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"
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) || is_Phi(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 /* This is needed because we create bads sometimes */
88 ir_graph *irg = get_irn_irg(block);
89 return new_r_Bad(irg);
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 wether 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)
152 const ir_edge_t *edge;
153 const ir_edge_t *next;
155 /* no need to do anything */
156 if (orig_val == second_val)
159 irg = get_irn_irg(orig_val);
160 inc_irg_visited(irg);
162 mode = get_irn_mode(orig_val);
163 set_irn_link(orig_block, orig_val);
164 mark_irn_visited(orig_block);
166 ssa_second_def_block = second_block;
167 ssa_second_def = second_val;
169 /* Only fix the users of the first, i.e. the original node */
170 foreach_out_edge_safe(orig_val, edge, next) {
171 ir_node *user = get_edge_src_irn(edge);
172 int j = get_edge_src_pos(edge);
173 ir_node *user_block = get_nodes_block(user);
180 DB((dbg, LEVEL_3, ">>> Fixing user %+F (pred %d == %+F)\n", user, j, get_irn_n(user, j)));
183 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
184 newval = search_def_and_create_phis(pred_block, mode, 1);
186 newval = search_def_and_create_phis(user_block, mode, 1);
189 /* don't fix newly created Phis from the SSA construction */
190 if (newval != user) {
191 DB((dbg, LEVEL_4, ">>>> Setting input %d of %+F to %+F\n", j, user, newval));
192 set_irn_n(user, j, newval);
197 static void split_critical_edge(ir_node *block, int pos)
199 ir_graph *irg = get_irn_irg(block);
204 in[0] = get_Block_cfgpred(block, pos);
205 new_block = new_r_Block(irg, 1, in);
206 new_jmp = new_r_Jmp(new_block);
207 set_Block_cfgpred(block, pos, new_jmp);
210 typedef struct jumpthreading_env_t {
212 ir_node *cmp; /**< The Compare node that might be partial evaluated */
213 pn_Cmp pnc; /**< The Compare mode of the Compare node. */
216 ir_visited_t visited_nr;
218 ir_node *cnst_pred; /**< the block before the constant */
219 int cnst_pos; /**< the pos to the constant block (needed to
220 kill that edge later) */
221 } jumpthreading_env_t;
223 static ir_node *copy_and_fix_node(const jumpthreading_env_t *env,
224 ir_node *block, ir_node *copy_block, int j,
230 /* we can evaluate Phis right now, all other nodes get copied */
232 copy = get_Phi_pred(node, j);
233 /* we might have to evaluate a Phi-cascade */
234 if (get_irn_visited(copy) >= env->visited_nr) {
235 copy = get_irn_link(copy);
238 copy = exact_copy(node);
239 set_nodes_block(copy, copy_block);
241 assert(get_irn_mode(copy) != mode_X);
243 arity = get_irn_arity(copy);
244 for (i = 0; i < arity; ++i) {
245 ir_node *pred = get_irn_n(copy, i);
248 if (get_nodes_block(pred) != block)
251 if (get_irn_visited(pred) >= env->visited_nr) {
252 new_pred = get_irn_link(pred);
254 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
256 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
257 set_irn_n(copy, i, new_pred);
261 set_irn_link(node, copy);
262 set_irn_visited(node, env->visited_nr);
267 static void copy_and_fix(const jumpthreading_env_t *env, ir_node *block,
268 ir_node *copy_block, int j)
270 const ir_edge_t *edge;
272 /* Look at all nodes in the cond_block and copy them into pred */
273 foreach_out_edge(block, edge) {
274 ir_node *node = get_edge_src_irn(edge);
278 /* ignore control flow */
279 mode = get_irn_mode(node);
280 if (mode == mode_X || is_Cond(node))
283 /* we may not copy mode_b nodes, because this could produce Phi with
284 * mode_bs which can't be handled in all backends. Instead we duplicate
285 * the node and move it to its users */
286 if (mode == mode_b) {
287 const ir_edge_t *edge, *next;
291 assert(is_Proj(node));
293 pred = get_Proj_pred(node);
294 pn = get_Proj_proj(node);
296 foreach_out_edge_safe(node, edge, next) {
298 ir_node *user = get_edge_src_irn(edge);
299 int pos = get_edge_src_pos(edge);
300 ir_node *user_block = get_nodes_block(user);
302 if (user_block == block)
305 cmp_copy = exact_copy(pred);
306 set_nodes_block(cmp_copy, user_block);
307 copy = new_r_Proj(cmp_copy, mode_b, pn);
308 set_irn_n(user, pos, copy);
314 copy = copy_and_fix_node(env, block, copy_block, j, node);
316 /* we might hit values in blocks that have already been processed by a
317 * recursive find_phi_with_const() call */
318 assert(get_irn_visited(copy) <= env->visited_nr);
319 if (get_irn_visited(copy) >= env->visited_nr) {
320 ir_node *prev_copy = get_irn_link(copy);
321 if (prev_copy != NULL)
322 set_irn_link(node, prev_copy);
326 /* fix data-flow (and reconstruct SSA if needed) */
327 foreach_out_edge(block, edge) {
328 ir_node *node = get_edge_src_irn(edge);
332 mode = get_irn_mode(node);
333 if (mode == mode_X || is_Cond(node))
340 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
342 copy_node = get_irn_link(node);
343 construct_ssa(block, node, copy_block, copy_node);
348 * returns whether the cmp evaluates to true or false, or can't be evaluated!
349 * 1: true, 0: false, -1: can't evaluate
351 * @param pnc the compare mode of the Compare
352 * @param tv_left the left tarval
353 * @param tv_right the right tarval
355 static int eval_cmp_tv(pn_Cmp pnc, tarval *tv_left, tarval *tv_right)
357 pn_Cmp cmp_result = tarval_cmp(tv_left, tv_right);
359 /* does the compare evaluate to true? */
360 if (cmp_result == pn_Cmp_False)
362 if ((cmp_result & pnc) != cmp_result)
369 * returns whether the cmp evaluates to true or false according to vrp
370 * information , or can't be evaluated!
371 * 1: true, 0: false, -1: can't evaluate
373 * @param pnc the compare mode of the Compare
374 * @param left the left node
375 * @param right the right node
377 static int eval_cmp_vrp(pn_Cmp pnc, ir_node *left, ir_node *right)
379 pn_Cmp cmp_result = vrp_cmp(left, right);
380 /* does the compare evaluate to true? */
381 if (cmp_result == pn_Cmp_False) {
384 if ((cmp_result & pnc) != cmp_result) {
385 if ((cmp_result & pnc) != 0) {
393 * returns whether the cmp evaluates to true or false, or can't be evaluated!
394 * 1: true, 0: false, -1: can't evaluate
396 * @param env the environment
397 * @param cand the candidate node, either a Const or a Confirm
399 static int eval_cmp(jumpthreading_env_t *env, ir_node *cand)
401 if (is_Const(cand)) {
402 tarval *tv_cand = get_Const_tarval(cand);
403 tarval *tv_cmp = get_Const_tarval(env->cnst);
405 return eval_cmp_tv(env->pnc, tv_cand, tv_cmp);
406 } else { /* a Confirm */
407 tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->pnc);
409 if (res == tarval_bad)
411 return res == tarval_b_true;
416 * Check for Const or Confirm with Const.
418 static int is_Const_or_Confirm(const ir_node *node)
420 if (is_Confirm(node))
421 node = get_Confirm_bound(node);
422 return is_Const(node);
426 * get the tarval of a Const or Confirm with
428 static tarval *get_Const_or_Confirm_tarval(const ir_node *node)
430 if (is_Confirm(node)) {
431 if (get_Confirm_bound(node))
432 node = get_Confirm_bound(node);
434 return get_Const_tarval(node);
437 static ir_node *find_const_or_confirm(jumpthreading_env_t *env, ir_node *jump,
440 ir_node *block = get_nodes_block(jump);
442 if (irn_visited_else_mark(value))
445 if (is_Const_or_Confirm(value)) {
446 if (eval_cmp(env, value) <= 0) {
452 "> Found jump threading candidate %+F->%+F\n",
453 env->true_block, block
456 /* adjust true_block to point directly towards our jump */
457 add_pred(env->true_block, jump);
459 split_critical_edge(env->true_block, 0);
461 /* we need a bigger visited nr when going back */
470 /* the Phi has to be in the same Block as the Jmp */
471 if (get_nodes_block(value) != block) {
475 arity = get_irn_arity(value);
476 for (i = 0; i < arity; ++i) {
478 ir_node *phi_pred = get_Phi_pred(value, i);
479 ir_node *cfgpred = get_Block_cfgpred(block, i);
481 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
482 if (copy_block == NULL)
485 /* copy duplicated nodes in copy_block and fix SSA */
486 copy_and_fix(env, block, copy_block, i);
488 if (copy_block == get_nodes_block(cfgpred)) {
489 env->cnst_pred = block;
493 /* return now as we can't process more possibilities in 1 run */
501 static ir_node *find_candidate(jumpthreading_env_t *env, ir_node *jump,
504 ir_node *block = get_nodes_block(jump);
506 if (irn_visited_else_mark(value)) {
510 if (is_Const_or_Confirm(value)) {
511 tarval *tv = get_Const_or_Confirm_tarval(value);
518 "> Found jump threading candidate %+F->%+F\n",
519 env->true_block, block
522 /* adjust true_block to point directly towards our jump */
523 add_pred(env->true_block, jump);
525 split_critical_edge(env->true_block, 0);
527 /* we need a bigger visited nr when going back */
535 /* the Phi has to be in the same Block as the Jmp */
536 if (get_nodes_block(value) != block)
539 arity = get_irn_arity(value);
540 for (i = 0; i < arity; ++i) {
542 ir_node *phi_pred = get_Phi_pred(value, i);
543 ir_node *cfgpred = get_Block_cfgpred(block, i);
545 copy_block = find_candidate(env, cfgpred, phi_pred);
546 if (copy_block == NULL)
549 /* copy duplicated nodes in copy_block and fix SSA */
550 copy_and_fix(env, block, copy_block, i);
552 if (copy_block == get_nodes_block(cfgpred)) {
553 env->cnst_pred = block;
557 /* return now as we can't process more possibilities in 1 run */
561 if (is_Proj(value)) {
565 ir_node *cmp = get_Proj_pred(value);
569 left = get_Cmp_left(cmp);
570 right = get_Cmp_right(cmp);
571 pnc = get_Proj_proj(value);
573 /* we assume that the constant is on the right side, swap left/right
575 if (is_Const(left)) {
580 pnc = get_inversed_pnc(pnc);
583 if (!is_Const(right))
586 if (get_nodes_block(left) != block) {
590 /* negate condition when we're looking for the false block */
591 if (env->tv == tarval_b_false) {
592 pnc = get_negated_pnc(pnc, get_irn_mode(right));
595 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
600 return find_const_or_confirm(env, jump, left);
607 * Block-walker: searches for the following construct
609 * Const or Phi with constants
619 static void thread_jumps(ir_node* block, void* data)
621 jumpthreading_env_t env;
627 int selector_evaluated;
628 const ir_edge_t *edge, *next;
633 if (get_Block_n_cfgpreds(block) != 1)
636 projx = get_Block_cfgpred(block, 0);
639 assert(get_irn_mode(projx) == mode_X);
641 cond = get_Proj_pred(projx);
645 selector = get_Cond_selector(cond);
646 /* TODO handle switch Conds */
647 if (get_irn_mode(selector) != mode_b)
650 /* handle cases that can be immediately evaluated */
651 selector_evaluated = -1;
652 if (is_Proj(selector)) {
653 ir_node *cmp = get_Proj_pred(selector);
655 ir_node *left = get_Cmp_left(cmp);
656 ir_node *right = get_Cmp_right(cmp);
657 if (is_Const(left) && is_Const(right)) {
658 int pnc = get_Proj_proj(selector);
659 tarval *tv_left = get_Const_tarval(left);
660 tarval *tv_right = get_Const_tarval(right);
662 selector_evaluated = eval_cmp_tv(pnc, tv_left, tv_right);
664 if (selector_evaluated < 0) {
665 /* This is only the case if the predecessor nodes are not
666 * constant or the comparison could not be evaluated.
667 * Try with VRP information now.
669 int pnc = get_Proj_proj(selector);
671 selector_evaluated = eval_cmp_vrp(pnc, left, right);
674 } else if (is_Const_or_Confirm(selector)) {
675 tarval *tv = get_Const_or_Confirm_tarval(selector);
676 if (tv == tarval_b_true) {
677 selector_evaluated = 1;
679 assert(tv == tarval_b_false);
680 selector_evaluated = 0;
684 env.cnst_pred = NULL;
685 if (get_Proj_proj(projx) == pn_Cond_false) {
686 env.tv = tarval_b_false;
687 if (selector_evaluated >= 0)
688 selector_evaluated = !selector_evaluated;
690 env.tv = tarval_b_true;
693 if (selector_evaluated == 0) {
694 ir_graph *irg = get_irn_irg(block);
695 bad = new_r_Bad(irg);
696 exchange(projx, bad);
699 } else if (selector_evaluated == 1) {
700 dbg_info *dbgi = get_irn_dbg_info(selector);
701 ir_node *jmp = new_rd_Jmp(dbgi, get_nodes_block(projx));
702 DBG_OPT_JUMPTHREADING(projx, jmp);
703 exchange(projx, jmp);
708 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
709 env.true_block = block;
710 irg = get_irn_irg(block);
711 inc_irg_visited(irg);
712 env.visited_nr = get_irg_visited(irg);
714 copy_block = find_candidate(&env, projx, selector);
715 if (copy_block == NULL)
718 /* we have to remove the edge towards the pred as the pred now
719 * jumps into the true_block. We also have to shorten Phis
720 * in our block because of this */
721 bad = new_r_Bad(irg);
722 cnst_pos = env.cnst_pos;
725 foreach_out_edge_safe(env.cnst_pred, edge, next) {
726 ir_node *node = get_edge_src_irn(edge);
729 set_Phi_pred(node, cnst_pos, bad);
732 set_Block_cfgpred(env.cnst_pred, cnst_pos, bad);
734 /* the graph is changed now */
738 void opt_jumpthreading(ir_graph* irg)
742 FIRM_DBG_REGISTER(dbg, "firm.opt.jumpthreading");
744 DB((dbg, LEVEL_1, "===> Performing jumpthreading on %+F\n", irg));
746 remove_critical_cf_edges(irg);
749 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
754 irg_block_walk_graph(irg, thread_jumps, NULL, &rerun);
758 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
761 /* control flow changed, some blocks may become dead */
762 set_irg_outs_inconsistent(irg);
763 set_irg_doms_inconsistent(irg);
764 set_irg_extblk_inconsistent(irg);
765 set_irg_loopinfo_inconsistent(irg);
766 set_irg_entity_usage_state(irg, ir_entity_usage_not_computed);
768 /* Dead code might be created. Optimize it away as it is dangerous
769 * to call optimize_df() an dead code. */
774 /* Creates an ir_graph pass for opt_jumpthreading. */
775 ir_graph_pass_t *opt_jumpthreading_pass(const char *name)
777 return def_graph_pass(name ? name : "jumpthreading", opt_jumpthreading);
778 } /* opt_jumpthreading_pass */