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 Copy minimization driver.
23 * @author Daniel Grund
26 * Main file for the optimization reducing the copies needed for:
28 * - Register-constrained nodes
29 * - Two-address code instructions
35 #include "execfreq_t.h"
37 #include "iredges_t.h"
42 #include "irprintf_t.h"
46 #include "raw_bitset.h"
51 #include "becopyopt_t.h"
52 #include "becopystat.h"
56 #include "beintlive_t.h"
62 #include "bestatevent.h"
66 #include "lc_opts_enum.h"
71 #define DUMP_ALL 2 * DUMP_APPEL - 1
73 #define COST_FUNC_FREQ 1
74 #define COST_FUNC_LOOP 2
75 #define COST_FUNC_ALL_ONE 3
78 * Flags for dumping the IFG.
81 CO_IFG_DUMP_COLORS = 1 << 0, /**< Dump the graph colored. */
82 CO_IFG_DUMP_LABELS = 1 << 1, /**< Dump node/edge labels. */
83 CO_IFG_DUMP_SHAPE = 1 << 2, /**< Give constrained nodes special shapes. */
84 CO_IFG_DUMP_CONSTR = 1 << 3, /**< Dump the node constraints in the label. */
87 static int co_get_costs_loop_depth(const ir_node *root, int pos);
88 static int co_get_costs_exec_freq(const ir_node *root, int pos);
89 static int co_get_costs_all_one(const ir_node *root, int pos);
91 static unsigned dump_flags = 0;
92 static unsigned style_flags = CO_IFG_DUMP_COLORS;
93 static int do_stats = 0;
94 static cost_fct_t cost_func = co_get_costs_exec_freq;
95 static int improve = 1;
97 static const lc_opt_enum_mask_items_t dump_items[] = {
98 { "before", DUMP_BEFORE },
99 { "after", DUMP_AFTER },
100 { "appel", DUMP_APPEL },
105 static const lc_opt_enum_mask_items_t style_items[] = {
106 { "color", CO_IFG_DUMP_COLORS },
107 { "labels", CO_IFG_DUMP_LABELS },
108 { "constr", CO_IFG_DUMP_CONSTR },
109 { "shape", CO_IFG_DUMP_SHAPE },
110 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
114 typedef int (*opt_funcptr)(void);
115 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
116 { "freq", (opt_funcptr) co_get_costs_exec_freq },
117 { "loop", (opt_funcptr) co_get_costs_loop_depth },
118 { "one", (opt_funcptr) co_get_costs_all_one },
122 static lc_opt_enum_mask_var_t dump_var = {
123 &dump_flags, dump_items
126 static lc_opt_enum_mask_var_t style_var = {
127 &style_flags, style_items
130 static lc_opt_enum_func_ptr_var_t cost_func_var = {
131 (opt_funcptr*) &cost_func, cost_func_items
134 static const lc_opt_table_entry_t options[] = {
135 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
136 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
137 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
138 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
139 LC_OPT_ENT_BOOL ("improve", "run heur1 before if algo can exploit start solutions", &improve),
143 static be_module_list_entry_t *copyopts = NULL;
144 static const co_algo_info *selected_copyopt = NULL;
146 void be_register_copyopt(const char *name, co_algo_info *copyopt)
148 if (selected_copyopt == NULL)
149 selected_copyopt = copyopt;
150 be_add_module_to_list(©opts, name, copyopt);
153 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt)
154 void be_init_copyopt(void)
156 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
157 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
158 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
159 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
161 lc_opt_add_table(co_grp, options);
162 be_add_module_list_opt(co_grp, "algo", "select copy optimization algo",
163 ©opts, (void**) &selected_copyopt);
166 static int void_algo(copy_opt_t *co)
172 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone)
173 void be_init_copynone(void)
175 static co_algo_info copyheur = {
179 be_register_copyopt("none", ©heur);
182 #undef QUICK_AND_DIRTY_HACK
184 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
187 return be_ifg_connected(env->ifg, a, b);
189 be_lv_t *lv = be_get_irg_liveness(env->irg);
190 return be_values_interfere(lv, a, b);
195 /******************************************************************************
198 | | __ ___ _ __ ___ _ __ __ _| |
199 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
200 | |__| | __/ | | | __/ | | (_| | |
201 \_____|\___|_| |_|\___|_| \__,_|_|
203 ******************************************************************************/
205 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
208 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
210 const char *s1, *s2, *s3;
214 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
216 co = XMALLOCZ(copy_opt_t);
217 co->cenv = chordal_env;
218 co->irg = chordal_env->irg;
219 co->cls = chordal_env->cls;
220 co->get_costs = get_costs;
223 s2 = get_entity_name(get_irg_entity(co->irg));
224 s3 = chordal_env->cls->name;
225 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
226 co->name = XMALLOCN(char, len);
227 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
232 void free_copy_opt(copy_opt_t *co)
239 * Checks if a node is optimizable, viz. has something to do with coalescing
240 * @param irn The irn to check
242 static int co_is_optimizable_root(ir_node *irn)
244 const arch_register_req_t *req;
246 if (arch_irn_is_ignore(irn))
249 if (is_Reg_Phi(irn) || is_Perm_Proj(irn))
252 req = arch_get_irn_register_req(irn);
253 if (is_2addr_code(req))
260 * Computes the costs of a copy according to loop depth
261 * @param pos the argument position of arg in the root arguments
262 * @return Must be >= 0 in all cases.
264 static int co_get_costs_loop_depth(const ir_node *root, int pos)
266 ir_node *block = get_nodes_block(root);
271 block = get_Block_cfgpred_block(block, pos);
273 loop = get_irn_loop(block);
275 int d = get_loop_depth(loop);
283 static ir_execfreq_int_factors factors;
286 * Computes the costs of a copy according to execution frequency
287 * @param pos the argument position of arg in the root arguments
288 * @return Must be >= 0 in all cases.
290 static int co_get_costs_exec_freq(const ir_node *root, int pos)
292 ir_node *root_bl = get_nodes_block(root);
294 = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
295 int res = get_block_execfreq_int(&factors, copy_bl);
297 /* don't allow values smaller than one. */
298 return res < 1 ? 1 : res;
302 * All costs equal 1. Using this will reduce the _number_ of copies.
303 * @param co The copy opt object.
304 * @return Must be >= 0 in all cases.
306 static int co_get_costs_all_one(const ir_node *root, int pos)
313 /******************************************************************************
314 ____ _ _ _ _ _ _____ _
315 / __ \ | | | | | | (_) | / ____| |
316 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
317 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
318 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
319 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
322 ******************************************************************************/
325 * Determines a maximum weighted independent set with respect to
326 * the interference and conflict edges of all nodes in a qnode.
328 static int ou_max_ind_set_costs(unit_t *ou)
330 be_chordal_env_t *chordal_env = ou->co->cenv;
331 ir_node **safe, **unsafe;
332 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
334 int curr_weight, best_weight = 0;
336 /* assign the nodes into two groups.
337 * safe: node has no interference, hence it is in every max stable set.
338 * unsafe: node has an interference
340 safe = ALLOCAN(ir_node*, ou->node_count - 1);
343 unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
344 unsafe_costs = ALLOCAN(int, ou->node_count - 1);
346 for (i=1; i<ou->node_count; ++i) {
348 for (o=1; o<ou->node_count; ++o) {
351 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
352 unsafe_costs[unsafe_count] = ou->costs[i];
353 unsafe[unsafe_count] = ou->nodes[i];
360 safe_costs += ou->costs[i];
361 safe[safe_count++] = ou->nodes[i];
366 /* now compute the best set out of the unsafe nodes*/
367 if (unsafe_count > MIS_HEUR_TRIGGER) {
368 bitset_t *best = bitset_alloca(unsafe_count);
369 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
370 for (i=0; i<unsafe_count; ++i) {
372 /* check if it is a stable set */
373 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
374 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
375 bitset_clear(best, i); /* clear the bit and try next one */
379 /* compute the weight */
380 bitset_foreach(best, pos)
381 best_weight += unsafe_costs[pos];
383 /* Exact Algorithm: Brute force */
384 curr = bitset_alloca(unsafe_count);
385 bitset_set_all(curr);
386 while (bitset_popcount(curr) != 0) {
387 /* check if curr is a stable set */
388 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
389 for (o=bitset_next_set(curr, i+1); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to qnode_max_ind_set(): NOT (curr, i) */
390 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
393 /* if we arrive here, we have a stable set */
394 /* compute the weight of the stable set*/
396 bitset_foreach(curr, pos)
397 curr_weight += unsafe_costs[pos];
400 if (curr_weight > best_weight) {
401 best_weight = curr_weight;
409 return safe_costs+best_weight;
412 static void co_collect_units(ir_node *irn, void *env)
414 const arch_register_req_t *req;
415 copy_opt_t *co = (copy_opt_t*)env;
418 if (get_irn_mode(irn) == mode_T)
420 req = arch_get_irn_register_req(irn);
421 if (req->cls != co->cls)
423 if (!co_is_optimizable_root(irn))
426 /* Init a new unit */
427 unit = XMALLOCZ(unit_t);
429 unit->node_count = 1;
430 INIT_LIST_HEAD(&unit->queue);
432 /* Phi with some/all of its arguments */
433 if (is_Reg_Phi(irn)) {
437 arity = get_irn_arity(irn);
438 unit->nodes = XMALLOCN(ir_node*, arity + 1);
439 unit->costs = XMALLOCN(int, arity + 1);
440 unit->nodes[0] = irn;
443 for (i=0; i<arity; ++i) {
445 ir_node *arg = get_irn_n(irn, i);
447 assert(arch_get_irn_reg_class(arg) == co->cls && "Argument not in same register class.");
450 if (nodes_interfere(co->cenv, irn, arg)) {
451 unit->inevitable_costs += co->get_costs(irn, i);
455 /* Else insert the argument of the phi to the members of this ou */
456 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
458 if (arch_irn_is_ignore(arg))
461 /* Check if arg has occurred at a prior position in the arg/list */
463 for (o=1; o<unit->node_count; ++o) {
464 if (unit->nodes[o] == arg) {
470 if (!arg_pos) { /* a new argument */
471 /* insert node, set costs */
472 unit->nodes[unit->node_count] = arg;
473 unit->costs[unit->node_count] = co->get_costs(irn, i);
475 } else { /* arg has occurred before in same phi */
476 /* increase costs for existing arg */
477 unit->costs[arg_pos] += co->get_costs(irn, i);
480 unit->nodes = XREALLOC(unit->nodes, ir_node*, unit->node_count);
481 unit->costs = XREALLOC(unit->costs, int, unit->node_count);
482 } else if (is_Perm_Proj(irn)) {
483 /* Proj of a perm with corresponding arg */
484 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
485 unit->nodes = XMALLOCN(ir_node*, 2);
486 unit->costs = XMALLOCN(int, 2);
487 unit->node_count = 2;
488 unit->nodes[0] = irn;
489 unit->nodes[1] = get_Perm_src(irn);
490 unit->costs[1] = co->get_costs(irn, -1);
492 /* Src == Tgt of a 2-addr-code instruction */
493 if (is_2addr_code(req)) {
494 const unsigned other = req->other_same;
498 for (i = 0; (1U << i) <= other; ++i) {
499 if (other & (1U << i)) {
500 ir_node *o = get_irn_n(skip_Proj(irn), i);
501 if (arch_irn_is_ignore(o))
503 if (nodes_interfere(co->cenv, irn, o))
512 unit->nodes = XMALLOCN(ir_node*, count);
513 unit->costs = XMALLOCN(int, count);
514 unit->node_count = count;
515 unit->nodes[k++] = irn;
517 for (i = 0; 1U << i <= other; ++i) {
518 if (other & (1U << i)) {
519 ir_node *o = get_irn_n(skip_Proj(irn), i);
520 if (!arch_irn_is_ignore(o) &&
521 !nodes_interfere(co->cenv, irn, o)) {
523 unit->costs[k] = co->get_costs(irn, -1);
530 assert(0 && "This is not an optimizable node!");
534 /* Insert the new unit at a position according to its costs */
535 if (unit->node_count > 1) {
537 struct list_head *tmp;
539 /* Determine the maximum costs this unit can cause: all_nodes_cost */
540 for (i=1; i<unit->node_count; ++i) {
541 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
542 unit->all_nodes_costs += unit->costs[i];
545 /* Determine the minimal costs this unit will cause: min_nodes_costs */
546 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
547 /* Insert the new ou according to its sort_key */
549 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
551 list_add(&unit->units, tmp);
557 #ifdef QUICK_AND_DIRTY_HACK
559 static int compare_ous(const void *k1, const void *k2)
561 const unit_t *u1 = *((const unit_t **) k1);
562 const unit_t *u2 = *((const unit_t **) k2);
563 int i, o, u1_has_constr, u2_has_constr;
564 arch_register_req_t req;
566 /* Units with constraints come first */
568 for (i=0; i<u1->node_count; ++i) {
569 arch_get_irn_register_req(&req, u1->nodes[i]);
570 if (arch_register_req_is(&req, limited)) {
577 for (i=0; i<u2->node_count; ++i) {
578 arch_get_irn_register_req(&req, u2->nodes[i]);
579 if (arch_register_req_is(&req, limited)) {
585 if (u1_has_constr != u2_has_constr)
586 return u2_has_constr - u1_has_constr;
588 /* Now check, whether the two units are connected */
590 for (i=0; i<u1->node_count; ++i)
591 for (o=0; o<u2->node_count; ++o)
592 if (u1->nodes[i] == u2->nodes[o])
596 /* After all, the sort key decides. Greater keys come first. */
597 return u2->sort_key - u1->sort_key;
602 * Sort the ou's according to constraints and their sort_key
604 static void co_sort_units(copy_opt_t *co)
606 int i, count = 0, costs;
609 /* get the number of ous, remove them form the list and fill the array */
610 list_for_each_entry(unit_t, ou, &co->units, units)
612 ous = ALLOCAN(unit_t, count);
614 costs = co_get_max_copy_costs(co);
617 list_for_each_entry(unit_t, ou, &co->units, units)
620 INIT_LIST_HEAD(&co->units);
622 assert(count == i && list_empty(&co->units));
624 for (i=0; i<count; ++i)
625 ir_printf("%+F\n", ous[i]->nodes[0]);
627 qsort(ous, count, sizeof(*ous), compare_ous);
630 for (i=0; i<count; ++i)
631 ir_printf("%+F\n", ous[i]->nodes[0]);
633 /* reinsert into list in correct order */
634 for (i=0; i<count; ++i)
635 list_add_tail(&ous[i]->units, &co->units);
637 assert(costs == co_get_max_copy_costs(co));
641 void co_build_ou_structure(copy_opt_t *co)
643 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
644 INIT_LIST_HEAD(&co->units);
645 irg_walk_graph(co->irg, co_collect_units, NULL, co);
646 #ifdef QUICK_AND_DIRTY_HACK
651 void co_free_ou_structure(copy_opt_t *co)
654 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
659 co->units.next = NULL;
662 /* co_solve_heuristic() is implemented in becopyheur.c */
664 int co_get_max_copy_costs(const copy_opt_t *co)
670 list_for_each_entry(unit_t, curr, &co->units, units) {
671 res += curr->inevitable_costs;
672 for (i=1; i<curr->node_count; ++i)
673 res += curr->costs[i];
678 int co_get_inevit_copy_costs(const copy_opt_t *co)
684 list_for_each_entry(unit_t, curr, &co->units, units)
685 res += curr->inevitable_costs;
689 int co_get_copy_costs(const copy_opt_t *co)
695 list_for_each_entry(unit_t, curr, &co->units, units) {
696 int root_col = get_irn_col(curr->nodes[0]);
697 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
698 res += curr->inevitable_costs;
699 for (i=1; i<curr->node_count; ++i) {
700 int arg_col = get_irn_col(curr->nodes[i]);
701 if (root_col != arg_col) {
702 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
703 res += curr->costs[i];
710 int co_get_lower_bound(const copy_opt_t *co)
716 list_for_each_entry(unit_t, curr, &co->units, units)
717 res += curr->inevitable_costs + curr->min_nodes_costs;
721 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
723 bitset_t *seen = bitset_malloc(get_irg_last_idx(co->irg));
725 memset(stat, 0, sizeof(stat[0]));
727 /* count affinity edges. */
728 co_gs_foreach_aff_node(co, an) {
729 stat->aff_nodes += 1;
730 bitset_set(seen, get_irn_idx(an->irn));
731 co_gs_foreach_neighb(an, neigh) {
732 if (!bitset_is_set(seen, get_irn_idx(neigh->irn))) {
733 stat->aff_edges += 1;
734 stat->max_costs += neigh->costs;
736 if (get_irn_col(an->irn) != get_irn_col(neigh->irn)) {
737 stat->costs += neigh->costs;
738 stat->unsatisfied_edges += 1;
741 if (nodes_interfere(co->cenv, an->irn, neigh->irn)) {
743 stat->inevit_costs += neigh->costs;
753 /******************************************************************************
755 / ____| | | / ____| |
756 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
757 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
758 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
759 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
762 ******************************************************************************/
764 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size)
766 const affinity_node_t *n1 = (const affinity_node_t*)k1;
767 const affinity_node_t *n2 = (const affinity_node_t*)k2;
770 return (n1->irn != n2->irn);
773 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
775 affinity_node_t new_node, *node;
781 new_node.neighbours = NULL;
782 node = set_insert(affinity_node_t, co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
784 for (nbr = node->neighbours; nbr; nbr = nbr->next)
785 if (nbr->irn == n2) {
790 /* if we did not find n2 in n1's neighbourhood insert it */
792 nbr = OALLOC(&co->obst, neighb_t);
795 nbr->next = node->neighbours;
797 node->neighbours = nbr;
801 /* now nbr points to n1's neighbour-entry of n2 */
805 static inline void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
807 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
808 add_edge(co, n1, n2, costs);
809 add_edge(co, n2, n1, costs);
813 static void build_graph_walker(ir_node *irn, void *env)
815 const arch_register_req_t *req;
816 copy_opt_t *co = (copy_opt_t*)env;
819 if (get_irn_mode(irn) == mode_T)
821 req = arch_get_irn_register_req(irn);
822 if (req->cls != co->cls || arch_irn_is_ignore(irn))
825 if (is_Reg_Phi(irn)) { /* Phis */
826 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
827 ir_node *arg = get_irn_n(irn, pos);
828 add_edges(co, irn, arg, co->get_costs(irn, pos));
830 } else if (is_Perm_Proj(irn)) { /* Perms */
831 ir_node *arg = get_Perm_src(irn);
832 add_edges(co, irn, arg, co->get_costs(irn, -1));
833 } else { /* 2-address code */
834 if (is_2addr_code(req)) {
835 const unsigned other = req->other_same;
838 for (i = 0; 1U << i <= other; ++i) {
839 if (other & (1U << i)) {
840 ir_node *other = get_irn_n(skip_Proj(irn), i);
841 if (!arch_irn_is_ignore(other))
842 add_edges(co, irn, other, co->get_costs(irn, -1));
849 void co_build_graph_structure(copy_opt_t *co)
851 obstack_init(&co->obst);
852 co->nodes = new_set(compare_affinity_node_t, 32);
854 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
857 void co_free_graph_structure(copy_opt_t *co)
862 obstack_free(&co->obst, NULL);
866 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn)
868 affinity_node_t new_node, *n;
873 n = set_find(affinity_node_t, co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
875 return (n->degree > 0);
880 static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b)
882 ir_node *nodes[] = { a, b };
883 bitset_t *constr[] = { NULL, NULL };
886 constr[0] = bitset_alloca(co->cls->n_regs);
887 constr[1] = bitset_alloca(co->cls->n_regs);
889 for (j = 0; j < 2; ++j) {
890 const arch_register_req_t *req = arch_get_irn_register_req(nodes[j]);
891 if (arch_register_req_is(req, limited))
892 rbitset_copy_to_bitset(req->limited, constr[j]);
894 bitset_set_all(constr[j]);
898 return !bitset_intersect(constr[0], constr[1]);
902 * Dump the interference graph according to the Appel/George coalescing contest file format.
903 * See: http://www.cs.princeton.edu/~appel/coalesce/format.html
904 * @note Requires graph structure.
905 * @param co The copy opt object.
906 * @param f A file to dump to.
908 static void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
910 be_ifg_t *ifg = co->cenv->ifg;
911 int *color_map = ALLOCAN(int, co->cls->n_regs);
912 int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
913 ir_graph *irg = co->irg;
914 be_irg_t *birg = be_birg_from_irg(irg);
918 neighbours_iter_t nit;
923 for (i = 0; i < co->cls->n_regs; ++i) {
924 const arch_register_t *reg = &co->cls->regs[i];
925 if (rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
926 color_map[i] = n_regs++;
933 * n contains the first node number.
934 * the values below n are the pre-colored register nodes
938 be_ifg_foreach_node(ifg, &it, irn) {
939 if (arch_irn_is_ignore(irn))
941 node_map[get_irn_idx(irn)] = n++;
944 fprintf(f, "%d %d\n", n, n_regs);
946 be_ifg_foreach_node(ifg, &it, irn) {
947 if (!arch_irn_is_ignore(irn)) {
948 int idx = node_map[get_irn_idx(irn)];
949 affinity_node_t *a = get_affinity_info(co, irn);
950 const arch_register_req_t *req = arch_get_irn_register_req(irn);
953 if (arch_register_req_is(req, limited)) {
954 for (i = 0; i < co->cls->n_regs; ++i) {
955 if (!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
956 fprintf(f, "%d %d -1\n", color_map[i], idx);
960 be_ifg_foreach_neighbour(ifg, &nit, irn, adj) {
961 if (!arch_irn_is_ignore(adj) &&
962 !co_dump_appel_disjoint_constraints(co, irn, adj)) {
963 int adj_idx = node_map[get_irn_idx(adj)];
965 fprintf(f, "%d %d -1\n", idx, adj_idx);
970 co_gs_foreach_neighb(a, n) {
971 if (!arch_irn_is_ignore(n->irn)) {
972 int n_idx = node_map[get_irn_idx(n->irn)];
974 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
984 static FILE *my_open(const be_chordal_env_t *env, const char *prefix,
991 const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
993 n = strlen(cup_name);
994 tu_name = XMALLOCN(char, n + 1);
995 strcpy(tu_name, cup_name);
996 for (i = 0; i < n; ++i)
997 if (tu_name[i] == '.')
1001 ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
1003 result = fopen(buf, "wt");
1004 if (result == NULL) {
1005 panic("Couldn't open '%s' for writing.", buf);
1011 void co_driver(be_chordal_env_t *cenv)
1013 ir_timer_t *timer = ir_timer_new();
1014 co_complete_stats_t before, after;
1016 int was_optimal = 0;
1018 assert(selected_copyopt);
1020 /* skip copymin if algo is 'none' */
1021 if (selected_copyopt->copyopt == void_algo)
1024 be_assure_live_chk(cenv->irg);
1026 co = new_copy_opt(cenv, cost_func);
1027 co_build_ou_structure(co);
1028 co_build_graph_structure(co);
1030 co_complete_stats(co, &before);
1032 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1033 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1034 be_stat_ev_ull("co_max_costs", before.max_costs);
1035 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1036 be_stat_ev_ull("co_aff_int", before.aff_int);
1038 be_stat_ev_ull("co_init_costs", before.costs);
1039 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1041 if (dump_flags & DUMP_BEFORE) {
1042 FILE *f = my_open(cenv, "", "-before.vcg");
1043 be_dump_ifg_co(f, co, style_flags & CO_IFG_DUMP_LABELS, style_flags & CO_IFG_DUMP_COLORS);
1047 /* if the algo can improve results, provide an initial solution with heur1 */
1048 if (improve && selected_copyopt->can_improve_existing) {
1049 co_complete_stats_t stats;
1051 /* produce a heuristic solution */
1052 co_solve_heuristic(co);
1054 /* do the stats and provide the current costs */
1055 co_complete_stats(co, &stats);
1056 be_stat_ev_ull("co_prepare_costs", stats.costs);
1059 /* perform actual copy minimization */
1060 ir_timer_reset_and_start(timer);
1061 was_optimal = selected_copyopt->copyopt(co);
1062 ir_timer_stop(timer);
1064 be_stat_ev("co_time", ir_timer_elapsed_msec(timer));
1065 be_stat_ev_ull("co_optimal", was_optimal);
1066 ir_timer_free(timer);
1068 if (dump_flags & DUMP_AFTER) {
1069 FILE *f = my_open(cenv, "", "-after.vcg");
1070 be_dump_ifg_co(f, co, style_flags & CO_IFG_DUMP_LABELS, style_flags & CO_IFG_DUMP_COLORS);
1074 co_complete_stats(co, &after);
1077 unsigned long long optimizable_costs = after.max_costs - after.inevit_costs;
1078 unsigned long long evitable = after.costs - after.inevit_costs;
1080 ir_printf("%30F ", cenv->irg);
1081 printf("%10s %10llu%10llu%10llu", cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1083 if (optimizable_costs > 0)
1084 printf("%10llu %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1086 printf("%10llu %5s\n", after.costs, "-");
1089 /* Dump the interference graph in Appel's format. */
1090 if (dump_flags & DUMP_APPEL) {
1091 FILE *f = my_open(cenv, "", ".apl");
1092 fprintf(f, "# %llu %llu\n", after.costs, after.unsatisfied_edges);
1093 co_dump_appel_graph(co, f);
1097 be_stat_ev_ull("co_after_costs", after.costs);
1098 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1100 co_free_graph_structure(co);
1101 co_free_ou_structure(co);