4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
25 #include "iredges_t.h"
28 #include "irphase_t.h"
29 #include "irprintf_t.h"
36 #include "becopyopt_t.h"
37 #include "becopystat.h"
40 #include "besched_t.h"
41 #include "benodesets.h"
42 #include "bestatevent.h"
47 #define DUMP_ALL 2 * DUMP_APPEL - 1
49 #define COST_FUNC_FREQ 1
50 #define COST_FUNC_LOOP 2
51 #define COST_FUNC_ALL_ONE 3
53 static int dump_flags = 0;
54 static int style_flags = 0;
55 static int do_stats = 0;
56 static cost_fct_t cost_func = co_get_costs_exec_freq;
57 static int algo = CO_ALGO_HEUR2;
58 static int improve = 1;
61 static const lc_opt_enum_mask_items_t dump_items[] = {
62 { "before", DUMP_BEFORE },
63 { "after", DUMP_AFTER },
64 { "appel", DUMP_APPEL },
69 static const lc_opt_enum_mask_items_t style_items[] = {
70 { "color", CO_IFG_DUMP_COLORS },
71 { "labels", CO_IFG_DUMP_LABELS },
72 { "constr", CO_IFG_DUMP_CONSTR },
73 { "shape", CO_IFG_DUMP_SHAPE },
74 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
78 static const lc_opt_enum_mask_items_t algo_items[] = {
79 { "none", CO_ALGO_NONE },
80 { "heur", CO_ALGO_HEUR },
81 { "heur2", CO_ALGO_HEUR2 },
82 { "heur3", CO_ALGO_HEUR3 },
83 { "ilp", CO_ALGO_ILP },
87 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
88 { "freq", co_get_costs_exec_freq },
89 { "loop", co_get_costs_loop_depth },
90 { "one", co_get_costs_all_one },
94 static lc_opt_enum_mask_var_t dump_var = {
95 &dump_flags, dump_items
98 static lc_opt_enum_mask_var_t style_var = {
99 &style_flags, style_items
102 static lc_opt_enum_mask_var_t algo_var = {
106 static lc_opt_enum_func_ptr_var_t cost_func_var = {
107 &cost_func, cost_func_items
110 static const lc_opt_table_entry_t options[] = {
111 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo (heur, heur2, heur3, ilp)", &algo_var),
112 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function (freq, loop, one)", &cost_func_var),
113 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
114 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
115 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
116 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
120 /* Insert additional options registration functions here. */
121 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
122 extern void be_co2_register_options(lc_opt_entry_t *grp);
123 extern void be_co3_register_options(lc_opt_entry_t *grp);
125 void co_register_options(lc_opt_entry_t *grp)
127 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
128 lc_opt_add_table(co_grp, options);
130 be_co2_register_options(co_grp);
131 be_co3_register_options(co_grp);
133 be_co_ilp_register_options(co_grp);
139 #undef QUICK_AND_DIRTY_HACK
141 /******************************************************************************
144 | | __ ___ _ __ ___ _ __ __ _| |
145 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
146 | |__| | __/ | | | __/ | | (_| | |
147 \_____|\___|_| |_|\___|_| \__,_|_|
149 ******************************************************************************/
151 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
153 void be_copy_opt_init(void) {
156 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
158 const char *s1, *s2, *s3;
162 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
164 co = xcalloc(1, sizeof(*co));
165 co->cenv = chordal_env;
166 co->aenv = chordal_env->birg->main_env->arch_env;
167 co->irg = chordal_env->irg;
168 co->cls = chordal_env->cls;
169 co->get_costs = get_costs;
171 s1 = get_irp_prog_name();
172 s2 = get_entity_name(get_irg_entity(co->irg));
173 s3 = chordal_env->cls->name;
174 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
175 co->name = xmalloc(len);
176 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
181 void free_copy_opt(copy_opt_t *co) {
186 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
187 arch_register_req_t req;
188 const arch_register_t *reg;
190 if (arch_irn_is(co->aenv, irn, ignore))
193 reg = arch_get_irn_register(co->aenv, irn);
194 if (arch_register_type_is(reg, ignore))
197 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
203 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
204 const ir_edge_t *edge;
205 const arch_register_t *reg;
207 assert(0 && "Is buggy and obsolete. Do not use");
209 if (arch_irn_is(co->aenv, irn, ignore))
212 reg = arch_get_irn_register(co->aenv, irn);
213 if (arch_register_type_is(reg, ignore))
216 foreach_out_edge(irn, edge) {
217 ir_node *n = edge->src;
219 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
220 arch_register_req_t req;
221 arch_get_register_req(co->aenv, &req, n, -1);
224 is_Perm(co->aenv, n) ||
225 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
234 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
237 ir_node *root_block = get_nodes_block(root);
240 /* for phis the copies are placed in the corresponding pred-block */
241 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
243 /* a perm places the copy in the same block as it resides */
244 loop = get_irn_loop(root_block);
247 int d = get_loop_depth(loop);
253 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
255 ir_node *root_bl = get_nodes_block(root);
256 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
257 res = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
259 /* don't allow values smaller than one. */
260 return res < 1 ? 1 : res;
264 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
268 /******************************************************************************
269 ____ _ _ _ _ _ _____ _
270 / __ \ | | | | | | (_) | / ____| |
271 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
272 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
273 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
274 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
277 ******************************************************************************/
280 * Determines a maximum weighted independent set with respect to
281 * the interference and conflict edges of all nodes in a qnode.
283 static int ou_max_ind_set_costs(unit_t *ou) {
284 be_chordal_env_t *chordal_env = ou->co->cenv;
285 ir_node **safe, **unsafe;
286 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
288 int max, pos, curr_weight, best_weight = 0;
290 /* assign the nodes into two groups.
291 * safe: node has no interference, hence it is in every max stable set.
292 * unsafe: node has an interference
294 safe = alloca((ou->node_count-1) * sizeof(*safe));
297 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
298 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
300 for(i=1; i<ou->node_count; ++i) {
302 for(o=1; o<ou->node_count; ++o) {
305 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
306 unsafe_costs[unsafe_count] = ou->costs[i];
307 unsafe[unsafe_count] = ou->nodes[i];
314 safe_costs += ou->costs[i];
315 safe[safe_count++] = ou->nodes[i];
320 /* now compute the best set out of the unsafe nodes*/
321 if (unsafe_count > MIS_HEUR_TRIGGER) {
322 bitset_t *best = bitset_alloca(unsafe_count);
323 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
324 for (i=0; i<unsafe_count; ++i) {
326 /* check if it is a stable set */
327 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
328 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
329 bitset_clear(best, i); /* clear the bit and try next one */
333 /* compute the weight */
334 bitset_foreach(best, pos)
335 best_weight += unsafe_costs[pos];
337 /* Exact Algorithm: Brute force */
338 curr = bitset_alloca(unsafe_count);
339 bitset_set_all(curr);
340 while ((max = bitset_popcnt(curr)) != 0) {
341 /* check if curr is a stable set */
342 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
343 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) */
344 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
347 /* if we arrive here, we have a stable set */
348 /* compute the weigth of the stable set*/
350 bitset_foreach(curr, pos)
351 curr_weight += unsafe_costs[pos];
354 if (curr_weight > best_weight) {
355 best_weight = curr_weight;
363 return safe_costs+best_weight;
366 static void co_collect_units(ir_node *irn, void *env) {
367 copy_opt_t *co = env;
369 arch_register_req_t req;
371 if (!is_curr_reg_class(co, irn))
373 if (!co_is_optimizable_root(co, irn))
376 /* Init a new unit */
377 unit = xcalloc(1, sizeof(*unit));
379 unit->node_count = 1;
380 INIT_LIST_HEAD(&unit->queue);
382 /* Phi with some/all of its arguments */
383 if (is_Reg_Phi(irn)) {
387 arity = get_irn_arity(irn);
388 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
389 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
390 unit->nodes[0] = irn;
393 for (i=0; i<arity; ++i) {
395 ir_node *arg = get_irn_n(irn, i);
397 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
400 if (nodes_interfere(co->cenv, irn, arg)) {
401 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
405 /* Else insert the argument of the phi to the members of this ou */
406 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
408 /* Check if arg has occurred at a prior position in the arg/list */
410 for (o=0; o<unit->node_count; ++o)
411 if (unit->nodes[o] == arg) {
416 if (!arg_pos) { /* a new argument */
417 /* insert node, set costs */
418 unit->nodes[unit->node_count] = arg;
419 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
421 } else { /* arg has occured before in same phi */
422 /* increase costs for existing arg */
423 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
426 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
427 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
430 /* Proj of a perm with corresponding arg */
431 if (is_Perm_Proj(co->aenv, irn)) {
432 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
433 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
434 unit->costs = xmalloc(2 * sizeof(*unit->costs));
435 unit->node_count = 2;
436 unit->nodes[0] = irn;
437 unit->nodes[1] = get_Perm_src(irn);
438 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
441 /* Src == Tgt of a 2-addr-code instruction */
442 if (is_2addr_code(co->aenv, irn, &req)) {
443 ir_node *other = req.other_same;
444 if (!nodes_interfere(co->cenv, irn, other)) {
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] = other;
450 unit->costs[1] = co->get_costs(co, irn, other, -1);
453 assert(0 && "This is not an optimizable node!");
455 /* Insert the new unit at a position according to its costs */
456 if (unit->node_count > 1) {
458 struct list_head *tmp;
460 /* Determine the maximum costs this unit can cause: all_nodes_cost */
461 for(i=1; i<unit->node_count; ++i) {
462 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
463 unit->all_nodes_costs += unit->costs[i];
466 /* Determine the minimal costs this unit will cause: min_nodes_costs */
467 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
468 /* Insert the new ou according to its sort_key */
470 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
472 list_add(&unit->units, tmp);
478 #ifdef QUICK_AND_DIRTY_HACK
480 static int compare_ous(const void *k1, const void *k2) {
481 const unit_t *u1 = *((const unit_t **) k1);
482 const unit_t *u2 = *((const unit_t **) k2);
483 int i, o, u1_has_constr, u2_has_constr;
484 arch_register_req_t req;
485 const arch_env_t *aenv = u1->co->aenv;
487 /* Units with constraints come first */
489 for (i=0; i<u1->node_count; ++i) {
490 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
491 if (arch_register_req_is(&req, limited)) {
498 for (i=0; i<u2->node_count; ++i) {
499 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
500 if (arch_register_req_is(&req, limited)) {
506 if (u1_has_constr != u2_has_constr)
507 return u2_has_constr - u1_has_constr;
509 /* Now check, whether the two units are connected */
511 for (i=0; i<u1->node_count; ++i)
512 for (o=0; o<u2->node_count; ++o)
513 if (u1->nodes[i] == u2->nodes[o])
517 /* After all, the sort key decides. Greater keys come first. */
518 return u2->sort_key - u1->sort_key;
523 * Sort the ou's according to constraints and their sort_key
525 static void co_sort_units(copy_opt_t *co) {
526 int i, count = 0, costs;
529 /* get the number of ous, remove them form the list and fill the array */
530 list_for_each_entry(unit_t, ou, &co->units, units)
532 ous = alloca(count * sizeof(*ous));
534 costs = co_get_max_copy_costs(co);
537 list_for_each_entry(unit_t, ou, &co->units, units)
540 INIT_LIST_HEAD(&co->units);
542 assert(count == i && list_empty(&co->units));
544 for (i=0; i<count; ++i)
545 ir_printf("%+F\n", ous[i]->nodes[0]);
547 qsort(ous, count, sizeof(*ous), compare_ous);
550 for (i=0; i<count; ++i)
551 ir_printf("%+F\n", ous[i]->nodes[0]);
553 /* reinsert into list in correct order */
554 for (i=0; i<count; ++i)
555 list_add_tail(&ous[i]->units, &co->units);
557 assert(costs == co_get_max_copy_costs(co));
561 void co_build_ou_structure(copy_opt_t *co) {
562 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
563 INIT_LIST_HEAD(&co->units);
564 irg_walk_graph(co->irg, co_collect_units, NULL, co);
565 #ifdef QUICK_AND_DIRTY_HACK
570 void co_free_ou_structure(copy_opt_t *co) {
573 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
578 co->units.next = NULL;
581 /* co_solve_heuristic() is implemented in becopyheur.c */
583 int co_get_max_copy_costs(const copy_opt_t *co) {
589 list_for_each_entry(unit_t, curr, &co->units, units) {
590 res += curr->inevitable_costs;
591 for (i=1; i<curr->node_count; ++i)
592 res += curr->costs[i];
597 int co_get_inevit_copy_costs(const copy_opt_t *co) {
603 list_for_each_entry(unit_t, curr, &co->units, units)
604 res += curr->inevitable_costs;
608 int co_get_copy_costs(const copy_opt_t *co) {
614 list_for_each_entry(unit_t, curr, &co->units, units) {
615 int root_col = get_irn_col(co, curr->nodes[0]);
616 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
617 res += curr->inevitable_costs;
618 for (i=1; i<curr->node_count; ++i) {
619 int arg_col = get_irn_col(co, curr->nodes[i]);
620 if (root_col != arg_col) {
621 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
622 res += curr->costs[i];
629 int co_get_lower_bound(const copy_opt_t *co) {
635 list_for_each_entry(unit_t, curr, &co->units, units)
636 res += curr->inevitable_costs + curr->min_nodes_costs;
640 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
642 bitset_t *seen = bitset_irg_malloc(co->irg);
645 memset(stat, 0, sizeof(stat[0]));
647 /* count affinity edges. */
648 co_gs_foreach_aff_node(co, an) {
650 stat->aff_nodes += 1;
651 bitset_add_irn(seen, an->irn);
652 co_gs_foreach_neighb(an, neigh) {
653 if(!bitset_contains_irn(seen, neigh->irn)) {
654 stat->aff_edges += 1;
655 stat->max_costs += neigh->costs;
657 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
658 stat->costs += neigh->costs;
659 stat->unsatisfied_edges += 1;
662 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
664 stat->inevit_costs += neigh->costs;
674 /******************************************************************************
676 / ____| | | / ____| |
677 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
678 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
679 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
680 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
683 ******************************************************************************/
685 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
686 const affinity_node_t *n1 = k1;
687 const affinity_node_t *n2 = k2;
689 return (n1->irn != n2->irn);
692 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
693 affinity_node_t new_node, *node;
694 neighb_t new_nbr, *nbr;
699 new_node.neighbours = NULL;
700 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
703 for (nbr = node->neighbours; nbr; nbr = nbr->next)
704 if (nbr->irn == n2) {
709 /* if we did not find n2 in n1's neighbourhood insert it */
711 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
712 nbr = obstack_finish(&co->obst);
715 nbr->next = node->neighbours;
716 node->neighbours = nbr;
720 /* now nbr points to n1's neighbour-entry of n2 */
724 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
725 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
726 add_edge(co, n1, n2, costs);
727 add_edge(co, n2, n1, costs);
731 static void build_graph_walker(ir_node *irn, void *env) {
732 copy_opt_t *co = env;
734 arch_register_req_t req;
735 const arch_register_t *reg;
737 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
740 reg = arch_get_irn_register(co->aenv, irn);
741 if (arch_register_type_is(reg, ignore))
746 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
747 ir_node *arg = get_irn_n(irn, pos);
748 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
752 else if (is_Perm_Proj(co->aenv, irn)) {
753 ir_node *arg = get_Perm_src(irn);
754 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
758 else if (is_2addr_code(co->aenv, irn, &req))
759 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
762 void co_build_graph_structure(copy_opt_t *co) {
763 obstack_init(&co->obst);
764 co->nodes = new_set(compare_affinity_node_t, 32);
766 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
769 void co_free_graph_structure(copy_opt_t *co) {
773 obstack_free(&co->obst, NULL);
777 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
779 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
780 affinity_node_t new_node, *n;
785 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
787 return (n->degree > 0);
792 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
794 be_ifg_t *ifg = co->cenv->ifg;
795 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
796 bitset_t *adm = bitset_alloca(co->cls->n_regs);
803 for(i = 0; i < co->cls->n_regs; ++i) {
804 const arch_register_t *reg = &co->cls->regs[i];
805 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
809 * n contains the first node number.
810 * the values below n are the pre-colored register nodes
813 it = be_ifg_nodes_iter_alloca(ifg);
814 nit = be_ifg_neighbours_iter_alloca(ifg);
817 be_ifg_foreach_node(ifg, it, irn) {
818 if(!arch_irn_is(co->aenv, irn, ignore))
819 set_irn_link(irn, INT_TO_PTR(n++));
822 fprintf(f, "%d %d\n", n, n_regs);
824 be_ifg_foreach_node(ifg, it, irn) {
825 if(!arch_irn_is(co->aenv, irn, ignore)) {
826 int idx = PTR_TO_INT(get_irn_link(irn));
827 affinity_node_t *a = get_affinity_info(co, irn);
829 arch_register_req_t req;
832 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
833 if(arch_register_req_is(&req, limited)) {
834 bitset_clear_all(adm);
835 req.limited(req.limited_env, adm);
836 for(i = 0; i < co->cls->n_regs; ++i)
837 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
838 fprintf(f, "%d %d -1\n", color_map[i], idx);
843 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
844 if(!arch_irn_is(co->aenv, adj, ignore)) {
845 int adj_idx = PTR_TO_INT(get_irn_link(adj));
847 fprintf(f, "%d %d -1\n", idx, adj_idx);
854 co_gs_foreach_neighb(a, n) {
855 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
856 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
858 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
866 typedef struct _appel_clique_walker_t {
868 const copy_opt_t *co;
875 } appel_clique_walker_t;
877 typedef struct _appel_block_info_t {
887 } appel_block_info_t;
889 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
892 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
893 int res = (int) freq;
894 return res == 0 ? 1 : res;
896 ir_loop *loop = get_irn_loop(bl);
898 int d = get_loop_depth(loop);
905 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
907 appel_block_info_t *res = NULL;
910 appel_clique_walker_t *d = (void *) phase;
911 res = phase_alloc(phase, sizeof(res[0]));
912 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
913 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
914 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
915 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
916 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
917 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
923 typedef struct _insn_list_t {
925 struct list_head list;
928 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
930 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
933 for(i = 0; i < bli->n_live_end; ++i)
934 if(bli->live_end[i] == irn)
935 return bli->live_end_nr[i];
940 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
942 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
948 foreach_pset(live, irn)
949 live_arr[n_live++] = irn;
951 /* dump the live after clique */
953 for(j = 0; j < n_live; ++j) {
956 for(k = j + 1; k < n_live; ++k) {
957 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
959 fprintf(env->f, "\n");
963 /* dump the affinities */
964 for(j = 0; !env->dumb && j < n_live; ++j) {
965 ir_node *irn = live_arr[j];
966 int old_nr = PTR_TO_INT(get_irn_link(irn));
968 /* if the node was already live in the last insn dump the affinity */
969 if(old_nr > start_nr) {
970 int weight = appel_aff_weight(env, bl);
971 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
975 /* set the current numbers into the link field. */
976 for(j = 0; j < n_live; ++j) {
977 ir_node *irn = live_arr[j];
978 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
981 return curr_nr + n_live;
984 static void appel_walker(ir_node *bl, void *data)
986 appel_clique_walker_t *env = data;
987 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
988 struct obstack *obst = &env->obst;
989 void *base = obstack_base(obst);
990 pset *live = pset_new_ptr_default();
994 int start_nr = env->curr_nr;
995 int curr_nr = start_nr;
997 be_insn_env_t insn_env;
1002 insn_env.aenv = env->co->aenv;
1003 insn_env.cls = env->co->cls;
1004 insn_env.obst = obst;
1005 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1007 /* Guess how many insns will be in this block. */
1008 sched_foreach(bl, irn)
1012 insns = malloc(n_nodes * sizeof(insns[0]));
1014 /* Put all insns in an array. */
1015 irn = sched_first(bl);
1016 while(!sched_is_end(irn)) {
1018 insn = be_scan_insn(&insn_env, irn);
1019 insns[n_insns++] = insn;
1020 irn = insn->next_insn;
1023 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1024 be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
1026 /* Generate the bad and ugly. */
1027 for(i = n_insns - 1; i >= 0; --i) {
1028 be_insn_t *insn = insns[i];
1030 /* The first live set has to be saved in the block border set. */
1031 if(i == n_insns - 1) {
1033 foreach_pset(live, irn) {
1034 bli->live_end[j] = irn;
1035 bli->live_end_nr[j] = curr_nr + j;
1038 bli->n_live_end = j;
1042 for(j = 0; j < insn->use_start; ++j) {
1043 ir_node *op = insn->ops[j].carrier;
1044 bitset_t *adm = insn->ops[j].regs;
1048 if(!insn->ops[j].has_constraints)
1052 foreach_pset(live, irn) {
1060 assert(nr < pset_count(live));
1062 for(k = 0; k < env->co->cls->n_regs; ++k) {
1063 int mapped_col = env->color_map[k];
1064 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1065 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1070 /* dump the clique and update the stuff. */
1071 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1073 /* remove all defs. */
1074 for(j = 0; j < insn->use_start; ++j)
1075 pset_remove_ptr(live, insn->ops[j].carrier);
1077 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1078 bli->phi[bli->n_phi] = insn->irn;
1079 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1085 for(j = insn->use_start; j < insn->n_ops; ++j)
1086 pset_insert_ptr(live, insn->ops[j].carrier);
1089 /* print the start clique. */
1090 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1093 foreach_pset(live, irn) {
1094 bli->live_in[i] = irn;
1095 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1102 obstack_free(obst, base);
1103 env->curr_nr = curr_nr;
1106 static void appel_inter_block_aff(ir_node *bl, void *data)
1108 appel_clique_walker_t *env = data;
1109 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1113 for(i = 0; i < bli->n_live_in; ++i) {
1114 ir_node *irn = bli->live_in[i];
1116 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1117 ir_node *pred = get_Block_cfgpred_block(bl, j);
1119 int nr = appel_get_live_end_nr(env, pred, irn);
1121 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1125 for(i = 0; i < bli->n_phi; ++i) {
1126 ir_node *irn = bli->phi[i];
1128 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1129 ir_node *pred = get_Block_cfgpred_block(bl, j);
1130 ir_node *op = get_irn_n(irn, j);
1132 int nr = appel_get_live_end_nr(env, pred, op);
1134 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1140 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1144 appel_clique_walker_t env;
1145 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1147 be_liveness_recompute(co->cenv->lv);
1148 obstack_init(&env.obst);
1149 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1150 env.curr_nr = co->cls->n_regs;
1154 bitset_copy(adm, co->cenv->ignore_colors);
1155 bitset_flip_all(adm);
1157 /* Make color map. */
1158 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1159 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1160 const arch_register_t *reg = &co->cls->regs[i];
1161 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1165 env.curr_nr = n_colors;
1166 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1167 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1169 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1171 /* make the first k nodes interfere */
1172 for(i = 0; i < n_colors; ++i) {
1174 for(j = i + 1; j < n_colors; ++j)
1175 fprintf(f, "%d %d -1 ", i, j);
1180 env.curr_nr = n_colors;
1181 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1182 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1183 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1184 obstack_free(&env.obst, NULL);
1188 ___ _____ ____ ____ ___ _____ ____ _
1189 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1190 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1191 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1192 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1196 static const char *get_dot_color_name(int col)
1198 static const char *names[] = {
1232 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1235 typedef struct _co_ifg_dump_t {
1236 const copy_opt_t *co;
1240 static void ifg_dump_graph_attr(FILE *f, void *self)
1242 fprintf(f, "overlap=scale");
1245 static int ifg_is_dump_node(void *self, ir_node *irn)
1247 co_ifg_dump_t *cod = self;
1248 return !arch_irn_is(cod->co->aenv, irn, ignore);
1251 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1253 co_ifg_dump_t *env = self;
1254 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1255 arch_register_req_t req;
1258 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1259 limited = arch_register_req_is(&req, limited);
1261 if(env->flags & CO_IFG_DUMP_LABELS) {
1262 ir_fprintf(f, "label=\"%+F", irn);
1264 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1265 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1266 req.limited(req.limited_env, bs);
1267 ir_fprintf(f, "\\n%B", bs);
1269 ir_fprintf(f, "\" ");
1273 fprintf(f, "label=\"\" shape=point " );
1275 if(env->flags & CO_IFG_DUMP_SHAPE)
1276 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1278 if(env->flags & CO_IFG_DUMP_COLORS)
1279 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1282 static void ifg_dump_at_end(FILE *file, void *self)
1284 co_ifg_dump_t *env = self;
1287 co_gs_foreach_aff_node(env->co, a) {
1288 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1289 unsigned aidx = get_irn_idx(a->irn);
1292 co_gs_foreach_neighb(a, n) {
1293 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1294 unsigned nidx = get_irn_idx(n->irn);
1297 const char *color = nr == ar ? "blue" : "red";
1298 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1299 if(env->flags & CO_IFG_DUMP_LABELS)
1300 fprintf(file, "label=\"%d\" ", n->costs);
1301 if(env->flags & CO_IFG_DUMP_COLORS)
1302 fprintf(file, "color=%s ", color);
1304 fprintf(file, "style=dotted");
1305 fprintf(file, "];\n");
1312 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1314 ifg_dump_graph_attr,
1323 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1329 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1333 void co_solve_park_moon(copy_opt_t *opt)
1338 static int void_algo(copy_opt_t *co)
1345 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1346 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1347 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1348 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1355 int can_improve_existing;
1358 static co_algo_info_t algos[] = {
1359 { void_algo, "none", 0 },
1360 { co_solve_heuristic, "heur1", 0 },
1361 { co_solve_heuristic_new, "heur2", 0 },
1362 { co_solve_heuristic_java, "heur3", 0 },
1364 { co_solve_ilp2, "ilp", 1 },
1371 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1372 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1373 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1374 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1378 void co_driver(be_chordal_env_t *cenv)
1380 co_complete_stats_t before, after;
1382 co_algo_t *algo_func;
1383 int was_optimal = 0;
1385 if(algo < 0 || algo >= CO_ALGO_LAST)
1388 co = new_copy_opt(cenv, cost_func);
1389 co_build_ou_structure(co);
1390 co_build_graph_structure(co);
1392 co_complete_stats(co, &before);
1394 be_stat_ev("co_aff_nodes", before.aff_nodes);
1395 be_stat_ev("co_aff_edges", before.aff_edges);
1396 be_stat_ev("co_max_costs", before.max_costs);
1397 be_stat_ev("co_inevit_costs", before.inevit_costs);
1398 be_stat_ev("co_aff_int", before.aff_int);
1400 be_stat_ev("co_init_costs", before.costs);
1401 be_stat_ev("co_init_unsat", before.unsatisfied_edges);
1403 /* Dump the interference graph in Appel's format. */
1404 if(dump_flags & DUMP_APPEL) {
1405 FILE *f = be_chordal_open(cenv, "", ".apl");
1406 co_dump_appel_graph(co, f);
1410 if(dump_flags & DUMP_BEFORE) {
1411 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1412 co_dump_ifg_dot(co, f, style_flags);
1416 /* if the algo can improve results, provide an initial solution with heur3 */
1417 if(improve && algos[algo].can_improve_existing) {
1418 co_complete_stats_t stats;
1420 /* produce a heuristical solution */
1421 co_solve_heuristic_java(co);
1423 /* do the stats and provide the current costs */
1424 co_complete_stats(co, &stats);
1425 be_stat_ev("co_prepare_costs", stats.costs);
1428 algo_func = algos[algo].algo;
1429 was_optimal = algo_func(co);
1430 be_stat_ev("co_optimal", was_optimal);
1432 if(dump_flags & DUMP_AFTER) {
1433 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1434 co_dump_ifg_dot(co, f, style_flags);
1438 co_complete_stats(co, &after);
1441 int optimizable_costs = after.max_costs - after.inevit_costs;
1442 int evitable = after.costs - after.inevit_costs;
1444 ir_printf("%30F %10s %10d%10d%10d", cenv->irg, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1446 if(optimizable_costs > 0)
1447 printf("%10d %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1449 printf("%10d %5s\n", after.costs, "-");
1452 be_stat_ev("co_after_costs", after.costs);
1453 be_stat_ev("co_after_unsat", after.unsatisfied_edges);
1455 co_free_graph_structure(co);
1456 co_free_ou_structure(co);