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
27 * Main file for the optimization reducing the copies needed for:
29 * - Register-constrained nodes
30 * - Two-address code instructions
38 #include "raw_bitset.h"
44 #include "iredges_t.h"
46 #include "irphase_t.h"
47 #include "irprintf_t.h"
54 #include "beintlive_t.h"
55 #include "becopyopt_t.h"
56 #include "becopystat.h"
60 #include "bestatevent.h"
65 #include "lc_opts_enum.h"
70 #define DUMP_ALL 2 * DUMP_APPEL - 1
72 #define COST_FUNC_FREQ 1
73 #define COST_FUNC_LOOP 2
74 #define COST_FUNC_ALL_ONE 3
76 static unsigned dump_flags = 0;
77 static unsigned style_flags = 0;
78 static unsigned do_stats = 0;
79 static cost_fct_t cost_func = co_get_costs_exec_freq;
80 static int improve = 1;
82 static const lc_opt_enum_mask_items_t dump_items[] = {
83 { "before", DUMP_BEFORE },
84 { "after", DUMP_AFTER },
85 { "appel", DUMP_APPEL },
90 static const lc_opt_enum_mask_items_t style_items[] = {
91 { "color", CO_IFG_DUMP_COLORS },
92 { "labels", CO_IFG_DUMP_LABELS },
93 { "constr", CO_IFG_DUMP_CONSTR },
94 { "shape", CO_IFG_DUMP_SHAPE },
95 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
99 typedef int (*opt_funcptr)(void);
101 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
102 { "freq", (opt_funcptr) co_get_costs_exec_freq },
103 { "loop", (opt_funcptr) co_get_costs_loop_depth },
104 { "one", (opt_funcptr) co_get_costs_all_one },
108 static lc_opt_enum_mask_var_t dump_var = {
109 &dump_flags, dump_items
112 static lc_opt_enum_mask_var_t style_var = {
113 &style_flags, style_items
116 static lc_opt_enum_func_ptr_var_t cost_func_var = {
117 (opt_funcptr*) &cost_func, cost_func_items
120 static const lc_opt_table_entry_t options[] = {
121 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
122 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
123 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
124 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
125 LC_OPT_ENT_BOOL ("improve", "run heur1 before if algo can exploit start solutions", &improve),
129 static be_module_list_entry_t *copyopts = NULL;
130 static const co_algo_info *selected_copyopt = NULL;
132 void be_register_copyopt(const char *name, co_algo_info *copyopt)
134 if (selected_copyopt == NULL)
135 selected_copyopt = copyopt;
136 be_add_module_to_list(©opts, name, copyopt);
139 void be_init_copyopt(void)
141 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
142 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
143 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
144 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
146 lc_opt_add_table(co_grp, options);
147 be_add_module_list_opt(co_grp, "algo", "select copy optimization algo",
148 ©opts, (void**) &selected_copyopt);
151 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt);
153 static int void_algo(copy_opt_t *co)
159 void be_init_copynone(void)
161 static co_algo_info copyheur = {
165 be_register_copyopt("none", ©heur);
168 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone);
170 #undef QUICK_AND_DIRTY_HACK
172 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
175 return be_ifg_connected(env->ifg, a, b);
177 return be_values_interfere(env->birg->lv, a, b);
181 /******************************************************************************
184 | | __ ___ _ __ ___ _ __ __ _| |
185 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
186 | |__| | __/ | | | __/ | | (_| | |
187 \_____|\___|_| |_|\___|_| \__,_|_|
189 ******************************************************************************/
191 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
194 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
196 const char *s1, *s2, *s3;
200 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
202 co = XMALLOCZ(copy_opt_t);
203 co->cenv = chordal_env;
204 co->irg = chordal_env->irg;
205 co->cls = chordal_env->cls;
206 co->get_costs = get_costs;
209 s2 = get_entity_name(get_irg_entity(co->irg));
210 s3 = chordal_env->cls->name;
211 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
212 co->name = XMALLOCN(char, len);
213 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
218 void free_copy_opt(copy_opt_t *co) {
224 * Checks if a node is optimizable, viz. has something to do with coalescing
225 * @param irn The irn to check
227 static int co_is_optimizable_root(ir_node *irn)
229 const arch_register_req_t *req;
230 const arch_register_t *reg;
232 if (arch_irn_is_ignore(irn))
235 reg = arch_get_irn_register(irn);
236 if (arch_register_type_is(reg, ignore))
239 if (is_Reg_Phi(irn) || is_Perm_Proj(irn))
242 req = arch_get_register_req_out(irn);
243 if (is_2addr_code(req))
249 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
252 ir_node *root_block = get_nodes_block(root);
257 /* for phis the copies are placed in the corresponding pred-block */
258 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
260 /* a perm places the copy in the same block as it resides */
261 loop = get_irn_loop(root_block);
264 int d = get_loop_depth(loop);
270 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
272 ir_node *root_bl = get_nodes_block(root);
273 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
275 res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
277 /* don't allow values smaller than one. */
278 return res < 1 ? 1 : res;
282 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
290 /******************************************************************************
291 ____ _ _ _ _ _ _____ _
292 / __ \ | | | | | | (_) | / ____| |
293 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
294 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
295 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
296 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
299 ******************************************************************************/
302 * Determines a maximum weighted independent set with respect to
303 * the interference and conflict edges of all nodes in a qnode.
305 static int ou_max_ind_set_costs(unit_t *ou) {
306 be_chordal_env_t *chordal_env = ou->co->cenv;
307 ir_node **safe, **unsafe;
308 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
311 int max, curr_weight, best_weight = 0;
313 /* assign the nodes into two groups.
314 * safe: node has no interference, hence it is in every max stable set.
315 * unsafe: node has an interference
317 safe = ALLOCAN(ir_node*, ou->node_count - 1);
320 unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
321 unsafe_costs = ALLOCAN(int, ou->node_count - 1);
323 for(i=1; i<ou->node_count; ++i) {
325 for(o=1; o<ou->node_count; ++o) {
328 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
329 unsafe_costs[unsafe_count] = ou->costs[i];
330 unsafe[unsafe_count] = ou->nodes[i];
337 safe_costs += ou->costs[i];
338 safe[safe_count++] = ou->nodes[i];
343 /* now compute the best set out of the unsafe nodes*/
344 if (unsafe_count > MIS_HEUR_TRIGGER) {
345 bitset_t *best = bitset_alloca(unsafe_count);
346 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
347 for (i=0; i<unsafe_count; ++i) {
349 /* check if it is a stable set */
350 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
351 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
352 bitset_clear(best, i); /* clear the bit and try next one */
356 /* compute the weight */
357 bitset_foreach(best, pos)
358 best_weight += unsafe_costs[pos];
360 /* Exact Algorithm: Brute force */
361 curr = bitset_alloca(unsafe_count);
362 bitset_set_all(curr);
363 while ((max = bitset_popcnt(curr)) != 0) {
364 /* check if curr is a stable set */
365 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
366 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) */
367 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
370 /* if we arrive here, we have a stable set */
371 /* compute the weigth of the stable set*/
373 bitset_foreach(curr, pos)
374 curr_weight += unsafe_costs[pos];
377 if (curr_weight > best_weight) {
378 best_weight = curr_weight;
386 return safe_costs+best_weight;
389 static void co_collect_units(ir_node *irn, void *env)
391 const arch_register_req_t *req = arch_get_register_req_out(irn);
392 copy_opt_t *co = env;
395 if (req->cls != co->cls)
397 if (!co_is_optimizable_root(irn))
400 /* Init a new unit */
401 unit = XMALLOCZ(unit_t);
403 unit->node_count = 1;
404 INIT_LIST_HEAD(&unit->queue);
406 /* Phi with some/all of its arguments */
407 if (is_Reg_Phi(irn)) {
411 arity = get_irn_arity(irn);
412 unit->nodes = XMALLOCN(ir_node*, arity + 1);
413 unit->costs = XMALLOCN(int, arity + 1);
414 unit->nodes[0] = irn;
417 for (i=0; i<arity; ++i) {
419 ir_node *arg = get_irn_n(irn, i);
421 assert(arch_get_irn_reg_class_out(arg) == co->cls && "Argument not in same register class.");
424 if (nodes_interfere(co->cenv, irn, arg)) {
425 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
429 /* Else insert the argument of the phi to the members of this ou */
430 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
432 if (arch_irn_is_ignore(arg))
435 /* Check if arg has occurred at a prior position in the arg/list */
437 for (o=1; o<unit->node_count; ++o) {
438 if (unit->nodes[o] == arg) {
444 if (!arg_pos) { /* a new argument */
445 /* insert node, set costs */
446 unit->nodes[unit->node_count] = arg;
447 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
449 } else { /* arg has occurred before in same phi */
450 /* increase costs for existing arg */
451 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
454 unit->nodes = XREALLOC(unit->nodes, ir_node*, unit->node_count);
455 unit->costs = XREALLOC(unit->costs, int, unit->node_count);
456 } else if (is_Perm_Proj(irn)) {
457 /* Proj of a perm with corresponding arg */
458 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
459 unit->nodes = XMALLOCN(ir_node*, 2);
460 unit->costs = XMALLOCN(int, 2);
461 unit->node_count = 2;
462 unit->nodes[0] = irn;
463 unit->nodes[1] = get_Perm_src(irn);
464 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
466 /* Src == Tgt of a 2-addr-code instruction */
467 if (is_2addr_code(req)) {
468 const unsigned other = req->other_same;
472 for (i = 0; (1U << i) <= other; ++i) {
473 if (other & (1U << i)) {
474 ir_node *o = get_irn_n(skip_Proj(irn), i);
475 if (arch_irn_is_ignore(o))
477 if (nodes_interfere(co->cenv, irn, o))
486 unit->nodes = XMALLOCN(ir_node*, count);
487 unit->costs = XMALLOCN(int, count);
488 unit->node_count = count;
489 unit->nodes[k++] = irn;
491 for (i = 0; 1U << i <= other; ++i) {
492 if (other & (1U << i)) {
493 ir_node *o = get_irn_n(skip_Proj(irn), i);
494 if (!arch_irn_is_ignore(o) &&
495 !nodes_interfere(co->cenv, irn, o)) {
497 unit->costs[k] = co->get_costs(co, irn, o, -1);
504 assert(0 && "This is not an optimizable node!");
508 /* Insert the new unit at a position according to its costs */
509 if (unit->node_count > 1) {
511 struct list_head *tmp;
513 /* Determine the maximum costs this unit can cause: all_nodes_cost */
514 for(i=1; i<unit->node_count; ++i) {
515 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
516 unit->all_nodes_costs += unit->costs[i];
519 /* Determine the minimal costs this unit will cause: min_nodes_costs */
520 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
521 /* Insert the new ou according to its sort_key */
523 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
525 list_add(&unit->units, tmp);
531 #ifdef QUICK_AND_DIRTY_HACK
533 static int compare_ous(const void *k1, const void *k2) {
534 const unit_t *u1 = *((const unit_t **) k1);
535 const unit_t *u2 = *((const unit_t **) k2);
536 int i, o, u1_has_constr, u2_has_constr;
537 arch_register_req_t req;
539 /* Units with constraints come first */
541 for (i=0; i<u1->node_count; ++i) {
542 arch_get_register_req_out(&req, u1->nodes[i]);
543 if (arch_register_req_is(&req, limited)) {
550 for (i=0; i<u2->node_count; ++i) {
551 arch_get_register_req_out(&req, u2->nodes[i]);
552 if (arch_register_req_is(&req, limited)) {
558 if (u1_has_constr != u2_has_constr)
559 return u2_has_constr - u1_has_constr;
561 /* Now check, whether the two units are connected */
563 for (i=0; i<u1->node_count; ++i)
564 for (o=0; o<u2->node_count; ++o)
565 if (u1->nodes[i] == u2->nodes[o])
569 /* After all, the sort key decides. Greater keys come first. */
570 return u2->sort_key - u1->sort_key;
575 * Sort the ou's according to constraints and their sort_key
577 static void co_sort_units(copy_opt_t *co) {
578 int i, count = 0, costs;
581 /* get the number of ous, remove them form the list and fill the array */
582 list_for_each_entry(unit_t, ou, &co->units, units)
584 ous = ALLOCAN(unit_t, count);
586 costs = co_get_max_copy_costs(co);
589 list_for_each_entry(unit_t, ou, &co->units, units)
592 INIT_LIST_HEAD(&co->units);
594 assert(count == i && list_empty(&co->units));
596 for (i=0; i<count; ++i)
597 ir_printf("%+F\n", ous[i]->nodes[0]);
599 qsort(ous, count, sizeof(*ous), compare_ous);
602 for (i=0; i<count; ++i)
603 ir_printf("%+F\n", ous[i]->nodes[0]);
605 /* reinsert into list in correct order */
606 for (i=0; i<count; ++i)
607 list_add_tail(&ous[i]->units, &co->units);
609 assert(costs == co_get_max_copy_costs(co));
613 void co_build_ou_structure(copy_opt_t *co) {
614 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
615 INIT_LIST_HEAD(&co->units);
616 irg_walk_graph(co->irg, co_collect_units, NULL, co);
617 #ifdef QUICK_AND_DIRTY_HACK
622 void co_free_ou_structure(copy_opt_t *co) {
625 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
630 co->units.next = NULL;
633 /* co_solve_heuristic() is implemented in becopyheur.c */
635 int co_get_max_copy_costs(const copy_opt_t *co) {
641 list_for_each_entry(unit_t, curr, &co->units, units) {
642 res += curr->inevitable_costs;
643 for (i=1; i<curr->node_count; ++i)
644 res += curr->costs[i];
649 int co_get_inevit_copy_costs(const copy_opt_t *co) {
655 list_for_each_entry(unit_t, curr, &co->units, units)
656 res += curr->inevitable_costs;
660 int co_get_copy_costs(const copy_opt_t *co) {
666 list_for_each_entry(unit_t, curr, &co->units, units) {
667 int root_col = get_irn_col(curr->nodes[0]);
668 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
669 res += curr->inevitable_costs;
670 for (i=1; i<curr->node_count; ++i) {
671 int arg_col = get_irn_col(curr->nodes[i]);
672 if (root_col != arg_col) {
673 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
674 res += curr->costs[i];
681 int co_get_lower_bound(const copy_opt_t *co) {
687 list_for_each_entry(unit_t, curr, &co->units, units)
688 res += curr->inevitable_costs + curr->min_nodes_costs;
692 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
694 bitset_t *seen = bitset_irg_malloc(co->irg);
697 memset(stat, 0, sizeof(stat[0]));
699 /* count affinity edges. */
700 co_gs_foreach_aff_node(co, an) {
702 stat->aff_nodes += 1;
703 bitset_add_irn(seen, an->irn);
704 co_gs_foreach_neighb(an, neigh) {
705 if(!bitset_contains_irn(seen, neigh->irn)) {
706 stat->aff_edges += 1;
707 stat->max_costs += neigh->costs;
709 if (get_irn_col(an->irn) != get_irn_col(neigh->irn)) {
710 stat->costs += neigh->costs;
711 stat->unsatisfied_edges += 1;
714 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
716 stat->inevit_costs += neigh->costs;
726 /******************************************************************************
728 / ____| | | / ____| |
729 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
730 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
731 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
732 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
735 ******************************************************************************/
737 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
738 const affinity_node_t *n1 = k1;
739 const affinity_node_t *n2 = k2;
742 return (n1->irn != n2->irn);
745 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
746 affinity_node_t new_node, *node;
752 new_node.neighbours = NULL;
753 node = set_insert(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
755 for (nbr = node->neighbours; nbr; nbr = nbr->next)
756 if (nbr->irn == n2) {
761 /* if we did not find n2 in n1's neighbourhood insert it */
763 nbr = OALLOC(&co->obst, neighb_t);
766 nbr->next = node->neighbours;
768 node->neighbours = nbr;
772 /* now nbr points to n1's neighbour-entry of n2 */
776 static inline void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
777 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
778 add_edge(co, n1, n2, costs);
779 add_edge(co, n2, n1, costs);
783 static void build_graph_walker(ir_node *irn, void *env) {
784 const arch_register_req_t *req = arch_get_register_req_out(irn);
785 copy_opt_t *co = env;
787 const arch_register_t *reg;
789 if (req->cls != co->cls || arch_irn_is_ignore(irn))
792 reg = arch_get_irn_register(irn);
793 if (arch_register_type_is(reg, ignore))
796 if (is_Reg_Phi(irn)) { /* Phis */
797 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
798 ir_node *arg = get_irn_n(irn, pos);
799 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
801 } else if (is_Perm_Proj(irn)) { /* Perms */
802 ir_node *arg = get_Perm_src(irn);
803 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
804 } else { /* 2-address code */
805 if (is_2addr_code(req)) {
806 const unsigned other = req->other_same;
809 for (i = 0; 1U << i <= other; ++i) {
810 if (other & (1U << i)) {
811 ir_node *other = get_irn_n(skip_Proj(irn), i);
812 if (!arch_irn_is_ignore(other))
813 add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
820 void co_build_graph_structure(copy_opt_t *co) {
821 obstack_init(&co->obst);
822 co->nodes = new_set(compare_affinity_node_t, 32);
824 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
827 void co_free_graph_structure(copy_opt_t *co) {
831 obstack_free(&co->obst, NULL);
835 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
837 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
838 affinity_node_t new_node, *n;
843 n = set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
845 return (n->degree > 0);
850 static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b)
852 ir_node *nodes[] = { a, b };
853 bitset_t *constr[] = { NULL, NULL };
856 constr[0] = bitset_alloca(co->cls->n_regs);
857 constr[1] = bitset_alloca(co->cls->n_regs);
859 for (j = 0; j < 2; ++j) {
860 const arch_register_req_t *req = arch_get_register_req_out(nodes[j]);
861 if(arch_register_req_is(req, limited))
862 rbitset_copy_to_bitset(req->limited, constr[j]);
864 bitset_set_all(constr[j]);
868 return !bitset_intersect(constr[0], constr[1]);
871 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
873 be_ifg_t *ifg = co->cenv->ifg;
874 int *color_map = ALLOCAN(int, co->cls->n_regs);
875 int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
883 for(i = 0; i < co->cls->n_regs; ++i) {
884 const arch_register_t *reg = &co->cls->regs[i];
885 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
889 * n contains the first node number.
890 * the values below n are the pre-colored register nodes
893 it = be_ifg_nodes_iter_alloca(ifg);
894 nit = be_ifg_neighbours_iter_alloca(ifg);
897 be_ifg_foreach_node(ifg, it, irn) {
898 if (arch_irn_is_ignore(irn))
900 node_map[get_irn_idx(irn)] = n++;
903 fprintf(f, "%d %d\n", n, n_regs);
905 be_ifg_foreach_node(ifg, it, irn) {
906 if (!arch_irn_is_ignore(irn)) {
907 int idx = node_map[get_irn_idx(irn)];
908 affinity_node_t *a = get_affinity_info(co, irn);
909 const arch_register_req_t *req = arch_get_register_req_out(irn);
912 if(arch_register_req_is(req, limited)) {
913 for(i = 0; i < co->cls->n_regs; ++i) {
914 if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
915 fprintf(f, "%d %d -1\n", color_map[i], idx);
919 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
920 if (!arch_irn_is_ignore(adj) &&
921 !co_dump_appel_disjoint_constraints(co, irn, adj)) {
922 int adj_idx = node_map[get_irn_idx(adj)];
924 fprintf(f, "%d %d -1\n", idx, adj_idx);
931 co_gs_foreach_neighb(a, n) {
932 if (!arch_irn_is_ignore(n->irn)) {
933 int n_idx = node_map[get_irn_idx(n->irn)];
935 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
946 ___ _____ ____ ____ ___ _____ ____ _
947 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
948 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
949 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
950 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
954 static const char *get_dot_color_name(size_t col)
956 static const char *names[] = {
990 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
993 typedef struct _co_ifg_dump_t {
994 const copy_opt_t *co;
998 static void ifg_dump_graph_attr(FILE *f, void *self)
1001 fprintf(f, "overlap=scale");
1004 static int ifg_is_dump_node(void *self, ir_node *irn)
1007 return !arch_irn_is_ignore(irn);
1010 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1012 co_ifg_dump_t *env = self;
1013 const arch_register_t *reg = arch_get_irn_register(irn);
1014 const arch_register_req_t *req = arch_get_register_req_out(irn);
1015 int limited = arch_register_req_is(req, limited);
1017 if(env->flags & CO_IFG_DUMP_LABELS) {
1018 ir_fprintf(f, "label=\"%+F", irn);
1020 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1021 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1022 rbitset_copy_to_bitset(req->limited, bs);
1023 ir_fprintf(f, "\\n%B", bs);
1025 ir_fprintf(f, "\" ");
1027 fprintf(f, "label=\"\" shape=point " );
1030 if(env->flags & CO_IFG_DUMP_SHAPE)
1031 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1033 if(env->flags & CO_IFG_DUMP_COLORS)
1034 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1037 static void ifg_dump_at_end(FILE *file, void *self)
1039 co_ifg_dump_t *env = self;
1042 co_gs_foreach_aff_node(env->co, a) {
1043 const arch_register_t *ar = arch_get_irn_register(a->irn);
1044 unsigned aidx = get_irn_idx(a->irn);
1047 co_gs_foreach_neighb(a, n) {
1048 const arch_register_t *nr = arch_get_irn_register(n->irn);
1049 unsigned nidx = get_irn_idx(n->irn);
1052 const char *color = nr == ar ? "blue" : "red";
1053 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1054 if(env->flags & CO_IFG_DUMP_LABELS)
1055 fprintf(file, "label=\"%d\" ", n->costs);
1056 if(env->flags & CO_IFG_DUMP_COLORS)
1057 fprintf(file, "color=%s ", color);
1059 fprintf(file, "style=dotted");
1060 fprintf(file, "];\n");
1067 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1069 ifg_dump_graph_attr,
1078 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1084 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1088 void co_solve_park_moon(copy_opt_t *opt)
1095 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1096 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1097 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1098 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1102 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
1109 n = strlen(env->birg->main_env->cup_name);
1110 tu_name = XMALLOCN(char, n + 1);
1111 strcpy(tu_name, env->birg->main_env->cup_name);
1112 for (i = 0; i < n; ++i)
1113 if (tu_name[i] == '.')
1117 ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
1119 result = fopen(buf, "wt");
1120 if(result == NULL) {
1121 panic("Couldn't open '%s' for writing.", buf);
1127 void co_driver(be_chordal_env_t *cenv)
1129 ir_timer_t *timer = ir_timer_new();
1130 co_complete_stats_t before, after;
1132 int was_optimal = 0;
1134 assert(selected_copyopt);
1136 /* skip copymin if algo is 'none' */
1137 if(selected_copyopt->copyopt == void_algo)
1140 be_liveness_assure_chk(be_get_birg_liveness(cenv->birg));
1142 co = new_copy_opt(cenv, cost_func);
1143 co_build_ou_structure(co);
1144 co_build_graph_structure(co);
1146 co_complete_stats(co, &before);
1148 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1149 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1150 be_stat_ev_ull("co_max_costs", before.max_costs);
1151 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1152 be_stat_ev_ull("co_aff_int", before.aff_int);
1154 be_stat_ev_ull("co_init_costs", before.costs);
1155 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1157 if (dump_flags & DUMP_BEFORE) {
1158 FILE *f = my_open(cenv, "", "-before.dot");
1159 co_dump_ifg_dot(co, f, style_flags);
1163 /* if the algo can improve results, provide an initial solution with heur1 */
1164 if (improve && selected_copyopt->can_improve_existing) {
1165 co_complete_stats_t stats;
1167 /* produce a heuristic solution */
1168 co_solve_heuristic(co);
1170 /* do the stats and provide the current costs */
1171 co_complete_stats(co, &stats);
1172 be_stat_ev_ull("co_prepare_costs", stats.costs);
1175 /* perform actual copy minimization */
1176 ir_timer_reset_and_start(timer);
1177 was_optimal = selected_copyopt->copyopt(co);
1178 ir_timer_stop(timer);
1180 be_stat_ev("co_time", ir_timer_elapsed_msec(timer));
1181 be_stat_ev_ull("co_optimal", was_optimal);
1182 ir_timer_free(timer);
1184 if (dump_flags & DUMP_AFTER) {
1185 FILE *f = my_open(cenv, "", "-after.dot");
1186 co_dump_ifg_dot(co, f, style_flags);
1190 co_complete_stats(co, &after);
1193 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1194 ulong64 evitable = after.costs - after.inevit_costs;
1196 ir_printf("%30F ", cenv->irg);
1197 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1199 if(optimizable_costs > 0)
1200 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1202 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1205 /* Dump the interference graph in Appel's format. */
1206 if (dump_flags & DUMP_APPEL) {
1207 FILE *f = my_open(cenv, "", ".apl");
1208 fprintf(f, "# %lld %lld\n", after.costs, after.unsatisfied_edges);
1209 co_dump_appel_graph(co, f);
1213 be_stat_ev_ull("co_after_costs", after.costs);
1214 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1216 co_free_graph_structure(co);
1217 co_free_ou_structure(co);