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.
23 # include "irnode_t.h"
24 # include "irgraph_t.h"
31 # include "pdeq.h" /* Fuer code placement */
34 # include "irbackedge_t.h"
36 /* Defined in iropt.c */
37 pset *new_identities (void);
38 void del_identities (pset *value_table);
39 void add_identities (pset *value_table, ir_node *node);
41 /********************************************************************/
42 /* apply optimizations of iropt to all nodes. */
43 /********************************************************************/
45 static void init_link (ir_node *n, void *env) {
46 set_irn_link(n, NULL);
49 #if 0 /* Old version. Avoids Ids.
50 This is not necessary: we do a postwalk, and get_irn_n
51 removes ids anyways. So it's much cheaper to call the
52 optimization less often and use the exchange() algorithm. */
54 optimize_in_place_wrapper (ir_node *n, void *env) {
56 ir_node *optimized, *old;
58 irn_arity = get_irn_arity(n);
59 for (i = 0; i < irn_arity; i++) {
60 /* get_irn_n skips Id nodes, so comparison old != optimized does not
61 show all optimizations. Therefore always set new predecessor. */
62 old = get_irn_n(n, i);
63 optimized = optimize_in_place_2(old);
64 set_irn_n(n, i, optimized);
67 if (get_irn_op(n) == op_Block) {
68 optimized = optimize_in_place_2(n);
69 if (optimized != n) exchange (n, optimized);
74 optimize_in_place_wrapper (ir_node *n, void *env) {
75 ir_node *optimized = optimize_in_place_2(n);
76 if (optimized != n) exchange (n, optimized);
83 local_optimize_graph (ir_graph *irg) {
84 ir_graph *rem = current_ir_graph;
85 current_ir_graph = irg;
87 /* Handle graph state */
88 assert(get_irg_phase_state(irg) != phase_building);
89 if (get_opt_global_cse())
90 set_irg_pinned(current_ir_graph, floats);
91 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
92 set_irg_outs_inconsistent(current_ir_graph);
93 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
94 set_irg_dom_inconsistent(current_ir_graph);
96 /* Clean the value_table in irg for the cse. */
97 del_identities(irg->value_table);
98 irg->value_table = new_identities();
100 /* walk over the graph */
101 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
103 current_ir_graph = rem;
106 /********************************************************************/
107 /* Routines for dead node elimination / copying garbage collection */
108 /* of the obstack. */
109 /********************************************************************/
111 /* Remeber the new node in the old node by using a field all nodes have. */
113 set_new_node (ir_node *old, ir_node *new)
118 /* Get this new node, before the old node is forgotton.*/
119 static INLINE ir_node *
120 get_new_node (ir_node * n)
125 /* We use the block_visited flag to mark that we have computed the
126 number of useful predecessors for this block.
127 Further we encode the new arity in this flag in the old blocks.
128 Remembering the arity is useful, as it saves a lot of pointer
129 accesses. This function is called for all Phi and Block nodes
132 compute_new_arity(ir_node *b) {
133 int i, res, irn_arity;
136 irg_v = get_irg_block_visited(current_ir_graph);
137 block_v = get_Block_block_visited(b);
138 if (block_v >= irg_v) {
139 /* we computed the number of preds for this block and saved it in the
141 return block_v - irg_v;
143 /* compute the number of good predecessors */
144 res = irn_arity = get_irn_arity(b);
145 for (i = 0; i < irn_arity; i++)
146 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
147 /* save it in the flag. */
148 set_Block_block_visited(b, irg_v + res);
153 static INLINE void new_backedge_info(ir_node *n) {
154 switch(get_irn_opcode(n)) {
156 n->attr.block.cg_backedge = NULL;
157 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
160 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
163 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
169 /* Copies the node to the new obstack. The Ins of the new node point to
170 the predecessors on the old obstack. For block/phi nodes not all
171 predecessors might be copied. n->link points to the new node.
172 For Phi and Block nodes the function allocates in-arrays with an arity
173 only for useful predecessors. The arity is determined by counting
174 the non-bad predecessors of the block. */
176 copy_node (ir_node *n, void *env) {
180 /* The end node looses it's flexible in array. This doesn't matter,
181 as dead node elimination builds End by hand, inlineing doesn't use
183 //assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC));
185 if (get_irn_opcode(n) == iro_Block) {
187 new_arity = compute_new_arity(n);
188 n->attr.block.graph_arr = NULL;
190 block = get_nodes_Block(n);
191 if (get_irn_opcode(n) == iro_Phi) {
192 new_arity = compute_new_arity(block);
194 new_arity = get_irn_arity(n);
197 nn = new_ir_node(get_irn_dbg_info(n),
204 /* Copy the attributes. These might point to additional data. If this
205 was allocated on the old obstack the pointers now are dangling. This
206 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
208 new_backedge_info(nn);
211 /* printf("\n old node: "); DDMSG2(n);
212 printf(" new node: "); DDMSG2(nn); */
216 /* Copies new predecessors of old node to new node remembered in link.
217 Spare the Bad predecessors of Phi and Block nodes. */
219 copy_preds (ir_node *n, void *env) {
223 nn = get_new_node(n);
225 /* printf("\n old node: "); DDMSG2(n);
226 printf(" new node: "); DDMSG2(nn);
227 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
229 if (get_irn_opcode(n) == iro_Block) {
230 /* Don't copy Bad nodes. */
232 irn_arity = get_irn_arity(n);
233 for (i = 0; i < irn_arity; i++)
234 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
235 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
236 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
239 /* repair the block visited flag from above misuse. Repair it in both
240 graphs so that the old one can still be used. */
241 set_Block_block_visited(nn, 0);
242 set_Block_block_visited(n, 0);
243 /* Local optimization could not merge two subsequent blocks if
244 in array contained Bads. Now it's possible.
245 We don't call optimize_in_place as it requires
246 that the fields in ir_graph are set properly. */
247 if ((get_opt_control_flow_straightening()) &&
248 (get_Block_n_cfgpreds(nn) == 1) &&
249 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp))
250 exchange(nn, get_nodes_Block(get_Block_cfgpred(nn, 0)));
251 } else if (get_irn_opcode(n) == iro_Phi) {
252 /* Don't copy node if corresponding predecessor in block is Bad.
253 The Block itself should not be Bad. */
254 block = get_nodes_Block(n);
255 set_irn_n (nn, -1, get_new_node(block));
257 irn_arity = get_irn_arity(n);
258 for (i = 0; i < irn_arity; i++)
259 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
260 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
261 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
264 /* If the pre walker reached this Phi after the post walker visited the
265 block block_visited is > 0. */
266 set_Block_block_visited(get_nodes_Block(n), 0);
267 /* Compacting the Phi's ins might generate Phis with only one
269 if (get_irn_arity(n) == 1)
270 exchange(n, get_irn_n(n, 0));
272 irn_arity = get_irn_arity(n);
273 for (i = -1; i < irn_arity; i++)
274 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
276 /* Now the new node is complete. We can add it to the hash table for cse.
277 @@@ inlinening aborts if we identify End. Why? */
278 if(get_irn_op(nn) != op_End)
279 add_identities (current_ir_graph->value_table, nn);
282 /* Copies the graph recursively, compacts the keepalive of the end node. */
285 ir_node *oe, *ne; /* old end, new end */
286 ir_node *ka; /* keep alive */
289 oe = get_irg_end(current_ir_graph);
290 /* copy the end node by hand, allocate dynamic in array! */
291 ne = new_ir_node(get_irn_dbg_info(oe),
298 /* Copy the attributes. Well, there might be some in the future... */
300 set_new_node(oe, ne);
302 /* copy the live nodes */
303 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
304 /* copy_preds for the end node ... */
305 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
307 /** ... and now the keep alives. **/
308 /* First pick the not marked block nodes and walk them. We must pick these
309 first as else we will oversee blocks reachable from Phis. */
310 irn_arity = get_irn_arity(oe);
311 for (i = 0; i < irn_arity; i++) {
312 ka = get_irn_n(oe, i);
313 if ((get_irn_op(ka) == op_Block) &&
314 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
315 /* We must keep the block alive and copy everything reachable */
316 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
317 irg_walk(ka, copy_node, copy_preds, NULL);
318 add_End_keepalive(ne, get_new_node(ka));
322 /* Now pick the Phis. Here we will keep all! */
323 irn_arity = get_irn_arity(oe);
324 for (i = 0; i < irn_arity; i++) {
325 ka = get_irn_n(oe, i);
326 if ((get_irn_op(ka) == op_Phi)) {
327 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
328 /* We didn't copy the Phi yet. */
329 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
330 irg_walk(ka, copy_node, copy_preds, NULL);
332 add_End_keepalive(ne, get_new_node(ka));
337 /* Copies the graph reachable from current_ir_graph->end to the obstack
338 in current_ir_graph and fixes the environment.
339 Then fixes the fields in current_ir_graph containing nodes of the
342 copy_graph_env (void) {
344 /* Not all nodes remembered in current_ir_graph might be reachable
345 from the end node. Assure their link is set to NULL, so that
346 we can test whether new nodes have been computed. */
347 set_irn_link(get_irg_frame (current_ir_graph), NULL);
348 set_irn_link(get_irg_globals(current_ir_graph), NULL);
349 set_irn_link(get_irg_args (current_ir_graph), NULL);
351 /* we use the block walk flag for removing Bads from Blocks ins. */
352 inc_irg_block_visited(current_ir_graph);
357 /* fix the fields in current_ir_graph */
358 old_end = get_irg_end(current_ir_graph);
359 set_irg_end (current_ir_graph, get_new_node(old_end));
361 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
362 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
363 copy_node (get_irg_frame(current_ir_graph), NULL);
364 copy_preds(get_irg_frame(current_ir_graph), NULL);
366 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
367 copy_node (get_irg_globals(current_ir_graph), NULL);
368 copy_preds(get_irg_globals(current_ir_graph), NULL);
370 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
371 copy_node (get_irg_args(current_ir_graph), NULL);
372 copy_preds(get_irg_args(current_ir_graph), NULL);
374 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
376 set_irg_start_block(current_ir_graph,
377 get_new_node(get_irg_start_block(current_ir_graph)));
378 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
379 set_irg_globals(current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
380 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
381 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
382 copy_node(get_irg_bad(current_ir_graph), NULL);
383 copy_preds(get_irg_bad(current_ir_graph), NULL);
385 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
386 /* GL removed: we need unknown with mode for analyses.
387 if (get_irn_link(get_irg_unknown(current_ir_graph)) == NULL) {
388 copy_node(get_irg_unknown(current_ir_graph), NULL);
389 copy_preds(get_irg_unknown(current_ir_graph), NULL);
391 set_irg_unknown(current_ir_graph, get_new_node(get_irg_unknown(current_ir_graph)));
395 /* Copies all reachable nodes to a new obstack. Removes bad inputs
396 from block nodes and the corresponding inputs from Phi nodes.
397 Merges single exit blocks with single entry blocks and removes
399 Adds all new nodes to a new hash table for cse. Does not
400 perform cse, so the hash table might contain common subexpressions. */
401 /* Amroq call this emigrate() */
403 dead_node_elimination(ir_graph *irg) {
405 struct obstack *graveyard_obst = NULL;
406 struct obstack *rebirth_obst = NULL;
408 /* Remember external state of current_ir_graph. */
409 rem = current_ir_graph;
410 current_ir_graph = irg;
412 /* Handle graph state */
413 assert(get_irg_phase_state(current_ir_graph) != phase_building);
414 free_outs(current_ir_graph);
416 /* @@@ so far we loose loops when copying */
417 set_irg_loop(current_ir_graph, NULL);
419 if (get_optimize() && get_opt_dead_node_elimination()) {
421 /* A quiet place, where the old obstack can rest in peace,
422 until it will be cremated. */
423 graveyard_obst = irg->obst;
425 /* A new obstack, where the reachable nodes will be copied to. */
426 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
427 current_ir_graph->obst = rebirth_obst;
428 obstack_init (current_ir_graph->obst);
430 /* We also need a new hash table for cse */
431 del_identities (irg->value_table);
432 irg->value_table = new_identities ();
434 /* Copy the graph from the old to the new obstack */
437 /* Free memory from old unoptimized obstack */
438 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
439 xfree (graveyard_obst); /* ... then free it. */
442 current_ir_graph = rem;
445 /* Relink bad predeseccors of a block and store the old in array to the
446 link field. This function is called by relink_bad_predecessors().
447 The array of link field starts with the block operand at position 0.
448 If block has bad predecessors, create a new in array without bad preds.
449 Otherwise let in array untouched. */
450 static void relink_bad_block_predecessors(ir_node *n, void *env) {
451 ir_node **new_in, *irn;
452 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
454 /* if link field of block is NULL, look for bad predecessors otherwise
455 this is allready done */
456 if (get_irn_op(n) == op_Block &&
457 get_irn_link(n) == NULL) {
459 /* save old predecessors in link field (position 0 is the block operand)*/
460 set_irn_link(n, (void *)get_irn_in(n));
462 /* count predecessors without bad nodes */
463 old_irn_arity = get_irn_arity(n);
464 for (i = 0; i < old_irn_arity; i++)
465 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
467 /* arity changing: set new predecessors without bad nodes */
468 if (new_irn_arity < old_irn_arity) {
469 /* get new predecessor array without Block predecessor */
470 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
472 /* set new predeseccors in array */
475 for (i = 1; i < old_irn_arity; i++) {
476 irn = get_irn_n(n, i);
477 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
480 } /* ir node has bad predecessors */
482 } /* Block is not relinked */
485 /* Relinks Bad predecesors from Bocks and Phis called by walker
486 remove_bad_predecesors(). If n is a Block, call
487 relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
488 function of Phi's Block. If this block has bad predecessors, relink preds
490 static void relink_bad_predecessors(ir_node *n, void *env) {
491 ir_node *block, **old_in;
492 int i, old_irn_arity, new_irn_arity;
494 /* relink bad predeseccors of a block */
495 if (get_irn_op(n) == op_Block)
496 relink_bad_block_predecessors(n, env);
498 /* If Phi node relink its block and its predecessors */
499 if (get_irn_op(n) == op_Phi) {
501 /* Relink predeseccors of phi's block */
502 block = get_nodes_Block(n);
503 if (get_irn_link(block) == NULL)
504 relink_bad_block_predecessors(block, env);
506 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
507 old_irn_arity = ARR_LEN(old_in);
509 /* Relink Phi predeseccors if count of predeseccors changed */
510 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
511 /* set new predeseccors in array
512 n->in[0] remains the same block */
514 for(i = 1; i < old_irn_arity; i++)
515 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
517 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
520 } /* n is a Phi node */
523 /* Removes Bad Bad predecesors from Blocks and the corresponding
524 inputs to Phi nodes as in dead_node_elimination but without
526 On walking up set the link field to NULL, on walking down call
527 relink_bad_predecessors() (This function stores the old in array
528 to the link field and sets a new in array if arity of predecessors
530 void remove_bad_predecessors(ir_graph *irg) {
531 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
535 /**********************************************************************/
536 /* Funcionality for inlining */
537 /**********************************************************************/
539 /* Copy node for inlineing. Updates attributes that change when
540 * inlineing but not for dead node elimination.
542 * Copies the node by calling copy_node and then updates the entity if
543 * it's a local one. env must be a pointer of the frame type of the
544 * inlined procedure. The new entities must be in the link field of
547 copy_node_inline (ir_node *n, void *env) {
549 type *frame_tp = (type *)env;
552 if (get_irn_op(n) == op_Sel) {
553 new = get_new_node (n);
554 assert(get_irn_op(new) == op_Sel);
555 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
556 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
558 } else if (get_irn_op(n) == op_Block) {
559 new = get_new_node (n);
560 new->attr.block.irg = current_ir_graph;
565 void inline_method(ir_node *call, ir_graph *called_graph) {
567 ir_node *post_call, *post_bl;
569 ir_node *end, *end_bl;
573 ir_node *cf_op = NULL, *bl;
574 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
577 if (!get_optimize() || !get_opt_inline()) return;
578 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
579 rem_opt = get_optimize();
582 /* Handle graph state */
583 assert(get_irg_phase_state(current_ir_graph) != phase_building);
584 assert(get_irg_pinned(current_ir_graph) == pinned);
585 assert(get_irg_pinned(called_graph) == pinned);
586 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
587 set_irg_outs_inconsistent(current_ir_graph);
589 /* -- Check preconditions -- */
590 assert(get_irn_op(call) == op_Call);
591 /* @@@ does not work for InterfaceIII.java after cgana
592 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
593 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
594 get_Call_type(call)));
596 assert(get_type_tpop(get_Call_type(call)) == type_method);
597 if (called_graph == current_ir_graph) {
598 set_optimize(rem_opt);
603 the procedure and later replaces the Start node of the called graph.
604 Post_call is the old Call node and collects the results of the called
605 graph. Both will end up being a tuple. -- */
606 post_bl = get_nodes_Block(call);
607 set_irg_current_block(current_ir_graph, post_bl);
608 /* XxMxPxP of Start + parameter of Call */
610 in[1] = get_Call_mem(call);
611 in[2] = get_irg_frame(current_ir_graph);
612 in[3] = get_irg_globals(current_ir_graph);
613 in[4] = new_Tuple (get_Call_n_params(call), get_Call_param_arr(call));
614 pre_call = new_Tuple(5, in);
618 The new block gets the ins of the old block, pre_call and all its
619 predecessors and all Phi nodes. -- */
620 part_block(pre_call);
622 /* -- Prepare state for dead node elimination -- */
623 /* Visited flags in calling irg must be >= flag in called irg.
624 Else walker and arity computation will not work. */
625 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
626 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
627 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
628 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
629 /* Set pre_call as new Start node in link field of the start node of
630 calling graph and pre_calls block as new block for the start block
632 Further mark these nodes so that they are not visited by the
634 set_irn_link(get_irg_start(called_graph), pre_call);
635 set_irn_visited(get_irg_start(called_graph),
636 get_irg_visited(current_ir_graph));
637 set_irn_link(get_irg_start_block(called_graph),
638 get_nodes_Block(pre_call));
639 set_irn_visited(get_irg_start_block(called_graph),
640 get_irg_visited(current_ir_graph));
642 /* Initialize for compaction of in arrays */
643 inc_irg_block_visited(current_ir_graph);
645 /* -- Replicate local entities of the called_graph -- */
646 /* copy the entities. */
647 called_frame = get_irg_frame_type(called_graph);
648 for (i = 0; i < get_class_n_members(called_frame); i++) {
649 entity *new_ent, *old_ent;
650 old_ent = get_class_member(called_frame, i);
651 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
652 set_entity_link(old_ent, new_ent);
655 /* visited is > than that of called graph. With this trick visited will
656 remain unchanged so that an outer walker, e.g., searching the call nodes
657 to inline, calling this inline will not visit the inlined nodes. */
658 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
660 /* -- Performing dead node elimination inlines the graph -- */
661 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
663 /* @@@ endless loops are not copied!! -- they should be, I think... */
664 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
665 get_irg_frame_type(called_graph));
667 /* Repair called_graph */
668 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
669 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
670 set_Block_block_visited(get_irg_start_block(called_graph), 0);
672 /* -- Merge the end of the inlined procedure with the call site -- */
673 /* We will turn the old Call node into a Tuple with the following
676 0: Phi of all Memories of Return statements.
677 1: Jmp from new Block that merges the control flow from all exception
678 predecessors of the old end block.
679 2: Tuple of all arguments.
680 3: Phi of Exception memories.
683 /* -- Precompute some values -- */
684 end_bl = get_new_node(get_irg_end_block(called_graph));
685 end = get_new_node(get_irg_end(called_graph));
686 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
687 n_res = get_method_n_ress(get_Call_type(call));
689 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
690 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
692 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
694 /* -- archive keepalives -- */
695 irn_arity = get_irn_arity(end);
696 for (i = 0; i < irn_arity; i++)
697 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
699 /* The new end node will die. We need not free as the in array is on the obstack:
700 copy_node only generated 'D' arrays. */
703 Return nodes by Jump nodes. -- */
705 for (i = 0; i < arity; i++) {
707 ret = get_irn_n(end_bl, i);
708 if (get_irn_op(ret) == op_Return) {
709 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
713 set_irn_in(post_bl, n_ret, cf_pred);
716 turned into a tuple. -- */
717 turn_into_tuple(post_call, 4);
718 /* First the Memory-Phi */
720 for (i = 0; i < arity; i++) {
721 ret = get_irn_n(end_bl, i);
722 if (get_irn_op(ret) == op_Return) {
723 cf_pred[n_ret] = get_Return_mem(ret);
727 phi = new_Phi(n_ret, cf_pred, mode_M);
728 set_Tuple_pred(call, 0, phi);
729 /* Conserve Phi-list for further inlinings -- but might be optimized */
730 if (get_nodes_Block(phi) == post_bl) {
731 set_irn_link(phi, get_irn_link(post_bl));
732 set_irn_link(post_bl, phi);
734 /* Now the real results */
736 for (j = 0; j < n_res; j++) {
738 for (i = 0; i < arity; i++) {
739 ret = get_irn_n(end_bl, i);
740 if (get_irn_op(ret) == op_Return) {
741 cf_pred[n_ret] = get_Return_res(ret, j);
745 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
747 /* Conserve Phi-list for further inlinings -- but might be optimized */
748 if (get_nodes_Block(phi) == post_bl) {
749 set_irn_link(phi, get_irn_link(post_bl));
750 set_irn_link(post_bl, phi);
753 set_Tuple_pred(call, 2, new_Tuple(n_res, res_pred));
755 set_Tuple_pred(call, 2, new_Bad());
757 /* Finally the exception control flow. We need to add a Phi node to
758 collect the memory containing the exception objects. Further we need
759 to add another block to get a correct representation of this Phi. To
760 this block we add a Jmp that resolves into the X output of the Call
761 when the Call is turned into a tuple. */
763 for (i = 0; i < arity; i++) {
765 ret = get_irn_n(end_bl, i);
766 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
767 cf_pred[n_exc] = ret;
772 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
773 set_Tuple_pred(call, 1, new_Jmp());
774 /* The Phi for the memories with the exception objects */
776 for (i = 0; i < arity; i++) {
778 ret = skip_Proj(get_irn_n(end_bl, i));
779 if (get_irn_op(ret) == op_Call) {
780 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
782 } else if (is_fragile_op(ret)) {
783 /* We rely that all cfops have the memory output at the same position. */
784 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
786 } else if (get_irn_op(ret) == op_Raise) {
787 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
791 set_Tuple_pred(call, 3, new_Phi(n_exc, cf_pred, mode_M));
793 set_Tuple_pred(call, 1, new_Bad());
794 set_Tuple_pred(call, 3, new_Bad());
800 /* -- If the exception control flow from the inlined Call directly
801 branched to the end block we now have the following control
802 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
803 remove the Jmp along with it's empty block and add Jmp's
804 predecessors as predecessors of this end block. No problem if
805 there is no exception, because then branches Bad to End which
807 /* find the problematic predecessor of the end block. */
808 end_bl = get_irg_end_block(current_ir_graph);
809 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
810 cf_op = get_Block_cfgpred(end_bl, i);
811 if (get_irn_op(cf_op) == op_Proj) {
812 cf_op = get_Proj_pred(cf_op);
813 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
814 // There are unoptimized tuples from inlineing before when no exc
815 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
816 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
817 assert(get_irn_op(cf_op) == op_Jmp);
823 if (i < get_Block_n_cfgpreds(end_bl)) {
824 bl = get_nodes_Block(cf_op);
825 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
826 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
827 for (j = 0; j < i; j++)
828 cf_pred[j] = get_Block_cfgpred(end_bl, j);
829 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
830 cf_pred[j] = get_Block_cfgpred(bl, j-i);
831 for (j = j; j < arity; j++)
832 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
833 set_irn_in(end_bl, arity, cf_pred);
835 // Remove the exception pred from post-call Tuple.
836 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
840 /* -- Turn cse back on. -- */
841 set_optimize(rem_opt);
844 /********************************************************************/
845 /* Apply inlineing to small methods. */
846 /********************************************************************/
850 /* It makes no sense to inline too many calls in one procedure. Anyways,
851 I didn't get a version with NEW_ARR_F to run. */
852 #define MAX_INLINE 1024
854 static void collect_calls(ir_node *call, void *env) {
855 ir_node **calls = (ir_node **)env;
858 ir_graph *called_irg;
860 if (get_irn_op(call) != op_Call) return;
862 addr = get_Call_ptr(call);
863 if (get_irn_op(addr) == op_Const) {
864 /* Check whether the constant is the pointer to a compiled entity. */
865 tv = get_Const_tarval(addr);
866 if (tarval_to_entity(tv)) {
867 called_irg = get_entity_irg(tarval_to_entity(tv));
868 if (called_irg && pos < MAX_INLINE) {
869 /* The Call node calls a locally defined method. Remember to inline. */
877 /* Inlines all small methods at call sites where the called address comes
878 from a Const node that references the entity representing the called
880 The size argument is a rough measure for the code size of the method:
881 Methods where the obstack containing the firm graph is smaller than
883 void inline_small_irgs(ir_graph *irg, int size) {
885 ir_node *calls[MAX_INLINE];
886 ir_graph *rem = current_ir_graph;
888 if (!(get_optimize() && get_opt_inline())) return;
890 current_ir_graph = irg;
891 /* Handle graph state */
892 assert(get_irg_phase_state(current_ir_graph) != phase_building);
894 /* Find Call nodes to inline.
895 (We can not inline during a walk of the graph, as inlineing the same
896 method several times changes the visited flag of the walked graph:
897 after the first inlineing visited of the callee equals visited of
898 the caller. With the next inlineing both are increased.) */
900 irg_walk(get_irg_end(irg), NULL, collect_calls, (void *) calls);
902 if ((pos > 0) && (pos < MAX_INLINE)) {
903 /* There are calls to inline */
904 collect_phiprojs(irg);
905 for (i = 0; i < pos; i++) {
908 tv = get_Const_tarval(get_Call_ptr(calls[i]));
909 callee = get_entity_irg(tarval_to_entity(tv));
910 if ((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) {
911 inline_method(calls[i], callee);
916 current_ir_graph = rem;
920 /********************************************************************/
921 /* Code Placement. Pinns all floating nodes to a block where they */
922 /* will be executed only if needed. */
923 /********************************************************************/
925 static pdeq *worklist; /* worklist of ir_node*s */
927 /* Find the earliest correct block for N. --- Place N into the
928 same Block as its dominance-deepest Input. */
930 place_floats_early (ir_node *n)
932 int i, start, irn_arity;
934 /* we must not run into an infinite loop */
935 assert (irn_not_visited(n));
938 /* Place floating nodes. */
939 if (get_op_pinned(get_irn_op(n)) == floats) {
941 ir_node *b = new_Bad(); /* The block to place this node in */
943 assert(get_irn_op(n) != op_Block);
945 if ((get_irn_op(n) == op_Const) ||
946 (get_irn_op(n) == op_SymConst) ||
948 (get_irn_op(n) == op_Unknown)) {
949 /* These nodes will not be placed by the loop below. */
950 b = get_irg_start_block(current_ir_graph);
954 /* find the block for this node. */
955 irn_arity = get_irn_arity(n);
956 for (i = 0; i < irn_arity; i++) {
957 ir_node *dep = get_irn_n(n, i);
959 if ((irn_not_visited(dep)) &&
960 (get_op_pinned(get_irn_op(dep)) == floats)) {
961 place_floats_early (dep);
963 /* Because all loops contain at least one pinned node, now all
964 our inputs are either pinned or place_early has already
965 been finished on them. We do not have any unfinished inputs! */
966 dep_block = get_nodes_Block(dep);
967 if ((!is_Bad(dep_block)) &&
968 (get_Block_dom_depth(dep_block) > depth)) {
970 depth = get_Block_dom_depth(dep_block);
972 /* Avoid that the node is placed in the Start block */
973 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
974 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
975 assert(b != get_irg_start_block(current_ir_graph));
979 set_nodes_Block(n, b);
982 /* Add predecessors of non floating nodes on worklist. */
983 start = (get_irn_op(n) == op_Block) ? 0 : -1;
984 irn_arity = get_irn_arity(n);
985 for (i = start; i < irn_arity; i++) {
986 ir_node *pred = get_irn_n(n, i);
987 if (irn_not_visited(pred)) {
988 pdeq_putr (worklist, pred);
993 /* Floating nodes form subgraphs that begin at nodes as Const, Load,
994 Start, Call and end at pinned nodes as Store, Call. Place_early
995 places all floating nodes reachable from its argument through floating
996 nodes and adds all beginnings at pinned nodes to the worklist. */
997 static INLINE void place_early (void) {
999 inc_irg_visited(current_ir_graph);
1001 /* this inits the worklist */
1002 place_floats_early (get_irg_end(current_ir_graph));
1004 /* Work the content of the worklist. */
1005 while (!pdeq_empty (worklist)) {
1006 ir_node *n = pdeq_getl (worklist);
1007 if (irn_not_visited(n)) place_floats_early (n);
1010 set_irg_outs_inconsistent(current_ir_graph);
1011 current_ir_graph->pinned = pinned;
1015 /* deepest common dominance ancestor of DCA and CONSUMER of PRODUCER */
1017 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1019 ir_node *block = NULL;
1021 /* Compute the latest block into which we can place a node so that it is
1023 if (get_irn_op(consumer) == op_Phi) {
1024 /* our comsumer is a Phi-node, the effective use is in all those
1025 blocks through which the Phi-node reaches producer */
1027 ir_node *phi_block = get_nodes_Block(consumer);
1028 irn_arity = get_irn_arity(consumer);
1029 for (i = 0; i < irn_arity; i++) {
1030 if (get_irn_n(consumer, i) == producer) {
1031 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1035 assert(is_no_Block(consumer));
1036 block = get_nodes_Block(consumer);
1039 /* Compute the deepest common ancestor of block and dca. */
1041 if (!dca) return block;
1042 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1043 block = get_Block_idom(block);
1044 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1045 dca = get_Block_idom(dca);
1046 while (block != dca)
1047 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1052 static INLINE int get_irn_loop_depth(ir_node *n) {
1053 return get_loop_depth(get_irn_loop(n));
1056 /* Move n to a block with less loop depth than it's current block. The
1057 new block must be dominated by early. */
1059 move_out_of_loops (ir_node *n, ir_node *early)
1061 ir_node *best, *dca;
1065 /* Find the region deepest in the dominator tree dominating
1066 dca with the least loop nesting depth, but still dominated
1067 by our early placement. */
1068 dca = get_nodes_Block(n);
1070 while (dca != early) {
1071 dca = get_Block_idom(dca);
1072 if (!dca) break; /* should we put assert(dca)? */
1073 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1077 if (best != get_nodes_Block(n)) {
1079 printf("Moving out of loop: "); DDMN(n);
1080 printf(" Outermost block: "); DDMN(early);
1081 printf(" Best block: "); DDMN(best);
1082 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1084 set_nodes_Block(n, best);
1088 /* Find the latest legal block for N and place N into the
1089 `optimal' Block between the latest and earliest legal block.
1090 The `optimal' block is the dominance-deepest block of those
1091 with the least loop-nesting-depth. This places N out of as many
1092 loops as possible and then makes it as controldependant as
1095 place_floats_late (ir_node *n)
1100 assert (irn_not_visited(n)); /* no multiple placement */
1102 /* no need to place block nodes, control nodes are already placed. */
1103 if ((get_irn_op(n) != op_Block) &&
1105 (get_irn_mode(n) != mode_X)) {
1106 /* Remember the early palacement of this block to move it
1107 out of loop no further than the early placement. */
1108 early = get_nodes_Block(n);
1109 /* Assure that our users are all placed, except the Phi-nodes.
1110 --- Each dataflow cycle contains at least one Phi-node. We
1111 have to break the `user has to be placed before the
1112 producer' dependance cycle and the Phi-nodes are the
1113 place to do so, because we need to base our placement on the
1114 final region of our users, which is OK with Phi-nodes, as they
1115 are pinned, and they never have to be placed after a
1116 producer of one of their inputs in the same block anyway. */
1117 for (i = 0; i < get_irn_n_outs(n); i++) {
1118 ir_node *succ = get_irn_out(n, i);
1119 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1120 place_floats_late (succ);
1123 /* We have to determine the final block of this node... except for
1125 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1126 (get_irn_op(n) != op_Const) &&
1127 (get_irn_op(n) != op_SymConst)) {
1128 ir_node *dca = NULL; /* deepest common ancestor in the
1129 dominator tree of all nodes'
1130 blocks depending on us; our final
1131 placement has to dominate DCA. */
1132 for (i = 0; i < get_irn_n_outs(n); i++) {
1133 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1135 set_nodes_Block(n, dca);
1137 move_out_of_loops (n, early);
1141 mark_irn_visited(n);
1143 /* Add predecessors of all non-floating nodes on list. (Those of floating
1144 nodes are placeded already and therefore are marked.) */
1145 for (i = 0; i < get_irn_n_outs(n); i++) {
1146 if (irn_not_visited(get_irn_out(n, i))) {
1147 pdeq_putr (worklist, get_irn_out(n, i));
1152 static INLINE void place_late(void) {
1154 inc_irg_visited(current_ir_graph);
1156 /* This fills the worklist initially. */
1157 place_floats_late(get_irg_start_block(current_ir_graph));
1158 /* And now empty the worklist again... */
1159 while (!pdeq_empty (worklist)) {
1160 ir_node *n = pdeq_getl (worklist);
1161 if (irn_not_visited(n)) place_floats_late(n);
1165 void place_code(ir_graph *irg) {
1166 ir_graph *rem = current_ir_graph;
1167 current_ir_graph = irg;
1169 if (!(get_optimize() && get_opt_global_cse())) return;
1171 /* Handle graph state */
1172 assert(get_irg_phase_state(irg) != phase_building);
1173 if (get_irg_dom_state(irg) != dom_consistent)
1176 construct_backedges(irg);
1178 /* Place all floating nodes as early as possible. This guarantees
1179 a legal code placement. */
1180 worklist = new_pdeq ();
1183 /* place_early invalidates the outs, place_late needs them. */
1185 /* Now move the nodes down in the dominator tree. This reduces the
1186 unnecessary executions of the node. */
1189 set_irg_outs_inconsistent(current_ir_graph);
1190 del_pdeq (worklist);
1191 current_ir_graph = rem;
1196 /********************************************************************/
1197 /* Control flow optimization. */
1198 /* Removes Bad control flow predecessors and empty blocks. A block */
1199 /* is empty if it contains only a Jmp node. */
1200 /* Blocks can only be removed if they are not needed for the */
1201 /* semantics of Phi nodes. */
1202 /********************************************************************/
1204 /* Removes Tuples from Block control flow predecessors.
1205 Optimizes blocks with equivalent_node().
1206 Replaces n by Bad if n is unreachable control flow. */
1207 static void merge_blocks(ir_node *n, void *env) {
1209 set_irn_link(n, NULL);
1211 if (get_irn_op(n) == op_Block) {
1213 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1214 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go throug.
1215 A different order of optimizations might cause problems. */
1216 if (get_opt_normalize())
1217 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1218 } else if (get_optimize() && (get_irn_mode(n) == mode_X)) {
1219 /* We will soon visit a block. Optimize it before visiting! */
1220 ir_node *b = get_nodes_Block(n);
1221 ir_node *new = equivalent_node(b);
1222 while (irn_not_visited(b) && (!is_Bad(new)) && (new != b)) {
1223 /* We would have to run gigo if new is bad, so we
1224 promote it directly below. */
1225 assert(((b == new) ||
1226 get_opt_control_flow_straightening() ||
1227 get_opt_control_flow_weak_simplification()) &&
1228 ("strange flag setting"));
1231 new = equivalent_node(b);
1233 /* GL @@@ get_opt_normalize hinzugefuegt, 5.5.2003 */
1234 if (is_Bad(new) && get_opt_normalize()) exchange (n, new_Bad());
1238 /* Collects all Phi nodes in link list of Block.
1239 Marks all blocks "block_visited" if they contain a node other
1241 static void collect_nodes(ir_node *n, void *env) {
1242 if (is_no_Block(n)) {
1243 ir_node *b = get_nodes_Block(n);
1245 if ((get_irn_op(n) == op_Phi)) {
1246 /* Collect Phi nodes to compact ins along with block's ins. */
1247 set_irn_link(n, get_irn_link(b));
1249 } else if (get_irn_op(n) != op_Jmp) { /* Check for non empty block. */
1250 mark_Block_block_visited(b);
1255 /* Returns true if pred is pred of block */
1256 static int is_pred_of(ir_node *pred, ir_node *b) {
1258 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1259 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1260 if (b_pred == pred) return 1;
1265 static int test_whether_dispensable(ir_node *b, int pos) {
1266 int i, j, n_preds = 1;
1267 int dispensable = 1;
1268 ir_node *cfop = get_Block_cfgpred(b, pos);
1269 ir_node *pred = get_nodes_Block(cfop);
1271 if (get_Block_block_visited(pred) + 1
1272 < get_irg_block_visited(current_ir_graph)) {
1273 if (!get_optimize() || !get_opt_control_flow_strong_simplification()) {
1274 /* Mark block so that is will not be removed. */
1275 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1278 /* Seems to be empty. */
1279 if (!get_irn_link(b)) {
1280 /* There are no Phi nodes ==> dispensable. */
1281 n_preds = get_Block_n_cfgpreds(pred);
1283 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1284 Work preds < pos as if they were already removed. */
1285 for (i = 0; i < pos; i++) {
1286 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1287 if (get_Block_block_visited(b_pred) + 1
1288 < get_irg_block_visited(current_ir_graph)) {
1289 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1290 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1291 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1294 if (is_pred_of(b_pred, pred)) dispensable = 0;
1297 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1298 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1299 if (is_pred_of(b_pred, pred)) dispensable = 0;
1302 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1305 n_preds = get_Block_n_cfgpreds(pred);
1313 static void optimize_blocks(ir_node *b, void *env) {
1314 int i, j, k, max_preds, n_preds;
1315 ir_node *pred, *phi;
1318 /* Count the number of predecessor if this block is merged with pred blocks
1321 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1322 max_preds += test_whether_dispensable(b, i);
1324 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1327 printf(" working on "); DDMN(b);
1328 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1329 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1330 if (is_Bad(get_Block_cfgpred(b, i))) {
1331 printf(" removing Bad %i\n ", i);
1332 } else if (get_Block_block_visited(pred) +1
1333 < get_irg_block_visited(current_ir_graph)) {
1334 printf(" removing pred %i ", i); DDMN(pred);
1335 } else { printf(" Nothing to do for "); DDMN(pred); }
1337 * end Debug output **/
1339 /** Fix the Phi nodes **/
1340 phi = get_irn_link(b);
1342 assert(get_irn_op(phi) == op_Phi);
1343 /* Find the new predecessors for the Phi */
1345 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1346 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1347 if (is_Bad(get_Block_cfgpred(b, i))) {
1349 } else if (get_Block_block_visited(pred) +1
1350 < get_irg_block_visited(current_ir_graph)) {
1351 /* It's an empty block and not yet visited. */
1352 ir_node *phi_pred = get_Phi_pred(phi, i);
1353 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1354 if (get_nodes_Block(phi_pred) == pred) {
1355 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1356 in[n_preds] = get_Phi_pred(phi_pred, j);
1358 in[n_preds] = phi_pred;
1362 /* The Phi_pred node is replaced now if it is a Phi.
1363 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1364 Daher muss der Phiknoten durch den neuen ersetzt werden.
1365 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1366 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1367 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1369 if (get_nodes_Block(phi_pred) == pred) {
1370 /* remove the Phi as it might be kept alive. Further there
1371 might be other users. */
1372 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1375 in[n_preds] = get_Phi_pred(phi, i);
1380 set_irn_in(phi, n_preds, in);
1382 phi = get_irn_link(phi);
1386 This happens only if merge between loop backedge and single loop entry. **/
1387 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1388 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1389 if (get_Block_block_visited(pred) +1
1390 < get_irg_block_visited(current_ir_graph)) {
1391 phi = get_irn_link(pred);
1393 if (get_irn_op(phi) == op_Phi) {
1394 set_nodes_Block(phi, b);
1397 for (i = 0; i < k; i++) {
1398 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1399 if (is_Bad(get_Block_cfgpred(b, i))) {
1401 } else if (get_Block_block_visited(pred) +1
1402 < get_irg_block_visited(current_ir_graph)) {
1403 /* It's an empty block and not yet visited. */
1404 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1405 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1406 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1407 Anweisungen.) Trotzdem tuts bisher!! */
1416 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1417 in[n_preds] = get_Phi_pred(phi, i);
1420 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1421 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1422 if (is_Bad(get_Block_cfgpred(b, i))) {
1424 } else if (get_Block_block_visited(pred) +1
1425 < get_irg_block_visited(current_ir_graph)) {
1426 /* It's an empty block and not yet visited. */
1427 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1436 set_irn_in(phi, n_preds, in);
1438 phi = get_irn_link(phi);
1443 /** Fix the block **/
1445 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1446 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1447 if (is_Bad(get_Block_cfgpred(b, i))) {
1449 } else if (get_Block_block_visited(pred) +1
1450 < get_irg_block_visited(current_ir_graph)) {
1451 /* It's an empty block and not yet visited. */
1452 assert(get_Block_n_cfgpreds(b) > 1);
1453 /* Else it should be optimized by equivalent_node. */
1454 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1455 in[n_preds] = get_Block_cfgpred(pred, j);
1458 /* Remove block as it might be kept alive. */
1459 exchange(pred, b/*new_Bad()*/);
1461 in[n_preds] = get_Block_cfgpred(b, i);
1465 set_irn_in(b, n_preds, in);
1469 void optimize_cf(ir_graph *irg) {
1472 ir_node *end = get_irg_end(irg);
1473 ir_graph *rem = current_ir_graph;
1474 current_ir_graph = irg;
1476 /* Handle graph state */
1477 assert(get_irg_phase_state(irg) != phase_building);
1478 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1479 set_irg_outs_inconsistent(current_ir_graph);
1480 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1481 set_irg_dom_inconsistent(current_ir_graph);
1483 /* Use block visited flag to mark non-empty blocks. */
1484 inc_irg_block_visited(irg);
1485 irg_walk(end, merge_blocks, collect_nodes, NULL);
1487 /* Optimize the standard code. */
1488 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1490 /* Walk all keep alives, optimize them if block, add to new in-array
1491 for end if useful. */
1492 in = NEW_ARR_F (ir_node *, 1);
1493 in[0] = get_nodes_Block(end);
1494 inc_irg_visited(current_ir_graph);
1495 for(i = 0; i < get_End_n_keepalives(end); i++) {
1496 ir_node *ka = get_End_keepalive(end, i);
1497 if (irn_not_visited(ka)) {
1498 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1499 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1500 get_irg_block_visited(current_ir_graph)-1);
1501 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1502 mark_irn_visited(ka);
1503 ARR_APP1 (ir_node *, in, ka);
1504 } else if (get_irn_op(ka) == op_Phi) {
1505 mark_irn_visited(ka);
1506 ARR_APP1 (ir_node *, in, ka);
1510 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1513 current_ir_graph = rem;
1518 * Called by walker of remove_critical_cf_edges.
1520 * Place an empty block to an edge between a blocks of multiple
1521 * predecessors and a block of multiple sucessors.
1524 * @param env Envirnment of walker. This field is unused and has
1527 static void walk_critical_cf_edges(ir_node *n, void *env) {
1529 ir_node *pre, *block, **in, *jmp;
1531 /* Block has multiple predecessors */
1532 if ((op_Block == get_irn_op(n)) &&
1533 (get_irn_arity(n) > 1)) {
1534 arity = get_irn_arity(n);
1536 if (n == get_irg_end_block(current_ir_graph))
1537 return; // No use to add a block here.
1539 for (i=0; i<arity; i++) {
1540 pre = get_irn_n(n, i);
1541 /* Predecessor has multiple sucessors. Insert new flow edge */
1542 if ((NULL != pre) &&
1543 (op_Proj == get_irn_op(pre)) &&
1544 op_Raise != get_irn_op(skip_Proj(pre))) {
1546 /* set predecessor array for new block */
1547 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1548 /* set predecessor of new block */
1550 block = new_Block(1, in);
1551 /* insert new jmp node to new block */
1552 switch_block(block);
1555 /* set sucessor of new block */
1556 set_irn_n(n, i, jmp);
1558 } /* predecessor has multiple sucessors */
1559 } /* for all predecessors */
1560 } /* n is a block */
1563 void remove_critical_cf_edges(ir_graph *irg) {
1564 if (get_opt_critical_edges())
1565 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);