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
40 #include "raw_bitset.h"
46 #include "iredges_t.h"
49 #include "irphase_t.h"
50 #include "irprintf_t.h"
57 #include "beintlive_t.h"
58 #include "becopyopt_t.h"
59 #include "becopystat.h"
62 #include "besched_t.h"
63 #include "bejavacoal.h"
64 #include "bestatevent.h"
69 #include "lc_opts_enum.h"
74 #define DUMP_ALL 2 * DUMP_APPEL - 1
76 #define COST_FUNC_FREQ 1
77 #define COST_FUNC_LOOP 2
78 #define COST_FUNC_ALL_ONE 3
80 static unsigned dump_flags = 0;
81 static unsigned style_flags = 0;
82 static unsigned do_stats = 0;
83 static cost_fct_t cost_func = co_get_costs_exec_freq;
84 static unsigned algo = CO_ALGO_HEUR4;
85 static int improve = 1;
87 static const lc_opt_enum_mask_items_t dump_items[] = {
88 { "before", DUMP_BEFORE },
89 { "after", DUMP_AFTER },
90 { "appel", DUMP_APPEL },
95 static const lc_opt_enum_mask_items_t style_items[] = {
96 { "color", CO_IFG_DUMP_COLORS },
97 { "labels", CO_IFG_DUMP_LABELS },
98 { "constr", CO_IFG_DUMP_CONSTR },
99 { "shape", CO_IFG_DUMP_SHAPE },
100 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
104 static const lc_opt_enum_mask_items_t algo_items[] = {
105 { "none", CO_ALGO_NONE },
106 { "heur", CO_ALGO_HEUR },
107 { "heur2", CO_ALGO_HEUR2 },
108 { "heur3", CO_ALGO_HEUR3 },
109 { "heur4", CO_ALGO_HEUR4 },
110 { "ilp", CO_ALGO_ILP },
114 typedef int (*opt_funcptr)(void);
116 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
117 { "freq", (opt_funcptr) co_get_costs_exec_freq },
118 { "loop", (opt_funcptr) co_get_costs_loop_depth },
119 { "one", (opt_funcptr) co_get_costs_all_one },
123 static lc_opt_enum_mask_var_t dump_var = {
124 &dump_flags, dump_items
127 static lc_opt_enum_mask_var_t style_var = {
128 &style_flags, style_items
131 static lc_opt_enum_mask_var_t algo_var = {
135 static lc_opt_enum_func_ptr_var_t cost_func_var = {
136 (opt_funcptr*) &cost_func, cost_func_items
139 static const lc_opt_table_entry_t options[] = {
140 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var),
141 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
142 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
143 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
144 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
145 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
149 /* Insert additional options registration functions here. */
150 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
151 extern void be_co2_register_options(lc_opt_entry_t *grp);
152 extern void be_co3_register_options(lc_opt_entry_t *grp);
154 void be_init_copycoal(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);
164 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
166 #undef QUICK_AND_DIRTY_HACK
168 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
171 return be_ifg_connected(env->ifg, a, b);
173 return values_interfere(env->birg, a, b);
177 /******************************************************************************
180 | | __ ___ _ __ ___ _ __ __ _| |
181 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
182 | |__| | __/ | | | __/ | | (_| | |
183 \_____|\___|_| |_|\___|_| \__,_|_|
185 ******************************************************************************/
187 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
190 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
192 const char *s1, *s2, *s3;
196 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
198 co = xcalloc(1, sizeof(*co));
199 co->cenv = chordal_env;
200 co->aenv = chordal_env->birg->main_env->arch_env;
201 co->irg = chordal_env->irg;
202 co->cls = chordal_env->cls;
203 co->get_costs = get_costs;
205 s1 = get_irp_prog_name();
206 s2 = get_entity_name(get_irg_entity(co->irg));
207 s3 = chordal_env->cls->name;
208 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
209 co->name = xmalloc(len);
210 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
215 void free_copy_opt(copy_opt_t *co) {
220 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
221 const arch_register_req_t *req;
222 const arch_register_t *reg;
224 if (arch_irn_is(co->aenv, irn, ignore))
227 reg = arch_get_irn_register(co->aenv, irn);
228 if (arch_register_type_is(reg, ignore))
231 req = arch_get_register_req(co->aenv, irn, -1);
232 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req))
238 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
241 ir_node *root_block = get_nodes_block(root);
246 /* for phis the copies are placed in the corresponding pred-block */
247 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
249 /* a perm places the copy in the same block as it resides */
250 loop = get_irn_loop(root_block);
253 int d = get_loop_depth(loop);
259 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
261 ir_node *root_bl = get_nodes_block(root);
262 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
264 res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
266 /* don't allow values smaller than one. */
267 return res < 1 ? 1 : res;
271 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
279 /******************************************************************************
280 ____ _ _ _ _ _ _____ _
281 / __ \ | | | | | | (_) | / ____| |
282 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
283 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
284 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
285 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
288 ******************************************************************************/
291 * Determines a maximum weighted independent set with respect to
292 * the interference and conflict edges of all nodes in a qnode.
294 static int ou_max_ind_set_costs(unit_t *ou) {
295 be_chordal_env_t *chordal_env = ou->co->cenv;
296 ir_node **safe, **unsafe;
297 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
300 int max, curr_weight, best_weight = 0;
302 /* assign the nodes into two groups.
303 * safe: node has no interference, hence it is in every max stable set.
304 * unsafe: node has an interference
306 safe = alloca((ou->node_count-1) * sizeof(*safe));
309 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
310 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
312 for(i=1; i<ou->node_count; ++i) {
314 for(o=1; o<ou->node_count; ++o) {
317 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
318 unsafe_costs[unsafe_count] = ou->costs[i];
319 unsafe[unsafe_count] = ou->nodes[i];
326 safe_costs += ou->costs[i];
327 safe[safe_count++] = ou->nodes[i];
332 /* now compute the best set out of the unsafe nodes*/
333 if (unsafe_count > MIS_HEUR_TRIGGER) {
334 bitset_t *best = bitset_alloca(unsafe_count);
335 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
336 for (i=0; i<unsafe_count; ++i) {
338 /* check if it is a stable set */
339 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
340 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
341 bitset_clear(best, i); /* clear the bit and try next one */
345 /* compute the weight */
346 bitset_foreach(best, pos)
347 best_weight += unsafe_costs[pos];
349 /* Exact Algorithm: Brute force */
350 curr = bitset_alloca(unsafe_count);
351 bitset_set_all(curr);
352 while ((max = bitset_popcnt(curr)) != 0) {
353 /* check if curr is a stable set */
354 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
355 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) */
356 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
359 /* if we arrive here, we have a stable set */
360 /* compute the weigth of the stable set*/
362 bitset_foreach(curr, pos)
363 curr_weight += unsafe_costs[pos];
366 if (curr_weight > best_weight) {
367 best_weight = curr_weight;
375 return safe_costs+best_weight;
378 static void co_collect_units(ir_node *irn, void *env) {
379 copy_opt_t *co = env;
382 if (!is_curr_reg_class(co, irn))
384 if (!co_is_optimizable_root(co, irn))
387 /* Init a new unit */
388 unit = xcalloc(1, sizeof(*unit));
390 unit->node_count = 1;
391 INIT_LIST_HEAD(&unit->queue);
393 /* Phi with some/all of its arguments */
394 if (is_Reg_Phi(irn)) {
398 arity = get_irn_arity(irn);
399 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
400 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
401 unit->nodes[0] = irn;
404 for (i=0; i<arity; ++i) {
406 ir_node *arg = get_irn_n(irn, i);
408 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
411 if (nodes_interfere(co->cenv, irn, arg)) {
412 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
416 /* Else insert the argument of the phi to the members of this ou */
417 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
419 if (! arch_irn_is(co->aenv, arg, ignore)) {
420 /* Check if arg has occurred at a prior position in the arg/list */
422 for (o=1; o<unit->node_count; ++o) {
423 if (unit->nodes[o] == arg) {
429 if (!arg_pos) { /* a new argument */
430 /* insert node, set costs */
431 unit->nodes[unit->node_count] = arg;
432 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
434 } else { /* arg has occurred before in same phi */
435 /* increase costs for existing arg */
436 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
440 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
441 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
442 } else if (is_Perm_Proj(co->aenv, irn)) {
443 /* Proj of a perm with corresponding arg */
444 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
445 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
446 unit->costs = xmalloc(2 * sizeof(*unit->costs));
447 unit->node_count = 2;
448 unit->nodes[0] = irn;
449 unit->nodes[1] = get_Perm_src(irn);
450 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
452 const arch_register_req_t *req =
453 arch_get_register_req(co->aenv, irn, -1);
455 /* Src == Tgt of a 2-addr-code instruction */
456 if (is_2addr_code(req)) {
457 const unsigned other = req->other_same;
461 for (i = 0; (1U << i) <= other; ++i) {
462 if (other & (1U << i)) {
463 ir_node *o = get_irn_n(skip_Proj(irn), i);
464 if (!arch_irn_is(co->aenv, o, ignore) &&
465 !nodes_interfere(co->cenv, irn, o)) {
474 unit->nodes = xmalloc(count * sizeof(*unit->nodes));
475 unit->costs = xmalloc(count * sizeof(*unit->costs));
476 unit->node_count = count;
477 unit->nodes[k++] = irn;
479 for (i = 0; 1U << i <= other; ++i) {
480 if (other & (1U << i)) {
481 ir_node *o = get_irn_n(skip_Proj(irn), i);
482 if (!arch_irn_is(co->aenv, o, ignore) &&
483 !nodes_interfere(co->cenv, irn, o)) {
485 unit->costs[k] = co->get_costs(co, irn, o, -1);
492 assert(0 && "This is not an optimizable node!");
496 /* Insert the new unit at a position according to its costs */
497 if (unit->node_count > 1) {
499 struct list_head *tmp;
501 /* Determine the maximum costs this unit can cause: all_nodes_cost */
502 for(i=1; i<unit->node_count; ++i) {
503 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
504 unit->all_nodes_costs += unit->costs[i];
507 /* Determine the minimal costs this unit will cause: min_nodes_costs */
508 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
509 /* Insert the new ou according to its sort_key */
511 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
513 list_add(&unit->units, tmp);
519 #ifdef QUICK_AND_DIRTY_HACK
521 static int compare_ous(const void *k1, const void *k2) {
522 const unit_t *u1 = *((const unit_t **) k1);
523 const unit_t *u2 = *((const unit_t **) k2);
524 int i, o, u1_has_constr, u2_has_constr;
525 arch_register_req_t req;
526 const arch_env_t *aenv = u1->co->aenv;
528 /* Units with constraints come first */
530 for (i=0; i<u1->node_count; ++i) {
531 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
532 if (arch_register_req_is(&req, limited)) {
539 for (i=0; i<u2->node_count; ++i) {
540 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
541 if (arch_register_req_is(&req, limited)) {
547 if (u1_has_constr != u2_has_constr)
548 return u2_has_constr - u1_has_constr;
550 /* Now check, whether the two units are connected */
552 for (i=0; i<u1->node_count; ++i)
553 for (o=0; o<u2->node_count; ++o)
554 if (u1->nodes[i] == u2->nodes[o])
558 /* After all, the sort key decides. Greater keys come first. */
559 return u2->sort_key - u1->sort_key;
564 * Sort the ou's according to constraints and their sort_key
566 static void co_sort_units(copy_opt_t *co) {
567 int i, count = 0, costs;
570 /* get the number of ous, remove them form the list and fill the array */
571 list_for_each_entry(unit_t, ou, &co->units, units)
573 ous = alloca(count * sizeof(*ous));
575 costs = co_get_max_copy_costs(co);
578 list_for_each_entry(unit_t, ou, &co->units, units)
581 INIT_LIST_HEAD(&co->units);
583 assert(count == i && list_empty(&co->units));
585 for (i=0; i<count; ++i)
586 ir_printf("%+F\n", ous[i]->nodes[0]);
588 qsort(ous, count, sizeof(*ous), compare_ous);
591 for (i=0; i<count; ++i)
592 ir_printf("%+F\n", ous[i]->nodes[0]);
594 /* reinsert into list in correct order */
595 for (i=0; i<count; ++i)
596 list_add_tail(&ous[i]->units, &co->units);
598 assert(costs == co_get_max_copy_costs(co));
602 void co_build_ou_structure(copy_opt_t *co) {
603 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
604 INIT_LIST_HEAD(&co->units);
605 irg_walk_graph(co->irg, co_collect_units, NULL, co);
606 #ifdef QUICK_AND_DIRTY_HACK
611 void co_free_ou_structure(copy_opt_t *co) {
614 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
619 co->units.next = NULL;
622 /* co_solve_heuristic() is implemented in becopyheur.c */
624 int co_get_max_copy_costs(const copy_opt_t *co) {
630 list_for_each_entry(unit_t, curr, &co->units, units) {
631 res += curr->inevitable_costs;
632 for (i=1; i<curr->node_count; ++i)
633 res += curr->costs[i];
638 int co_get_inevit_copy_costs(const copy_opt_t *co) {
644 list_for_each_entry(unit_t, curr, &co->units, units)
645 res += curr->inevitable_costs;
649 int co_get_copy_costs(const copy_opt_t *co) {
655 list_for_each_entry(unit_t, curr, &co->units, units) {
656 int root_col = get_irn_col(co, curr->nodes[0]);
657 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
658 res += curr->inevitable_costs;
659 for (i=1; i<curr->node_count; ++i) {
660 int arg_col = get_irn_col(co, curr->nodes[i]);
661 if (root_col != arg_col) {
662 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
663 res += curr->costs[i];
670 int co_get_lower_bound(const copy_opt_t *co) {
676 list_for_each_entry(unit_t, curr, &co->units, units)
677 res += curr->inevitable_costs + curr->min_nodes_costs;
681 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
683 bitset_t *seen = bitset_irg_malloc(co->irg);
686 memset(stat, 0, sizeof(stat[0]));
688 /* count affinity edges. */
689 co_gs_foreach_aff_node(co, an) {
691 stat->aff_nodes += 1;
692 bitset_add_irn(seen, an->irn);
693 co_gs_foreach_neighb(an, neigh) {
694 if(!bitset_contains_irn(seen, neigh->irn)) {
695 stat->aff_edges += 1;
696 stat->max_costs += neigh->costs;
698 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
699 stat->costs += neigh->costs;
700 stat->unsatisfied_edges += 1;
703 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
705 stat->inevit_costs += neigh->costs;
715 /******************************************************************************
717 / ____| | | / ____| |
718 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
719 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
720 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
721 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
724 ******************************************************************************/
726 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
727 const affinity_node_t *n1 = k1;
728 const affinity_node_t *n2 = k2;
731 return (n1->irn != n2->irn);
734 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
735 affinity_node_t new_node, *node;
741 new_node.neighbours = NULL;
742 node = set_insert(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
744 for (nbr = node->neighbours; nbr; nbr = nbr->next)
745 if (nbr->irn == n2) {
750 /* if we did not find n2 in n1's neighbourhood insert it */
752 nbr = obstack_alloc(&co->obst, sizeof(*nbr));
755 nbr->next = node->neighbours;
757 node->neighbours = nbr;
761 /* now nbr points to n1's neighbour-entry of n2 */
765 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
766 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
767 add_edge(co, n1, n2, costs);
768 add_edge(co, n2, n1, costs);
772 static void build_graph_walker(ir_node *irn, void *env) {
773 copy_opt_t *co = env;
775 const arch_register_t *reg;
777 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
780 reg = arch_get_irn_register(co->aenv, irn);
781 if (arch_register_type_is(reg, ignore))
784 if (is_Reg_Phi(irn)) { /* Phis */
785 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
786 ir_node *arg = get_irn_n(irn, pos);
787 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
790 else if (is_Perm_Proj(co->aenv, irn)) { /* Perms */
791 ir_node *arg = get_Perm_src(irn);
792 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
794 else { /* 2-address code */
795 const arch_register_req_t *req = arch_get_register_req(co->aenv, irn, -1);
796 if (is_2addr_code(req)) {
797 const unsigned other = req->other_same;
800 for (i = 0; 1U << i <= other; ++i) {
801 if (other & (1U << i)) {
802 ir_node *other = get_irn_n(skip_Proj(irn), i);
803 if (! arch_irn_is(co->aenv, other, ignore))
804 add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
811 void co_build_graph_structure(copy_opt_t *co) {
812 obstack_init(&co->obst);
813 co->nodes = new_set(compare_affinity_node_t, 32);
815 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
818 void co_free_graph_structure(copy_opt_t *co) {
822 obstack_free(&co->obst, NULL);
826 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
828 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
829 affinity_node_t new_node, *n;
834 n = set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
836 return (n->degree > 0);
841 static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b)
843 ir_node *nodes[] = { a, b };
844 bitset_t *constr[] = { NULL, NULL };
845 const arch_register_req_t *req;
848 constr[0] = bitset_alloca(co->cls->n_regs);
849 constr[1] = bitset_alloca(co->cls->n_regs);
851 for (j = 0; j < 2; ++j) {
852 req = arch_get_register_req(co->aenv, nodes[j], BE_OUT_POS(0));
853 if(arch_register_req_is(req, limited))
854 rbitset_copy_to_bitset(req->limited, constr[j]);
856 bitset_set_all(constr[j]);
860 return !bitset_intersect(constr[0], constr[1]);
863 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
865 be_ifg_t *ifg = co->cenv->ifg;
866 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
867 int *node_map = xmalloc((get_irg_last_idx(co->irg) + 1) * sizeof(node_map[0]));
875 for(i = 0; i < co->cls->n_regs; ++i) {
876 const arch_register_t *reg = &co->cls->regs[i];
877 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
881 * n contains the first node number.
882 * the values below n are the pre-colored register nodes
885 it = be_ifg_nodes_iter_alloca(ifg);
886 nit = be_ifg_neighbours_iter_alloca(ifg);
889 be_ifg_foreach_node(ifg, it, irn) {
890 if(!arch_irn_is(co->aenv, irn, ignore))
891 node_map[get_irn_idx(irn)] = n++;
894 fprintf(f, "%d %d\n", n, n_regs);
896 be_ifg_foreach_node(ifg, it, irn) {
897 if(!arch_irn_is(co->aenv, irn, ignore)) {
898 int idx = node_map[get_irn_idx(irn)];
899 affinity_node_t *a = get_affinity_info(co, irn);
901 const arch_register_req_t *req;
904 req = arch_get_register_req(co->aenv, irn, BE_OUT_POS(0));
905 if(arch_register_req_is(req, limited)) {
906 for(i = 0; i < co->cls->n_regs; ++i) {
907 if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
908 fprintf(f, "%d %d -1\n", color_map[i], idx);
912 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
913 if(!arch_irn_is(co->aenv, adj, ignore) && !co_dump_appel_disjoint_constraints(co, irn, adj)) {
914 int adj_idx = node_map[get_irn_idx(adj)];
916 fprintf(f, "%d %d -1\n", idx, adj_idx);
923 co_gs_foreach_neighb(a, n) {
924 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
925 int n_idx = node_map[get_irn_idx(n->irn)];
927 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
938 ___ _____ ____ ____ ___ _____ ____ _
939 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
940 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
941 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
942 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
946 static const char *get_dot_color_name(size_t col)
948 static const char *names[] = {
982 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
985 typedef struct _co_ifg_dump_t {
986 const copy_opt_t *co;
990 static void ifg_dump_graph_attr(FILE *f, void *self)
993 fprintf(f, "overlap=scale");
996 static int ifg_is_dump_node(void *self, ir_node *irn)
998 co_ifg_dump_t *cod = self;
999 return !arch_irn_is(cod->co->aenv, irn, ignore);
1002 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1004 co_ifg_dump_t *env = self;
1005 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1006 const arch_register_req_t *req;
1009 req = arch_get_register_req(env->co->aenv, irn, BE_OUT_POS(0));
1010 limited = arch_register_req_is(req, limited);
1012 if(env->flags & CO_IFG_DUMP_LABELS) {
1013 ir_fprintf(f, "label=\"%+F", irn);
1015 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1016 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1017 rbitset_copy_to_bitset(req->limited, bs);
1018 ir_fprintf(f, "\\n%B", bs);
1020 ir_fprintf(f, "\" ");
1022 fprintf(f, "label=\"\" shape=point " );
1025 if(env->flags & CO_IFG_DUMP_SHAPE)
1026 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1028 if(env->flags & CO_IFG_DUMP_COLORS)
1029 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1032 static void ifg_dump_at_end(FILE *file, void *self)
1034 co_ifg_dump_t *env = self;
1037 co_gs_foreach_aff_node(env->co, a) {
1038 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1039 unsigned aidx = get_irn_idx(a->irn);
1042 co_gs_foreach_neighb(a, n) {
1043 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1044 unsigned nidx = get_irn_idx(n->irn);
1047 const char *color = nr == ar ? "blue" : "red";
1048 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1049 if(env->flags & CO_IFG_DUMP_LABELS)
1050 fprintf(file, "label=\"%d\" ", n->costs);
1051 if(env->flags & CO_IFG_DUMP_COLORS)
1052 fprintf(file, "color=%s ", color);
1054 fprintf(file, "style=dotted");
1055 fprintf(file, "];\n");
1062 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1064 ifg_dump_graph_attr,
1073 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1079 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1083 void co_solve_park_moon(copy_opt_t *opt)
1088 static int void_algo(copy_opt_t *co)
1096 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1097 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1098 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1099 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1106 int can_improve_existing;
1109 static co_algo_info_t algos[] = {
1110 { void_algo, "none", 0 },
1111 { co_solve_heuristic, "heur1", 0 },
1112 { co_solve_heuristic_new, "heur2", 0 },
1114 { co_solve_heuristic_java, "heur3", 0 },
1116 { NULL, "heur3", 0 },
1118 { co_solve_heuristic_mst, "heur4", 0 },
1120 { co_solve_ilp2, "ilp", 1 },
1129 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1130 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1131 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1132 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1136 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
1143 n = strlen(env->birg->main_env->cup_name);
1144 tu_name = xmalloc((n + 1) * sizeof(*tu_name));
1145 strcpy(tu_name, env->birg->main_env->cup_name);
1146 for (i = 0; i < n; ++i)
1147 if (tu_name[i] == '.')
1151 ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
1153 result = fopen(buf, "wt");
1154 if(result == NULL) {
1155 panic("Couldn't open '%s' for writing.", buf);
1161 void co_driver(be_chordal_env_t *cenv)
1163 ir_timer_t *timer = ir_timer_register("firm.be.copyopt", "runtime");
1164 co_complete_stats_t before, after;
1166 co_algo_t *algo_func;
1167 int was_optimal = 0;
1169 if (algo >= CO_ALGO_LAST)
1172 be_liveness_assure_chk(be_get_birg_liveness(cenv->birg));
1174 co = new_copy_opt(cenv, cost_func);
1175 co_build_ou_structure(co);
1176 co_build_graph_structure(co);
1178 co_complete_stats(co, &before);
1180 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1181 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1182 be_stat_ev_ull("co_max_costs", before.max_costs);
1183 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1184 be_stat_ev_ull("co_aff_int", before.aff_int);
1186 be_stat_ev_ull("co_init_costs", before.costs);
1187 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1189 if (dump_flags & DUMP_BEFORE) {
1190 FILE *f = my_open(cenv, "", "-before.dot");
1191 co_dump_ifg_dot(co, f, style_flags);
1195 /* if the algo can improve results, provide an initial solution with heur3 */
1196 if (improve && algos[algo].can_improve_existing) {
1197 co_complete_stats_t stats;
1199 /* produce a heuristic solution */
1201 co_solve_heuristic_java(co);
1203 co_solve_heuristic(co);
1204 #endif /* WITH_JVM */
1206 /* do the stats and provide the current costs */
1207 co_complete_stats(co, &stats);
1208 be_stat_ev_ull("co_prepare_costs", stats.costs);
1212 /* start the JVM here so that it does not tamper the timing. */
1213 if (algo == CO_ALGO_HEUR3)
1214 be_java_coal_start_jvm();
1215 #endif /* WITH_JVM */
1217 algo_func = algos[algo].algo;
1219 /* perform actual copy minimization */
1220 ir_timer_reset_and_start(timer);
1221 was_optimal = algo_func(co);
1222 ir_timer_stop(timer);
1224 be_stat_ev("co_time", ir_timer_elapsed_msec(timer));
1225 be_stat_ev_ull("co_optimal", was_optimal);
1227 if (dump_flags & DUMP_AFTER) {
1228 FILE *f = my_open(cenv, "", "-after.dot");
1229 co_dump_ifg_dot(co, f, style_flags);
1233 co_complete_stats(co, &after);
1236 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1237 ulong64 evitable = after.costs - after.inevit_costs;
1239 ir_printf("%30F ", cenv->irg);
1240 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1242 if(optimizable_costs > 0)
1243 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1245 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1248 /* Dump the interference graph in Appel's format. */
1249 if (dump_flags & DUMP_APPEL) {
1250 FILE *f = my_open(cenv, "", ".apl");
1251 fprintf(f, "# %lld %lld\n", after.costs, after.unsatisfied_edges);
1252 co_dump_appel_graph(co, f);
1256 be_stat_ev_ull("co_after_costs", after.costs);
1257 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1259 co_free_graph_structure(co);
1260 co_free_ou_structure(co);