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 Partial condition evaluation
24 * @author Christoph Mallon, Matthias Braun
31 #include "iroptimize.h"
43 #include "iredges_t.h"
47 #include "opt_confirms.h"
51 DEBUG_ONLY(static firm_dbg_module_t *dbg);
54 * Add the new predecessor x to node node, which is either a Block or a Phi
56 static void add_pred(ir_node* node, ir_node* x)
62 assert(is_Block(node) || is_Phi(node));
64 n = get_irn_arity(node);
65 NEW_ARR_A(ir_node*, ins, n + 1);
66 for (i = 0; i < n; i++)
67 ins[i] = get_irn_n(node, i);
69 set_irn_in(node, n + 1, ins);
72 static ir_node *ssa_second_def;
73 static ir_node *ssa_second_def_block;
75 static ir_node *search_def_and_create_phis(ir_node *block, ir_mode *mode,
84 /* This is needed because we create bads sometimes */
88 /* the other defs can't be marked for cases where a user of the original
89 * value is in the same block as the alternative definition.
90 * In this case we mustn't use the alternative definition.
91 * So we keep a flag that indicated wether we walked at least 1 block
92 * away and may use the alternative definition */
93 if (block == ssa_second_def_block && !first) {
94 return ssa_second_def;
97 /* already processed this block? */
98 if(irn_visited(block)) {
99 ir_node *value = (ir_node*) get_irn_link(block);
103 irg = get_irn_irg(block);
104 assert(block != get_irg_start_block(irg));
106 /* a Block with only 1 predecessor needs no Phi */
107 n_cfgpreds = get_Block_n_cfgpreds(block);
108 if(n_cfgpreds == 1) {
109 ir_node *pred_block = get_Block_cfgpred_block(block, 0);
110 ir_node *value = search_def_and_create_phis(pred_block, mode, 0);
112 set_irn_link(block, value);
113 mark_irn_visited(block);
117 /* create a new Phi */
118 NEW_ARR_A(ir_node*, in, n_cfgpreds);
119 for(i = 0; i < n_cfgpreds; ++i)
120 in[i] = new_Unknown(mode);
122 phi = new_r_Phi(irg, block, n_cfgpreds, in, mode);
123 set_irn_link(block, phi);
124 mark_irn_visited(block);
126 /* set Phi predecessors */
127 for(i = 0; i < n_cfgpreds; ++i) {
128 ir_node *pred_block = get_Block_cfgpred_block(block, i);
129 ir_node *pred_val = search_def_and_create_phis(pred_block, mode, 0);
131 set_irn_n(phi, i, pred_val);
138 * Given a set of values this function constructs SSA-form for the users of the
139 * first value (the users are determined through the out-edges of the value).
140 * Uses the irn_visited flags. Works without using the dominance tree.
142 static void construct_ssa(ir_node *orig_block, ir_node *orig_val,
143 ir_node *second_block, ir_node *second_val)
147 const ir_edge_t *edge;
148 const ir_edge_t *next;
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, next) {
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);
192 static void split_critical_edge(ir_node *block, int pos) {
193 ir_graph *irg = get_irn_irg(block);
198 in[0] = get_Block_cfgpred(block, pos);
199 new_block = new_r_Block(irg, 1, in);
200 new_jmp = new_r_Jmp(irg, new_block);
201 set_Block_cfgpred(block, pos, new_jmp);
204 typedef struct condeval_env_t {
206 ir_node *cmp; /**< The Compare node that might be partial evaluated */
207 pn_Cmp pnc; /**< The Compare mode of the Compare node. */
210 ir_visited_t visited_nr;
212 ir_node *cnst_pred; /**< the block before the constant */
213 int cnst_pos; /**< the pos to the constant block (needed to
214 kill that edge later) */
217 static ir_node *copy_and_fix_node(const condeval_env_t *env, ir_node *block,
218 ir_node *copy_block, int j, ir_node *node) {
222 /* we can evaluate Phis right now, all other nodes get copied */
224 copy = get_Phi_pred(node, j);
225 /* we might have to evaluate a Phi-cascade */
226 if(get_irn_visited(copy) >= env->visited_nr) {
227 copy = get_irn_link(copy);
230 copy = exact_copy(node);
231 set_nodes_block(copy, copy_block);
233 assert(get_irn_mode(copy) != mode_X);
235 arity = get_irn_arity(copy);
236 for(i = 0; i < arity; ++i) {
237 ir_node *pred = get_irn_n(copy, i);
240 if(get_nodes_block(pred) != block)
243 if(get_irn_visited(pred) >= env->visited_nr) {
244 new_pred = get_irn_link(pred);
246 new_pred = copy_and_fix_node(env, block, copy_block, j, pred);
248 DB((dbg, LEVEL_2, ">> Set Pred of %+F to %+F\n", copy, new_pred));
249 set_irn_n(copy, i, new_pred);
253 set_irn_link(node, copy);
254 set_irn_visited(node, env->visited_nr);
259 static void copy_and_fix(const condeval_env_t *env, ir_node *block,
260 ir_node *copy_block, int j) {
261 const ir_edge_t *edge;
263 /* Look at all nodes in the cond_block and copy them into pred */
264 foreach_out_edge(block, edge) {
265 ir_node *node = get_edge_src_irn(edge);
269 if (is_Block(node)) {
270 /* Block->Block edge, should be the MacroBlock edge */
271 assert(get_Block_MacroBlock(node) == block && "Block->Block edge found");
275 /* ignore control flow */
276 mode = get_irn_mode(node);
277 if (mode == mode_X || is_Cond(node))
280 /* we may not copy mode_b nodes, because this could produce Phi with
281 * mode_bs which can't be handled in all backends. Instead we duplicate
282 * the node and move it to its users */
283 if (mode == mode_b) {
284 const ir_edge_t *edge, *next;
288 assert(is_Proj(node));
290 pred = get_Proj_pred(node);
291 pn = get_Proj_proj(node);
293 foreach_out_edge_safe(node, edge, next) {
295 ir_node *user = get_edge_src_irn(edge);
296 int pos = get_edge_src_pos(edge);
297 ir_node *user_block = get_nodes_block(user);
299 if(user_block == block)
302 cmp_copy = exact_copy(pred);
303 set_nodes_block(cmp_copy, user_block);
304 copy = new_r_Proj(current_ir_graph, user_block, cmp_copy, mode_b, pn);
305 set_irn_n(user, pos, copy);
311 copy = copy_and_fix_node(env, block, copy_block, j, node);
313 /* we might hit values in blocks that have already been processed by a
314 * recursive find_phi_with_const() call */
315 assert(get_irn_visited(copy) <= env->visited_nr);
316 if(get_irn_visited(copy) >= env->visited_nr) {
317 ir_node *prev_copy = get_irn_link(copy);
318 if(prev_copy != NULL)
319 set_irn_link(node, prev_copy);
323 /* fix data-flow (and reconstruct SSA if needed) */
324 foreach_out_edge(block, edge) {
325 ir_node *node = get_edge_src_irn(edge);
329 if (is_Block(node)) {
330 /* Block->Block edge, should be the MacroBlock edge */
331 assert(get_Block_MacroBlock(node) == block && "Block->Block edge found");
335 mode = get_irn_mode(node);
336 if (mode == mode_X || is_Cond(node))
343 DB((dbg, LEVEL_2, ">> Fixing users of %+F\n", node));
345 copy_node = get_irn_link(node);
346 construct_ssa(block, node, copy_block, copy_node);
351 * returns whether the cmp evaluates to true or false, or can't be evaluated!
352 * 1: true, 0: false, -1: can't evaluate
354 * @param pnc the compare mode of the Compare
355 * @param tv_left the left tarval
356 * @param tv_right the right tarval
358 static int eval_cmp_tv(pn_Cmp pnc, tarval *tv_left, tarval *tv_right) {
359 pn_Cmp cmp_result = tarval_cmp(tv_left, tv_right);
361 /* does the compare evaluate to true? */
362 if (cmp_result == pn_Cmp_False)
364 if ((cmp_result & pnc) != cmp_result)
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 env the environment
375 * @param cand the candidate node, either a Const or a Confirm
377 static int eval_cmp(condeval_env_t *env, ir_node *cand) {
378 if (is_Const(cand)) {
379 tarval *tv_cand = get_Const_tarval(cand);
380 tarval *tv_cmp = get_Const_tarval(env->cnst);
382 return eval_cmp_tv(env->pnc, tv_cand, tv_cmp);
383 } else { /* a Confirm */
384 tarval *res = computed_value_Cmp_Confirm(env->cmp, cand, env->cnst, env->pnc);
386 if (res == tarval_bad)
388 return res == tarval_b_true;
393 * Check for Const or Confirm with Const.
395 static int is_Const_or_Confirm(const ir_node *node) {
396 if (is_Confirm(node))
397 node = get_Confirm_bound(node);
398 return is_Const(node);
402 * get the tarval of a Const or Confirm with
404 static tarval *get_Const_or_Confirm_tarval(const ir_node *node) {
405 if (is_Confirm(node)) {
406 if (get_Confirm_bound(node))
407 node = get_Confirm_bound(node);
409 return get_Const_tarval(node);
412 static ir_node *find_const_or_confirm(condeval_env_t *env, ir_node *jump, ir_node *value)
414 ir_node *block = get_nodes_block(jump);
416 if (irn_visited(value))
418 mark_irn_visited(value);
420 if (is_Const_or_Confirm(value)) {
421 if (eval_cmp(env, value) <= 0) {
427 "> Found condition evaluation candidate %+F->%+F\n",
428 env->true_block, block
431 /* adjust true_block to point directly towards our jump */
432 add_pred(env->true_block, jump);
434 split_critical_edge(env->true_block, 0);
436 /* we need a bigger visited nr when going back */
445 /* the Phi has to be in the same Block as the Jmp */
446 if(get_nodes_block(value) != block) {
450 arity = get_irn_arity(value);
451 for(i = 0; i < arity; ++i) {
453 ir_node *phi_pred = get_Phi_pred(value, i);
454 ir_node *cfgpred = get_Block_cfgpred(block, i);
456 copy_block = find_const_or_confirm(env, cfgpred, phi_pred);
457 if(copy_block == NULL)
460 /* copy duplicated nodes in copy_block and fix SSA */
461 copy_and_fix(env, block, copy_block, i);
463 if(copy_block == get_nodes_block(cfgpred)) {
464 env->cnst_pred = block;
468 /* return now as we can't process more possibilities in 1 run */
476 static ir_node *find_candidate(condeval_env_t *env, ir_node *jump,
479 ir_node *block = get_nodes_block(jump);
481 if(irn_visited(value)) {
484 mark_irn_visited(value);
486 if (is_Const_or_Confirm(value)) {
487 tarval *tv = get_Const_or_Confirm_tarval(value);
494 "> Found condition evaluation candidate %+F->%+F\n",
495 env->true_block, block
498 /* adjust true_block to point directly towards our jump */
499 add_pred(env->true_block, jump);
501 split_critical_edge(env->true_block, 0);
503 /* we need a bigger visited nr when going back */
511 /* the Phi has to be in the same Block as the Jmp */
512 if(get_nodes_block(value) != block)
515 arity = get_irn_arity(value);
516 for(i = 0; i < arity; ++i) {
518 ir_node *phi_pred = get_Phi_pred(value, i);
519 ir_node *cfgpred = get_Block_cfgpred(block, i);
521 copy_block = find_candidate(env, cfgpred, phi_pred);
522 if(copy_block == NULL)
525 /* copy duplicated nodes in copy_block and fix SSA */
526 copy_and_fix(env, block, copy_block, i);
528 if(copy_block == get_nodes_block(cfgpred)) {
529 env->cnst_pred = block;
533 /* return now as we can't process more possibilities in 1 run */
541 ir_node *cmp = get_Proj_pred(value);
545 left = get_Cmp_left(cmp);
546 right = get_Cmp_right(cmp);
547 pnc = get_Proj_proj(value);
549 /* we assume that the constant is on the right side, swap left/right
556 pnc = get_inversed_pnc(pnc);
562 if(get_nodes_block(left) != block) {
566 /* negate condition when we're looking for the false block */
567 if(env->tv == tarval_b_false) {
568 pnc = get_negated_pnc(pnc, get_irn_mode(right));
571 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
576 return find_const_or_confirm(env, jump, left);
583 * Block-walker: searches for the following construct
585 * Const or Phi with constants
595 static void cond_eval(ir_node* block, void* data)
603 int selector_evaluated;
604 const ir_edge_t *edge, *next;
608 if(get_Block_n_cfgpreds(block) != 1)
611 projx = get_Block_cfgpred(block, 0);
614 assert(get_irn_mode(projx) == mode_X);
616 cond = get_Proj_pred(projx);
620 selector = get_Cond_selector(cond);
621 /* TODO handle switch Conds */
622 if (get_irn_mode(selector) != mode_b)
625 /* handle cases that can be immediately evaluated */
626 selector_evaluated = -1;
627 if(is_Proj(selector)) {
628 ir_node *cmp = get_Proj_pred(selector);
630 ir_node *left = get_Cmp_left(cmp);
631 ir_node *right = get_Cmp_right(cmp);
632 if(is_Const(left) && is_Const(right)) {
633 int pnc = get_Proj_proj(selector);
634 tarval *tv_left = get_Const_tarval(left);
635 tarval *tv_right = get_Const_tarval(right);
637 selector_evaluated = eval_cmp_tv(pnc, tv_left, tv_right);
638 if(selector_evaluated < 0)
642 } else if (is_Const_or_Confirm(selector)) {
643 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) {
663 exchange(projx, bad);
666 } else if(selector_evaluated == 1) {
667 dbg_info *dbgi = get_irn_dbg_info(selector);
668 ir_node *jmp = new_rd_Jmp(dbgi, current_ir_graph, get_nodes_block(projx));
669 exchange(projx, jmp);
674 /* (recursively) look if a pred of a Phi is a constant or a Confirm */
675 env.true_block = block;
676 inc_irg_visited(current_ir_graph);
677 env.visited_nr = get_irg_visited(current_ir_graph);
679 copy_block = find_candidate(&env, projx, selector);
680 if (copy_block == NULL)
683 /* we have to remove the edge towards the pred as the pred now
684 * jumps into the true_block. We also have to shorten Phis
685 * in our block because of this */
687 cnst_pos = env.cnst_pos;
690 foreach_out_edge_safe(env.cnst_pred, edge, next) {
691 ir_node *node = get_edge_src_irn(edge);
694 set_Phi_pred(node, cnst_pos, bad);
697 set_Block_cfgpred(env.cnst_pred, cnst_pos, bad);
699 /* the graph is changed now */
703 void opt_cond_eval(ir_graph* irg)
707 FIRM_DBG_REGISTER(dbg, "firm.opt.condeval");
709 DB((dbg, LEVEL_1, "===> Performing condition evaluation on %+F\n", irg));
711 remove_critical_cf_edges(irg);
712 normalize_proj_nodes(irg);
715 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
720 irg_block_walk_graph(irg, cond_eval, NULL, &rerun);
724 ir_free_resources(irg, IR_RESOURCE_IRN_LINK | IR_RESOURCE_IRN_VISITED);
727 /* control flow changed, some blocks may become dead */
728 set_irg_outs_inconsistent(irg);
729 set_irg_doms_inconsistent(irg);
730 set_irg_extblk_inconsistent(irg);
731 set_irg_loopinfo_inconsistent(irg);
733 /* Dead code might be created. Optimize it away as it is dangerous
734 * to call optimize_df() an dead code. */