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)
62 int const n = get_Block_n_cfgpreds(node);
63 NEW_ARR_A(ir_node*, ins, n + 1);
64 for (int i = 0; i < n; i++)
65 ins[i] = get_irn_n(node, i);
67 set_irn_in(node, n + 1, ins);
70 static ir_node *ssa_second_def;
71 static ir_node *ssa_second_def_block;
73 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode,
83 /* In case of a bad input to a block we need to return the bad value */
85 ir_graph *irg = get_irn_irg(block);
86 return new_r_Bad(irg, mode);
89 /* the other defs can't be marked for cases where a user of the original
90 * value is in the same block as the alternative definition.
91 * In this case we mustn't use the alternative definition.
92 * So we keep a flag that indicated whether we walked at least 1 block
93 * away and may use the alternative definition */
94 if (block == ssa_second_def_block && !first) {
95 return ssa_second_def;
98 /* already processed this block? */
99 if (irn_visited(block)) {
100 ir_node *value = (ir_node*) get_irn_link(block);
104 irg = get_irn_irg(block);
105 assert(block != get_irg_start_block(irg));
107 /* a Block with only 1 predecessor needs no Phi */
108 n_cfgpreds = get_Block_n_cfgpreds(block);
109 if (n_cfgpreds == 1) {
110 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
111 ir_node *value = search_def_and_create_phis(pred_block, mode, 0);
113 set_irn_link(block, value);
114 mark_irn_visited(block);
118 /* create a new Phi */
119 NEW_ARR_A(ir_node*, in, n_cfgpreds);
120 dummy = new_r_Dummy(irg, mode);
121 for (i = 0; i < n_cfgpreds; ++i)
124 phi = new_r_Phi(block, n_cfgpreds, in, mode);
125 set_irn_link(block, phi);
126 mark_irn_visited(block);
128 /* set Phi predecessors */
129 for (i = 0; i < n_cfgpreds; ++i) {
130 ir_node *pred_block = get_Block_cfgpred_block(block, i);
131 ir_node *pred_val = search_def_and_create_phis(pred_block, mode, 0);
133 set_irn_n(phi, i, pred_val);
140 * Given a set of values this function constructs SSA-form for the users of the
141 * first value (the users are determined through the out-edges of the value).
142 * Uses the irn_visited flags. Works without using the dominance tree.
144 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
145 ir_node *second_block, ir_node *second_val)
150 /* no need to do anything */
151 if (orig_val == second_val)
154 irg = get_irn_irg(orig_val);
155 inc_irg_visited(irg);
157 mode = get_irn_mode(orig_val);
158 set_irn_link(orig_block, orig_val);
159 mark_irn_visited(orig_block);
161 ssa_second_def_block = second_block;
162 ssa_second_def = second_val;
164 /* Only fix the users of the first, i.e. the original node */
165 foreach_out_edge_safe(orig_val, edge) {
166 ir_node *user = get_edge_src_irn(edge);
167 int j = get_edge_src_pos(edge);
168 ir_node *user_block = get_nodes_block(user);
175 DB((dbg, LEVEL_3, ">>> Fixing user %+F (pred %d == %+F)\n", user, j, get_irn_n(user, j)));
178 ir_node *pred_block = get_Block_cfgpred_block(user_block, j);
179 newval = search_def_and_create_phis(pred_block, mode, 1);
181 newval = search_def_and_create_phis(user_block, mode, 1);
184 /* don't fix newly created Phis from the SSA construction */
185 if (newval != user) {
186 DB((dbg, LEVEL_4, ">>>> Setting input %d of %+F to %+F\n", j, user, newval));
187 set_irn_n(user, j, newval);
193 * jumpthreading produces critical edges, e.g. B-C:
200 * By splitting this critical edge more threadings might be possible.
202 static void split_critical_edge(ir_node *block, int pos)
204 ir_graph *irg = get_irn_irg(block);
209 in[0] = get_Block_cfgpred(block, pos);
210 new_block = new_r_Block(irg, 1, in);
211 new_jmp = new_r_Jmp(new_block);
212 set_Block_cfgpred(block, pos, new_jmp);
215 typedef struct jumpthreading_env_t {
217 ir_node *cmp; /**< The Compare node that might be partial evaluated */
218 ir_relation relation; /**< The Compare mode of the Compare node. */
221 ir_visited_t visited_nr;
223 ir_node *cnst_pred; /**< the block before the constant */
224 int cnst_pos; /**< the pos to the constant block (needed to
225 kill that edge later) */
226 } jumpthreading_env_t;
228 static ir_node *copy_and_fix_node(const jumpthreading_env_t *env,
229 ir_node *block, ir_node *copy_block, int j,
235 /* we can evaluate Phis right now, all other nodes get copied */
237 copy = get_Phi_pred(node, j);
238 /* we might have to evaluate a Phi-cascade */
239 if (get_irn_visited(copy) >= env->visited_nr) {
240 copy = (ir_node*)get_irn_link(copy);
243 copy = exact_copy(node);
244 set_nodes_block(copy, copy_block);
246 assert(get_irn_mode(copy) != mode_X);
248 arity = get_irn_arity(copy);
249 for (i = 0; i < arity; ++i) {
250 ir_node *pred = get_irn_n(copy, i);
253 if (get_nodes_block(pred) != block)
256 if (get_irn_visited(pred) >= env->visited_nr) {
257 new_pred = (ir_node*)get_irn_link(pred);
259 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
261 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
262 set_irn_n(copy, i, new_pred);
266 set_irn_link(node, copy);
267 set_irn_visited(node, env->visited_nr);
272 static void copy_and_fix(const jumpthreading_env_t *env, ir_node *block,
273 ir_node *copy_block, int j)
275 /* Look at all nodes in the cond_block and copy them into pred */
276 foreach_out_edge(block, edge) {
277 ir_node *node = get_edge_src_irn(edge);
282 /* edge is a Keep edge. If the end block is unreachable via normal
283 * control flow, we must maintain end's reachability with Keeps.
285 keep_alive(copy_block);
288 /* ignore control flow */
289 mode = get_irn_mode(node);
290 if (mode == mode_X || is_Cond(node) || is_Switch(node))
293 /* we may not copy mode_b nodes, because this could produce Phi with
294 * mode_bs which can't be handled in all backends. Instead we duplicate
295 * the node and move it to its users */
296 if (mode == mode_b) {
297 ir_node *const pred = get_Proj_pred(node);
298 long const pn = get_Proj_proj(node);
300 foreach_out_edge_safe(node, edge) {
302 ir_node *user = get_edge_src_irn(edge);
303 int pos = get_edge_src_pos(edge);
304 ir_node *user_block = get_nodes_block(user);
306 if (user_block == block)
309 cmp_copy = exact_copy(pred);
310 set_nodes_block(cmp_copy, user_block);
311 copy = new_r_Proj(cmp_copy, mode_b, pn);
312 set_irn_n(user, pos, copy);
318 copy = copy_and_fix_node(env, block, copy_block, j, node);
320 /* we might hit values in blocks that have already been processed by a
321 * recursive find_phi_with_const() call */
322 assert(get_irn_visited(copy) <= env->visited_nr);
323 if (get_irn_visited(copy) >= env->visited_nr) {
324 ir_node *prev_copy = (ir_node*)get_irn_link(copy);
325 if (prev_copy != NULL)
326 set_irn_link(node, prev_copy);
330 /* fix data-flow (and reconstruct SSA if needed) */
331 foreach_out_edge(block, edge) {
332 ir_node *node = get_edge_src_irn(edge);
336 mode = get_irn_mode(node);
337 if (mode == mode_X || is_Cond(node) || is_Switch(node))
344 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
346 copy_node = (ir_node*)get_irn_link(node);
347 construct_ssa(block, node, copy_block, copy_node);
350 /* make sure new nodes are kept alive if old nodes were */
351 ir_graph *irg = get_irn_irg(block);
352 ir_node *end = get_irg_end(irg);
353 for (int i = 0, arity = get_End_n_keepalives(end); i < arity; ++i) {
354 ir_node *keep = get_End_keepalive(end, i);
355 if (get_irn_visited(keep) < env->visited_nr || is_Block(keep))
357 ir_node *copy = get_irn_link(keep);
358 add_End_keepalive(end, copy);
363 * returns whether the cmp evaluates to true or false, or can't be evaluated!
364 * 1: true, 0: false, -1: can't evaluate
366 * @param relation the compare mode of the Compare
367 * @param tv_left the left tarval
368 * @param tv_right the right tarval
370 static int eval_cmp_tv(ir_relation relation, ir_tarval *tv_left,
373 ir_relation cmp_result = tarval_cmp(tv_left, tv_right);
375 /* does the compare evaluate to true? */
376 if (cmp_result == ir_relation_false)
378 if ((cmp_result & relation) != 0)
385 * returns whether the cmp evaluates to true or false, or can't be evaluated!
386 * 1: true, 0: false, -1: can't evaluate
388 * @param env the environment
389 * @param cand the candidate node, either a Const or a Confirm
391 static int eval_cmp(jumpthreading_env_t *env, ir_node *cand)
393 if (is_Const(cand)) {
394 ir_tarval *tv_cand = get_Const_tarval(cand);
395 ir_tarval *tv_cmp = get_Const_tarval(env->cnst);
397 return eval_cmp_tv(env->relation, tv_cand, tv_cmp);
398 } else { /* a Confirm */
399 ir_tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->relation);
401 if (res == tarval_bad)
403 return res == tarval_b_true;
408 * Check for Const or Confirm with Const.
410 static int is_Const_or_Confirm(const ir_node *node)
412 if (is_Confirm(node))
413 node = get_Confirm_bound(node);
414 return is_Const(node);
418 * get the tarval of a Const or Confirm with
420 static ir_tarval *get_Const_or_Confirm_tarval(const ir_node *node)
422 if (is_Confirm(node)) {
423 if (get_Confirm_bound(node))
424 node = get_Confirm_bound(node);
426 return get_Const_tarval(node);
429 static ir_node *find_const_or_confirm(jumpthreading_env_t *env, ir_node *jump,
432 ir_node *block = get_nodes_block(jump);
434 if (irn_visited_else_mark(value))
437 if (is_Const_or_Confirm(value)) {
438 if (eval_cmp(env, value) <= 0)
443 "> Found jump threading candidate %+F->%+F\n",
444 block, env->true_block
447 /* adjust true_block to point directly towards our jump */
448 add_pred(env->true_block, jump);
450 split_critical_edge(env->true_block, 0);
452 /* we need a bigger visited nr when going back */
461 /* the Phi has to be in the same Block as the Jmp */
462 if (get_nodes_block(value) != block)
465 arity = get_irn_arity(value);
466 for (i = 0; i < arity; ++i) {
468 ir_node *phi_pred = get_Phi_pred(value, i);
469 ir_node *cfgpred = get_Block_cfgpred(block, i);
471 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
472 if (copy_block == NULL)
475 /* copy duplicated nodes in copy_block and fix SSA */
476 copy_and_fix(env, block, copy_block, i);
478 if (copy_block == get_nodes_block(cfgpred)) {
479 env->cnst_pred = block;
483 /* return now as we can't process more possibilities in 1 run */
491 static ir_node *find_candidate(jumpthreading_env_t *env, ir_node *jump,
494 ir_node *block = get_nodes_block(jump);
496 if (irn_visited_else_mark(value)) {
500 if (is_Const_or_Confirm(value)) {
501 ir_tarval *tv = get_Const_or_Confirm_tarval(value);
508 "> Found jump threading candidate %+F->%+F\n",
509 block, env->true_block
512 /* adjust true_block to point directly towards our jump */
513 add_pred(env->true_block, jump);
515 split_critical_edge(env->true_block, 0);
517 /* we need a bigger visited nr when going back */
525 /* the Phi has to be in the same Block as the Jmp */
526 if (get_nodes_block(value) != block)
529 arity = get_irn_arity(value);
530 for (i = 0; i < arity; ++i) {
532 ir_node *phi_pred = get_Phi_pred(value, i);
533 ir_node *cfgpred = get_Block_cfgpred(block, i);
535 copy_block = find_candidate(env, cfgpred, phi_pred);
536 if (copy_block == NULL)
539 /* copy duplicated nodes in copy_block and fix SSA */
540 copy_and_fix(env, block, copy_block, i);
542 if (copy_block == get_nodes_block(cfgpred)) {
543 env->cnst_pred = block;
547 /* return now as we can't process more possibilities in 1 run */
552 ir_node *cmp = value;
553 ir_node *left = get_Cmp_left(cmp);
554 ir_node *right = get_Cmp_right(cmp);
555 ir_relation relation = get_Cmp_relation(cmp);
557 /* we assume that the constant is on the right side, swap left/right
559 if (is_Const(left)) {
564 relation = get_inversed_relation(relation);
567 if (!is_Const(right))
570 if (get_nodes_block(left) != block)
573 /* negate condition when we're looking for the false block */
574 if (env->tv == tarval_b_false) {
575 relation = get_negated_relation(relation);
578 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
580 env->relation = relation;
583 return find_const_or_confirm(env, jump, left);
590 * Block-walker: searches for the following construct
592 * Const or Phi with constants
602 static void thread_jumps(ir_node* block, void* data)
604 jumpthreading_env_t env;
605 bool *changed = (bool*)data;
610 int selector_evaluated;
615 /* we do not deal with Phis, so restrict this to exactly one cfgpred */
616 if (get_Block_n_cfgpreds(block) != 1)
619 projx = get_Block_cfgpred(block, 0);
622 assert(get_irn_mode(projx) == mode_X);
624 cond = get_Proj_pred(projx);
625 /* TODO handle switch Conds */
629 /* handle cases that can be immediately evaluated */
630 selector = get_Cond_selector(cond);
631 selector_evaluated = -1;
632 if (is_Cmp(selector)) {
633 ir_node *left = get_Cmp_left(selector);
634 ir_node *right = get_Cmp_right(selector);
635 if (is_Const(left) && is_Const(right)) {
636 ir_relation relation = get_Cmp_relation(selector);
637 ir_tarval *tv_left = get_Const_tarval(left);
638 ir_tarval *tv_right = get_Const_tarval(right);
640 selector_evaluated = eval_cmp_tv(relation, tv_left, tv_right);
642 } else if (is_Const_or_Confirm(selector)) {
643 ir_tarval *tv = get_Const_or_Confirm_tarval(selector);
644 if (tv == tarval_b_true) {
645 selector_evaluated = 1;
647 assert(tv == tarval_b_false);
648 selector_evaluated = 0;
652 env.cnst_pred = NULL;
653 if (get_Proj_proj(projx) == pn_Cond_false) {
654 env.tv = tarval_b_false;
655 if (selector_evaluated >= 0)
656 selector_evaluated = !selector_evaluated;
658 env.tv = tarval_b_true;
661 if (selector_evaluated == 0) {
662 ir_graph *irg = get_irn_irg(block);
663 ir_node *bad = new_r_Bad(irg, mode_X);
664 exchange(projx, bad);
667 } else if (selector_evaluated == 1) {
668 dbg_info *dbgi = get_irn_dbg_info(selector);
669 ir_node *jmp = new_rd_Jmp(dbgi, get_nodes_block(projx));
670 DBG_OPT_JUMPTHREADING(projx, jmp);
671 exchange(projx, jmp);
676 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
677 env.true_block = block;
678 irg = get_irn_irg(block);
679 inc_irg_visited(irg);
680 env.visited_nr = get_irg_visited(irg);
682 copy_block = find_candidate(&env, projx, selector);
683 if (copy_block == NULL)
686 /* We might thread the condition block of an infinite loop,
687 * such that there is no path to End anymore. */
690 /* we have to remove the edge towards the pred as the pred now
691 * jumps into the true_block. We also have to shorten Phis
692 * in our block because of this */
693 badX = new_r_Bad(irg, mode_X);
694 cnst_pos = env.cnst_pos;
697 foreach_out_edge_safe(env.cnst_pred, edge) {
698 ir_node *node = get_edge_src_irn(edge);
701 ir_node *bad = new_r_Bad(irg, get_irn_mode(node));
702 set_Phi_pred(node, cnst_pos, bad);
706 set_Block_cfgpred(env.cnst_pred, cnst_pos, badX);
708 /* the graph is changed now */
712 void opt_jumpthreading(ir_graph* irg)
717 assure_irg_properties(irg,
718 IR_GRAPH_PROPERTY_NO_UNREACHABLE_CODE
719 | IR_GRAPH_PROPERTY_CONSISTENT_OUT_EDGES
720 | IR_GRAPH_PROPERTY_NO_CRITICAL_EDGES);
722 FIRM_DBG_REGISTER(dbg, "firm.opt.jumpthreading");
724 DB((dbg, LEVEL_1, "===> Performing jumpthreading on %+F\n", irg));
726 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
731 irg_block_walk_graph(irg, thread_jumps, NULL, &rerun);
735 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
737 confirm_irg_properties(irg,
738 changed ? IR_GRAPH_PROPERTIES_NONE : IR_GRAPH_PROPERTIES_ALL);
741 /* Creates an ir_graph pass for opt_jumpthreading. */
742 ir_graph_pass_t *opt_jumpthreading_pass(const char *name)
744 return def_graph_pass(name ? name : "jumpthreading", opt_jumpthreading);