2 * Copyright (C) 1995-2007 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 "benodesets.h"
64 #include "bejavacoal.h"
65 #include "bestatevent.h"
69 #include <libcore/lc_timing.h>
70 #include <libcore/lc_opts.h>
71 #include <libcore/lc_opts_enum.h>
76 #define DUMP_ALL 2 * DUMP_APPEL - 1
78 #define COST_FUNC_FREQ 1
79 #define COST_FUNC_LOOP 2
80 #define COST_FUNC_ALL_ONE 3
82 static unsigned dump_flags = 0;
83 static unsigned style_flags = 0;
84 static unsigned do_stats = 0;
85 static cost_fct_t cost_func = co_get_costs_exec_freq;
86 static unsigned algo = CO_ALGO_HEUR4;
87 static int improve = 1;
89 static const lc_opt_enum_mask_items_t dump_items[] = {
90 { "before", DUMP_BEFORE },
91 { "after", DUMP_AFTER },
92 { "appel", DUMP_APPEL },
97 static const lc_opt_enum_mask_items_t style_items[] = {
98 { "color", CO_IFG_DUMP_COLORS },
99 { "labels", CO_IFG_DUMP_LABELS },
100 { "constr", CO_IFG_DUMP_CONSTR },
101 { "shape", CO_IFG_DUMP_SHAPE },
102 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
106 static const lc_opt_enum_mask_items_t algo_items[] = {
107 { "none", CO_ALGO_NONE },
108 { "heur", CO_ALGO_HEUR },
109 { "heur2", CO_ALGO_HEUR2 },
111 { "heur3", CO_ALGO_HEUR3 },
112 #endif /* WITH_JVM */
113 { "heur4", CO_ALGO_HEUR4 },
115 { "ilp", CO_ALGO_ILP },
116 #endif /* WITH_ILP */
120 typedef int (*opt_funcptr)(void);
122 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
123 { "freq", (opt_funcptr) co_get_costs_exec_freq },
124 { "loop", (opt_funcptr) co_get_costs_loop_depth },
125 { "one", (opt_funcptr) co_get_costs_all_one },
129 static lc_opt_enum_mask_var_t dump_var = {
130 &dump_flags, dump_items
133 static lc_opt_enum_mask_var_t style_var = {
134 &style_flags, style_items
137 static lc_opt_enum_mask_var_t algo_var = {
141 static lc_opt_enum_func_ptr_var_t cost_func_var = {
142 (opt_funcptr*) &cost_func, cost_func_items
145 static const lc_opt_table_entry_t options[] = {
146 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var),
147 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
148 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
149 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
150 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
151 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
155 /* Insert additional options registration functions here. */
156 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
157 extern void be_co2_register_options(lc_opt_entry_t *grp);
158 extern void be_co3_register_options(lc_opt_entry_t *grp);
160 void be_init_copycoal(void)
162 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
163 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
164 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
165 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
167 lc_opt_add_table(co_grp, options);
170 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
172 #undef QUICK_AND_DIRTY_HACK
174 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
177 return be_ifg_connected(env->ifg, a, b);
179 return values_interfere(env->birg, a, b);
183 /******************************************************************************
186 | | __ ___ _ __ ___ _ __ __ _| |
187 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
188 | |__| | __/ | | | __/ | | (_| | |
189 \_____|\___|_| |_|\___|_| \__,_|_|
191 ******************************************************************************/
193 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
196 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
198 const char *s1, *s2, *s3;
202 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
204 co = xcalloc(1, sizeof(*co));
205 co->cenv = chordal_env;
206 co->aenv = chordal_env->birg->main_env->arch_env;
207 co->irg = chordal_env->irg;
208 co->cls = chordal_env->cls;
209 co->get_costs = get_costs;
211 s1 = get_irp_prog_name();
212 s2 = get_entity_name(get_irg_entity(co->irg));
213 s3 = chordal_env->cls->name;
214 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
215 co->name = xmalloc(len);
216 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
221 void free_copy_opt(copy_opt_t *co) {
226 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
227 const arch_register_req_t *req;
228 const arch_register_t *reg;
230 if (arch_irn_is(co->aenv, irn, ignore))
233 reg = arch_get_irn_register(co->aenv, irn);
234 if (arch_register_type_is(reg, ignore))
237 req = arch_get_register_req(co->aenv, irn, -1);
238 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req))
244 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
247 ir_node *root_block = get_nodes_block(root);
250 /* for phis the copies are placed in the corresponding pred-block */
251 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
253 /* a perm places the copy in the same block as it resides */
254 loop = get_irn_loop(root_block);
257 int d = get_loop_depth(loop);
263 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
265 ir_node *root_bl = get_nodes_block(root);
266 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
267 res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
269 /* don't allow values smaller than one. */
270 return res < 1 ? 1 : res;
274 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
278 /******************************************************************************
279 ____ _ _ _ _ _ _____ _
280 / __ \ | | | | | | (_) | / ____| |
281 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
282 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
283 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
284 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
287 ******************************************************************************/
290 * Determines a maximum weighted independent set with respect to
291 * the interference and conflict edges of all nodes in a qnode.
293 static int ou_max_ind_set_costs(unit_t *ou) {
294 be_chordal_env_t *chordal_env = ou->co->cenv;
295 ir_node **safe, **unsafe;
296 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
298 int max, pos, curr_weight, best_weight = 0;
300 /* assign the nodes into two groups.
301 * safe: node has no interference, hence it is in every max stable set.
302 * unsafe: node has an interference
304 safe = alloca((ou->node_count-1) * sizeof(*safe));
307 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
308 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
310 for(i=1; i<ou->node_count; ++i) {
312 for(o=1; o<ou->node_count; ++o) {
315 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
316 unsafe_costs[unsafe_count] = ou->costs[i];
317 unsafe[unsafe_count] = ou->nodes[i];
324 safe_costs += ou->costs[i];
325 safe[safe_count++] = ou->nodes[i];
330 /* now compute the best set out of the unsafe nodes*/
331 if (unsafe_count > MIS_HEUR_TRIGGER) {
332 bitset_t *best = bitset_alloca(unsafe_count);
333 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
334 for (i=0; i<unsafe_count; ++i) {
336 /* check if it is a stable set */
337 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
338 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
339 bitset_clear(best, i); /* clear the bit and try next one */
343 /* compute the weight */
344 bitset_foreach(best, pos)
345 best_weight += unsafe_costs[pos];
347 /* Exact Algorithm: Brute force */
348 curr = bitset_alloca(unsafe_count);
349 bitset_set_all(curr);
350 while ((max = bitset_popcnt(curr)) != 0) {
351 /* check if curr is a stable set */
352 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
353 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) */
354 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
357 /* if we arrive here, we have a stable set */
358 /* compute the weigth of the stable set*/
360 bitset_foreach(curr, pos)
361 curr_weight += unsafe_costs[pos];
364 if (curr_weight > best_weight) {
365 best_weight = curr_weight;
373 return safe_costs+best_weight;
376 static void co_collect_units(ir_node *irn, void *env) {
377 copy_opt_t *co = env;
380 if (!is_curr_reg_class(co, irn))
382 if (!co_is_optimizable_root(co, irn))
385 /* Init a new unit */
386 unit = xcalloc(1, sizeof(*unit));
388 unit->node_count = 1;
389 INIT_LIST_HEAD(&unit->queue);
391 /* Phi with some/all of its arguments */
392 if (is_Reg_Phi(irn)) {
396 arity = get_irn_arity(irn);
397 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
398 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
399 unit->nodes[0] = irn;
402 for (i=0; i<arity; ++i) {
404 ir_node *arg = get_irn_n(irn, i);
406 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
409 if (nodes_interfere(co->cenv, irn, arg)) {
410 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
414 /* Else insert the argument of the phi to the members of this ou */
415 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
417 if (! arch_irn_is(co->aenv, arg, ignore)) {
418 /* Check if arg has occurred at a prior position in the arg/list */
420 for (o=1; o<unit->node_count; ++o) {
421 if (unit->nodes[o] == arg) {
427 if (!arg_pos) { /* a new argument */
428 /* insert node, set costs */
429 unit->nodes[unit->node_count] = arg;
430 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
432 } else { /* arg has occured before in same phi */
433 /* increase costs for existing arg */
434 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
438 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
439 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
442 /* Proj of a perm with corresponding arg */
443 if (is_Perm_Proj(co->aenv, irn)) {
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 ir_node *other = get_irn_n(skip_Proj(irn), req->other_same);
458 if (!arch_irn_is(co->aenv, other, ignore) &&
459 !nodes_interfere(co->cenv, irn, other)) {
460 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
461 unit->costs = xmalloc(2 * sizeof(*unit->costs));
462 unit->node_count = 2;
463 unit->nodes[0] = irn;
464 unit->nodes[1] = other;
465 unit->costs[1] = co->get_costs(co, irn, other, -1);
468 assert(0 && "This is not an optimizable node!");
472 /* Insert the new unit at a position according to its costs */
473 if (unit->node_count > 1) {
475 struct list_head *tmp;
477 /* Determine the maximum costs this unit can cause: all_nodes_cost */
478 for(i=1; i<unit->node_count; ++i) {
479 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
480 unit->all_nodes_costs += unit->costs[i];
483 /* Determine the minimal costs this unit will cause: min_nodes_costs */
484 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
485 /* Insert the new ou according to its sort_key */
487 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
489 list_add(&unit->units, tmp);
495 #ifdef QUICK_AND_DIRTY_HACK
497 static int compare_ous(const void *k1, const void *k2) {
498 const unit_t *u1 = *((const unit_t **) k1);
499 const unit_t *u2 = *((const unit_t **) k2);
500 int i, o, u1_has_constr, u2_has_constr;
501 arch_register_req_t req;
502 const arch_env_t *aenv = u1->co->aenv;
504 /* Units with constraints come first */
506 for (i=0; i<u1->node_count; ++i) {
507 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
508 if (arch_register_req_is(&req, limited)) {
515 for (i=0; i<u2->node_count; ++i) {
516 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
517 if (arch_register_req_is(&req, limited)) {
523 if (u1_has_constr != u2_has_constr)
524 return u2_has_constr - u1_has_constr;
526 /* Now check, whether the two units are connected */
528 for (i=0; i<u1->node_count; ++i)
529 for (o=0; o<u2->node_count; ++o)
530 if (u1->nodes[i] == u2->nodes[o])
534 /* After all, the sort key decides. Greater keys come first. */
535 return u2->sort_key - u1->sort_key;
540 * Sort the ou's according to constraints and their sort_key
542 static void co_sort_units(copy_opt_t *co) {
543 int i, count = 0, costs;
546 /* get the number of ous, remove them form the list and fill the array */
547 list_for_each_entry(unit_t, ou, &co->units, units)
549 ous = alloca(count * sizeof(*ous));
551 costs = co_get_max_copy_costs(co);
554 list_for_each_entry(unit_t, ou, &co->units, units)
557 INIT_LIST_HEAD(&co->units);
559 assert(count == i && list_empty(&co->units));
561 for (i=0; i<count; ++i)
562 ir_printf("%+F\n", ous[i]->nodes[0]);
564 qsort(ous, count, sizeof(*ous), compare_ous);
567 for (i=0; i<count; ++i)
568 ir_printf("%+F\n", ous[i]->nodes[0]);
570 /* reinsert into list in correct order */
571 for (i=0; i<count; ++i)
572 list_add_tail(&ous[i]->units, &co->units);
574 assert(costs == co_get_max_copy_costs(co));
578 void co_build_ou_structure(copy_opt_t *co) {
579 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
580 INIT_LIST_HEAD(&co->units);
581 irg_walk_graph(co->irg, co_collect_units, NULL, co);
582 #ifdef QUICK_AND_DIRTY_HACK
587 void co_free_ou_structure(copy_opt_t *co) {
590 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
595 co->units.next = NULL;
598 /* co_solve_heuristic() is implemented in becopyheur.c */
600 int co_get_max_copy_costs(const copy_opt_t *co) {
606 list_for_each_entry(unit_t, curr, &co->units, units) {
607 res += curr->inevitable_costs;
608 for (i=1; i<curr->node_count; ++i)
609 res += curr->costs[i];
614 int co_get_inevit_copy_costs(const copy_opt_t *co) {
620 list_for_each_entry(unit_t, curr, &co->units, units)
621 res += curr->inevitable_costs;
625 int co_get_copy_costs(const copy_opt_t *co) {
631 list_for_each_entry(unit_t, curr, &co->units, units) {
632 int root_col = get_irn_col(co, curr->nodes[0]);
633 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
634 res += curr->inevitable_costs;
635 for (i=1; i<curr->node_count; ++i) {
636 int arg_col = get_irn_col(co, curr->nodes[i]);
637 if (root_col != arg_col) {
638 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
639 res += curr->costs[i];
646 int co_get_lower_bound(const copy_opt_t *co) {
652 list_for_each_entry(unit_t, curr, &co->units, units)
653 res += curr->inevitable_costs + curr->min_nodes_costs;
657 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
659 bitset_t *seen = bitset_irg_malloc(co->irg);
662 memset(stat, 0, sizeof(stat[0]));
664 /* count affinity edges. */
665 co_gs_foreach_aff_node(co, an) {
667 stat->aff_nodes += 1;
668 bitset_add_irn(seen, an->irn);
669 co_gs_foreach_neighb(an, neigh) {
670 if(!bitset_contains_irn(seen, neigh->irn)) {
671 stat->aff_edges += 1;
672 stat->max_costs += neigh->costs;
674 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
675 stat->costs += neigh->costs;
676 stat->unsatisfied_edges += 1;
679 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
681 stat->inevit_costs += neigh->costs;
691 /******************************************************************************
693 / ____| | | / ____| |
694 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
695 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
696 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
697 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
700 ******************************************************************************/
702 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
703 const affinity_node_t *n1 = k1;
704 const affinity_node_t *n2 = k2;
706 return (n1->irn != n2->irn);
709 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
710 affinity_node_t new_node, *node;
716 new_node.neighbours = NULL;
717 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
719 for (nbr = node->neighbours; nbr; nbr = nbr->next)
720 if (nbr->irn == n2) {
725 /* if we did not find n2 in n1's neighbourhood insert it */
727 nbr = obstack_alloc(&co->obst, sizeof(*nbr));
730 nbr->next = node->neighbours;
732 node->neighbours = nbr;
736 /* now nbr points to n1's neighbour-entry of n2 */
740 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
741 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
742 add_edge(co, n1, n2, costs);
743 add_edge(co, n2, n1, costs);
747 static void build_graph_walker(ir_node *irn, void *env) {
748 copy_opt_t *co = env;
750 const arch_register_t *reg;
752 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
755 reg = arch_get_irn_register(co->aenv, irn);
756 if (arch_register_type_is(reg, ignore))
761 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
762 ir_node *arg = get_irn_n(irn, pos);
763 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
767 else if (is_Perm_Proj(co->aenv, irn)) {
768 ir_node *arg = get_Perm_src(irn);
769 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
774 const arch_register_req_t *req =
775 arch_get_register_req(co->aenv, irn, -1);
776 if (is_2addr_code(req)) {
777 ir_node *other = get_irn_n(skip_Proj(irn), req->other_same);
778 if (! arch_irn_is(co->aenv, other, ignore))
779 add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
784 void co_build_graph_structure(copy_opt_t *co) {
785 obstack_init(&co->obst);
786 co->nodes = new_set(compare_affinity_node_t, 32);
788 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
791 void co_free_graph_structure(copy_opt_t *co) {
795 obstack_free(&co->obst, NULL);
799 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
801 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
802 affinity_node_t new_node, *n;
807 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
809 return (n->degree > 0);
814 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
816 be_ifg_t *ifg = co->cenv->ifg;
817 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
824 for(i = 0; i < co->cls->n_regs; ++i) {
825 const arch_register_t *reg = &co->cls->regs[i];
826 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
830 * n contains the first node number.
831 * the values below n are the pre-colored register nodes
834 it = be_ifg_nodes_iter_alloca(ifg);
835 nit = be_ifg_neighbours_iter_alloca(ifg);
838 be_ifg_foreach_node(ifg, it, irn) {
839 if(!arch_irn_is(co->aenv, irn, ignore))
840 set_irn_link(irn, INT_TO_PTR(n++));
843 fprintf(f, "%d %d\n", n, n_regs);
845 be_ifg_foreach_node(ifg, it, irn) {
846 if(!arch_irn_is(co->aenv, irn, ignore)) {
847 int idx = PTR_TO_INT(get_irn_link(irn));
848 affinity_node_t *a = get_affinity_info(co, irn);
850 const arch_register_req_t *req;
853 req = arch_get_register_req(co->aenv, irn, BE_OUT_POS(0));
854 if(arch_register_req_is(req, limited)) {
855 for(i = 0; i < co->cls->n_regs; ++i) {
856 if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
857 fprintf(f, "%d %d -1\n", color_map[i], idx);
862 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
863 if(!arch_irn_is(co->aenv, adj, ignore)) {
864 int adj_idx = PTR_TO_INT(get_irn_link(adj));
866 fprintf(f, "%d %d -1\n", idx, adj_idx);
873 co_gs_foreach_neighb(a, n) {
874 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
875 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
877 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
885 typedef struct _appel_clique_walker_t {
887 const copy_opt_t *co;
894 } appel_clique_walker_t;
896 typedef struct _appel_block_info_t {
906 } appel_block_info_t;
908 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
911 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
912 int res = (int) freq;
913 return res == 0 ? 1 : res;
915 ir_loop *loop = get_irn_loop(bl);
917 int d = get_loop_depth(loop);
924 static void *appel_clique_walker_irn_init(ir_phase *phase, ir_node *irn, void *old)
926 appel_block_info_t *res = NULL;
929 appel_clique_walker_t *d = (void *) phase;
930 res = phase_alloc(phase, sizeof(res[0]));
931 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
932 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
933 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
934 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
935 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
936 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
942 typedef struct _insn_list_t {
944 struct list_head list;
947 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
949 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
952 for(i = 0; i < bli->n_live_end; ++i)
953 if(bli->live_end[i] == irn)
954 return bli->live_end_nr[i];
959 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
961 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
967 foreach_pset(live, irn)
968 live_arr[n_live++] = irn;
970 /* dump the live after clique */
972 for(j = 0; j < n_live; ++j) {
975 for(k = j + 1; k < n_live; ++k) {
976 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
978 fprintf(env->f, "\n");
982 /* dump the affinities */
983 for(j = 0; !env->dumb && j < n_live; ++j) {
984 ir_node *irn = live_arr[j];
985 int old_nr = PTR_TO_INT(get_irn_link(irn));
987 /* if the node was already live in the last insn dump the affinity */
988 if(old_nr > start_nr) {
989 int weight = appel_aff_weight(env, bl);
990 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
994 /* set the current numbers into the link field. */
995 for(j = 0; j < n_live; ++j) {
996 ir_node *irn = live_arr[j];
997 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
1000 return curr_nr + n_live;
1003 static void appel_walker(ir_node *bl, void *data)
1005 appel_clique_walker_t *env = data;
1006 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
1007 struct obstack *obst = &env->obst;
1008 void *base = obstack_base(obst);
1009 pset *live = pset_new_ptr_default();
1010 be_lv_t *lv = env->co->cenv->birg->lv;
1014 int start_nr = env->curr_nr;
1015 int curr_nr = start_nr;
1017 be_insn_env_t insn_env;
1022 insn_env.aenv = env->co->aenv;
1023 insn_env.cls = env->co->cls;
1024 insn_env.obst = obst;
1025 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1027 /* Guess how many insns will be in this block. */
1028 sched_foreach(bl, irn)
1032 insns = malloc(n_nodes * sizeof(insns[0]));
1034 /* Put all insns in an array. */
1035 irn = sched_first(bl);
1036 while(!sched_is_end(irn)) {
1038 insn = be_scan_insn(&insn_env, irn);
1039 insns[n_insns++] = insn;
1040 irn = insn->next_insn;
1043 DBG((dbg, LEVEL_2, "%+F\n", bl));
1044 be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
1046 /* Generate the bad and ugly. */
1047 for(i = n_insns - 1; i >= 0; --i) {
1048 be_insn_t *insn = insns[i];
1050 /* The first live set has to be saved in the block border set. */
1051 if(i == n_insns - 1) {
1053 foreach_pset(live, irn) {
1054 bli->live_end[j] = irn;
1055 bli->live_end_nr[j] = curr_nr + j;
1058 bli->n_live_end = j;
1062 for(j = 0; j < insn->use_start; ++j) {
1063 ir_node *op = insn->ops[j].carrier;
1064 bitset_t *adm = insn->ops[j].regs;
1068 if(!insn->ops[j].has_constraints)
1072 foreach_pset(live, irn) {
1080 assert(nr < pset_count(live));
1082 for(k = 0; k < env->co->cls->n_regs; ++k) {
1083 int mapped_col = env->color_map[k];
1084 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1085 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1090 /* dump the clique and update the stuff. */
1091 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1093 /* remove all defs. */
1094 for(j = 0; j < insn->use_start; ++j)
1095 pset_remove_ptr(live, insn->ops[j].carrier);
1097 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1098 bli->phi[bli->n_phi] = insn->irn;
1099 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1105 for(j = insn->use_start; j < insn->n_ops; ++j)
1106 pset_insert_ptr(live, insn->ops[j].carrier);
1109 /* print the start clique. */
1110 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1113 foreach_pset(live, irn) {
1114 bli->live_in[i] = irn;
1115 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1122 obstack_free(obst, base);
1123 env->curr_nr = curr_nr;
1126 static void appel_inter_block_aff(ir_node *bl, void *data)
1128 appel_clique_walker_t *env = data;
1129 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1133 for(i = 0; i < bli->n_live_in; ++i) {
1134 ir_node *irn = bli->live_in[i];
1136 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1137 ir_node *pred = get_Block_cfgpred_block(bl, j);
1139 int nr = appel_get_live_end_nr(env, pred, irn);
1141 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1145 for(i = 0; i < bli->n_phi; ++i) {
1146 ir_node *irn = bli->phi[i];
1148 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1149 ir_node *pred = get_Block_cfgpred_block(bl, j);
1150 ir_node *op = get_irn_n(irn, j);
1152 int nr = appel_get_live_end_nr(env, pred, op);
1154 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1160 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1164 appel_clique_walker_t env;
1165 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1166 be_lv_t *lv = co->cenv->birg->lv;
1168 be_liveness_recompute(lv);
1169 obstack_init(&env.obst);
1170 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init, NULL);
1171 env.curr_nr = co->cls->n_regs;
1175 bitset_copy(adm, co->cenv->ignore_colors);
1176 bitset_flip_all(adm);
1178 /* Make color map. */
1179 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1180 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1181 const arch_register_t *reg = &co->cls->regs[i];
1182 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1186 env.curr_nr = n_colors;
1187 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1188 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1190 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1192 /* make the first k nodes interfere */
1193 for(i = 0; i < n_colors; ++i) {
1195 for(j = i + 1; j < n_colors; ++j)
1196 fprintf(f, "%d %d -1 ", i, j);
1201 env.curr_nr = n_colors;
1202 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1203 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1204 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1205 obstack_free(&env.obst, NULL);
1209 ___ _____ ____ ____ ___ _____ ____ _
1210 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1211 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1212 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1213 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1217 static const char *get_dot_color_name(int col)
1219 static const char *names[] = {
1253 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1256 typedef struct _co_ifg_dump_t {
1257 const copy_opt_t *co;
1261 static void ifg_dump_graph_attr(FILE *f, void *self)
1263 fprintf(f, "overlap=scale");
1266 static int ifg_is_dump_node(void *self, ir_node *irn)
1268 co_ifg_dump_t *cod = self;
1269 return !arch_irn_is(cod->co->aenv, irn, ignore);
1272 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1274 co_ifg_dump_t *env = self;
1275 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1276 const arch_register_req_t *req;
1279 req = arch_get_register_req(env->co->aenv, irn, BE_OUT_POS(0));
1280 limited = arch_register_req_is(req, limited);
1282 if(env->flags & CO_IFG_DUMP_LABELS) {
1283 ir_fprintf(f, "label=\"%+F", irn);
1287 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1288 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1289 req.limited(req.limited_env, bs);
1290 ir_fprintf(f, "\\n%B", bs);
1293 ir_fprintf(f, "\" ");
1295 fprintf(f, "label=\"\" shape=point " );
1298 if(env->flags & CO_IFG_DUMP_SHAPE)
1299 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1301 if(env->flags & CO_IFG_DUMP_COLORS)
1302 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1305 static void ifg_dump_at_end(FILE *file, void *self)
1307 co_ifg_dump_t *env = self;
1310 co_gs_foreach_aff_node(env->co, a) {
1311 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1312 unsigned aidx = get_irn_idx(a->irn);
1315 co_gs_foreach_neighb(a, n) {
1316 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1317 unsigned nidx = get_irn_idx(n->irn);
1320 const char *color = nr == ar ? "blue" : "red";
1321 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1322 if(env->flags & CO_IFG_DUMP_LABELS)
1323 fprintf(file, "label=\"%d\" ", n->costs);
1324 if(env->flags & CO_IFG_DUMP_COLORS)
1325 fprintf(file, "color=%s ", color);
1327 fprintf(file, "style=dotted");
1328 fprintf(file, "];\n");
1335 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1337 ifg_dump_graph_attr,
1346 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1352 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1356 void co_solve_park_moon(copy_opt_t *opt)
1361 static int void_algo(copy_opt_t *co)
1368 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1369 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1370 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1371 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1378 int can_improve_existing;
1381 static co_algo_info_t algos[] = {
1382 { void_algo, "none", 0 },
1383 { co_solve_heuristic, "heur1", 0 },
1384 { co_solve_heuristic_new, "heur2", 0 },
1385 { co_solve_heuristic_java, "heur3", 0 },
1386 { co_solve_heuristic_mst, "heur4", 0 },
1388 { co_solve_ilp2, "ilp", 1 },
1395 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1396 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1397 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1398 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1402 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
1407 ir_snprintf(buf, sizeof(buf), "%s%F_%s%s", prefix, env->irg, env->cls->name, suffix);
1408 result = fopen(buf, "wt");
1409 if(result == NULL) {
1410 panic("Couldn't open '%s' for writing.", buf);
1416 void co_driver(be_chordal_env_t *cenv)
1418 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1419 co_complete_stats_t before, after;
1421 co_algo_t *algo_func;
1422 int was_optimal = 0;
1424 if (algo < 0 || algo >= CO_ALGO_LAST)
1427 co = new_copy_opt(cenv, cost_func);
1428 co_build_ou_structure(co);
1429 co_build_graph_structure(co);
1431 co_complete_stats(co, &before);
1433 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1434 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1435 be_stat_ev_ull("co_max_costs", before.max_costs);
1436 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1437 be_stat_ev_ull("co_aff_int", before.aff_int);
1439 be_stat_ev_ull("co_init_costs", before.costs);
1440 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1442 /* Dump the interference graph in Appel's format. */
1443 if (dump_flags & DUMP_APPEL) {
1444 FILE *f = my_open(cenv, "", ".apl");
1445 co_dump_appel_graph(co, f);
1449 if (dump_flags & DUMP_BEFORE) {
1450 FILE *f = my_open(cenv, "", "-before.dot");
1451 co_dump_ifg_dot(co, f, style_flags);
1455 /* if the algo can improve results, provide an initial solution with heur3 */
1456 if (improve && algos[algo].can_improve_existing) {
1457 co_complete_stats_t stats;
1459 /* produce a heuristic solution */
1461 co_solve_heuristic_java(co);
1463 co_solve_heuristic(co);
1464 #endif /* WITH_JVM */
1466 /* do the stats and provide the current costs */
1467 co_complete_stats(co, &stats);
1468 be_stat_ev_ull("co_prepare_costs", stats.costs);
1472 /* start the JVM here so that it does not tamper the timing. */
1473 if (algo == CO_ALGO_HEUR3)
1474 be_java_coal_start_jvm();
1475 #endif /* WITH_JVM */
1477 algo_func = algos[algo].algo;
1479 /* perform actual copy minimization */
1480 lc_timer_reset_and_start(timer);
1481 was_optimal = algo_func(co);
1482 lc_timer_stop(timer);
1484 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1485 be_stat_ev_ull("co_optimal", was_optimal);
1487 if (dump_flags & DUMP_AFTER) {
1488 FILE *f = my_open(cenv, "", "-after.dot");
1489 co_dump_ifg_dot(co, f, style_flags);
1493 co_complete_stats(co, &after);
1496 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1497 ulong64 evitable = after.costs - after.inevit_costs;
1499 ir_printf("%30F ", cenv->irg);
1500 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1502 if(optimizable_costs > 0)
1503 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1505 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1508 be_stat_ev_ull("co_after_costs", after.costs);
1509 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1511 co_free_graph_structure(co);
1512 co_free_ou_structure(co);