called by walker
624 * remove_bad_predecesors(). If n is a Block, call
625 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
626 * function of Phi's Block. If this block has bad predecessors, relink preds
629 static void relink_bad_predecessors(ir_node *n, void *env) {
630 ir_node *block, **old_in;
631 int i, old_irn_arity, new_irn_arity;
633 /* relink bad predecessors of a block */
634 if (get_irn_op(n) == op_Block)
635 relink_bad_block_predecessors(n, env);
637 /* If Phi node relink its block and its predecessors */
638 if (get_irn_op(n) == op_Phi) {
640 /* Relink predecessors of phi's block */
641 block = get_nodes_block(n);
642 if (get_irn_link(block) == NULL)
643 relink_bad_block_predecessors(block, env);
645 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
646 old_irn_arity = ARR_LEN(old_in);
648 /* Relink Phi predecessors if count of predecessors changed */
649 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
650 /* set new predecessors in array
651 n->in[0] remains the same block */
653 for(i = 1; i < old_irn_arity; i++)
654 if (!is_Bad((ir_node *)old_in[i])) {
655 n->in[new_irn_arity] = n->in[i];
656 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
660 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
661 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
664 } /* n is a Phi node */
668 * Removes Bad Bad predecessors from Blocks and the corresponding
669 * inputs to Phi nodes as in dead_node_elimination but without
671 * On walking up set the link field to NULL, on walking down call
672 * relink_bad_predecessors() (This function stores the old in array
673 * to the link field and sets a new in array if arity of predecessors
676 void remove_bad_predecessors(ir_graph *irg) {
677 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
681 /*--------------------------------------------------------------------*/
682 /* Functionality for inlining */
683 /*--------------------------------------------------------------------*/
686 * Copy node for inlineing. Updates attributes that change when
687 * inlineing but not for dead node elimination.
689 * Copies the node by calling copy_node() and then updates the entity if
690 * it's a local one. env must be a pointer of the frame type of the
691 * inlined procedure. The new entities must be in the link field of
695 copy_node_inline (ir_node *n, void *env) {
697 type *frame_tp = (type *)env;
700 if (get_irn_op(n) == op_Sel) {
701 new = get_new_node (n);
702 assert(get_irn_op(new) == op_Sel);
703 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
704 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
706 } else if (get_irn_op(n) == op_Block) {
707 new = get_new_node (n);
708 new->attr.block.irg = current_ir_graph;
712 static void find_addr(ir_node *node, void *env)
714 if (get_irn_opcode(node) == iro_Proj) {
715 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
721 * currently, we cannot inline two cases:
722 * - call with compound arguments
723 * - graphs that take the address of a parameter
725 * check these conditions here
727 static int can_inline(ir_node *call, ir_graph *called_graph)
729 type *call_type = get_Call_type(call);
730 int params, ress, i, res;
731 assert(is_Method_type(call_type));
733 params = get_method_n_params(call_type);
734 ress = get_method_n_ress(call_type);
737 for (i = 0; i < params; ++i) {
738 type *p_type = get_method_param_type(call_type, i);
740 if (is_compound_type(p_type))
745 for (i = 0; i < ress; ++i) {
746 type *r_type = get_method_res_type(call_type, i);
748 if (is_compound_type(r_type))
753 irg_walk_graph(called_graph, find_addr, NULL, &res);
758 int inline_method(ir_node *call, ir_graph *called_graph) {
760 ir_node *post_call, *post_bl;
762 ir_node *end, *end_bl;
766 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
769 irg_inline_property prop = get_irg_inline_property(called_graph);
771 if ( (prop != irg_inline_forced) &&
772 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
774 /* Do not inline variadic functions. */
775 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
778 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
779 get_method_n_params(get_Call_type(call)));
782 * currently, we cannot inline two cases:
783 * - call with compound arguments
784 * - graphs that take the address of a parameter
786 if (! can_inline(call, called_graph))
789 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
790 rem_opt = get_opt_optimize();
793 /* Handle graph state */
794 assert(get_irg_phase_state(current_ir_graph) != phase_building);
795 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
796 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
797 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
798 set_irg_outs_inconsistent(current_ir_graph);
799 set_irg_loopinfo_inconsistent(current_ir_graph);
800 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
802 /* -- Check preconditions -- */
803 assert(get_irn_op(call) == op_Call);
804 /* @@@ does not work for InterfaceIII.java after cgana
805 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
806 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
807 get_Call_type(call)));
809 assert(get_type_tpop(get_Call_type(call)) == type_method);
810 if (called_graph == current_ir_graph) {
811 set_optimize(rem_opt);
815 /* here we know we WILL inline, so inform the statistics */
816 hook_inline(call, called_graph);
818 /* -- Decide how to handle exception control flow: Is there a handler
819 for the Call node, or do we branch directly to End on an exception?
821 0 There is a handler.
823 2 Exception handling not represented in Firm. -- */
825 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
826 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
827 assert(get_irn_op(proj) == op_Proj);
828 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
829 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
831 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
832 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
833 else { exc_handling = 2; } /* !Mproj && !Xproj */
838 the procedure and later replaces the Start node of the called graph.
839 Post_call is the old Call node and collects the results of the called
840 graph. Both will end up being a tuple. -- */
841 post_bl = get_nodes_block(call);
842 set_irg_current_block(current_ir_graph, post_bl);
843 /* XxMxPxP of Start + parameter of Call */
844 in[pn_Start_X_initial_exec] = new_Jmp();
845 in[pn_Start_M] = get_Call_mem(call);
846 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
847 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
848 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
849 /* in[pn_Start_P_value_arg_base] = ??? */
850 pre_call = new_Tuple(5, in);
854 The new block gets the ins of the old block, pre_call and all its
855 predecessors and all Phi nodes. -- */
856 part_block(pre_call);
858 /* -- Prepare state for dead node elimination -- */
859 /* Visited flags in calling irg must be >= flag in called irg.
860 Else walker and arity computation will not work. */
861 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
862 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
863 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
864 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
865 /* Set pre_call as new Start node in link field of the start node of
866 calling graph and pre_calls block as new block for the start block
868 Further mark these nodes so that they are not visited by the
870 set_irn_link(get_irg_start(called_graph), pre_call);
871 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
872 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
873 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
874 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
875 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
877 /* Initialize for compaction of in arrays */
878 inc_irg_block_visited(current_ir_graph);
880 /* -- Replicate local entities of the called_graph -- */
881 /* copy the entities. */
882 called_frame = get_irg_frame_type(called_graph);
883 for (i = 0; i < get_class_n_members(called_frame); i++) {
884 entity *new_ent, *old_ent;
885 old_ent = get_class_member(called_frame, i);
886 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
887 set_entity_link(old_ent, new_ent);
890 /* visited is > than that of called graph. With this trick visited will
891 remain unchanged so that an outer walker, e.g., searching the call nodes
892 to inline, calling this inline will not visit the inlined nodes. */
893 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
895 /* -- Performing dead node elimination inlines the graph -- */
896 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
898 /* @@@ endless loops are not copied!! -- they should be, I think... */
899 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
900 get_irg_frame_type(called_graph));
902 /* Repair called_graph */
903 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
904 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
905 set_Block_block_visited(get_irg_start_block(called_graph), 0);
907 /* -- Merge the end of the inlined procedure with the call site -- */
908 /* We will turn the old Call node into a Tuple with the following
911 0: Phi of all Memories of Return statements.
912 1: Jmp from new Block that merges the control flow from all exception
913 predecessors of the old end block.
914 2: Tuple of all arguments.
915 3: Phi of Exception memories.
916 In case the old Call directly branches to End on an exception we don't
917 need the block merging all exceptions nor the Phi of the exception
921 /* -- Precompute some values -- */
922 end_bl = get_new_node(get_irg_end_block(called_graph));
923 end = get_new_node(get_irg_end(called_graph));
924 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
925 n_res = get_method_n_ress(get_Call_type(call));
927 res_pred = xmalloc (n_res * sizeof(*res_pred));
928 cf_pred = xmalloc (arity * sizeof(*res_pred));
930 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
932 /* -- archive keepalives -- */
933 irn_arity = get_irn_arity(end);
934 for (i = 0; i < irn_arity; i++)
935 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
937 /* The new end node will die. We need not free as the in array is on the obstack:
938 copy_node() only generated 'D' arrays. */
940 /* -- Replace Return nodes by Jump nodes. -- */
942 for (i = 0; i < arity; i++) {
944 ret = get_irn_n(end_bl, i);
945 if (get_irn_op(ret) == op_Return) {
946 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
950 set_irn_in(post_bl, n_ret, cf_pred);
952 /* -- Build a Tuple for all results of the method.
953 Add Phi node if there was more than one Return. -- */
954 turn_into_tuple(post_call, 4);
955 /* First the Memory-Phi */
957 for (i = 0; i < arity; i++) {
958 ret = get_irn_n(end_bl, i);
959 if (get_irn_op(ret) == op_Return) {
960 cf_pred[n_ret] = get_Return_mem(ret);
964 phi = new_Phi(n_ret, cf_pred, mode_M);
965 set_Tuple_pred(call, pn_Call_M_regular, phi);
966 /* Conserve Phi-list for further inlinings -- but might be optimized */
967 if (get_nodes_block(phi) == post_bl) {
968 set_irn_link(phi, get_irn_link(post_bl));
969 set_irn_link(post_bl, phi);
971 /* Now the real results */
973 for (j = 0; j < n_res; j++) {
975 for (i = 0; i < arity; i++) {
976 ret = get_irn_n(end_bl, i);
977 if (get_irn_op(ret) == op_Return) {
978 cf_pred[n_ret] = get_Return_res(ret, j);
983 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
987 /* Conserve Phi-list for further inlinings -- but might be optimized */
988 if (get_nodes_block(phi) == post_bl) {
989 set_irn_link(phi, get_irn_link(post_bl));
990 set_irn_link(post_bl, phi);
993 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
995 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
997 /* Finally the exception control flow.
998 We have two (three) possible situations:
999 First if the Call branches to an exception handler: We need to add a Phi node to
1000 collect the memory containing the exception objects. Further we need
1001 to add another block to get a correct representation of this Phi. To
1002 this block we add a Jmp that resolves into the X output of the Call
1003 when the Call is turned into a tuple.
1004 Second the Call branches to End, the exception is not handled. Just
1005 add all inlined exception branches to the End node.
1006 Third: there is no Exception edge at all. Handle as case two. */
1007 if (exc_handling == 0) {
1009 for (i = 0; i < arity; i++) {
1011 ret = get_irn_n(end_bl, i);
1012 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1013 cf_pred[n_exc] = ret;
1018 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1019 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1020 /* The Phi for the memories with the exception objects */
1022 for (i = 0; i < arity; i++) {
1024 ret = skip_Proj(get_irn_n(end_bl, i));
1025 if (get_irn_op(ret) == op_Call) {
1026 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1028 } else if (is_fragile_op(ret)) {
1029 /* We rely that all cfops have the memory output at the same position. */
1030 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1032 } else if (get_irn_op(ret) == op_Raise) {
1033 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1037 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1039 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1040 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1043 ir_node *main_end_bl;
1044 int main_end_bl_arity;
1045 ir_node **end_preds;
1047 /* assert(exc_handling == 1 || no exceptions. ) */
1049 for (i = 0; i < arity; i++) {
1050 ir_node *ret = get_irn_n(end_bl, i);
1052 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1053 cf_pred[n_exc] = ret;
1057 main_end_bl = get_irg_end_block(current_ir_graph);
1058 main_end_bl_arity = get_irn_arity(main_end_bl);
1059 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1061 for (i = 0; i < main_end_bl_arity; ++i)
1062 end_preds[i] = get_irn_n(main_end_bl, i);
1063 for (i = 0; i < n_exc; ++i)
1064 end_preds[main_end_bl_arity + i] = cf_pred[i];
1065 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1066 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1067 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1073 #if 0 /* old. now better, correcter, faster implementation. */
1075 /* -- If the exception control flow from the inlined Call directly
1076 branched to the end block we now have the following control
1077 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1078 remove the Jmp along with it's empty block and add Jmp's
1079 predecessors as predecessors of this end block. No problem if
1080 there is no exception, because then branches Bad to End which
1082 @@@ can't we know this beforehand: by getting the Proj(1) from
1083 the Call link list and checking whether it goes to Proj. */
1084 /* find the problematic predecessor of the end block. */
1085 end_bl = get_irg_end_block(current_ir_graph);
1086 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1087 cf_op = get_Block_cfgpred(end_bl, i);
1088 if (get_irn_op(cf_op) == op_Proj) {
1089 cf_op = get_Proj_pred(cf_op);
1090 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1091 /* There are unoptimized tuples from inlineing before when no exc */
1092 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1093 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1094 assert(get_irn_op(cf_op) == op_Jmp);
1100 if (i < get_Block_n_cfgpreds(end_bl)) {
1101 bl = get_nodes_block(cf_op);
1102 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1103 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1104 for (j = 0; j < i; j++)
1105 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1106 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1107 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1108 for (j = j; j < arity; j++)
1109 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1110 set_irn_in(end_bl, arity, cf_pred);
1112 /* Remove the exception pred from post-call Tuple. */
1113 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1118 /* -- Turn CSE back on. -- */
1119 set_optimize(rem_opt);
1124 /********************************************************************/
1125 /* Apply inlineing to small methods. */
1126 /********************************************************************/
1128 /* It makes no sense to inline too many calls in one procedure. Anyways,
1129 I didn't get a version with NEW_ARR_F to run. */
1130 #define MAX_INLINE 1024
1133 * environment for inlining small irgs
1135 typedef struct _inline_env_t {
1137 ir_node *calls[MAX_INLINE];
1141 * Returns the irg called from a Call node. If the irg is not
1142 * known, NULL is returned.
1144 static ir_graph *get_call_called_irg(ir_node *call) {
1146 ir_graph *called_irg = NULL;
1148 assert(get_irn_op(call) == op_Call);
1150 addr = get_Call_ptr(call);
1151 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1152 called_irg = get_entity_irg(get_SymConst_entity(addr));
1158 static void collect_calls(ir_node *call, void *env) {
1161 if (get_irn_op(call) != op_Call) return;
1163 addr = get_Call_ptr(call);
1165 if (get_irn_op(addr) == op_SymConst) {
1166 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1167 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1168 inline_env_t *ienv = (inline_env_t *)env;
1169 if (called_irg && ienv->pos < MAX_INLINE) {
1170 /* The Call node calls a locally defined method. Remember to inline. */
1171 ienv->calls[ienv->pos++] = call;
1178 * Inlines all small methods at call sites where the called address comes
1179 * from a Const node that references the entity representing the called
1181 * The size argument is a rough measure for the code size of the method:
1182 * Methods where the obstack containing the firm graph is smaller than
1185 void inline_small_irgs(ir_graph *irg, int size) {
1187 ir_graph *rem = current_ir_graph;
1188 inline_env_t env /* = {0, NULL}*/;
1190 if (!(get_opt_optimize() && get_opt_inline())) return;
1192 current_ir_graph = irg;
1193 /* Handle graph state */
1194 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1195 free_callee_info(current_ir_graph);
1197 /* Find Call nodes to inline.
1198 (We can not inline during a walk of the graph, as inlineing the same
1199 method several times changes the visited flag of the walked graph:
1200 after the first inlineing visited of the callee equals visited of
1201 the caller. With the next inlineing both are increased.) */
1203 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1205 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1206 /* There are calls to inline */
1207 collect_phiprojs(irg);
1208 for (i = 0; i < env.pos; i++) {
1210 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1211 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1212 (get_irg_inline_property(callee) == irg_inline_forced)) {
1213 inline_method(env.calls[i], callee);
1218 current_ir_graph = rem;
1222 * Environment for inlining irgs.
1225 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1226 int n_nodes_orig; /**< for statistics */
1227 eset *call_nodes; /**< All call nodes in this graph */
1229 int n_call_nodes_orig; /**< for statistics */
1230 int n_callers; /**< Number of known graphs that call this graphs. */
1231 int n_callers_orig; /**< for statistics */
1235 * Allocate a new environment for inlining.
1237 static inline_irg_env *new_inline_irg_env(void) {
1238 inline_irg_env *env = xmalloc(sizeof(*env));
1239 env->n_nodes = -2; /* do not count count Start, End */
1240 env->n_nodes_orig = -2; /* do not count Start, End */
1241 env->call_nodes = eset_create();
1242 env->n_call_nodes = 0;
1243 env->n_call_nodes_orig = 0;
1245 env->n_callers_orig = 0;
1250 * destroy an environment for inlining.
1252 static void free_inline_irg_env(inline_irg_env *env) {
1253 eset_destroy(env->call_nodes);
1258 * post-walker: collect all calls in the inline-environment
1259 * of a graph and sum some statistics.
1261 static void collect_calls2(ir_node *call, void *env) {
1262 inline_irg_env *x = (inline_irg_env *)env;
1263 ir_op *op = get_irn_op(call);
1266 /* count meaningful nodes in irg */
1267 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1272 if (op != op_Call) return;
1274 /* collect all call nodes */
1275 eset_insert(x->call_nodes, call);
1277 x->n_call_nodes_orig++;
1279 /* count all static callers */
1280 callee = get_call_called_irg(call);
1282 inline_irg_env *callee_env = get_irg_link(callee);
1283 callee_env->n_callers++;
1284 callee_env->n_callers_orig++;
1289 * Returns TRUE if the number of callers in 0 in the irg's environment,
1290 * hence this irg is a leave.
1292 INLINE static int is_leave(ir_graph *irg) {
1293 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1297 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1299 INLINE static int is_smaller(ir_graph *callee, int size) {
1300 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1305 * Inlines small leave methods at call sites where the called address comes
1306 * from a Const node that references the entity representing the called
1308 * The size argument is a rough measure for the code size of the method:
1309 * Methods where the obstack containing the firm graph is smaller than
1312 void inline_leave_functions(int maxsize, int leavesize, int size) {
1313 inline_irg_env *env;
1314 int i, n_irgs = get_irp_n_irgs();
1315 ir_graph *rem = current_ir_graph;
1318 if (!(get_opt_optimize() && get_opt_inline())) return;
1320 /* extend all irgs by a temporary data structure for inlining. */
1321 for (i = 0; i < n_irgs; ++i)
1322 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1324 /* Precompute information in temporary data structure. */
1325 for (i = 0; i < n_irgs; ++i) {
1326 current_ir_graph = get_irp_irg(i);
1327 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1328 free_callee_info(current_ir_graph);
1330 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1331 get_irg_link(current_ir_graph));
1334 /* -- and now inline. -- */
1336 /* Inline leaves recursively -- we might construct new leaves. */
1337 while (did_inline) {
1340 for (i = 0; i < n_irgs; ++i) {
1342 int phiproj_computed = 0;
1344 current_ir_graph = get_irp_irg(i);
1345 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1347 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1350 if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
1351 callee = get_call_called_irg(call);
1353 if (env->n_nodes > maxsize) continue; // break;
1355 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1356 if (!phiproj_computed) {
1357 phiproj_computed = 1;
1358 collect_phiprojs(current_ir_graph);
1360 did_inline = inline_method(call, callee);
1363 /* Do some statistics */
1364 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1365 env->n_call_nodes --;
1366 env->n_nodes += callee_env->n_nodes;
1367 callee_env->n_callers--;
1374 /* inline other small functions. */
1375 for (i = 0; i < n_irgs; ++i) {
1378 int phiproj_computed = 0;
1380 current_ir_graph = get_irp_irg(i);
1381 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1383 /* we can not walk and change a set, nor remove from it.
1385 walkset = env->call_nodes;
1386 env->call_nodes = eset_create();
1387 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1390 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1391 callee = get_call_called_irg(call);
1394 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1395 (get_irg_inline_property(callee) == irg_inline_forced))) {
1396 if (!phiproj_computed) {
1397 phiproj_computed = 1;
1398 collect_phiprojs(current_ir_graph);
1400 if (inline_method(call, callee)) {
1401 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1402 env->n_call_nodes--;
1403 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1404 env->n_call_nodes += callee_env->n_call_nodes;
1405 env->n_nodes += callee_env->n_nodes;
1406 callee_env->n_callers--;
1409 eset_insert(env->call_nodes, call);
1412 eset_destroy(walkset);
1415 for (i = 0; i < n_irgs; ++i) {
1416 current_ir_graph = get_irp_irg(i);
1418 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1419 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1420 (env->n_callers_orig != env->n_callers))
1421 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1422 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1423 env->n_callers_orig, env->n_callers,
1424 get_entity_name(get_irg_entity(current_ir_graph)));
1426 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1429 current_ir_graph = rem;
1432 /*******************************************************************/
1433 /* Code Placement. Pins all floating nodes to a block where they */
1434 /* will be executed only if needed. */
1435 /*******************************************************************/
1438 * Returns non-zero, is a block is not reachable from Start.
1440 * @param block the block to test
1443 is_Block_unreachable(ir_node *block) {
1444 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1448 * Find the earliest correct block for N. --- Place N into the
1449 * same Block as its dominance-deepest Input.
1451 * We have to avoid calls to get_nodes_block() here
1452 * because the graph is floating.
1454 * move_out_of_loops() expects that place_floats_early() have placed
1455 * all "living" nodes into a living block. That's why we must
1456 * move nodes in dead block with "live" successors into a valid
1458 * We move them just into the same block as it's successor (or
1459 * in case of a Phi into the effective use block). For Phi successors,
1460 * this may still be a dead block, but then there is no real use, as
1461 * the control flow will be dead later.
1464 place_floats_early(ir_node *n, pdeq *worklist)
1468 /* we must not run into an infinite loop */
1469 assert(irn_not_visited(n));
1470 mark_irn_visited(n);
1472 /* Place floating nodes. */
1473 if (get_irn_pinned(n) == op_pin_state_floats) {
1474 ir_node *curr_block = get_irn_n(n, -1);
1475 int in_dead_block = is_Block_unreachable(curr_block);
1477 ir_node *b = NULL; /* The block to place this node in */
1479 assert(get_irn_op(n) != op_Block);
1481 if ((get_irn_op(n) == op_Const) ||
1482 (get_irn_op(n) == op_SymConst) ||
1484 (get_irn_op(n) == op_Unknown)) {
1485 /* These nodes will not be placed by the loop below. */
1486 b = get_irg_start_block(current_ir_graph);
1490 /* find the block for this node. */
1491 irn_arity = get_irn_arity(n);
1492 for (i = 0; i < irn_arity; i++) {
1493 ir_node *pred = get_irn_n(n, i);
1494 ir_node *pred_block;
1496 if ((irn_not_visited(pred))
1497 && (get_irn_pinned(pred) == op_pin_state_floats)) {
1500 * If the current node is NOT in a dead block, but one of its
1501 * predecessors is, we must move the predecessor to a live block.
1502 * Such thing can happen, if global CSE chose a node from a dead block.
1503 * We move it simple to our block.
1504 * Note that neither Phi nor End nodes are floating, so we don't
1505 * need to handle them here.
1507 if (! in_dead_block) {
1508 if (get_irn_pinned(pred) == op_pin_state_floats &&
1509 is_Block_unreachable(get_irn_n(pred, -1)))
1510 set_nodes_block(pred, curr_block);
1512 place_floats_early(pred, worklist);
1516 * A node in the Bad block must stay in the bad block,
1517 * so don't compute a new block for it.
1522 /* Because all loops contain at least one op_pin_state_pinned node, now all
1523 our inputs are either op_pin_state_pinned or place_early() has already
1524 been finished on them. We do not have any unfinished inputs! */
1525 pred_block = get_irn_n(pred, -1);
1526 if ((!is_Block_dead(pred_block)) &&
1527 (get_Block_dom_depth(pred_block) > depth)) {
1529 depth = get_Block_dom_depth(pred_block);
1531 /* Avoid that the node is placed in the Start block */
1532 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1533 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1534 assert(b != get_irg_start_block(current_ir_graph));
1539 set_nodes_block(n, b);
1543 * Add predecessors of non floating nodes and non-floating predecessors
1544 * of floating nodes to worklist and fix their blocks if the are in dead block.
1546 irn_arity = get_irn_arity(n);
1548 if (get_irn_op(n) == op_End) {
1550 * Simplest case: End node. Predecessors are keep-alives,
1551 * no need to move out of dead block.
1553 for (i = -1; i < irn_arity; ++i) {
1554 ir_node *pred = get_irn_n(n, i);
1555 if (irn_not_visited(pred))
1556 pdeq_putr(worklist, pred);
1559 else if (is_Block(n)) {
1561 * Blocks: Predecessors are control flow, no need to move
1562 * them out of dead block.
1564 for (i = irn_arity - 1; i >= 0; --i) {
1565 ir_node *pred = get_irn_n(n, i);
1566 if (irn_not_visited(pred))
1567 pdeq_putr(worklist, pred);
1570 else if (is_Phi(n)) {
1572 ir_node *curr_block = get_irn_n(n, -1);
1573 int in_dead_block = is_Block_unreachable(curr_block);
1576 * Phi nodes: move nodes from dead blocks into the effective use
1577 * of the Phi-input if the Phi is not in a bad block.
1579 pred = get_irn_n(n, -1);
1580 if (irn_not_visited(pred))
1581 pdeq_putr(worklist, pred);
1583 for (i = irn_arity - 1; i >= 0; --i) {
1584 ir_node *pred = get_irn_n(n, i);
1586 if (irn_not_visited(pred)) {
1587 if (! in_dead_block &&
1588 get_irn_pinned(pred) == op_pin_state_floats &&
1589 is_Block_unreachable(get_irn_n(pred, -1))) {
1590 set_nodes_block(pred, get_Block_cfgpred_block(curr_block, i));
1592 pdeq_putr(worklist, pred);
1598 ir_node *curr_block = get_irn_n(n, -1);
1599 int in_dead_block = is_Block_unreachable(curr_block);
1602 * All other nodes: move nodes from dead blocks into the same block.
1604 pred = get_irn_n(n, -1);
1605 if (irn_not_visited(pred))
1606 pdeq_putr(worklist, pred);
1608 for (i = irn_arity - 1; i >= 0; --i) {
1609 ir_node *pred = get_irn_n(n, i);
1611 if (irn_not_visited(pred)) {
1612 if (! in_dead_block &&
1613 get_irn_pinned(pred) == op_pin_state_floats &&
1614 is_Block_unreachable(get_irn_n(pred, -1))) {
1615 set_nodes_block(pred, curr_block);
1617 pdeq_putr(worklist, pred);
1624 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1625 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1626 * places all floating nodes reachable from its argument through floating
1627 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1629 static INLINE void place_early(pdeq *worklist) {
1631 inc_irg_visited(current_ir_graph);
1633 /* this inits the worklist */
1634 place_floats_early(get_irg_end(current_ir_graph), worklist);
1636 /* Work the content of the worklist. */
1637 while (!pdeq_empty(worklist)) {
1638 ir_node *n = pdeq_getl(worklist);
1639 if (irn_not_visited(n))
1640 place_floats_early(n, worklist);
1643 set_irg_outs_inconsistent(current_ir_graph);
1644 set_irg_pinned(current_ir_graph, op_pin_state_pinned);
1648 * Compute the deepest common ancestor of block and dca.
1650 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1654 /* we do not want to place nodes in dead blocks */
1655 if (is_Block_dead(block))
1658 /* We found a first legal placement. */
1659 if (!dca) return block;
1661 /* Find a placement that is dominates both, dca and block. */
1662 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1663 block = get_Block_idom(block);
1665 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1666 dca = get_Block_idom(dca);
1669 while (block != dca)
1670 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1675 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1676 * I.e., DCA is the block where we might place PRODUCER.
1677 * A data flow edge points from producer to consumer.
1680 consumer_dom_dca(ir_node *dca, ir_node *consumer, ir_node *producer)
1682 ir_node *block = NULL;
1684 /* Compute the latest block into which we can place a node so that it is
1686 if (get_irn_op(consumer) == op_Phi) {
1687 /* our consumer is a Phi-node, the effective use is in all those
1688 blocks through which the Phi-node reaches producer */
1690 ir_node *phi_block = get_nodes_block(consumer);
1691 irn_arity = get_irn_arity(consumer);
1693 for (i = 0; i < irn_arity; i++) {
1694 if (get_irn_n(consumer, i) == producer) {
1695 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1697 if (! is_Block_unreachable(new_block))
1698 block = calc_dca(block, new_block);
1703 block = get_irn_n(producer, -1);
1706 assert(is_no_Block(consumer));
1707 block = get_nodes_block(consumer);
1710 /* Compute the deepest common ancestor of block and dca. */
1711 return calc_dca(dca, block);
1714 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1716 static INLINE int get_irn_loop_depth(ir_node *n) {
1717 return get_loop_depth(get_irn_loop(n));
1721 * Move n to a block with less loop depth than it's current block. The
1722 * new block must be dominated by early.
1724 * @param n the node that should be moved
1725 * @param early the earliest block we can n move to
1728 move_out_of_loops (ir_node *n, ir_node *early)
1730 ir_node *best, *dca;
1734 /* Find the region deepest in the dominator tree dominating
1735 dca with the least loop nesting depth, but still dominated
1736 by our early placement. */
1737 dca = get_nodes_block(n);
1740 while (dca != early) {
1741 dca = get_Block_idom(dca);
1742 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1743 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1747 if (best != get_nodes_block(n)) {
1749 printf("Moving out of loop: "); DDMN(n);
1750 printf(" Outermost block: "); DDMN(early);
1751 printf(" Best block: "); DDMN(best);
1752 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1754 set_nodes_block(n, best);
1759 * Find the latest legal block for N and place N into the
1760 * `optimal' Block between the latest and earliest legal block.
1761 * The `optimal' block is the dominance-deepest block of those
1762 * with the least loop-nesting-depth. This places N out of as many
1763 * loops as possible and then makes it as control dependent as
1767 place_floats_late(ir_node *n, pdeq *worklist)
1772 assert(irn_not_visited(n)); /* no multiple placement */
1774 mark_irn_visited(n);
1776 /* no need to place block nodes, control nodes are already placed. */
1777 if ((get_irn_op(n) != op_Block) &&
1779 (get_irn_mode(n) != mode_X)) {
1780 /* Remember the early_blk placement of this block to move it
1781 out of loop no further than the early_blk placement. */
1782 early_blk = get_irn_n(n, -1);
1785 * BEWARE: Here we also get code, that is live, but
1786 * was in a dead block. If the node is life, but because
1787 * of CSE in a dead block, we still might need it.
1790 /* Assure that our users are all placed, except the Phi-nodes.
1791 --- Each data flow cycle contains at least one Phi-node. We
1792 have to break the `user has to be placed before the
1793 producer' dependence cycle and the Phi-nodes are the
1794 place to do so, because we need to base our placement on the
1795 final region of our users, which is OK with Phi-nodes, as they
1796 are op_pin_state_pinned, and they never have to be placed after a
1797 producer of one of their inputs in the same block anyway. */
1798 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1799 ir_node *succ = get_irn_out(n, i);
1800 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1801 place_floats_late(succ, worklist);
1804 if (! is_Block_dead(early_blk)) {
1805 /* do only move things that where not dead */
1807 /* We have to determine the final block of this node... except for
1809 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1810 (get_irn_op(n) != op_Const) &&
1811 (get_irn_op(n) != op_SymConst)) {
1812 ir_node *dca = NULL; /* deepest common ancestor in the
1813 dominator tree of all nodes'
1814 blocks depending on us; our final
1815 placement has to dominate DCA. */
1816 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1817 ir_node *succ = get_irn_out(n, i);
1820 if (get_irn_op(succ) == op_End) {
1822 * This consumer is the End node, a keep alive edge.
1823 * This is not a real consumer, so we ignore it
1828 /* ignore if succ is in dead code */
1829 succ_blk = get_irn_n(succ, -1);
1830 if (is_Block_unreachable(succ_blk))
1832 dca = consumer_dom_dca(dca, succ, n);
1835 set_nodes_block(n, dca);
1836 move_out_of_loops(n, early_blk);
1842 /* Add predecessors of all non-floating nodes on list. (Those of floating
1843 nodes are placed already and therefore are marked.) */
1844 for (i = 0; i < get_irn_n_outs(n); i++) {
1845 ir_node *succ = get_irn_out(n, i);
1846 if (irn_not_visited(get_irn_out(n, i))) {
1847 pdeq_putr(worklist, succ);
1852 static INLINE void place_late(pdeq *worklist) {
1854 inc_irg_visited(current_ir_graph);
1856 /* This fills the worklist initially. */
1857 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1859 /* And now empty the worklist again... */
1860 while (!pdeq_empty(worklist)) {
1861 ir_node *n = pdeq_getl(worklist);
1862 if (irn_not_visited(n))
1863 place_floats_late(n, worklist);
1867 void place_code(ir_graph *irg) {
1869 ir_graph *rem = current_ir_graph;
1871 current_ir_graph = irg;
1873 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1875 /* Handle graph state */
1876 assert(get_irg_phase_state(irg) != phase_building);
1877 if (get_irg_dom_state(irg) != dom_consistent)
1880 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1881 free_loop_information(irg);
1882 construct_backedges(irg);
1885 /* Place all floating nodes as early as possible. This guarantees
1886 a legal code placement. */
1887 worklist = new_pdeq();
1888 place_early(worklist);
1890 /* place_early() invalidates the outs, place_late needs them. */
1891 compute_irg_outs(irg);
1893 /* Now move the nodes down in the dominator tree. This reduces the
1894 unnecessary executions of the node. */
1895 place_late(worklist);
1897 set_irg_outs_inconsistent(current_ir_graph);
1898 set_irg_loopinfo_inconsistent(current_ir_graph);
1900 current_ir_graph = rem;
1904 * Called by walker of remove_critical_cf_edges().
1906 * Place an empty block to an edge between a blocks of multiple
1907 * predecessors and a block of multiple successors.
1910 * @param env Environment of walker. This field is unused and has
1913 static void walk_critical_cf_edges(ir_node *n, void *env) {
1915 ir_node *pre, *block, **in, *jmp;
1917 /* Block has multiple predecessors */
1918 if ((op_Block == get_irn_op(n)) &&
1919 (get_irn_arity(n) > 1)) {
1920 arity = get_irn_arity(n);
1922 if (n == get_irg_end_block(current_ir_graph))
1923 return; /* No use to add a block here. */
1925 for (i=0; i<arity; i++) {
1926 pre = get_irn_n(n, i);
1927 /* Predecessor has multiple successors. Insert new flow edge */
1928 if ((NULL != pre) &&
1929 (op_Proj == get_irn_op(pre)) &&
1930 op_Raise != get_irn_op(skip_Proj(pre))) {
1932 /* set predecessor array for new block */
1933 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1934 /* set predecessor of new block */
1936 block = new_Block(1, in);
1937 /* insert new jmp node to new block */
1938 set_cur_block(block);
1941 /* set successor of new block */
1942 set_irn_n(n, i, jmp);
1944 } /* predecessor has multiple successors */
1945 } /* for all predecessors */
1946 } /* n is a block */
1949 void remove_critical_cf_edges(ir_graph *irg) {
1950 if (get_opt_critical_edges())
1951 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);