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.
22 #include "irgraph_t.h"
34 #include "pdeq.h" /* Fuer code placement */
38 #include "irbackedge_t.h"
44 /* Defined in iropt.c */
45 pset *new_identities (void);
46 void del_identities (pset *value_table);
47 void add_identities (pset *value_table, ir_node *node);
49 /*------------------------------------------------------------------*/
50 /* apply optimizations of iropt to all nodes. */
51 /*------------------------------------------------------------------*/
53 static void init_link (ir_node *n, void *env) {
54 set_irn_link(n, NULL);
57 #if 0 /* Old version. Avoids Ids.
58 This is not necessary: we do a postwalk, and get_irn_n
59 removes ids anyways. So it's much cheaper to call the
60 optimization less often and use the exchange() algorithm. */
62 optimize_in_place_wrapper (ir_node *n, void *env) {
64 ir_node *optimized, *old;
66 irn_arity = get_irn_arity(n);
67 for (i = 0; i < irn_arity; i++) {
68 /* get_irn_n skips Id nodes, so comparison old != optimized does not
69 show all optimizations. Therefore always set new predecessor. */
70 old = get_irn_intra_n(n, i);
71 optimized = optimize_in_place_2(old);
72 set_irn_n(n, i, optimized);
75 if (get_irn_op(n) == op_Block) {
76 optimized = optimize_in_place_2(n);
77 if (optimized != n) exchange (n, optimized);
82 optimize_in_place_wrapper (ir_node *n, void *env) {
83 ir_node *optimized = optimize_in_place_2(n);
84 if (optimized != n) exchange (n, optimized);
91 local_optimize_graph (ir_graph *irg) {
92 ir_graph *rem = current_ir_graph;
93 current_ir_graph = irg;
95 /* Handle graph state */
96 assert(get_irg_phase_state(irg) != phase_building);
97 if (get_opt_global_cse())
98 set_irg_pinned(current_ir_graph, floats);
99 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
100 set_irg_outs_inconsistent(current_ir_graph);
101 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
102 set_irg_dom_inconsistent(current_ir_graph);
103 set_irg_loopinfo_inconsistent(current_ir_graph);
106 /* Clean the value_table in irg for the cse. */
107 del_identities(irg->value_table);
108 irg->value_table = new_identities();
110 /* walk over the graph */
111 irg_walk(irg->end, init_link, optimize_in_place_wrapper, NULL);
113 current_ir_graph = rem;
116 /*------------------------------------------------------------------*/
117 /* Routines for dead node elimination / copying garbage collection */
118 /* of the obstack. */
119 /*------------------------------------------------------------------*/
122 * Remember the new node in the old node by using a field all nodes have.
125 set_new_node (ir_node *old, ir_node *new)
131 * Get this new node, before the old node is forgotton.
133 static INLINE ir_node *
134 get_new_node (ir_node * n)
140 * We use the block_visited flag to mark that we have computed the
141 * number of useful predecessors for this block.
142 * Further we encode the new arity in this flag in the old blocks.
143 * Remembering the arity is useful, as it saves a lot of pointer
144 * accesses. This function is called for all Phi and Block nodes
148 compute_new_arity(ir_node *b) {
149 int i, res, irn_arity;
152 irg_v = get_irg_block_visited(current_ir_graph);
153 block_v = get_Block_block_visited(b);
154 if (block_v >= irg_v) {
155 /* we computed the number of preds for this block and saved it in the
157 return block_v - irg_v;
159 /* compute the number of good predecessors */
160 res = irn_arity = get_irn_arity(b);
161 for (i = 0; i < irn_arity; i++)
162 if (get_irn_opcode(get_irn_n(b, i)) == iro_Bad) res--;
163 /* save it in the flag. */
164 set_Block_block_visited(b, irg_v + res);
169 /* TODO: add an ir_op operation */
170 static INLINE void new_backedge_info(ir_node *n) {
171 switch(get_irn_opcode(n)) {
173 n->attr.block.cg_backedge = NULL;
174 n->attr.block.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
177 n->attr.phi_backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
180 n->attr.filter.backedge = new_backedge_arr(current_ir_graph->obst, get_irn_arity(n));
187 * Copies the node to the new obstack. The Ins of the new node point to
188 * the predecessors on the old obstack. For block/phi nodes not all
189 * predecessors might be copied. n->link points to the new node.
190 * For Phi and Block nodes the function allocates in-arrays with an arity
191 * only for useful predecessors. The arity is determined by counting
192 * the non-bad predecessors of the block.
195 copy_node (ir_node *n, void *env) {
198 opcode op = get_irn_opcode(n);
199 /* The end node looses it's flexible in array. This doesn't matter,
200 as dead node elimination builds End by hand, inlineing doesn't use
202 /* assert(n->op == op_End || ((_ARR_DESCR(n->in))->cookie != ARR_F_MAGIC)); */
205 /* node copied already */
207 } else if (op == iro_Block) {
209 new_arity = compute_new_arity(n);
210 n->attr.block.graph_arr = NULL;
212 block = get_nodes_Block(n);
213 if (get_irn_opcode(n) == iro_Phi) {
214 new_arity = compute_new_arity(block);
216 new_arity = get_irn_arity(n);
219 nn = new_ir_node(get_irn_dbg_info(n),
226 /* Copy the attributes. These might point to additional data. If this
227 was allocated on the old obstack the pointers now are dangling. This
228 frees e.g. the memory of the graph_arr allocated in new_immBlock. */
230 new_backedge_info(nn);
233 /* printf("\n old node: "); DDMSG2(n);
234 printf(" new node: "); DDMSG2(nn); */
239 * Copies new predecessors of old node to new node remembered in link.
240 * Spare the Bad predecessors of Phi and Block nodes.
243 copy_preds (ir_node *n, void *env) {
247 nn = get_new_node(n);
249 /* printf("\n old node: "); DDMSG2(n);
250 printf(" new node: "); DDMSG2(nn);
251 printf(" arities: old: %d, new: %d\n", get_irn_arity(n), get_irn_arity(nn)); */
253 if (get_irn_opcode(n) == iro_Block) {
254 /* Don't copy Bad nodes. */
256 irn_arity = get_irn_arity(n);
257 for (i = 0; i < irn_arity; i++)
258 if (get_irn_opcode(get_irn_n(n, i)) != iro_Bad) {
259 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
260 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
263 /* repair the block visited flag from above misuse. Repair it in both
264 graphs so that the old one can still be used. */
265 set_Block_block_visited(nn, 0);
266 set_Block_block_visited(n, 0);
267 /* Local optimization could not merge two subsequent blocks if
268 in array contained Bads. Now it's possible.
269 We don't call optimize_in_place as it requires
270 that the fields in ir_graph are set properly. */
271 if ((get_opt_control_flow_straightening()) &&
272 (get_Block_n_cfgpreds(nn) == 1) &&
273 (get_irn_op(get_Block_cfgpred(nn, 0)) == op_Jmp)) {
274 ir_node *old = get_nodes_Block(get_Block_cfgpred(nn, 0));
276 /* Jmp jumps into the block it is in -- deal self cycle. */
277 assert(is_Bad(get_new_node(get_irg_bad(current_ir_graph))));
278 exchange(nn, get_new_node(get_irg_bad(current_ir_graph)));
283 } else if (get_irn_opcode(n) == iro_Phi) {
284 /* Don't copy node if corresponding predecessor in block is Bad.
285 The Block itself should not be Bad. */
286 block = get_nodes_Block(n);
287 set_irn_n (nn, -1, get_new_node(block));
289 irn_arity = get_irn_arity(n);
290 for (i = 0; i < irn_arity; i++)
291 if (get_irn_opcode(get_irn_n(block, i)) != iro_Bad) {
292 set_irn_n (nn, j, get_new_node(get_irn_n(n, i)));
293 /*if (is_backedge(n, i)) set_backedge(nn, j);*/
296 /* If the pre walker reached this Phi after the post walker visited the
297 block block_visited is > 0. */
298 set_Block_block_visited(get_nodes_Block(n), 0);
299 /* Compacting the Phi's ins might generate Phis with only one
301 if (get_irn_arity(n) == 1)
302 exchange(n, get_irn_n(n, 0));
304 irn_arity = get_irn_arity(n);
305 for (i = -1; i < irn_arity; i++)
306 set_irn_n (nn, i, get_new_node(get_irn_n(n, i)));
308 /* Now the new node is complete. We can add it to the hash table for cse.
309 @@@ inlinening aborts if we identify End. Why? */
310 if(get_irn_op(nn) != op_End)
311 add_identities (current_ir_graph->value_table, nn);
315 * Copies the graph recursively, compacts the keepalive of the end node.
319 ir_node *oe, *ne, *ob, *nb; /* old end, new end, old bad, new bad */
320 ir_node *ka; /* keep alive */
323 oe = get_irg_end(current_ir_graph);
324 /* copy the end node by hand, allocate dynamic in array! */
325 ne = new_ir_node(get_irn_dbg_info(oe),
332 /* Copy the attributes. Well, there might be some in the future... */
334 set_new_node(oe, ne);
336 ob = get_irg_bad(current_ir_graph);
337 nb = new_ir_node(get_irn_dbg_info(ob),
344 set_new_node(ob, nb);
346 /* copy the live nodes */
347 irg_walk(get_nodes_Block(oe), copy_node, copy_preds, NULL);
348 /* copy_preds for the end node ... */
349 set_nodes_Block(ne, get_new_node(get_nodes_Block(oe)));
350 set_nodes_Block(nb, get_new_node(get_nodes_Block(ob)));
352 /*- ... and now the keep alives. -*/
353 /* First pick the not marked block nodes and walk them. We must pick these
354 first as else we will oversee blocks reachable from Phis. */
355 irn_arity = get_irn_arity(oe);
356 for (i = 0; i < irn_arity; i++) {
357 ka = get_irn_intra_n(oe, i);
358 if ((get_irn_op(ka) == op_Block) &&
359 (get_irn_visited(ka) < get_irg_visited(current_ir_graph))) {
360 /* We must keep the block alive and copy everything reachable */
361 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
362 irg_walk(ka, copy_node, copy_preds, NULL);
363 add_End_keepalive(ne, get_new_node(ka));
367 /* Now pick the Phis. Here we will keep all! */
368 irn_arity = get_irn_arity(oe);
369 for (i = 0; i < irn_arity; i++) {
370 ka = get_irn_intra_n(oe, i);
371 if ((get_irn_op(ka) == op_Phi)) {
372 if (get_irn_visited(ka) < get_irg_visited(current_ir_graph)) {
373 /* We didn't copy the Phi yet. */
374 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
375 irg_walk(ka, copy_node, copy_preds, NULL);
377 add_End_keepalive(ne, get_new_node(ka));
383 * Copies the graph reachable from current_ir_graph->end to the obstack
384 * in current_ir_graph and fixes the environment.
385 * Then fixes the fields in current_ir_graph containing nodes of the
389 copy_graph_env (void) {
391 /* Not all nodes remembered in current_ir_graph might be reachable
392 from the end node. Assure their link is set to NULL, so that
393 we can test whether new nodes have been computed. */
394 set_irn_link(get_irg_frame (current_ir_graph), NULL);
395 set_irn_link(get_irg_globals (current_ir_graph), NULL);
396 set_irn_link(get_irg_args (current_ir_graph), NULL);
397 set_irn_link(get_irg_initial_mem(current_ir_graph), NULL);
399 /* we use the block walk flag for removing Bads from Blocks ins. */
400 inc_irg_block_visited(current_ir_graph);
405 /* fix the fields in current_ir_graph */
406 old_end = get_irg_end(current_ir_graph);
407 set_irg_end (current_ir_graph, get_new_node(old_end));
408 set_irg_end_except (current_ir_graph, get_irg_end(current_ir_graph));
409 set_irg_end_reg (current_ir_graph, get_irg_end(current_ir_graph));
411 set_irg_end_block (current_ir_graph, get_new_node(get_irg_end_block(current_ir_graph)));
412 if (get_irn_link(get_irg_frame(current_ir_graph)) == NULL) {
413 copy_node (get_irg_frame(current_ir_graph), NULL);
414 copy_preds(get_irg_frame(current_ir_graph), NULL);
416 if (get_irn_link(get_irg_globals(current_ir_graph)) == NULL) {
417 copy_node (get_irg_globals(current_ir_graph), NULL);
418 copy_preds(get_irg_globals(current_ir_graph), NULL);
420 if (get_irn_link(get_irg_initial_mem(current_ir_graph)) == NULL) {
421 copy_node (get_irg_initial_mem(current_ir_graph), NULL);
422 copy_preds(get_irg_initial_mem(current_ir_graph), NULL);
424 if (get_irn_link(get_irg_args(current_ir_graph)) == NULL) {
425 copy_node (get_irg_args(current_ir_graph), NULL);
426 copy_preds(get_irg_args(current_ir_graph), NULL);
428 set_irg_start (current_ir_graph, get_new_node(get_irg_start(current_ir_graph)));
430 set_irg_start_block(current_ir_graph,
431 get_new_node(get_irg_start_block(current_ir_graph)));
432 set_irg_frame (current_ir_graph, get_new_node(get_irg_frame(current_ir_graph)));
433 set_irg_globals (current_ir_graph, get_new_node(get_irg_globals(current_ir_graph)));
434 set_irg_initial_mem(current_ir_graph, get_new_node(get_irg_initial_mem(current_ir_graph)));
435 set_irg_args (current_ir_graph, get_new_node(get_irg_args(current_ir_graph)));
437 if (get_irn_link(get_irg_bad(current_ir_graph)) == NULL) {
438 copy_node(get_irg_bad(current_ir_graph), NULL);
439 copy_preds(get_irg_bad(current_ir_graph), NULL);
441 set_irg_bad(current_ir_graph, get_new_node(get_irg_bad(current_ir_graph)));
445 * Copies all reachable nodes to a new obstack. Removes bad inputs
446 * from block nodes and the corresponding inputs from Phi nodes.
447 * Merges single exit blocks with single entry blocks and removes
449 * Adds all new nodes to a new hash table for cse. Does not
450 * perform cse, so the hash table might contain common subexpressions.
453 dead_node_elimination(ir_graph *irg) {
455 int rem_ipview = interprocedural_view;
456 struct obstack *graveyard_obst = NULL;
457 struct obstack *rebirth_obst = NULL;
459 /* inform statistics that we started a dead-node elimination run */
460 stat_dead_node_elim_start(irg);
462 /* Remember external state of current_ir_graph. */
463 rem = current_ir_graph;
464 current_ir_graph = irg;
465 interprocedural_view = 0;
467 /* Handle graph state */
468 assert(get_irg_phase_state(current_ir_graph) != phase_building);
469 free_callee_info(current_ir_graph);
470 free_outs(current_ir_graph);
471 /* @@@ so far we loose loops when copying */
472 free_loop_information(current_ir_graph);
474 if (get_opt_optimize() && get_opt_dead_node_elimination()) {
476 /* A quiet place, where the old obstack can rest in peace,
477 until it will be cremated. */
478 graveyard_obst = irg->obst;
480 /* A new obstack, where the reachable nodes will be copied to. */
481 rebirth_obst = (struct obstack *) xmalloc (sizeof (struct obstack));
482 current_ir_graph->obst = rebirth_obst;
483 obstack_init (current_ir_graph->obst);
485 /* We also need a new hash table for cse */
486 del_identities (irg->value_table);
487 irg->value_table = new_identities ();
489 /* Copy the graph from the old to the new obstack */
492 /* Free memory from old unoptimized obstack */
493 obstack_free(graveyard_obst, 0); /* First empty the obstack ... */
494 xfree (graveyard_obst); /* ... then free it. */
497 /* inform statistics that the run is over */
498 stat_dead_node_elim_stop(irg);
500 current_ir_graph = rem;
501 interprocedural_view = rem_ipview;
505 * Relink bad predeseccors of a block and store the old in array to the
506 * link field. This function is called by relink_bad_predecessors().
507 * The array of link field starts with the block operand at position 0.
508 * If block has bad predecessors, create a new in array without bad preds.
509 * Otherwise let in array untouched.
511 static void relink_bad_block_predecessors(ir_node *n, void *env) {
512 ir_node **new_in, *irn;
513 int i, new_irn_n, old_irn_arity, new_irn_arity = 0;
515 /* if link field of block is NULL, look for bad predecessors otherwise
516 this is allready done */
517 if (get_irn_op(n) == op_Block &&
518 get_irn_link(n) == NULL) {
520 /* save old predecessors in link field (position 0 is the block operand)*/
521 set_irn_link(n, (void *)get_irn_in(n));
523 /* count predecessors without bad nodes */
524 old_irn_arity = get_irn_arity(n);
525 for (i = 0; i < old_irn_arity; i++)
526 if (!is_Bad(get_irn_n(n, i))) new_irn_arity++;
528 /* arity changing: set new predecessors without bad nodes */
529 if (new_irn_arity < old_irn_arity) {
530 /* get new predecessor array without Block predecessor */
531 new_in = NEW_ARR_D (ir_node *, current_ir_graph->obst, (new_irn_arity+1));
533 /* set new predeseccors in array */
536 for (i = 1; i < old_irn_arity; i++) {
537 irn = get_irn_n(n, i);
538 if (!is_Bad(irn)) new_in[new_irn_n++] = irn;
541 } /* ir node has bad predecessors */
543 } /* Block is not relinked */
547 * Relinks Bad predecesors from Bocks and Phis called by walker
548 * remove_bad_predecesors(). If n is a Block, call
549 * relink_bad_block_redecessors(). If n is a Phinode, call also the relinking
550 * function of Phi's Block. If this block has bad predecessors, relink preds
553 static void relink_bad_predecessors(ir_node *n, void *env) {
554 ir_node *block, **old_in;
555 int i, old_irn_arity, new_irn_arity;
557 /* relink bad predeseccors of a block */
558 if (get_irn_op(n) == op_Block)
559 relink_bad_block_predecessors(n, env);
561 /* If Phi node relink its block and its predecessors */
562 if (get_irn_op(n) == op_Phi) {
564 /* Relink predeseccors of phi's block */
565 block = get_nodes_Block(n);
566 if (get_irn_link(block) == NULL)
567 relink_bad_block_predecessors(block, env);
569 old_in = (ir_node **)get_irn_link(block); /* Of Phi's Block */
570 old_irn_arity = ARR_LEN(old_in);
572 /* Relink Phi predeseccors if count of predeseccors changed */
573 if (old_irn_arity != ARR_LEN(get_irn_in(block))) {
574 /* set new predeseccors in array
575 n->in[0] remains the same block */
577 for(i = 1; i < old_irn_arity; i++)
578 if (!is_Bad((ir_node *)old_in[i])) n->in[new_irn_arity++] = n->in[i];
580 ARR_SETLEN(ir_node *, n->in, new_irn_arity);
583 } /* n is a Phi node */
587 * Removes Bad Bad predecesors from Blocks and the corresponding
588 * inputs to Phi nodes as in dead_node_elimination but without
590 * On walking up set the link field to NULL, on walking down call
591 * relink_bad_predecessors() (This function stores the old in array
592 * to the link field and sets a new in array if arity of predecessors
595 void remove_bad_predecessors(ir_graph *irg) {
596 irg_walk_graph(irg, init_link, relink_bad_predecessors, NULL);
600 /*--------------------------------------------------------------------*/
601 /* Funcionality for inlining */
602 /*--------------------------------------------------------------------*/
605 * Copy node for inlineing. Updates attributes that change when
606 * inlineing but not for dead node elimination.
608 * Copies the node by calling copy_node and then updates the entity if
609 * it's a local one. env must be a pointer of the frame type of the
610 * inlined procedure. The new entities must be in the link field of
614 copy_node_inline (ir_node *n, void *env) {
616 type *frame_tp = (type *)env;
619 if (get_irn_op(n) == op_Sel) {
620 new = get_new_node (n);
621 assert(get_irn_op(new) == op_Sel);
622 if (get_entity_owner(get_Sel_entity(n)) == frame_tp) {
623 set_Sel_entity(new, get_entity_link(get_Sel_entity(n)));
625 } else if (get_irn_op(n) == op_Block) {
626 new = get_new_node (n);
627 new->attr.block.irg = current_ir_graph;
631 static void find_addr(ir_node *node, void *env)
633 if (get_irn_opcode(node) == iro_Proj) {
634 if (get_Proj_proj(node) == pn_Start_P_value_arg_base)
640 * currently, we cannot inline two cases:
641 * - call with compound arguments
642 * - graphs that take the address of a parameter
644 * check these condition here
646 static int can_inline(ir_node *call, ir_graph *called_graph)
648 type *call_type = get_Call_type(call);
649 int params, ress, i, res;
651 assert(is_method_type(call_type));
653 params = get_method_n_params(call_type);
654 ress = get_method_n_ress(call_type);
657 for (i = 0; i < params; ++i) {
658 type *p_type = get_method_param_type(call_type, i);
660 if (is_compound_type(p_type))
665 for (i = 0; i < ress; ++i) {
666 type *r_type = get_method_res_type(call_type, i);
668 if (is_compound_type(r_type))
673 irg_walk_graph(called_graph, find_addr, NULL, &res);
678 int inline_method(ir_node *call, ir_graph *called_graph) {
680 ir_node *post_call, *post_bl;
682 ir_node *end, *end_bl;
686 int arity, n_ret, n_exc, n_res, i, j, rem_opt, irn_arity;
689 irg_inline_property prop = get_irg_inline_property(called_graph);
691 if ( (prop != irg_inline_forced) && (!get_opt_optimize() || !get_opt_inline() ||
692 (prop == irg_inline_forbidden))) return 0;
695 * currently, we cannot inline two cases:
696 * - call with compound arguments
697 * - graphs that take the address of a parameter
699 if (! can_inline(call, called_graph))
702 /* -- Turn off optimizations, this can cause problems when allocating new nodes. -- */
703 rem_opt = get_opt_optimize();
706 /* Handle graph state */
707 assert(get_irg_phase_state(current_ir_graph) != phase_building);
708 assert(get_irg_pinned(current_ir_graph) == pinned);
709 assert(get_irg_pinned(called_graph) == pinned);
710 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
711 set_irg_outs_inconsistent(current_ir_graph);
712 set_irg_loopinfo_inconsistent(current_ir_graph);
714 /* -- Check preconditions -- */
715 assert(get_irn_op(call) == op_Call);
716 /* @@@ does not work for InterfaceIII.java after cgana
717 assert(get_Call_type(call) == get_entity_type(get_irg_ent(called_graph)));
718 assert(smaller_type(get_entity_type(get_irg_ent(called_graph)),
719 get_Call_type(call)));
721 assert(get_type_tpop(get_Call_type(call)) == type_method);
722 if (called_graph == current_ir_graph) {
723 set_optimize(rem_opt);
727 /* here we know we WILL inline, so inform the statistics */
728 stat_inline(call, called_graph);
730 /* -- Decide how to handle exception control flow: Is there a handler
731 for the Call node, or do we branch directly to End on an exception?
732 exc_handling: 0 There is a handler.
734 2 Exception handling not represented in Firm. -- */
736 ir_node *proj, *Mproj = NULL, *Xproj = NULL;
737 for (proj = (ir_node *)get_irn_link(call); proj; proj = (ir_node *)get_irn_link(proj)) {
738 assert(get_irn_op(proj) == op_Proj);
739 if (get_Proj_proj(proj) == pn_Call_X_except) Xproj = proj;
740 if (get_Proj_proj(proj) == pn_Call_M_except) Mproj = proj;
742 if (Mproj) { assert(Xproj); exc_handling = 0; } /* Mproj */
743 else if (Xproj) { exc_handling = 1; } /* !Mproj && Xproj */
744 else { exc_handling = 2; } /* !Mproj && !Xproj */
749 the procedure and later replaces the Start node of the called graph.
750 Post_call is the old Call node and collects the results of the called
751 graph. Both will end up being a tuple. -- */
752 post_bl = get_nodes_Block(call);
753 set_irg_current_block(current_ir_graph, post_bl);
754 /* XxMxPxP of Start + parameter of Call */
755 in[pn_Start_X_initial_exec] = new_Jmp();
756 in[pn_Start_M] = get_Call_mem(call);
757 in[pn_Start_P_frame_base] = get_irg_frame(current_ir_graph);
758 in[pn_Start_P_globals] = get_irg_globals(current_ir_graph);
759 in[pn_Start_T_args] = new_Tuple(get_Call_n_params(call), get_Call_param_arr(call));
760 /* in[pn_Start_P_value_arg_base] = ??? */
761 pre_call = new_Tuple(5, in);
765 The new block gets the ins of the old block, pre_call and all its
766 predecessors and all Phi nodes. -- */
767 part_block(pre_call);
769 /* -- Prepare state for dead node elimination -- */
770 /* Visited flags in calling irg must be >= flag in called irg.
771 Else walker and arity computation will not work. */
772 if (get_irg_visited(current_ir_graph) <= get_irg_visited(called_graph))
773 set_irg_visited(current_ir_graph, get_irg_visited(called_graph)+1);
774 if (get_irg_block_visited(current_ir_graph)< get_irg_block_visited(called_graph))
775 set_irg_block_visited(current_ir_graph, get_irg_block_visited(called_graph));
776 /* Set pre_call as new Start node in link field of the start node of
777 calling graph and pre_calls block as new block for the start block
779 Further mark these nodes so that they are not visited by the
781 set_irn_link(get_irg_start(called_graph), pre_call);
782 set_irn_visited(get_irg_start(called_graph), get_irg_visited(current_ir_graph));
783 set_irn_link(get_irg_start_block(called_graph), get_nodes_Block(pre_call));
784 set_irn_visited(get_irg_start_block(called_graph), get_irg_visited(current_ir_graph));
785 set_irn_link(get_irg_bad(called_graph), get_irg_bad(current_ir_graph));
786 set_irn_visited(get_irg_bad(called_graph), get_irg_visited(current_ir_graph));
788 /* Initialize for compaction of in arrays */
789 inc_irg_block_visited(current_ir_graph);
791 /* -- Replicate local entities of the called_graph -- */
792 /* copy the entities. */
793 called_frame = get_irg_frame_type(called_graph);
794 for (i = 0; i < get_class_n_members(called_frame); i++) {
795 entity *new_ent, *old_ent;
796 old_ent = get_class_member(called_frame, i);
797 new_ent = copy_entity_own(old_ent, get_cur_frame_type());
798 set_entity_link(old_ent, new_ent);
801 /* visited is > than that of called graph. With this trick visited will
802 remain unchanged so that an outer walker, e.g., searching the call nodes
803 to inline, calling this inline will not visit the inlined nodes. */
804 set_irg_visited(current_ir_graph, get_irg_visited(current_ir_graph)-1);
806 /* -- Performing dead node elimination inlines the graph -- */
807 /* Copies the nodes to the obstack of current_ir_graph. Updates links to new
809 /* @@@ endless loops are not copied!! -- they should be, I think... */
810 irg_walk(get_irg_end(called_graph), copy_node_inline, copy_preds,
811 get_irg_frame_type(called_graph));
813 /* Repair called_graph */
814 set_irg_visited(called_graph, get_irg_visited(current_ir_graph));
815 set_irg_block_visited(called_graph, get_irg_block_visited(current_ir_graph));
816 set_Block_block_visited(get_irg_start_block(called_graph), 0);
818 /* -- Merge the end of the inlined procedure with the call site -- */
819 /* We will turn the old Call node into a Tuple with the following
822 0: Phi of all Memories of Return statements.
823 1: Jmp from new Block that merges the control flow from all exception
824 predecessors of the old end block.
825 2: Tuple of all arguments.
826 3: Phi of Exception memories.
827 In case the old Call directly branches to End on an exception we don't
828 need the block merging all exceptions nor the Phi of the exception
832 /* -- Precompute some values -- */
833 end_bl = get_new_node(get_irg_end_block(called_graph));
834 end = get_new_node(get_irg_end(called_graph));
835 arity = get_irn_arity(end_bl); /* arity = n_exc + n_ret */
836 n_res = get_method_n_ress(get_Call_type(call));
838 res_pred = (ir_node **) malloc (n_res * sizeof (ir_node *));
839 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
841 set_irg_current_block(current_ir_graph, post_bl); /* just to make sure */
843 /* -- archive keepalives -- */
844 irn_arity = get_irn_arity(end);
845 for (i = 0; i < irn_arity; i++)
846 add_End_keepalive(get_irg_end(current_ir_graph), get_irn_n(end, i));
848 /* The new end node will die. We need not free as the in array is on the obstack:
849 copy_node only generated 'D' arrays. */
851 /* -- Replace Return nodes by Jump nodes. -- */
853 for (i = 0; i < arity; i++) {
855 ret = get_irn_n(end_bl, i);
856 if (get_irn_op(ret) == op_Return) {
857 cf_pred[n_ret] = new_r_Jmp(current_ir_graph, get_nodes_Block(ret));
861 set_irn_in(post_bl, n_ret, cf_pred);
863 /* -- Build a Tuple for all results of the method.
864 Add Phi node if there was more than one Return. -- */
865 turn_into_tuple(post_call, 4);
866 /* First the Memory-Phi */
868 for (i = 0; i < arity; i++) {
869 ret = get_irn_n(end_bl, i);
870 if (get_irn_op(ret) == op_Return) {
871 cf_pred[n_ret] = get_Return_mem(ret);
875 phi = new_Phi(n_ret, cf_pred, mode_M);
876 set_Tuple_pred(call, pn_Call_M_regular, phi);
877 /* Conserve Phi-list for further inlinings -- but might be optimized */
878 if (get_nodes_Block(phi) == post_bl) {
879 set_irn_link(phi, get_irn_link(post_bl));
880 set_irn_link(post_bl, phi);
882 /* Now the real results */
884 for (j = 0; j < n_res; j++) {
886 for (i = 0; i < arity; i++) {
887 ret = get_irn_n(end_bl, i);
888 if (get_irn_op(ret) == op_Return) {
889 cf_pred[n_ret] = get_Return_res(ret, j);
894 phi = new_Phi(n_ret, cf_pred, get_irn_mode(cf_pred[0]));
898 /* Conserve Phi-list for further inlinings -- but might be optimized */
899 if (get_nodes_Block(phi) == post_bl) {
900 set_irn_link(phi, get_irn_link(post_bl));
901 set_irn_link(post_bl, phi);
904 set_Tuple_pred(call, pn_Call_T_result, new_Tuple(n_res, res_pred));
906 set_Tuple_pred(call, pn_Call_T_result, new_Bad());
908 /* Finally the exception control flow.
909 We have two (three) possible situations:
910 First if the Call branches to an exception handler: We need to add a Phi node to
911 collect the memory containing the exception objects. Further we need
912 to add another block to get a correct representation of this Phi. To
913 this block we add a Jmp that resolves into the X output of the Call
914 when the Call is turned into a tuple.
915 Second the Call branches to End, the exception is not handled. Just
916 add all inlined exception branches to the End node.
917 Third: there is no Exception edge at all. Handle as case two. */
918 if (exc_handling == 0) {
920 for (i = 0; i < arity; i++) {
922 ret = get_irn_n(end_bl, i);
923 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
924 cf_pred[n_exc] = ret;
929 new_Block(n_exc, cf_pred); /* watch it: current_block is changed! */
930 set_Tuple_pred(call, pn_Call_X_except, new_Jmp());
931 /* The Phi for the memories with the exception objects */
933 for (i = 0; i < arity; i++) {
935 ret = skip_Proj(get_irn_n(end_bl, i));
936 if (get_irn_op(ret) == op_Call) {
937 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 3);
939 } else if (is_fragile_op(ret)) {
940 /* We rely that all cfops have the memory output at the same position. */
941 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 0);
943 } else if (get_irn_op(ret) == op_Raise) {
944 cf_pred[n_exc] = new_r_Proj(current_ir_graph, get_nodes_Block(ret), ret, mode_M, 1);
948 set_Tuple_pred(call, pn_Call_M_except, new_Phi(n_exc, cf_pred, mode_M));
950 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
951 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
954 ir_node *main_end_bl;
955 int main_end_bl_arity;
958 /* assert(exc_handling == 1 || no exceptions. ) */
960 for (i = 0; i < arity; i++) {
961 ir_node *ret = get_irn_n(end_bl, i);
963 if (is_fragile_op(skip_Proj(ret)) || (get_irn_op(skip_Proj(ret)) == op_Raise)) {
964 cf_pred[n_exc] = ret;
968 main_end_bl = get_irg_end_block(current_ir_graph);
969 main_end_bl_arity = get_irn_arity(main_end_bl);
970 end_preds = (ir_node **) malloc ((n_exc + main_end_bl_arity) * sizeof (ir_node *));
972 for (i = 0; i < main_end_bl_arity; ++i)
973 end_preds[i] = get_irn_n(main_end_bl, i);
974 for (i = 0; i < n_exc; ++i)
975 end_preds[main_end_bl_arity + i] = cf_pred[i];
976 set_irn_in(main_end_bl, n_exc + main_end_bl_arity, end_preds);
977 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
978 set_Tuple_pred(call, pn_Call_M_except, new_Bad());
984 #if 0 /* old. now better, correcter, faster implementation. */
986 /* -- If the exception control flow from the inlined Call directly
987 branched to the end block we now have the following control
988 flow predecessor pattern: ProjX -> Tuple -> Jmp. We must
989 remove the Jmp along with it's empty block and add Jmp's
990 predecessors as predecessors of this end block. No problem if
991 there is no exception, because then branches Bad to End which
993 @@@ can't we know this beforehand: by getting the Proj(1) from
994 the Call link list and checking whether it goes to Proj. */
995 /* find the problematic predecessor of the end block. */
996 end_bl = get_irg_end_block(current_ir_graph);
997 for (i = 0; i < get_Block_n_cfgpreds(end_bl); i++) {
998 cf_op = get_Block_cfgpred(end_bl, i);
999 if (get_irn_op(cf_op) == op_Proj) {
1000 cf_op = get_Proj_pred(cf_op);
1001 if ((get_irn_op(cf_op) == op_Tuple) && (cf_op == call)) {
1002 /* There are unoptimized tuples from inlineing before when no exc */
1003 assert(get_Proj_proj(get_Block_cfgpred(end_bl, i)) == pn_Call_X_except);
1004 cf_op = get_Tuple_pred(cf_op, pn_Call_X_except);
1005 assert(get_irn_op(cf_op) == op_Jmp);
1011 if (i < get_Block_n_cfgpreds(end_bl)) {
1012 bl = get_nodes_Block(cf_op);
1013 arity = get_Block_n_cfgpreds(end_bl) + get_Block_n_cfgpreds(bl) - 1;
1014 cf_pred = (ir_node **) malloc (arity * sizeof (ir_node *));
1015 for (j = 0; j < i; j++)
1016 cf_pred[j] = get_Block_cfgpred(end_bl, j);
1017 for (j = j; j < i + get_Block_n_cfgpreds(bl); j++)
1018 cf_pred[j] = get_Block_cfgpred(bl, j-i);
1019 for (j = j; j < arity; j++)
1020 cf_pred[j] = get_Block_cfgpred(end_bl, j-get_Block_n_cfgpreds(bl) +1);
1021 set_irn_in(end_bl, arity, cf_pred);
1023 /* Remove the exception pred from post-call Tuple. */
1024 set_Tuple_pred(call, pn_Call_X_except, new_Bad());
1029 /* -- Turn cse back on. -- */
1030 set_optimize(rem_opt);
1035 /********************************************************************/
1036 /* Apply inlineing to small methods. */
1037 /********************************************************************/
1039 /* It makes no sense to inline too many calls in one procedure. Anyways,
1040 I didn't get a version with NEW_ARR_F to run. */
1041 #define MAX_INLINE 1024
1044 * environment for inlining small irgs
1046 typedef struct _inline_env_t {
1048 ir_node *calls[MAX_INLINE];
1052 * Returns the irg called from a Call node. If the irg is not
1053 * known, NULL is returned.
1055 static ir_graph *get_call_called_irg(ir_node *call) {
1058 ir_graph *called_irg = NULL;
1060 assert(get_irn_op(call) == op_Call);
1062 addr = get_Call_ptr(call);
1063 if (get_irn_op(addr) == op_Const) {
1064 /* Check whether the constant is the pointer to a compiled entity. */
1065 tv = get_Const_tarval(addr);
1066 if (get_tarval_entity(tv))
1067 called_irg = get_entity_irg(get_tarval_entity(tv));
1068 } else if (get_irn_op(addr) == op_SymConst && get_SymConst_kind(addr) == symconst_addr_ent) {
1069 called_irg = get_entity_irg(get_SymConst_entity(addr));
1074 static void collect_calls(ir_node *call, void *env) {
1075 inline_env_t *ienv = env;
1078 ir_graph *called_irg;
1080 if (get_irn_op(call) != op_Call) return;
1082 addr = get_Call_ptr(call);
1083 if (get_irn_op(addr) == op_Const) {
1084 /* Check whether the constant is the pointer to a compiled entity. */
1085 tv = get_Const_tarval(addr);
1086 if (get_tarval_entity(tv)) {
1087 called_irg = get_entity_irg(get_tarval_entity(tv));
1088 if (called_irg && ienv->pos < MAX_INLINE) {
1089 /* The Call node calls a locally defined method. Remember to inline. */
1090 ienv->calls[ienv->pos++] = call;
1097 * Inlines all small methods at call sites where the called address comes
1098 * from a Const node that references the entity representing the called
1100 * The size argument is a rough measure for the code size of the method:
1101 * Methods where the obstack containing the firm graph is smaller than
1104 void inline_small_irgs(ir_graph *irg, int size) {
1106 ir_graph *rem = current_ir_graph;
1109 if (!(get_opt_optimize() && get_opt_inline())) return;
1111 current_ir_graph = irg;
1112 /* Handle graph state */
1113 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1114 free_callee_info(current_ir_graph);
1116 /* Find Call nodes to inline.
1117 (We can not inline during a walk of the graph, as inlineing the same
1118 method several times changes the visited flag of the walked graph:
1119 after the first inlineing visited of the callee equals visited of
1120 the caller. With the next inlineing both are increased.) */
1122 irg_walk(get_irg_end(irg), NULL, collect_calls, &env);
1124 if ((env.pos > 0) && (env.pos < MAX_INLINE)) {
1125 /* There are calls to inline */
1126 collect_phiprojs(irg);
1127 for (i = 0; i < env.pos; i++) {
1130 tv = get_Const_tarval(get_Call_ptr(env.calls[i]));
1131 callee = get_entity_irg(get_tarval_entity(tv));
1132 if (((_obstack_memory_used(callee->obst) - obstack_room(callee->obst)) < size) ||
1133 (get_irg_inline_property(callee) == irg_inline_forced)) {
1134 inline_method(env.calls[i], callee);
1139 current_ir_graph = rem;
1143 * Environment for inlining irgs.
1146 int n_nodes; /**< Nodes in graph except Id, Tuple, Proj, Start, End */
1147 int n_nodes_orig; /**< for statistics */
1148 eset *call_nodes; /**< All call nodes in this graph */
1150 int n_call_nodes_orig; /**< for statistics */
1151 int n_callers; /**< Number of known graphs that call this graphs. */
1152 int n_callers_orig; /**< for statistics */
1155 static inline_irg_env *new_inline_irg_env(void) {
1156 inline_irg_env *env = malloc(sizeof(inline_irg_env));
1157 env->n_nodes = -2; /* uncount Start, End */
1158 env->n_nodes_orig = -2; /* uncount Start, End */
1159 env->call_nodes = eset_create();
1160 env->n_call_nodes = 0;
1161 env->n_call_nodes_orig = 0;
1163 env->n_callers_orig = 0;
1167 static void free_inline_irg_env(inline_irg_env *env) {
1168 eset_destroy(env->call_nodes);
1172 static void collect_calls2(ir_node *call, void *env) {
1173 inline_irg_env *x = (inline_irg_env *)env;
1174 ir_op *op = get_irn_op(call);
1177 /* count nodes in irg */
1178 if (op != op_Proj && op != op_Tuple && op != op_Sync) {
1183 if (op != op_Call) return;
1185 /* collect all call nodes */
1186 eset_insert(x->call_nodes, (void *)call);
1188 x->n_call_nodes_orig++;
1190 /* count all static callers */
1191 callee = get_call_called_irg(call);
1193 ((inline_irg_env *)get_irg_link(callee))->n_callers++;
1194 ((inline_irg_env *)get_irg_link(callee))->n_callers_orig++;
1198 INLINE static int is_leave(ir_graph *irg) {
1199 return (((inline_irg_env *)get_irg_link(irg))->n_call_nodes == 0);
1202 INLINE static int is_smaller(ir_graph *callee, int size) {
1203 return (((inline_irg_env *)get_irg_link(callee))->n_nodes < size);
1208 * Inlines small leave methods at call sites where the called address comes
1209 * from a Const node that references the entity representing the called
1211 * The size argument is a rough measure for the code size of the method:
1212 * Methods where the obstack containing the firm graph is smaller than
1215 void inline_leave_functions(int maxsize, int leavesize, int size) {
1216 inline_irg_env *env;
1217 int i, n_irgs = get_irp_n_irgs();
1218 ir_graph *rem = current_ir_graph;
1221 if (!(get_opt_optimize() && get_opt_inline())) return;
1223 /* extend all irgs by a temporary data structure for inlineing. */
1224 for (i = 0; i < n_irgs; ++i)
1225 set_irg_link(get_irp_irg(i), new_inline_irg_env());
1227 /* Precompute information in temporary data structure. */
1228 for (i = 0; i < n_irgs; ++i) {
1229 current_ir_graph = get_irp_irg(i);
1230 assert(get_irg_phase_state(current_ir_graph) != phase_building);
1231 free_callee_info(current_ir_graph);
1233 irg_walk(get_irg_end(current_ir_graph), NULL, collect_calls2,
1234 get_irg_link(current_ir_graph));
1238 Inline leaves recursively -- we might construct new leaves. */
1239 /* int itercnt = 1; */
1240 while (did_inline) {
1241 /* printf("iteration %d\n", itercnt++); */
1243 for (i = 0; i < n_irgs; ++i) {
1246 int phiproj_computed = 0;
1248 current_ir_graph = get_irp_irg(i);
1249 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1251 /* we can not walk and change a set, nor remove from it.
1253 walkset = env->call_nodes;
1254 env->call_nodes = eset_create();
1255 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1256 inline_irg_env *callee_env;
1257 ir_graph *callee = get_call_called_irg(call);
1259 if (env->n_nodes > maxsize) break;
1261 ((is_leave(callee) && is_smaller(callee, leavesize)) ||
1262 (get_irg_inline_property(callee) == irg_inline_forced))) {
1263 if (!phiproj_computed) {
1264 phiproj_computed = 1;
1265 collect_phiprojs(current_ir_graph);
1267 callee_env = (inline_irg_env *)get_irg_link(callee);
1268 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1269 /* get_entity_name(get_irg_entity(callee))); */
1270 if (inline_method(call, callee)) {
1272 env->n_call_nodes--;
1273 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1274 env->n_call_nodes += callee_env->n_call_nodes;
1275 env->n_nodes += callee_env->n_nodes;
1276 callee_env->n_callers--;
1279 eset_insert(env->call_nodes, call);
1282 eset_destroy(walkset);
1286 /* printf("Non leaves\n"); */
1287 /* inline other small functions. */
1288 for (i = 0; i < n_irgs; ++i) {
1291 int phiproj_computed = 0;
1293 current_ir_graph = get_irp_irg(i);
1294 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1296 /* we can not walk and change a set, nor remove from it.
1298 walkset = env->call_nodes;
1299 env->call_nodes = eset_create();
1300 for (call = eset_first(walkset); call; call = eset_next(walkset)) {
1301 inline_irg_env *callee_env;
1302 ir_graph *callee = get_call_called_irg(call);
1304 if (env->n_nodes > maxsize) break;
1305 if (callee && is_smaller(callee, size)) {
1306 if (!phiproj_computed) {
1307 phiproj_computed = 1;
1308 collect_phiprojs(current_ir_graph);
1310 callee_env = (inline_irg_env *)get_irg_link(callee);
1311 /* printf(" %s: Inlineing %s.\n", get_entity_name(get_irg_entity(current_ir_graph)), */
1312 /* get_entity_name(get_irg_entity(callee))); */
1313 if (inline_method(call, callee)) {
1315 env->n_call_nodes--;
1316 eset_insert_all(env->call_nodes, callee_env->call_nodes);
1317 env->n_call_nodes += callee_env->n_call_nodes;
1318 env->n_nodes += callee_env->n_nodes;
1319 callee_env->n_callers--;
1322 eset_insert(env->call_nodes, call);
1325 eset_destroy(walkset);
1328 for (i = 0; i < n_irgs; ++i) {
1329 current_ir_graph = get_irp_irg(i);
1331 env = (inline_irg_env *)get_irg_link(current_ir_graph);
1332 if ((env->n_call_nodes_orig != env->n_call_nodes) ||
1333 (env->n_callers_orig != env->n_callers))
1334 printf("Nodes:%3d ->%3d, calls:%3d ->%3d, callers:%3d ->%3d, -- %s\n",
1335 env->n_nodes_orig, env->n_nodes, env->n_call_nodes_orig, env->n_call_nodes,
1336 env->n_callers_orig, env->n_callers,
1337 get_entity_name(get_irg_entity(current_ir_graph)));
1339 free_inline_irg_env((inline_irg_env *)get_irg_link(current_ir_graph));
1342 current_ir_graph = rem;
1345 /*******************************************************************/
1346 /* Code Placement. Pins all floating nodes to a block where they */
1347 /* will be executed only if needed. */
1348 /*******************************************************************/
1351 * Find the earliest correct block for N. --- Place N into the
1352 * same Block as its dominance-deepest Input.
1355 place_floats_early(ir_node *n, pdeq *worklist)
1357 int i, start, irn_arity;
1359 /* we must not run into an infinite loop */
1360 assert (irn_not_visited(n));
1361 mark_irn_visited(n);
1363 /* Place floating nodes. */
1364 if (get_op_pinned(get_irn_op(n)) == floats) {
1366 ir_node *b = new_Bad(); /* The block to place this node in */
1367 int bad_recursion = is_Bad(get_nodes_block(n));
1369 assert(get_irn_op(n) != op_Block);
1371 if ((get_irn_op(n) == op_Const) ||
1372 (get_irn_op(n) == op_SymConst) ||
1374 (get_irn_op(n) == op_Unknown)) {
1375 /* These nodes will not be placed by the loop below. */
1376 b = get_irg_start_block(current_ir_graph);
1380 /* find the block for this node. */
1381 irn_arity = get_irn_arity(n);
1382 for (i = 0; i < irn_arity; i++) {
1383 ir_node *dep = get_irn_n(n, i);
1386 if ((irn_not_visited(dep))
1387 && (get_op_pinned(get_irn_op(dep)) == floats)) {
1388 place_floats_early(dep, worklist);
1392 * A node in the Bad block must stay in the bad block,
1393 * so don't compute a new block for it.
1398 /* Because all loops contain at least one pinned node, now all
1399 our inputs are either pinned or place_early has already
1400 been finished on them. We do not have any unfinished inputs! */
1401 dep_block = get_nodes_Block(dep);
1402 if ((!is_Bad(dep_block)) &&
1403 (get_Block_dom_depth(dep_block) > depth)) {
1405 depth = get_Block_dom_depth(dep_block);
1407 /* Avoid that the node is placed in the Start block */
1408 if ((depth == 1) && (get_Block_dom_depth(get_nodes_Block(n)) > 1)) {
1409 b = get_Block_cfg_out(get_irg_start_block(current_ir_graph), 0);
1410 assert(b != get_irg_start_block(current_ir_graph));
1414 set_nodes_Block(n, b);
1417 /* Add predecessors of non floating nodes on worklist. */
1418 start = (get_irn_op(n) == op_Block) ? 0 : -1;
1419 irn_arity = get_irn_arity(n);
1420 for (i = start; i < irn_arity; i++) {
1421 ir_node *pred = get_irn_n(n, i);
1422 if (irn_not_visited(pred)) {
1423 pdeq_putr (worklist, pred);
1429 * Floating nodes form subgraphs that begin at nodes as Const, Load,
1430 * Start, Call and that end at pinned nodes as Store, Call. Place_early
1431 * places all floating nodes reachable from its argument through floating
1432 * nodes and adds all beginnings at pinned nodes to the worklist.
1434 static INLINE void place_early(pdeq* worklist) {
1436 inc_irg_visited(current_ir_graph);
1438 /* this inits the worklist */
1439 place_floats_early(get_irg_end(current_ir_graph), worklist);
1441 /* Work the content of the worklist. */
1442 while (!pdeq_empty (worklist)) {
1443 ir_node *n = pdeq_getl (worklist);
1444 if (irn_not_visited(n)) place_floats_early(n, worklist);
1447 set_irg_outs_inconsistent(current_ir_graph);
1448 current_ir_graph->pinned = pinned;
1452 /** deepest common dominance ancestor of DCA and CONSUMER of PRODUCER. */
1454 consumer_dom_dca (ir_node *dca, ir_node *consumer, ir_node *producer)
1456 ir_node *block = NULL;
1458 /* Compute the latest block into which we can place a node so that it is
1460 if (get_irn_op(consumer) == op_Phi) {
1461 /* our consumer is a Phi-node, the effective use is in all those
1462 blocks through which the Phi-node reaches producer */
1464 ir_node *phi_block = get_nodes_Block(consumer);
1465 irn_arity = get_irn_arity(consumer);
1466 for (i = 0; i < irn_arity; i++) {
1467 if (get_irn_n(consumer, i) == producer) {
1468 block = get_nodes_Block(get_Block_cfgpred(phi_block, i));
1472 assert(is_no_Block(consumer));
1473 block = get_nodes_Block(consumer);
1476 /* Compute the deepest common ancestor of block and dca. */
1478 if (!dca) return block;
1479 while (get_Block_dom_depth(block) > get_Block_dom_depth(dca))
1480 block = get_Block_idom(block);
1481 while (get_Block_dom_depth(dca) > get_Block_dom_depth(block))
1482 dca = get_Block_idom(dca);
1483 while (block != dca)
1484 { block = get_Block_idom(block); dca = get_Block_idom(dca); }
1489 static INLINE int get_irn_loop_depth(ir_node *n) {
1490 return get_loop_depth(get_irn_loop(n));
1494 * Move n to a block with less loop depth than it's current block. The
1495 * new block must be dominated by early.
1498 move_out_of_loops (ir_node *n, ir_node *early)
1500 ir_node *best, *dca;
1504 /* Find the region deepest in the dominator tree dominating
1505 dca with the least loop nesting depth, but still dominated
1506 by our early placement. */
1507 dca = get_nodes_Block(n);
1509 while (dca != early) {
1510 dca = get_Block_idom(dca);
1511 if (!dca) break; /* should we put assert(dca)? */
1512 if (get_irn_loop_depth(dca) < get_irn_loop_depth(best)) {
1516 if (best != get_nodes_Block(n)) {
1518 printf("Moving out of loop: "); DDMN(n);
1519 printf(" Outermost block: "); DDMN(early);
1520 printf(" Best block: "); DDMN(best);
1521 printf(" Innermost block: "); DDMN(get_nodes_Block(n));
1523 set_nodes_Block(n, best);
1528 * Find the latest legal block for N and place N into the
1529 * `optimal' Block between the latest and earliest legal block.
1530 * The `optimal' block is the dominance-deepest block of those
1531 * with the least loop-nesting-depth. This places N out of as many
1532 * loops as possible and then makes it as control dependant as
1536 place_floats_late(ir_node *n, pdeq *worklist)
1541 assert (irn_not_visited(n)); /* no multiple placement */
1543 /* no need to place block nodes, control nodes are already placed. */
1544 if ((get_irn_op(n) != op_Block) &&
1546 (get_irn_mode(n) != mode_X)) {
1547 /* Remember the early placement of this block to move it
1548 out of loop no further than the early placement. */
1549 early = get_nodes_Block(n);
1550 /* Assure that our users are all placed, except the Phi-nodes.
1551 --- Each data flow cycle contains at least one Phi-node. We
1552 have to break the `user has to be placed before the
1553 producer' dependence cycle and the Phi-nodes are the
1554 place to do so, because we need to base our placement on the
1555 final region of our users, which is OK with Phi-nodes, as they
1556 are pinned, and they never have to be placed after a
1557 producer of one of their inputs in the same block anyway. */
1558 for (i = 0; i < get_irn_n_outs(n); i++) {
1559 ir_node *succ = get_irn_out(n, i);
1560 if (irn_not_visited(succ) && (get_irn_op(succ) != op_Phi))
1561 place_floats_late(succ, worklist);
1564 /* We have to determine the final block of this node... except for
1566 if ((get_op_pinned(get_irn_op(n)) == floats) &&
1567 (get_irn_op(n) != op_Const) &&
1568 (get_irn_op(n) != op_SymConst)) {
1569 ir_node *dca = NULL; /* deepest common ancestor in the
1570 dominator tree of all nodes'
1571 blocks depending on us; our final
1572 placement has to dominate DCA. */
1573 for (i = 0; i < get_irn_n_outs(n); i++) {
1574 dca = consumer_dom_dca (dca, get_irn_out(n, i), n);
1576 set_nodes_Block(n, dca);
1578 move_out_of_loops (n, early);
1582 mark_irn_visited(n);
1584 /* Add predecessors of all non-floating nodes on list. (Those of floating
1585 nodes are placeded already and therefore are marked.) */
1586 for (i = 0; i < get_irn_n_outs(n); i++) {
1587 if (irn_not_visited(get_irn_out(n, i))) {
1588 pdeq_putr (worklist, get_irn_out(n, i));
1593 static INLINE void place_late(pdeq *worklist) {
1595 inc_irg_visited(current_ir_graph);
1597 /* This fills the worklist initially. */
1598 place_floats_late(get_irg_start_block(current_ir_graph), worklist);
1599 /* And now empty the worklist again... */
1600 while (!pdeq_empty (worklist)) {
1601 ir_node *n = pdeq_getl (worklist);
1602 if (irn_not_visited(n)) place_floats_late(n, worklist);
1606 void place_code(ir_graph *irg) {
1608 ir_graph *rem = current_ir_graph;
1610 current_ir_graph = irg;
1612 if (!(get_opt_optimize() && get_opt_global_cse())) return;
1614 /* Handle graph state */
1615 assert(get_irg_phase_state(irg) != phase_building);
1616 if (get_irg_dom_state(irg) != dom_consistent)
1619 if (1 || get_irg_loopinfo_state(irg) != loopinfo_consistent) {
1620 free_loop_information(irg);
1621 construct_backedges(irg);
1624 /* Place all floating nodes as early as possible. This guarantees
1625 a legal code placement. */
1626 worklist = new_pdeq();
1627 place_early(worklist);
1629 /* place_early invalidates the outs, place_late needs them. */
1631 /* Now move the nodes down in the dominator tree. This reduces the
1632 unnecessary executions of the node. */
1633 place_late(worklist);
1635 set_irg_outs_inconsistent(current_ir_graph);
1636 set_irg_loopinfo_inconsistent(current_ir_graph);
1638 current_ir_graph = rem;
1643 /********************************************************************/
1644 /* Control flow optimization. */
1645 /* Removes Bad control flow predecessors and empty blocks. A block */
1646 /* is empty if it contains only a Jmp node. */
1647 /* Blocks can only be removed if they are not needed for the */
1648 /* semantics of Phi nodes. */
1649 /********************************************************************/
1652 * Removes Tuples from Block control flow predecessors.
1653 * Optimizes blocks with equivalent_node().
1654 * Replaces n by Bad if n is unreachable control flow.
1656 static void merge_blocks(ir_node *n, void *env) {
1658 set_irn_link(n, NULL);
1660 if (get_irn_op(n) == op_Block) {
1662 for (i = 0; i < get_Block_n_cfgpreds(n); i++)
1663 /* GL @@@ : is this possible? if (get_opt_normalize()) -- added, all tests go through.
1664 A different order of optimizations might cause problems. */
1665 if (get_opt_normalize())
1666 set_Block_cfgpred(n, i, skip_Tuple(get_Block_cfgpred(n, i)));
1667 } else if (get_opt_optimize() && (get_irn_mode(n) == mode_X)) {
1668 /* We will soon visit a block. Optimize it before visiting! */
1669 ir_node *b = get_nodes_Block(n);
1670 ir_node *new_node = equivalent_node(b);
1671 while (irn_not_visited(b) && (!is_Bad(new_node)) && (new_node != b)) {
1672 /* We would have to run gigo if new is bad, so we
1673 promote it directly below. */
1674 assert(((b == new_node) ||
1675 get_opt_control_flow_straightening() ||
1676 get_opt_control_flow_weak_simplification()) &&
1677 ("strange flag setting"));
1678 exchange (b, new_node);
1680 new_node = equivalent_node(b);
1682 if (is_Bad(new_node) && get_opt_normalize()) exchange(n, new_Bad());
1687 * Collects all Phi nodes in link list of Block.
1688 * Marks all blocks "block_visited" if they contain a node other
1691 * Colelct Conds and its Projs in the cond_list
1693 static void collect_nodes(ir_node *n, void *env) {
1694 ir_node **cond_list = env;
1696 if (is_no_Block(n)) {
1697 ir_node *b = get_nodes_Block(n);
1699 switch (get_irn_opcode(n)) {
1702 /* Collect Phi nodes to compact ins along with block's ins. */
1703 set_irn_link(n, get_irn_link(b));
1709 ir_node *value = get_Cond_selector(n);
1710 tarval *tv = computed_value(value);
1712 if (tv != tarval_bad && mode_is_int(get_tarval_mode(tv))) {
1713 set_irn_link(n, *cond_list);
1716 else { /* mark the cond, we cannot optimize it */
1717 set_irn_link(n, NULL);
1724 ir_node *cond = get_Proj_pred(n);
1726 if (get_irn_op(cond) == op_Cond) {
1727 void *link = get_irn_link(cond);
1730 /* the cond node can be optimized, collect Proj */
1731 set_irn_link(n, link);
1732 set_irn_link(cond, n);
1742 if ((get_irn_op(n) != op_Jmp) && !is_Bad(b)) { /* Check for non empty block. */
1743 mark_Block_block_visited(b);
1748 /** Returns true if pred is predecessor of block. */
1749 static int is_pred_of(ir_node *pred, ir_node *b) {
1751 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1752 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1753 if (b_pred == pred) return 1;
1758 static int test_whether_dispensable(ir_node *b, int pos) {
1759 int i, j, n_preds = 1;
1760 int dispensable = 1;
1761 ir_node *cfop = get_Block_cfgpred(b, pos);
1762 ir_node *pred = get_nodes_Block(cfop);
1764 if (get_Block_block_visited(pred) + 1
1765 < get_irg_block_visited(current_ir_graph)) {
1766 if (!get_opt_optimize() || !get_opt_control_flow_strong_simplification()) {
1767 /* Mark block so that is will not be removed. */
1768 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1771 /* Seems to be empty. */
1772 if (!get_irn_link(b)) {
1773 /* There are no Phi nodes ==> dispensable. */
1774 n_preds = get_Block_n_cfgpreds(pred);
1776 /* b's pred blocks and pred's pred blocks must be pairwise disjunct.
1777 Work preds < pos as if they were already removed. */
1778 for (i = 0; i < pos; i++) {
1779 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1780 if (get_Block_block_visited(b_pred) + 1
1781 < get_irg_block_visited(current_ir_graph)) {
1782 for (j = 0; j < get_Block_n_cfgpreds(b_pred); j++) {
1783 ir_node *b_pred_pred = get_nodes_Block(get_Block_cfgpred(b_pred, j));
1784 if (is_pred_of(b_pred_pred, pred)) dispensable = 0;
1787 if (is_pred_of(b_pred, pred)) dispensable = 0;
1790 for (i = pos +1; i < get_Block_n_cfgpreds(b); i++) {
1791 ir_node *b_pred = get_nodes_Block(get_Block_cfgpred(b, i));
1792 if (is_pred_of(b_pred, pred)) dispensable = 0;
1795 set_Block_block_visited(pred, get_irg_block_visited(current_ir_graph)-1);
1798 n_preds = get_Block_n_cfgpreds(pred);
1806 static void optimize_blocks(ir_node *b, void *env) {
1807 int i, j, k, max_preds, n_preds;
1808 ir_node *pred, *phi;
1811 /* Count the number of predecessor if this block is merged with pred blocks
1814 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1815 max_preds += test_whether_dispensable(b, i);
1817 in = (ir_node **) malloc(max_preds * sizeof(ir_node *));
1820 printf(" working on "); DDMN(b);
1821 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1822 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1823 if (is_Bad(get_Block_cfgpred(b, i))) {
1824 printf(" removing Bad %i\n ", i);
1825 } else if (get_Block_block_visited(pred) +1
1826 < get_irg_block_visited(current_ir_graph)) {
1827 printf(" removing pred %i ", i); DDMN(pred);
1828 } else { printf(" Nothing to do for "); DDMN(pred); }
1830 * end Debug output -*/
1832 /*- Fix the Phi nodes -*/
1833 phi = get_irn_link(b);
1835 assert(get_irn_op(phi) == op_Phi);
1836 /* Find the new predecessors for the Phi */
1838 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1839 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1840 if (is_Bad(get_Block_cfgpred(b, i))) {
1842 } else if (get_Block_block_visited(pred) +1
1843 < get_irg_block_visited(current_ir_graph)) {
1844 /* It's an empty block and not yet visited. */
1845 ir_node *phi_pred = get_Phi_pred(phi, i);
1846 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1847 if (get_nodes_Block(phi_pred) == pred) {
1848 assert(get_irn_op(phi_pred) == op_Phi); /* Block is empty!! */
1849 in[n_preds] = get_Phi_pred(phi_pred, j);
1851 in[n_preds] = phi_pred;
1855 /* The Phi_pred node is replaced now if it is a Phi.
1856 In Schleifen kann offenbar der entfernte Phi Knoten legal verwendet werden.
1857 Daher muss der Phiknoten durch den neuen ersetzt werden.
1858 Weiter muss der alte Phiknoten entfernt werden (durch ersetzen oder
1859 durch einen Bad) damit er aus den keep_alive verschwinden kann.
1860 Man sollte also, falls keine Schleife vorliegt, exchange mit new_Bad
1862 if (get_nodes_Block(phi_pred) == pred) {
1863 /* remove the Phi as it might be kept alive. Further there
1864 might be other users. */
1865 exchange(phi_pred, phi); /* geht, ist aber doch semantisch falsch! Warum?? */
1868 in[n_preds] = get_Phi_pred(phi, i);
1873 set_irn_in(phi, n_preds, in);
1875 phi = get_irn_link(phi);
1878 /*- This happens only if merge between loop backedge and single loop entry. -*/
1879 for (k = 0; k < get_Block_n_cfgpreds(b); k++) {
1880 pred = get_nodes_Block(get_Block_cfgpred(b, k));
1881 if (get_Block_block_visited(pred)+1 < get_irg_block_visited(current_ir_graph)) {
1882 phi = get_irn_link(pred);
1884 if (get_irn_op(phi) == op_Phi) {
1885 set_nodes_Block(phi, b);
1888 for (i = 0; i < k; i++) {
1889 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1890 if (is_Bad(get_Block_cfgpred(b, i))) {
1892 } else if (get_Block_block_visited(pred) +1
1893 < get_irg_block_visited(current_ir_graph)) {
1894 /* It's an empty block and not yet visited. */
1895 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1896 /* @@@ Hier brauche ich Schleifeninformation!!! Kontrollflusskante
1897 muss Rueckwaertskante sein! (An allen vier in[n_preds] = phi
1898 Anweisungen.) Trotzdem tuts bisher!! */
1907 for (i = 0; i < get_Phi_n_preds(phi); i++) {
1908 in[n_preds] = get_Phi_pred(phi, i);
1911 for (i = k+1; i < get_Block_n_cfgpreds(b); i++) {
1912 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1913 if (is_Bad(get_Block_cfgpred(b, i))) {
1915 } else if (get_Block_block_visited(pred) +1
1916 < get_irg_block_visited(current_ir_graph)) {
1917 /* It's an empty block and not yet visited. */
1918 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1927 set_irn_in(phi, n_preds, in);
1929 phi = get_irn_link(phi);
1934 /*- Fix the block -*/
1936 for (i = 0; i < get_Block_n_cfgpreds(b); i++) {
1937 pred = get_nodes_Block(get_Block_cfgpred(b, i));
1938 if (is_Bad(get_Block_cfgpred(b, i))) {
1940 } else if (get_Block_block_visited(pred) +1
1941 < get_irg_block_visited(current_ir_graph)) {
1942 /* It's an empty block and not yet visited. */
1943 assert(get_Block_n_cfgpreds(b) > 1);
1944 /* Else it should be optimized by equivalent_node. */
1945 for (j = 0; j < get_Block_n_cfgpreds(pred); j++) {
1946 in[n_preds] = get_Block_cfgpred(pred, j);
1949 /* Remove block as it might be kept alive. */
1950 exchange(pred, b/*new_Bad()*/);
1952 in[n_preds] = get_Block_cfgpred(b, i);
1956 set_irn_in(b, n_preds, in);
1961 * an impossible ir_node * != NULL
1963 static const char _anchor;
1964 #define ANCHOR ((ir_node *)&_anchor)
1967 * optimize constant cond with projs
1969 * If we come here, we should have 3 situations:
1971 * 1.) Only one Proj exists: This should be the default Proj, convert to jump
1972 * 2.) Two Proj's exists: One is the taken, the other the not-taken default proj
1973 * 3.) More than 2 Proj's: local_optimize() did not run or internal error do nothing
1975 static void optimize_const_conds(ir_node *cond_list)
1977 ir_node *cond, *next;
1979 for (cond = cond_list; cond != ANCHOR; cond = next) {
1980 ir_node *block, *jmp;
1981 ir_node *proj = NULL, *def_proj = NULL;
1984 for (next = get_irn_link(cond); next != ANCHOR; next = get_irn_link(next)) {
1985 if (get_irn_op(next) == op_Cond)
1988 assert(get_irn_op(next) == op_Proj && "something neither Cond not Proj in cond-list");
1989 assert(get_Proj_pred(next) == cond && "cond-list corrupt");
1991 if (get_Proj_proj(next) == get_Cond_defaultProj(cond))
2001 block = get_nodes_block(def_proj);
2002 jmp = new_rd_Jmp(get_irn_dbg_info(def_proj), current_ir_graph, block);
2003 exchange(def_proj, jmp);
2005 else if (num == 2) {
2006 assert(def_proj && proj);
2008 exchange(def_proj, new_Bad());
2010 block = get_nodes_block(proj);
2011 jmp = new_rd_Jmp(get_irn_dbg_info(proj), current_ir_graph, block);
2012 exchange(proj, jmp);
2017 void optimize_cf(ir_graph *irg) {
2020 ir_node *end = get_irg_end(irg);
2021 ir_graph *rem = current_ir_graph;
2022 current_ir_graph = irg;
2023 ir_node *cond_list = ANCHOR;
2025 /* Handle graph state */
2026 assert(get_irg_phase_state(irg) != phase_building);
2027 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
2028 set_irg_outs_inconsistent(current_ir_graph);
2029 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
2030 set_irg_dom_inconsistent(current_ir_graph);
2032 /* Use block visited flag to mark non-empty blocks. */
2033 inc_irg_block_visited(irg);
2034 irg_walk(end, merge_blocks, collect_nodes, &cond_list);
2036 /* now, optimize switches */
2037 if (get_opt_unreachable_code())
2038 optimize_const_conds(cond_list);
2040 /* Optimize the standard code. */
2041 irg_block_walk(get_irg_end_block(irg), optimize_blocks, NULL, NULL);
2043 /* Walk all keep alives, optimize them if block, add to new in-array
2044 for end if useful. */
2045 in = NEW_ARR_F (ir_node *, 1);
2046 in[0] = get_nodes_Block(end);
2047 inc_irg_visited(current_ir_graph);
2048 for(i = 0; i < get_End_n_keepalives(end); i++) {
2049 ir_node *ka = get_End_keepalive(end, i);
2050 if (irn_not_visited(ka)) {
2051 if ((get_irn_op(ka) == op_Block) && Block_not_block_visited(ka)) {
2052 set_irg_block_visited(current_ir_graph, /* Don't walk all the way to Start. */
2053 get_irg_block_visited(current_ir_graph)-1);
2054 irg_block_walk(ka, optimize_blocks, NULL, NULL);
2055 mark_irn_visited(ka);
2056 ARR_APP1 (ir_node *, in, ka);
2057 } else if (get_irn_op(ka) == op_Phi) {
2058 mark_irn_visited(ka);
2059 ARR_APP1 (ir_node *, in, ka);
2063 /* DEL_ARR_F(end->in); GL @@@ tut nicht ! */
2066 current_ir_graph = rem;
2071 * Called by walker of remove_critical_cf_edges().
2073 * Place an empty block to an edge between a blocks of multiple
2074 * predecessors and a block of multiple successors.
2077 * @param env Environment of walker. This field is unused and has
2080 static void walk_critical_cf_edges(ir_node *n, void *env) {
2082 ir_node *pre, *block, **in, *jmp;
2084 arity = get_irn_arity(n);
2086 /* Block has multiple predecessors */
2087 if (op_Block == get_irn_op(n) && arity > 1) {
2089 if (n == get_irg_end_block(current_ir_graph))
2090 return; /* No use to add a block here. */
2092 for (i = 0; i < arity; ++i) {
2093 pre = get_irn_n(n, i);
2094 /* Predecessor has multiple successors. Insert new flow edge */
2095 if ((NULL != pre) &&
2096 (op_Proj == get_irn_op(pre)) &&
2097 op_Raise != get_irn_op(skip_Proj(pre))) {
2099 /* set predecessor array for new block */
2100 in = NEW_ARR_D (ir_node *, current_ir_graph->obst, 1);
2101 /* set predecessor of new block */
2103 block = new_Block(1, in);
2104 /* insert new jmp node to new block */
2105 switch_block(block);
2108 /* set successor of new block */
2109 set_irn_n(n, i, jmp);
2111 } /* predecessor has multiple successors */
2112 } /* for all predecessors */
2113 } /* n is a block */
2116 void remove_critical_cf_edges(ir_graph *irg) {
2117 if (get_opt_critical_edges())
2118 irg_walk_graph(irg, NULL, walk_critical_cf_edges, NULL);