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 /* Defined in iropt.c */
45 pset *new_identities (void);
46 void del_identities (pset *value_table);
47 void add_identities (pset *value_table, ir_node *node);
49 /*------------------------------------------------------------------*/
50 /* apply optimizations of iropt to all nodes. */
51 /*------------------------------------------------------------------*/
53 static void init_link (ir_node *n, void *env) {
54 set_irn_link(n, NULL);
57 #if 0 /* Old version. Avoids Ids.
58 This is not necessary: we do a postwalk, and get_irn_n
59 removes ids anyways. So it's much cheaper to call the
60 optimization less often and use the exchange() algorithm. */
62 optimize_in_place_wrapper (ir_node *n, void *env) {
64 ir_node *optimized, *old;
66 irn_arity = get_irn_arity(n);
67 for (i = 0; i < irn_arity; i++) {
68 /* get_irn_n skips Id nodes, so comparison old != optimized does not
69 show all optimizations. Therefore always set new predecessor. */
70 old = get_irn_intra_n(n, i);
71 optimized = optimize_in_place_2(old);
72 set_irn_n(n, i, optimized);
75 if (get_irn_op(n) == op_Block) {
76 optimized = optimize_in_place_2(n);
77 if (optimized != n) exchange (n, optimized);
82 optimize_in_place_wrapper (ir_node *n, void *env) {
83 ir_node *optimized = optimize_in_place_2(n);
84 if (optimized != n) exchange (n, optimized);
89 static INLINE void do_local_optimize(ir_node *n) {
90 /* Handle graph state */
91 assert(get_irg_phase_state(current_ir_graph) != phase_building);
92 if (get_opt_global_cse())
93 set_irg_pinned(current_ir_graph, op_pin_state_floats);
94 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
95 set_irg_outs_inconsistent(current_ir_graph);
96 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
97 set_irg_dom_inconsistent(current_ir_graph);
98 set_irg_loopinfo_inconsistent(current_ir_graph);
101 /* Clean the value_table in irg for the cse. */
102 del_identities(current_ir_graph->value_table);
103 current_ir_graph->value_table = new_identities();
105 /* walk over the graph */
106 irg_walk(n, init_link, optimize_in_place_wrapper, NULL);
109 void local_optimize_node(ir_node *n) {
110 ir_graph *rem = current_ir_graph;
111 current_ir_graph = get_irn_irg(n);
113 do_local_optimize(n);
115 current_ir_graph = rem;
120 local_optimize_graph (ir_graph *irg) {
121 ir_graph *rem = current_ir_graph;
122 current_ir_graph = irg;
124 do_local_optimize(irg->end);
126 current_ir_graph = rem;
130 /*------------------------------------------------------------------*/
131 /* Routines for dead node elimination / copying garbage collection */
132 /* of the obstack. */
133 /*------------------------------------------------------------------*/
136 * Remember the new node in the old node by using a field all nodes have.
139 set_new_node (ir_node *old, ir_node *new)
145 * Get this new node, before the old node is forgotton.
147 static INLINE ir_node *
148 get_new_node (ir_node * n)
154 * We use the block_visited flag to mark that we have computed the
155 * number of useful predecessors for this block.
156 * Further we encode the new arity in this flag in the old blocks.
157 * Remembering the arity is useful, as it saves a lot of pointer
158 * accesses. This function is called for all Phi and Block nodes
162 compute_new_arity(ir_node *b) {
163 int i, res, irn_arity;
166 irg_v = get_irg_block_visited(current_ir_graph);
167 block_v = get_Block_block_visited(b);
168 if (block_v >= irg_v) {
169 /* we computed the number of preds for this block and saved it in the
171 return block_v - irg_v;
173 /* compute the number of good predecessors */
174 res = irn_arity = get_irn_arity(b);
175 for (i = 0; i < irn_arity; i++)
176 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
177 /* save it in the flag. */
178 set_Block_block_visited(b, irg_v + res);
183 /* TODO: add an ir_op operation */
184 static INLINE void new_backedge_info(ir_node *n) {
185 switch(get_irn_opcode(n)) {
187 n->attr.block.cg_backedge = NULL;
188 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
191 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
194 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
201 * Copies the node to the new obstack. The Ins of the new node point to
202 * the predecessors on the old obstack. For block/phi nodes not all
203 * predecessors might be copied. n->link points to the new node.
204 * For Phi and Block nodes the function allocates in-arrays with an arity
205 * only for useful predecessors. The arity is determined by counting
206 * the non-bad predecessors of the block.
208 * @param n The node to be copied
209 * @param env if non-NULL, the node number attribute will be copied to the new node
211 * Note: Also used for loop unrolling.
214 firm_copy_node (ir_node *n, void *env) {
217 opcode op = get_irn_opcode(n);
218 int copy_node_nr = env != NULL;
220 /* The end node looses it's flexible in array. This doesn't matter,
221 as dead node elimination builds End by hand, inlineing doesn't use
223 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
226 /* node copied already */
228 } else if (op == iro_Block) {
230 new_arity = compute_new_arity(n);
231 n->attr.block.graph_arr = NULL;
233 block = get_nodes_block(n);
234 if (get_irn_opcode(n) == iro_Phi) {
235 new_arity = compute_new_arity(block);
237 new_arity = get_irn_arity(n);
240 nn = new_ir_node(get_irn_dbg_info(n),
247 /* Copy the attributes. These might point to additional data. If this
248 was allocated on the old obstack the pointers now are dangling. This
249 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
250 copy_node_attr(n, nn);
251 new_backedge_info(nn);
256 /* for easier debugging, we want to copy the node numbers too */
257 nn->node_nr = n->node_nr;
261 /* printf("\n old node: "); DDMSG2(n);
262 printf(" new node: "); DDMSG2(nn); */
266 * Copies new predecessors of old node to new node remembered in link.
267 * Spare the Bad predecessors of Phi and Block nodes.
270 copy_preds (ir_node *n, void *env) {
274 nn = get_new_node(n);
276 /* printf("\n old node: "); DDMSG2(n);
277 printf(" new node: "); DDMSG2(nn);
278 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
280 if (get_irn_opcode(n) == iro_Block) {
281 /* Don't copy Bad nodes. */
283 irn_arity = get_irn_arity(n);
284 for (i = 0; i < irn_arity; i++)
285 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
286 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
287 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
290 /* repair the block visited flag from above misuse. Repair it in both
291 graphs so that the old one can still be used. */
292 set_Block_block_visited(nn, 0);
293 set_Block_block_visited(n, 0);
294 /* Local optimization could not merge two subsequent blocks if
295 in array contained Bads. Now it's possible.
296 We don't call optimize_in_place as it requires
297 that the fields in ir_graph are set properly. */
298 if ((get_opt_control_flow_straightening()) &&
299 (get_Block_n_cfgpreds(nn) == 1) &&
300 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
301 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
303 /* Jmp jumps into the block it is in -- deal self cycle. */
304 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
305 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
310 } else if (get_irn_opcode(n) == iro_Phi) {
311 /* Don't copy node if corresponding predecessor in block is Bad.
312 The Block itself should not be Bad. */
313 block = get_nodes_block(n);
314 set_irn_n (nn, -1, get_new_node(block));
316 irn_arity = get_irn_arity(n);
317 for (i = 0; i < irn_arity; i++)
318 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
319 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
320 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
323 /* If the pre walker reached this Phi after the post walker visited the
324 block block_visited is > 0. */
325 set_Block_block_visited(get_nodes_block(n), 0);
326 /* Compacting the Phi's ins might generate Phis with only one
328 if (get_irn_arity(n) == 1)
329 exchange(n, get_irn_n(n, 0));
331 irn_arity = get_irn_arity(n);
332 for (i = -1; i < irn_arity; i++)
333 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
335 /* Now the new node is complete. We can add it to the hash table for cse.
336 @@@ inlinening aborts if we identify End. Why? */
337 if(get_irn_op(nn) != op_End)
338 add_identities (current_ir_graph->value_table, nn);
342 * Copies the graph recursively, compacts the keepalive of the end node.
344 * @param copy_node_nr If non-zero, the node number will be copied
347 copy_graph (int copy_node_nr) {
348 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
349 ir_node *ka; /* keep alive */
352 oe = get_irg_end(current_ir_graph);
353 /* copy the end node by hand, allocate dynamic in array! */
354 ne = new_ir_node(get_irn_dbg_info(oe),
361 /* Copy the attributes. Well, there might be some in the future... */
362 copy_node_attr(oe, ne);
363 set_new_node(oe, ne);
365 /* copy the Bad node */
366 ob = get_irg_bad(current_ir_graph);
367 nb = new_ir_node(get_irn_dbg_info(ob),
374 set_new_node(ob, nb);
376 /* copy the NoMem node */
377 om = get_irg_no_mem(current_ir_graph);
378 nm = new_ir_node(get_irn_dbg_info(om),
385 set_new_node(om, nm);
387 /* copy the live nodes */
388 irg_walk(get_nodes_block(oe), firm_copy_node, copy_preds, (void *)copy_node_nr);
389 /* copy_preds for the end node ... */
390 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
392 /*- ... and now the keep alives. -*/
393 /* First pick the not marked block nodes and walk them. We must pick these
394 first as else we will oversee blocks reachable from Phis. */
395 irn_arity = get_irn_arity(oe);
396 for (i = 0; i < irn_arity; i++) {
397 ka = get_irn_intra_n(oe, i);
398 if ((get_irn_op(ka) == op_Block) &&
399 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
400 /* We must keep the block alive and copy everything reachable */
401 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
402 irg_walk(ka, firm_copy_node, copy_preds, (void *)copy_node_nr);
403 add_End_keepalive(ne, get_new_node(ka));
407 /* Now pick the Phis. Here we will keep all! */
408 irn_arity = get_irn_arity(oe);
409 for (i = 0; i < irn_arity; i++) {
410 ka = get_irn_intra_n(oe, i);
411 if ((get_irn_op(ka) == op_Phi)) {
412 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
413 /* We didn't copy the Phi yet. */
414 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
415 irg_walk(ka, firm_copy_node, copy_preds, (void *)copy_node_nr);
417 add_End_keepalive(ne, get_new_node(ka));
421 /* start block sometimes only reached after keep alives */
422 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
423 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
427 * Copies the graph reachable from current_ir_graph->end to the obstack
428 * in current_ir_graph and fixes the environment.
429 * Then fixes the fields in current_ir_graph containing nodes of the
432 * @param copy_node_nr If non-zero, the node number will be copied
435 copy_graph_env (int copy_node_nr) {
437 /* Not all nodes remembered in current_ir_graph might be reachable
438 from the end node. Assure their link is set to NULL, so that
439 we can test whether new nodes have been computed. */
440 set_irn_link(get_irg_frame (current_ir_graph), NULL);
441 set_irn_link(get_irg_globals (current_ir_graph), NULL);
442 set_irn_link(get_irg_args (current_ir_graph), NULL);
443 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
444 set_irn_link(get_irg_no_mem (current_ir_graph), NULL);
446 /* we use the block walk flag for removing Bads from Blocks ins. */
447 inc_irg_block_visited(current_ir_graph);
450 copy_graph(copy_node_nr);
452 /* fix the fields in current_ir_graph */
453 old_end = get_irg_end(current_ir_graph);
454 set_irg_end (current_ir_graph, get_new_node(old_end));
455 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
456 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
458 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
459 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
460 firm_copy_node (get_irg_frame(current_ir_graph), (void *)copy_node_nr);
461 copy_preds(get_irg_frame(current_ir_graph), NULL);
463 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
464 firm_copy_node (get_irg_globals(current_ir_graph), (void *)copy_node_nr);
465 copy_preds(get_irg_globals(current_ir_graph), NULL);
467 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
468 firm_copy_node (get_irg_initial_mem(current_ir_graph), (void *)copy_node_nr);
469 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
471 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
472 firm_copy_node (get_irg_args(current_ir_graph), (void *)copy_node_nr);
473 copy_preds(get_irg_args(current_ir_graph), NULL);
475 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
477 set_irg_start_block(current_ir_graph,
478 get_new_node(get_irg_start_block(current_ir_graph)));
479 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
480 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
481 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
482 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
484 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
485 firm_copy_node(get_irg_bad(current_ir_graph), (void *)copy_node_nr);
486 copy_preds(get_irg_bad(current_ir_graph), NULL);
488 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
490 if (get_irn_link(get_irg_no_mem(current_ir_graph)) == NULL) {
491 firm_copy_node(get_irg_no_mem(current_ir_graph), (void *)copy_node_nr);
492 copy_preds(get_irg_no_mem(current_ir_graph), NULL);
494 set_irg_no_mem(current_ir_graph, get_new_node(get_irg_no_mem(current_ir_graph)));
498 * Copies all reachable nodes to a new obstack. Removes bad inputs
499 * from block nodes and the corresponding inputs from Phi nodes.
500 * Merges single exit blocks with single entry blocks and removes
502 * Adds all new nodes to a new hash table for cse. Does not
503 * perform cse, so the hash table might contain common subexpressions.
506 dead_node_elimination(ir_graph *irg) {
508 int rem_ipview = get_interprocedural_view();
509 struct obstack *graveyard_obst = NULL;
510 struct obstack *rebirth_obst = NULL;
512 /* inform statistics that we started a dead-node elimination run */
513 hook_dead_node_elim_start(irg);
515 /* Remember external state of current_ir_graph. */
516 rem = current_ir_graph;
517 current_ir_graph = irg;
518 set_interprocedural_view(false);
520 /* Handle graph state */
521 assert(get_irg_phase_state(current_ir_graph) != phase_building);
522 free_callee_info(current_ir_graph);
523 free_outs(current_ir_graph);
524 /* @@@ so far we loose loops when copying */
525 free_loop_information(current_ir_graph);
527 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
529 /* A quiet place, where the old obstack can rest in peace,
530 until it will be cremated. */
531 graveyard_obst = irg->obst;
533 /* A new obstack, where the reachable nodes will be copied to. */
534 rebirth_obst = xmalloc (sizeof(*rebirth_obst));
535 current_ir_graph->obst = rebirth_obst;
536 obstack_init (current_ir_graph->obst);
538 /* We also need a new hash table for cse */
539 del_identities (irg->value_table);
540 irg->value_table = new_identities ();
542 /* Copy the graph from the old to the new obstack */
545 /* Free memory from old unoptimized obstack */
546 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
547 xfree (graveyard_obst); /* ... then free it. */
550 /* inform statistics that the run is over */
551 hook_dead_node_elim_stop(irg);
553 current_ir_graph = rem;
554 set_interprocedural_view(rem_ipview);
558 * Relink bad predeseccors of a block and store the old in array to the
559 * link field. This function is called by relink_bad_predecessors().
560 * The array of link field starts with the block operand at position 0.
561 * If block has bad predecessors, create a new in array without bad preds.
562 * Otherwise let in array untouched.
564 static void relink_bad_block_predecessors(ir_node *n, void *env) {
565 ir_node **new_in, *irn;
566 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
568 /* if link field of block is NULL, look for bad predecessors otherwise
569 this is allready done */
570 if (get_irn_op(n) == op_Block &&
571 get_irn_link(n) == NULL) {
573 /* save old predecessors in link field (position 0 is the block operand)*/
574 set_irn_link(n, (void *)get_irn_in(n));
576 /* count predecessors without bad nodes */
577 old_irn_arity = get_irn_arity(n);
578 for (i = 0; i < old_irn_arity; i++)
579 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
581 /* arity changing: set new predecessors without bad nodes */
582 if (new_irn_arity < old_irn_arity) {
583 /* Get new predecessor array. We do not resize the array, as we must
584 keep the old one to update Phis. */
585 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
587 /* set new predeseccors in array */
590 for (i = 0; i < old_irn_arity; i++) {
591 irn = get_irn_n(n, i);
593 new_in[new_irn_n] = irn;
594 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
598 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
599 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
602 } /* ir node has bad predecessors */
604 } /* Block is not relinked */
608 * Relinks Bad predecesors from Bocks and Phis called by walker
609 * remove_bad_predecesors(). If n is a Block, call
610 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
611 * function of Phi's Block. If this block has bad predecessors, relink preds
614 static void relink_bad_predecessors(ir_node *n, void *env) {
615 ir_node *block, **old_in;
616 int i, old_irn_arity, new_irn_arity;
618 /* relink bad predeseccors of a block */
619 if (get_irn_op(n) == op_Block)
620 relink_bad_block_predecessors(n, env);
622 /* If Phi node relink its block and its predecessors */
623 if (get_irn_op(n) == op_Phi) {
625 /* Relink predeseccors of phi's block */
626 block = get_nodes_block(n);
627 if (get_irn_link(block) == NULL)
628 relink_bad_block_predecessors(block, env);
630 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
631 old_irn_arity = ARR_LEN(old_in);
633 /* Relink Phi predeseccors if count of predeseccors changed */
634 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
635 /* set new predeseccors in array
636 n->in[0] remains the same block */
638 for(i = 1; i < old_irn_arity; i++)
639 if (!is_Bad((ir_node *)old_in[i])) {
640 n->in[new_irn_arity] = n->in[i];
641 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
645 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
646 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
649 } /* n is a Phi node */
653 * Removes Bad Bad predecesors from Blocks and the corresponding
654 * inputs to Phi nodes as in dead_node_elimination but without
656 * On walking up set the link field to NULL, on walking down call
657 * relink_bad_predecessors() (This function stores the old in array
658 * to the link field and sets a new in array if arity of predecessors
661 void remove_bad_predecessors(ir_graph *irg) {
662 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
666 /*--------------------------------------------------------------------*/
667 /* Funcionality for inlining */
668 /*--------------------------------------------------------------------*/
671 * Copy node for inlineing. Updates attributes that change when
672 * inlineing but not for dead node elimination.
674 * Copies the node by calling firm_copy_node and then updates the entity if
675 * it's a local one. env must be a pointer of the frame type of the
676 * inlined procedure. The new entities must be in the link field of
680 copy_node_inline (ir_node *n, void *env) {
682 type *frame_tp = (type *)env;
684 firm_copy_node(n, NULL);
685 if (get_irn_op(n) == op_Sel) {
686 new = get_new_node (n);
687 assert(get_irn_op(new) == op_Sel);
688 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
689 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
691 } else if (get_irn_op(n) == op_Block) {
692 new = get_new_node (n);
693 new->attr.block.irg = current_ir_graph;
697 static void find_addr(ir_node *node, void *env)
699 if (get_irn_opcode(node) == iro_Proj) {
700 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
706 * currently, we cannot inline two cases:
707 * - call with compound arguments
708 * - graphs that take the address of a parameter
710 * check these conditions here
712 static int can_inline(ir_node *call, ir_graph *called_graph)
714 type *call_type = get_Call_type(call);
715 int params, ress, i, res;
716 assert(is_Method_type(call_type));
718 params = get_method_n_params(call_type);
719 ress = get_method_n_ress(call_type);
722 for (i = 0; i < params; ++i) {
723 type *p_type = get_method_param_type(call_type, i);
725 if (is_compound_type(p_type))
730 for (i = 0; i < ress; ++i) {
731 type *r_type = get_method_res_type(call_type, i);
733 if (is_compound_type(r_type))
738 irg_walk_graph(called_graph, find_addr, NULL, &res);
743 int inline_method(ir_node *call, ir_graph *called_graph) {
745 ir_node *post_call, *post_bl;
747 ir_node *end, *end_bl;
751 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
754 irg_inline_property prop = get_irg_inline_property(called_graph);
756 if ( (prop != irg_inline_forced) &&
757 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
759 /* Do not inline variadic functions. */
760 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
763 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
764 get_method_n_params(get_Call_type(call)));
767 * currently, we cannot inline two cases:
768 * - call with compound arguments
769 * - graphs that take the address of a parameter
771 if (! can_inline(call, called_graph))
774 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
775 rem_opt = get_opt_optimize();
778 /* Handle graph state */
779 assert(get_irg_phase_state(current_ir_graph) != phase_building);
780 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
781 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
782 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
783 set_irg_outs_inconsistent(current_ir_graph);
784 set_irg_loopinfo_inconsistent(current_ir_graph);
785 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
787 /* -- Check preconditions -- */
788 assert(get_irn_op(call) == op_Call);
789 /* @@@ does not work for InterfaceIII.java after cgana
790 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
791 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
792 get_Call_type(call)));
794 assert(get_type_tpop(get_Call_type(call)) == type_method);
795 if (called_graph == current_ir_graph) {
796 set_optimize(rem_opt);
800 /* here we know we WILL inline, so inform the statistics */
801 hook_inline(call, called_graph);
803 /* -- Decide how to handle exception control flow: Is there a handler
804 for the Call node, or do we branch directly to End on an exception?
806 0 There is a handler.
808 2 Exception handling not represented in Firm. -- */
810 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
811 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
812 assert(get_irn_op(proj) == op_Proj);
813 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
814 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
816 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
817 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
818 else { exc_handling = 2; } /* !Mproj && !Xproj */
823 the procedure and later replaces the Start node of the called graph.
824 Post_call is the old Call node and collects the results of the called
825 graph. Both will end up being a tuple. -- */
826 post_bl = get_nodes_block(call);
827 set_irg_current_block(current_ir_graph, post_bl);
828 /* XxMxPxP of Start + parameter of Call */
829 in[pn_Start_X_initial_exec] = new_Jmp();
830 in[pn_Start_M] = get_Call_mem(call);
831 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
832 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
833 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
834 /* in[pn_Start_P_value_arg_base] = ??? */
835 pre_call = new_Tuple(5, in);
839 The new block gets the ins of the old block, pre_call and all its
840 predecessors and all Phi nodes. -- */
841 part_block(pre_call);
843 /* -- Prepare state for dead node elimination -- */
844 /* Visited flags in calling irg must be >= flag in called irg.
845 Else walker and arity computation will not work. */
846 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
847 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
848 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
849 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
850 /* Set pre_call as new Start node in link field of the start node of
851 calling graph and pre_calls block as new block for the start block
853 Further mark these nodes so that they are not visited by the
855 set_irn_link(get_irg_start(called_graph), pre_call);
856 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
857 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
858 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
859 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
860 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
862 /* Initialize for compaction of in arrays */
863 inc_irg_block_visited(current_ir_graph);
865 /* -- Replicate local entities of the called_graph -- */
866 /* copy the entities. */
867 called_frame = get_irg_frame_type(called_graph);
868 for (i = 0; i < get_class_n_members(called_frame); i++) {
869 entity *new_ent, *old_ent;
870 old_ent = get_class_member(called_frame, i);
871 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
872 set_entity_link(old_ent, new_ent);
875 /* visited is > than that of called graph. With this trick visited will
876 remain unchanged so that an outer walker, e.g., searching the call nodes
877 to inline, calling this inline will not visit the inlined nodes. */
878 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
880 /* -- Performing dead node elimination inlines the graph -- */
881 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
883 /* @@@ endless loops are not copied!! -- they should be, I think... */
884 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
885 get_irg_frame_type(called_graph));
887 /* Repair called_graph */
888 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
889 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
890 set_Block_block_visited(get_irg_start_block(called_graph), 0);
892 /* -- Merge the end of the inlined procedure with the call site -- */
893 /* We will turn the old Call node into a Tuple with the following
896 0: Phi of all Memories of Return statements.
897 1: Jmp from new Block that merges the control flow from all exception
898 predecessors of the old end block.
899 2: Tuple of all arguments.
900 3: Phi of Exception memories.
901 In case the old Call directly branches to End on an exception we don't
902 need the block merging all exceptions nor the Phi of the exception
906 /* -- Precompute some values -- */
907 end_bl = get_new_node(get_irg_end_block(called_graph));
908 end = get_new_node(get_irg_end(called_graph));
909 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
910 n_res = get_method_n_ress(get_Call_type(call));
912 res_pred = xmalloc (n_res * sizeof(*res_pred));
913 cf_pred = xmalloc (arity * sizeof(*res_pred));
915 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
917 /* -- archive keepalives -- */
918 irn_arity = get_irn_arity(end);
919 for (i = 0; i < irn_arity; i++)
920 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
922 /* The new end node will die. We need not free as the in array is on the obstack:
923 firm_copy_node only generated 'D' arrays. */
925 /* -- Replace Return nodes by Jump nodes. -- */
927 for (i = 0; i < arity; i++) {
929 ret = get_irn_n(end_bl, i);
930 if (get_irn_op(ret) == op_Return) {
931 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
935 set_irn_in(post_bl, n_ret, cf_pred);
937 /* -- Build a Tuple for all results of the method.
938 Add Phi node if there was more than one Return. -- */
939 turn_into_tuple(post_call, 4);
940 /* First the Memory-Phi */
942 for (i = 0; i < arity; i++) {
943 ret = get_irn_n(end_bl, i);
944 if (get_irn_op(ret) == op_Return) {
945 cf_pred[n_ret] = get_Return_mem(ret);
949 phi = new_Phi(n_ret, cf_pred, mode_M);
950 set_Tuple_pred(call, pn_Call_M_regular, phi);
951 /* Conserve Phi-list for further inlinings -- but might be optimized */
952 if (get_nodes_block(phi) == post_bl) {
953 set_irn_link(phi, get_irn_link(post_bl));
954 set_irn_link(post_bl, phi);
956 /* Now the real results */
958 for (j = 0; j < n_res; j++) {
960 for (i = 0; i < arity; i++) {
961 ret = get_irn_n(end_bl, i);
962 if (get_irn_op(ret) == op_Return) {
963 cf_pred[n_ret] = get_Return_res(ret, j);
968 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
972 /* Conserve Phi-list for further inlinings -- but might be optimized */
973 if (get_nodes_block(phi) == post_bl) {
974 set_irn_link(phi, get_irn_link(post_bl));
975 set_irn_link(post_bl, phi);
978 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
980 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
982 /* Finally the exception control flow.
983 We have two (three) possible situations:
984 First if the Call branches to an exception handler: We need to add a Phi node to
985 collect the memory containing the exception objects. Further we need
986 to add another block to get a correct representation of this Phi. To
987 this block we add a Jmp that resolves into the X output of the Call
988 when the Call is turned into a tuple.
989 Second the Call branches to End, the exception is not handled. Just
990 add all inlined exception branches to the End node.
991 Third: there is no Exception edge at all. Handle as case two. */
992 if (exc_handling == 0) {
994 for (i = 0; i < arity; i++) {
996 ret = get_irn_n(end_bl, i);
997 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
998 cf_pred[n_exc] = ret;
1003 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1004 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1005 /* The Phi for the memories with the exception objects */
1007 for (i = 0; i < arity; i++) {
1009 ret = skip_Proj(get_irn_n(end_bl, i));
1010 if (get_irn_op(ret) == op_Call) {
1011 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1013 } else if (is_fragile_op(ret)) {
1014 /* We rely that all cfops have the memory output at the same position. */
1015 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1017 } else if (get_irn_op(ret) == op_Raise) {
1018 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1022 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1024 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1025 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1028 ir_node *main_end_bl;
1029 int main_end_bl_arity;
1030 ir_node **end_preds;
1032 /* assert(exc_handling == 1 || no exceptions. ) */
1034 for (i = 0; i < arity; i++) {
1035 ir_node *ret = get_irn_n(end_bl, i);
1037 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1038 cf_pred[n_exc] = ret;
1042 main_end_bl = get_irg_end_block(current_ir_graph);
1043 main_end_bl_arity = get_irn_arity(main_end_bl);
1044 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1046 for (i = 0; i < main_end_bl_arity; ++i)
1047 end_preds[i] = get_irn_n(main_end_bl, i);
1048 for (i = 0; i < n_exc; ++i)
1049 end_preds[main_end_bl_arity + i] = cf_pred[i];
1050 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1051 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1052 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1058 #if 0 /* old. now better, correcter, faster implementation. */
1060 /* -- If the exception control flow from the inlined Call directly
1061 branched to the end block we now have the following control
1062 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1063 remove the Jmp along with it's empty block and add Jmp's
1064 predecessors as predecessors of this end block. No problem if
1065 there is no exception, because then branches Bad to End which
1067 @@@ can't we know this beforehand: by getting the Proj(1) from
1068 the Call link list and checking whether it goes to Proj. */
1069 /* find the problematic predecessor of the end block. */
1070 end_bl = get_irg_end_block(current_ir_graph);
1071 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1072 cf_op = get_Block_cfgpred(end_bl, i);
1073 if (get_irn_op(cf_op) == op_Proj) {
1074 cf_op = get_Proj_pred(cf_op);
1075 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1076 /* There are unoptimized tuples from inlineing before when no exc */
1077 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1078 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1079 assert(get_irn_op(cf_op) == op_Jmp);
1085 if (i < get_Block_n_cfgpreds(end_bl)) {
1086 bl = get_nodes_block(cf_op);
1087 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1088 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1089 for (j = 0; j < i; j++)
1090 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1091 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1092 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1093 for (j = j; j < arity; j++)
1094 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1095 set_irn_in(end_bl, arity, cf_pred);
1097 /* Remove the exception pred from post-call Tuple. */
1098 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1103 /* -- Turn cse back on. -- */
1104 set_optimize(rem_opt);
1109 /********************************************************************/
1110 /* Apply inlineing to small methods. */
1111 /********************************************************************/
1113 /* It makes no sense to inline too many calls in one procedure. Anyways,
1114 I didn't get a version with NEW_ARR_F to run. */
1115 #define MAX_INLINE 1024
1118 * environment for inlining small irgs
1120 typedef struct _inline_env_t {
1122 ir_node *calls[MAX_INLINE];
1126 * Returns the irg called from a Call node. If the irg is not
1127 * known, NULL is returned.
1129 static ir_graph *get_call_called_irg(ir_node *call) {
1131 ir_graph *called_irg = NULL;
1133 assert(get_irn_op(call) == op_Call);
1135 addr = get_Call_ptr(call);
1136 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1137 called_irg = get_entity_irg(get_SymConst_entity(addr));
1143 static void collect_calls(ir_node *call, void *env) {
1146 if (get_irn_op(call) != op_Call) return;
1148 addr = get_Call_ptr(call);
1150 if (get_irn_op(addr) == op_SymConst) {
1151 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1152 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1153 inline_env_t *ienv = (inline_env_t *)env;
1154 if (called_irg && ienv->pos < MAX_INLINE) {
1155 /* The Call node calls a locally defined method. Remember to inline. */
1156 ienv->calls[ienv->pos++] = call;
1163 * Inlines all small methods at call sites where the called address comes
1164 * from a Const node that references the entity representing the called
1166 * The size argument is a rough measure for the code size of the method:
1167 * Methods where the obstack containing the firm graph is smaller than
1170 void inline_small_irgs(ir_graph *irg, int size) {
1172 ir_graph *rem = current_ir_graph;
1173 inline_env_t env /* = {0, NULL}*/;
1175 if (!(get_opt_optimize() && get_opt_inline())) return;
1177 current_ir_graph = irg;
1178 /* Handle graph state */
1179 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1180 free_callee_info(current_ir_graph);
1182 /* Find Call nodes to inline.
1183 (We can not inline during a walk of the graph, as inlineing the same
1184 method several times changes the visited flag of the walked graph:
1185 after the first inlineing visited of the callee equals visited of
1186 the caller. With the next inlineing both are increased.) */
1188 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1190 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1191 /* There are calls to inline */
1192 collect_phiprojs(irg);
1193 for (i = 0; i < env.pos; i++) {
1195 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1196 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1197 (get_irg_inline_property(callee) == irg_inline_forced)) {
1198 inline_method(env.calls[i], callee);
1203 current_ir_graph = rem;
1207 * Environment for inlining irgs.
1210 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1211 int n_nodes_orig; /**< for statistics */
1212 eset *call_nodes; /**< All call nodes in this graph */
1214 int n_call_nodes_orig; /**< for statistics */
1215 int n_callers; /**< Number of known graphs that call this graphs. */
1216 int n_callers_orig; /**< for statistics */
1219 static inline_irg_env *new_inline_irg_env(void) {
1220 inline_irg_env *env = xmalloc(sizeof(*env));
1221 env->n_nodes = -2; /* uncount Start, End */
1222 env->n_nodes_orig = -2; /* uncount Start, End */
1223 env->call_nodes = eset_create();
1224 env->n_call_nodes = 0;
1225 env->n_call_nodes_orig = 0;
1227 env->n_callers_orig = 0;
1231 static void free_inline_irg_env(inline_irg_env *env) {
1232 eset_destroy(env->call_nodes);
1236 static void collect_calls2(ir_node *call, void *env) {
1237 inline_irg_env *x = (inline_irg_env *)env;
1238 ir_op *op = get_irn_op(call);
1241 /* count nodes in irg */
1242 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1247 if (op != op_Call) return;
1249 /* collect all call nodes */
1250 eset_insert(x->call_nodes, (void *)call);
1252 x->n_call_nodes_orig++;
1254 /* count all static callers */
1255 callee = get_call_called_irg(call);
1257 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1258 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1262 INLINE static int is_leave(ir_graph *irg) {
1263 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1266 INLINE static int is_smaller(ir_graph *callee, int size) {
1267 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1272 * Inlines small leave methods at call sites where the called address comes
1273 * from a Const node that references the entity representing the called
1275 * The size argument is a rough measure for the code size of the method:
1276 * Methods where the obstack containing the firm graph is smaller than
1279 void inline_leave_functions(int maxsize, int leavesize, int size) {
1280 inline_irg_env *env;
1281 int i, n_irgs = get_irp_n_irgs();
1282 ir_graph *rem = current_ir_graph;
1285 if (!(get_opt_optimize() && get_opt_inline())) return;
1287 /* extend all irgs by a temporary data structure for inlining. */
1288 for (i = 0; i < n_irgs; ++i)
1289 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1291 /* Precompute information in temporary data structure. */
1292 for (i = 0; i < n_irgs; ++i) {
1293 current_ir_graph = get_irp_irg(i);
1294 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1295 free_callee_info(current_ir_graph);
1297 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1298 get_irg_link(current_ir_graph));
1301 /* -- and now inline. -- */
1303 /* Inline leaves recursively -- we might construct new leaves. */
1304 while (did_inline) {
1307 for (i = 0; i < n_irgs; ++i) {
1309 int phiproj_computed = 0;
1311 current_ir_graph = get_irp_irg(i);
1312 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1314 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1317 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1318 callee = get_call_called_irg(call);
1320 if (env->n_nodes > maxsize) continue; // break;
1322 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1323 if (!phiproj_computed) {
1324 phiproj_computed = 1;
1325 collect_phiprojs(current_ir_graph);
1327 did_inline = inline_method(call, callee);
1330 /* Do some statistics */
1331 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1332 env->n_call_nodes --;
1333 env->n_nodes += callee_env->n_nodes;
1334 callee_env->n_callers--;
1341 /* inline other small functions. */
1342 for (i = 0; i < n_irgs; ++i) {
1345 int phiproj_computed = 0;
1347 current_ir_graph = get_irp_irg(i);
1348 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1350 /* we can not walk and change a set, nor remove from it.
1352 walkset = env->call_nodes;
1353 env->call_nodes = eset_create();
1354 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1357 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1358 callee = get_call_called_irg(call);
1361 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1362 (get_irg_inline_property(callee) == irg_inline_forced))) {
1363 if (!phiproj_computed) {
1364 phiproj_computed = 1;
1365 collect_phiprojs(current_ir_graph);
1367 if (inline_method(call, callee)) {
1368 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1369 env->n_call_nodes--;
1370 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1371 env->n_call_nodes += callee_env->n_call_nodes;
1372 env->n_nodes += callee_env->n_nodes;
1373 callee_env->n_callers--;
1376 eset_insert(env->call_nodes, call);
1379 eset_destroy(walkset);
1382 for (i = 0; i < n_irgs; ++i) {
1383 current_ir_graph = get_irp_irg(i);
1385 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1386 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1387 (env->n_callers_orig != env->n_callers))
1388 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1389 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1390 env->n_callers_orig, env->n_callers,
1391 get_entity_name(get_irg_entity(current_ir_graph)));
1393 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1396 current_ir_graph = rem;
1399 /*******************************************************************/
1400 /* Code Placement. Pins all floating nodes to a block where they */
1401 /* will be executed only if needed. */
1402 /*******************************************************************/
1405 * Find the earliest correct block for N. --- Place N into the
1406 * same Block as its dominance-deepest Input.
1409 place_floats_early(ir_node *n, pdeq *worklist)
1411 int i, start, irn_arity;
1413 /* we must not run into an infinite loop */
1414 assert (irn_not_visited(n));
1415 mark_irn_visited(n);
1417 /* Place floating nodes. */
1418 if (get_irn_pinned(n) == op_pin_state_floats) {
1420 ir_node *b = new_Bad(); /* The block to place this node in */
1421 int bad_recursion = is_Bad(get_nodes_block(n));
1423 assert(get_irn_op(n) != op_Block);
1425 if ((get_irn_op(n) == op_Const) ||
1426 (get_irn_op(n) == op_SymConst) ||
1428 (get_irn_op(n) == op_Unknown)) {
1429 /* These nodes will not be placed by the loop below. */
1430 b = get_irg_start_block(current_ir_graph);
1434 /* find the block for this node. */
1435 irn_arity = get_irn_arity(n);
1436 for (i = 0; i < irn_arity; i++) {
1437 ir_node *dep = get_irn_n(n, i);
1440 if ((irn_not_visited(dep))
1441 && (get_irn_pinned(dep) == op_pin_state_floats)) {
1442 place_floats_early(dep, worklist);
1446 * A node in the Bad block must stay in the bad block,
1447 * so don't compute a new block for it.
1452 /* Because all loops contain at least one op_pin_state_pinned node, now all
1453 our inputs are either op_pin_state_pinned or place_early has already
1454 been finished on them. We do not have any unfinished inputs! */
1455 dep_block = get_nodes_block(dep);
1456 if ((!is_Bad(dep_block)) &&
1457 (get_Block_dom_depth(dep_block) > depth)) {
1459 depth = get_Block_dom_depth(dep_block);
1461 /* Avoid that the node is placed in the Start block */
1462 if ((depth == 1) && (get_Block_dom_depth(get_nodes_block(n)) > 1)) {
1463 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1464 assert(b != get_irg_start_block(current_ir_graph));
1468 set_nodes_block(n, b);
1471 /* Add predecessors of non floating nodes on worklist. */
1472 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1473 irn_arity = get_irn_arity(n);
1474 for (i = start; i < irn_arity; i++) {
1475 ir_node *pred = get_irn_n(n, i);
1476 if (irn_not_visited(pred)) {
1477 pdeq_putr (worklist, pred);
1483 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1484 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1485 * places all floating nodes reachable from its argument through floating
1486 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1488 static INLINE void place_early(pdeq *worklist) {
1490 inc_irg_visited(current_ir_graph);
1492 /* this inits the worklist */
1493 place_floats_early(get_irg_end(current_ir_graph), worklist);
1495 /* Work the content of the worklist. */
1496 while (!pdeq_empty (worklist)) {
1497 ir_node *n = pdeq_getl (worklist);
1498 if (irn_not_visited(n)) place_floats_early(n, worklist);
1501 set_irg_outs_inconsistent(current_ir_graph);
1502 current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
1505 /** Compute the deepest common ancestor of block and dca. */
1506 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1509 if (!dca) return block;
1510 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1511 block = get_Block_idom(block);
1512 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1513 dca = get_Block_idom(dca);
1515 while (block != dca)
1516 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1521 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1522 * I.e., DCA is the block where we might place PRODUCER.
1523 * A data flow edge points from producer to consumer.
1526 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1528 ir_node *block = NULL;
1530 /* Compute the latest block into which we can place a node so that it is
1532 if (get_irn_op(consumer) == op_Phi) {
1533 /* our consumer is a Phi-node, the effective use is in all those
1534 blocks through which the Phi-node reaches producer */
1536 ir_node *phi_block = get_nodes_block(consumer);
1537 irn_arity = get_irn_arity(consumer);
1539 for (i = 0; i < irn_arity; i++) {
1540 if (get_irn_n(consumer, i) == producer) {
1541 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1543 block = calc_dca(block, new_block);
1547 assert(is_no_Block(consumer));
1548 block = get_nodes_block(consumer);
1551 /* Compute the deepest common ancestor of block and dca. */
1552 return calc_dca(dca, block);
1555 static INLINE int get_irn_loop_depth(ir_node *n) {
1556 return get_loop_depth(get_irn_loop(n));
1560 * Move n to a block with less loop depth than it's current block. The
1561 * new block must be dominated by early.
1564 move_out_of_loops (ir_node *n, ir_node *early)
1566 ir_node *best, *dca;
1570 /* Find the region deepest in the dominator tree dominating
1571 dca with the least loop nesting depth, but still dominated
1572 by our early placement. */
1573 dca = get_nodes_block(n);
1575 while (dca != early) {
1576 dca = get_Block_idom(dca);
1577 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1578 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1582 if (best != get_nodes_block(n)) {
1584 printf("Moving out of loop: "); DDMN(n);
1585 printf(" Outermost block: "); DDMN(early);
1586 printf(" Best block: "); DDMN(best);
1587 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1589 set_nodes_block(n, best);
1594 * Find the latest legal block for N and place N into the
1595 * `optimal' Block between the latest and earliest legal block.
1596 * The `optimal' block is the dominance-deepest block of those
1597 * with the least loop-nesting-depth. This places N out of as many
1598 * loops as possible and then makes it as control dependant as
1602 place_floats_late(ir_node *n, pdeq *worklist)
1607 assert (irn_not_visited(n)); /* no multiple placement */
1609 mark_irn_visited(n);
1611 /* no need to place block nodes, control nodes are already placed. */
1612 if ((get_irn_op(n) != op_Block) &&
1614 (get_irn_mode(n) != mode_X)) {
1615 /* Remember the early placement of this block to move it
1616 out of loop no further than the early placement. */
1617 early = get_nodes_block(n);
1619 /* Do not move code not reachable from Start. For
1620 * these we could not compute dominator information. */
1621 if (is_Bad(early) || get_Block_dom_depth(early) == -1)
1624 /* Assure that our users are all placed, except the Phi-nodes.
1625 --- Each data flow cycle contains at least one Phi-node. We
1626 have to break the `user has to be placed before the
1627 producer' dependence cycle and the Phi-nodes are the
1628 place to do so, because we need to base our placement on the
1629 final region of our users, which is OK with Phi-nodes, as they
1630 are op_pin_state_pinned, and they never have to be placed after a
1631 producer of one of their inputs in the same block anyway. */
1632 for (i = 0; i < get_irn_n_outs(n); i++) {
1633 ir_node *succ = get_irn_out(n, i);
1634 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1635 place_floats_late(succ, worklist);
1638 /* We have to determine the final block of this node... except for
1640 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1641 (get_irn_op(n) != op_Const) &&
1642 (get_irn_op(n) != op_SymConst)) {
1643 ir_node *dca = NULL; /* deepest common ancestor in the
1644 dominator tree of all nodes'
1645 blocks depending on us; our final
1646 placement has to dominate DCA. */
1647 for (i = 0; i < get_irn_n_outs(n); i++) {
1648 ir_node *out = get_irn_out(n, i);
1649 /* ignore if out is in dead code */
1650 ir_node *outbl = get_nodes_block(out);
1651 if (is_Bad(outbl) || get_Block_dom_depth(outbl) == -1)
1653 dca = consumer_dom_dca (dca, out, n);
1656 set_nodes_block(n, dca);
1658 move_out_of_loops (n, early);
1660 /* else all outs are in dead code */
1664 /* Add predecessors of all non-floating nodes on list. (Those of floating
1665 nodes are placeded already and therefore are marked.) */
1666 for (i = 0; i < get_irn_n_outs(n); i++) {
1667 ir_node *succ = get_irn_out(n, i);
1668 if (irn_not_visited(get_irn_out(n, i))) {
1669 pdeq_putr (worklist, succ);
1674 static INLINE void place_late(pdeq *worklist) {
1676 inc_irg_visited(current_ir_graph);
1678 /* This fills the worklist initially. */
1679 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1681 /* And now empty the worklist again... */
1682 while (!pdeq_empty (worklist)) {
1683 ir_node *n = pdeq_getl (worklist);
1684 if (irn_not_visited(n)) place_floats_late(n, worklist);
1688 void place_code(ir_graph *irg) {
1690 ir_graph *rem = current_ir_graph;
1692 current_ir_graph = irg;
1694 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1696 /* Handle graph state */
1697 assert(get_irg_phase_state(irg) != phase_building);
1698 if (get_irg_dom_state(irg) != dom_consistent)
1701 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1702 free_loop_information(irg);
1703 construct_backedges(irg);
1706 /* Place all floating nodes as early as possible. This guarantees
1707 a legal code placement. */
1708 worklist = new_pdeq();
1709 place_early(worklist);
1711 /* place_early invalidates the outs, place_late needs them. */
1713 /* Now move the nodes down in the dominator tree. This reduces the
1714 unnecessary executions of the node. */
1715 place_late(worklist);
1717 set_irg_outs_inconsistent(current_ir_graph);
1718 set_irg_loopinfo_inconsistent(current_ir_graph);
1720 current_ir_graph = rem;
1724 * Called by walker of remove_critical_cf_edges().
1726 * Place an empty block to an edge between a blocks of multiple
1727 * predecessors and a block of multiple successors.
1730 * @param env Environment of walker. This field is unused and has
1733 static void walk_critical_cf_edges(ir_node *n, void *env) {
1735 ir_node *pre, *block, **in, *jmp;
1737 /* Block has multiple predecessors */
1738 if ((op_Block == get_irn_op(n)) &&
1739 (get_irn_arity(n) > 1)) {
1740 arity = get_irn_arity(n);
1742 if (n == get_irg_end_block(current_ir_graph))
1743 return; /* No use to add a block here. */
1745 for (i=0; i<arity; i++) {
1746 pre = get_irn_n(n, i);
1747 /* Predecessor has multiple successors. Insert new flow edge */
1748 if ((NULL != pre) &&
1749 (op_Proj == get_irn_op(pre)) &&
1750 op_Raise != get_irn_op(skip_Proj(pre))) {
1752 /* set predecessor array for new block */
1753 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1754 /* set predecessor of new block */
1756 block = new_Block(1, in);
1757 /* insert new jmp node to new block */
1758 set_cur_block(block);
1761 /* set successor of new block */
1762 set_irn_n(n, i, jmp);
1764 } /* predecessor has multiple successors */
1765 } /* for all predecessors */
1766 } /* n is a block */
1769 void remove_critical_cf_edges(ir_graph *irg) {
1770 if (get_opt_critical_edges())
1771 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);