3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
21 #include "irgraph_t.h"
33 #include "pdeq.h" /* Fuer code placement */
38 #include "irbackedge_t.h"
44 #include "iredges_t.h"
47 /* Defined in iropt.c */
48 pset *new_identities (void);
49 void del_identities (pset *value_table);
50 void add_identities (pset *value_table, ir_node *node);
52 /*------------------------------------------------------------------*/
53 /* apply optimizations of iropt to all nodes. */
54 /*------------------------------------------------------------------*/
56 static void init_link (ir_node *n, void *env) {
57 set_irn_link(n, NULL);
60 #if 0 /* Old version. Avoids Ids.
61 This is not necessary: we do a post walk, and get_irn_n
62 removes ids anyways. So it's much cheaper to call the
63 optimization less often and use the exchange() algorithm. */
65 optimize_in_place_wrapper (ir_node *n, void *env) {
67 ir_node *optimized, *old;
69 irn_arity = get_irn_arity(n);
70 for (i = 0; i < irn_arity; i++) {
71 /* get_irn_n skips Id nodes, so comparison old != optimized does not
72 show all optimizations. Therefore always set new predecessor. */
73 old = get_irn_intra_n(n, i);
74 optimized = optimize_in_place_2(old);
75 set_irn_n(n, i, optimized);
78 if (get_irn_op(n) == op_Block) {
79 optimized = optimize_in_place_2(n);
80 if (optimized != n) exchange (n, optimized);
85 optimize_in_place_wrapper (ir_node *n, void *env) {
86 ir_node *optimized = optimize_in_place_2(n);
87 if (optimized != n) exchange (n, optimized);
92 static INLINE void do_local_optimize(ir_node *n) {
93 /* Handle graph state */
94 assert(get_irg_phase_state(current_ir_graph) != phase_building);
95 if (get_opt_global_cse())
96 set_irg_pinned(current_ir_graph, op_pin_state_floats);
97 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
98 set_irg_outs_inconsistent(current_ir_graph);
99 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
100 set_irg_dom_inconsistent(current_ir_graph);
101 set_irg_loopinfo_inconsistent(current_ir_graph);
104 /* Clean the value_table in irg for the CSE. */
105 del_identities(current_ir_graph->value_table);
106 current_ir_graph->value_table = new_identities();
108 /* walk over the graph */
109 irg_walk(n, init_link, optimize_in_place_wrapper, NULL);
112 void local_optimize_node(ir_node *n) {
113 ir_graph *rem = current_ir_graph;
114 current_ir_graph = get_irn_irg(n);
116 do_local_optimize(n);
118 current_ir_graph = rem;
122 local_optimize_graph (ir_graph *irg) {
123 ir_graph *rem = current_ir_graph;
124 current_ir_graph = irg;
126 do_local_optimize(irg->end);
128 current_ir_graph = rem;
132 /*------------------------------------------------------------------*/
133 /* Routines for dead node elimination / copying garbage collection */
134 /* of the obstack. */
135 /*------------------------------------------------------------------*/
138 * Remember the new node in the old node by using a field all nodes have.
141 set_new_node (ir_node *old, ir_node *new)
147 * Get this new node, before the old node is forgotton.
149 static INLINE ir_node *
150 get_new_node (ir_node * n)
156 * We use the block_visited flag to mark that we have computed the
157 * number of useful predecessors for this block.
158 * Further we encode the new arity in this flag in the old blocks.
159 * Remembering the arity is useful, as it saves a lot of pointer
160 * accesses. This function is called for all Phi and Block nodes
164 compute_new_arity(ir_node *b) {
165 int i, res, irn_arity;
168 irg_v = get_irg_block_visited(current_ir_graph);
169 block_v = get_Block_block_visited(b);
170 if (block_v >= irg_v) {
171 /* we computed the number of preds for this block and saved it in the
173 return block_v - irg_v;
175 /* compute the number of good predecessors */
176 res = irn_arity = get_irn_arity(b);
177 for (i = 0; i < irn_arity; i++)
178 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
179 /* save it in the flag. */
180 set_Block_block_visited(b, irg_v + res);
185 /* TODO: add an ir_op operation */
186 static INLINE void new_backedge_info(ir_node *n) {
187 switch(get_irn_opcode(n)) {
189 n->attr.block.cg_backedge = NULL;
190 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
193 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
196 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
203 * Copies the node to the new obstack. The Ins of the new node point to
204 * the predecessors on the old obstack. For block/phi nodes not all
205 * predecessors might be copied. n->link points to the new node.
206 * For Phi and Block nodes the function allocates in-arrays with an arity
207 * only for useful predecessors. The arity is determined by counting
208 * the non-bad predecessors of the block.
210 * @param n The node to be copied
211 * @param env if non-NULL, the node number attribute will be copied to the new node
213 * Note: Also used for loop unrolling.
215 static void copy_node(ir_node *n, void *env) {
218 ir_op *op = get_irn_op(n);
219 int copy_node_nr = env != NULL;
221 /* The end node looses it's flexible in array. This doesn't matter,
222 as dead node elimination builds End by hand, inlineing doesn't use
224 /* assert(op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
227 /* node copied already */
229 } else if (op == op_Block) {
231 new_arity = compute_new_arity(n);
232 n->attr.block.graph_arr = NULL;
234 block = get_nodes_block(n);
236 new_arity = compute_new_arity(block);
238 new_arity = get_irn_arity(n);
241 nn = new_ir_node(get_irn_dbg_info(n),
248 /* Copy the attributes. These might point to additional data. If this
249 was allocated on the old obstack the pointers now are dangling. This
250 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
251 copy_node_attr(n, nn);
252 new_backedge_info(nn);
256 /* for easier debugging, we want to copy the node numbers too */
257 nn->node_nr = n->node_nr;
265 * Copies new predecessors of old node to new node remembered in link.
266 * Spare the Bad predecessors of Phi and Block nodes.
269 copy_preds (ir_node *n, void *env) {
273 nn = get_new_node(n);
275 /* printf("\n old node: "); DDMSG2(n);
276 printf(" new node: "); DDMSG2(nn);
277 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
280 /* Don't copy Bad nodes. */
282 irn_arity = get_irn_arity(n);
283 for (i = 0; i < irn_arity; i++)
284 if (! is_Bad(get_irn_n(n, i))) {
285 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
286 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
289 /* repair the block visited flag from above misuse. Repair it in both
290 graphs so that the old one can still be used. */
291 set_Block_block_visited(nn, 0);
292 set_Block_block_visited(n, 0);
293 /* Local optimization could not merge two subsequent blocks if
294 in array contained Bads. Now it's possible.
295 We don't call optimize_in_place as it requires
296 that the fields in ir_graph are set properly. */
297 if ((get_opt_control_flow_straightening()) &&
298 (get_Block_n_cfgpreds(nn) == 1) &&
299 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
300 ir_node *old = get_nodes_block(get_Block_cfgpred(nn, 0));
302 /* Jmp jumps into the block it is in -- deal self cycle. */
303 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
304 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
309 } else if (get_irn_op(n) == op_Phi) {
310 /* Don't copy node if corresponding predecessor in block is Bad.
311 The Block itself should not be Bad. */
312 block = get_nodes_block(n);
313 set_irn_n (nn, -1, get_new_node(block));
315 irn_arity = get_irn_arity(n);
316 for (i = 0; i < irn_arity; i++)
317 if (! is_Bad(get_irn_n(block, i))) {
318 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
319 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
322 /* If the pre walker reached this Phi after the post walker visited the
323 block block_visited is > 0. */
324 set_Block_block_visited(get_nodes_block(n), 0);
325 /* Compacting the Phi's ins might generate Phis with only one
327 if (get_irn_arity(nn) == 1)
328 exchange(nn, get_irn_n(nn, 0));
330 irn_arity = get_irn_arity(n);
331 for (i = -1; i < irn_arity; i++)
332 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
334 /* Now the new node is complete. We can add it to the hash table for CSE.
335 @@@ inlinening aborts if we identify End. Why? */
336 if (get_irn_op(nn) != op_End)
337 add_identities (current_ir_graph->value_table, nn);
341 * Copies the graph recursively, compacts the keepalive of the end node.
343 * @param copy_node_nr If non-zero, the node number will be copied
346 copy_graph (int copy_node_nr) {
347 ir_node *oe, *ne, *ob, *nb, *om, *nm; /* old end, new end, old bad, new bad, old NoMem, new NoMem */
348 ir_node *ka; /* keep alive */
351 oe = get_irg_end(current_ir_graph);
352 /* copy the end node by hand, allocate dynamic in array! */
353 ne = new_ir_node(get_irn_dbg_info(oe),
360 /* Copy the attributes. Well, there might be some in the future... */
361 copy_node_attr(oe, ne);
362 set_new_node(oe, ne);
364 /* copy the Bad node */
365 ob = get_irg_bad(current_ir_graph);
366 nb = new_ir_node(get_irn_dbg_info(ob),
373 set_new_node(ob, nb);
375 /* copy the NoMem node */
376 om = get_irg_no_mem(current_ir_graph);
377 nm = new_ir_node(get_irn_dbg_info(om),
384 set_new_node(om, nm);
386 /* copy the live nodes */
387 irg_walk(get_nodes_block(oe), copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
388 /* copy_preds for the end node ... */
389 set_nodes_block(ne, get_new_node(get_nodes_block(oe)));
391 /*- ... and now the keep alives. -*/
392 /* First pick the not marked block nodes and walk them. We must pick these
393 first as else we will oversee blocks reachable from Phis. */
394 irn_arity = get_irn_arity(oe);
395 for (i = 0; i < irn_arity; i++) {
396 ka = get_irn_intra_n(oe, i);
397 if ((get_irn_op(ka) == op_Block) &&
398 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
399 /* We must keep the block alive and copy everything reachable */
400 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
401 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
402 add_End_keepalive(ne, get_new_node(ka));
406 /* Now pick the Phis. Here we will keep all! */
407 irn_arity = get_irn_arity(oe);
408 for (i = 0; i < irn_arity; i++) {
409 ka = get_irn_intra_n(oe, i);
410 if ((get_irn_op(ka) == op_Phi)) {
411 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
412 /* We didn't copy the Phi yet. */
413 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
414 irg_walk(ka, copy_node, copy_preds, INT_TO_PTR(copy_node_nr));
416 add_End_keepalive(ne, get_new_node(ka));
420 /* start block sometimes only reached after keep alives */
421 set_nodes_block(nb, get_new_node(get_nodes_block(ob)));
422 set_nodes_block(nm, get_new_node(get_nodes_block(om)));
426 * Copies the graph reachable from current_ir_graph->end to the obstack
427 * in current_ir_graph and fixes the environment.
428 * Then fixes the fields in current_ir_graph containing nodes of the
431 * @param copy_node_nr If non-zero, the node number will be copied
434 copy_graph_env (int copy_node_nr) {
436 /* Not all nodes remembered in current_ir_graph might be reachable
437 from the end node. Assure their link is set to NULL, so that
438 we can test whether new nodes have been computed. */
439 set_irn_link(get_irg_frame (current_ir_graph), NULL);
440 set_irn_link(get_irg_globals (current_ir_graph), NULL);
441 set_irn_link(get_irg_args (current_ir_graph), NULL);
442 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
443 set_irn_link(get_irg_no_mem (current_ir_graph), NULL);
445 /* we use the block walk flag for removing Bads from Blocks ins. */
446 inc_irg_block_visited(current_ir_graph);
449 copy_graph(copy_node_nr);
451 /* fix the fields in current_ir_graph */
452 old_end = get_irg_end(current_ir_graph);
453 set_irg_end (current_ir_graph, get_new_node(old_end));
454 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
455 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
457 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
458 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
459 copy_node (get_irg_frame(current_ir_graph), INT_TO_PTR(copy_node_nr));
460 copy_preds(get_irg_frame(current_ir_graph), NULL);
462 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
463 copy_node (get_irg_globals(current_ir_graph), INT_TO_PTR(copy_node_nr));
464 copy_preds(get_irg_globals(current_ir_graph), NULL);
466 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
467 copy_node (get_irg_initial_mem(current_ir_graph), INT_TO_PTR(copy_node_nr));
468 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
470 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
471 copy_node (get_irg_args(current_ir_graph), INT_TO_PTR(copy_node_nr));
472 copy_preds(get_irg_args(current_ir_graph), NULL);
474 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
476 set_irg_start_block(current_ir_graph,
477 get_new_node(get_irg_start_block(current_ir_graph)));
478 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
479 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
480 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
481 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
483 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
484 copy_node(get_irg_bad(current_ir_graph), INT_TO_PTR(copy_node_nr));
485 copy_preds(get_irg_bad(current_ir_graph), NULL);
487 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
489 if (get_irn_link(get_irg_no_mem(current_ir_graph)) == NULL) {
490 copy_node(get_irg_no_mem(current_ir_graph), INT_TO_PTR(copy_node_nr));
491 copy_preds(get_irg_no_mem(current_ir_graph), NULL);
493 set_irg_no_mem(current_ir_graph, get_new_node(get_irg_no_mem(current_ir_graph)));
497 * Copies all reachable nodes to a new obstack. Removes bad inputs
498 * from block nodes and the corresponding inputs from Phi nodes.
499 * Merges single exit blocks with single entry blocks and removes
501 * Adds all new nodes to a new hash table for CSE. Does not
502 * perform CSE, so the hash table might contain common subexpressions.
505 dead_node_elimination(ir_graph *irg) {
507 int rem_ipview = get_interprocedural_view();
508 struct obstack *graveyard_obst = NULL;
509 struct obstack *rebirth_obst = NULL;
511 edges_init_graph(irg);
513 /* inform statistics that we started a dead-node elimination run */
514 hook_dead_node_elim_start(irg);
516 /* Remember external state of current_ir_graph. */
517 rem = current_ir_graph;
518 current_ir_graph = irg;
519 set_interprocedural_view(false);
521 /* Handle graph state */
522 assert(get_irg_phase_state(current_ir_graph) != phase_building);
523 free_callee_info(current_ir_graph);
524 free_irg_outs(current_ir_graph);
526 /* @@@ so far we loose loops when copying */
527 free_loop_information(current_ir_graph);
529 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
531 /* A quiet place, where the old obstack can rest in peace,
532 until it will be cremated. */
533 graveyard_obst = irg->obst;
535 /* A new obstack, where the reachable nodes will be copied to. */
536 rebirth_obst = xmalloc (sizeof(*rebirth_obst));
537 current_ir_graph->obst = rebirth_obst;
538 obstack_init (current_ir_graph->obst);
540 /* We also need a new hash table for cse */
541 del_identities (irg->value_table);
542 irg->value_table = new_identities ();
544 /* Copy the graph from the old to the new obstack */
547 /* Free memory from old unoptimized obstack */
548 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
549 xfree (graveyard_obst); /* ... then free it. */
552 /* inform statistics that the run is over */
553 hook_dead_node_elim_stop(irg);
555 current_ir_graph = rem;
556 set_interprocedural_view(rem_ipview);
560 * Relink bad predecessors of a block and store the old in array to the
561 * link field. This function is called by relink_bad_predecessors().
562 * The array of link field starts with the block operand at position 0.
563 * If block has bad predecessors, create a new in array without bad preds.
564 * Otherwise let in array untouched.
566 static void relink_bad_block_predecessors(ir_node *n, void *env) {
567 ir_node **new_in, *irn;
568 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
570 /* if link field of block is NULL, look for bad predecessors otherwise
571 this is already done */
572 if (get_irn_op(n) == op_Block &&
573 get_irn_link(n) == NULL) {
575 /* save old predecessors in link field (position 0 is the block operand)*/
576 set_irn_link(n, get_irn_in(n));
578 /* count predecessors without bad nodes */
579 old_irn_arity = get_irn_arity(n);
580 for (i = 0; i < old_irn_arity; i++)
581 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
583 /* arity changing: set new predecessors without bad nodes */
584 if (new_irn_arity < old_irn_arity) {
585 /* Get new predecessor array. We do not resize the array, as we must
586 keep the old one to update Phis. */
587 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
589 /* set new predecessors in array */
592 for (i = 0; i < old_irn_arity; i++) {
593 irn = get_irn_n(n, i);
595 new_in[new_irn_n] = irn;
596 is_backedge(n, i) ? set_backedge(n, new_irn_n-1) : set_not_backedge(n, new_irn_n-1);
600 //ARR_SETLEN(int, n->attr.block.backedge, new_irn_arity);
601 ARR_SHRINKLEN(n->attr.block.backedge, new_irn_arity);
604 } /* ir node has bad predecessors */
606 } /* Block is not relinked */
610 * Relinks Bad predecessors from Blocks and Phis called by walker
611 * remove_bad_predecesors(). If n is a Block, call
612 * relink_bad_block_redecessors(). If n is a Phi-node, call also the relinking
613 * function of Phi's Block. If this block has bad predecessors, relink preds
616 static void relink_bad_predecessors(ir_node *n, void *env) {
617 ir_node *block, **old_in;
618 int i, old_irn_arity, new_irn_arity;
620 /* relink bad predecessors of a block */
621 if (get_irn_op(n) == op_Block)
622 relink_bad_block_predecessors(n, env);
624 /* If Phi node relink its block and its predecessors */
625 if (get_irn_op(n) == op_Phi) {
627 /* Relink predecessors of phi's block */
628 block = get_nodes_block(n);
629 if (get_irn_link(block) == NULL)
630 relink_bad_block_predecessors(block, env);
632 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
633 old_irn_arity = ARR_LEN(old_in);
635 /* Relink Phi predecessors if count of predecessors changed */
636 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
637 /* set new predecessors in array
638 n->in[0] remains the same block */
640 for(i = 1; i < old_irn_arity; i++)
641 if (!is_Bad((ir_node *)old_in[i])) {
642 n->in[new_irn_arity] = n->in[i];
643 is_backedge(n, i) ? set_backedge(n, new_irn_arity) : set_not_backedge(n, new_irn_arity);
647 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
648 ARR_SETLEN(int, n->attr.phi_backedge, new_irn_arity);
651 } /* n is a Phi node */
655 * Removes Bad Bad predecessors from Blocks and the corresponding
656 * inputs to Phi nodes as in dead_node_elimination but without
658 * On walking up set the link field to NULL, on walking down call
659 * relink_bad_predecessors() (This function stores the old in array
660 * to the link field and sets a new in array if arity of predecessors
663 void remove_bad_predecessors(ir_graph *irg) {
664 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
668 /*--------------------------------------------------------------------*/
669 /* Funcionality for inlining */
670 /*--------------------------------------------------------------------*/
673 * Copy node for inlineing. Updates attributes that change when
674 * inlineing but not for dead node elimination.
676 * Copies the node by calling copy_node() and then updates the entity if
677 * it's a local one. env must be a pointer of the frame type of the
678 * inlined procedure. The new entities must be in the link field of
682 copy_node_inline (ir_node *n, void *env) {
684 type *frame_tp = (type *)env;
687 if (get_irn_op(n) == op_Sel) {
688 new = get_new_node (n);
689 assert(get_irn_op(new) == op_Sel);
690 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
691 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
693 } else if (get_irn_op(n) == op_Block) {
694 new = get_new_node (n);
695 new->attr.block.irg = current_ir_graph;
699 static void find_addr(ir_node *node, void *env)
701 if (get_irn_opcode(node) == iro_Proj) {
702 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
708 * currently, we cannot inline two cases:
709 * - call with compound arguments
710 * - graphs that take the address of a parameter
712 * check these conditions here
714 static int can_inline(ir_node *call, ir_graph *called_graph)
716 type *call_type = get_Call_type(call);
717 int params, ress, i, res;
718 assert(is_Method_type(call_type));
720 params = get_method_n_params(call_type);
721 ress = get_method_n_ress(call_type);
724 for (i = 0; i < params; ++i) {
725 type *p_type = get_method_param_type(call_type, i);
727 if (is_compound_type(p_type))
732 for (i = 0; i < ress; ++i) {
733 type *r_type = get_method_res_type(call_type, i);
735 if (is_compound_type(r_type))
740 irg_walk_graph(called_graph, find_addr, NULL, &res);
745 int inline_method(ir_node *call, ir_graph *called_graph) {
747 ir_node *post_call, *post_bl;
749 ir_node *end, *end_bl;
753 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
756 irg_inline_property prop = get_irg_inline_property(called_graph);
758 if ( (prop != irg_inline_forced) &&
759 (!get_opt_optimize() || !get_opt_inline() || (prop == irg_inline_forbidden))) return 0;
761 /* Do not inline variadic functions. */
762 if (get_method_variadicity(get_entity_type(get_irg_entity(called_graph))) == variadicity_variadic)
765 assert(get_method_n_params(get_entity_type(get_irg_entity(called_graph))) ==
766 get_method_n_params(get_Call_type(call)));
769 * currently, we cannot inline two cases:
770 * - call with compound arguments
771 * - graphs that take the address of a parameter
773 if (! can_inline(call, called_graph))
776 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
777 rem_opt = get_opt_optimize();
780 /* Handle graph state */
781 assert(get_irg_phase_state(current_ir_graph) != phase_building);
782 assert(get_irg_pinned(current_ir_graph) == op_pin_state_pinned);
783 assert(get_irg_pinned(called_graph) == op_pin_state_pinned);
784 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
785 set_irg_outs_inconsistent(current_ir_graph);
786 set_irg_loopinfo_inconsistent(current_ir_graph);
787 set_irg_callee_info_state(current_ir_graph, irg_callee_info_inconsistent);
789 /* -- Check preconditions -- */
790 assert(get_irn_op(call) == op_Call);
791 /* @@@ does not work for InterfaceIII.java after cgana
792 assert(get_Call_type(call) == get_entity_type(get_irg_entity(called_graph)));
793 assert(smaller_type(get_entity_type(get_irg_entity(called_graph)),
794 get_Call_type(call)));
796 assert(get_type_tpop(get_Call_type(call)) == type_method);
797 if (called_graph == current_ir_graph) {
798 set_optimize(rem_opt);
802 /* here we know we WILL inline, so inform the statistics */
803 hook_inline(call, called_graph);
805 /* -- Decide how to handle exception control flow: Is there a handler
806 for the Call node, or do we branch directly to End on an exception?
808 0 There is a handler.
810 2 Exception handling not represented in Firm. -- */
812 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
813 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
814 assert(get_irn_op(proj) == op_Proj);
815 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
816 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
818 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
819 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
820 else { exc_handling = 2; } /* !Mproj && !Xproj */
825 the procedure and later replaces the Start node of the called graph.
826 Post_call is the old Call node and collects the results of the called
827 graph. Both will end up being a tuple. -- */
828 post_bl = get_nodes_block(call);
829 set_irg_current_block(current_ir_graph, post_bl);
830 /* XxMxPxP of Start + parameter of Call */
831 in[pn_Start_X_initial_exec] = new_Jmp();
832 in[pn_Start_M] = get_Call_mem(call);
833 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
834 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
835 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
836 /* in[pn_Start_P_value_arg_base] = ??? */
837 pre_call = new_Tuple(5, in);
841 The new block gets the ins of the old block, pre_call and all its
842 predecessors and all Phi nodes. -- */
843 part_block(pre_call);
845 /* -- Prepare state for dead node elimination -- */
846 /* Visited flags in calling irg must be >= flag in called irg.
847 Else walker and arity computation will not work. */
848 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
849 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
850 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
851 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
852 /* Set pre_call as new Start node in link field of the start node of
853 calling graph and pre_calls block as new block for the start block
855 Further mark these nodes so that they are not visited by the
857 set_irn_link(get_irg_start(called_graph), pre_call);
858 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
859 set_irn_link(get_irg_start_block(called_graph), get_nodes_block(pre_call));
860 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
861 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
862 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
864 /* Initialize for compaction of in arrays */
865 inc_irg_block_visited(current_ir_graph);
867 /* -- Replicate local entities of the called_graph -- */
868 /* copy the entities. */
869 called_frame = get_irg_frame_type(called_graph);
870 for (i = 0; i < get_class_n_members(called_frame); i++) {
871 entity *new_ent, *old_ent;
872 old_ent = get_class_member(called_frame, i);
873 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
874 set_entity_link(old_ent, new_ent);
877 /* visited is > than that of called graph. With this trick visited will
878 remain unchanged so that an outer walker, e.g., searching the call nodes
879 to inline, calling this inline will not visit the inlined nodes. */
880 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
882 /* -- Performing dead node elimination inlines the graph -- */
883 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
885 /* @@@ endless loops are not copied!! -- they should be, I think... */
886 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
887 get_irg_frame_type(called_graph));
889 /* Repair called_graph */
890 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
891 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
892 set_Block_block_visited(get_irg_start_block(called_graph), 0);
894 /* -- Merge the end of the inlined procedure with the call site -- */
895 /* We will turn the old Call node into a Tuple with the following
898 0: Phi of all Memories of Return statements.
899 1: Jmp from new Block that merges the control flow from all exception
900 predecessors of the old end block.
901 2: Tuple of all arguments.
902 3: Phi of Exception memories.
903 In case the old Call directly branches to End on an exception we don't
904 need the block merging all exceptions nor the Phi of the exception
908 /* -- Precompute some values -- */
909 end_bl = get_new_node(get_irg_end_block(called_graph));
910 end = get_new_node(get_irg_end(called_graph));
911 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
912 n_res = get_method_n_ress(get_Call_type(call));
914 res_pred = xmalloc (n_res * sizeof(*res_pred));
915 cf_pred = xmalloc (arity * sizeof(*res_pred));
917 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
919 /* -- archive keepalives -- */
920 irn_arity = get_irn_arity(end);
921 for (i = 0; i < irn_arity; i++)
922 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
924 /* The new end node will die. We need not free as the in array is on the obstack:
925 copy_node() only generated 'D' arrays. */
927 /* -- Replace Return nodes by Jump nodes. -- */
929 for (i = 0; i < arity; i++) {
931 ret = get_irn_n(end_bl, i);
932 if (get_irn_op(ret) == op_Return) {
933 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_block(ret));
937 set_irn_in(post_bl, n_ret, cf_pred);
939 /* -- Build a Tuple for all results of the method.
940 Add Phi node if there was more than one Return. -- */
941 turn_into_tuple(post_call, 4);
942 /* First the Memory-Phi */
944 for (i = 0; i < arity; i++) {
945 ret = get_irn_n(end_bl, i);
946 if (get_irn_op(ret) == op_Return) {
947 cf_pred[n_ret] = get_Return_mem(ret);
951 phi = new_Phi(n_ret, cf_pred, mode_M);
952 set_Tuple_pred(call, pn_Call_M_regular, phi);
953 /* Conserve Phi-list for further inlinings -- but might be optimized */
954 if (get_nodes_block(phi) == post_bl) {
955 set_irn_link(phi, get_irn_link(post_bl));
956 set_irn_link(post_bl, phi);
958 /* Now the real results */
960 for (j = 0; j < n_res; j++) {
962 for (i = 0; i < arity; i++) {
963 ret = get_irn_n(end_bl, i);
964 if (get_irn_op(ret) == op_Return) {
965 cf_pred[n_ret] = get_Return_res(ret, j);
970 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
974 /* Conserve Phi-list for further inlinings -- but might be optimized */
975 if (get_nodes_block(phi) == post_bl) {
976 set_irn_link(phi, get_irn_link(post_bl));
977 set_irn_link(post_bl, phi);
980 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
982 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
984 /* Finally the exception control flow.
985 We have two (three) possible situations:
986 First if the Call branches to an exception handler: We need to add a Phi node to
987 collect the memory containing the exception objects. Further we need
988 to add another block to get a correct representation of this Phi. To
989 this block we add a Jmp that resolves into the X output of the Call
990 when the Call is turned into a tuple.
991 Second the Call branches to End, the exception is not handled. Just
992 add all inlined exception branches to the End node.
993 Third: there is no Exception edge at all. Handle as case two. */
994 if (exc_handling == 0) {
996 for (i = 0; i < arity; i++) {
998 ret = get_irn_n(end_bl, i);
999 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1000 cf_pred[n_exc] = ret;
1005 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
1006 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
1007 /* The Phi for the memories with the exception objects */
1009 for (i = 0; i < arity; i++) {
1011 ret = skip_Proj(get_irn_n(end_bl, i));
1012 if (get_irn_op(ret) == op_Call) {
1013 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 3);
1015 } else if (is_fragile_op(ret)) {
1016 /* We rely that all cfops have the memory output at the same position. */
1017 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 0);
1019 } else if (get_irn_op(ret) == op_Raise) {
1020 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_block(ret), ret, mode_M, 1);
1024 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
1026 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1027 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1030 ir_node *main_end_bl;
1031 int main_end_bl_arity;
1032 ir_node **end_preds;
1034 /* assert(exc_handling == 1 || no exceptions. ) */
1036 for (i = 0; i < arity; i++) {
1037 ir_node *ret = get_irn_n(end_bl, i);
1039 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
1040 cf_pred[n_exc] = ret;
1044 main_end_bl = get_irg_end_block(current_ir_graph);
1045 main_end_bl_arity = get_irn_arity(main_end_bl);
1046 end_preds = xmalloc ((n_exc + main_end_bl_arity) * sizeof(*end_preds));
1048 for (i = 0; i < main_end_bl_arity; ++i)
1049 end_preds[i] = get_irn_n(main_end_bl, i);
1050 for (i = 0; i < n_exc; ++i)
1051 end_preds[main_end_bl_arity + i] = cf_pred[i];
1052 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
1053 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1054 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
1060 #if 0 /* old. now better, correcter, faster implementation. */
1062 /* -- If the exception control flow from the inlined Call directly
1063 branched to the end block we now have the following control
1064 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
1065 remove the Jmp along with it's empty block and add Jmp's
1066 predecessors as predecessors of this end block. No problem if
1067 there is no exception, because then branches Bad to End which
1069 @@@ can't we know this beforehand: by getting the Proj(1) from
1070 the Call link list and checking whether it goes to Proj. */
1071 /* find the problematic predecessor of the end block. */
1072 end_bl = get_irg_end_block(current_ir_graph);
1073 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
1074 cf_op = get_Block_cfgpred(end_bl, i);
1075 if (get_irn_op(cf_op) == op_Proj) {
1076 cf_op = get_Proj_pred(cf_op);
1077 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1078 /* There are unoptimized tuples from inlineing before when no exc */
1079 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1080 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1081 assert(get_irn_op(cf_op) == op_Jmp);
1087 if (i < get_Block_n_cfgpreds(end_bl)) {
1088 bl = get_nodes_block(cf_op);
1089 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1090 cf_pred = xmalloc (arity * sizeof(*cf_pred));
1091 for (j = 0; j < i; j++)
1092 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1093 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1094 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1095 for (j = j; j < arity; j++)
1096 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1097 set_irn_in(end_bl, arity, cf_pred);
1099 /* Remove the exception pred from post-call Tuple. */
1100 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1105 /* -- Turn CSE back on. -- */
1106 set_optimize(rem_opt);
1111 /********************************************************************/
1112 /* Apply inlineing to small methods. */
1113 /********************************************************************/
1115 /* It makes no sense to inline too many calls in one procedure. Anyways,
1116 I didn't get a version with NEW_ARR_F to run. */
1117 #define MAX_INLINE 1024
1120 * environment for inlining small irgs
1122 typedef struct _inline_env_t {
1124 ir_node *calls[MAX_INLINE];
1128 * Returns the irg called from a Call node. If the irg is not
1129 * known, NULL is returned.
1131 static ir_graph *get_call_called_irg(ir_node *call) {
1133 ir_graph *called_irg = NULL;
1135 assert(get_irn_op(call) == op_Call);
1137 addr = get_Call_ptr(call);
1138 if ((get_irn_op(addr) == op_SymConst) && (get_SymConst_kind (addr) == symconst_addr_ent)) {
1139 called_irg = get_entity_irg(get_SymConst_entity(addr));
1145 static void collect_calls(ir_node *call, void *env) {
1148 if (get_irn_op(call) != op_Call) return;
1150 addr = get_Call_ptr(call);
1152 if (get_irn_op(addr) == op_SymConst) {
1153 if (get_SymConst_kind(addr) == symconst_addr_ent) {
1154 ir_graph *called_irg = get_entity_irg(get_SymConst_entity(addr));
1155 inline_env_t *ienv = (inline_env_t *)env;
1156 if (called_irg && ienv->pos < MAX_INLINE) {
1157 /* The Call node calls a locally defined method. Remember to inline. */
1158 ienv->calls[ienv->pos++] = call;
1165 * Inlines all small methods at call sites where the called address comes
1166 * from a Const node that references the entity representing the called
1168 * The size argument is a rough measure for the code size of the method:
1169 * Methods where the obstack containing the firm graph is smaller than
1172 void inline_small_irgs(ir_graph *irg, int size) {
1174 ir_graph *rem = current_ir_graph;
1175 inline_env_t env /* = {0, NULL}*/;
1177 if (!(get_opt_optimize() && get_opt_inline())) return;
1179 current_ir_graph = irg;
1180 /* Handle graph state */
1181 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1182 free_callee_info(current_ir_graph);
1184 /* Find Call nodes to inline.
1185 (We can not inline during a walk of the graph, as inlineing the same
1186 method several times changes the visited flag of the walked graph:
1187 after the first inlineing visited of the callee equals visited of
1188 the caller. With the next inlineing both are increased.) */
1190 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1192 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1193 /* There are calls to inline */
1194 collect_phiprojs(irg);
1195 for (i = 0; i < env.pos; i++) {
1197 callee = get_entity_irg(get_SymConst_entity(get_Call_ptr(env.calls[i])));
1198 if (((_obstack_memory_used(callee->obst) - (int)obstack_room(callee->obst)) < size) ||
1199 (get_irg_inline_property(callee) == irg_inline_forced)) {
1200 inline_method(env.calls[i], callee);
1205 current_ir_graph = rem;
1209 * Environment for inlining irgs.
1212 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1213 int n_nodes_orig; /**< for statistics */
1214 eset *call_nodes; /**< All call nodes in this graph */
1216 int n_call_nodes_orig; /**< for statistics */
1217 int n_callers; /**< Number of known graphs that call this graphs. */
1218 int n_callers_orig; /**< for statistics */
1222 * Allocate a new nvironment for inlining.
1224 static inline_irg_env *new_inline_irg_env(void) {
1225 inline_irg_env *env = xmalloc(sizeof(*env));
1226 env->n_nodes = -2; /* do not count count Start, End */
1227 env->n_nodes_orig = -2; /* do not count Start, End */
1228 env->call_nodes = eset_create();
1229 env->n_call_nodes = 0;
1230 env->n_call_nodes_orig = 0;
1232 env->n_callers_orig = 0;
1237 * destroy an environment for inlining.
1239 static void free_inline_irg_env(inline_irg_env *env) {
1240 eset_destroy(env->call_nodes);
1245 * post-walker: collect all calls in the inline-environment
1246 * of a graph and sum some statistics.
1248 static void collect_calls2(ir_node *call, void *env) {
1249 inline_irg_env *x = (inline_irg_env *)env;
1250 ir_op *op = get_irn_op(call);
1253 /* count meaningful nodes in irg */
1254 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1259 if (op != op_Call) return;
1261 /* collect all call nodes */
1262 eset_insert(x->call_nodes, call);
1264 x->n_call_nodes_orig++;
1266 /* count all static callers */
1267 callee = get_call_called_irg(call);
1269 inline_irg_env *callee_env = get_irg_link(callee);
1270 callee_env->n_callers++;
1271 callee_env->n_callers_orig++;
1276 * Returns TRUE if the number of callers in 0 in the irg's environment,
1277 * hence this irg is a leave.
1279 INLINE static int is_leave(ir_graph *irg) {
1280 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1284 * Returns TRUE if the number of callers is smaller size in the irg's environment.
1286 INLINE static int is_smaller(ir_graph *callee, int size) {
1287 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1292 * Inlines small leave methods at call sites where the called address comes
1293 * from a Const node that references the entity representing the called
1295 * The size argument is a rough measure for the code size of the method:
1296 * Methods where the obstack containing the firm graph is smaller than
1299 void inline_leave_functions(int maxsize, int leavesize, int size) {
1300 inline_irg_env *env;
1301 int i, n_irgs = get_irp_n_irgs();
1302 ir_graph *rem = current_ir_graph;
1305 if (!(get_opt_optimize() && get_opt_inline())) return;
1307 /* extend all irgs by a temporary data structure for inlining. */
1308 for (i = 0; i < n_irgs; ++i)
1309 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1311 /* Precompute information in temporary data structure. */
1312 for (i = 0; i < n_irgs; ++i) {
1313 current_ir_graph = get_irp_irg(i);
1314 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1315 free_callee_info(current_ir_graph);
1317 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1318 get_irg_link(current_ir_graph));
1321 /* -- and now inline. -- */
1323 /* Inline leaves recursively -- we might construct new leaves. */
1324 while (did_inline) {
1327 for (i = 0; i < n_irgs; ++i) {
1329 int phiproj_computed = 0;
1331 current_ir_graph = get_irp_irg(i);
1332 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1334 for (call = eset_first(env->call_nodes); call; call = eset_next(env->call_nodes)) {
1337 if (get_irn_op(call) == op_Tuple) continue; /* We already have inlined this call. */
1338 callee = get_call_called_irg(call);
1340 if (env->n_nodes > maxsize) continue; // break;
1342 if (callee && (is_leave(callee) && is_smaller(callee, leavesize))) {
1343 if (!phiproj_computed) {
1344 phiproj_computed = 1;
1345 collect_phiprojs(current_ir_graph);
1347 did_inline = inline_method(call, callee);
1350 /* Do some statistics */
1351 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1352 env->n_call_nodes --;
1353 env->n_nodes += callee_env->n_nodes;
1354 callee_env->n_callers--;
1361 /* inline other small functions. */
1362 for (i = 0; i < n_irgs; ++i) {
1365 int phiproj_computed = 0;
1367 current_ir_graph = get_irp_irg(i);
1368 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1370 /* we can not walk and change a set, nor remove from it.
1372 walkset = env->call_nodes;
1373 env->call_nodes = eset_create();
1374 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1377 if (get_irn_op(call) == op_Tuple) continue; /* We already inlined. */
1378 callee = get_call_called_irg(call);
1381 ((is_smaller(callee, size) && (env->n_nodes < maxsize)) || /* small function */
1382 (get_irg_inline_property(callee) == irg_inline_forced))) {
1383 if (!phiproj_computed) {
1384 phiproj_computed = 1;
1385 collect_phiprojs(current_ir_graph);
1387 if (inline_method(call, callee)) {
1388 inline_irg_env *callee_env = (inline_irg_env *)get_irg_link(callee);
1389 env->n_call_nodes--;
1390 eset_insert_all(env->call_nodes, callee_env->call_nodes); /* @@@ ??? This are the wrong nodes !? Not the copied ones. */
1391 env->n_call_nodes += callee_env->n_call_nodes;
1392 env->n_nodes += callee_env->n_nodes;
1393 callee_env->n_callers--;
1396 eset_insert(env->call_nodes, call);
1399 eset_destroy(walkset);
1402 for (i = 0; i < n_irgs; ++i) {
1403 current_ir_graph = get_irp_irg(i);
1405 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1406 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1407 (env->n_callers_orig != env->n_callers))
1408 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1409 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1410 env->n_callers_orig, env->n_callers,
1411 get_entity_name(get_irg_entity(current_ir_graph)));
1413 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1416 current_ir_graph = rem;
1419 /*******************************************************************/
1420 /* Code Placement. Pins all floating nodes to a block where they */
1421 /* will be executed only if needed. */
1422 /*******************************************************************/
1425 * Returns non-zero, is a block is not reachable from Start.
1428 is_Block_unreachable(ir_node *block) {
1429 return is_Block_dead(block) || get_Block_dom_depth(block) < 0;
1433 * Find the earliest correct block for N. --- Place N into the
1434 * same Block as its dominance-deepest Input.
1436 * We have to avoid calls to get_nodes_block() here
1437 * because the graph is floating.
1440 place_floats_early(ir_node *n, pdeq *worklist)
1442 int i, start, irn_arity;
1444 /* we must not run into an infinite loop */
1445 assert (irn_not_visited(n));
1446 mark_irn_visited(n);
1448 /* Place floating nodes. */
1449 if (get_irn_pinned(n) == op_pin_state_floats) {
1451 ir_node *b = NULL; /* The block to place this node in */
1452 int bad_recursion = is_Block_unreachable(get_irn_n(n, -1));
1454 assert(get_irn_op(n) != op_Block);
1456 if ((get_irn_op(n) == op_Const) ||
1457 (get_irn_op(n) == op_SymConst) ||
1459 (get_irn_op(n) == op_Unknown)) {
1460 /* These nodes will not be placed by the loop below. */
1461 b = get_irg_start_block(current_ir_graph);
1465 /* find the block for this node. */
1466 irn_arity = get_irn_arity(n);
1467 for (i = 0; i < irn_arity; i++) {
1468 ir_node *dep = get_irn_n(n, i);
1471 if ((irn_not_visited(dep))
1472 && (get_irn_pinned(dep) == op_pin_state_floats)) {
1473 place_floats_early(dep, worklist);
1477 * A node in the Bad block must stay in the bad block,
1478 * so don't compute a new block for it.
1483 /* Because all loops contain at least one op_pin_state_pinned node, now all
1484 our inputs are either op_pin_state_pinned or place_early() has already
1485 been finished on them. We do not have any unfinished inputs! */
1486 dep_block = get_irn_n(dep, -1);
1487 if ((!is_Block_dead(dep_block)) &&
1488 (get_Block_dom_depth(dep_block) > depth)) {
1490 depth = get_Block_dom_depth(dep_block);
1492 /* Avoid that the node is placed in the Start block */
1493 if ((depth == 1) && (get_Block_dom_depth(get_irn_n(n, -1)) > 1)) {
1494 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1495 assert(b != get_irg_start_block(current_ir_graph));
1500 set_nodes_block(n, b);
1503 /* Add predecessors of non floating nodes on worklist. */
1504 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1505 irn_arity = get_irn_arity(n);
1506 for (i = start; i < irn_arity; i++) {
1507 ir_node *pred = get_irn_n(n, i);
1508 if (irn_not_visited(pred)) {
1509 pdeq_putr (worklist, pred);
1515 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1516 * Start, Call and that end at op_pin_state_pinned nodes as Store, Call. Place_early
1517 * places all floating nodes reachable from its argument through floating
1518 * nodes and adds all beginnings at op_pin_state_pinned nodes to the worklist.
1520 static INLINE void place_early(pdeq *worklist) {
1522 inc_irg_visited(current_ir_graph);
1524 /* this inits the worklist */
1525 place_floats_early(get_irg_end(current_ir_graph), worklist);
1527 /* Work the content of the worklist. */
1528 while (!pdeq_empty (worklist)) {
1529 ir_node *n = pdeq_getl (worklist);
1530 if (irn_not_visited(n)) place_floats_early(n, worklist);
1533 set_irg_outs_inconsistent(current_ir_graph);
1534 current_ir_graph->op_pin_state_pinned = op_pin_state_pinned;
1538 * Compute the deepest common ancestor of block and dca.
1540 static ir_node *calc_dca(ir_node *dca, ir_node *block)
1544 /* we do not want to place nodes in dead blocks */
1545 if (is_Block_dead(block))
1548 /* We found a first legal placement. */
1549 if (!dca) return block;
1551 /* Find a placement that is dominates both, dca and block. */
1552 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1553 block = get_Block_idom(block);
1555 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block)) {
1556 dca = get_Block_idom(dca);
1559 while (block != dca)
1560 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1565 /** Deepest common dominance ancestor of DCA and CONSUMER of PRODUCER.
1566 * I.e., DCA is the block where we might place PRODUCER.
1567 * A data flow edge points from producer to consumer.
1570 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1572 ir_node *block = NULL;
1574 /* Compute the latest block into which we can place a node so that it is
1576 if (get_irn_op(consumer) == op_Phi) {
1577 /* our consumer is a Phi-node, the effective use is in all those
1578 blocks through which the Phi-node reaches producer */
1580 ir_node *phi_block = get_nodes_block(consumer);
1581 irn_arity = get_irn_arity(consumer);
1583 for (i = 0; i < irn_arity; i++) {
1584 if (get_irn_n(consumer, i) == producer) {
1585 ir_node *new_block = get_nodes_block(get_Block_cfgpred(phi_block, i));
1587 if (! is_Block_unreachable(new_block))
1588 block = calc_dca(block, new_block);
1593 block = get_irn_n(producer, -1);
1596 assert(is_no_Block(consumer));
1597 block = get_nodes_block(consumer);
1600 /* Compute the deepest common ancestor of block and dca. */
1601 return calc_dca(dca, block);
1604 /* FIXME: the name clashes here with the function from ana/field_temperature.c
1606 static INLINE int get_irn_loop_depth(ir_node *n) {
1607 return get_loop_depth(get_irn_loop(n));
1611 * Move n to a block with less loop depth than it's current block. The
1612 * new block must be dominated by early.
1614 * @param n the node that should be moved
1615 * @param early the earliest block we can n move to
1618 move_out_of_loops (ir_node *n, ir_node *early)
1620 ir_node *best, *dca;
1624 /* Find the region deepest in the dominator tree dominating
1625 dca with the least loop nesting depth, but still dominated
1626 by our early placement. */
1627 dca = get_nodes_block(n);
1630 while (dca != early) {
1631 dca = get_Block_idom(dca);
1632 if (!dca || is_Bad(dca)) break; /* may be Bad if not reachable from Start */
1633 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1637 if (best != get_nodes_block(n)) {
1639 printf("Moving out of loop: "); DDMN(n);
1640 printf(" Outermost block: "); DDMN(early);
1641 printf(" Best block: "); DDMN(best);
1642 printf(" Innermost block: "); DDMN(get_nodes_block(n));
1644 set_nodes_block(n, best);
1649 * Find the latest legal block for N and place N into the
1650 * `optimal' Block between the latest and earliest legal block.
1651 * The `optimal' block is the dominance-deepest block of those
1652 * with the least loop-nesting-depth. This places N out of as many
1653 * loops as possible and then makes it as control dependent as
1657 place_floats_late(ir_node *n, pdeq *worklist)
1662 assert (irn_not_visited(n)); /* no multiple placement */
1664 mark_irn_visited(n);
1666 /* no need to place block nodes, control nodes are already placed. */
1667 if ((get_irn_op(n) != op_Block) &&
1669 (get_irn_mode(n) != mode_X)) {
1670 /* Remember the early placement of this block to move it
1671 out of loop no further than the early placement. */
1672 early = get_irn_n(n, -1);
1675 * BEWARE: Here we also get code, that is live, but
1676 * was in a dead block. If the node is life, but because
1677 * of CSE in a dead block, we still might need it.
1680 /* Assure that our users are all placed, except the Phi-nodes.
1681 --- Each data flow cycle contains at least one Phi-node. We
1682 have to break the `user has to be placed before the
1683 producer' dependence cycle and the Phi-nodes are the
1684 place to do so, because we need to base our placement on the
1685 final region of our users, which is OK with Phi-nodes, as they
1686 are op_pin_state_pinned, and they never have to be placed after a
1687 producer of one of their inputs in the same block anyway. */
1688 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1689 ir_node *succ = get_irn_out(n, i);
1690 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1691 place_floats_late(succ, worklist);
1694 /* We have to determine the final block of this node... except for
1696 if ((get_irn_pinned(n) == op_pin_state_floats) &&
1697 (get_irn_op(n) != op_Const) &&
1698 (get_irn_op(n) != op_SymConst)) {
1699 ir_node *dca = NULL; /* deepest common ancestor in the
1700 dominator tree of all nodes'
1701 blocks depending on us; our final
1702 placement has to dominate DCA. */
1703 for (i = get_irn_n_outs(n) - 1; i >= 0; --i) {
1704 ir_node *out = get_irn_out(n, i);
1707 if (get_irn_op(out) == op_End) {
1709 * This consumer is the End node, a keep alive edge.
1710 * This is not a real consumer, so we ignore it
1715 /* ignore if out is in dead code */
1716 outbl = get_irn_n(out, -1);
1717 if (is_Block_unreachable(outbl))
1719 dca = consumer_dom_dca(dca, out, n);
1722 set_nodes_block(n, dca);
1723 move_out_of_loops (n, early);
1725 /* else all outs are in dead code */
1729 /* Add predecessors of all non-floating nodes on list. (Those of floating
1730 nodes are placed already and therefore are marked.) */
1731 for (i = 0; i < get_irn_n_outs(n); i++) {
1732 ir_node *succ = get_irn_out(n, i);
1733 if (irn_not_visited(get_irn_out(n, i))) {
1734 pdeq_putr (worklist, succ);
1739 static INLINE void place_late(pdeq *worklist) {
1741 inc_irg_visited(current_ir_graph);
1743 /* This fills the worklist initially. */
1744 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1746 /* And now empty the worklist again... */
1747 while (!pdeq_empty (worklist)) {
1748 ir_node *n = pdeq_getl (worklist);
1749 if (irn_not_visited(n)) place_floats_late(n, worklist);
1753 void place_code(ir_graph *irg) {
1755 ir_graph *rem = current_ir_graph;
1757 current_ir_graph = irg;
1759 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1761 /* Handle graph state */
1762 assert(get_irg_phase_state(irg) != phase_building);
1763 if (get_irg_dom_state(irg) != dom_consistent)
1766 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1767 free_loop_information(irg);
1768 construct_backedges(irg);
1771 /* Place all floating nodes as early as possible. This guarantees
1772 a legal code placement. */
1773 worklist = new_pdeq();
1774 place_early(worklist);
1776 /* place_early() invalidates the outs, place_late needs them. */
1777 compute_irg_outs(irg);
1778 /* Now move the nodes down in the dominator tree. This reduces the
1779 unnecessary executions of the node. */
1780 place_late(worklist);
1782 set_irg_outs_inconsistent(current_ir_graph);
1783 set_irg_loopinfo_inconsistent(current_ir_graph);
1785 current_ir_graph = rem;
1789 * Called by walker of remove_critical_cf_edges().
1791 * Place an empty block to an edge between a blocks of multiple
1792 * predecessors and a block of multiple successors.
1795 * @param env Environment of walker. This field is unused and has
1798 static void walk_critical_cf_edges(ir_node *n, void *env) {
1800 ir_node *pre, *block, **in, *jmp;
1802 /* Block has multiple predecessors */
1803 if ((op_Block == get_irn_op(n)) &&
1804 (get_irn_arity(n) > 1)) {
1805 arity = get_irn_arity(n);
1807 if (n == get_irg_end_block(current_ir_graph))
1808 return; /* No use to add a block here. */
1810 for (i=0; i<arity; i++) {
1811 pre = get_irn_n(n, i);
1812 /* Predecessor has multiple successors. Insert new flow edge */
1813 if ((NULL != pre) &&
1814 (op_Proj == get_irn_op(pre)) &&
1815 op_Raise != get_irn_op(skip_Proj(pre))) {
1817 /* set predecessor array for new block */
1818 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1819 /* set predecessor of new block */
1821 block = new_Block(1, in);
1822 /* insert new jmp node to new block */
1823 set_cur_block(block);
1826 /* set successor of new block */
1827 set_irn_n(n, i, jmp);
1829 } /* predecessor has multiple successors */
1830 } /* for all predecessors */
1831 } /* n is a block */
1834 void remove_critical_cf_edges(ir_graph *irg) {
1835 if (get_opt_critical_edges())
1836 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);