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
212 copy_node (ir_node *n, void *env) {
215 opcode op = get_irn_opcode(n);
216 int copy_node_nr = env != NULL;
218 /* The end node looses it's flexible in array. This doesn't matter,
219 as dead node elimination builds End by hand, inlineing doesn't use
221 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
224 /* node copied already */
226 } else if (op == iro_Block) {
228 new_arity = compute_new_arity(n);
229 n->attr.block.graph_arr = NULL;
231 block = get_nodes_block(n);
232 if (get_irn_opcode(n) == iro_Phi) {
233 new_arity = compute_new_arity(block);
235 new_arity = get_irn_arity(n);
238 nn = new_ir_node(get_irn_dbg_info(n),
245 /* Copy the attributes. These might point to additional data. If this
246 was allocated on the old obstack the pointers now are dangling. This
247 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
249 new_backedge_info(nn);
254 /* for easier debugging, we want to copy the node numbers too */
255 nn->node_nr = n->node_nr;
259 /* printf("\n old node: "); DDMSG2(n);
260 printf(" new node: "); DDMSG2(nn); */
264 * Copies new predecessors of old node to new node remembered in link.
265 * Spare the Bad predecessors of Phi and Block nodes.
268 copy_preds (ir_node *n, void *env) {
272 nn = get_new_node(n);
274 /* printf("\n old node: "); DDMSG2(n);
275 printf(" new node: "); DDMSG2(nn);
276 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
278 if (get_irn_opcode(n) == iro_Block) {
279 /* Don't copy Bad nodes. */
281 irn_arity = get_irn_arity(n);
282 for (i = 0; i < irn_arity; i++)
283 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
284 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
285 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
288 /* repair the block visited flag from above misuse. Repair it in both
289 graphs so that the old one can still be used. */
290 set_Block_block_visited(nn, 0);
291 set_Block_block_visited(n, 0);
292 /* Local optimization could not merge two subsequent blocks if
293 in array contained Bads. Now it's possible.
294 We don't call optimize_in_place as it requires
295 that the fields in ir_graph are set properly. */
296 if ((get_opt_control_flow_straightening()) &&
297 (get_Block_n_cfgpreds(nn) == 1) &&
298 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
299 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
301 /* Jmp jumps into the block it is in -- deal self cycle. */
302 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
303 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
308 } else if (get_irn_opcode(n) == iro_Phi) {
309 /* Don't copy node if corresponding predecessor in block is Bad.
310 The Block itself should not be Bad. */
311 block = get_nodes_block(n);
312 set_irn_n (nn, -1, get_new_node(block));
314 irn_arity = get_irn_arity(n);
315 for (i = 0; i < irn_arity; i++)
316 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
317 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
318 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
321 /* If the pre walker reached this Phi after the post walker visited the
322 block block_visited is > 0. */
323 set_Block_block_visited(get_nodes_block(n), 0);
324 /* Compacting the Phi's ins might generate Phis with only one
326 if (get_irn_arity(n) == 1)
327 exchange(n, get_irn_n(n, 0));
329 irn_arity = get_irn_arity(n);
330 for (i = -1; i < irn_arity; i++)
331 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
333 /* Now the new node is complete. We can add it to the hash table for cse.
334 @@@ inlinening aborts if we identify End. Why? */
335 if(get_irn_op(nn) != op_End)
336 add_identities (current_ir_graph->value_table, nn);
340 * Copies the graph recursively, compacts the keepalive of the end node.
342 * @param copy_node_nr If non-zero, the node number will be copied
345 copy_graph (int copy_node_nr) {
346 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
347 ir_node *ka; /* keep alive */
350 oe = get_irg_end(current_ir_graph);
351 /* copy the end node by hand, allocate dynamic in array! */
352 ne = new_ir_node(get_irn_dbg_info(oe),
359 /* Copy the attributes. Well, there might be some in the future... */
361 set_new_node(oe, ne);
363 /* copy the Bad node */
364 ob = get_irg_bad(current_ir_graph);
365 nb = new_ir_node(get_irn_dbg_info(ob),
372 set_new_node(ob, nb);
374 /* copy the NoMem node */
375 om = get_irg_no_mem(current_ir_graph);
376 nm = new_ir_node(get_irn_dbg_info(om),
383 set_new_node(om, nm);
385 /* copy the live nodes */
386 irg_walk(get_nodes_block(oe), copy_node, copy_preds, (void *)copy_node_nr);
387 /* copy_preds for the end node ... */
388 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
390 /*- ... and now the keep alives. -*/
391 /* First pick the not marked block nodes and walk them. We must pick these
392 first as else we will oversee blocks reachable from Phis. */
393 irn_arity = get_irn_arity(oe);
394 for (i = 0; i < irn_arity; i++) {
395 ka = get_irn_intra_n(oe, i);
396 if ((get_irn_op(ka) == op_Block) &&
397 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
398 /* We must keep the block alive and copy everything reachable */
399 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
400 irg_walk(ka, copy_node, copy_preds, (void *)copy_node_nr);
401 add_End_keepalive(ne, get_new_node(ka));
405 /* Now pick the Phis. Here we will keep all! */
406 irn_arity = get_irn_arity(oe);
407 for (i = 0; i < irn_arity; i++) {
408 ka = get_irn_intra_n(oe, i);
409 if ((get_irn_op(ka) == op_Phi)) {
410 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
411 /* We didn't copy the Phi yet. */
412 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
413 irg_walk(ka, copy_node, copy_preds, (void *)copy_node_nr);
415 add_End_keepalive(ne, get_new_node(ka));
419 /* start block sometimes only reached after keep alives */
420 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
421 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
425 * Copies the graph reachable from current_ir_graph->end to the obstack
426 * in current_ir_graph and fixes the environment.
427 * Then fixes the fields in current_ir_graph containing nodes of the
430 * @param copy_node_nr If non-zero, the node number will be copied
433 copy_graph_env (int copy_node_nr) {
435 /* Not all nodes remembered in current_ir_graph might be reachable
436 from the end node. Assure their link is set to NULL, so that
437 we can test whether new nodes have been computed. */
438 set_irn_link(get_irg_frame (current_ir_graph), NULL);
439 set_irn_link(get_irg_globals (current_ir_graph), NULL);
440 set_irn_link(get_irg_args (current_ir_graph), NULL);
441 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
442 set_irn_link(get_irg_no_mem (current_ir_graph), NULL);
444 /* we use the block walk flag for removing Bads from Blocks ins. */
445 inc_irg_block_visited(current_ir_graph);
448 copy_graph(copy_node_nr);
450 /* fix the fields in current_ir_graph */
451 old_end = get_irg_end(current_ir_graph);
452 set_irg_end (current_ir_graph, get_new_node(old_end));
453 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
454 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
456 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
457 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
458 copy_node (get_irg_frame(current_ir_graph), (void *)copy_node_nr);
459 copy_preds(get_irg_frame(current_ir_graph), NULL);
461 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
462 copy_node (get_irg_globals(current_ir_graph), (void *)copy_node_nr);
463 copy_preds(get_irg_globals(current_ir_graph), NULL);
465 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
466 copy_node (get_irg_initial_mem(current_ir_graph), (void *)copy_node_nr);
467 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
469 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
470 copy_node (get_irg_args(current_ir_graph), (void *)copy_node_nr);
471 copy_preds(get_irg_args(current_ir_graph), NULL);
473 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
475 set_irg_start_block(current_ir_graph,
476 get_new_node(get_irg_start_block(current_ir_graph)));
477 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
478 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
479 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
480 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
482 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
483 copy_node(get_irg_bad(current_ir_graph), (void *)copy_node_nr);
484 copy_preds(get_irg_bad(current_ir_graph), NULL);
486 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
488 if (get_irn_link(get_irg_no_mem(current_ir_graph)) == NULL) {
489 copy_node(get_irg_no_mem(current_ir_graph), (void *)copy_node_nr);
490 copy_preds(get_irg_no_mem(current_ir_graph), NULL);
492 set_irg_no_mem(current_ir_graph, get_new_node(get_irg_no_mem(current_ir_graph)));
496 * Copies all reachable nodes to a new obstack. Removes bad inputs
497 * from block nodes and the corresponding inputs from Phi nodes.
498 * Merges single exit blocks with single entry blocks and removes
500 * Adds all new nodes to a new hash table for cse. Does not
501 * perform cse, so the hash table might contain common subexpressions.
504 dead_node_elimination(ir_graph *irg) {
506 int rem_ipview = get_interprocedural_view();
507 struct obstack *graveyard_obst = NULL;
508 struct obstack *rebirth_obst = NULL;
510 /* inform statistics that we started a dead-node elimination run */
511 stat_dead_node_elim_start(irg);
513 /* Remember external state of current_ir_graph. */
514 rem = current_ir_graph;
515 current_ir_graph = irg;
516 set_interprocedural_view(false);
518 /* Handle graph state */
519 assert(get_irg_phase_state(current_ir_graph) != phase_building);
520 free_callee_info(current_ir_graph);
521 free_outs(current_ir_graph);
522 /* @@@ so far we loose loops when copying */
523 free_loop_information(current_ir_graph);
525 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
527 /* A quiet place, where the old obstack can rest in peace,
528 until it will be cremated. */
529 graveyard_obst = irg->obst;
531 /* A new obstack, where the reachable nodes will be copied to. */
532 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
533 current_ir_graph->obst = rebirth_obst;
534 obstack_init (current_ir_graph->obst);
536 /* We also need a new hash table for cse */
537 del_identities (irg->value_table);
538 irg->value_table = new_identities ();
540 /* Copy the graph from the old to the new obstack */
543 /* Free memory from old unoptimized obstack */
544 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
545 xfree (graveyard_obst); /* ... then free it. */
548 /* inform statistics that the run is over */
549 stat_dead_node_elim_stop(irg);
551 current_ir_graph = rem;
552 set_interprocedural_view(rem_ipview);
556 * Relink bad predeseccors of a block and store the old in array to the
557 * link field. This function is called by relink_bad_predecessors().
558 * The array of link field starts with the block operand at position 0.
559 * If block has bad predecessors, create a new in array without bad preds.
560 * Otherwise let in array untouched.
562 static void relink_bad_block_predecessors(ir_node *n, void *env) {
563 ir_node **new_in, *irn;
564 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
566 /* if link field of block is NULL, look for bad predecessors otherwise
567 this is allready done */
568 if (get_irn_op(n) == op_Block &&
569 get_irn_link(n) == NULL) {
571 /* save old predecessors in link field (position 0 is the block operand)*/
572 set_irn_link(n, (void *)get_irn_in(n));
574 /* count predecessors without bad nodes */
575 old_irn_arity = get_irn_arity(n);
576 for (i = 0; i < old_irn_arity; i++)
577 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
579 /* arity changing: set new predecessors without bad nodes */
580 if (new_irn_arity < old_irn_arity) {
581 /* Get new predecessor array. We do not resize the array, as we must
582 keep the old one to update Phis. */
583 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
585 /* set new predeseccors in array */
588 for (i = 0; i < old_irn_arity; i++) {
589 irn = get_irn_n(n, i);
591 new_in[new_irn_n] = irn;
592 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
596 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
597 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
600 } /* ir node has bad predecessors */
602 } /* Block is not relinked */
606 * Relinks Bad predecesors from Bocks and Phis called by walker
607 * remove_bad_predecesors(). If n is a Block, call
608 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
609 * function of Phi's Block. If this block has bad predecessors, relink preds
612 static void relink_bad_predecessors(ir_node *n, void *env) {
613 ir_node *block, **old_in;
614 int i, old_irn_arity, new_irn_arity;
616 /* relink bad predeseccors of a block */
617 if (get_irn_op(n) == op_Block)
618 relink_bad_block_predecessors(n, env);
620 /* If Phi node relink its block and its predecessors */
621 if (get_irn_op(n) == op_Phi) {
623 /* Relink predeseccors of phi's block */
624 block = get_nodes_block(n);
625 if (get_irn_link(block) == NULL)
626 relink_bad_block_predecessors(block, env);
628 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
629 old_irn_arity = ARR_LEN(old_in);
631 /* Relink Phi predeseccors if count of predeseccors changed */
632 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
633 /* set new predeseccors in array
634 n->in[0] remains the same block */
636 for(i = 1; i < old_irn_arity; i++)
637 if (!is_Bad((ir_node *)old_in[i])) {
638 n->in[new_irn_arity] = n->in[i];
639 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
643 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
644 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
647 } /* n is a Phi node */
651 * Removes Bad Bad predecesors from Blocks and the corresponding
652 * inputs to Phi nodes as in dead_node_elimination but without
654 * On walking up set the link field to NULL, on walking down call
655 * relink_bad_predecessors() (This function stores the old in array
656 * to the link field and sets a new in array if arity of predecessors
659 void remove_bad_predecessors(ir_graph *irg) {
660 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
664 /*--------------------------------------------------------------------*/
665 /* Funcionality for inlining */
666 /*--------------------------------------------------------------------*/
669 * Copy node for inlineing. Updates attributes that change when
670 * inlineing but not for dead node elimination.
672 * Copies the node by calling copy_node and then updates the entity if
673 * it's a local one. env must be a pointer of the frame type of the
674 * inlined procedure. The new entities must be in the link field of
678 copy_node_inline (ir_node *n, void *env) {
680 type *frame_tp = (type *)env;
683 if (get_irn_op(n) == op_Sel) {
684 new = get_new_node (n);
685 assert(get_irn_op(new) == op_Sel);
686 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
687 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
689 } else if (get_irn_op(n) == op_Block) {
690 new = get_new_node (n);
691 new->attr.block.irg = current_ir_graph;
695 static void find_addr(ir_node *node, void *env)
697 if (get_irn_opcode(node) == iro_Proj) {
698 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
704 * currently, we cannot inline two cases:
705 * - call with compound arguments
706 * - graphs that take the address of a parameter
708 * check these conditions here
710 static int can_inline(ir_node *call, ir_graph *called_graph)
712 type *call_type = get_Call_type(call);
713 int params, ress, i, res;
714 assert(is_method_type(call_type));
716 params = get_method_n_params(call_type);
717 ress = get_method_n_ress(call_type);
720 for (i = 0; i < params; ++i) {
721 type *p_type = get_method_param_type(call_type, i);
723 if (is_compound_type(p_type))
728 for (i = 0; i < ress; ++i) {
729 type *r_type = get_method_res_type(call_type, i);
731 if (is_compound_type(r_type))
736 irg_walk_graph(called_graph, find_addr, NULL, &res);
741 int inline_method(ir_node *call, ir_graph *called_graph) {
743 ir_node *post_call, *post_bl;
745 ir_node *end, *end_bl;
749 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
752 irg_inline_property prop = get_irg_inline_property(called_graph);
754 if ( (prop != irg_inline_forced) &&
755 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
757 /* Do not inline variadic functions. */
758 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
761 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
762 get_method_n_params(get_Call_type(call)));
765 * currently, we cannot inline two cases:
766 * - call with compound arguments
767 * - graphs that take the address of a parameter
769 if (! can_inline(call, called_graph))
772 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
773 rem_opt = get_opt_optimize();
776 /* Handle graph state */
777 assert(get_irg_phase_state(current_ir_graph) != phase_building);
778 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
779 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
780 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
781 set_irg_outs_inconsistent(current_ir_graph);
782 set_irg_loopinfo_inconsistent(current_ir_graph);
783 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
785 /* -- Check preconditions -- */
786 assert(get_irn_op(call) == op_Call);
787 /* @@@ does not work for InterfaceIII.java after cgana
788 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
789 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
790 get_Call_type(call)));
792 assert(get_type_tpop(get_Call_type(call)) == type_method);
793 if (called_graph == current_ir_graph) {
794 set_optimize(rem_opt);
798 /* here we know we WILL inline, so inform the statistics */
799 stat_inline(call, called_graph);
801 /* -- Decide how to handle exception control flow: Is there a handler
802 for the Call node, or do we branch directly to End on an exception?
804 0 There is a handler.
806 2 Exception handling not represented in Firm. -- */
808 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
809 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
810 assert(get_irn_op(proj) == op_Proj);
811 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
812 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
814 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
815 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
816 else { exc_handling = 2; } /* !Mproj && !Xproj */
821 the procedure and later replaces the Start node of the called graph.
822 Post_call is the old Call node and collects the results of the called
823 graph. Both will end up being a tuple. -- */
824 post_bl = get_nodes_block(call);
825 set_irg_current_block(current_ir_graph, post_bl);
826 /* XxMxPxP of Start + parameter of Call */
827 in[pn_Start_X_initial_exec] = new_Jmp();
828 in[pn_Start_M] = get_Call_mem(call);
829 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
830 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
831 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
832 /* in[pn_Start_P_value_arg_base] = ??? */
833 pre_call = new_Tuple(5, in);
837 The new block gets the ins of the old block, pre_call and all its
838 predecessors and all Phi nodes. -- */
839 part_block(pre_call);
841 /* -- Prepare state for dead node elimination -- */
842 /* Visited flags in calling irg must be >= flag in called irg.
843 Else walker and arity computation will not work. */
844 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
845 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
846 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
847 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
848 /* Set pre_call as new Start node in link field of the start node of
849 calling graph and pre_calls block as new block for the start block
851 Further mark these nodes so that they are not visited by the
853 set_irn_link(get_irg_start(called_graph), pre_call);
854 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
855 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
856 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
857 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
858 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
860 /* Initialize for compaction of in arrays */
861 inc_irg_block_visited(current_ir_graph);
863 /* -- Replicate local entities of the called_graph -- */
864 /* copy the entities. */
865 called_frame = get_irg_frame_type(called_graph);
866 for (i = 0; i < get_class_n_members(called_frame); i++) {
867 entity *new_ent, *old_ent;
868 old_ent = get_class_member(called_frame, i);
869 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
870 set_entity_link(old_ent, new_ent);
873 /* visited is > than that of called graph. With this trick visited will
874 remain unchanged so that an outer walker, e.g., searching the call nodes
875 to inline, calling this inline will not visit the inlined nodes. */
876 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
878 /* -- Performing dead node elimination inlines the graph -- */
879 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
881 /* @@@ endless loops are not copied!! -- they should be, I think... */
882 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
883 get_irg_frame_type(called_graph));
885 /* Repair called_graph */
886 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
887 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
888 set_Block_block_visited(get_irg_start_block(called_graph), 0);
890 /* -- Merge the end of the inlined procedure with the call site -- */
891 /* We will turn the old Call node into a Tuple with the following
894 0: Phi of all Memories of Return statements.
895 1: Jmp from new Block that merges the control flow from all exception
896 predecessors of the old end block.
897 2: Tuple of all arguments.
898 3: Phi of Exception memories.
899 In case the old Call directly branches to End on an exception we don't
900 need the block merging all exceptions nor the Phi of the exception
904 /* -- Precompute some values -- */
905 end_bl = get_new_node(get_irg_end_block(called_graph));
906 end = get_new_node(get_irg_end(called_graph));
907 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
908 n_res = get_method_n_ress(get_Call_type(call));
910 res_pred = (ir_node **) xmalloc (n_res * sizeof (ir_node *));
911 cf_pred = (ir_node **) xmalloc (arity * sizeof (ir_node *));
913 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
915 /* -- archive keepalives -- */
916 irn_arity = get_irn_arity(end);
917 for (i = 0; i < irn_arity; i++)
918 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
920 /* The new end node will die. We need not free as the in array is on the obstack:
921 copy_node only generated 'D' arrays. */
923 /* -- Replace Return nodes by Jump nodes. -- */
925 for (i = 0; i < arity; i++) {
927 ret = get_irn_n(end_bl, i);
928 if (get_irn_op(ret) == op_Return) {
929 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
933 set_irn_in(post_bl, n_ret, cf_pred);
935 /* -- Build a Tuple for all results of the method.
936 Add Phi node if there was more than one Return. -- */
937 turn_into_tuple(post_call, 4);
938 /* First the Memory-Phi */
940 for (i = 0; i < arity; i++) {
941 ret = get_irn_n(end_bl, i);
942 if (get_irn_op(ret) == op_Return) {
943 cf_pred[n_ret] = get_Return_mem(ret);
947 phi = new_Phi(n_ret, cf_pred, mode_M);
948 set_Tuple_pred(call, pn_Call_M_regular, phi);
949 /* Conserve Phi-list for further inlinings -- but might be optimized */
950 if (get_nodes_block(phi) == post_bl) {
951 set_irn_link(phi, get_irn_link(post_bl));
952 set_irn_link(post_bl, phi);
954 /* Now the real results */
956 for (j = 0; j < n_res; j++) {
958 for (i = 0; i < arity; i++) {
959 ret = get_irn_n(end_bl, i);
960 if (get_irn_op(ret) == op_Return) {
961 cf_pred[n_ret] = get_Return_res(ret, j);
966 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
970 /* Conserve Phi-list for further inlinings -- but might be optimized */
971 if (get_nodes_block(phi) == post_bl) {
972 set_irn_link(phi, get_irn_link(post_bl));
973 set_irn_link(post_bl, phi);
976 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
978 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
980 /* Finally the exception control flow.
981 We have two (three) possible situations:
982 First if the Call branches to an exception handler: We need to add a Phi node to
983 collect the memory containing the exception objects. Further we need
984 to add another block to get a correct representation of this Phi. To
985 this block we add a Jmp that resolves into the X output of the Call
986 when the Call is turned into a tuple.
987 Second the Call branches to End, the exception is not handled. Just
988 add all inlined exception branches to the End node.
989 Third: there is no Exception edge at all. Handle as case two. */
990 if (exc_handling == 0) {
992 for (i = 0; i < arity; i++) {
994 ret = get_irn_n(end_bl, i);
995 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
996 cf_pred[n_exc] = ret;
1001 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1002 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1003 /* The Phi for the memories with the exception objects */
1005 for (i = 0; i < arity; i++) {
1007 ret = skip_Proj(get_irn_n(end_bl, i));
1008 if (get_irn_op(ret) == op_Call) {
1009 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1011 } else if (is_fragile_op(ret)) {
1012 /* We rely that all cfops have the memory output at the same position. */
1013 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1015 } else if (get_irn_op(ret) == op_Raise) {
1016 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1020 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1022 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1023 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1026 ir_node *main_end_bl;
1027 int main_end_bl_arity;
1028 ir_node **end_preds;
1030 /* assert(exc_handling == 1 || no exceptions. ) */
1032 for (i = 0; i < arity; i++) {
1033 ir_node *ret = get_irn_n(end_bl, i);
1035 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1036 cf_pred[n_exc] = ret;
1040 main_end_bl = get_irg_end_block(current_ir_graph);
1041 main_end_bl_arity = get_irn_arity(main_end_bl);
1042 end_preds = (ir_node **) xmalloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
1044 for (i = 0; i < main_end_bl_arity; ++i)
1045 end_preds[i] = get_irn_n(main_end_bl, i);
1046 for (i = 0; i < n_exc; ++i)
1047 end_preds[main_end_bl_arity + i] = cf_pred[i];
1048 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1049 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1050 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1056 #if 0 /* old. now better, correcter, faster implementation. */
1058 /* -- If the exception control flow from the inlined Call directly
1059 branched to the end block we now have the following control
1060 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1061 remove the Jmp along with it's empty block and add Jmp's
1062 predecessors as predecessors of this end block. No problem if
1063 there is no exception, because then branches Bad to End which
1065 @@@ can't we know this beforehand: by getting the Proj(1) from
1066 the Call link list and checking whether it goes to Proj. */
1067 /* find the problematic predecessor of the end block. */
1068 end_bl = get_irg_end_block(current_ir_graph);
1069 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1070 cf_op = get_Block_cfgpred(end_bl, i);
1071 if (get_irn_op(cf_op) == op_Proj) {
1072 cf_op = get_Proj_pred(cf_op);
1073 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1074 /* There are unoptimized tuples from inlineing before when no exc */
1075 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1076 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1077 assert(get_irn_op(cf_op) == op_Jmp);
1083 if (i < get_Block_n_cfgpreds(end_bl)) {
1084 bl = get_nodes_block(cf_op);
1085 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1086 cf_pred = (ir_node **) xmalloc (arity * sizeof (ir_node *));
1087 for (j = 0; j < i; j++)
1088 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1089 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1090 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1091 for (j = j; j < arity; j++)
1092 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1093 set_irn_in(end_bl, arity, cf_pred);
1095 /* Remove the exception pred from post-call Tuple. */
1096 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1101 /* -- Turn cse back on. -- */
1102 set_optimize(rem_opt);
1107 /********************************************************************/
1108 /* Apply inlineing to small methods. */
1109 /********************************************************************/
1111 /* It makes no sense to inline too many calls in one procedure. Anyways,
1112 I didn't get a version with NEW_ARR_F to run. */
1113 #define MAX_INLINE 1024
1116 * environment for inlining small irgs
1118 typedef struct _inline_env_t {
1120 ir_node *calls[MAX_INLINE];
1124 * Returns the irg called from a Call node. If the irg is not
1125 * known, NULL is returned.
1127 static ir_graph *get_call_called_irg(ir_node *call) {
1129 ir_graph *called_irg = NULL;
1131 assert(get_irn_op(call) == op_Call);
1133 addr = get_Call_ptr(call);
1134 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1135 called_irg = get_entity_irg(get_SymConst_entity(addr));
1141 static void collect_calls(ir_node *call, void *env) {
1144 if (get_irn_op(call) != op_Call) return;
1146 addr = get_Call_ptr(call);
1148 if (get_irn_op(addr) == op_SymConst) {
1149 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1150 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1151 inline_env_t *ienv = (inline_env_t *)env;
1152 if (called_irg && ienv->pos < MAX_INLINE) {
1153 /* The Call node calls a locally defined method. Remember to inline. */
1154 ienv->calls[ienv->pos++] = call;
1161 * Inlines all small methods at call sites where the called address comes
1162 * from a Const node that references the entity representing the called
1164 * The size argument is a rough measure for the code size of the method:
1165 * Methods where the obstack containing the firm graph is smaller than
1168 void inline_small_irgs(ir_graph *irg, int size) {
1170 ir_graph *rem = current_ir_graph;
1171 inline_env_t env /* = {0, NULL}*/;
1173 if (!(get_opt_optimize() && get_opt_inline())) return;
1175 current_ir_graph = irg;
1176 /* Handle graph state */
1177 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1178 free_callee_info(current_ir_graph);
1180 /* Find Call nodes to inline.
1181 (We can not inline during a walk of the graph, as inlineing the same
1182 method several times changes the visited flag of the walked graph:
1183 after the first inlineing visited of the callee equals visited of
1184 the caller. With the next inlineing both are increased.) */
1186 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1188 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1189 /* There are calls to inline */
1190 collect_phiprojs(irg);
1191 for (i = 0; i < env.pos; i++) {
1193 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1194 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1195 (get_irg_inline_property(callee) == irg_inline_forced)) {
1196 inline_method(env.calls[i], callee);
1201 current_ir_graph = rem;
1205 * Environment for inlining irgs.
1208 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1209 int n_nodes_orig; /**< for statistics */
1210 eset *call_nodes; /**< All call nodes in this graph */
1212 int n_call_nodes_orig; /**< for statistics */
1213 int n_callers; /**< Number of known graphs that call this graphs. */
1214 int n_callers_orig; /**< for statistics */
1217 static inline_irg_env *new_inline_irg_env(void) {
1218 inline_irg_env *env = xmalloc(sizeof(inline_irg_env));
1219 env->n_nodes = -2; /* uncount Start, End */
1220 env->n_nodes_orig = -2; /* uncount Start, End */
1221 env->call_nodes = eset_create();
1222 env->n_call_nodes = 0;
1223 env->n_call_nodes_orig = 0;
1225 env->n_callers_orig = 0;
1229 static void free_inline_irg_env(inline_irg_env *env) {
1230 eset_destroy(env->call_nodes);
1234 static void collect_calls2(ir_node *call, void *env) {
1235 inline_irg_env *x = (inline_irg_env *)env;
1236 ir_op *op = get_irn_op(call);
1239 /* count nodes in irg */
1240 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1245 if (op != op_Call) return;
1247 /* collect all call nodes */
1248 eset_insert(x->call_nodes, (void *)call);
1250 x->n_call_nodes_orig++;
1252 /* count all static callers */
1253 callee = get_call_called_irg(call);
1255 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1256 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1260 INLINE static int is_leave(ir_graph *irg) {
1261 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1264 INLINE static int is_smaller(ir_graph *callee, int size) {
1265 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1270 * Inlines small leave methods at call sites where the called address comes
1271 * from a Const node that references the entity representing the called
1273 * The size argument is a rough measure for the code size of the method:
1274 * Methods where the obstack containing the firm graph is smaller than
1277 void inline_leave_functions(int maxsize, int leavesize, int size) {
1278 inline_irg_env *env;
1279 int i, n_irgs = get_irp_n_irgs();
1280 ir_graph *rem = current_ir_graph;
1283 if (!(get_opt_optimize() && get_opt_inline())) return;
1285 /* extend all irgs by a temporary data structure for inlineing. */
1286 for (i = 0; i < n_irgs; ++i)
1287 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1289 /* Precompute information in temporary data structure. */
1290 for (i = 0; i < n_irgs; ++i) {
1291 current_ir_graph = get_irp_irg(i);
1292 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1293 free_callee_info(current_ir_graph);
1295 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1296 get_irg_link(current_ir_graph));
1299 /* -- and now inline. -- */
1301 /* Inline leaves recursively -- we might construct new leaves. */
1302 while (did_inline) {
1305 for (i = 0; i < n_irgs; ++i) {
1307 int phiproj_computed = 0;
1309 current_ir_graph = get_irp_irg(i);
1310 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1312 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1313 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1314 ir_graph *callee = get_call_called_irg(call);
1316 if (env->n_nodes > maxsize) continue; // break;
1318 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1319 if (!phiproj_computed) {
1320 phiproj_computed = 1;
1321 collect_phiprojs(current_ir_graph);
1323 did_inline = inline_method(call, callee);
1326 /* Do some statistics */
1327 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1328 env->n_call_nodes --;
1329 env->n_nodes += callee_env->n_nodes;
1330 callee_env->n_callers--;
1337 /* inline other small functions. */
1338 for (i = 0; i < n_irgs; ++i) {
1341 int phiproj_computed = 0;
1343 current_ir_graph = get_irp_irg(i);
1344 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1346 /* we can not walk and change a set, nor remove from it.
1348 walkset = env->call_nodes;
1349 env->call_nodes = eset_create();
1350 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1351 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1352 ir_graph *callee = get_call_called_irg(call);
1355 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1356 (get_irg_inline_property(callee) == irg_inline_forced))) {
1357 if (!phiproj_computed) {
1358 phiproj_computed = 1;
1359 collect_phiprojs(current_ir_graph);
1361 if (inline_method(call, callee)) {
1362 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1363 env->n_call_nodes--;
1364 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1365 env->n_call_nodes += callee_env->n_call_nodes;
1366 env->n_nodes += callee_env->n_nodes;
1367 callee_env->n_callers--;
1370 eset_insert(env->call_nodes, call);
1373 eset_destroy(walkset);
1376 for (i = 0; i < n_irgs; ++i) {
1377 current_ir_graph = get_irp_irg(i);
1379 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1380 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1381 (env->n_callers_orig != env->n_callers))
1382 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1383 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1384 env->n_callers_orig, env->n_callers,
1385 get_entity_name(get_irg_entity(current_ir_graph)));
1387 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1390 current_ir_graph = rem;
1393 /*******************************************************************/
1394 /* Code Placement. Pins all floating nodes to a block where they */
1395 /* will be executed only if needed. */
1396 /*******************************************************************/
1399 * Find the earliest correct block for N. --- Place N into the
1400 * same Block as its dominance-deepest Input.
1403 place_floats_early(ir_node *n, pdeq *worklist)
1405 int i, start, irn_arity;
1407 /* we must not run into an infinite loop */
1408 assert (irn_not_visited(n));
1409 mark_irn_visited(n);
1411 /* Place floating nodes. */
1412 if (get_irn_pinned(n) == op_pin_state_floats) {
1414 ir_node *b = new_Bad(); /* The block to place this node in */
1415 int bad_recursion = is_Bad(get_nodes_block(n));
1417 assert(get_irn_op(n) != op_Block);
1419 if ((get_irn_op(n) == op_Const) ||
1420 (get_irn_op(n) == op_SymConst) ||
1422 (get_irn_op(n) == op_Unknown)) {
1423 /* These nodes will not be placed by the loop below. */
1424 b = get_irg_start_block(current_ir_graph);
1428 /* find the block for this node. */
1429 irn_arity = get_irn_arity(n);
1430 for (i = 0; i < irn_arity; i++) {
1431 ir_node *dep = get_irn_n(n, i);
1434 if ((irn_not_visited(dep))
1435 && (get_irn_pinned(dep) == op_pin_state_floats)) {
1436 place_floats_early(dep, worklist);
1440 * A node in the Bad block must stay in the bad block,
1441 * so don't compute a new block for it.
1446 /* Because all loops contain at least one op_pin_state_pinned node, now all
1447 our inputs are either op_pin_state_pinned or place_early has already
1448 been finished on them. We do not have any unfinished inputs! */
1449 dep_block = get_nodes_block(dep);
1450 if ((!is_Bad(dep_block)) &&
1451 (get_Block_dom_depth(dep_block) > depth)) {
1453 depth = get_Block_dom_depth(dep_block);
1455 /* Avoid that the node is placed in the Start block */
1456 if ((depth == 1) && (get_Block_dom_depth(get_nodes_block(n)) > 1)) {
1457 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1458 assert(b != get_irg_start_block(current_ir_graph));
1462 set_nodes_block(n, b);
1465 /* Add predecessors of non floating nodes on worklist. */
1466 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1467 irn_arity = get_irn_arity(n);
1468 for (i = start; i < irn_arity; i++) {
1469 ir_node *pred = get_irn_n(n, i);
1470 if (irn_not_visited(pred)) {
1471 pdeq_putr (worklist, pred);
1477 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1478 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1479 * places all floating nodes reachable from its argument through floating
1480 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1482 static INLINE void place_early(pdeq *worklist) {
1484 inc_irg_visited(current_ir_graph);
1486 /* this inits the worklist */
1487 place_floats_early(get_irg_end(current_ir_graph), worklist);
1489 /* Work the content of the worklist. */
1490 while (!pdeq_empty (worklist)) {
1491 ir_node *n = pdeq_getl (worklist);
1492 if (irn_not_visited(n)) place_floats_early(n, worklist);
1495 set_irg_outs_inconsistent(current_ir_graph);
1496 current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
1499 /** Compute the deepest common ancestor of block and dca. */
1500 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1503 if (!dca) return block;
1504 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1505 block = get_Block_idom(block);
1506 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1507 dca = get_Block_idom(dca);
1509 while (block != dca)
1510 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1515 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1516 * I.e., DCA is the block where we might place PRODUCER.
1517 * A data flow edge points from producer to consumer.
1520 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1522 ir_node *block = NULL;
1524 /* Compute the latest block into which we can place a node so that it is
1526 if (get_irn_op(consumer) == op_Phi) {
1527 /* our consumer is a Phi-node, the effective use is in all those
1528 blocks through which the Phi-node reaches producer */
1530 ir_node *phi_block = get_nodes_block(consumer);
1531 irn_arity = get_irn_arity(consumer);
1533 for (i = 0; i < irn_arity; i++) {
1534 if (get_irn_n(consumer, i) == producer) {
1535 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1537 block = calc_dca(block, new_block);
1541 assert(is_no_Block(consumer));
1542 block = get_nodes_block(consumer);
1545 /* Compute the deepest common ancestor of block and dca. */
1546 return calc_dca(dca, block);
1549 static INLINE int get_irn_loop_depth(ir_node *n) {
1550 return get_loop_depth(get_irn_loop(n));
1554 * Move n to a block with less loop depth than it's current block. The
1555 * new block must be dominated by early.
1558 move_out_of_loops (ir_node *n, ir_node *early)
1560 ir_node *best, *dca;
1564 /* Find the region deepest in the dominator tree dominating
1565 dca with the least loop nesting depth, but still dominated
1566 by our early placement. */
1567 dca = get_nodes_block(n);
1569 while (dca != early) {
1570 dca = get_Block_idom(dca);
1571 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1572 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1576 if (best != get_nodes_block(n)) {
1578 printf("Moving out of loop: "); DDMN(n);
1579 printf(" Outermost block: "); DDMN(early);
1580 printf(" Best block: "); DDMN(best);
1581 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1583 set_nodes_block(n, best);
1588 * Find the latest legal block for N and place N into the
1589 * `optimal' Block between the latest and earliest legal block.
1590 * The `optimal' block is the dominance-deepest block of those
1591 * with the least loop-nesting-depth. This places N out of as many
1592 * loops as possible and then makes it as control dependant as
1596 place_floats_late(ir_node *n, pdeq *worklist)
1601 assert (irn_not_visited(n)); /* no multiple placement */
1603 mark_irn_visited(n);
1605 /* no need to place block nodes, control nodes are already placed. */
1606 if ((get_irn_op(n) != op_Block) &&
1608 (get_irn_mode(n) != mode_X)) {
1609 /* Remember the early placement of this block to move it
1610 out of loop no further than the early placement. */
1611 early = get_nodes_block(n);
1613 /* Do not move code not reachable from Start. For
1614 * these we could not compute dominator information. */
1615 if (is_Bad(early) || get_Block_dom_depth(early) == -1)
1618 /* Assure that our users are all placed, except the Phi-nodes.
1619 --- Each data flow cycle contains at least one Phi-node. We
1620 have to break the `user has to be placed before the
1621 producer' dependence cycle and the Phi-nodes are the
1622 place to do so, because we need to base our placement on the
1623 final region of our users, which is OK with Phi-nodes, as they
1624 are op_pin_state_pinned, and they never have to be placed after a
1625 producer of one of their inputs in the same block anyway. */
1626 for (i = 0; i < get_irn_n_outs(n); i++) {
1627 ir_node *succ = get_irn_out(n, i);
1628 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1629 place_floats_late(succ, worklist);
1632 /* We have to determine the final block of this node... except for
1634 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1635 (get_irn_op(n) != op_Const) &&
1636 (get_irn_op(n) != op_SymConst)) {
1637 ir_node *dca = NULL; /* deepest common ancestor in the
1638 dominator tree of all nodes'
1639 blocks depending on us; our final
1640 placement has to dominate DCA. */
1641 for (i = 0; i < get_irn_n_outs(n); i++) {
1642 ir_node *out = get_irn_out(n, i);
1643 /* ignore if out is in dead code */
1644 ir_node *outbl = get_nodes_block(out);
1645 if (is_Bad(outbl) || get_Block_dom_depth(outbl) == -1)
1647 dca = consumer_dom_dca (dca, out, n);
1650 set_nodes_block(n, dca);
1652 move_out_of_loops (n, early);
1654 /* else all outs are in dead code */
1658 /* Add predecessors of all non-floating nodes on list. (Those of floating
1659 nodes are placeded already and therefore are marked.) */
1660 for (i = 0; i < get_irn_n_outs(n); i++) {
1661 ir_node *succ = get_irn_out(n, i);
1662 if (irn_not_visited(get_irn_out(n, i))) {
1663 pdeq_putr (worklist, succ);
1668 static INLINE void place_late(pdeq *worklist) {
1670 inc_irg_visited(current_ir_graph);
1672 /* This fills the worklist initially. */
1673 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1675 /* And now empty the worklist again... */
1676 while (!pdeq_empty (worklist)) {
1677 ir_node *n = pdeq_getl (worklist);
1678 if (irn_not_visited(n)) place_floats_late(n, worklist);
1682 void place_code(ir_graph *irg) {
1684 ir_graph *rem = current_ir_graph;
1686 current_ir_graph = irg;
1688 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1690 /* Handle graph state */
1691 assert(get_irg_phase_state(irg) != phase_building);
1692 if (get_irg_dom_state(irg) != dom_consistent)
1695 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1696 free_loop_information(irg);
1697 construct_backedges(irg);
1700 /* Place all floating nodes as early as possible. This guarantees
1701 a legal code placement. */
1702 worklist = new_pdeq();
1703 place_early(worklist);
1705 /* place_early invalidates the outs, place_late needs them. */
1707 /* Now move the nodes down in the dominator tree. This reduces the
1708 unnecessary executions of the node. */
1709 place_late(worklist);
1711 set_irg_outs_inconsistent(current_ir_graph);
1712 set_irg_loopinfo_inconsistent(current_ir_graph);
1714 current_ir_graph = rem;
1718 * Called by walker of remove_critical_cf_edges().
1720 * Place an empty block to an edge between a blocks of multiple
1721 * predecessors and a block of multiple successors.
1724 * @param env Environment of walker. This field is unused and has
1727 static void walk_critical_cf_edges(ir_node *n, void *env) {
1729 ir_node *pre, *block, **in, *jmp;
1731 /* Block has multiple predecessors */
1732 if ((op_Block == get_irn_op(n)) &&
1733 (get_irn_arity(n) > 1)) {
1734 arity = get_irn_arity(n);
1736 if (n == get_irg_end_block(current_ir_graph))
1737 return; /* No use to add a block here. */
1739 for (i=0; i<arity; i++) {
1740 pre = get_irn_n(n, i);
1741 /* Predecessor has multiple successors. Insert new flow edge */
1742 if ((NULL != pre) &&
1743 (op_Proj == get_irn_op(pre)) &&
1744 op_Raise != get_irn_op(skip_Proj(pre))) {
1746 /* set predecessor array for new block */
1747 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1748 /* set predecessor of new block */
1750 block = new_Block(1, in);
1751 /* insert new jmp node to new block */
1752 set_cur_block(block);
1755 /* set successor of new block */
1756 set_irn_n(n, i, jmp);
1758 } /* predecessor has multiple successors */
1759 } /* for all predecessors */
1760 } /* n is a block */
1763 void remove_critical_cf_edges(ir_graph *irg) {
1764 if (get_opt_critical_edges())
1765 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);