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 Path-Sensitive Jump Threading
24 * @author Christoph Mallon, Matthias Braun
29 #include "iroptimize.h"
41 #include "iredges_t.h"
45 #include "opt_confirms.h"
46 #include "iropt_dbg.h"
50 DEBUG_ONLY(static firm_dbg_module_t *dbg);
53 * Add the new predecessor x to node node, which is either a Block or a Phi
55 static void add_pred(ir_node* node, ir_node* x)
61 assert(is_Block(node) || is_Phi(node));
63 n = get_irn_arity(node);
64 NEW_ARR_A(ir_node*, ins, n + 1);
65 for (i = 0; i < n; i++)
66 ins[i] = get_irn_n(node, i);
68 set_irn_in(node, n + 1, ins);
71 static ir_node *ssa_second_def;
72 static ir_node *ssa_second_def_block;
74 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode,
83 /* This is needed because we create bads sometimes */
87 /* the other defs can't be marked for cases where a user of the original
88 * value is in the same block as the alternative definition.
89 * In this case we mustn't use the alternative definition.
90 * So we keep a flag that indicated wether we walked at least 1 block
91 * away and may use the alternative definition */
92 if (block == ssa_second_def_block && !first) {
93 return ssa_second_def;
96 /* already processed this block? */
97 if (irn_visited(block)) {
98 ir_node *value = (ir_node*) get_irn_link(block);
102 irg = get_irn_irg(block);
103 assert(block != get_irg_start_block(irg));
105 /* a Block with only 1 predecessor needs no Phi */
106 n_cfgpreds = get_Block_n_cfgpreds(block);
107 if (n_cfgpreds == 1) {
108 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
109 ir_node *value = search_def_and_create_phis(pred_block, mode, 0);
111 set_irn_link(block, value);
112 mark_irn_visited(block);
116 /* create a new Phi */
117 NEW_ARR_A(ir_node*, in, n_cfgpreds);
118 for(i = 0; i < n_cfgpreds; ++i)
119 in[i] = new_Unknown(mode);
121 phi = new_r_Phi(block, n_cfgpreds, in, mode);
122 set_irn_link(block, phi);
123 mark_irn_visited(block);
125 /* set Phi predecessors */
126 for(i = 0; i < n_cfgpreds; ++i) {
127 ir_node *pred_block = get_Block_cfgpred_block(block, i);
128 ir_node *pred_val = search_def_and_create_phis(pred_block, mode, 0);
130 set_irn_n(phi, i, pred_val);
137 * Given a set of values this function constructs SSA-form for the users of the
138 * first value (the users are determined through the out-edges of the value).
139 * Uses the irn_visited flags. Works without using the dominance tree.
141 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
142 ir_node *second_block, ir_node *second_val)
146 const ir_edge_t *edge;
147 const ir_edge_t *next;
149 /* no need to do anything */
150 if (orig_val == second_val)
153 irg = get_irn_irg(orig_val);
154 inc_irg_visited(irg);
156 mode = get_irn_mode(orig_val);
157 set_irn_link(orig_block, orig_val);
158 mark_irn_visited(orig_block);
160 ssa_second_def_block = second_block;
161 ssa_second_def = second_val;
163 /* Only fix the users of the first, i.e. the original node */
164 foreach_out_edge_safe(orig_val, edge, next) {
165 ir_node *user = get_edge_src_irn(edge);
166 int j = get_edge_src_pos(edge);
167 ir_node *user_block = get_nodes_block(user);
174 DB((dbg, LEVEL_3, ">>> Fixing user %+F (pred %d == %+F)\n", user, j, get_irn_n(user, j)));
177 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
178 newval = search_def_and_create_phis(pred_block, mode, 1);
180 newval = search_def_and_create_phis(user_block, mode, 1);
183 /* don't fix newly created Phis from the SSA construction */
184 if (newval != user) {
185 DB((dbg, LEVEL_4, ">>>> Setting input %d of %+F to %+F\n", j, user, newval));
186 set_irn_n(user, j, newval);
191 static void split_critical_edge(ir_node *block, int pos) {
192 ir_graph *irg = get_irn_irg(block);
197 in[0] = get_Block_cfgpred(block, pos);
198 new_block = new_r_Block(irg, 1, in);
199 new_jmp = new_r_Jmp(new_block);
200 set_Block_cfgpred(block, pos, new_jmp);
203 typedef struct jumpthreading_env_t {
205 ir_node *cmp; /**< The Compare node that might be partial evaluated */
206 pn_Cmp pnc; /**< The Compare mode of the Compare node. */
209 ir_visited_t visited_nr;
211 ir_node *cnst_pred; /**< the block before the constant */
212 int cnst_pos; /**< the pos to the constant block (needed to
213 kill that edge later) */
214 } jumpthreading_env_t;
216 static ir_node *copy_and_fix_node(const jumpthreading_env_t *env,
217 ir_node *block, ir_node *copy_block, int j,
223 /* we can evaluate Phis right now, all other nodes get copied */
225 copy = get_Phi_pred(node, j);
226 /* we might have to evaluate a Phi-cascade */
227 if (get_irn_visited(copy) >= env->visited_nr) {
228 copy = get_irn_link(copy);
231 copy = exact_copy(node);
232 set_nodes_block(copy, copy_block);
234 assert(get_irn_mode(copy) != mode_X);
236 arity = get_irn_arity(copy);
237 for(i = 0; i < arity; ++i) {
238 ir_node *pred = get_irn_n(copy, i);
241 if (get_nodes_block(pred) != block)
244 if (get_irn_visited(pred) >= env->visited_nr) {
245 new_pred = get_irn_link(pred);
247 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
249 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
250 set_irn_n(copy, i, new_pred);
254 set_irn_link(node, copy);
255 set_irn_visited(node, env->visited_nr);
260 static void copy_and_fix(const jumpthreading_env_t *env, ir_node *block,
261 ir_node *copy_block, int j)
263 const ir_edge_t *edge;
265 /* Look at all nodes in the cond_block and copy them into pred */
266 foreach_out_edge(block, edge) {
267 ir_node *node = get_edge_src_irn(edge);
271 if (is_Block(node)) {
272 /* Block->Block edge, should be the MacroBlock edge */
273 assert(get_Block_MacroBlock(node) == block && "Block->Block edge found");
277 /* ignore control flow */
278 mode = get_irn_mode(node);
279 if (mode == mode_X || is_Cond(node))
282 /* we may not copy mode_b nodes, because this could produce Phi with
283 * mode_bs which can't be handled in all backends. Instead we duplicate
284 * the node and move it to its users */
285 if (mode == mode_b) {
286 const ir_edge_t *edge, *next;
290 assert(is_Proj(node));
292 pred = get_Proj_pred(node);
293 pn = get_Proj_proj(node);
295 foreach_out_edge_safe(node, edge, next) {
297 ir_node *user = get_edge_src_irn(edge);
298 int pos = get_edge_src_pos(edge);
299 ir_node *user_block = get_nodes_block(user);
301 if (user_block == block)
304 cmp_copy = exact_copy(pred);
305 set_nodes_block(cmp_copy, user_block);
306 copy = new_r_Proj(current_ir_graph, user_block, cmp_copy, mode_b, pn);
307 set_irn_n(user, pos, copy);
313 copy = copy_and_fix_node(env, block, copy_block, j, node);
315 /* we might hit values in blocks that have already been processed by a
316 * recursive find_phi_with_const() call */
317 assert(get_irn_visited(copy) <= env->visited_nr);
318 if (get_irn_visited(copy) >= env->visited_nr) {
319 ir_node *prev_copy = get_irn_link(copy);
320 if (prev_copy != NULL)
321 set_irn_link(node, prev_copy);
325 /* fix data-flow (and reconstruct SSA if needed) */
326 foreach_out_edge(block, edge) {
327 ir_node *node = get_edge_src_irn(edge);
331 if (is_Block(node)) {
332 /* Block->Block edge, should be the MacroBlock edge */
333 assert(get_Block_MacroBlock(node) == block && "Block->Block edge found");
337 mode = get_irn_mode(node);
338 if (mode == mode_X || is_Cond(node))
345 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
347 copy_node = get_irn_link(node);
348 construct_ssa(block, node, copy_block, copy_node);
353 * returns whether the cmp evaluates to true or false, or can't be evaluated!
354 * 1: true, 0: false, -1: can't evaluate
356 * @param pnc the compare mode of the Compare
357 * @param tv_left the left tarval
358 * @param tv_right the right tarval
360 static int eval_cmp_tv(pn_Cmp pnc, tarval *tv_left, tarval *tv_right)
362 pn_Cmp cmp_result = tarval_cmp(tv_left, tv_right);
364 /* does the compare evaluate to true? */
365 if (cmp_result == pn_Cmp_False)
367 if ((cmp_result & pnc) != cmp_result)
374 * returns whether the cmp evaluates to true or false, or can't be evaluated!
375 * 1: true, 0: false, -1: can't evaluate
377 * @param env the environment
378 * @param cand the candidate node, either a Const or a Confirm
380 static int eval_cmp(jumpthreading_env_t *env, ir_node *cand)
382 if (is_Const(cand)) {
383 tarval *tv_cand = get_Const_tarval(cand);
384 tarval *tv_cmp = get_Const_tarval(env->cnst);
386 return eval_cmp_tv(env->pnc, tv_cand, tv_cmp);
387 } else { /* a Confirm */
388 tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->pnc);
390 if (res == tarval_bad)
392 return res == tarval_b_true;
397 * Check for Const or Confirm with Const.
399 static int is_Const_or_Confirm(const ir_node *node)
401 if (is_Confirm(node))
402 node = get_Confirm_bound(node);
403 return is_Const(node);
407 * get the tarval of a Const or Confirm with
409 static tarval *get_Const_or_Confirm_tarval(const ir_node *node)
411 if (is_Confirm(node)) {
412 if (get_Confirm_bound(node))
413 node = get_Confirm_bound(node);
415 return get_Const_tarval(node);
418 static ir_node *find_const_or_confirm(jumpthreading_env_t *env, ir_node *jump,
421 ir_node *block = get_nodes_block(jump);
423 if (irn_visited_else_mark(value))
426 if (is_Const_or_Confirm(value)) {
427 if (eval_cmp(env, value) <= 0) {
433 "> Found jump threading candidate %+F->%+F\n",
434 env->true_block, block
437 /* adjust true_block to point directly towards our jump */
438 add_pred(env->true_block, jump);
440 split_critical_edge(env->true_block, 0);
442 /* we need a bigger visited nr when going back */
451 /* the Phi has to be in the same Block as the Jmp */
452 if (get_nodes_block(value) != block) {
456 arity = get_irn_arity(value);
457 for(i = 0; i < arity; ++i) {
459 ir_node *phi_pred = get_Phi_pred(value, i);
460 ir_node *cfgpred = get_Block_cfgpred(block, i);
462 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
463 if (copy_block == NULL)
466 /* copy duplicated nodes in copy_block and fix SSA */
467 copy_and_fix(env, block, copy_block, i);
469 if (copy_block == get_nodes_block(cfgpred)) {
470 env->cnst_pred = block;
474 /* return now as we can't process more possibilities in 1 run */
482 static ir_node *find_candidate(jumpthreading_env_t *env, ir_node *jump,
485 ir_node *block = get_nodes_block(jump);
487 if (irn_visited_else_mark(value)) {
491 if (is_Const_or_Confirm(value)) {
492 tarval *tv = get_Const_or_Confirm_tarval(value);
499 "> Found jump threading candidate %+F->%+F\n",
500 env->true_block, block
503 /* adjust true_block to point directly towards our jump */
504 add_pred(env->true_block, jump);
506 split_critical_edge(env->true_block, 0);
508 /* we need a bigger visited nr when going back */
516 /* the Phi has to be in the same Block as the Jmp */
517 if (get_nodes_block(value) != block)
520 arity = get_irn_arity(value);
521 for(i = 0; i < arity; ++i) {
523 ir_node *phi_pred = get_Phi_pred(value, i);
524 ir_node *cfgpred = get_Block_cfgpred(block, i);
526 copy_block = find_candidate(env, cfgpred, phi_pred);
527 if (copy_block == NULL)
530 /* copy duplicated nodes in copy_block and fix SSA */
531 copy_and_fix(env, block, copy_block, i);
533 if (copy_block == get_nodes_block(cfgpred)) {
534 env->cnst_pred = block;
538 /* return now as we can't process more possibilities in 1 run */
542 if (is_Proj(value)) {
546 ir_node *cmp = get_Proj_pred(value);
550 left = get_Cmp_left(cmp);
551 right = get_Cmp_right(cmp);
552 pnc = get_Proj_proj(value);
554 /* we assume that the constant is on the right side, swap left/right
556 if (is_Const(left)) {
561 pnc = get_inversed_pnc(pnc);
564 if (!is_Const(right))
567 if (get_nodes_block(left) != block) {
571 /* negate condition when we're looking for the false block */
572 if (env->tv == tarval_b_false) {
573 pnc = get_negated_pnc(pnc, get_irn_mode(right));
576 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
581 return find_const_or_confirm(env, jump, left);
588 * Block-walker: searches for the following construct
590 * Const or Phi with constants
600 static void thread_jumps(ir_node* block, void* data)
602 jumpthreading_env_t env;
608 int selector_evaluated;
609 const ir_edge_t *edge, *next;
613 if (get_Block_n_cfgpreds(block) != 1)
616 projx = get_Block_cfgpred(block, 0);
619 assert(get_irn_mode(projx) == mode_X);
621 cond = get_Proj_pred(projx);
625 selector = get_Cond_selector(cond);
626 /* TODO handle switch Conds */
627 if (get_irn_mode(selector) != mode_b)
630 /* handle cases that can be immediately evaluated */
631 selector_evaluated = -1;
632 if (is_Proj(selector)) {
633 ir_node *cmp = get_Proj_pred(selector);
635 ir_node *left = get_Cmp_left(cmp);
636 ir_node *right = get_Cmp_right(cmp);
637 if (is_Const(left) && is_Const(right)) {
638 int pnc = get_Proj_proj(selector);
639 tarval *tv_left = get_Const_tarval(left);
640 tarval *tv_right = get_Const_tarval(right);
642 selector_evaluated = eval_cmp_tv(pnc, tv_left, tv_right);
643 if (selector_evaluated < 0)
647 } else if (is_Const_or_Confirm(selector)) {
648 tarval *tv = get_Const_or_Confirm_tarval(selector);
649 if (tv == tarval_b_true) {
650 selector_evaluated = 1;
652 assert(tv == tarval_b_false);
653 selector_evaluated = 0;
657 env.cnst_pred = NULL;
658 if (get_Proj_proj(projx) == pn_Cond_false) {
659 env.tv = tarval_b_false;
660 if (selector_evaluated >= 0)
661 selector_evaluated = !selector_evaluated;
663 env.tv = tarval_b_true;
666 if (selector_evaluated == 0) {
668 exchange(projx, bad);
671 } else if (selector_evaluated == 1) {
672 dbg_info *dbgi = get_irn_dbg_info(selector);
673 ir_node *jmp = new_rd_Jmp(dbgi, get_nodes_block(projx));
674 DBG_OPT_JUMPTHREADING(projx, jmp);
675 exchange(projx, jmp);
680 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
681 env.true_block = block;
682 inc_irg_visited(current_ir_graph);
683 env.visited_nr = get_irg_visited(current_ir_graph);
685 copy_block = find_candidate(&env, projx, selector);
686 if (copy_block == NULL)
689 /* we have to remove the edge towards the pred as the pred now
690 * jumps into the true_block. We also have to shorten Phis
691 * in our block because of this */
693 cnst_pos = env.cnst_pos;
696 foreach_out_edge_safe(env.cnst_pred, edge, next) {
697 ir_node *node = get_edge_src_irn(edge);
700 set_Phi_pred(node, cnst_pos, bad);
703 set_Block_cfgpred(env.cnst_pred, cnst_pos, bad);
705 /* the graph is changed now */
709 void opt_jumpthreading(ir_graph* irg)
713 FIRM_DBG_REGISTER(dbg, "firm.opt.jumpthreading");
715 DB((dbg, LEVEL_1, "===> Performing jumpthreading on %+F\n", irg));
717 remove_critical_cf_edges(irg);
720 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
725 irg_block_walk_graph(irg, thread_jumps, NULL, &rerun);
729 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
732 /* control flow changed, some blocks may become dead */
733 set_irg_outs_inconsistent(irg);
734 set_irg_doms_inconsistent(irg);
735 set_irg_extblk_inconsistent(irg);
736 set_irg_loopinfo_inconsistent(irg);
737 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
739 /* Dead code might be created. Optimize it away as it is dangerous
740 * to call optimize_df() an dead code. */