3 * File name: ir/ir/irgopt.c
4 * Purpose: Optimizations for a whole ir graph, i.e., a procedure.
5 * Author: Christian Schaefer, Goetz Lindenmaier
6 * Modified by: Sebastian Felis
9 * Copyright: (c) 1998-2003 Universität Karlsruhe
10 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
23 # include "irnode_t.h"
24 # include "irgraph_t.h"
27 # include "ircons_t.h"
32 # include "pdeq.h" /* Fuer code placement */
35 # include "irbackedge_t.h"
36 # include "irflag_t.h"
37 # include "firmstat.h"
40 /* Defined in iropt.c */
41 pset *new_identities (void);
42 void del_identities (pset *value_table);
43 void add_identities (pset *value_table, ir_node *node);
45 /*------------------------------------------------------------------*/
46 /* apply optimizations of iropt to all nodes. */
47 /*------------------------------------------------------------------*/
49 static void init_link (ir_node *n, void *env) {
50 set_irn_link(n, NULL);
53 #if 0 /* Old version. Avoids Ids.
54 This is not necessary: we do a postwalk, and get_irn_n
55 removes ids anyways. So it's much cheaper to call the
56 optimization less often and use the exchange() algorithm. */
58 optimize_in_place_wrapper (ir_node *n, void *env) {
60 ir_node *optimized, *old;
62 irn_arity = get_irn_arity(n);
63 for (i = 0; i < irn_arity; i++) {
64 /* get_irn_n skips Id nodes, so comparison old != optimized does not
65 show all optimizations. Therefore always set new predecessor. */
66 old = get_irn_intra_n(n, i);
67 optimized = optimize_in_place_2(old);
68 set_irn_n(n, i, optimized);
71 if (get_irn_op(n) == op_Block) {
72 optimized = optimize_in_place_2(n);
73 if (optimized != n) exchange (n, optimized);
78 optimize_in_place_wrapper (ir_node *n, void *env) {
79 ir_node *optimized = optimize_in_place_2(n);
80 if (optimized != n) exchange (n, optimized);
87 local_optimize_graph (ir_graph *irg) {
88 ir_graph *rem = current_ir_graph;
89 current_ir_graph = irg;
91 /* Handle graph state */
92 assert(get_irg_phase_state(irg) != phase_building);
93 if (get_opt_global_cse())
94 set_irg_pinned(current_ir_graph, floats);
95 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
96 set_irg_outs_inconsistent(current_ir_graph);
97 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
98 set_irg_dom_inconsistent(current_ir_graph);
99 set_irg_loopinfo_inconsistent(current_ir_graph);
102 /* Clean the value_table in irg for the cse. */
103 del_identities(irg->value_table);
104 irg->value_table = new_identities();
106 /* walk over the graph */
107 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
109 current_ir_graph = rem;
112 /*------------------------------------------------------------------*/
113 /* Routines for dead node elimination / copying garbage collection */
114 /* of the obstack. */
115 /*------------------------------------------------------------------*/
118 * Remember the new node in the old node by using a field all nodes have.
121 set_new_node (ir_node *old, ir_node *new)
127 * Get this new node, before the old node is forgotton.
129 static INLINE ir_node *
130 get_new_node (ir_node * n)
136 * We use the block_visited flag to mark that we have computed the
137 * number of useful predecessors for this block.
138 * Further we encode the new arity in this flag in the old blocks.
139 * Remembering the arity is useful, as it saves a lot of pointer
140 * accesses. This function is called for all Phi and Block nodes
144 compute_new_arity(ir_node *b) {
145 int i, res, irn_arity;
148 irg_v = get_irg_block_visited(current_ir_graph);
149 block_v = get_Block_block_visited(b);
150 if (block_v >= irg_v) {
151 /* we computed the number of preds for this block and saved it in the
153 return block_v - irg_v;
155 /* compute the number of good predecessors */
156 res = irn_arity = get_irn_arity(b);
157 for (i = 0; i < irn_arity; i++)
158 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
159 /* save it in the flag. */
160 set_Block_block_visited(b, irg_v + res);
165 /* TODO: add an ir_op operation */
166 static INLINE void new_backedge_info(ir_node *n) {
167 switch(get_irn_opcode(n)) {
169 n->attr.block.cg_backedge = NULL;
170 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
173 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
176 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
183 * Copies the node to the new obstack. The Ins of the new node point to
184 * the predecessors on the old obstack. For block/phi nodes not all
185 * predecessors might be copied. n->link points to the new node.
186 * For Phi and Block nodes the function allocates in-arrays with an arity
187 * only for useful predecessors. The arity is determined by counting
188 * the non-bad predecessors of the block.
191 copy_node (ir_node *n, void *env) {
194 opcode op = get_irn_opcode(n);
195 /* The end node looses it's flexible in array. This doesn't matter,
196 as dead node elimination builds End by hand, inlineing doesn't use
198 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
201 /* node copied already */
203 } else if (op == iro_Block) {
205 new_arity = compute_new_arity(n);
206 n->attr.block.graph_arr = NULL;
208 block = get_nodes_Block(n);
209 if (get_irn_opcode(n) == iro_Phi) {
210 new_arity = compute_new_arity(block);
212 new_arity = get_irn_arity(n);
215 nn = new_ir_node(get_irn_dbg_info(n),
222 /* Copy the attributes. These might point to additional data. If this
223 was allocated on the old obstack the pointers now are dangling. This
224 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
226 new_backedge_info(nn);
229 /* printf("\n old node: "); DDMSG2(n);
230 printf(" new node: "); DDMSG2(nn); */
235 * Copies new predecessors of old node to new node remembered in link.
236 * Spare the Bad predecessors of Phi and Block nodes.
239 copy_preds (ir_node *n, void *env) {
243 nn = get_new_node(n);
245 /* printf("\n old node: "); DDMSG2(n);
246 printf(" new node: "); DDMSG2(nn);
247 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
249 if (get_irn_opcode(n) == iro_Block) {
250 /* Don't copy Bad nodes. */
252 irn_arity = get_irn_arity(n);
253 for (i = 0; i < irn_arity; i++)
254 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
255 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
256 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
259 /* repair the block visited flag from above misuse. Repair it in both
260 graphs so that the old one can still be used. */
261 set_Block_block_visited(nn, 0);
262 set_Block_block_visited(n, 0);
263 /* Local optimization could not merge two subsequent blocks if
264 in array contained Bads. Now it's possible.
265 We don't call optimize_in_place as it requires
266 that the fields in ir_graph are set properly. */
267 if ((get_opt_control_flow_straightening()) &&
268 (get_Block_n_cfgpreds(nn) == 1) &&
269 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
270 ir_node *old = get_nodes_Block(get_Block_cfgpred(nn, 0));
272 /* Jmp jumps into the block it is in -- deal self cycle. */
273 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
274 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
279 } else if (get_irn_opcode(n) == iro_Phi) {
280 /* Don't copy node if corresponding predecessor in block is Bad.
281 The Block itself should not be Bad. */
282 block = get_nodes_Block(n);
283 set_irn_n (nn, -1, get_new_node(block));
285 irn_arity = get_irn_arity(n);
286 for (i = 0; i < irn_arity; i++)
287 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
288 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
289 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
292 /* If the pre walker reached this Phi after the post walker visited the
293 block block_visited is > 0. */
294 set_Block_block_visited(get_nodes_Block(n), 0);
295 /* Compacting the Phi's ins might generate Phis with only one
297 if (get_irn_arity(n) == 1)
298 exchange(n, get_irn_n(n, 0));
300 irn_arity = get_irn_arity(n);
301 for (i = -1; i < irn_arity; i++)
302 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
304 /* Now the new node is complete. We can add it to the hash table for cse.
305 @@@ inlinening aborts if we identify End. Why? */
306 if(get_irn_op(nn) != op_End)
307 add_identities (current_ir_graph->value_table, nn);
311 * Copies the graph recursively, compacts the keepalive of the end node.
315 ir_node *oe, *ne, *ob, *nb; /* old end, new end, old bad, new bad */
316 ir_node *ka; /* keep alive */
319 oe = get_irg_end(current_ir_graph);
320 /* copy the end node by hand, allocate dynamic in array! */
321 ne = new_ir_node(get_irn_dbg_info(oe),
328 /* Copy the attributes. Well, there might be some in the future... */
330 set_new_node(oe, ne);
332 ob = get_irg_bad(current_ir_graph);
333 nb = new_ir_node(get_irn_dbg_info(ob),
340 set_new_node(ob, nb);
342 /* copy the live nodes */
343 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
344 /* copy_preds for the end node ... */
345 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
346 set_nodes_Block(nb, get_new_node(get_nodes_Block(ob)));
348 /*- ... and now the keep alives. -*/
349 /* First pick the not marked block nodes and walk them. We must pick these
350 first as else we will oversee blocks reachable from Phis. */
351 irn_arity = get_irn_arity(oe);
352 for (i = 0; i < irn_arity; i++) {
353 ka = get_irn_intra_n(oe, i);
354 if ((get_irn_op(ka) == op_Block) &&
355 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
356 /* We must keep the block alive and copy everything reachable */
357 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
358 irg_walk(ka, copy_node, copy_preds, NULL);
359 add_End_keepalive(ne, get_new_node(ka));
363 /* Now pick the Phis. Here we will keep all! */
364 irn_arity = get_irn_arity(oe);
365 for (i = 0; i < irn_arity; i++) {
366 ka = get_irn_intra_n(oe, i);
367 if ((get_irn_op(ka) == op_Phi)) {
368 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
369 /* We didn't copy the Phi yet. */
370 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
371 irg_walk(ka, copy_node, copy_preds, NULL);
373 add_End_keepalive(ne, get_new_node(ka));
379 * Copies the graph reachable from current_ir_graph->end to the obstack
380 * in current_ir_graph and fixes the environment.
381 * Then fixes the fields in current_ir_graph containing nodes of the
385 copy_graph_env (void) {
387 /* Not all nodes remembered in current_ir_graph might be reachable
388 from the end node. Assure their link is set to NULL, so that
389 we can test whether new nodes have been computed. */
390 set_irn_link(get_irg_frame (current_ir_graph), NULL);
391 set_irn_link(get_irg_globals (current_ir_graph), NULL);
392 set_irn_link(get_irg_args (current_ir_graph), NULL);
393 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
395 /* we use the block walk flag for removing Bads from Blocks ins. */
396 inc_irg_block_visited(current_ir_graph);
401 /* fix the fields in current_ir_graph */
402 old_end = get_irg_end(current_ir_graph);
403 set_irg_end (current_ir_graph, get_new_node(old_end));
404 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
405 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
407 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
408 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
409 copy_node (get_irg_frame(current_ir_graph), NULL);
410 copy_preds(get_irg_frame(current_ir_graph), NULL);
412 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
413 copy_node (get_irg_globals(current_ir_graph), NULL);
414 copy_preds(get_irg_globals(current_ir_graph), NULL);
416 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
417 copy_node (get_irg_initial_mem(current_ir_graph), NULL);
418 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
420 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
421 copy_node (get_irg_args(current_ir_graph), NULL);
422 copy_preds(get_irg_args(current_ir_graph), NULL);
424 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
426 set_irg_start_block(current_ir_graph,
427 get_new_node(get_irg_start_block(current_ir_graph)));
428 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
429 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
430 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
431 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
433 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
434 copy_node(get_irg_bad(current_ir_graph), NULL);
435 copy_preds(get_irg_bad(current_ir_graph), NULL);
437 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
441 * Copies all reachable nodes to a new obstack. Removes bad inputs
442 * from block nodes and the corresponding inputs from Phi nodes.
443 * Merges single exit blocks with single entry blocks and removes
445 * Adds all new nodes to a new hash table for cse. Does not
446 * perform cse, so the hash table might contain common subexpressions.
449 dead_node_elimination(ir_graph *irg) {
451 int rem_ipview = interprocedural_view;
452 struct obstack *graveyard_obst = NULL;
453 struct obstack *rebirth_obst = NULL;
455 /* inform statistics that we started a dead-node elimination run */
456 stat_dead_node_elim_start(irg);
458 /* Remember external state of current_ir_graph. */
459 rem = current_ir_graph;
460 current_ir_graph = irg;
461 interprocedural_view = 0;
463 /* Handle graph state */
464 assert(get_irg_phase_state(current_ir_graph) != phase_building);
465 free_callee_info(current_ir_graph);
466 free_outs(current_ir_graph);
467 /* @@@ so far we loose loops when copying */
468 free_loop_information(current_ir_graph);
470 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
472 /* A quiet place, where the old obstack can rest in peace,
473 until it will be cremated. */
474 graveyard_obst = irg->obst;
476 /* A new obstack, where the reachable nodes will be copied to. */
477 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
478 current_ir_graph->obst = rebirth_obst;
479 obstack_init (current_ir_graph->obst);
481 /* We also need a new hash table for cse */
482 del_identities (irg->value_table);
483 irg->value_table = new_identities ();
485 /* Copy the graph from the old to the new obstack */
488 /* Free memory from old unoptimized obstack */
489 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
490 xfree (graveyard_obst); /* ... then free it. */
493 /* inform statistics that the run is over */
494 stat_dead_node_elim_stop(irg);
496 current_ir_graph = rem;
497 interprocedural_view = rem_ipview;
501 * Relink bad predeseccors of a block and store the old in array to the
502 * link field. This function is called by relink_bad_predecessors().
503 * The array of link field starts with the block operand at position 0.
504 * If block has bad predecessors, create a new in array without bad preds.
505 * Otherwise let in array untouched.
507 static void relink_bad_block_predecessors(ir_node *n, void *env) {
508 ir_node **new_in, *irn;
509 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
511 /* if link field of block is NULL, look for bad predecessors otherwise
512 this is allready done */
513 if (get_irn_op(n) == op_Block &&
514 get_irn_link(n) == NULL) {
516 /* save old predecessors in link field (position 0 is the block operand)*/
517 set_irn_link(n, (void *)get_irn_in(n));
519 /* count predecessors without bad nodes */
520 old_irn_arity = get_irn_arity(n);
521 for (i = 0; i < old_irn_arity; i++)
522 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
524 /* arity changing: set new predecessors without bad nodes */
525 if (new_irn_arity < old_irn_arity) {
526 /* get new predecessor array without Block predecessor */
527 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
529 /* set new predeseccors in array */
532 for (i = 1; i < old_irn_arity; i++) {
533 irn = get_irn_n(n, i);
534 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
537 } /* ir node has bad predecessors */
539 } /* Block is not relinked */
543 * Relinks Bad predecesors from Bocks and Phis called by walker
544 * remove_bad_predecesors(). If n is a Block, call
545 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
546 * function of Phi's Block. If this block has bad predecessors, relink preds
549 static void relink_bad_predecessors(ir_node *n, void *env) {
550 ir_node *block, **old_in;
551 int i, old_irn_arity, new_irn_arity;
553 /* relink bad predeseccors of a block */
554 if (get_irn_op(n) == op_Block)
555 relink_bad_block_predecessors(n, env);
557 /* If Phi node relink its block and its predecessors */
558 if (get_irn_op(n) == op_Phi) {
560 /* Relink predeseccors of phi's block */
561 block = get_nodes_Block(n);
562 if (get_irn_link(block) == NULL)
563 relink_bad_block_predecessors(block, env);
565 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
566 old_irn_arity = ARR_LEN(old_in);
568 /* Relink Phi predeseccors if count of predeseccors changed */
569 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
570 /* set new predeseccors in array
571 n->in[0] remains the same block */
573 for(i = 1; i < old_irn_arity; i++)
574 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
576 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
579 } /* n is a Phi node */
583 * Removes Bad Bad predecesors from Blocks and the corresponding
584 * inputs to Phi nodes as in dead_node_elimination but without
586 * On walking up set the link field to NULL, on walking down call
587 * relink_bad_predecessors() (This function stores the old in array
588 * to the link field and sets a new in array if arity of predecessors
591 void remove_bad_predecessors(ir_graph *irg) {
592 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
596 /*--------------------------------------------------------------------*/
597 /* Funcionality for inlining */
598 /*--------------------------------------------------------------------*/
601 * Copy node for inlineing. Updates attributes that change when
602 * inlineing but not for dead node elimination.
604 * Copies the node by calling copy_node and then updates the entity if
605 * it's a local one. env must be a pointer of the frame type of the
606 * inlined procedure. The new entities must be in the link field of
610 copy_node_inline (ir_node *n, void *env) {
612 type *frame_tp = (type *)env;
615 if (get_irn_op(n) == op_Sel) {
616 new = get_new_node (n);
617 assert(get_irn_op(new) == op_Sel);
618 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
619 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
621 } else if (get_irn_op(n) == op_Block) {
622 new = get_new_node (n);
623 new->attr.block.irg = current_ir_graph;
627 static void find_addr(ir_node *node, void *env)
629 if (get_irn_opcode(node) == iro_Proj) {
630 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
636 * currently, we cannot inline two cases:
637 * - call with compound arguments
638 * - graphs that take the address of a parameter
640 * check these condition here
642 static int can_inline(ir_node *call, ir_graph *called_graph)
644 type *call_type = get_Call_type(call);
645 int params, ress, i, res;
647 assert(is_method_type(call_type));
649 params = get_method_n_params(call_type);
650 ress = get_method_n_ress(call_type);
653 for (i = 0; i < params; ++i) {
654 type *p_type = get_method_param_type(call_type, i);
656 if (is_compound_type(p_type))
661 for (i = 0; i < ress; ++i) {
662 type *r_type = get_method_res_type(call_type, i);
664 if (is_compound_type(r_type))
669 irg_walk_graph(called_graph, find_addr, NULL, &res);
674 int inline_method(ir_node *call, ir_graph *called_graph) {
676 ir_node *post_call, *post_bl;
678 ir_node *end, *end_bl;
682 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
685 irg_inline_property prop = get_irg_inline_property(called_graph);
687 if ( (prop != irg_inline_forced) && (!get_opt_optimize() || !get_opt_inline() ||
688 (prop == irg_inline_forbidden))) return 0;
692 * currently, we cannot inline two cases:
693 * - call with compound arguments
694 * - graphs that take the address of a parameter
696 if (! can_inline(call, called_graph))
699 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
700 rem_opt = get_opt_optimize();
703 /* Handle graph state */
704 assert(get_irg_phase_state(current_ir_graph) != phase_building);
705 assert(get_irg_pinned(current_ir_graph) == pinned);
706 assert(get_irg_pinned(called_graph) == pinned);
707 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
708 set_irg_outs_inconsistent(current_ir_graph);
709 set_irg_loopinfo_inconsistent(current_ir_graph);
711 /* -- Check preconditions -- */
712 assert(get_irn_op(call) == op_Call);
713 /* @@@ does not work for InterfaceIII.java after cgana
714 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
715 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
716 get_Call_type(call)));
718 assert(get_type_tpop(get_Call_type(call)) == type_method);
719 if (called_graph == current_ir_graph) {
720 set_optimize(rem_opt);
724 /* here we know we WILL inline, so inform the statistics */
725 stat_inline(call, called_graph);
727 /* -- Decide how to handle exception control flow: Is there a handler
728 for the Call node, or do we branch directly to End on an exception?
729 exc_handling: 0 There is a handler.
731 2 Exception handling not represented in Firm. -- */
733 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
734 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
735 assert(get_irn_op(proj) == op_Proj);
736 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
737 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
739 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
740 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
741 else { exc_handling = 2; } /* !Mproj && !Xproj */
746 the procedure and later replaces the Start node of the called graph.
747 Post_call is the old Call node and collects the results of the called
748 graph. Both will end up being a tuple. -- */
749 post_bl = get_nodes_Block(call);
750 set_irg_current_block(current_ir_graph, post_bl);
751 /* XxMxPxP of Start + parameter of Call */
752 in[pn_Start_X_initial_exec] = new_Jmp();
753 in[pn_Start_M] = get_Call_mem(call);
754 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
755 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
756 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
757 /* in[pn_Start_P_value_arg_base] = ??? */
758 pre_call = new_Tuple(5, in);
762 The new block gets the ins of the old block, pre_call and all its
763 predecessors and all Phi nodes. -- */
764 part_block(pre_call);
766 /* -- Prepare state for dead node elimination -- */
767 /* Visited flags in calling irg must be >= flag in called irg.
768 Else walker and arity computation will not work. */
769 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
770 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
771 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
772 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
773 /* Set pre_call as new Start node in link field of the start node of
774 calling graph and pre_calls block as new block for the start block
776 Further mark these nodes so that they are not visited by the
778 set_irn_link(get_irg_start(called_graph), pre_call);
779 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
780 set_irn_link(get_irg_start_block(called_graph), get_nodes_Block(pre_call));
781 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
782 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
783 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
785 /* Initialize for compaction of in arrays */
786 inc_irg_block_visited(current_ir_graph);
788 /* -- Replicate local entities of the called_graph -- */
789 /* copy the entities. */
790 called_frame = get_irg_frame_type(called_graph);
791 for (i = 0; i < get_class_n_members(called_frame); i++) {
792 entity *new_ent, *old_ent;
793 old_ent = get_class_member(called_frame, i);
794 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
795 set_entity_link(old_ent, new_ent);
798 /* visited is > than that of called graph. With this trick visited will
799 remain unchanged so that an outer walker, e.g., searching the call nodes
800 to inline, calling this inline will not visit the inlined nodes. */
801 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
803 /* -- Performing dead node elimination inlines the graph -- */
804 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
806 /* @@@ endless loops are not copied!! -- they should be, I think... */
807 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
808 get_irg_frame_type(called_graph));
810 /* Repair called_graph */
811 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
812 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
813 set_Block_block_visited(get_irg_start_block(called_graph), 0);
815 /* -- Merge the end of the inlined procedure with the call site -- */
816 /* We will turn the old Call node into a Tuple with the following
819 0: Phi of all Memories of Return statements.
820 1: Jmp from new Block that merges the control flow from all exception
821 predecessors of the old end block.
822 2: Tuple of all arguments.
823 3: Phi of Exception memories.
824 In case the old Call directly branches to End on an exception we don't
825 need the block merging all exceptions nor the Phi of the exception
829 /* -- Precompute some values -- */
830 end_bl = get_new_node(get_irg_end_block(called_graph));
831 end = get_new_node(get_irg_end(called_graph));
832 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
833 n_res = get_method_n_ress(get_Call_type(call));
835 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
836 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
838 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
840 /* -- archive keepalives -- */
841 irn_arity = get_irn_arity(end);
842 for (i = 0; i < irn_arity; i++)
843 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
845 /* The new end node will die. We need not free as the in array is on the obstack:
846 copy_node only generated 'D' arrays. */
848 /* -- Replace Return nodes by Jump nodes. -- */
850 for (i = 0; i < arity; i++) {
852 ret = get_irn_n(end_bl, i);
853 if (get_irn_op(ret) == op_Return) {
854 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
858 set_irn_in(post_bl, n_ret, cf_pred);
860 /* -- Build a Tuple for all results of the method.
861 Add Phi node if there was more than one Return. -- */
862 turn_into_tuple(post_call, 4);
863 /* First the Memory-Phi */
865 for (i = 0; i < arity; i++) {
866 ret = get_irn_n(end_bl, i);
867 if (get_irn_op(ret) == op_Return) {
868 cf_pred[n_ret] = get_Return_mem(ret);
872 phi = new_Phi(n_ret, cf_pred, mode_M);
873 set_Tuple_pred(call, pn_Call_M_regular, phi);
874 /* Conserve Phi-list for further inlinings -- but might be optimized */
875 if (get_nodes_Block(phi) == post_bl) {
876 set_irn_link(phi, get_irn_link(post_bl));
877 set_irn_link(post_bl, phi);
879 /* Now the real results */
881 for (j = 0; j < n_res; j++) {
883 for (i = 0; i < arity; i++) {
884 ret = get_irn_n(end_bl, i);
885 if (get_irn_op(ret) == op_Return) {
886 cf_pred[n_ret] = get_Return_res(ret, j);
891 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
895 /* Conserve Phi-list for further inlinings -- but might be optimized */
896 if (get_nodes_Block(phi) == post_bl) {
897 set_irn_link(phi, get_irn_link(post_bl));
898 set_irn_link(post_bl, phi);
901 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
903 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
905 /* Finally the exception control flow.
906 We have two (three) possible situations:
907 First if the Call branches to an exception handler: We need to add a Phi node to
908 collect the memory containing the exception objects. Further we need
909 to add another block to get a correct representation of this Phi. To
910 this block we add a Jmp that resolves into the X output of the Call
911 when the Call is turned into a tuple.
912 Second the Call branches to End, the exception is not handled. Just
913 add all inlined exception branches to the End node.
914 Third: there is no Exception edge at all. Handle as case two. */
915 if (exc_handling == 0) {
917 for (i = 0; i < arity; i++) {
919 ret = get_irn_n(end_bl, i);
920 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
921 cf_pred[n_exc] = ret;
926 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
927 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
928 /* The Phi for the memories with the exception objects */
930 for (i = 0; i < arity; i++) {
932 ret = skip_Proj(get_irn_n(end_bl, i));
933 if (get_irn_op(ret) == op_Call) {
934 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
936 } else if (is_fragile_op(ret)) {
937 /* We rely that all cfops have the memory output at the same position. */
938 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
940 } else if (get_irn_op(ret) == op_Raise) {
941 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
945 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
947 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
948 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
951 ir_node *main_end_bl;
952 int main_end_bl_arity;
955 /* assert(exc_handling == 1 || no exceptions. ) */
957 for (i = 0; i < arity; i++) {
958 ir_node *ret = get_irn_n(end_bl, i);
960 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
961 cf_pred[n_exc] = ret;
965 main_end_bl = get_irg_end_block(current_ir_graph);
966 main_end_bl_arity = get_irn_arity(main_end_bl);
967 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
969 for (i = 0; i < main_end_bl_arity; ++i)
970 end_preds[i] = get_irn_n(main_end_bl, i);
971 for (i = 0; i < n_exc; ++i)
972 end_preds[main_end_bl_arity + i] = cf_pred[i];
973 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
974 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
975 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
981 #if 0 /* old. now better, correcter, faster implementation. */
983 /* -- If the exception control flow from the inlined Call directly
984 branched to the end block we now have the following control
985 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
986 remove the Jmp along with it's empty block and add Jmp's
987 predecessors as predecessors of this end block. No problem if
988 there is no exception, because then branches Bad to End which
990 @@@ can't we know this beforehand: by getting the Proj(1) from
991 the Call link list and checking whether it goes to Proj. */
992 /* find the problematic predecessor of the end block. */
993 end_bl = get_irg_end_block(current_ir_graph);
994 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
995 cf_op = get_Block_cfgpred(end_bl, i);
996 if (get_irn_op(cf_op) == op_Proj) {
997 cf_op = get_Proj_pred(cf_op);
998 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
999 /* There are unoptimized tuples from inlineing before when no exc */
1000 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1001 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1002 assert(get_irn_op(cf_op) == op_Jmp);
1008 if (i < get_Block_n_cfgpreds(end_bl)) {
1009 bl = get_nodes_Block(cf_op);
1010 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1011 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
1012 for (j = 0; j < i; j++)
1013 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1014 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1015 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1016 for (j = j; j < arity; j++)
1017 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1018 set_irn_in(end_bl, arity, cf_pred);
1020 /* Remove the exception pred from post-call Tuple. */
1021 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1026 /* -- Turn cse back on. -- */
1027 set_optimize(rem_opt);
1032 /********************************************************************/
1033 /* Apply inlineing to small methods. */
1034 /********************************************************************/
1036 /* It makes no sense to inline too many calls in one procedure. Anyways,
1037 I didn't get a version with NEW_ARR_F to run. */
1038 #define MAX_INLINE 1024
1041 * environment for inlining small irgs
1043 typedef struct _inline_env_t {
1045 ir_node *calls[MAX_INLINE];
1049 * Returns the irg called from a Call node. If the irg is not
1050 * known, NULL is returned.
1052 static ir_graph *get_call_called_irg(ir_node *call) {
1055 ir_graph *called_irg = NULL;
1057 assert(get_irn_op(call) == op_Call);
1059 addr = get_Call_ptr(call);
1060 if (get_irn_op(addr) == op_Const) {
1061 /* Check whether the constant is the pointer to a compiled entity. */
1062 tv = get_Const_tarval(addr);
1063 if (tarval_to_entity(tv))
1064 called_irg = get_entity_irg(tarval_to_entity(tv));
1069 static void collect_calls(ir_node *call, void *env) {
1070 inline_env_t *ienv = env;
1073 ir_graph *called_irg;
1075 if (get_irn_op(call) != op_Call) return;
1077 addr = get_Call_ptr(call);
1078 if (get_irn_op(addr) == op_Const) {
1079 /* Check whether the constant is the pointer to a compiled entity. */
1080 tv = get_Const_tarval(addr);
1081 if (tarval_to_entity(tv)) {
1082 called_irg = get_entity_irg(tarval_to_entity(tv));
1083 if (called_irg && ienv->pos < MAX_INLINE) {
1084 /* The Call node calls a locally defined method. Remember to inline. */
1085 ienv->calls[ienv->pos++] = call;
1092 * Inlines all small methods at call sites where the called address comes
1093 * from a Const node that references the entity representing the called
1095 * The size argument is a rough measure for the code size of the method:
1096 * Methods where the obstack containing the firm graph is smaller than
1099 void inline_small_irgs(ir_graph *irg, int size) {
1101 ir_graph *rem = current_ir_graph;
1104 if (!(get_opt_optimize() && get_opt_inline())) return;
1106 current_ir_graph = irg;
1107 /* Handle graph state */
1108 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1109 free_callee_info(current_ir_graph);
1111 /* Find Call nodes to inline.
1112 (We can not inline during a walk of the graph, as inlineing the same
1113 method several times changes the visited flag of the walked graph:
1114 after the first inlineing visited of the callee equals visited of
1115 the caller. With the next inlineing both are increased.) */
1117 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1119 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1120 /* There are calls to inline */
1121 collect_phiprojs(irg);
1122 for (i = 0; i < env.pos; i++) {
1125 tv = get_Const_tarval(get_Call_ptr(env.calls[i]));
1126 callee = get_entity_irg(tarval_to_entity(tv));
1127 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1128 (get_irg_inline_property(callee) == irg_inline_forced)) {
1129 inline_method(env.calls[i], callee);
1134 current_ir_graph = rem;
1138 * Environment for inlining irgs.
1141 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1142 int n_nodes_orig; /**< for statistics */
1143 eset *call_nodes; /**< All call nodes in this graph */
1145 int n_call_nodes_orig; /**< for statistics */
1146 int n_callers; /**< Number of known graphs that call this graphs. */
1147 int n_callers_orig; /**< for statistics */
1150 static inline_irg_env *new_inline_irg_env(void) {
1151 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1152 env->n_nodes = -2; /* uncount Start, End */
1153 env->n_nodes_orig = -2; /* uncount Start, End */
1154 env->call_nodes = eset_create();
1155 env->n_call_nodes = 0;
1156 env->n_call_nodes_orig = 0;
1158 env->n_callers_orig = 0;
1162 static void free_inline_irg_env(inline_irg_env *env) {
1163 eset_destroy(env->call_nodes);
1167 static void collect_calls2(ir_node *call, void *env) {
1168 inline_irg_env *x = (inline_irg_env *)env;
1169 ir_op *op = get_irn_op(call);
1172 /* count nodes in irg */
1173 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1178 if (op != op_Call) return;
1180 /* collect all call nodes */
1181 eset_insert(x->call_nodes, (void *)call);
1183 x->n_call_nodes_orig++;
1185 /* count all static callers */
1186 callee = get_call_called_irg(call);
1188 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1189 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1193 INLINE static int is_leave(ir_graph *irg) {
1194 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1197 INLINE static int is_smaller(ir_graph *callee, int size) {
1198 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1203 * Inlines small leave methods at call sites where the called address comes
1204 * from a Const node that references the entity representing the called
1206 * The size argument is a rough measure for the code size of the method:
1207 * Methods where the obstack containing the firm graph is smaller than
1210 void inline_leave_functions(int maxsize, int leavesize, int size) {
1211 inline_irg_env *env;
1212 int i, n_irgs = get_irp_n_irgs();
1213 ir_graph *rem = current_ir_graph;
1216 if (!(get_opt_optimize() && get_opt_inline())) return;
1218 /* extend all irgs by a temporary data structure for inlineing. */
1219 for (i = 0; i < n_irgs; ++i)
1220 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1222 /* Precompute information in temporary data structure. */
1223 for (i = 0; i < n_irgs; ++i) {
1224 current_ir_graph = get_irp_irg(i);
1225 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1226 free_callee_info(current_ir_graph);
1228 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1229 get_irg_link(current_ir_graph));
1233 Inline leaves recursively -- we might construct new leaves. */
1234 /* int itercnt = 1; */
1235 while (did_inline) {
1236 /* printf("iteration %d\n", itercnt++); */
1238 for (i = 0; i < n_irgs; ++i) {
1241 int phiproj_computed = 0;
1243 current_ir_graph = get_irp_irg(i);
1244 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1246 /* we can not walk and change a set, nor remove from it.
1248 walkset = env->call_nodes;
1249 env->call_nodes = eset_create();
1250 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1251 inline_irg_env *callee_env;
1252 ir_graph *callee = get_call_called_irg(call);
1254 if (env->n_nodes > maxsize) break;
1256 ((is_leave(callee) && is_smaller(callee, leavesize)) ||
1257 (get_irg_inline_property(callee) == irg_inline_forced))) {
1258 if (!phiproj_computed) {
1259 phiproj_computed = 1;
1260 collect_phiprojs(current_ir_graph);
1262 callee_env = (inline_irg_env *)get_irg_link(callee);
1263 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1264 /* get_entity_name(get_irg_entity(callee))); */
1265 if (inline_method(call, callee)) {
1267 env->n_call_nodes--;
1268 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1269 env->n_call_nodes += callee_env->n_call_nodes;
1270 env->n_nodes += callee_env->n_nodes;
1271 callee_env->n_callers--;
1274 eset_insert(env->call_nodes, call);
1277 eset_destroy(walkset);
1281 /* printf("Non leaves\n"); */
1282 /* inline other small functions. */
1283 for (i = 0; i < n_irgs; ++i) {
1286 int phiproj_computed = 0;
1288 current_ir_graph = get_irp_irg(i);
1289 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1291 /* we can not walk and change a set, nor remove from it.
1293 walkset = env->call_nodes;
1294 env->call_nodes = eset_create();
1295 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1296 inline_irg_env *callee_env;
1297 ir_graph *callee = get_call_called_irg(call);
1299 if (env->n_nodes > maxsize) break;
1300 if (callee && is_smaller(callee, size)) {
1301 if (!phiproj_computed) {
1302 phiproj_computed = 1;
1303 collect_phiprojs(current_ir_graph);
1305 callee_env = (inline_irg_env *)get_irg_link(callee);
1306 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1307 /* get_entity_name(get_irg_entity(callee))); */
1308 if (inline_method(call, callee)) {
1310 env->n_call_nodes--;
1311 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1312 env->n_call_nodes += callee_env->n_call_nodes;
1313 env->n_nodes += callee_env->n_nodes;
1314 callee_env->n_callers--;
1317 eset_insert(env->call_nodes, call);
1320 eset_destroy(walkset);
1323 for (i = 0; i < n_irgs; ++i) {
1324 current_ir_graph = get_irp_irg(i);
1326 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1327 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1328 (env->n_callers_orig != env->n_callers))
1329 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1330 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1331 env->n_callers_orig, env->n_callers,
1332 get_entity_name(get_irg_entity(current_ir_graph)));
1334 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1337 current_ir_graph = rem;
1340 /*******************************************************************/
1341 /* Code Placement. Pins all floating nodes to a block where they */
1342 /* will be executed only if needed. */
1343 /*******************************************************************/
1346 * Find the earliest correct block for N. --- Place N into the
1347 * same Block as its dominance-deepest Input.
1350 place_floats_early(ir_node *n, pdeq *worklist)
1352 int i, start, irn_arity;
1354 /* we must not run into an infinite loop */
1355 assert (irn_not_visited(n));
1356 mark_irn_visited(n);
1358 /* Place floating nodes. */
1359 if (get_op_pinned(get_irn_op(n)) == floats) {
1361 ir_node *b = new_Bad(); /* The block to place this node in */
1362 int bad_recursion = is_Bad(get_nodes_block(n));
1364 assert(get_irn_op(n) != op_Block);
1366 if ((get_irn_op(n) == op_Const) ||
1367 (get_irn_op(n) == op_SymConst) ||
1369 (get_irn_op(n) == op_Unknown)) {
1370 /* These nodes will not be placed by the loop below. */
1371 b = get_irg_start_block(current_ir_graph);
1375 /* find the block for this node. */
1376 irn_arity = get_irn_arity(n);
1377 for (i = 0; i < irn_arity; i++) {
1378 ir_node *dep = get_irn_n(n, i);
1381 if ((irn_not_visited(dep))
1382 && (get_op_pinned(get_irn_op(dep)) == floats)) {
1383 place_floats_early(dep, worklist);
1387 * A node in the Bad block must stay in the bad block,
1388 * so don't compute a new block for it.
1393 /* Because all loops contain at least one pinned node, now all
1394 our inputs are either pinned or place_early has already
1395 been finished on them. We do not have any unfinished inputs! */
1396 dep_block = get_nodes_Block(dep);
1397 if ((!is_Bad(dep_block)) &&
1398 (get_Block_dom_depth(dep_block) > depth)) {
1400 depth = get_Block_dom_depth(dep_block);
1402 /* Avoid that the node is placed in the Start block */
1403 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
1404 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1405 assert(b != get_irg_start_block(current_ir_graph));
1409 set_nodes_Block(n, b);
1412 /* Add predecessors of non floating nodes on worklist. */
1413 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1414 irn_arity = get_irn_arity(n);
1415 for (i = start; i < irn_arity; i++) {
1416 ir_node *pred = get_irn_n(n, i);
1417 if (irn_not_visited(pred)) {
1418 pdeq_putr (worklist, pred);
1424 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1425 * Start, Call and that end at pinned nodes as Store, Call. Place_early
1426 * places all floating nodes reachable from its argument through floating
1427 * nodes and adds all beginnings at pinned nodes to the worklist.
1429 static INLINE void place_early(pdeq* worklist) {
1431 inc_irg_visited(current_ir_graph);
1433 /* this inits the worklist */
1434 place_floats_early(get_irg_end(current_ir_graph), worklist);
1436 /* Work the content of the worklist. */
1437 while (!pdeq_empty (worklist)) {
1438 ir_node *n = pdeq_getl (worklist);
1439 if (irn_not_visited(n)) place_floats_early(n, worklist);
1442 set_irg_outs_inconsistent(current_ir_graph);
1443 current_ir_graph->pinned = pinned;
1447 /** deepest common dominance ancestor of DCA and CONSUMER of PRODUCER. */
1449 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1451 ir_node *block = NULL;
1453 /* Compute the latest block into which we can place a node so that it is
1455 if (get_irn_op(consumer) == op_Phi) {
1456 /* our consumer is a Phi-node, the effective use is in all those
1457 blocks through which the Phi-node reaches producer */
1459 ir_node *phi_block = get_nodes_Block(consumer);
1460 irn_arity = get_irn_arity(consumer);
1461 for (i = 0; i < irn_arity; i++) {
1462 if (get_irn_n(consumer, i) == producer) {
1463 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1467 assert(is_no_Block(consumer));
1468 block = get_nodes_Block(consumer);
1471 /* Compute the deepest common ancestor of block and dca. */
1473 if (!dca) return block;
1474 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1475 block = get_Block_idom(block);
1476 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1477 dca = get_Block_idom(dca);
1478 while (block != dca)
1479 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1484 static INLINE int get_irn_loop_depth(ir_node *n) {
1485 return get_loop_depth(get_irn_loop(n));
1489 * Move n to a block with less loop depth than it's current block. The
1490 * new block must be dominated by early.
1493 move_out_of_loops (ir_node *n, ir_node *early)
1495 ir_node *best, *dca;
1499 /* Find the region deepest in the dominator tree dominating
1500 dca with the least loop nesting depth, but still dominated
1501 by our early placement. */
1502 dca = get_nodes_Block(n);
1504 while (dca != early) {
1505 dca = get_Block_idom(dca);
1506 if (!dca) break; /* should we put assert(dca)? */
1507 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1511 if (best != get_nodes_Block(n)) {
1513 printf("Moving out of loop: "); DDMN(n);
1514 printf(" Outermost block: "); DDMN(early);
1515 printf(" Best block: "); DDMN(best);
1516 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1518 set_nodes_Block(n, best);
1523 * Find the latest legal block for N and place N into the
1524 * `optimal' Block between the latest and earliest legal block.
1525 * The `optimal' block is the dominance-deepest block of those
1526 * with the least loop-nesting-depth. This places N out of as many
1527 * loops as possible and then makes it as control dependant as
1531 place_floats_late(ir_node *n, pdeq *worklist)
1536 assert (irn_not_visited(n)); /* no multiple placement */
1538 /* no need to place block nodes, control nodes are already placed. */
1539 if ((get_irn_op(n) != op_Block) &&
1541 (get_irn_mode(n) != mode_X)) {
1542 /* Remember the early placement of this block to move it
1543 out of loop no further than the early placement. */
1544 early = get_nodes_Block(n);
1545 /* Assure that our users are all placed, except the Phi-nodes.
1546 --- Each data flow cycle contains at least one Phi-node. We
1547 have to break the `user has to be placed before the
1548 producer' dependence cycle and the Phi-nodes are the
1549 place to do so, because we need to base our placement on the
1550 final region of our users, which is OK with Phi-nodes, as they
1551 are pinned, and they never have to be placed after a
1552 producer of one of their inputs in the same block anyway. */
1553 for (i = 0; i < get_irn_n_outs(n); i++) {
1554 ir_node *succ = get_irn_out(n, i);
1555 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1556 place_floats_late(succ, worklist);
1559 /* We have to determine the final block of this node... except for
1561 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1562 (get_irn_op(n) != op_Const) &&
1563 (get_irn_op(n) != op_SymConst)) {
1564 ir_node *dca = NULL; /* deepest common ancestor in the
1565 dominator tree of all nodes'
1566 blocks depending on us; our final
1567 placement has to dominate DCA. */
1568 for (i = 0; i < get_irn_n_outs(n); i++) {
1569 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1571 set_nodes_Block(n, dca);
1573 move_out_of_loops (n, early);
1577 mark_irn_visited(n);
1579 /* Add predecessors of all non-floating nodes on list. (Those of floating
1580 nodes are placeded already and therefore are marked.) */
1581 for (i = 0; i < get_irn_n_outs(n); i++) {
1582 if (irn_not_visited(get_irn_out(n, i))) {
1583 pdeq_putr (worklist, get_irn_out(n, i));
1588 static INLINE void place_late(pdeq* worklist) {
1590 inc_irg_visited(current_ir_graph);
1592 /* This fills the worklist initially. */
1593 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1594 /* And now empty the worklist again... */
1595 while (!pdeq_empty (worklist)) {
1596 ir_node *n = pdeq_getl (worklist);
1597 if (irn_not_visited(n)) place_floats_late(n, worklist);
1601 void place_code(ir_graph *irg) {
1603 ir_graph *rem = current_ir_graph;
1605 current_ir_graph = irg;
1607 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1609 /* Handle graph state */
1610 assert(get_irg_phase_state(irg) != phase_building);
1611 if (get_irg_dom_state(irg) != dom_consistent)
1614 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1615 free_loop_information(irg);
1616 construct_backedges(irg);
1619 /* Place all floating nodes as early as possible. This guarantees
1620 a legal code placement. */
1621 worklist = new_pdeq();
1622 place_early(worklist);
1624 /* place_early invalidates the outs, place_late needs them. */
1626 /* Now move the nodes down in the dominator tree. This reduces the
1627 unnecessary executions of the node. */
1628 place_late(worklist);
1630 set_irg_outs_inconsistent(current_ir_graph);
1631 set_irg_loopinfo_inconsistent(current_ir_graph);
1633 current_ir_graph = rem;
1638 /********************************************************************/
1639 /* Control flow optimization. */
1640 /* Removes Bad control flow predecessors and empty blocks. A block */
1641 /* is empty if it contains only a Jmp node. */
1642 /* Blocks can only be removed if they are not needed for the */
1643 /* semantics of Phi nodes. */
1644 /********************************************************************/
1647 * Removes Tuples from Block control flow predecessors.
1648 * Optimizes blocks with equivalent_node().
1649 * Replaces n by Bad if n is unreachable control flow.
1651 static void merge_blocks(ir_node *n, void *env) {
1653 set_irn_link(n, NULL);
1655 if (get_irn_op(n) == op_Block) {
1657 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1658 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
1659 A different order of optimizations might cause problems. */
1660 if (get_opt_normalize())
1661 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1662 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
1663 /* We will soon visit a block. Optimize it before visiting! */
1664 ir_node *b = get_nodes_Block(n);
1665 ir_node *new_node = equivalent_node(b);
1666 while (irn_not_visited(b) && (!is_Bad(new_node)) && (new_node != b)) {
1667 /* We would have to run gigo if new is bad, so we
1668 promote it directly below. */
1669 assert(((b == new_node) ||
1670 get_opt_control_flow_straightening() ||
1671 get_opt_control_flow_weak_simplification()) &&
1672 ("strange flag setting"));
1673 exchange (b, new_node);
1675 new_node = equivalent_node(b);
1677 if (is_Bad(new_node) && get_opt_normalize()) exchange(n, new_Bad());
1682 * Collects all Phi nodes in link list of Block.
1683 * Marks all blocks "block_visited" if they contain a node other
1686 static void collect_nodes(ir_node *n, void *env) {
1687 if (is_no_Block(n)) {
1688 ir_node *b = get_nodes_Block(n);
1690 if ((get_irn_op(n) == op_Phi)) {
1691 /* Collect Phi nodes to compact ins along with block's ins. */
1692 set_irn_link(n, get_irn_link(b));
1694 } else if ((get_irn_op(n) != op_Jmp) && !is_Bad(b)) { /* Check for non empty block. */
1695 mark_Block_block_visited(b);
1700 /** Returns true if pred is predecessor of block. */
1701 static int is_pred_of(ir_node *pred, ir_node *b) {
1703 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1704 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1705 if (b_pred == pred) return 1;
1710 static int test_whether_dispensable(ir_node *b, int pos) {
1711 int i, j, n_preds = 1;
1712 int dispensable = 1;
1713 ir_node *cfop = get_Block_cfgpred(b, pos);
1714 ir_node *pred = get_nodes_Block(cfop);
1716 if (get_Block_block_visited(pred) + 1
1717 < get_irg_block_visited(current_ir_graph)) {
1718 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
1719 /* Mark block so that is will not be removed. */
1720 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1723 /* Seems to be empty. */
1724 if (!get_irn_link(b)) {
1725 /* There are no Phi nodes ==> dispensable. */
1726 n_preds = get_Block_n_cfgpreds(pred);
1728 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1729 Work preds < pos as if they were already removed. */
1730 for (i = 0; i < pos; i++) {
1731 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1732 if (get_Block_block_visited(b_pred) + 1
1733 < get_irg_block_visited(current_ir_graph)) {
1734 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1735 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1736 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1739 if (is_pred_of(b_pred, pred)) dispensable = 0;
1742 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1743 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1744 if (is_pred_of(b_pred, pred)) dispensable = 0;
1747 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1750 n_preds = get_Block_n_cfgpreds(pred);
1758 static void optimize_blocks(ir_node *b, void *env) {
1759 int i, j, k, max_preds, n_preds;
1760 ir_node *pred, *phi;
1763 /* Count the number of predecessor if this block is merged with pred blocks
1766 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1767 max_preds += test_whether_dispensable(b, i);
1769 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1772 printf(" working on "); DDMN(b);
1773 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1774 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1775 if (is_Bad(get_Block_cfgpred(b, i))) {
1776 printf(" removing Bad %i\n ", i);
1777 } else if (get_Block_block_visited(pred) +1
1778 < get_irg_block_visited(current_ir_graph)) {
1779 printf(" removing pred %i ", i); DDMN(pred);
1780 } else { printf(" Nothing to do for "); DDMN(pred); }
1782 * end Debug output -*/
1784 /*- Fix the Phi nodes -*/
1785 phi = get_irn_link(b);
1787 assert(get_irn_op(phi) == op_Phi);
1788 /* Find the new predecessors for the Phi */
1790 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1791 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1792 if (is_Bad(get_Block_cfgpred(b, i))) {
1794 } else if (get_Block_block_visited(pred) +1
1795 < get_irg_block_visited(current_ir_graph)) {
1796 /* It's an empty block and not yet visited. */
1797 ir_node *phi_pred = get_Phi_pred(phi, i);
1798 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1799 if (get_nodes_Block(phi_pred) == pred) {
1800 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1801 in[n_preds] = get_Phi_pred(phi_pred, j);
1803 in[n_preds] = phi_pred;
1807 /* The Phi_pred node is replaced now if it is a Phi.
1808 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1809 Daher muss der Phiknoten durch den neuen ersetzt werden.
1810 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1811 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1812 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1814 if (get_nodes_Block(phi_pred) == pred) {
1815 /* remove the Phi as it might be kept alive. Further there
1816 might be other users. */
1817 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1820 in[n_preds] = get_Phi_pred(phi, i);
1825 set_irn_in(phi, n_preds, in);
1827 phi = get_irn_link(phi);
1831 This happens only if merge between loop backedge and single loop entry. -*/
1832 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1833 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1834 if (get_Block_block_visited(pred)+1 < get_irg_block_visited(current_ir_graph)) {
1835 phi = get_irn_link(pred);
1837 if (get_irn_op(phi) == op_Phi) {
1838 set_nodes_Block(phi, b);
1841 for (i = 0; i < k; i++) {
1842 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1843 if (is_Bad(get_Block_cfgpred(b, i))) {
1845 } else if (get_Block_block_visited(pred) +1
1846 < get_irg_block_visited(current_ir_graph)) {
1847 /* It's an empty block and not yet visited. */
1848 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1849 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1850 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1851 Anweisungen.) Trotzdem tuts bisher!! */
1860 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1861 in[n_preds] = get_Phi_pred(phi, i);
1864 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1865 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1866 if (is_Bad(get_Block_cfgpred(b, i))) {
1868 } else if (get_Block_block_visited(pred) +1
1869 < get_irg_block_visited(current_ir_graph)) {
1870 /* It's an empty block and not yet visited. */
1871 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1880 set_irn_in(phi, n_preds, in);
1882 phi = get_irn_link(phi);
1887 /*- Fix the block -*/
1889 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1890 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1891 if (is_Bad(get_Block_cfgpred(b, i))) {
1893 } else if (get_Block_block_visited(pred) +1
1894 < get_irg_block_visited(current_ir_graph)) {
1895 /* It's an empty block and not yet visited. */
1896 assert(get_Block_n_cfgpreds(b) > 1);
1897 /* Else it should be optimized by equivalent_node. */
1898 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1899 in[n_preds] = get_Block_cfgpred(pred, j);
1902 /* Remove block as it might be kept alive. */
1903 exchange(pred, b/*new_Bad()*/);
1905 in[n_preds] = get_Block_cfgpred(b, i);
1909 set_irn_in(b, n_preds, in);
1913 void optimize_cf(ir_graph *irg) {
1916 ir_node *end = get_irg_end(irg);
1917 ir_graph *rem = current_ir_graph;
1918 current_ir_graph = irg;
1920 /* Handle graph state */
1921 assert(get_irg_phase_state(irg) != phase_building);
1922 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
1923 set_irg_outs_inconsistent(current_ir_graph);
1924 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
1925 set_irg_dom_inconsistent(current_ir_graph);
1927 /* Use block visited flag to mark non-empty blocks. */
1928 inc_irg_block_visited(irg);
1929 irg_walk(end, merge_blocks, collect_nodes, NULL);
1931 /* Optimize the standard code. */
1932 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
1934 /* Walk all keep alives, optimize them if block, add to new in-array
1935 for end if useful. */
1936 in = NEW_ARR_F (ir_node *, 1);
1937 in[0] = get_nodes_Block(end);
1938 inc_irg_visited(current_ir_graph);
1939 for(i = 0; i < get_End_n_keepalives(end); i++) {
1940 ir_node *ka = get_End_keepalive(end, i);
1941 if (irn_not_visited(ka)) {
1942 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
1943 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
1944 get_irg_block_visited(current_ir_graph)-1);
1945 irg_block_walk(ka, optimize_blocks, NULL, NULL);
1946 mark_irn_visited(ka);
1947 ARR_APP1 (ir_node *, in, ka);
1948 } else if (get_irn_op(ka) == op_Phi) {
1949 mark_irn_visited(ka);
1950 ARR_APP1 (ir_node *, in, ka);
1954 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
1957 current_ir_graph = rem;
1962 * Called by walker of remove_critical_cf_edges().
1964 * Place an empty block to an edge between a blocks of multiple
1965 * predecessors and a block of multiple successors.
1968 * @param env Environment of walker. This field is unused and has
1971 static void walk_critical_cf_edges(ir_node *n, void *env) {
1973 ir_node *pre, *block, **in, *jmp;
1975 /* Block has multiple predecessors */
1976 if ((op_Block == get_irn_op(n)) &&
1977 (get_irn_arity(n) > 1)) {
1978 arity = get_irn_arity(n);
1980 if (n == get_irg_end_block(current_ir_graph))
1981 return; /* No use to add a block here. */
1983 for (i=0; i<arity; i++) {
1984 pre = get_irn_n(n, i);
1985 /* Predecessor has multiple successors. Insert new flow edge */
1986 if ((NULL != pre) &&
1987 (op_Proj == get_irn_op(pre)) &&
1988 op_Raise != get_irn_op(skip_Proj(pre))) {
1990 /* set predecessor array for new block */
1991 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
1992 /* set predecessor of new block */
1994 block = new_Block(1, in);
1995 /* insert new jmp node to new block */
1996 switch_block(block);
1999 /* set successor of new block */
2000 set_irn_n(n, i, jmp);
2002 } /* predecessor has multiple successors */
2003 } /* for all predecessors */
2004 } /* n is a block */
2007 void remove_critical_cf_edges(ir_graph *irg) {
2008 if (get_opt_critical_edges())
2009 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);