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
28 #include "iroptimize.h"
41 #include "iredges_t.h"
45 #include "iroptimize.h"
46 #include "iropt_dbg.h"
52 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
55 * Add the new predecessor x to node node, which is either a Block or a Phi
57 static void add_pred(ir_node* node, ir_node* x)
63 assert(is_Block(node));
65 n = get_irn_arity(node);
66 NEW_ARR_A(ir_node*, ins, n + 1);
67 for (i = 0; i < n; i++)
68 ins[i] = get_irn_n(node, i);
70 set_irn_in(node, n + 1, ins);
73 static ir_node *ssa_second_def;
74 static ir_node *ssa_second_def_block;
76 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode,
86 /* In case of a bad input to a block we need to return the bad value */
88 ir_graph *irg = get_irn_irg(block);
89 return new_r_Bad(irg, mode);
92 /* the other defs can't be marked for cases where a user of the original
93 * value is in the same block as the alternative definition.
94 * In this case we mustn't use the alternative definition.
95 * So we keep a flag that indicated whether we walked at least 1 block
96 * away and may use the alternative definition */
97 if (block == ssa_second_def_block && !first) {
98 return ssa_second_def;
101 /* already processed this block? */
102 if (irn_visited(block)) {
103 ir_node *value = (ir_node*) get_irn_link(block);
107 irg = get_irn_irg(block);
108 assert(block != get_irg_start_block(irg));
110 /* a Block with only 1 predecessor needs no Phi */
111 n_cfgpreds = get_Block_n_cfgpreds(block);
112 if (n_cfgpreds == 1) {
113 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
114 ir_node *value = search_def_and_create_phis(pred_block, mode, 0);
116 set_irn_link(block, value);
117 mark_irn_visited(block);
121 /* create a new Phi */
122 NEW_ARR_A(ir_node*, in, n_cfgpreds);
123 dummy = new_r_Dummy(irg, mode);
124 for (i = 0; i < n_cfgpreds; ++i)
127 phi = new_r_Phi(block, n_cfgpreds, in, mode);
128 set_irn_link(block, phi);
129 mark_irn_visited(block);
131 /* set Phi predecessors */
132 for (i = 0; i < n_cfgpreds; ++i) {
133 ir_node *pred_block = get_Block_cfgpred_block(block, i);
134 ir_node *pred_val = search_def_and_create_phis(pred_block, mode, 0);
136 set_irn_n(phi, i, pred_val);
143 * Given a set of values this function constructs SSA-form for the users of the
144 * first value (the users are determined through the out-edges of the value).
145 * Uses the irn_visited flags. Works without using the dominance tree.
147 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
148 ir_node *second_block, ir_node *second_val)
153 /* no need to do anything */
154 if (orig_val == second_val)
157 irg = get_irn_irg(orig_val);
158 inc_irg_visited(irg);
160 mode = get_irn_mode(orig_val);
161 set_irn_link(orig_block, orig_val);
162 mark_irn_visited(orig_block);
164 ssa_second_def_block = second_block;
165 ssa_second_def = second_val;
167 /* Only fix the users of the first, i.e. the original node */
168 foreach_out_edge_safe(orig_val, edge) {
169 ir_node *user = get_edge_src_irn(edge);
170 int j = get_edge_src_pos(edge);
171 ir_node *user_block = get_nodes_block(user);
178 DB((dbg, LEVEL_3, ">>> Fixing user %+F (pred %d == %+F)\n", user, j, get_irn_n(user, j)));
181 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
182 newval = search_def_and_create_phis(pred_block, mode, 1);
184 newval = search_def_and_create_phis(user_block, mode, 1);
187 /* don't fix newly created Phis from the SSA construction */
188 if (newval != user) {
189 DB((dbg, LEVEL_4, ">>>> Setting input %d of %+F to %+F\n", j, user, newval));
190 set_irn_n(user, j, newval);
196 * jumpthreading produces critical edges, e.g. B-C:
203 * By splitting this critical edge more threadings might be possible.
205 static void split_critical_edge(ir_node *block, int pos)
207 ir_graph *irg = get_irn_irg(block);
212 in[0] = get_Block_cfgpred(block, pos);
213 new_block = new_r_Block(irg, 1, in);
214 new_jmp = new_r_Jmp(new_block);
215 set_Block_cfgpred(block, pos, new_jmp);
218 typedef struct jumpthreading_env_t {
220 ir_node *cmp; /**< The Compare node that might be partial evaluated */
221 ir_relation relation; /**< The Compare mode of the Compare node. */
224 ir_visited_t visited_nr;
226 ir_node *cnst_pred; /**< the block before the constant */
227 int cnst_pos; /**< the pos to the constant block (needed to
228 kill that edge later) */
229 } jumpthreading_env_t;
231 static ir_node *copy_and_fix_node(const jumpthreading_env_t *env,
232 ir_node *block, ir_node *copy_block, int j,
238 /* we can evaluate Phis right now, all other nodes get copied */
240 copy = get_Phi_pred(node, j);
241 /* we might have to evaluate a Phi-cascade */
242 if (get_irn_visited(copy) >= env->visited_nr) {
243 copy = (ir_node*)get_irn_link(copy);
246 copy = exact_copy(node);
247 set_nodes_block(copy, copy_block);
249 assert(get_irn_mode(copy) != mode_X);
251 arity = get_irn_arity(copy);
252 for (i = 0; i < arity; ++i) {
253 ir_node *pred = get_irn_n(copy, i);
256 if (get_nodes_block(pred) != block)
259 if (get_irn_visited(pred) >= env->visited_nr) {
260 new_pred = (ir_node*)get_irn_link(pred);
262 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
264 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
265 set_irn_n(copy, i, new_pred);
269 set_irn_link(node, copy);
270 set_irn_visited(node, env->visited_nr);
275 static void copy_and_fix(const jumpthreading_env_t *env, ir_node *block,
276 ir_node *copy_block, int j)
278 /* Look at all nodes in the cond_block and copy them into pred */
279 foreach_out_edge(block, edge) {
280 ir_node *node = get_edge_src_irn(edge);
285 /* edge is a Keep edge. If the end block is unreachable via normal
286 * control flow, we must maintain end's reachability with Keeps.
288 keep_alive(copy_block);
291 /* ignore control flow */
292 mode = get_irn_mode(node);
293 if (mode == mode_X || is_Cond(node) || is_Switch(node))
296 /* we may not copy mode_b nodes, because this could produce Phi with
297 * mode_bs which can't be handled in all backends. Instead we duplicate
298 * the node and move it to its users */
299 if (mode == mode_b) {
303 assert(is_Proj(node));
305 pred = get_Proj_pred(node);
306 pn = get_Proj_proj(node);
308 foreach_out_edge_safe(node, edge) {
310 ir_node *user = get_edge_src_irn(edge);
311 int pos = get_edge_src_pos(edge);
312 ir_node *user_block = get_nodes_block(user);
314 if (user_block == block)
317 cmp_copy = exact_copy(pred);
318 set_nodes_block(cmp_copy, user_block);
319 copy = new_r_Proj(cmp_copy, mode_b, pn);
320 set_irn_n(user, pos, copy);
326 copy = copy_and_fix_node(env, block, copy_block, j, node);
328 /* we might hit values in blocks that have already been processed by a
329 * recursive find_phi_with_const() call */
330 assert(get_irn_visited(copy) <= env->visited_nr);
331 if (get_irn_visited(copy) >= env->visited_nr) {
332 ir_node *prev_copy = (ir_node*)get_irn_link(copy);
333 if (prev_copy != NULL)
334 set_irn_link(node, prev_copy);
338 /* fix data-flow (and reconstruct SSA if needed) */
339 foreach_out_edge(block, edge) {
340 ir_node *node = get_edge_src_irn(edge);
344 mode = get_irn_mode(node);
345 if (mode == mode_X || is_Cond(node) || is_Switch(node))
352 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
354 copy_node = (ir_node*)get_irn_link(node);
355 construct_ssa(block, node, copy_block, copy_node);
358 /* make sure new nodes are kept alive if old nodes were */
359 ir_graph *irg = get_irn_irg(block);
360 ir_node *end = get_irg_end(irg);
361 for (int i = 0, arity = get_End_n_keepalives(end); i < arity; ++i) {
362 ir_node *keep = get_End_keepalive(end, i);
363 if (get_irn_visited(keep) < env->visited_nr || is_Block(keep))
365 ir_node *copy = get_irn_link(keep);
366 add_End_keepalive(end, copy);
371 * returns whether the cmp evaluates to true or false, or can't be evaluated!
372 * 1: true, 0: false, -1: can't evaluate
374 * @param relation the compare mode of the Compare
375 * @param tv_left the left tarval
376 * @param tv_right the right tarval
378 static int eval_cmp_tv(ir_relation relation, ir_tarval *tv_left,
381 ir_relation cmp_result = tarval_cmp(tv_left, tv_right);
383 /* does the compare evaluate to true? */
384 if (cmp_result == ir_relation_false)
386 if ((cmp_result & relation) != 0)
393 /* Matze: disabled, check first if the compare still is correct */
396 * returns whether the cmp evaluates to true or false according to vrp
397 * information , or can't be evaluated!
398 * 1: true, 0: false, -1: can't evaluate
400 * @param relation the compare mode of the Compare
401 * @param left the left node
402 * @param right the right node
404 static int eval_cmp_vrp(ir_relation relation, ir_node *left, ir_node *right)
406 ir_relation cmp_result = vrp_cmp(left, right);
407 /* does the compare evaluate to true? */
408 if (cmp_result == ir_relation_false)
411 if ((cmp_result & relation) != cmp_result) {
412 if ((cmp_result & relation) != 0) {
422 * returns whether the cmp evaluates to true or false, or can't be evaluated!
423 * 1: true, 0: false, -1: can't evaluate
425 * @param env the environment
426 * @param cand the candidate node, either a Const or a Confirm
428 static int eval_cmp(jumpthreading_env_t *env, ir_node *cand)
430 if (is_Const(cand)) {
431 ir_tarval *tv_cand = get_Const_tarval(cand);
432 ir_tarval *tv_cmp = get_Const_tarval(env->cnst);
434 return eval_cmp_tv(env->relation, tv_cand, tv_cmp);
435 } else { /* a Confirm */
436 ir_tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->relation);
438 if (res == tarval_bad)
440 return res == tarval_b_true;
445 * Check for Const or Confirm with Const.
447 static int is_Const_or_Confirm(const ir_node *node)
449 if (is_Confirm(node))
450 node = get_Confirm_bound(node);
451 return is_Const(node);
455 * get the tarval of a Const or Confirm with
457 static ir_tarval *get_Const_or_Confirm_tarval(const ir_node *node)
459 if (is_Confirm(node)) {
460 if (get_Confirm_bound(node))
461 node = get_Confirm_bound(node);
463 return get_Const_tarval(node);
466 static ir_node *find_const_or_confirm(jumpthreading_env_t *env, ir_node *jump,
469 ir_node *block = get_nodes_block(jump);
471 if (irn_visited_else_mark(value))
474 if (is_Const_or_Confirm(value)) {
475 if (eval_cmp(env, value) <= 0)
480 "> Found jump threading candidate %+F->%+F\n",
481 block, env->true_block
484 /* adjust true_block to point directly towards our jump */
485 add_pred(env->true_block, jump);
487 split_critical_edge(env->true_block, 0);
489 /* we need a bigger visited nr when going back */
498 /* the Phi has to be in the same Block as the Jmp */
499 if (get_nodes_block(value) != block)
502 arity = get_irn_arity(value);
503 for (i = 0; i < arity; ++i) {
505 ir_node *phi_pred = get_Phi_pred(value, i);
506 ir_node *cfgpred = get_Block_cfgpred(block, i);
508 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
509 if (copy_block == NULL)
512 /* copy duplicated nodes in copy_block and fix SSA */
513 copy_and_fix(env, block, copy_block, i);
515 if (copy_block == get_nodes_block(cfgpred)) {
516 env->cnst_pred = block;
520 /* return now as we can't process more possibilities in 1 run */
528 static ir_node *find_candidate(jumpthreading_env_t *env, ir_node *jump,
531 ir_node *block = get_nodes_block(jump);
533 if (irn_visited_else_mark(value)) {
537 if (is_Const_or_Confirm(value)) {
538 ir_tarval *tv = get_Const_or_Confirm_tarval(value);
545 "> Found jump threading candidate %+F->%+F\n",
546 block, env->true_block
549 /* adjust true_block to point directly towards our jump */
550 add_pred(env->true_block, jump);
552 split_critical_edge(env->true_block, 0);
554 /* we need a bigger visited nr when going back */
562 /* the Phi has to be in the same Block as the Jmp */
563 if (get_nodes_block(value) != block)
566 arity = get_irn_arity(value);
567 for (i = 0; i < arity; ++i) {
569 ir_node *phi_pred = get_Phi_pred(value, i);
570 ir_node *cfgpred = get_Block_cfgpred(block, i);
572 copy_block = find_candidate(env, cfgpred, phi_pred);
573 if (copy_block == NULL)
576 /* copy duplicated nodes in copy_block and fix SSA */
577 copy_and_fix(env, block, copy_block, i);
579 if (copy_block == get_nodes_block(cfgpred)) {
580 env->cnst_pred = block;
584 /* return now as we can't process more possibilities in 1 run */
589 ir_node *cmp = value;
590 ir_node *left = get_Cmp_left(cmp);
591 ir_node *right = get_Cmp_right(cmp);
592 ir_relation relation = get_Cmp_relation(cmp);
594 /* we assume that the constant is on the right side, swap left/right
596 if (is_Const(left)) {
601 relation = get_inversed_relation(relation);
604 if (!is_Const(right))
607 if (get_nodes_block(left) != block)
610 /* negate condition when we're looking for the false block */
611 if (env->tv == tarval_b_false) {
612 relation = get_negated_relation(relation);
615 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
617 env->relation = relation;
620 return find_const_or_confirm(env, jump, left);
627 * Block-walker: searches for the following construct
629 * Const or Phi with constants
639 static void thread_jumps(ir_node* block, void* data)
641 jumpthreading_env_t env;
642 bool *changed = (bool*)data;
647 int selector_evaluated;
652 /* we do not deal with Phis, so restrict this to exactly one cfgpred */
653 if (get_Block_n_cfgpreds(block) != 1)
656 projx = get_Block_cfgpred(block, 0);
659 assert(get_irn_mode(projx) == mode_X);
661 cond = get_Proj_pred(projx);
662 /* TODO handle switch Conds */
666 /* handle cases that can be immediately evaluated */
667 selector = get_Cond_selector(cond);
668 selector_evaluated = -1;
669 if (is_Cmp(selector)) {
670 ir_node *left = get_Cmp_left(selector);
671 ir_node *right = get_Cmp_right(selector);
672 if (is_Const(left) && is_Const(right)) {
673 ir_relation relation = get_Cmp_relation(selector);
674 ir_tarval *tv_left = get_Const_tarval(left);
675 ir_tarval *tv_right = get_Const_tarval(right);
677 selector_evaluated = eval_cmp_tv(relation, tv_left, tv_right);
680 if (selector_evaluated < 0) {
681 /* This is only the case if the predecessor nodes are not
682 * constant or the comparison could not be evaluated.
683 * Try with VRP information now.
685 selector_evaluated = eval_cmp_vrp(relation, left, right);
688 } else if (is_Const_or_Confirm(selector)) {
689 ir_tarval *tv = get_Const_or_Confirm_tarval(selector);
690 if (tv == tarval_b_true) {
691 selector_evaluated = 1;
693 assert(tv == tarval_b_false);
694 selector_evaluated = 0;
698 env.cnst_pred = NULL;
699 if (get_Proj_proj(projx) == pn_Cond_false) {
700 env.tv = tarval_b_false;
701 if (selector_evaluated >= 0)
702 selector_evaluated = !selector_evaluated;
704 env.tv = tarval_b_true;
707 if (selector_evaluated == 0) {
708 ir_graph *irg = get_irn_irg(block);
709 ir_node *bad = new_r_Bad(irg, mode_X);
710 exchange(projx, bad);
713 } else if (selector_evaluated == 1) {
714 dbg_info *dbgi = get_irn_dbg_info(selector);
715 ir_node *jmp = new_rd_Jmp(dbgi, get_nodes_block(projx));
716 DBG_OPT_JUMPTHREADING(projx, jmp);
717 exchange(projx, jmp);
722 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
723 env.true_block = block;
724 irg = get_irn_irg(block);
725 inc_irg_visited(irg);
726 env.visited_nr = get_irg_visited(irg);
728 copy_block = find_candidate(&env, projx, selector);
729 if (copy_block == NULL)
732 /* We might thread the condition block of an infinite loop,
733 * such that there is no path to End anymore. */
736 /* we have to remove the edge towards the pred as the pred now
737 * jumps into the true_block. We also have to shorten Phis
738 * in our block because of this */
739 badX = new_r_Bad(irg, mode_X);
740 cnst_pos = env.cnst_pos;
743 foreach_out_edge_safe(env.cnst_pred, edge) {
744 ir_node *node = get_edge_src_irn(edge);
747 ir_node *bad = new_r_Bad(irg, get_irn_mode(node));
748 set_Phi_pred(node, cnst_pos, bad);
752 set_Block_cfgpred(env.cnst_pred, cnst_pos, badX);
754 /* the graph is changed now */
758 void opt_jumpthreading(ir_graph* irg)
763 assure_irg_properties(irg,
764 IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
765 | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
766 | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES);
768 FIRM_DBG_REGISTER(dbg, "firm.opt.jumpthreading");
770 DB((dbg, LEVEL_1, "===> Performing jumpthreading on %+F\n", irg));
772 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
777 irg_block_walk_graph(irg, thread_jumps, NULL, &rerun);
781 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
783 confirm_irg_properties(irg,
784 changed ? IR_GRAPH_PROPERTIES_NONE : IR_GRAPH_PROPERTIES_ALL);
787 /* Creates an ir_graph pass for opt_jumpthreading. */
788 ir_graph_pass_t *opt_jumpthreading_pass(const char *name)
790 return def_graph_pass(name ? name : "jumpthreading", opt_jumpthreading);
791 } /* opt_jumpthreading_pass */