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"
49 #include "firmstat_t.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)
154 /* no need to do anything */
155 if (orig_val == second_val)
158 irg = get_irn_irg(orig_val);
159 inc_irg_visited(irg);
161 mode = get_irn_mode(orig_val);
162 set_irn_link(orig_block, orig_val);
163 mark_irn_visited(orig_block);
165 ssa_second_def_block = second_block;
166 ssa_second_def = second_val;
168 /* Only fix the users of the first, i.e. the original node */
169 foreach_out_edge_safe(orig_val, edge) {
170 ir_node *user = get_edge_src_irn(edge);
171 int j = get_edge_src_pos(edge);
172 ir_node *user_block = get_nodes_block(user);
179 DB((dbg, LEVEL_3, ">>> Fixing user %+F (pred %d == %+F)\n", user, j, get_irn_n(user, j)));
182 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
183 newval = search_def_and_create_phis(pred_block, mode, 1);
185 newval = search_def_and_create_phis(user_block, mode, 1);
188 /* don't fix newly created Phis from the SSA construction */
189 if (newval != user) {
190 DB((dbg, LEVEL_4, ">>>> Setting input %d of %+F to %+F\n", j, user, newval));
191 set_irn_n(user, j, newval);
197 * jumpthreading produces critical edges, e.g. B-C:
204 * By splitting this critical edge more threadings might be possible.
206 static void split_critical_edge(ir_node *block, int pos)
208 ir_graph *irg = get_irn_irg(block);
213 in[0] = get_Block_cfgpred(block, pos);
214 new_block = new_r_Block(irg, 1, in);
215 new_jmp = new_r_Jmp(new_block);
216 set_Block_cfgpred(block, pos, new_jmp);
219 typedef struct jumpthreading_env_t {
221 ir_node *cmp; /**< The Compare node that might be partial evaluated */
222 ir_relation relation; /**< The Compare mode of the Compare node. */
225 ir_visited_t visited_nr;
227 ir_node *cnst_pred; /**< the block before the constant */
228 int cnst_pos; /**< the pos to the constant block (needed to
229 kill that edge later) */
230 } jumpthreading_env_t;
232 static ir_node *copy_and_fix_node(const jumpthreading_env_t *env,
233 ir_node *block, ir_node *copy_block, int j,
239 /* we can evaluate Phis right now, all other nodes get copied */
241 copy = get_Phi_pred(node, j);
242 /* we might have to evaluate a Phi-cascade */
243 if (get_irn_visited(copy) >= env->visited_nr) {
244 copy = (ir_node*)get_irn_link(copy);
247 copy = exact_copy(node);
248 set_nodes_block(copy, copy_block);
250 assert(get_irn_mode(copy) != mode_X);
252 arity = get_irn_arity(copy);
253 for (i = 0; i < arity; ++i) {
254 ir_node *pred = get_irn_n(copy, i);
257 if (get_nodes_block(pred) != block)
260 if (get_irn_visited(pred) >= env->visited_nr) {
261 new_pred = (ir_node*)get_irn_link(pred);
263 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
265 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
266 set_irn_n(copy, i, new_pred);
270 set_irn_link(node, copy);
271 set_irn_visited(node, env->visited_nr);
276 static void copy_and_fix(const jumpthreading_env_t *env, ir_node *block,
277 ir_node *copy_block, int j)
279 /* Look at all nodes in the cond_block and copy them into pred */
280 foreach_out_edge(block, edge) {
281 ir_node *node = get_edge_src_irn(edge);
286 /* edge is a Keep edge. If the end block is unreachable via normal
287 * control flow, we must maintain end's reachability with Keeps.
289 keep_alive(copy_block);
292 /* ignore control flow */
293 mode = get_irn_mode(node);
294 if (mode == mode_X || is_Cond(node) || is_Switch(node))
297 /* we may not copy mode_b nodes, because this could produce Phi with
298 * mode_bs which can't be handled in all backends. Instead we duplicate
299 * the node and move it to its users */
300 if (mode == mode_b) {
304 assert(is_Proj(node));
306 pred = get_Proj_pred(node);
307 pn = get_Proj_proj(node);
309 foreach_out_edge_safe(node, edge) {
311 ir_node *user = get_edge_src_irn(edge);
312 int pos = get_edge_src_pos(edge);
313 ir_node *user_block = get_nodes_block(user);
315 if (user_block == block)
318 cmp_copy = exact_copy(pred);
319 set_nodes_block(cmp_copy, user_block);
320 copy = new_r_Proj(cmp_copy, mode_b, pn);
321 set_irn_n(user, pos, copy);
327 copy = copy_and_fix_node(env, block, copy_block, j, node);
329 /* we might hit values in blocks that have already been processed by a
330 * recursive find_phi_with_const() call */
331 assert(get_irn_visited(copy) <= env->visited_nr);
332 if (get_irn_visited(copy) >= env->visited_nr) {
333 ir_node *prev_copy = (ir_node*)get_irn_link(copy);
334 if (prev_copy != NULL)
335 set_irn_link(node, prev_copy);
339 /* fix data-flow (and reconstruct SSA if needed) */
340 foreach_out_edge(block, edge) {
341 ir_node *node = get_edge_src_irn(edge);
345 mode = get_irn_mode(node);
346 if (mode == mode_X || is_Cond(node) || is_Switch(node))
353 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
355 copy_node = (ir_node*)get_irn_link(node);
356 construct_ssa(block, node, copy_block, copy_node);
359 /* make sure new nodes are kept alive if old nodes were */
360 ir_graph *irg = get_irn_irg(block);
361 ir_node *end = get_irg_end(irg);
362 for (int i = 0, arity = get_End_n_keepalives(end); i < arity; ++i) {
363 ir_node *keep = get_End_keepalive(end, i);
364 if (get_irn_visited(keep) < env->visited_nr || is_Block(keep))
366 ir_node *copy = get_irn_link(keep);
367 add_End_keepalive(end, copy);
372 * returns whether the cmp evaluates to true or false, or can't be evaluated!
373 * 1: true, 0: false, -1: can't evaluate
375 * @param relation the compare mode of the Compare
376 * @param tv_left the left tarval
377 * @param tv_right the right tarval
379 static int eval_cmp_tv(ir_relation relation, ir_tarval *tv_left,
382 ir_relation cmp_result = tarval_cmp(tv_left, tv_right);
384 /* does the compare evaluate to true? */
385 if (cmp_result == ir_relation_false)
387 if ((cmp_result & relation) != 0)
394 /* Matze: disabled, check first if the compare still is correct */
397 * returns whether the cmp evaluates to true or false according to vrp
398 * information , or can't be evaluated!
399 * 1: true, 0: false, -1: can't evaluate
401 * @param relation the compare mode of the Compare
402 * @param left the left node
403 * @param right the right node
405 static int eval_cmp_vrp(ir_relation relation, ir_node *left, ir_node *right)
407 ir_relation cmp_result = vrp_cmp(left, right);
408 /* does the compare evaluate to true? */
409 if (cmp_result == ir_relation_false)
412 if ((cmp_result & relation) != cmp_result) {
413 if ((cmp_result & relation) != 0) {
423 * returns whether the cmp evaluates to true or false, or can't be evaluated!
424 * 1: true, 0: false, -1: can't evaluate
426 * @param env the environment
427 * @param cand the candidate node, either a Const or a Confirm
429 static int eval_cmp(jumpthreading_env_t *env, ir_node *cand)
431 if (is_Const(cand)) {
432 ir_tarval *tv_cand = get_Const_tarval(cand);
433 ir_tarval *tv_cmp = get_Const_tarval(env->cnst);
435 return eval_cmp_tv(env->relation, tv_cand, tv_cmp);
436 } else { /* a Confirm */
437 ir_tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->relation);
439 if (res == tarval_bad)
441 return res == tarval_b_true;
446 * Check for Const or Confirm with Const.
448 static int is_Const_or_Confirm(const ir_node *node)
450 if (is_Confirm(node))
451 node = get_Confirm_bound(node);
452 return is_Const(node);
456 * get the tarval of a Const or Confirm with
458 static ir_tarval *get_Const_or_Confirm_tarval(const ir_node *node)
460 if (is_Confirm(node)) {
461 if (get_Confirm_bound(node))
462 node = get_Confirm_bound(node);
464 return get_Const_tarval(node);
467 static ir_node *find_const_or_confirm(jumpthreading_env_t *env, ir_node *jump,
470 ir_node *block = get_nodes_block(jump);
472 if (irn_visited_else_mark(value))
475 if (is_Const_or_Confirm(value)) {
476 if (eval_cmp(env, value) <= 0)
481 "> Found jump threading candidate %+F->%+F\n",
482 block, env->true_block
485 /* adjust true_block to point directly towards our jump */
486 add_pred(env->true_block, jump);
488 split_critical_edge(env->true_block, 0);
490 /* we need a bigger visited nr when going back */
499 /* the Phi has to be in the same Block as the Jmp */
500 if (get_nodes_block(value) != block)
503 arity = get_irn_arity(value);
504 for (i = 0; i < arity; ++i) {
506 ir_node *phi_pred = get_Phi_pred(value, i);
507 ir_node *cfgpred = get_Block_cfgpred(block, i);
509 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
510 if (copy_block == NULL)
513 /* copy duplicated nodes in copy_block and fix SSA */
514 copy_and_fix(env, block, copy_block, i);
516 if (copy_block == get_nodes_block(cfgpred)) {
517 env->cnst_pred = block;
521 /* return now as we can't process more possibilities in 1 run */
529 static ir_node *find_candidate(jumpthreading_env_t *env, ir_node *jump,
532 ir_node *block = get_nodes_block(jump);
534 if (irn_visited_else_mark(value)) {
538 if (is_Const_or_Confirm(value)) {
539 ir_tarval *tv = get_Const_or_Confirm_tarval(value);
546 "> Found jump threading candidate %+F->%+F\n",
547 block, env->true_block
550 /* adjust true_block to point directly towards our jump */
551 add_pred(env->true_block, jump);
553 split_critical_edge(env->true_block, 0);
555 /* we need a bigger visited nr when going back */
563 /* the Phi has to be in the same Block as the Jmp */
564 if (get_nodes_block(value) != block)
567 arity = get_irn_arity(value);
568 for (i = 0; i < arity; ++i) {
570 ir_node *phi_pred = get_Phi_pred(value, i);
571 ir_node *cfgpred = get_Block_cfgpred(block, i);
573 copy_block = find_candidate(env, cfgpred, phi_pred);
574 if (copy_block == NULL)
577 /* copy duplicated nodes in copy_block and fix SSA */
578 copy_and_fix(env, block, copy_block, i);
580 if (copy_block == get_nodes_block(cfgpred)) {
581 env->cnst_pred = block;
585 /* return now as we can't process more possibilities in 1 run */
590 ir_node *cmp = value;
591 ir_node *left = get_Cmp_left(cmp);
592 ir_node *right = get_Cmp_right(cmp);
593 ir_relation relation = get_Cmp_relation(cmp);
595 /* we assume that the constant is on the right side, swap left/right
597 if (is_Const(left)) {
602 relation = get_inversed_relation(relation);
605 if (!is_Const(right))
608 if (get_nodes_block(left) != block)
611 /* negate condition when we're looking for the false block */
612 if (env->tv == tarval_b_false) {
613 relation = get_negated_relation(relation);
616 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
618 env->relation = relation;
621 return find_const_or_confirm(env, jump, left);
628 * Block-walker: searches for the following construct
630 * Const or Phi with constants
640 static void thread_jumps(ir_node* block, void* data)
642 jumpthreading_env_t env;
643 bool *changed = (bool*)data;
648 int selector_evaluated;
653 /* we do not deal with Phis, so restrict this to exactly one cfgpred */
654 if (get_Block_n_cfgpreds(block) != 1)
657 projx = get_Block_cfgpred(block, 0);
660 assert(get_irn_mode(projx) == mode_X);
662 cond = get_Proj_pred(projx);
663 /* TODO handle switch Conds */
667 /* handle cases that can be immediately evaluated */
668 selector = get_Cond_selector(cond);
669 selector_evaluated = -1;
670 if (is_Cmp(selector)) {
671 ir_node *left = get_Cmp_left(selector);
672 ir_node *right = get_Cmp_right(selector);
673 if (is_Const(left) && is_Const(right)) {
674 ir_relation relation = get_Cmp_relation(selector);
675 ir_tarval *tv_left = get_Const_tarval(left);
676 ir_tarval *tv_right = get_Const_tarval(right);
678 selector_evaluated = eval_cmp_tv(relation, tv_left, tv_right);
681 if (selector_evaluated < 0) {
682 /* This is only the case if the predecessor nodes are not
683 * constant or the comparison could not be evaluated.
684 * Try with VRP information now.
686 selector_evaluated = eval_cmp_vrp(relation, left, right);
689 } else if (is_Const_or_Confirm(selector)) {
690 ir_tarval *tv = get_Const_or_Confirm_tarval(selector);
691 if (tv == tarval_b_true) {
692 selector_evaluated = 1;
694 assert(tv == tarval_b_false);
695 selector_evaluated = 0;
699 env.cnst_pred = NULL;
700 if (get_Proj_proj(projx) == pn_Cond_false) {
701 env.tv = tarval_b_false;
702 if (selector_evaluated >= 0)
703 selector_evaluated = !selector_evaluated;
705 env.tv = tarval_b_true;
708 if (selector_evaluated == 0) {
709 ir_graph *irg = get_irn_irg(block);
710 ir_node *bad = new_r_Bad(irg, mode_X);
711 exchange(projx, bad);
714 } else if (selector_evaluated == 1) {
715 dbg_info *dbgi = get_irn_dbg_info(selector);
716 ir_node *jmp = new_rd_Jmp(dbgi, get_nodes_block(projx));
717 DBG_OPT_JUMPTHREADING(projx, jmp);
718 exchange(projx, jmp);
723 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
724 env.true_block = block;
725 irg = get_irn_irg(block);
726 inc_irg_visited(irg);
727 env.visited_nr = get_irg_visited(irg);
729 copy_block = find_candidate(&env, projx, selector);
730 if (copy_block == NULL)
733 /* We might thread the condition block of an infinite loop,
734 * such that there is no path to End anymore. */
737 /* we have to remove the edge towards the pred as the pred now
738 * jumps into the true_block. We also have to shorten Phis
739 * in our block because of this */
740 badX = new_r_Bad(irg, mode_X);
741 cnst_pos = env.cnst_pos;
744 foreach_out_edge_safe(env.cnst_pred, edge) {
745 ir_node *node = get_edge_src_irn(edge);
748 ir_node *bad = new_r_Bad(irg, get_irn_mode(node));
749 set_Phi_pred(node, cnst_pos, bad);
753 set_Block_cfgpred(env.cnst_pred, cnst_pos, badX);
755 /* the graph is changed now */
759 void opt_jumpthreading(ir_graph* irg)
764 assure_irg_properties(irg,
765 IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
766 | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
767 | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES);
769 FIRM_DBG_REGISTER(dbg, "firm.opt.jumpthreading");
771 DB((dbg, LEVEL_1, "===> Performing jumpthreading on %+F\n", irg));
773 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
778 irg_block_walk_graph(irg, thread_jumps, NULL, &rerun);
782 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
784 confirm_irg_properties(irg,
785 changed ? IR_GRAPH_PROPERTIES_NONE : IR_GRAPH_PROPERTIES_ALL);
788 /* Creates an ir_graph pass for opt_jumpthreading. */
789 ir_graph_pass_t *opt_jumpthreading_pass(const char *name)
791 return def_graph_pass(name ? name : "jumpthreading", opt_jumpthreading);
792 } /* opt_jumpthreading_pass */