2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Global Value Numbering Partial Redundancy Elimination
23 * (VanDrunen Hosking 2004)
24 * @author Michael Beck
38 #include "irnodemap.h"
39 #include "irnodeset.h"
41 #include "iropt_dbg.h"
44 #include "irgraph_t.h"
48 /** Additional info we need for every block. */
49 typedef struct block_info {
50 ir_valueset_t *exp_gen; /**< The set of expression per block. */
51 ir_valueset_t *avail_out; /**< The Avail_out set for a block. */
52 ir_valueset_t *antic_in; /**< The Antic_in set for a block. */
53 ir_valueset_t *new_set; /**< The set of all new values for a block. */
54 ir_node *avail; /**< The get_map(avail, block) result. */
55 ir_node *block; /**< The Block of the block info. */
56 struct block_info *next; /**< Links all entries, so we can recover the sets easily. */
57 int not_found; /**< Non-zero, if avail was not found in this block. */
61 * A pair of nodes that must be exchanged.
62 * We must defer the exchange because our hash-sets cannot
63 * find an already replace node else.
65 typedef struct elim_pair {
66 ir_node *old_node; /**< The old node that will be replaced. */
67 ir_node *new_node; /**< The new node. */
68 struct elim_pair *next; /**< Links all entries in a list. */
69 int reason; /**< The reason for the replacement. */
72 /** The environment for the GVN-PRE algorithm */
73 typedef struct pre_env {
74 struct obstack *obst; /**< The obstack to allocate on. */
75 ir_node *start_block; /**< The start block of the current graph. */
76 ir_node *end_block; /**< The end block of the current graph */
77 block_info *list; /**< Links all block info entires for easier recovery. */
78 elim_pair *pairs; /**< A list of node pairs that must be eliminated. */
79 unsigned last_idx; /**< last node index of "old" nodes, all higher indexes are newly created once. */
80 char changes; /**< Non-zero, if calculation of Antic_in has changed. */
81 char first_iter; /**< non-zero for first iteration */
84 static pset *value_table;
85 static ir_nodemap_t value_map;
87 /** The debug module handle. */
88 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
90 /* ---------- Functions for Value sets ---------- */
92 /** computes dst = dst \/ src for value sets */
93 static void value_union(ir_valueset_t *dst, ir_valueset_t *src)
95 ir_valueset_iterator_t iter;
96 ir_node *value, *expr;
98 foreach_valueset(src, value, expr, iter) {
99 ir_valueset_insert(dst, value, expr);
104 /* ---------- Functions for Values ---------- */
107 * Add a node e representing the value v to the set.
109 * @param e a node representing an expression
110 * @param v a node representing a value
112 * @return the final value for the expression e
114 static ir_node *add(ir_node *e, ir_node *v)
117 ir_node *pred = get_Proj_pred(v);
118 ir_node *v_pred = identify_remember(value_table, pred);
120 if (v_pred != pred) {
121 /* must create a new value here */
122 v = new_r_Proj(current_ir_graph, get_nodes_block(v_pred), v_pred, get_irn_mode(v), get_Proj_proj(v));
125 v = identify_remember(value_table, v);
126 ir_nodemap_insert(&value_map, e, v);
131 * Lookup a value in a value set.
133 * @param e a node representing an expression
135 * @return a node representing the value or NULL if
136 * the given expression is not available
138 static ir_node *lookup(ir_node *e)
140 ir_node *value = ir_nodemap_get(&value_map, e);
142 return identify_remember(value_table, value);
147 * Return the block info of a block.
149 * @param block the block
151 static block_info *get_block_info(ir_node *block) {
152 return get_irn_link(block);
153 } /* get_block_info */
156 * Allocate a block info for a block.
158 * @param block the block
159 * @param env the environment
161 static void alloc_blk_info(ir_node *block, pre_env *env) {
162 block_info *info = obstack_alloc(env->obst, sizeof(*info));
164 set_irn_link(block, info);
165 info->exp_gen = ir_valueset_new(16);
166 info->avail_out = ir_valueset_new(16);
167 info->antic_in = ir_valueset_new(16);
168 info->new_set = NULL;
171 info->next = env->list;
174 } /* alloc_blk_info */
177 * Returns non-zero if a node is movable and a possible candidate for PRE.
181 static int is_nice_value(ir_node *n) {
185 n = get_Proj_pred(n);
186 if (get_irn_pinned(n) == op_pin_state_pinned)
188 mode = get_irn_mode(n);
189 if (!mode_is_data(mode)) {
190 if (! is_Div(n) && ! is_Mod(n) && ! is_DivMod(n))
192 if (! is_NoMem(get_fragile_op_mem(n)))
196 } /* is_nice_value */
202 * @param set the set to dump
203 * @param txt a text to describe the set
204 * @param block the owner block of the set
206 static void dump_value_set(ir_valueset_t *set, char *txt, ir_node *block) {
207 ir_valueset_iterator_t iter;
208 ir_node *value, *expr;
211 DB((dbg, LEVEL_2, "%s(%+F) = {\n", txt, block));
213 foreach_valueset(set, value, expr, iter) {
215 DB((dbg, LEVEL_2, "\n"));
217 DB((dbg, LEVEL_2, " %+F(%+F),", expr, value));
219 DB((dbg, LEVEL_2, " %+F,", expr));
222 DB((dbg, LEVEL_2, "\n}\n"));
223 } /* dump_value_set */
226 #define dump_value_set(set, txt, block)
227 #endif /* DEBUG_libfirm */
230 * Topological walker. Allocates block info for every block and place nodes in topological
231 * order into the nodes set.
233 static void topo_walker(ir_node *irn, void *ctx) {
240 /* the topological walker ensures that blocks are visited before anything else */
241 alloc_blk_info(irn, env);
244 /* GVN step: remember the value */
245 value = add(irn, irn);
247 /* no need to put constants into the sets: they are always redundant */
248 if (! is_nice_value(irn) || is_irn_constlike(irn))
251 /* Do not put mode_T nodes info the sets, or PhiT will be created
252 (which are not allowed in Firm). Instead, put the Proj's here only. */
253 if (get_irn_mode(irn) == mode_T)
256 /* place this node into the set of possible nodes of its block */
257 block = get_nodes_block(irn);
258 info = get_block_info(block);
260 ir_valueset_insert(info->exp_gen, value, irn);
264 * Computes Avail_out(block):
266 * Avail_in(block) = Avail_out(dom(block))
267 * Avail_out(block) = Avail_in(block) \/ Nodes(block)
270 * This function must be called in the top-down dominance order:
271 * Then, it computes Leader(Nodes(block)) instead of Nodes(block) !
273 * @param block the block
274 * @param ctx walker context
276 static void compute_avail_top_down(ir_node *block, void *ctx)
279 block_info *dom_info;
280 block_info *info = get_block_info(block);
283 /* we don't need the end block Avail */
284 if (block == env->end_block)
288 * First add all nodes from the dominator.
289 * This must be done to ensure that Antic_out contains the leader
290 * for every node. The root has no dominator.
292 if (block != env->start_block) {
293 dom_blk = get_Block_idom(block);
294 assert(is_Block(dom_blk));
296 dom_info = get_block_info(dom_blk);
299 value_union(info->avail_out, dom_info->avail_out);
301 value_union(info->avail_out, info->exp_gen);
303 dump_value_set(info->avail_out, "Avail_out", block);
304 } /* compute_avail_top_down */
307 * check if a node n is clean in block block.
310 * @param block the block
311 * @param set a value set, containing the already processed predecessors
313 static int is_clean_in_block(ir_node *n, ir_node *block, ir_valueset_t *set) {
319 if (! is_nice_value(n))
322 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
323 ir_node *pred = get_irn_n(n, i);
326 if (get_nodes_block(pred) != block)
332 if (! is_nice_value(pred))
335 value = lookup(pred);
338 if (! ir_valueset_lookup(set, value))
342 } /* is_clean_in_block */
345 * Implements phi_translate. Translate a expression above a Phi.
347 * @param node the node
348 * @param block the block in which the node is translated
349 * @param pos the input number of the destination block
350 * @param translated the valueset containing the other already translated nodes
352 * @return a node representing the translated value
354 static ir_node *phi_translate(ir_node *node, ir_node *block, int pos, ir_valueset_t *translated)
360 if (get_nodes_block(node) == block) {
361 /* a Phi inside our block */
362 return get_Phi_pred(node, pos);
364 /* already outside */
368 arity = get_irn_intra_arity(node);
370 /* check if the node has at least one Phi predecessor */
371 for (i = 0; i < arity; ++i) {
372 ir_node *pred = get_irn_intra_n(node, i);
373 ir_node *leader = lookup(pred);
376 leader = leader != NULL ? leader : pred;
377 trans = ir_valueset_lookup(translated, leader);
379 if ((trans != NULL && trans != leader) || (is_Phi(leader) && get_nodes_block(leader) == block))
383 /* no translation needed */
388 get_irn_dbg_info(node),
395 /* We need the attribute copy here, because the Hash value of a
396 node might depend on that. */
397 copy_node_attr(node, nn);
399 set_nodes_block(nn, get_nodes_block(node));
400 for (i = 0; i < arity; ++i) {
401 ir_node *pred = get_irn_intra_n(node, i);
402 ir_node *leader = lookup(pred);
405 leader = leader != NULL ? leader : pred;
406 trans = ir_valueset_lookup(translated, leader);
410 if (is_Phi(trans) && get_nodes_block(trans) == block)
411 set_irn_n(nn, i, get_Phi_pred(trans, pos));
413 set_irn_n(nn, i, trans);
415 nn = optimize_node(nn);
417 } /* phi_translate */
420 * Block-walker, computes Antic_in(block).
422 * @param block the block
423 * @param ctx the walker environment
425 static void compute_antic(ir_node *block, void *ctx) {
427 block_info *succ_info;
428 block_info *info = get_block_info(block);
429 ir_node *succ, *value, *expr;
431 ir_valueset_iterator_t iter;
433 /* no need for computations in start block */
434 if (block == env->start_block)
437 size = ir_valueset_size(info->antic_in);
439 /* the end block has no successor */
440 if (block != env->end_block) {
444 * This step puts all generated expression from the current
445 * current block into Antic_in.
446 * It is enough to do this in the first iteration only, because
447 * the set info->exp_gen is not changed anymore.
449 if (env->first_iter) {
450 foreach_valueset(info->exp_gen, value, expr, iter) {
451 ir_valueset_insert(info->antic_in, value, expr);
455 n_succ = get_Block_n_cfg_outs(block);
459 /* find blocks position in succ's block predecessors */
460 succ = get_Block_cfg_out(block, 0);
461 pos = get_Block_cfgpred_pos(succ, block);
464 succ_info = get_block_info(succ);
465 /* translate into list: we cannot insert into a set we iterate
466 * and succ might be equal to block for endless loops */
467 foreach_valueset(succ_info->antic_in, value, expr, iter) {
468 ir_node *trans = phi_translate(expr, succ, pos, info->antic_in);
470 if (is_clean_in_block(trans, block, info->antic_in))
471 ir_valueset_insert(info->antic_in, value, trans);
473 } else { /* n_succ > 1 */
475 block_info *succ0_info;
480 /* Select a successor to compute the disjoint of all Nodes
481 sets, it might be useful to select the block with the
482 smallest number of nodes. For simplicity we choose the
484 succ0 = get_Block_cfg_out(block, 0);
485 succ0_info = get_block_info(succ0);
486 foreach_valueset(succ0_info->antic_in, value, expr, iter) {
487 /* we need the disjoint */
488 for (i = 1; i < n_succ; ++i) {
489 ir_node *succ = get_Block_cfg_out(block, i);
490 block_info *succ_info = get_block_info(succ);
491 if (ir_valueset_lookup(succ_info->antic_in, value) == NULL)
495 /* we found a value that is common in all Antic_in(succ(b)),
496 put it in Antic_in(b) if the value is NOT already represented. */
497 if (is_clean_in_block(expr, block, info->antic_in))
498 ir_valueset_insert(info->antic_in, value, expr);
504 /* we do not need a clean here, because we ensure that only cleaned nodes are in exp_gen
505 * and all other sets */
507 dump_value_set(info->antic_in, "Antic_in", block);
508 if (size != ir_valueset_size(info->antic_in)) {
509 /* the Antic_in set has changed */
512 } /* compute_antic */
515 * Perform insertion of partially redundant values.
516 * For every Block node, do the following:
517 * 1. Propagate the NEW_SETS of the dominator into the current block.
518 * If the block has multiple predecessors,
519 * 2a. Iterate over the ANTIC expressions for the block to see if
520 * any of them are partially redundant.
521 * 2b. If so, insert them into the necessary predecessors to make
522 * the expression fully redundant.
523 * 2c. Insert a new Phi merging the values of the predecessors.
524 * 2d. Insert the new Phi, and the new expressions, into the
527 * @param block the block
528 * @param ctx the walker environment
530 static void insert_nodes(ir_node *block, void *ctx)
533 ir_node *value, *expr, *idom, *first_s, *worklist;
534 block_info *curr_info, *idom_info;
535 int pos, arity = get_irn_intra_arity(block);
536 int all_same, by_some, updated;
537 ir_valueset_iterator_t iter;
539 /* ensure that even the start block has a new_set */
540 curr_info = get_block_info(block);
541 if (curr_info->new_set)
542 ir_valueset_del(curr_info->new_set);
543 curr_info->new_set = ir_valueset_new(16);
545 if (block == env->start_block)
548 idom = get_Block_idom(block);
549 idom_info = get_block_info(idom);
551 /* update the new_sets */
553 dump_value_set(idom_info->new_set, "[New Set]", idom);
554 foreach_valueset(idom_info->new_set, value, expr, iter) {
555 ir_valueset_insert(curr_info->new_set, value, expr);
556 updated |= ir_valueset_replace(curr_info->avail_out, value, expr);
559 dump_value_set(curr_info->avail_out, "Updated [Avail_out]", block);
565 /* convert the set into a list. This allows the removal of
566 * elements from the set */
568 foreach_valueset(curr_info->antic_in, value, expr, iter) {
571 /* If the value was already computed in the dominator, then
572 it is totally redundant. Hence we have nothing to insert. */
573 if (ir_valueset_lookup(idom_info->avail_out, value)) {
574 // DB((dbg, LEVEL_2, "Found %+F from block %+F avail in dom %+F\n", v, block, idom));
583 /* for all predecessor blocks */
584 for (pos = 0; pos < arity; ++pos) {
585 block_info *pred_info;
586 ir_node *pred_blk = get_Block_cfgpred_block(block, pos);
587 ir_node *e_prime, *v_prime, *e_dprime;
589 /* ignore bad blocks. */
590 if (is_Bad(pred_blk))
593 e_prime = phi_translate(expr, block, pos, curr_info->avail_out);
594 v_prime = lookup(e_prime);
598 pred_info = get_block_info(pred_blk);
599 e_dprime = ir_valueset_lookup(pred_info->avail_out, v_prime);
601 if (e_dprime == NULL) {
602 pred_info->avail = e_prime;
603 pred_info->not_found = 1;
606 pred_info->avail = e_dprime;
607 pred_info->not_found = 0;
608 mode = get_irn_mode(e_dprime);
613 else if (first_s != e_dprime)
616 DB((dbg, LEVEL_2, "Found %+F from block %+F as %+F in pred %+F\n", expr, block, e_dprime, pred_blk));
620 /* If it's not the same value already existing along every predecessor, and
621 it's defined by some predecessor, it is partially redundant. */
622 if (! all_same && by_some) {
623 ir_node *phi, *l, **in;
625 DB((dbg, LEVEL_1, "Partial redundant %+F from block %+F found\n", expr, block));
627 in = XMALLOCN(ir_node*, arity);
628 /* for all predecessor blocks */
629 for (pos = 0; pos < arity; ++pos) {
630 ir_node *pred_blk = get_Block_cfgpred_block(block, pos);
631 block_info *pred_info = get_block_info(pred_blk);
633 /* ignore bad blocks. */
634 if (is_Bad(pred_blk)) {
639 /* ignore blocks that already have the expression */
640 if (pred_info->not_found) {
641 ir_node *e_prime = pred_info->avail;
643 if (!is_Phi(e_prime)) {
644 ir_node *proj_pred = NULL;
645 if (is_Proj(e_prime)) {
646 ir_node *pred = get_Proj_pred(e_prime);
647 mode = get_irn_mode(pred);
649 get_irn_dbg_info(pred),
650 current_ir_graph, pred_blk,
654 get_irn_in(pred) + 1);
655 copy_node_attr(pred, nn);
657 DB((dbg, LEVEL_1, "New node %+F in block %+F created\n", nn, pred_blk));
660 mode = get_irn_mode(e_prime);
662 get_irn_dbg_info(e_prime),
663 current_ir_graph, pred_blk,
666 get_irn_arity(e_prime),
667 get_irn_in(e_prime) + 1);
668 copy_node_attr(e_prime, nn);
669 if (proj_pred != NULL) {
670 set_Proj_pred(nn, proj_pred);
673 DB((dbg, LEVEL_1, "New node %+F in block %+F created\n", nn, pred_blk));
676 l = add(expr, value);
678 ir_valueset_insert(pred_info->avail_out, add(nn, l), nn);
679 pred_info->avail = nn;
682 in[pos] = pred_info->avail;
684 phi = new_r_Phi(current_ir_graph, block, arity, in, mode);
687 l = add(expr, value);
690 ir_valueset_replace(curr_info->avail_out, value, phi);
691 ir_valueset_insert(curr_info->new_set, value, phi);
693 DB((dbg, LEVEL_1, "New %+F for redundant %+F created\n", phi, expr));
694 ir_valueset_remove_iterator(curr_info->antic_in, &iter);
697 } /* node_set_foreach */
701 * Walker, change nodes by its value if different.
703 * We cannot do the changes right here, as this would change
704 * the hash values of the nodes in the avail_out set!
706 * @param irn the node
707 * @param ctx the walker environment
709 static void eliminate(ir_node *irn, void *ctx) {
712 if (is_no_Block(irn)) {
713 ir_node *block = get_nodes_block(irn);
714 block_info *bl = get_block_info(block);
715 ir_node *value = lookup(irn);
718 ir_node *expr = ir_valueset_lookup(bl->avail_out, value);
720 if (expr != NULL && expr != irn) {
721 elim_pair *p = obstack_alloc(env->obst, sizeof(*p));
725 p->next = env->pairs;
726 p->reason = get_irn_idx(expr) >= env->last_idx ? FS_OPT_GVN_PARTLY : FS_OPT_GVN_FULLY;
734 * Do all the recorded changes and optimize
735 * newly created Phi's.
737 * @param pairs list of elimination pairs
739 static void eliminate_nodes(elim_pair *pairs) {
742 for (p = pairs; p != NULL; p = p->next) {
743 /* might be already changed */
744 p->new_node = skip_Id(p->new_node);
746 DB((dbg, LEVEL_2, "Replacing %+F by %+F\n", p->old_node, p->new_node));
748 * PRE tends to create Phi(self, self, ... , x, self, self, ...)
749 * which we can optimize here
751 if (is_Phi(p->new_node)) {
755 for (i = get_irn_intra_arity(p->new_node) - 1; i >= 0; --i) {
756 ir_node *pred = get_irn_n(p->new_node, i);
758 if (pred != p->old_node) {
767 exchange(p->new_node, res);
771 DBG_OPT_GVN_PRE(p->old_node, p->new_node, p->reason);
772 exchange(p->old_node, p->new_node);
774 } /* eliminate_nodes */
777 * Argh: Endless loops cause problems, because the
778 * insert algorithm did not terminate. We get translated nodes that
779 * references the origin. These nodes are translated again and again...
781 * The current fix is to use post-dominance. This simple ignores
782 * endless loops, ie we cannot optimize them.
784 void do_gvn_pre(ir_graph *irg)
788 optimization_state_t state;
790 unsigned antic_iter, insert_iter;
791 ir_node *value, *expr;
793 /* register a debug mask */
794 FIRM_DBG_REGISTER(dbg, "firm.opt.gvn_pre");
795 firm_dbg_set_mask(dbg, 3);
797 /* edges will crash if enabled due to our allocate on other obstack trick */
798 edges_deactivate(irg);
800 value_table = new_identities();
801 ir_nodemap_init(&value_map);
806 a_env.start_block = get_irg_start_block(irg);
807 a_env.end_block = get_irg_end_block(irg);
810 /* Move Proj's into the same block as their args,
811 else we would assign the result to wrong blocks */
812 normalize_proj_nodes(irg);
814 /* critical edges MUST be removed */
815 remove_critical_cf_edges(irg);
817 /* we need dominator for Antic_out calculation */
819 assure_postdoms(irg);
820 /* we get all nodes of a block by following outs */
821 assure_irg_outs(irg);
824 * Switch on GCSE. We need it to correctly compute
825 * the value of a node, which is independent from
828 save_optimization_state(&state);
829 set_opt_global_cse(1);
831 DB((dbg, LEVEL_1, "Doing GVN-PRE for %+F\n", irg));
833 /* allocate block info for all blocks */
834 irg_walk_blkwise_dom_top_down(irg, NULL, topo_walker, &a_env);
836 /* clean the exp_gen set. Doing this here saves the cleanup in the iteration. */
837 for (bl_info = a_env.list; bl_info != NULL; bl_info = bl_info->next) {
838 ir_valueset_iterator_t iter;
840 foreach_valueset(bl_info->exp_gen, value, expr, iter) {
841 if (!is_clean_in_block(expr, bl_info->block, bl_info->exp_gen))
842 ir_valueset_remove_iterator(bl_info->exp_gen, &iter);
845 /* compute the available value sets for all blocks */
846 dom_tree_walk_irg(irg, compute_avail_top_down, NULL, &a_env);
848 /* compute the anticipated value sets for all blocks */
850 a_env.first_iter = 1;
852 /* we use the visited flag to mark non-clean nodes */
853 inc_irg_visited(irg);
855 DB((dbg, LEVEL_1, "Antic_in Iteration %d starts ...\n", ++antic_iter));
857 postdom_tree_walk_irg(irg, compute_antic, NULL, &a_env);
858 a_env.first_iter = 0;
859 DB((dbg, LEVEL_1, "------------------------\n"));
860 } while (a_env.changes != 0);
862 /* compute redundant expressions */
864 a_env.last_idx = get_irg_last_idx(irg);
866 DB((dbg, LEVEL_1, "Insert Iteration %d starts ...\n", ++insert_iter));
868 dom_tree_walk_irg(irg, insert_nodes, NULL, &a_env);
869 DB((dbg, LEVEL_1, "------------------------\n"));
870 } while (a_env.changes != 0);
872 /* last step: eliminate nodes */
873 irg_walk_graph(irg, NULL, eliminate, &a_env);
874 eliminate_nodes(a_env.pairs);
876 /* clean up: delete all sets */
877 for (bl_info = a_env.list; bl_info != NULL; bl_info = bl_info->next) {
878 ir_valueset_del(bl_info->exp_gen);
879 ir_valueset_del(bl_info->avail_out);
880 ir_valueset_del(bl_info->antic_in);
881 if (bl_info->new_set)
882 ir_valueset_del(bl_info->new_set);
884 del_identities(value_table);
885 ir_nodemap_destroy(&value_map);
886 obstack_free(&obst, NULL);
889 /* pin the graph again: This is needed due to the use of set_opt_global_cse(1) */
890 set_irg_pinned(irg, op_pin_state_pinned);
891 restore_optimization_state(&state);
894 set_irg_outs_inconsistent(irg);
895 set_irg_loopinfo_inconsistent(irg);