2 * Copyright (C) 1995-2010 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 Path-Sensitive Jump Threading
24 * @author Christoph Mallon, Matthias Braun
29 #include "iroptimize.h"
41 #include "iredges_t.h"
45 #include "iroptimize.h"
46 #include "iropt_dbg.h"
49 #include "opt_manage.h"
53 DEBUG_ONLY(static firm_dbg_module_t *dbg);
56 * Add the new predecessor x to node node, which is either a Block or a Phi
58 static void add_pred(ir_node* node, ir_node* x)
64 assert(is_Block(node));
66 n = get_irn_arity(node);
67 NEW_ARR_A(ir_node*, ins, n + 1);
68 for (i = 0; i < n; i++)
69 ins[i] = get_irn_n(node, i);
71 set_irn_in(node, n + 1, ins);
74 static ir_node *ssa_second_def;
75 static ir_node *ssa_second_def_block;
77 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode,
87 /* In case of a bad input to a block we need to return the bad value */
89 ir_graph *irg = get_irn_irg(block);
90 return new_r_Bad(irg, mode);
93 /* the other defs can't be marked for cases where a user of the original
94 * value is in the same block as the alternative definition.
95 * In this case we mustn't use the alternative definition.
96 * So we keep a flag that indicated whether we walked at least 1 block
97 * away and may use the alternative definition */
98 if (block == ssa_second_def_block && !first) {
99 return ssa_second_def;
102 /* already processed this block? */
103 if (irn_visited(block)) {
104 ir_node *value = (ir_node*) get_irn_link(block);
108 irg = get_irn_irg(block);
109 assert(block != get_irg_start_block(irg));
111 /* a Block with only 1 predecessor needs no Phi */
112 n_cfgpreds = get_Block_n_cfgpreds(block);
113 if (n_cfgpreds == 1) {
114 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
115 ir_node *value = search_def_and_create_phis(pred_block, mode, 0);
117 set_irn_link(block, value);
118 mark_irn_visited(block);
122 /* create a new Phi */
123 NEW_ARR_A(ir_node*, in, n_cfgpreds);
124 dummy = new_r_Dummy(irg, mode);
125 for (i = 0; i < n_cfgpreds; ++i)
128 phi = new_r_Phi(block, n_cfgpreds, in, mode);
129 set_irn_link(block, phi);
130 mark_irn_visited(block);
132 /* set Phi predecessors */
133 for (i = 0; i < n_cfgpreds; ++i) {
134 ir_node *pred_block = get_Block_cfgpred_block(block, i);
135 ir_node *pred_val = search_def_and_create_phis(pred_block, mode, 0);
137 set_irn_n(phi, i, pred_val);
144 * Given a set of values this function constructs SSA-form for the users of the
145 * first value (the users are determined through the out-edges of the value).
146 * Uses the irn_visited flags. Works without using the dominance tree.
148 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
149 ir_node *second_block, ir_node *second_val)
153 const ir_edge_t *edge;
154 const ir_edge_t *next;
156 /* no need to do anything */
157 if (orig_val == second_val)
160 irg = get_irn_irg(orig_val);
161 inc_irg_visited(irg);
163 mode = get_irn_mode(orig_val);
164 set_irn_link(orig_block, orig_val);
165 mark_irn_visited(orig_block);
167 ssa_second_def_block = second_block;
168 ssa_second_def = second_val;
170 /* Only fix the users of the first, i.e. the original node */
171 foreach_out_edge_safe(orig_val, edge, next) {
172 ir_node *user = get_edge_src_irn(edge);
173 int j = get_edge_src_pos(edge);
174 ir_node *user_block = get_nodes_block(user);
181 DB((dbg, LEVEL_3, ">>> Fixing user %+F (pred %d == %+F)\n", user, j, get_irn_n(user, j)));
184 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
185 newval = search_def_and_create_phis(pred_block, mode, 1);
187 newval = search_def_and_create_phis(user_block, mode, 1);
190 /* don't fix newly created Phis from the SSA construction */
191 if (newval != user) {
192 DB((dbg, LEVEL_4, ">>>> Setting input %d of %+F to %+F\n", j, user, newval));
193 set_irn_n(user, j, newval);
199 * jumpthreading produces critical edges, e.g. B-C:
206 * By splitting this critical edge more threadings might be possible.
208 static void split_critical_edge(ir_node *block, int pos)
210 ir_graph *irg = get_irn_irg(block);
215 in[0] = get_Block_cfgpred(block, pos);
216 new_block = new_r_Block(irg, 1, in);
217 new_jmp = new_r_Jmp(new_block);
218 set_Block_cfgpred(block, pos, new_jmp);
221 typedef struct jumpthreading_env_t {
223 ir_node *cmp; /**< The Compare node that might be partial evaluated */
224 ir_relation relation; /**< The Compare mode of the Compare node. */
227 ir_visited_t visited_nr;
229 ir_node *cnst_pred; /**< the block before the constant */
230 int cnst_pos; /**< the pos to the constant block (needed to
231 kill that edge later) */
232 } jumpthreading_env_t;
234 static ir_node *copy_and_fix_node(const jumpthreading_env_t *env,
235 ir_node *block, ir_node *copy_block, int j,
241 /* we can evaluate Phis right now, all other nodes get copied */
243 copy = get_Phi_pred(node, j);
244 /* we might have to evaluate a Phi-cascade */
245 if (get_irn_visited(copy) >= env->visited_nr) {
246 copy = (ir_node*)get_irn_link(copy);
249 copy = exact_copy(node);
250 set_nodes_block(copy, copy_block);
252 assert(get_irn_mode(copy) != mode_X);
254 arity = get_irn_arity(copy);
255 for (i = 0; i < arity; ++i) {
256 ir_node *pred = get_irn_n(copy, i);
259 if (get_nodes_block(pred) != block)
262 if (get_irn_visited(pred) >= env->visited_nr) {
263 new_pred = (ir_node*)get_irn_link(pred);
265 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
267 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
268 set_irn_n(copy, i, new_pred);
272 set_irn_link(node, copy);
273 set_irn_visited(node, env->visited_nr);
278 static void copy_and_fix(const jumpthreading_env_t *env, ir_node *block,
279 ir_node *copy_block, int j)
281 const ir_edge_t *edge;
283 /* Look at all nodes in the cond_block and copy them into pred */
284 foreach_out_edge(block, edge) {
285 ir_node *node = get_edge_src_irn(edge);
290 /* edge is a Keep edge. If the end block is unreachable via normal control flow,
291 * we must maintain end's reachability with Keeps.
293 keep_alive(copy_block);
296 /* ignore control flow */
297 mode = get_irn_mode(node);
298 if (mode == mode_X || is_Cond(node))
301 /* we may not copy mode_b nodes, because this could produce Phi with
302 * mode_bs which can't be handled in all backends. Instead we duplicate
303 * the node and move it to its users */
304 if (mode == mode_b) {
305 const ir_edge_t *edge, *next;
309 assert(is_Proj(node));
311 pred = get_Proj_pred(node);
312 pn = get_Proj_proj(node);
314 foreach_out_edge_safe(node, edge, next) {
316 ir_node *user = get_edge_src_irn(edge);
317 int pos = get_edge_src_pos(edge);
318 ir_node *user_block = get_nodes_block(user);
320 if (user_block == block)
323 cmp_copy = exact_copy(pred);
324 set_nodes_block(cmp_copy, user_block);
325 copy = new_r_Proj(cmp_copy, mode_b, pn);
326 set_irn_n(user, pos, copy);
332 copy = copy_and_fix_node(env, block, copy_block, j, node);
334 /* we might hit values in blocks that have already been processed by a
335 * recursive find_phi_with_const() call */
336 assert(get_irn_visited(copy) <= env->visited_nr);
337 if (get_irn_visited(copy) >= env->visited_nr) {
338 ir_node *prev_copy = (ir_node*)get_irn_link(copy);
339 if (prev_copy != NULL)
340 set_irn_link(node, prev_copy);
344 /* fix data-flow (and reconstruct SSA if needed) */
345 foreach_out_edge(block, edge) {
346 ir_node *node = get_edge_src_irn(edge);
350 mode = get_irn_mode(node);
351 if (mode == mode_X || is_Cond(node))
358 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
360 copy_node = (ir_node*)get_irn_link(node);
361 construct_ssa(block, node, copy_block, copy_node);
366 * returns whether the cmp evaluates to true or false, or can't be evaluated!
367 * 1: true, 0: false, -1: can't evaluate
369 * @param relation the compare mode of the Compare
370 * @param tv_left the left tarval
371 * @param tv_right the right tarval
373 static int eval_cmp_tv(ir_relation relation, ir_tarval *tv_left,
376 ir_relation cmp_result = tarval_cmp(tv_left, tv_right);
378 /* does the compare evaluate to true? */
379 if (cmp_result == ir_relation_false)
381 if ((cmp_result & relation) != 0)
388 /* Matze: disabled, check first if the compare still is correct */
391 * returns whether the cmp evaluates to true or false according to vrp
392 * information , or can't be evaluated!
393 * 1: true, 0: false, -1: can't evaluate
395 * @param relation the compare mode of the Compare
396 * @param left the left node
397 * @param right the right node
399 static int eval_cmp_vrp(ir_relation relation, ir_node *left, ir_node *right)
401 ir_relation cmp_result = vrp_cmp(left, right);
402 /* does the compare evaluate to true? */
403 if (cmp_result == ir_relation_false)
406 if ((cmp_result & relation) != cmp_result) {
407 if ((cmp_result & relation) != 0) {
417 * returns whether the cmp evaluates to true or false, or can't be evaluated!
418 * 1: true, 0: false, -1: can't evaluate
420 * @param env the environment
421 * @param cand the candidate node, either a Const or a Confirm
423 static int eval_cmp(jumpthreading_env_t *env, ir_node *cand)
425 if (is_Const(cand)) {
426 ir_tarval *tv_cand = get_Const_tarval(cand);
427 ir_tarval *tv_cmp = get_Const_tarval(env->cnst);
429 return eval_cmp_tv(env->relation, tv_cand, tv_cmp);
430 } else { /* a Confirm */
431 ir_tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->relation);
433 if (res == tarval_bad)
435 return res == tarval_b_true;
440 * Check for Const or Confirm with Const.
442 static int is_Const_or_Confirm(const ir_node *node)
444 if (is_Confirm(node))
445 node = get_Confirm_bound(node);
446 return is_Const(node);
450 * get the tarval of a Const or Confirm with
452 static ir_tarval *get_Const_or_Confirm_tarval(const ir_node *node)
454 if (is_Confirm(node)) {
455 if (get_Confirm_bound(node))
456 node = get_Confirm_bound(node);
458 return get_Const_tarval(node);
461 static ir_node *find_const_or_confirm(jumpthreading_env_t *env, ir_node *jump,
464 ir_node *block = get_nodes_block(jump);
466 if (irn_visited_else_mark(value))
469 if (is_Const_or_Confirm(value)) {
470 if (eval_cmp(env, value) <= 0)
475 "> Found jump threading candidate %+F->%+F\n",
476 env->true_block, block
479 /* adjust true_block to point directly towards our jump */
480 add_pred(env->true_block, jump);
482 split_critical_edge(env->true_block, 0);
484 /* we need a bigger visited nr when going back */
493 /* the Phi has to be in the same Block as the Jmp */
494 if (get_nodes_block(value) != block)
497 arity = get_irn_arity(value);
498 for (i = 0; i < arity; ++i) {
500 ir_node *phi_pred = get_Phi_pred(value, i);
501 ir_node *cfgpred = get_Block_cfgpred(block, i);
503 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
504 if (copy_block == NULL)
507 /* copy duplicated nodes in copy_block and fix SSA */
508 copy_and_fix(env, block, copy_block, i);
510 if (copy_block == get_nodes_block(cfgpred)) {
511 env->cnst_pred = block;
515 /* return now as we can't process more possibilities in 1 run */
523 static ir_node *find_candidate(jumpthreading_env_t *env, ir_node *jump,
526 ir_node *block = get_nodes_block(jump);
528 if (irn_visited_else_mark(value)) {
532 if (is_Const_or_Confirm(value)) {
533 ir_tarval *tv = get_Const_or_Confirm_tarval(value);
540 "> Found jump threading candidate %+F->%+F\n",
541 env->true_block, block
544 /* adjust true_block to point directly towards our jump */
545 add_pred(env->true_block, jump);
547 split_critical_edge(env->true_block, 0);
549 /* we need a bigger visited nr when going back */
557 /* the Phi has to be in the same Block as the Jmp */
558 if (get_nodes_block(value) != block)
561 arity = get_irn_arity(value);
562 for (i = 0; i < arity; ++i) {
564 ir_node *phi_pred = get_Phi_pred(value, i);
565 ir_node *cfgpred = get_Block_cfgpred(block, i);
567 copy_block = find_candidate(env, cfgpred, phi_pred);
568 if (copy_block == NULL)
571 /* copy duplicated nodes in copy_block and fix SSA */
572 copy_and_fix(env, block, copy_block, i);
574 if (copy_block == get_nodes_block(cfgpred)) {
575 env->cnst_pred = block;
579 /* return now as we can't process more possibilities in 1 run */
584 ir_node *cmp = value;
585 ir_node *left = get_Cmp_left(cmp);
586 ir_node *right = get_Cmp_right(cmp);
587 ir_relation relation = get_Cmp_relation(cmp);
589 /* we assume that the constant is on the right side, swap left/right
591 if (is_Const(left)) {
596 relation = get_inversed_relation(relation);
599 if (!is_Const(right))
602 if (get_nodes_block(left) != block)
605 /* negate condition when we're looking for the false block */
606 if (env->tv == tarval_b_false) {
607 relation = get_negated_relation(relation);
610 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
612 env->relation = relation;
615 return find_const_or_confirm(env, jump, left);
622 * Block-walker: searches for the following construct
624 * Const or Phi with constants
634 static void thread_jumps(ir_node* block, void* data)
636 jumpthreading_env_t env;
637 int *changed = (int*)data;
642 int selector_evaluated;
643 const ir_edge_t *edge, *next;
648 /* we do not deal with Phis, so restrict this to exactly one cfgpred */
649 if (get_Block_n_cfgpreds(block) != 1)
652 projx = get_Block_cfgpred(block, 0);
655 assert(get_irn_mode(projx) == mode_X);
657 cond = get_Proj_pred(projx);
661 selector = get_Cond_selector(cond);
662 /* TODO handle switch Conds */
663 if (get_irn_mode(selector) != mode_b)
666 /* handle cases that can be immediately evaluated */
667 selector_evaluated = -1;
668 if (is_Cmp(selector)) {
669 ir_node *left = get_Cmp_left(selector);
670 ir_node *right = get_Cmp_right(selector);
671 if (is_Const(left) && is_Const(right)) {
672 ir_relation relation = get_Cmp_relation(selector);
673 ir_tarval *tv_left = get_Const_tarval(left);
674 ir_tarval *tv_right = get_Const_tarval(right);
676 selector_evaluated = eval_cmp_tv(relation, tv_left, tv_right);
679 if (selector_evaluated < 0) {
680 /* This is only the case if the predecessor nodes are not
681 * constant or the comparison could not be evaluated.
682 * Try with VRP information now.
684 selector_evaluated = eval_cmp_vrp(relation, left, right);
687 } else if (is_Const_or_Confirm(selector)) {
688 ir_tarval *tv = get_Const_or_Confirm_tarval(selector);
689 if (tv == tarval_b_true) {
690 selector_evaluated = 1;
692 assert(tv == tarval_b_false);
693 selector_evaluated = 0;
697 env.cnst_pred = NULL;
698 if (get_Proj_proj(projx) == pn_Cond_false) {
699 env.tv = tarval_b_false;
700 if (selector_evaluated >= 0)
701 selector_evaluated = !selector_evaluated;
703 env.tv = tarval_b_true;
706 if (selector_evaluated == 0) {
707 ir_graph *irg = get_irn_irg(block);
708 ir_node *bad = new_r_Bad(irg, mode_X);
709 exchange(projx, bad);
712 } else if (selector_evaluated == 1) {
713 dbg_info *dbgi = get_irn_dbg_info(selector);
714 ir_node *jmp = new_rd_Jmp(dbgi, get_nodes_block(projx));
715 DBG_OPT_JUMPTHREADING(projx, jmp);
716 exchange(projx, jmp);
721 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
722 env.true_block = block;
723 irg = get_irn_irg(block);
724 inc_irg_visited(irg);
725 env.visited_nr = get_irg_visited(irg);
727 copy_block = find_candidate(&env, projx, selector);
728 if (copy_block == NULL)
731 /* We might thread the condition block of an infinite loop,
732 * such that there is no path to End anymore. */
735 /* we have to remove the edge towards the pred as the pred now
736 * jumps into the true_block. We also have to shorten Phis
737 * in our block because of this */
738 badX = new_r_Bad(irg, mode_X);
739 cnst_pos = env.cnst_pos;
742 foreach_out_edge_safe(env.cnst_pred, edge, next) {
743 ir_node *node = get_edge_src_irn(edge);
746 ir_node *bad = new_r_Bad(irg, get_irn_mode(node));
747 set_Phi_pred(node, cnst_pos, bad);
751 set_Block_cfgpred(env.cnst_pred, cnst_pos, badX);
753 /* the graph is changed now */
757 static ir_graph_state_t do_jumpthread(ir_graph* irg)
760 ir_graph_state_t res = 0;
762 FIRM_DBG_REGISTER(dbg, "firm.opt.jumpthreading");
764 DB((dbg, LEVEL_1, "===> Performing jumpthreading on %+F\n", irg));
766 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
771 irg_block_walk_graph(irg, thread_jumps, NULL, &rerun);
775 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
778 res |= IR_GRAPH_STATE_CONSISTENT_DOMINANCE | IR_GRAPH_STATE_CONSISTENT_ENTITY_USAGE;
784 optdesc_t opt_jumpthread = {
786 IR_GRAPH_STATE_NO_UNREACHABLE_BLOCKS | IR_GRAPH_STATE_CONSISTENT_OUT_EDGES | IR_GRAPH_STATE_NO_CRITICAL_EDGES,
790 void opt_jumpthreading(ir_graph* irg)
792 perform_irg_optimization(irg, &opt_jumpthread);
795 /* Creates an ir_graph pass for opt_jumpthreading. */
796 ir_graph_pass_t *opt_jumpthreading_pass(const char *name)
798 return def_graph_pass(name ? name : "jumpthreading", opt_jumpthreading);
799 } /* opt_jumpthreading_pass */