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"
46 /* Defined in iropt.c */
47 pset *new_identities (void);
48 void del_identities (pset *value_table);
49 void add_identities (pset *value_table, ir_node *node);
51 /*------------------------------------------------------------------*/
52 /* apply optimizations of iropt to all nodes. */
53 /*------------------------------------------------------------------*/
55 static void init_link (ir_node *n, void *env) {
56 set_irn_link(n, NULL);
59 #if 0 /* Old version. Avoids Ids.
60 This is not necessary: we do a postwalk, and get_irn_n
61 removes ids anyways. So it's much cheaper to call the
62 optimization less often and use the exchange() algorithm. */
64 optimize_in_place_wrapper (ir_node *n, void *env) {
66 ir_node *optimized, *old;
68 irn_arity = get_irn_arity(n);
69 for (i = 0; i < irn_arity; i++) {
70 /* get_irn_n skips Id nodes, so comparison old != optimized does not
71 show all optimizations. Therefore always set new predecessor. */
72 old = get_irn_intra_n(n, i);
73 optimized = optimize_in_place_2(old);
74 set_irn_n(n, i, optimized);
77 if (get_irn_op(n) == op_Block) {
78 optimized = optimize_in_place_2(n);
79 if (optimized != n) exchange (n, optimized);
84 optimize_in_place_wrapper (ir_node *n, void *env) {
85 ir_node *optimized = optimize_in_place_2(n);
86 if (optimized != n) exchange (n, optimized);
91 static INLINE void do_local_optimize(ir_node *n) {
92 /* Handle graph state */
93 assert(get_irg_phase_state(current_ir_graph) != phase_building);
94 if (get_opt_global_cse())
95 set_irg_pinned(current_ir_graph, op_pin_state_floats);
96 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
97 set_irg_outs_inconsistent(current_ir_graph);
98 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
99 set_irg_dom_inconsistent(current_ir_graph);
100 set_irg_loopinfo_inconsistent(current_ir_graph);
103 /* Clean the value_table in irg for the cse. */
104 del_identities(current_ir_graph->value_table);
105 current_ir_graph->value_table = new_identities();
107 /* walk over the graph */
108 irg_walk(n, init_link, optimize_in_place_wrapper, NULL);
111 void local_optimize_node(ir_node *n) {
112 ir_graph *rem = current_ir_graph;
113 current_ir_graph = get_irn_irg(n);
115 do_local_optimize(n);
117 current_ir_graph = rem;
121 local_optimize_graph (ir_graph *irg) {
122 ir_graph *rem = current_ir_graph;
123 current_ir_graph = irg;
125 do_local_optimize(irg->end);
127 current_ir_graph = rem;
131 /*------------------------------------------------------------------*/
132 /* Routines for dead node elimination / copying garbage collection */
133 /* of the obstack. */
134 /*------------------------------------------------------------------*/
137 * Remember the new node in the old node by using a field all nodes have.
140 set_new_node (ir_node *old, ir_node *new)
146 * Get this new node, before the old node is forgotton.
148 static INLINE ir_node *
149 get_new_node (ir_node * n)
155 * We use the block_visited flag to mark that we have computed the
156 * number of useful predecessors for this block.
157 * Further we encode the new arity in this flag in the old blocks.
158 * Remembering the arity is useful, as it saves a lot of pointer
159 * accesses. This function is called for all Phi and Block nodes
163 compute_new_arity(ir_node *b) {
164 int i, res, irn_arity;
167 irg_v = get_irg_block_visited(current_ir_graph);
168 block_v = get_Block_block_visited(b);
169 if (block_v >= irg_v) {
170 /* we computed the number of preds for this block and saved it in the
172 return block_v - irg_v;
174 /* compute the number of good predecessors */
175 res = irn_arity = get_irn_arity(b);
176 for (i = 0; i < irn_arity; i++)
177 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
178 /* save it in the flag. */
179 set_Block_block_visited(b, irg_v + res);
184 /* TODO: add an ir_op operation */
185 static INLINE void new_backedge_info(ir_node *n) {
186 switch(get_irn_opcode(n)) {
188 n->attr.block.cg_backedge = NULL;
189 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
192 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
195 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
202 * Copies the node to the new obstack. The Ins of the new node point to
203 * the predecessors on the old obstack. For block/phi nodes not all
204 * predecessors might be copied. n->link points to the new node.
205 * For Phi and Block nodes the function allocates in-arrays with an arity
206 * only for useful predecessors. The arity is determined by counting
207 * the non-bad predecessors of the block.
209 * @param n The node to be copied
210 * @param env if non-NULL, the node number attribute will be copied to the new node
212 * Note: Also used for loop unrolling.
215 firm_copy_node (ir_node *n, void *env) {
218 opcode op = get_irn_opcode(n);
219 int copy_node_nr = env != NULL;
221 /* The end node looses it's flexible in array. This doesn't matter,
222 as dead node elimination builds End by hand, inlineing doesn't use
224 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
227 /* node copied already */
229 } else if (op == iro_Block) {
231 new_arity = compute_new_arity(n);
232 n->attr.block.graph_arr = NULL;
234 block = get_nodes_block(n);
235 if (get_irn_opcode(n) == iro_Phi) {
236 new_arity = compute_new_arity(block);
238 new_arity = get_irn_arity(n);
241 nn = new_ir_node(get_irn_dbg_info(n),
248 /* Copy the attributes. These might point to additional data. If this
249 was allocated on the old obstack the pointers now are dangling. This
250 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
251 copy_node_attr(n, nn);
252 new_backedge_info(nn);
257 /* for easier debugging, we want to copy the node numbers too */
258 nn->node_nr = n->node_nr;
262 /* printf("\n old node: "); DDMSG2(n);
263 printf(" new node: "); DDMSG2(nn); */
267 * Copies new predecessors of old node to new node remembered in link.
268 * Spare the Bad predecessors of Phi and Block nodes.
271 copy_preds (ir_node *n, void *env) {
275 nn = get_new_node(n);
277 /* printf("\n old node: "); DDMSG2(n);
278 printf(" new node: "); DDMSG2(nn);
279 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
281 if (get_irn_opcode(n) == iro_Block) {
282 /* Don't copy Bad nodes. */
284 irn_arity = get_irn_arity(n);
285 for (i = 0; i < irn_arity; i++)
286 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
287 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
288 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
291 /* repair the block visited flag from above misuse. Repair it in both
292 graphs so that the old one can still be used. */
293 set_Block_block_visited(nn, 0);
294 set_Block_block_visited(n, 0);
295 /* Local optimization could not merge two subsequent blocks if
296 in array contained Bads. Now it's possible.
297 We don't call optimize_in_place as it requires
298 that the fields in ir_graph are set properly. */
299 if ((get_opt_control_flow_straightening()) &&
300 (get_Block_n_cfgpreds(nn) == 1) &&
301 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
302 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
304 /* Jmp jumps into the block it is in -- deal self cycle. */
305 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
306 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
311 } else if (get_irn_opcode(n) == iro_Phi) {
312 /* Don't copy node if corresponding predecessor in block is Bad.
313 The Block itself should not be Bad. */
314 block = get_nodes_block(n);
315 set_irn_n (nn, -1, get_new_node(block));
317 irn_arity = get_irn_arity(n);
318 for (i = 0; i < irn_arity; i++)
319 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
320 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
321 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
324 /* If the pre walker reached this Phi after the post walker visited the
325 block block_visited is > 0. */
326 set_Block_block_visited(get_nodes_block(n), 0);
327 /* Compacting the Phi's ins might generate Phis with only one
329 if (get_irn_arity(n) == 1)
330 exchange(n, get_irn_n(n, 0));
332 irn_arity = get_irn_arity(n);
333 for (i = -1; i < irn_arity; i++)
334 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
336 /* Now the new node is complete. We can add it to the hash table for cse.
337 @@@ inlinening aborts if we identify End. Why? */
338 if(get_irn_op(nn) != op_End)
339 add_identities (current_ir_graph->value_table, nn);
343 * Copies the graph recursively, compacts the keepalive of the end node.
345 * @param copy_node_nr If non-zero, the node number will be copied
348 copy_graph (int copy_node_nr) {
349 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
350 ir_node *ka; /* keep alive */
353 oe = get_irg_end(current_ir_graph);
354 /* copy the end node by hand, allocate dynamic in array! */
355 ne = new_ir_node(get_irn_dbg_info(oe),
362 /* Copy the attributes. Well, there might be some in the future... */
363 copy_node_attr(oe, ne);
364 set_new_node(oe, ne);
366 /* copy the Bad node */
367 ob = get_irg_bad(current_ir_graph);
368 nb = new_ir_node(get_irn_dbg_info(ob),
375 set_new_node(ob, nb);
377 /* copy the NoMem node */
378 om = get_irg_no_mem(current_ir_graph);
379 nm = new_ir_node(get_irn_dbg_info(om),
386 set_new_node(om, nm);
388 /* copy the live nodes */
389 irg_walk(get_nodes_block(oe), firm_copy_node, copy_preds, (void *)copy_node_nr);
390 /* copy_preds for the end node ... */
391 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
393 /*- ... and now the keep alives. -*/
394 /* First pick the not marked block nodes and walk them. We must pick these
395 first as else we will oversee blocks reachable from Phis. */
396 irn_arity = get_irn_arity(oe);
397 for (i = 0; i < irn_arity; i++) {
398 ka = get_irn_intra_n(oe, i);
399 if ((get_irn_op(ka) == op_Block) &&
400 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
401 /* We must keep the block alive and copy everything reachable */
402 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
403 irg_walk(ka, firm_copy_node, copy_preds, (void *)copy_node_nr);
404 add_End_keepalive(ne, get_new_node(ka));
408 /* Now pick the Phis. Here we will keep all! */
409 irn_arity = get_irn_arity(oe);
410 for (i = 0; i < irn_arity; i++) {
411 ka = get_irn_intra_n(oe, i);
412 if ((get_irn_op(ka) == op_Phi)) {
413 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
414 /* We didn't copy the Phi yet. */
415 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
416 irg_walk(ka, firm_copy_node, copy_preds, (void *)copy_node_nr);
418 add_End_keepalive(ne, get_new_node(ka));
422 /* start block sometimes only reached after keep alives */
423 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
424 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
428 * Copies the graph reachable from current_ir_graph->end to the obstack
429 * in current_ir_graph and fixes the environment.
430 * Then fixes the fields in current_ir_graph containing nodes of the
433 * @param copy_node_nr If non-zero, the node number will be copied
436 copy_graph_env (int copy_node_nr) {
438 /* Not all nodes remembered in current_ir_graph might be reachable
439 from the end node. Assure their link is set to NULL, so that
440 we can test whether new nodes have been computed. */
441 set_irn_link(get_irg_frame (current_ir_graph), NULL);
442 set_irn_link(get_irg_globals (current_ir_graph), NULL);
443 set_irn_link(get_irg_args (current_ir_graph), NULL);
444 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
445 set_irn_link(get_irg_no_mem (current_ir_graph), NULL);
447 /* we use the block walk flag for removing Bads from Blocks ins. */
448 inc_irg_block_visited(current_ir_graph);
451 copy_graph(copy_node_nr);
453 /* fix the fields in current_ir_graph */
454 old_end = get_irg_end(current_ir_graph);
455 set_irg_end (current_ir_graph, get_new_node(old_end));
456 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
457 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
459 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
460 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
461 firm_copy_node (get_irg_frame(current_ir_graph), (void *)copy_node_nr);
462 copy_preds(get_irg_frame(current_ir_graph), NULL);
464 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
465 firm_copy_node (get_irg_globals(current_ir_graph), (void *)copy_node_nr);
466 copy_preds(get_irg_globals(current_ir_graph), NULL);
468 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
469 firm_copy_node (get_irg_initial_mem(current_ir_graph), (void *)copy_node_nr);
470 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
472 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
473 firm_copy_node (get_irg_args(current_ir_graph), (void *)copy_node_nr);
474 copy_preds(get_irg_args(current_ir_graph), NULL);
476 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
478 set_irg_start_block(current_ir_graph,
479 get_new_node(get_irg_start_block(current_ir_graph)));
480 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
481 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
482 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
483 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
485 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
486 firm_copy_node(get_irg_bad(current_ir_graph), (void *)copy_node_nr);
487 copy_preds(get_irg_bad(current_ir_graph), NULL);
489 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
491 if (get_irn_link(get_irg_no_mem(current_ir_graph)) == NULL) {
492 firm_copy_node(get_irg_no_mem(current_ir_graph), (void *)copy_node_nr);
493 copy_preds(get_irg_no_mem(current_ir_graph), NULL);
495 set_irg_no_mem(current_ir_graph, get_new_node(get_irg_no_mem(current_ir_graph)));
499 * Copies all reachable nodes to a new obstack. Removes bad inputs
500 * from block nodes and the corresponding inputs from Phi nodes.
501 * Merges single exit blocks with single entry blocks and removes
503 * Adds all new nodes to a new hash table for cse. Does not
504 * perform cse, so the hash table might contain common subexpressions.
507 dead_node_elimination(ir_graph *irg) {
509 int rem_ipview = get_interprocedural_view();
510 struct obstack *graveyard_obst = NULL;
511 struct obstack *rebirth_obst = NULL;
513 edges_init_graph(irg);
515 /* inform statistics that we started a dead-node elimination run */
516 hook_dead_node_elim_start(irg);
518 /* Remember external state of current_ir_graph. */
519 rem = current_ir_graph;
520 current_ir_graph = irg;
521 set_interprocedural_view(false);
523 /* Handle graph state */
524 assert(get_irg_phase_state(current_ir_graph) != phase_building);
525 free_callee_info(current_ir_graph);
526 free_outs(current_ir_graph);
528 /* @@@ so far we loose loops when copying */
529 free_loop_information(current_ir_graph);
531 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
533 /* A quiet place, where the old obstack can rest in peace,
534 until it will be cremated. */
535 graveyard_obst = irg->obst;
537 /* A new obstack, where the reachable nodes will be copied to. */
538 rebirth_obst = xmalloc (sizeof(*rebirth_obst));
539 current_ir_graph->obst = rebirth_obst;
540 obstack_init (current_ir_graph->obst);
542 /* We also need a new hash table for cse */
543 del_identities (irg->value_table);
544 irg->value_table = new_identities ();
546 /* Copy the graph from the old to the new obstack */
549 /* Free memory from old unoptimized obstack */
550 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
551 xfree (graveyard_obst); /* ... then free it. */
554 /* inform statistics that the run is over */
555 hook_dead_node_elim_stop(irg);
557 current_ir_graph = rem;
558 set_interprocedural_view(rem_ipview);
562 * Relink bad predeseccors of a block and store the old in array to the
563 * link field. This function is called by relink_bad_predecessors().
564 * The array of link field starts with the block operand at position 0.
565 * If block has bad predecessors, create a new in array without bad preds.
566 * Otherwise let in array untouched.
568 static void relink_bad_block_predecessors(ir_node *n, void *env) {
569 ir_node **new_in, *irn;
570 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
572 /* if link field of block is NULL, look for bad predecessors otherwise
573 this is allready done */
574 if (get_irn_op(n) == op_Block &&
575 get_irn_link(n) == NULL) {
577 /* save old predecessors in link field (position 0 is the block operand)*/
578 set_irn_link(n, (void *)get_irn_in(n));
580 /* count predecessors without bad nodes */
581 old_irn_arity = get_irn_arity(n);
582 for (i = 0; i < old_irn_arity; i++)
583 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
585 /* arity changing: set new predecessors without bad nodes */
586 if (new_irn_arity < old_irn_arity) {
587 /* Get new predecessor array. We do not resize the array, as we must
588 keep the old one to update Phis. */
589 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
591 /* set new predeseccors in array */
594 for (i = 0; i < old_irn_arity; i++) {
595 irn = get_irn_n(n, i);
597 new_in[new_irn_n] = irn;
598 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
602 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
603 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
606 } /* ir node has bad predecessors */
608 } /* Block is not relinked */
612 * Relinks Bad predecesors from Bocks and Phis called by walker
613 * remove_bad_predecesors(). If n is a Block, call
614 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
615 * function of Phi's Block. If this block has bad predecessors, relink preds
618 static void relink_bad_predecessors(ir_node *n, void *env) {
619 ir_node *block, **old_in;
620 int i, old_irn_arity, new_irn_arity;
622 /* relink bad predeseccors of a block */
623 if (get_irn_op(n) == op_Block)
624 relink_bad_block_predecessors(n, env);
626 /* If Phi node relink its block and its predecessors */
627 if (get_irn_op(n) == op_Phi) {
629 /* Relink predeseccors of phi's block */
630 block = get_nodes_block(n);
631 if (get_irn_link(block) == NULL)
632 relink_bad_block_predecessors(block, env);
634 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
635 old_irn_arity = ARR_LEN(old_in);
637 /* Relink Phi predeseccors if count of predeseccors changed */
638 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
639 /* set new predeseccors in array
640 n->in[0] remains the same block */
642 for(i = 1; i < old_irn_arity; i++)
643 if (!is_Bad((ir_node *)old_in[i])) {
644 n->in[new_irn_arity] = n->in[i];
645 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
649 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
650 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
653 } /* n is a Phi node */
657 * Removes Bad Bad predecesors from Blocks and the corresponding
658 * inputs to Phi nodes as in dead_node_elimination but without
660 * On walking up set the link field to NULL, on walking down call
661 * relink_bad_predecessors() (This function stores the old in array
662 * to the link field and sets a new in array if arity of predecessors
665 void remove_bad_predecessors(ir_graph *irg) {
666 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
670 /*--------------------------------------------------------------------*/
671 /* Funcionality for inlining */
672 /*--------------------------------------------------------------------*/
675 * Copy node for inlineing. Updates attributes that change when
676 * inlineing but not for dead node elimination.
678 * Copies the node by calling firm_copy_node and then updates the entity if
679 * it's a local one. env must be a pointer of the frame type of the
680 * inlined procedure. The new entities must be in the link field of
684 copy_node_inline (ir_node *n, void *env) {
686 type *frame_tp = (type *)env;
688 firm_copy_node(n, NULL);
689 if (get_irn_op(n) == op_Sel) {
690 new = get_new_node (n);
691 assert(get_irn_op(new) == op_Sel);
692 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
693 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
695 } else if (get_irn_op(n) == op_Block) {
696 new = get_new_node (n);
697 new->attr.block.irg = current_ir_graph;
701 static void find_addr(ir_node *node, void *env)
703 if (get_irn_opcode(node) == iro_Proj) {
704 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
710 * currently, we cannot inline two cases:
711 * - call with compound arguments
712 * - graphs that take the address of a parameter
714 * check these conditions here
716 static int can_inline(ir_node *call, ir_graph *called_graph)
718 type *call_type = get_Call_type(call);
719 int params, ress, i, res;
720 assert(is_Method_type(call_type));
722 params = get_method_n_params(call_type);
723 ress = get_method_n_ress(call_type);
726 for (i = 0; i < params; ++i) {
727 type *p_type = get_method_param_type(call_type, i);
729 if (is_compound_type(p_type))
734 for (i = 0; i < ress; ++i) {
735 type *r_type = get_method_res_type(call_type, i);
737 if (is_compound_type(r_type))
742 irg_walk_graph(called_graph, find_addr, NULL, &res);
747 int inline_method(ir_node *call, ir_graph *called_graph) {
749 ir_node *post_call, *post_bl;
751 ir_node *end, *end_bl;
755 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
758 irg_inline_property prop = get_irg_inline_property(called_graph);
760 if ( (prop != irg_inline_forced) &&
761 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
763 /* Do not inline variadic functions. */
764 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
767 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
768 get_method_n_params(get_Call_type(call)));
771 * currently, we cannot inline two cases:
772 * - call with compound arguments
773 * - graphs that take the address of a parameter
775 if (! can_inline(call, called_graph))
778 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
779 rem_opt = get_opt_optimize();
782 /* Handle graph state */
783 assert(get_irg_phase_state(current_ir_graph) != phase_building);
784 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
785 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
786 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
787 set_irg_outs_inconsistent(current_ir_graph);
788 set_irg_loopinfo_inconsistent(current_ir_graph);
789 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
791 /* -- Check preconditions -- */
792 assert(get_irn_op(call) == op_Call);
793 /* @@@ does not work for InterfaceIII.java after cgana
794 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
795 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
796 get_Call_type(call)));
798 assert(get_type_tpop(get_Call_type(call)) == type_method);
799 if (called_graph == current_ir_graph) {
800 set_optimize(rem_opt);
804 /* here we know we WILL inline, so inform the statistics */
805 hook_inline(call, called_graph);
807 /* -- Decide how to handle exception control flow: Is there a handler
808 for the Call node, or do we branch directly to End on an exception?
810 0 There is a handler.
812 2 Exception handling not represented in Firm. -- */
814 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
815 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
816 assert(get_irn_op(proj) == op_Proj);
817 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
818 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
820 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
821 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
822 else { exc_handling = 2; } /* !Mproj && !Xproj */
827 the procedure and later replaces the Start node of the called graph.
828 Post_call is the old Call node and collects the results of the called
829 graph. Both will end up being a tuple. -- */
830 post_bl = get_nodes_block(call);
831 set_irg_current_block(current_ir_graph, post_bl);
832 /* XxMxPxP of Start + parameter of Call */
833 in[pn_Start_X_initial_exec] = new_Jmp();
834 in[pn_Start_M] = get_Call_mem(call);
835 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
836 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
837 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
838 /* in[pn_Start_P_value_arg_base] = ??? */
839 pre_call = new_Tuple(5, in);
843 The new block gets the ins of the old block, pre_call and all its
844 predecessors and all Phi nodes. -- */
845 part_block(pre_call);
847 /* -- Prepare state for dead node elimination -- */
848 /* Visited flags in calling irg must be >= flag in called irg.
849 Else walker and arity computation will not work. */
850 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
851 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
852 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
853 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
854 /* Set pre_call as new Start node in link field of the start node of
855 calling graph and pre_calls block as new block for the start block
857 Further mark these nodes so that they are not visited by the
859 set_irn_link(get_irg_start(called_graph), pre_call);
860 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
861 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
862 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
863 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
864 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
866 /* Initialize for compaction of in arrays */
867 inc_irg_block_visited(current_ir_graph);
869 /* -- Replicate local entities of the called_graph -- */
870 /* copy the entities. */
871 called_frame = get_irg_frame_type(called_graph);
872 for (i = 0; i < get_class_n_members(called_frame); i++) {
873 entity *new_ent, *old_ent;
874 old_ent = get_class_member(called_frame, i);
875 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
876 set_entity_link(old_ent, new_ent);
879 /* visited is > than that of called graph. With this trick visited will
880 remain unchanged so that an outer walker, e.g., searching the call nodes
881 to inline, calling this inline will not visit the inlined nodes. */
882 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
884 /* -- Performing dead node elimination inlines the graph -- */
885 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
887 /* @@@ endless loops are not copied!! -- they should be, I think... */
888 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
889 get_irg_frame_type(called_graph));
891 /* Repair called_graph */
892 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
893 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
894 set_Block_block_visited(get_irg_start_block(called_graph), 0);
896 /* -- Merge the end of the inlined procedure with the call site -- */
897 /* We will turn the old Call node into a Tuple with the following
900 0: Phi of all Memories of Return statements.
901 1: Jmp from new Block that merges the control flow from all exception
902 predecessors of the old end block.
903 2: Tuple of all arguments.
904 3: Phi of Exception memories.
905 In case the old Call directly branches to End on an exception we don't
906 need the block merging all exceptions nor the Phi of the exception
910 /* -- Precompute some values -- */
911 end_bl = get_new_node(get_irg_end_block(called_graph));
912 end = get_new_node(get_irg_end(called_graph));
913 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
914 n_res = get_method_n_ress(get_Call_type(call));
916 res_pred = xmalloc (n_res * sizeof(*res_pred));
917 cf_pred = xmalloc (arity * sizeof(*res_pred));
919 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
921 /* -- archive keepalives -- */
922 irn_arity = get_irn_arity(end);
923 for (i = 0; i < irn_arity; i++)
924 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
926 /* The new end node will die. We need not free as the in array is on the obstack:
927 firm_copy_node only generated 'D' arrays. */
929 /* -- Replace Return nodes by Jump nodes. -- */
931 for (i = 0; i < arity; i++) {
933 ret = get_irn_n(end_bl, i);
934 if (get_irn_op(ret) == op_Return) {
935 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
939 set_irn_in(post_bl, n_ret, cf_pred);
941 /* -- Build a Tuple for all results of the method.
942 Add Phi node if there was more than one Return. -- */
943 turn_into_tuple(post_call, 4);
944 /* First the Memory-Phi */
946 for (i = 0; i < arity; i++) {
947 ret = get_irn_n(end_bl, i);
948 if (get_irn_op(ret) == op_Return) {
949 cf_pred[n_ret] = get_Return_mem(ret);
953 phi = new_Phi(n_ret, cf_pred, mode_M);
954 set_Tuple_pred(call, pn_Call_M_regular, phi);
955 /* Conserve Phi-list for further inlinings -- but might be optimized */
956 if (get_nodes_block(phi) == post_bl) {
957 set_irn_link(phi, get_irn_link(post_bl));
958 set_irn_link(post_bl, phi);
960 /* Now the real results */
962 for (j = 0; j < n_res; j++) {
964 for (i = 0; i < arity; i++) {
965 ret = get_irn_n(end_bl, i);
966 if (get_irn_op(ret) == op_Return) {
967 cf_pred[n_ret] = get_Return_res(ret, j);
972 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
976 /* Conserve Phi-list for further inlinings -- but might be optimized */
977 if (get_nodes_block(phi) == post_bl) {
978 set_irn_link(phi, get_irn_link(post_bl));
979 set_irn_link(post_bl, phi);
982 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
984 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
986 /* Finally the exception control flow.
987 We have two (three) possible situations:
988 First if the Call branches to an exception handler: We need to add a Phi node to
989 collect the memory containing the exception objects. Further we need
990 to add another block to get a correct representation of this Phi. To
991 this block we add a Jmp that resolves into the X output of the Call
992 when the Call is turned into a tuple.
993 Second the Call branches to End, the exception is not handled. Just
994 add all inlined exception branches to the End node.
995 Third: there is no Exception edge at all. Handle as case two. */
996 if (exc_handling == 0) {
998 for (i = 0; i < arity; i++) {
1000 ret = get_irn_n(end_bl, i);
1001 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1002 cf_pred[n_exc] = ret;
1007 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1008 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1009 /* The Phi for the memories with the exception objects */
1011 for (i = 0; i < arity; i++) {
1013 ret = skip_Proj(get_irn_n(end_bl, i));
1014 if (get_irn_op(ret) == op_Call) {
1015 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1017 } else if (is_fragile_op(ret)) {
1018 /* We rely that all cfops have the memory output at the same position. */
1019 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1021 } else if (get_irn_op(ret) == op_Raise) {
1022 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1026 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1028 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1029 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1032 ir_node *main_end_bl;
1033 int main_end_bl_arity;
1034 ir_node **end_preds;
1036 /* assert(exc_handling == 1 || no exceptions. ) */
1038 for (i = 0; i < arity; i++) {
1039 ir_node *ret = get_irn_n(end_bl, i);
1041 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1042 cf_pred[n_exc] = ret;
1046 main_end_bl = get_irg_end_block(current_ir_graph);
1047 main_end_bl_arity = get_irn_arity(main_end_bl);
1048 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1050 for (i = 0; i < main_end_bl_arity; ++i)
1051 end_preds[i] = get_irn_n(main_end_bl, i);
1052 for (i = 0; i < n_exc; ++i)
1053 end_preds[main_end_bl_arity + i] = cf_pred[i];
1054 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1055 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1056 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1062 #if 0 /* old. now better, correcter, faster implementation. */
1064 /* -- If the exception control flow from the inlined Call directly
1065 branched to the end block we now have the following control
1066 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1067 remove the Jmp along with it's empty block and add Jmp's
1068 predecessors as predecessors of this end block. No problem if
1069 there is no exception, because then branches Bad to End which
1071 @@@ can't we know this beforehand: by getting the Proj(1) from
1072 the Call link list and checking whether it goes to Proj. */
1073 /* find the problematic predecessor of the end block. */
1074 end_bl = get_irg_end_block(current_ir_graph);
1075 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1076 cf_op = get_Block_cfgpred(end_bl, i);
1077 if (get_irn_op(cf_op) == op_Proj) {
1078 cf_op = get_Proj_pred(cf_op);
1079 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1080 /* There are unoptimized tuples from inlineing before when no exc */
1081 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1082 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1083 assert(get_irn_op(cf_op) == op_Jmp);
1089 if (i < get_Block_n_cfgpreds(end_bl)) {
1090 bl = get_nodes_block(cf_op);
1091 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1092 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1093 for (j = 0; j < i; j++)
1094 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1095 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1096 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1097 for (j = j; j < arity; j++)
1098 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1099 set_irn_in(end_bl, arity, cf_pred);
1101 /* Remove the exception pred from post-call Tuple. */
1102 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1107 /* -- Turn cse back on. -- */
1108 set_optimize(rem_opt);
1113 /********************************************************************/
1114 /* Apply inlineing to small methods. */
1115 /********************************************************************/
1117 /* It makes no sense to inline too many calls in one procedure. Anyways,
1118 I didn't get a version with NEW_ARR_F to run. */
1119 #define MAX_INLINE 1024
1122 * environment for inlining small irgs
1124 typedef struct _inline_env_t {
1126 ir_node *calls[MAX_INLINE];
1130 * Returns the irg called from a Call node. If the irg is not
1131 * known, NULL is returned.
1133 static ir_graph *get_call_called_irg(ir_node *call) {
1135 ir_graph *called_irg = NULL;
1137 assert(get_irn_op(call) == op_Call);
1139 addr = get_Call_ptr(call);
1140 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1141 called_irg = get_entity_irg(get_SymConst_entity(addr));
1147 static void collect_calls(ir_node *call, void *env) {
1150 if (get_irn_op(call) != op_Call) return;
1152 addr = get_Call_ptr(call);
1154 if (get_irn_op(addr) == op_SymConst) {
1155 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1156 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1157 inline_env_t *ienv = (inline_env_t *)env;
1158 if (called_irg && ienv->pos < MAX_INLINE) {
1159 /* The Call node calls a locally defined method. Remember to inline. */
1160 ienv->calls[ienv->pos++] = call;
1167 * Inlines all small methods at call sites where the called address comes
1168 * from a Const node that references the entity representing the called
1170 * The size argument is a rough measure for the code size of the method:
1171 * Methods where the obstack containing the firm graph is smaller than
1174 void inline_small_irgs(ir_graph *irg, int size) {
1176 ir_graph *rem = current_ir_graph;
1177 inline_env_t env /* = {0, NULL}*/;
1179 if (!(get_opt_optimize() && get_opt_inline())) return;
1181 current_ir_graph = irg;
1182 /* Handle graph state */
1183 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1184 free_callee_info(current_ir_graph);
1186 /* Find Call nodes to inline.
1187 (We can not inline during a walk of the graph, as inlineing the same
1188 method several times changes the visited flag of the walked graph:
1189 after the first inlineing visited of the callee equals visited of
1190 the caller. With the next inlineing both are increased.) */
1192 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1194 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1195 /* There are calls to inline */
1196 collect_phiprojs(irg);
1197 for (i = 0; i < env.pos; i++) {
1199 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1200 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1201 (get_irg_inline_property(callee) == irg_inline_forced)) {
1202 inline_method(env.calls[i], callee);
1207 current_ir_graph = rem;
1211 * Environment for inlining irgs.
1214 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1215 int n_nodes_orig; /**< for statistics */
1216 eset *call_nodes; /**< All call nodes in this graph */
1218 int n_call_nodes_orig; /**< for statistics */
1219 int n_callers; /**< Number of known graphs that call this graphs. */
1220 int n_callers_orig; /**< for statistics */
1223 static inline_irg_env *new_inline_irg_env(void) {
1224 inline_irg_env *env = xmalloc(sizeof(*env));
1225 env->n_nodes = -2; /* uncount Start, End */
1226 env->n_nodes_orig = -2; /* uncount Start, End */
1227 env->call_nodes = eset_create();
1228 env->n_call_nodes = 0;
1229 env->n_call_nodes_orig = 0;
1231 env->n_callers_orig = 0;
1235 static void free_inline_irg_env(inline_irg_env *env) {
1236 eset_destroy(env->call_nodes);
1240 static void collect_calls2(ir_node *call, void *env) {
1241 inline_irg_env *x = (inline_irg_env *)env;
1242 ir_op *op = get_irn_op(call);
1245 /* count nodes in irg */
1246 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1251 if (op != op_Call) return;
1253 /* collect all call nodes */
1254 eset_insert(x->call_nodes, (void *)call);
1256 x->n_call_nodes_orig++;
1258 /* count all static callers */
1259 callee = get_call_called_irg(call);
1261 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1262 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1266 INLINE static int is_leave(ir_graph *irg) {
1267 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1270 INLINE static int is_smaller(ir_graph *callee, int size) {
1271 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1276 * Inlines small leave methods at call sites where the called address comes
1277 * from a Const node that references the entity representing the called
1279 * The size argument is a rough measure for the code size of the method:
1280 * Methods where the obstack containing the firm graph is smaller than
1283 void inline_leave_functions(int maxsize, int leavesize, int size) {
1284 inline_irg_env *env;
1285 int i, n_irgs = get_irp_n_irgs();
1286 ir_graph *rem = current_ir_graph;
1289 if (!(get_opt_optimize() && get_opt_inline())) return;
1291 /* extend all irgs by a temporary data structure for inlining. */
1292 for (i = 0; i < n_irgs; ++i)
1293 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1295 /* Precompute information in temporary data structure. */
1296 for (i = 0; i < n_irgs; ++i) {
1297 current_ir_graph = get_irp_irg(i);
1298 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1299 free_callee_info(current_ir_graph);
1301 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1302 get_irg_link(current_ir_graph));
1305 /* -- and now inline. -- */
1307 /* Inline leaves recursively -- we might construct new leaves. */
1308 while (did_inline) {
1311 for (i = 0; i < n_irgs; ++i) {
1313 int phiproj_computed = 0;
1315 current_ir_graph = get_irp_irg(i);
1316 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1318 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1321 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1322 callee = get_call_called_irg(call);
1324 if (env->n_nodes > maxsize) continue; // break;
1326 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1327 if (!phiproj_computed) {
1328 phiproj_computed = 1;
1329 collect_phiprojs(current_ir_graph);
1331 did_inline = inline_method(call, callee);
1334 /* Do some statistics */
1335 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1336 env->n_call_nodes --;
1337 env->n_nodes += callee_env->n_nodes;
1338 callee_env->n_callers--;
1345 /* inline other small functions. */
1346 for (i = 0; i < n_irgs; ++i) {
1349 int phiproj_computed = 0;
1351 current_ir_graph = get_irp_irg(i);
1352 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1354 /* we can not walk and change a set, nor remove from it.
1356 walkset = env->call_nodes;
1357 env->call_nodes = eset_create();
1358 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1361 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1362 callee = get_call_called_irg(call);
1365 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1366 (get_irg_inline_property(callee) == irg_inline_forced))) {
1367 if (!phiproj_computed) {
1368 phiproj_computed = 1;
1369 collect_phiprojs(current_ir_graph);
1371 if (inline_method(call, callee)) {
1372 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1373 env->n_call_nodes--;
1374 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1375 env->n_call_nodes += callee_env->n_call_nodes;
1376 env->n_nodes += callee_env->n_nodes;
1377 callee_env->n_callers--;
1380 eset_insert(env->call_nodes, call);
1383 eset_destroy(walkset);
1386 for (i = 0; i < n_irgs; ++i) {
1387 current_ir_graph = get_irp_irg(i);
1389 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1390 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1391 (env->n_callers_orig != env->n_callers))
1392 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1393 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1394 env->n_callers_orig, env->n_callers,
1395 get_entity_name(get_irg_entity(current_ir_graph)));
1397 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1400 current_ir_graph = rem;
1403 /*******************************************************************/
1404 /* Code Placement. Pins all floating nodes to a block where they */
1405 /* will be executed only if needed. */
1406 /*******************************************************************/
1409 * Find the earliest correct block for N. --- Place N into the
1410 * same Block as its dominance-deepest Input.
1413 place_floats_early(ir_node *n, pdeq *worklist)
1415 int i, start, irn_arity;
1417 /* we must not run into an infinite loop */
1418 assert (irn_not_visited(n));
1419 mark_irn_visited(n);
1421 /* Place floating nodes. */
1422 if (get_irn_pinned(n) == op_pin_state_floats) {
1424 ir_node *b = new_Bad(); /* The block to place this node in */
1425 int bad_recursion = is_Bad(get_nodes_block(n));
1427 assert(get_irn_op(n) != op_Block);
1429 if ((get_irn_op(n) == op_Const) ||
1430 (get_irn_op(n) == op_SymConst) ||
1432 (get_irn_op(n) == op_Unknown)) {
1433 /* These nodes will not be placed by the loop below. */
1434 b = get_irg_start_block(current_ir_graph);
1438 /* find the block for this node. */
1439 irn_arity = get_irn_arity(n);
1440 for (i = 0; i < irn_arity; i++) {
1441 ir_node *dep = get_irn_n(n, i);
1444 if ((irn_not_visited(dep))
1445 && (get_irn_pinned(dep) == op_pin_state_floats)) {
1446 place_floats_early(dep, worklist);
1450 * A node in the Bad block must stay in the bad block,
1451 * so don't compute a new block for it.
1456 /* Because all loops contain at least one op_pin_state_pinned node, now all
1457 our inputs are either op_pin_state_pinned or place_early has already
1458 been finished on them. We do not have any unfinished inputs! */
1459 dep_block = get_nodes_block(dep);
1460 if ((!is_Bad(dep_block)) &&
1461 (get_Block_dom_depth(dep_block) > depth)) {
1463 depth = get_Block_dom_depth(dep_block);
1465 /* Avoid that the node is placed in the Start block */
1466 if ((depth == 1) && (get_Block_dom_depth(get_nodes_block(n)) > 1)) {
1467 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1468 assert(b != get_irg_start_block(current_ir_graph));
1472 set_nodes_block(n, b);
1475 /* Add predecessors of non floating nodes on worklist. */
1476 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1477 irn_arity = get_irn_arity(n);
1478 for (i = start; i < irn_arity; i++) {
1479 ir_node *pred = get_irn_n(n, i);
1480 if (irn_not_visited(pred)) {
1481 pdeq_putr (worklist, pred);
1487 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1488 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1489 * places all floating nodes reachable from its argument through floating
1490 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1492 static INLINE void place_early(pdeq *worklist) {
1494 inc_irg_visited(current_ir_graph);
1496 /* this inits the worklist */
1497 place_floats_early(get_irg_end(current_ir_graph), worklist);
1499 /* Work the content of the worklist. */
1500 while (!pdeq_empty (worklist)) {
1501 ir_node *n = pdeq_getl (worklist);
1502 if (irn_not_visited(n)) place_floats_early(n, worklist);
1505 set_irg_outs_inconsistent(current_ir_graph);
1506 current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
1509 /** Compute the deepest common ancestor of block and dca. */
1510 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1513 if (!dca) return block;
1514 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1515 block = get_Block_idom(block);
1516 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1517 dca = get_Block_idom(dca);
1519 while (block != dca)
1520 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1525 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1526 * I.e., DCA is the block where we might place PRODUCER.
1527 * A data flow edge points from producer to consumer.
1530 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1532 ir_node *block = NULL;
1534 /* Compute the latest block into which we can place a node so that it is
1536 if (get_irn_op(consumer) == op_Phi) {
1537 /* our consumer is a Phi-node, the effective use is in all those
1538 blocks through which the Phi-node reaches producer */
1540 ir_node *phi_block = get_nodes_block(consumer);
1541 irn_arity = get_irn_arity(consumer);
1543 for (i = 0; i < irn_arity; i++) {
1544 if (get_irn_n(consumer, i) == producer) {
1545 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1547 block = calc_dca(block, new_block);
1551 assert(is_no_Block(consumer));
1552 block = get_nodes_block(consumer);
1555 /* Compute the deepest common ancestor of block and dca. */
1556 return calc_dca(dca, block);
1559 static INLINE int get_irn_loop_depth(ir_node *n) {
1560 return get_loop_depth(get_irn_loop(n));
1564 * Move n to a block with less loop depth than it's current block. The
1565 * new block must be dominated by early.
1568 move_out_of_loops (ir_node *n, ir_node *early)
1570 ir_node *best, *dca;
1574 /* Find the region deepest in the dominator tree dominating
1575 dca with the least loop nesting depth, but still dominated
1576 by our early placement. */
1577 dca = get_nodes_block(n);
1579 while (dca != early) {
1580 dca = get_Block_idom(dca);
1581 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1582 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1586 if (best != get_nodes_block(n)) {
1588 printf("Moving out of loop: "); DDMN(n);
1589 printf(" Outermost block: "); DDMN(early);
1590 printf(" Best block: "); DDMN(best);
1591 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1593 set_nodes_block(n, best);
1598 * Find the latest legal block for N and place N into the
1599 * `optimal' Block between the latest and earliest legal block.
1600 * The `optimal' block is the dominance-deepest block of those
1601 * with the least loop-nesting-depth. This places N out of as many
1602 * loops as possible and then makes it as control dependant as
1606 place_floats_late(ir_node *n, pdeq *worklist)
1611 assert (irn_not_visited(n)); /* no multiple placement */
1613 mark_irn_visited(n);
1615 /* no need to place block nodes, control nodes are already placed. */
1616 if ((get_irn_op(n) != op_Block) &&
1618 (get_irn_mode(n) != mode_X)) {
1619 /* Remember the early placement of this block to move it
1620 out of loop no further than the early placement. */
1621 early = get_nodes_block(n);
1623 /* Do not move code not reachable from Start. For
1624 * these we could not compute dominator information. */
1625 if (is_Bad(early) || get_Block_dom_depth(early) == -1)
1628 /* Assure that our users are all placed, except the Phi-nodes.
1629 --- Each data flow cycle contains at least one Phi-node. We
1630 have to break the `user has to be placed before the
1631 producer' dependence cycle and the Phi-nodes are the
1632 place to do so, because we need to base our placement on the
1633 final region of our users, which is OK with Phi-nodes, as they
1634 are op_pin_state_pinned, and they never have to be placed after a
1635 producer of one of their inputs in the same block anyway. */
1636 for (i = 0; i < get_irn_n_outs(n); i++) {
1637 ir_node *succ = get_irn_out(n, i);
1638 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1639 place_floats_late(succ, worklist);
1642 /* We have to determine the final block of this node... except for
1644 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1645 (get_irn_op(n) != op_Const) &&
1646 (get_irn_op(n) != op_SymConst)) {
1647 ir_node *dca = NULL; /* deepest common ancestor in the
1648 dominator tree of all nodes'
1649 blocks depending on us; our final
1650 placement has to dominate DCA. */
1651 for (i = 0; i < get_irn_n_outs(n); i++) {
1652 ir_node *out = get_irn_out(n, i);
1653 /* ignore if out is in dead code */
1654 ir_node *outbl = get_nodes_block(out);
1655 if (is_Bad(outbl) || get_Block_dom_depth(outbl) == -1)
1657 dca = consumer_dom_dca (dca, out, n);
1660 set_nodes_block(n, dca);
1662 move_out_of_loops (n, early);
1664 /* else all outs are in dead code */
1668 /* Add predecessors of all non-floating nodes on list. (Those of floating
1669 nodes are placeded already and therefore are marked.) */
1670 for (i = 0; i < get_irn_n_outs(n); i++) {
1671 ir_node *succ = get_irn_out(n, i);
1672 if (irn_not_visited(get_irn_out(n, i))) {
1673 pdeq_putr (worklist, succ);
1678 static INLINE void place_late(pdeq *worklist) {
1680 inc_irg_visited(current_ir_graph);
1682 /* This fills the worklist initially. */
1683 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1685 /* And now empty the worklist again... */
1686 while (!pdeq_empty (worklist)) {
1687 ir_node *n = pdeq_getl (worklist);
1688 if (irn_not_visited(n)) place_floats_late(n, worklist);
1692 void place_code(ir_graph *irg) {
1694 ir_graph *rem = current_ir_graph;
1696 current_ir_graph = irg;
1698 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1700 /* Handle graph state */
1701 assert(get_irg_phase_state(irg) != phase_building);
1702 if (get_irg_dom_state(irg) != dom_consistent)
1705 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1706 free_loop_information(irg);
1707 construct_backedges(irg);
1710 /* Place all floating nodes as early as possible. This guarantees
1711 a legal code placement. */
1712 worklist = new_pdeq();
1713 place_early(worklist);
1715 /* place_early invalidates the outs, place_late needs them. */
1717 /* Now move the nodes down in the dominator tree. This reduces the
1718 unnecessary executions of the node. */
1719 place_late(worklist);
1721 set_irg_outs_inconsistent(current_ir_graph);
1722 set_irg_loopinfo_inconsistent(current_ir_graph);
1724 current_ir_graph = rem;
1728 * Called by walker of remove_critical_cf_edges().
1730 * Place an empty block to an edge between a blocks of multiple
1731 * predecessors and a block of multiple successors.
1734 * @param env Environment of walker. This field is unused and has
1737 static void walk_critical_cf_edges(ir_node *n, void *env) {
1739 ir_node *pre, *block, **in, *jmp;
1741 /* Block has multiple predecessors */
1742 if ((op_Block == get_irn_op(n)) &&
1743 (get_irn_arity(n) > 1)) {
1744 arity = get_irn_arity(n);
1746 if (n == get_irg_end_block(current_ir_graph))
1747 return; /* No use to add a block here. */
1749 for (i=0; i<arity; i++) {
1750 pre = get_irn_n(n, i);
1751 /* Predecessor has multiple successors. Insert new flow edge */
1752 if ((NULL != pre) &&
1753 (op_Proj == get_irn_op(pre)) &&
1754 op_Raise != get_irn_op(skip_Proj(pre))) {
1756 /* set predecessor array for new block */
1757 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1758 /* set predecessor of new block */
1760 block = new_Block(1, in);
1761 /* insert new jmp node to new block */
1762 set_cur_block(block);
1765 /* set successor of new block */
1766 set_irn_n(n, i, jmp);
1768 } /* predecessor has multiple successors */
1769 } /* for all predecessors */
1770 } /* n is a block */
1773 void remove_critical_cf_edges(ir_graph *irg) {
1774 if (get_opt_critical_edges())
1775 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);