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.
213 void copy_node (ir_node *n, void *env) {
216 opcode op = get_irn_opcode(n);
217 int copy_node_nr = env != NULL;
219 /* The end node looses it's flexible in array. This doesn't matter,
220 as dead node elimination builds End by hand, inlineing doesn't use
222 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
225 /* node copied already */
227 } else if (op == iro_Block) {
229 new_arity = compute_new_arity(n);
230 n->attr.block.graph_arr = NULL;
232 block = get_nodes_block(n);
233 if (get_irn_opcode(n) == iro_Phi) {
234 new_arity = compute_new_arity(block);
236 new_arity = get_irn_arity(n);
239 nn = new_ir_node(get_irn_dbg_info(n),
246 /* Copy the attributes. These might point to additional data. If this
247 was allocated on the old obstack the pointers now are dangling. This
248 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
249 copy_node_attr(n, nn);
250 new_backedge_info(nn);
255 /* for easier debugging, we want to copy the node numbers too */
256 nn->node_nr = n->node_nr;
260 /* printf("\n old node: "); DDMSG2(n);
261 printf(" new node: "); DDMSG2(nn); */
265 * Copies new predecessors of old node to new node remembered in link.
266 * Spare the Bad predecessors of Phi and Block nodes.
269 copy_preds (ir_node *n, void *env) {
273 nn = get_new_node(n);
275 /* printf("\n old node: "); DDMSG2(n);
276 printf(" new node: "); DDMSG2(nn);
277 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
279 if (get_irn_opcode(n) == iro_Block) {
280 /* Don't copy Bad nodes. */
282 irn_arity = get_irn_arity(n);
283 for (i = 0; i < irn_arity; i++)
284 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
285 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
286 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
289 /* repair the block visited flag from above misuse. Repair it in both
290 graphs so that the old one can still be used. */
291 set_Block_block_visited(nn, 0);
292 set_Block_block_visited(n, 0);
293 /* Local optimization could not merge two subsequent blocks if
294 in array contained Bads. Now it's possible.
295 We don't call optimize_in_place as it requires
296 that the fields in ir_graph are set properly. */
297 if ((get_opt_control_flow_straightening()) &&
298 (get_Block_n_cfgpreds(nn) == 1) &&
299 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
300 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
302 /* Jmp jumps into the block it is in -- deal self cycle. */
303 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
304 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
309 } else if (get_irn_opcode(n) == iro_Phi) {
310 /* Don't copy node if corresponding predecessor in block is Bad.
311 The Block itself should not be Bad. */
312 block = get_nodes_block(n);
313 set_irn_n (nn, -1, get_new_node(block));
315 irn_arity = get_irn_arity(n);
316 for (i = 0; i < irn_arity; i++)
317 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
318 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
319 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
322 /* If the pre walker reached this Phi after the post walker visited the
323 block block_visited is > 0. */
324 set_Block_block_visited(get_nodes_block(n), 0);
325 /* Compacting the Phi's ins might generate Phis with only one
327 if (get_irn_arity(n) == 1)
328 exchange(n, get_irn_n(n, 0));
330 irn_arity = get_irn_arity(n);
331 for (i = -1; i < irn_arity; i++)
332 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
334 /* Now the new node is complete. We can add it to the hash table for cse.
335 @@@ inlinening aborts if we identify End. Why? */
336 if(get_irn_op(nn) != op_End)
337 add_identities (current_ir_graph->value_table, nn);
341 * Copies the graph recursively, compacts the keepalive of the end node.
343 * @param copy_node_nr If non-zero, the node number will be copied
346 copy_graph (int copy_node_nr) {
347 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
348 ir_node *ka; /* keep alive */
351 oe = get_irg_end(current_ir_graph);
352 /* copy the end node by hand, allocate dynamic in array! */
353 ne = new_ir_node(get_irn_dbg_info(oe),
360 /* Copy the attributes. Well, there might be some in the future... */
361 copy_node_attr(oe, ne);
362 set_new_node(oe, ne);
364 /* copy the Bad node */
365 ob = get_irg_bad(current_ir_graph);
366 nb = new_ir_node(get_irn_dbg_info(ob),
373 set_new_node(ob, nb);
375 /* copy the NoMem node */
376 om = get_irg_no_mem(current_ir_graph);
377 nm = new_ir_node(get_irn_dbg_info(om),
384 set_new_node(om, nm);
386 /* copy the live nodes */
387 irg_walk(get_nodes_block(oe), copy_node, copy_preds, (void *)copy_node_nr);
388 /* copy_preds for the end node ... */
389 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
391 /*- ... and now the keep alives. -*/
392 /* First pick the not marked block nodes and walk them. We must pick these
393 first as else we will oversee blocks reachable from Phis. */
394 irn_arity = get_irn_arity(oe);
395 for (i = 0; i < irn_arity; i++) {
396 ka = get_irn_intra_n(oe, i);
397 if ((get_irn_op(ka) == op_Block) &&
398 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
399 /* We must keep the block alive and copy everything reachable */
400 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
401 irg_walk(ka, copy_node, copy_preds, (void *)copy_node_nr);
402 add_End_keepalive(ne, get_new_node(ka));
406 /* Now pick the Phis. Here we will keep all! */
407 irn_arity = get_irn_arity(oe);
408 for (i = 0; i < irn_arity; i++) {
409 ka = get_irn_intra_n(oe, i);
410 if ((get_irn_op(ka) == op_Phi)) {
411 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
412 /* We didn't copy the Phi yet. */
413 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
414 irg_walk(ka, copy_node, copy_preds, (void *)copy_node_nr);
416 add_End_keepalive(ne, get_new_node(ka));
420 /* start block sometimes only reached after keep alives */
421 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
422 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
426 * Copies the graph reachable from current_ir_graph->end to the obstack
427 * in current_ir_graph and fixes the environment.
428 * Then fixes the fields in current_ir_graph containing nodes of the
431 * @param copy_node_nr If non-zero, the node number will be copied
434 copy_graph_env (int copy_node_nr) {
436 /* Not all nodes remembered in current_ir_graph might be reachable
437 from the end node. Assure their link is set to NULL, so that
438 we can test whether new nodes have been computed. */
439 set_irn_link(get_irg_frame (current_ir_graph), NULL);
440 set_irn_link(get_irg_globals (current_ir_graph), NULL);
441 set_irn_link(get_irg_args (current_ir_graph), NULL);
442 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
443 set_irn_link(get_irg_no_mem (current_ir_graph), NULL);
445 /* we use the block walk flag for removing Bads from Blocks ins. */
446 inc_irg_block_visited(current_ir_graph);
449 copy_graph(copy_node_nr);
451 /* fix the fields in current_ir_graph */
452 old_end = get_irg_end(current_ir_graph);
453 set_irg_end (current_ir_graph, get_new_node(old_end));
454 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
455 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
457 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
458 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
459 copy_node (get_irg_frame(current_ir_graph), (void *)copy_node_nr);
460 copy_preds(get_irg_frame(current_ir_graph), NULL);
462 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
463 copy_node (get_irg_globals(current_ir_graph), (void *)copy_node_nr);
464 copy_preds(get_irg_globals(current_ir_graph), NULL);
466 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
467 copy_node (get_irg_initial_mem(current_ir_graph), (void *)copy_node_nr);
468 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
470 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
471 copy_node (get_irg_args(current_ir_graph), (void *)copy_node_nr);
472 copy_preds(get_irg_args(current_ir_graph), NULL);
474 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
476 set_irg_start_block(current_ir_graph,
477 get_new_node(get_irg_start_block(current_ir_graph)));
478 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
479 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
480 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
481 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
483 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
484 copy_node(get_irg_bad(current_ir_graph), (void *)copy_node_nr);
485 copy_preds(get_irg_bad(current_ir_graph), NULL);
487 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
489 if (get_irn_link(get_irg_no_mem(current_ir_graph)) == NULL) {
490 copy_node(get_irg_no_mem(current_ir_graph), (void *)copy_node_nr);
491 copy_preds(get_irg_no_mem(current_ir_graph), NULL);
493 set_irg_no_mem(current_ir_graph, get_new_node(get_irg_no_mem(current_ir_graph)));
497 * Copies all reachable nodes to a new obstack. Removes bad inputs
498 * from block nodes and the corresponding inputs from Phi nodes.
499 * Merges single exit blocks with single entry blocks and removes
501 * Adds all new nodes to a new hash table for cse. Does not
502 * perform cse, so the hash table might contain common subexpressions.
505 dead_node_elimination(ir_graph *irg) {
507 int rem_ipview = get_interprocedural_view();
508 struct obstack *graveyard_obst = NULL;
509 struct obstack *rebirth_obst = NULL;
511 /* inform statistics that we started a dead-node elimination run */
512 hook_dead_node_elim_start(irg);
514 /* Remember external state of current_ir_graph. */
515 rem = current_ir_graph;
516 current_ir_graph = irg;
517 set_interprocedural_view(false);
519 /* Handle graph state */
520 assert(get_irg_phase_state(current_ir_graph) != phase_building);
521 free_callee_info(current_ir_graph);
522 free_outs(current_ir_graph);
523 /* @@@ so far we loose loops when copying */
524 free_loop_information(current_ir_graph);
526 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
528 /* A quiet place, where the old obstack can rest in peace,
529 until it will be cremated. */
530 graveyard_obst = irg->obst;
532 /* A new obstack, where the reachable nodes will be copied to. */
533 rebirth_obst = xmalloc (sizeof(*rebirth_obst));
534 current_ir_graph->obst = rebirth_obst;
535 obstack_init (current_ir_graph->obst);
537 /* We also need a new hash table for cse */
538 del_identities (irg->value_table);
539 irg->value_table = new_identities ();
541 /* Copy the graph from the old to the new obstack */
544 /* Free memory from old unoptimized obstack */
545 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
546 xfree (graveyard_obst); /* ... then free it. */
549 /* inform statistics that the run is over */
550 hook_dead_node_elim_stop(irg);
552 current_ir_graph = rem;
553 set_interprocedural_view(rem_ipview);
557 * Relink bad predeseccors of a block and store the old in array to the
558 * link field. This function is called by relink_bad_predecessors().
559 * The array of link field starts with the block operand at position 0.
560 * If block has bad predecessors, create a new in array without bad preds.
561 * Otherwise let in array untouched.
563 static void relink_bad_block_predecessors(ir_node *n, void *env) {
564 ir_node **new_in, *irn;
565 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
567 /* if link field of block is NULL, look for bad predecessors otherwise
568 this is allready done */
569 if (get_irn_op(n) == op_Block &&
570 get_irn_link(n) == NULL) {
572 /* save old predecessors in link field (position 0 is the block operand)*/
573 set_irn_link(n, (void *)get_irn_in(n));
575 /* count predecessors without bad nodes */
576 old_irn_arity = get_irn_arity(n);
577 for (i = 0; i < old_irn_arity; i++)
578 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
580 /* arity changing: set new predecessors without bad nodes */
581 if (new_irn_arity < old_irn_arity) {
582 /* Get new predecessor array. We do not resize the array, as we must
583 keep the old one to update Phis. */
584 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
586 /* set new predeseccors in array */
589 for (i = 0; i < old_irn_arity; i++) {
590 irn = get_irn_n(n, i);
592 new_in[new_irn_n] = irn;
593 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
597 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
598 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
601 } /* ir node has bad predecessors */
603 } /* Block is not relinked */
607 * Relinks Bad predecesors from Bocks and Phis called by walker
608 * remove_bad_predecesors(). If n is a Block, call
609 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
610 * function of Phi's Block. If this block has bad predecessors, relink preds
613 static void relink_bad_predecessors(ir_node *n, void *env) {
614 ir_node *block, **old_in;
615 int i, old_irn_arity, new_irn_arity;
617 /* relink bad predeseccors of a block */
618 if (get_irn_op(n) == op_Block)
619 relink_bad_block_predecessors(n, env);
621 /* If Phi node relink its block and its predecessors */
622 if (get_irn_op(n) == op_Phi) {
624 /* Relink predeseccors of phi's block */
625 block = get_nodes_block(n);
626 if (get_irn_link(block) == NULL)
627 relink_bad_block_predecessors(block, env);
629 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
630 old_irn_arity = ARR_LEN(old_in);
632 /* Relink Phi predeseccors if count of predeseccors changed */
633 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
634 /* set new predeseccors in array
635 n->in[0] remains the same block */
637 for(i = 1; i < old_irn_arity; i++)
638 if (!is_Bad((ir_node *)old_in[i])) {
639 n->in[new_irn_arity] = n->in[i];
640 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
644 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
645 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
648 } /* n is a Phi node */
652 * Removes Bad Bad predecesors from Blocks and the corresponding
653 * inputs to Phi nodes as in dead_node_elimination but without
655 * On walking up set the link field to NULL, on walking down call
656 * relink_bad_predecessors() (This function stores the old in array
657 * to the link field and sets a new in array if arity of predecessors
660 void remove_bad_predecessors(ir_graph *irg) {
661 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
665 /*--------------------------------------------------------------------*/
666 /* Funcionality for inlining */
667 /*--------------------------------------------------------------------*/
670 * Copy node for inlineing. Updates attributes that change when
671 * inlineing but not for dead node elimination.
673 * Copies the node by calling copy_node and then updates the entity if
674 * it's a local one. env must be a pointer of the frame type of the
675 * inlined procedure. The new entities must be in the link field of
679 copy_node_inline (ir_node *n, void *env) {
681 type *frame_tp = (type *)env;
684 if (get_irn_op(n) == op_Sel) {
685 new = get_new_node (n);
686 assert(get_irn_op(new) == op_Sel);
687 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
688 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
690 } else if (get_irn_op(n) == op_Block) {
691 new = get_new_node (n);
692 new->attr.block.irg = current_ir_graph;
696 static void find_addr(ir_node *node, void *env)
698 if (get_irn_opcode(node) == iro_Proj) {
699 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
705 * currently, we cannot inline two cases:
706 * - call with compound arguments
707 * - graphs that take the address of a parameter
709 * check these conditions here
711 static int can_inline(ir_node *call, ir_graph *called_graph)
713 type *call_type = get_Call_type(call);
714 int params, ress, i, res;
715 assert(is_Method_type(call_type));
717 params = get_method_n_params(call_type);
718 ress = get_method_n_ress(call_type);
721 for (i = 0; i < params; ++i) {
722 type *p_type = get_method_param_type(call_type, i);
724 if (is_compound_type(p_type))
729 for (i = 0; i < ress; ++i) {
730 type *r_type = get_method_res_type(call_type, i);
732 if (is_compound_type(r_type))
737 irg_walk_graph(called_graph, find_addr, NULL, &res);
742 int inline_method(ir_node *call, ir_graph *called_graph) {
744 ir_node *post_call, *post_bl;
746 ir_node *end, *end_bl;
750 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
753 irg_inline_property prop = get_irg_inline_property(called_graph);
755 if ( (prop != irg_inline_forced) &&
756 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
758 /* Do not inline variadic functions. */
759 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
762 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
763 get_method_n_params(get_Call_type(call)));
766 * currently, we cannot inline two cases:
767 * - call with compound arguments
768 * - graphs that take the address of a parameter
770 if (! can_inline(call, called_graph))
773 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
774 rem_opt = get_opt_optimize();
777 /* Handle graph state */
778 assert(get_irg_phase_state(current_ir_graph) != phase_building);
779 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
780 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
781 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
782 set_irg_outs_inconsistent(current_ir_graph);
783 set_irg_loopinfo_inconsistent(current_ir_graph);
784 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
786 /* -- Check preconditions -- */
787 assert(get_irn_op(call) == op_Call);
788 /* @@@ does not work for InterfaceIII.java after cgana
789 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
790 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
791 get_Call_type(call)));
793 assert(get_type_tpop(get_Call_type(call)) == type_method);
794 if (called_graph == current_ir_graph) {
795 set_optimize(rem_opt);
799 /* here we know we WILL inline, so inform the statistics */
800 hook_inline(call, called_graph);
802 /* -- Decide how to handle exception control flow: Is there a handler
803 for the Call node, or do we branch directly to End on an exception?
805 0 There is a handler.
807 2 Exception handling not represented in Firm. -- */
809 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
810 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
811 assert(get_irn_op(proj) == op_Proj);
812 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
813 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
815 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
816 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
817 else { exc_handling = 2; } /* !Mproj && !Xproj */
822 the procedure and later replaces the Start node of the called graph.
823 Post_call is the old Call node and collects the results of the called
824 graph. Both will end up being a tuple. -- */
825 post_bl = get_nodes_block(call);
826 set_irg_current_block(current_ir_graph, post_bl);
827 /* XxMxPxP of Start + parameter of Call */
828 in[pn_Start_X_initial_exec] = new_Jmp();
829 in[pn_Start_M] = get_Call_mem(call);
830 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
831 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
832 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
833 /* in[pn_Start_P_value_arg_base] = ??? */
834 pre_call = new_Tuple(5, in);
838 The new block gets the ins of the old block, pre_call and all its
839 predecessors and all Phi nodes. -- */
840 part_block(pre_call);
842 /* -- Prepare state for dead node elimination -- */
843 /* Visited flags in calling irg must be >= flag in called irg.
844 Else walker and arity computation will not work. */
845 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
846 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
847 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
848 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
849 /* Set pre_call as new Start node in link field of the start node of
850 calling graph and pre_calls block as new block for the start block
852 Further mark these nodes so that they are not visited by the
854 set_irn_link(get_irg_start(called_graph), pre_call);
855 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
856 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
857 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
858 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
859 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
861 /* Initialize for compaction of in arrays */
862 inc_irg_block_visited(current_ir_graph);
864 /* -- Replicate local entities of the called_graph -- */
865 /* copy the entities. */
866 called_frame = get_irg_frame_type(called_graph);
867 for (i = 0; i < get_class_n_members(called_frame); i++) {
868 entity *new_ent, *old_ent;
869 old_ent = get_class_member(called_frame, i);
870 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
871 set_entity_link(old_ent, new_ent);
874 /* visited is > than that of called graph. With this trick visited will
875 remain unchanged so that an outer walker, e.g., searching the call nodes
876 to inline, calling this inline will not visit the inlined nodes. */
877 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
879 /* -- Performing dead node elimination inlines the graph -- */
880 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
882 /* @@@ endless loops are not copied!! -- they should be, I think... */
883 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
884 get_irg_frame_type(called_graph));
886 /* Repair called_graph */
887 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
888 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
889 set_Block_block_visited(get_irg_start_block(called_graph), 0);
891 /* -- Merge the end of the inlined procedure with the call site -- */
892 /* We will turn the old Call node into a Tuple with the following
895 0: Phi of all Memories of Return statements.
896 1: Jmp from new Block that merges the control flow from all exception
897 predecessors of the old end block.
898 2: Tuple of all arguments.
899 3: Phi of Exception memories.
900 In case the old Call directly branches to End on an exception we don't
901 need the block merging all exceptions nor the Phi of the exception
905 /* -- Precompute some values -- */
906 end_bl = get_new_node(get_irg_end_block(called_graph));
907 end = get_new_node(get_irg_end(called_graph));
908 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
909 n_res = get_method_n_ress(get_Call_type(call));
911 res_pred = xmalloc (n_res * sizeof(*res_pred));
912 cf_pred = xmalloc (arity * sizeof(*res_pred));
914 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
916 /* -- archive keepalives -- */
917 irn_arity = get_irn_arity(end);
918 for (i = 0; i < irn_arity; i++)
919 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
921 /* The new end node will die. We need not free as the in array is on the obstack:
922 copy_node only generated 'D' arrays. */
924 /* -- Replace Return nodes by Jump nodes. -- */
926 for (i = 0; i < arity; i++) {
928 ret = get_irn_n(end_bl, i);
929 if (get_irn_op(ret) == op_Return) {
930 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
934 set_irn_in(post_bl, n_ret, cf_pred);
936 /* -- Build a Tuple for all results of the method.
937 Add Phi node if there was more than one Return. -- */
938 turn_into_tuple(post_call, 4);
939 /* First the Memory-Phi */
941 for (i = 0; i < arity; i++) {
942 ret = get_irn_n(end_bl, i);
943 if (get_irn_op(ret) == op_Return) {
944 cf_pred[n_ret] = get_Return_mem(ret);
948 phi = new_Phi(n_ret, cf_pred, mode_M);
949 set_Tuple_pred(call, pn_Call_M_regular, phi);
950 /* Conserve Phi-list for further inlinings -- but might be optimized */
951 if (get_nodes_block(phi) == post_bl) {
952 set_irn_link(phi, get_irn_link(post_bl));
953 set_irn_link(post_bl, phi);
955 /* Now the real results */
957 for (j = 0; j < n_res; j++) {
959 for (i = 0; i < arity; i++) {
960 ret = get_irn_n(end_bl, i);
961 if (get_irn_op(ret) == op_Return) {
962 cf_pred[n_ret] = get_Return_res(ret, j);
967 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
971 /* Conserve Phi-list for further inlinings -- but might be optimized */
972 if (get_nodes_block(phi) == post_bl) {
973 set_irn_link(phi, get_irn_link(post_bl));
974 set_irn_link(post_bl, phi);
977 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
979 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
981 /* Finally the exception control flow.
982 We have two (three) possible situations:
983 First if the Call branches to an exception handler: We need to add a Phi node to
984 collect the memory containing the exception objects. Further we need
985 to add another block to get a correct representation of this Phi. To
986 this block we add a Jmp that resolves into the X output of the Call
987 when the Call is turned into a tuple.
988 Second the Call branches to End, the exception is not handled. Just
989 add all inlined exception branches to the End node.
990 Third: there is no Exception edge at all. Handle as case two. */
991 if (exc_handling == 0) {
993 for (i = 0; i < arity; i++) {
995 ret = get_irn_n(end_bl, i);
996 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
997 cf_pred[n_exc] = ret;
1002 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1003 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1004 /* The Phi for the memories with the exception objects */
1006 for (i = 0; i < arity; i++) {
1008 ret = skip_Proj(get_irn_n(end_bl, i));
1009 if (get_irn_op(ret) == op_Call) {
1010 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1012 } else if (is_fragile_op(ret)) {
1013 /* We rely that all cfops have the memory output at the same position. */
1014 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1016 } else if (get_irn_op(ret) == op_Raise) {
1017 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1021 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1023 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1024 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1027 ir_node *main_end_bl;
1028 int main_end_bl_arity;
1029 ir_node **end_preds;
1031 /* assert(exc_handling == 1 || no exceptions. ) */
1033 for (i = 0; i < arity; i++) {
1034 ir_node *ret = get_irn_n(end_bl, i);
1036 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1037 cf_pred[n_exc] = ret;
1041 main_end_bl = get_irg_end_block(current_ir_graph);
1042 main_end_bl_arity = get_irn_arity(main_end_bl);
1043 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1045 for (i = 0; i < main_end_bl_arity; ++i)
1046 end_preds[i] = get_irn_n(main_end_bl, i);
1047 for (i = 0; i < n_exc; ++i)
1048 end_preds[main_end_bl_arity + i] = cf_pred[i];
1049 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1050 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1051 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1057 #if 0 /* old. now better, correcter, faster implementation. */
1059 /* -- If the exception control flow from the inlined Call directly
1060 branched to the end block we now have the following control
1061 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1062 remove the Jmp along with it's empty block and add Jmp's
1063 predecessors as predecessors of this end block. No problem if
1064 there is no exception, because then branches Bad to End which
1066 @@@ can't we know this beforehand: by getting the Proj(1) from
1067 the Call link list and checking whether it goes to Proj. */
1068 /* find the problematic predecessor of the end block. */
1069 end_bl = get_irg_end_block(current_ir_graph);
1070 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1071 cf_op = get_Block_cfgpred(end_bl, i);
1072 if (get_irn_op(cf_op) == op_Proj) {
1073 cf_op = get_Proj_pred(cf_op);
1074 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1075 /* There are unoptimized tuples from inlineing before when no exc */
1076 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1077 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1078 assert(get_irn_op(cf_op) == op_Jmp);
1084 if (i < get_Block_n_cfgpreds(end_bl)) {
1085 bl = get_nodes_block(cf_op);
1086 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1087 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1088 for (j = 0; j < i; j++)
1089 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1090 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1091 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1092 for (j = j; j < arity; j++)
1093 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1094 set_irn_in(end_bl, arity, cf_pred);
1096 /* Remove the exception pred from post-call Tuple. */
1097 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1102 /* -- Turn cse back on. -- */
1103 set_optimize(rem_opt);
1108 /********************************************************************/
1109 /* Apply inlineing to small methods. */
1110 /********************************************************************/
1112 /* It makes no sense to inline too many calls in one procedure. Anyways,
1113 I didn't get a version with NEW_ARR_F to run. */
1114 #define MAX_INLINE 1024
1117 * environment for inlining small irgs
1119 typedef struct _inline_env_t {
1121 ir_node *calls[MAX_INLINE];
1125 * Returns the irg called from a Call node. If the irg is not
1126 * known, NULL is returned.
1128 static ir_graph *get_call_called_irg(ir_node *call) {
1130 ir_graph *called_irg = NULL;
1132 assert(get_irn_op(call) == op_Call);
1134 addr = get_Call_ptr(call);
1135 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1136 called_irg = get_entity_irg(get_SymConst_entity(addr));
1142 static void collect_calls(ir_node *call, void *env) {
1145 if (get_irn_op(call) != op_Call) return;
1147 addr = get_Call_ptr(call);
1149 if (get_irn_op(addr) == op_SymConst) {
1150 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1151 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1152 inline_env_t *ienv = (inline_env_t *)env;
1153 if (called_irg && ienv->pos < MAX_INLINE) {
1154 /* The Call node calls a locally defined method. Remember to inline. */
1155 ienv->calls[ienv->pos++] = call;
1162 * Inlines all small methods at call sites where the called address comes
1163 * from a Const node that references the entity representing the called
1165 * The size argument is a rough measure for the code size of the method:
1166 * Methods where the obstack containing the firm graph is smaller than
1169 void inline_small_irgs(ir_graph *irg, int size) {
1171 ir_graph *rem = current_ir_graph;
1172 inline_env_t env /* = {0, NULL}*/;
1174 if (!(get_opt_optimize() && get_opt_inline())) return;
1176 current_ir_graph = irg;
1177 /* Handle graph state */
1178 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1179 free_callee_info(current_ir_graph);
1181 /* Find Call nodes to inline.
1182 (We can not inline during a walk of the graph, as inlineing the same
1183 method several times changes the visited flag of the walked graph:
1184 after the first inlineing visited of the callee equals visited of
1185 the caller. With the next inlineing both are increased.) */
1187 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1189 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1190 /* There are calls to inline */
1191 collect_phiprojs(irg);
1192 for (i = 0; i < env.pos; i++) {
1194 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1195 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1196 (get_irg_inline_property(callee) == irg_inline_forced)) {
1197 inline_method(env.calls[i], callee);
1202 current_ir_graph = rem;
1206 * Environment for inlining irgs.
1209 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1210 int n_nodes_orig; /**< for statistics */
1211 eset *call_nodes; /**< All call nodes in this graph */
1213 int n_call_nodes_orig; /**< for statistics */
1214 int n_callers; /**< Number of known graphs that call this graphs. */
1215 int n_callers_orig; /**< for statistics */
1218 static inline_irg_env *new_inline_irg_env(void) {
1219 inline_irg_env *env = xmalloc(sizeof(*env));
1220 env->n_nodes = -2; /* uncount Start, End */
1221 env->n_nodes_orig = -2; /* uncount Start, End */
1222 env->call_nodes = eset_create();
1223 env->n_call_nodes = 0;
1224 env->n_call_nodes_orig = 0;
1226 env->n_callers_orig = 0;
1230 static void free_inline_irg_env(inline_irg_env *env) {
1231 eset_destroy(env->call_nodes);
1235 static void collect_calls2(ir_node *call, void *env) {
1236 inline_irg_env *x = (inline_irg_env *)env;
1237 ir_op *op = get_irn_op(call);
1240 /* count nodes in irg */
1241 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1246 if (op != op_Call) return;
1248 /* collect all call nodes */
1249 eset_insert(x->call_nodes, (void *)call);
1251 x->n_call_nodes_orig++;
1253 /* count all static callers */
1254 callee = get_call_called_irg(call);
1256 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1257 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1261 INLINE static int is_leave(ir_graph *irg) {
1262 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1265 INLINE static int is_smaller(ir_graph *callee, int size) {
1266 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1271 * Inlines small leave methods at call sites where the called address comes
1272 * from a Const node that references the entity representing the called
1274 * The size argument is a rough measure for the code size of the method:
1275 * Methods where the obstack containing the firm graph is smaller than
1278 void inline_leave_functions(int maxsize, int leavesize, int size) {
1279 inline_irg_env *env;
1280 int i, n_irgs = get_irp_n_irgs();
1281 ir_graph *rem = current_ir_graph;
1284 if (!(get_opt_optimize() && get_opt_inline())) return;
1286 /* extend all irgs by a temporary data structure for inlining. */
1287 for (i = 0; i < n_irgs; ++i)
1288 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1290 /* Precompute information in temporary data structure. */
1291 for (i = 0; i < n_irgs; ++i) {
1292 current_ir_graph = get_irp_irg(i);
1293 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1294 free_callee_info(current_ir_graph);
1296 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1297 get_irg_link(current_ir_graph));
1300 /* -- and now inline. -- */
1302 /* Inline leaves recursively -- we might construct new leaves. */
1303 while (did_inline) {
1306 for (i = 0; i < n_irgs; ++i) {
1308 int phiproj_computed = 0;
1310 current_ir_graph = get_irp_irg(i);
1311 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1313 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1316 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1317 callee = get_call_called_irg(call);
1319 if (env->n_nodes > maxsize) continue; // break;
1321 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1322 if (!phiproj_computed) {
1323 phiproj_computed = 1;
1324 collect_phiprojs(current_ir_graph);
1326 did_inline = inline_method(call, callee);
1329 /* Do some statistics */
1330 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1331 env->n_call_nodes --;
1332 env->n_nodes += callee_env->n_nodes;
1333 callee_env->n_callers--;
1340 /* inline other small functions. */
1341 for (i = 0; i < n_irgs; ++i) {
1344 int phiproj_computed = 0;
1346 current_ir_graph = get_irp_irg(i);
1347 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1349 /* we can not walk and change a set, nor remove from it.
1351 walkset = env->call_nodes;
1352 env->call_nodes = eset_create();
1353 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1356 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1357 callee = get_call_called_irg(call);
1360 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1361 (get_irg_inline_property(callee) == irg_inline_forced))) {
1362 if (!phiproj_computed) {
1363 phiproj_computed = 1;
1364 collect_phiprojs(current_ir_graph);
1366 if (inline_method(call, callee)) {
1367 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1368 env->n_call_nodes--;
1369 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1370 env->n_call_nodes += callee_env->n_call_nodes;
1371 env->n_nodes += callee_env->n_nodes;
1372 callee_env->n_callers--;
1375 eset_insert(env->call_nodes, call);
1378 eset_destroy(walkset);
1381 for (i = 0; i < n_irgs; ++i) {
1382 current_ir_graph = get_irp_irg(i);
1384 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1385 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1386 (env->n_callers_orig != env->n_callers))
1387 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1388 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1389 env->n_callers_orig, env->n_callers,
1390 get_entity_name(get_irg_entity(current_ir_graph)));
1392 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1395 current_ir_graph = rem;
1398 /*******************************************************************/
1399 /* Code Placement. Pins all floating nodes to a block where they */
1400 /* will be executed only if needed. */
1401 /*******************************************************************/
1404 * Find the earliest correct block for N. --- Place N into the
1405 * same Block as its dominance-deepest Input.
1408 place_floats_early(ir_node *n, pdeq *worklist)
1410 int i, start, irn_arity;
1412 /* we must not run into an infinite loop */
1413 assert (irn_not_visited(n));
1414 mark_irn_visited(n);
1416 /* Place floating nodes. */
1417 if (get_irn_pinned(n) == op_pin_state_floats) {
1419 ir_node *b = new_Bad(); /* The block to place this node in */
1420 int bad_recursion = is_Bad(get_nodes_block(n));
1422 assert(get_irn_op(n) != op_Block);
1424 if ((get_irn_op(n) == op_Const) ||
1425 (get_irn_op(n) == op_SymConst) ||
1427 (get_irn_op(n) == op_Unknown)) {
1428 /* These nodes will not be placed by the loop below. */
1429 b = get_irg_start_block(current_ir_graph);
1433 /* find the block for this node. */
1434 irn_arity = get_irn_arity(n);
1435 for (i = 0; i < irn_arity; i++) {
1436 ir_node *dep = get_irn_n(n, i);
1439 if ((irn_not_visited(dep))
1440 && (get_irn_pinned(dep) == op_pin_state_floats)) {
1441 place_floats_early(dep, worklist);
1445 * A node in the Bad block must stay in the bad block,
1446 * so don't compute a new block for it.
1451 /* Because all loops contain at least one op_pin_state_pinned node, now all
1452 our inputs are either op_pin_state_pinned or place_early has already
1453 been finished on them. We do not have any unfinished inputs! */
1454 dep_block = get_nodes_block(dep);
1455 if ((!is_Bad(dep_block)) &&
1456 (get_Block_dom_depth(dep_block) > depth)) {
1458 depth = get_Block_dom_depth(dep_block);
1460 /* Avoid that the node is placed in the Start block */
1461 if ((depth == 1) && (get_Block_dom_depth(get_nodes_block(n)) > 1)) {
1462 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1463 assert(b != get_irg_start_block(current_ir_graph));
1467 set_nodes_block(n, b);
1470 /* Add predecessors of non floating nodes on worklist. */
1471 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1472 irn_arity = get_irn_arity(n);
1473 for (i = start; i < irn_arity; i++) {
1474 ir_node *pred = get_irn_n(n, i);
1475 if (irn_not_visited(pred)) {
1476 pdeq_putr (worklist, pred);
1482 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1483 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1484 * places all floating nodes reachable from its argument through floating
1485 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1487 static INLINE void place_early(pdeq *worklist) {
1489 inc_irg_visited(current_ir_graph);
1491 /* this inits the worklist */
1492 place_floats_early(get_irg_end(current_ir_graph), worklist);
1494 /* Work the content of the worklist. */
1495 while (!pdeq_empty (worklist)) {
1496 ir_node *n = pdeq_getl (worklist);
1497 if (irn_not_visited(n)) place_floats_early(n, worklist);
1500 set_irg_outs_inconsistent(current_ir_graph);
1501 current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
1504 /** Compute the deepest common ancestor of block and dca. */
1505 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1508 if (!dca) return block;
1509 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1510 block = get_Block_idom(block);
1511 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1512 dca = get_Block_idom(dca);
1514 while (block != dca)
1515 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1520 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1521 * I.e., DCA is the block where we might place PRODUCER.
1522 * A data flow edge points from producer to consumer.
1525 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1527 ir_node *block = NULL;
1529 /* Compute the latest block into which we can place a node so that it is
1531 if (get_irn_op(consumer) == op_Phi) {
1532 /* our consumer is a Phi-node, the effective use is in all those
1533 blocks through which the Phi-node reaches producer */
1535 ir_node *phi_block = get_nodes_block(consumer);
1536 irn_arity = get_irn_arity(consumer);
1538 for (i = 0; i < irn_arity; i++) {
1539 if (get_irn_n(consumer, i) == producer) {
1540 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1542 block = calc_dca(block, new_block);
1546 assert(is_no_Block(consumer));
1547 block = get_nodes_block(consumer);
1550 /* Compute the deepest common ancestor of block and dca. */
1551 return calc_dca(dca, block);
1554 static INLINE int get_irn_loop_depth(ir_node *n) {
1555 return get_loop_depth(get_irn_loop(n));
1559 * Move n to a block with less loop depth than it's current block. The
1560 * new block must be dominated by early.
1563 move_out_of_loops (ir_node *n, ir_node *early)
1565 ir_node *best, *dca;
1569 /* Find the region deepest in the dominator tree dominating
1570 dca with the least loop nesting depth, but still dominated
1571 by our early placement. */
1572 dca = get_nodes_block(n);
1574 while (dca != early) {
1575 dca = get_Block_idom(dca);
1576 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1577 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1581 if (best != get_nodes_block(n)) {
1583 printf("Moving out of loop: "); DDMN(n);
1584 printf(" Outermost block: "); DDMN(early);
1585 printf(" Best block: "); DDMN(best);
1586 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1588 set_nodes_block(n, best);
1593 * Find the latest legal block for N and place N into the
1594 * `optimal' Block between the latest and earliest legal block.
1595 * The `optimal' block is the dominance-deepest block of those
1596 * with the least loop-nesting-depth. This places N out of as many
1597 * loops as possible and then makes it as control dependant as
1601 place_floats_late(ir_node *n, pdeq *worklist)
1606 assert (irn_not_visited(n)); /* no multiple placement */
1608 mark_irn_visited(n);
1610 /* no need to place block nodes, control nodes are already placed. */
1611 if ((get_irn_op(n) != op_Block) &&
1613 (get_irn_mode(n) != mode_X)) {
1614 /* Remember the early placement of this block to move it
1615 out of loop no further than the early placement. */
1616 early = get_nodes_block(n);
1618 /* Do not move code not reachable from Start. For
1619 * these we could not compute dominator information. */
1620 if (is_Bad(early) || get_Block_dom_depth(early) == -1)
1623 /* Assure that our users are all placed, except the Phi-nodes.
1624 --- Each data flow cycle contains at least one Phi-node. We
1625 have to break the `user has to be placed before the
1626 producer' dependence cycle and the Phi-nodes are the
1627 place to do so, because we need to base our placement on the
1628 final region of our users, which is OK with Phi-nodes, as they
1629 are op_pin_state_pinned, and they never have to be placed after a
1630 producer of one of their inputs in the same block anyway. */
1631 for (i = 0; i < get_irn_n_outs(n); i++) {
1632 ir_node *succ = get_irn_out(n, i);
1633 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1634 place_floats_late(succ, worklist);
1637 /* We have to determine the final block of this node... except for
1639 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1640 (get_irn_op(n) != op_Const) &&
1641 (get_irn_op(n) != op_SymConst)) {
1642 ir_node *dca = NULL; /* deepest common ancestor in the
1643 dominator tree of all nodes'
1644 blocks depending on us; our final
1645 placement has to dominate DCA. */
1646 for (i = 0; i < get_irn_n_outs(n); i++) {
1647 ir_node *out = get_irn_out(n, i);
1648 /* ignore if out is in dead code */
1649 ir_node *outbl = get_nodes_block(out);
1650 if (is_Bad(outbl) || get_Block_dom_depth(outbl) == -1)
1652 dca = consumer_dom_dca (dca, out, n);
1655 set_nodes_block(n, dca);
1657 move_out_of_loops (n, early);
1659 /* else all outs are in dead code */
1663 /* Add predecessors of all non-floating nodes on list. (Those of floating
1664 nodes are placeded already and therefore are marked.) */
1665 for (i = 0; i < get_irn_n_outs(n); i++) {
1666 ir_node *succ = get_irn_out(n, i);
1667 if (irn_not_visited(get_irn_out(n, i))) {
1668 pdeq_putr (worklist, succ);
1673 static INLINE void place_late(pdeq *worklist) {
1675 inc_irg_visited(current_ir_graph);
1677 /* This fills the worklist initially. */
1678 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1680 /* And now empty the worklist again... */
1681 while (!pdeq_empty (worklist)) {
1682 ir_node *n = pdeq_getl (worklist);
1683 if (irn_not_visited(n)) place_floats_late(n, worklist);
1687 void place_code(ir_graph *irg) {
1689 ir_graph *rem = current_ir_graph;
1691 current_ir_graph = irg;
1693 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1695 /* Handle graph state */
1696 assert(get_irg_phase_state(irg) != phase_building);
1697 if (get_irg_dom_state(irg) != dom_consistent)
1700 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1701 free_loop_information(irg);
1702 construct_backedges(irg);
1705 /* Place all floating nodes as early as possible. This guarantees
1706 a legal code placement. */
1707 worklist = new_pdeq();
1708 place_early(worklist);
1710 /* place_early invalidates the outs, place_late needs them. */
1712 /* Now move the nodes down in the dominator tree. This reduces the
1713 unnecessary executions of the node. */
1714 place_late(worklist);
1716 set_irg_outs_inconsistent(current_ir_graph);
1717 set_irg_loopinfo_inconsistent(current_ir_graph);
1719 current_ir_graph = rem;
1723 * Called by walker of remove_critical_cf_edges().
1725 * Place an empty block to an edge between a blocks of multiple
1726 * predecessors and a block of multiple successors.
1729 * @param env Environment of walker. This field is unused and has
1732 static void walk_critical_cf_edges(ir_node *n, void *env) {
1734 ir_node *pre, *block, **in, *jmp;
1736 /* Block has multiple predecessors */
1737 if ((op_Block == get_irn_op(n)) &&
1738 (get_irn_arity(n) > 1)) {
1739 arity = get_irn_arity(n);
1741 if (n == get_irg_end_block(current_ir_graph))
1742 return; /* No use to add a block here. */
1744 for (i=0; i<arity; i++) {
1745 pre = get_irn_n(n, i);
1746 /* Predecessor has multiple successors. Insert new flow edge */
1747 if ((NULL != pre) &&
1748 (op_Proj == get_irn_op(pre)) &&
1749 op_Raise != get_irn_op(skip_Proj(pre))) {
1751 /* set predecessor array for new block */
1752 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1753 /* set predecessor of new block */
1755 block = new_Block(1, in);
1756 /* insert new jmp node to new block */
1757 set_cur_block(block);
1760 /* set successor of new block */
1761 set_irn_n(n, i, jmp);
1763 } /* predecessor has multiple successors */
1764 } /* for all predecessors */
1765 } /* n is a block */
1768 void remove_critical_cf_edges(ir_graph *irg) {
1769 if (get_opt_critical_edges())
1770 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);