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;
60 static const lc_opt_enum_mask_items_t dump_items[] = {
61 { "before", DUMP_BEFORE },
62 { "after", DUMP_AFTER },
63 { "appel", DUMP_APPEL },
68 static const lc_opt_enum_mask_items_t style_items[] = {
69 { "color", CO_IFG_DUMP_COLORS },
70 { "labels", CO_IFG_DUMP_LABELS },
71 { "constr", CO_IFG_DUMP_CONSTR },
72 { "shape", CO_IFG_DUMP_SHAPE },
73 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
77 static const lc_opt_enum_mask_items_t algo_items[] = {
78 { "none", CO_ALGO_NONE },
79 { "heur", CO_ALGO_HEUR },
80 { "heur2", CO_ALGO_HEUR2 },
81 { "heur3", CO_ALGO_HEUR3 },
82 { "ilp", CO_ALGO_ILP },
86 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
87 { "freq", co_get_costs_exec_freq },
88 { "loop", co_get_costs_loop_depth },
89 { "one", co_get_costs_all_one },
93 static lc_opt_enum_mask_var_t dump_var = {
94 &dump_flags, dump_items
97 static lc_opt_enum_mask_var_t style_var = {
98 &style_flags, style_items
101 static lc_opt_enum_mask_var_t algo_var = {
105 static lc_opt_enum_func_ptr_var_t cost_func_var = {
106 &cost_func, cost_func_items
109 static const lc_opt_table_entry_t options[] = {
110 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo (heur, heur2, heur3, ilp)", &algo_var),
111 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function (freq, loop, one)", &cost_func_var),
112 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
113 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
114 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
118 /* Insert additional options registration functions here. */
119 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
120 extern void be_co2_register_options(lc_opt_entry_t *grp);
121 extern void be_co3_register_options(lc_opt_entry_t *grp);
123 void co_register_options(lc_opt_entry_t *grp)
125 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
126 lc_opt_add_table(co_grp, options);
128 be_co2_register_options(co_grp);
129 be_co3_register_options(co_grp);
131 be_co_ilp_register_options(co_grp);
137 #undef QUICK_AND_DIRTY_HACK
139 /******************************************************************************
142 | | __ ___ _ __ ___ _ __ __ _| |
143 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
144 | |__| | __/ | | | __/ | | (_| | |
145 \_____|\___|_| |_|\___|_| \__,_|_|
147 ******************************************************************************/
149 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
151 void be_copy_opt_init(void) {
154 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
156 const char *s1, *s2, *s3;
160 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
162 co = xcalloc(1, sizeof(*co));
163 co->cenv = chordal_env;
164 co->aenv = chordal_env->birg->main_env->arch_env;
165 co->irg = chordal_env->irg;
166 co->cls = chordal_env->cls;
167 co->get_costs = get_costs;
169 s1 = get_irp_prog_name();
170 s2 = get_entity_name(get_irg_entity(co->irg));
171 s3 = chordal_env->cls->name;
172 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
173 co->name = xmalloc(len);
174 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
179 void free_copy_opt(copy_opt_t *co) {
184 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
185 arch_register_req_t req;
186 const arch_register_t *reg;
188 if (arch_irn_is(co->aenv, irn, ignore))
191 reg = arch_get_irn_register(co->aenv, irn);
192 if (arch_register_type_is(reg, ignore))
195 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
201 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
202 const ir_edge_t *edge;
203 const arch_register_t *reg;
205 assert(0 && "Is buggy and obsolete. Do not use");
207 if (arch_irn_is(co->aenv, irn, ignore))
210 reg = arch_get_irn_register(co->aenv, irn);
211 if (arch_register_type_is(reg, ignore))
214 foreach_out_edge(irn, edge) {
215 ir_node *n = edge->src;
217 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
218 arch_register_req_t req;
219 arch_get_register_req(co->aenv, &req, n, -1);
222 is_Perm(co->aenv, n) ||
223 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
232 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
235 ir_node *root_block = get_nodes_block(root);
238 /* for phis the copies are placed in the corresponding pred-block */
239 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
241 /* a perm places the copy in the same block as it resides */
242 loop = get_irn_loop(root_block);
245 int d = get_loop_depth(loop);
251 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
253 ir_node *root_bl = get_nodes_block(root);
254 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
255 res = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
257 /* don't allow values smaller than one. */
258 return res < 1 ? 1 : res;
262 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
266 /******************************************************************************
267 ____ _ _ _ _ _ _____ _
268 / __ \ | | | | | | (_) | / ____| |
269 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
270 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
271 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
272 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
275 ******************************************************************************/
278 * Determines a maximum weighted independent set with respect to
279 * the interference and conflict edges of all nodes in a qnode.
281 static int ou_max_ind_set_costs(unit_t *ou) {
282 be_chordal_env_t *chordal_env = ou->co->cenv;
283 ir_node **safe, **unsafe;
284 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
286 int max, pos, curr_weight, best_weight = 0;
288 /* assign the nodes into two groups.
289 * safe: node has no interference, hence it is in every max stable set.
290 * unsafe: node has an interference
292 safe = alloca((ou->node_count-1) * sizeof(*safe));
295 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
296 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
298 for(i=1; i<ou->node_count; ++i) {
300 for(o=1; o<ou->node_count; ++o) {
303 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
304 unsafe_costs[unsafe_count] = ou->costs[i];
305 unsafe[unsafe_count] = ou->nodes[i];
312 safe_costs += ou->costs[i];
313 safe[safe_count++] = ou->nodes[i];
318 /* now compute the best set out of the unsafe nodes*/
319 if (unsafe_count > MIS_HEUR_TRIGGER) {
320 bitset_t *best = bitset_alloca(unsafe_count);
321 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
322 for (i=0; i<unsafe_count; ++i) {
324 /* check if it is a stable set */
325 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
326 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
327 bitset_clear(best, i); /* clear the bit and try next one */
331 /* compute the weight */
332 bitset_foreach(best, pos)
333 best_weight += unsafe_costs[pos];
335 /* Exact Algorithm: Brute force */
336 curr = bitset_alloca(unsafe_count);
337 bitset_set_all(curr);
338 while ((max = bitset_popcnt(curr)) != 0) {
339 /* check if curr is a stable set */
340 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
341 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) */
342 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
345 /* if we arrive here, we have a stable set */
346 /* compute the weigth of the stable set*/
348 bitset_foreach(curr, pos)
349 curr_weight += unsafe_costs[pos];
352 if (curr_weight > best_weight) {
353 best_weight = curr_weight;
361 return safe_costs+best_weight;
364 static void co_collect_units(ir_node *irn, void *env) {
365 copy_opt_t *co = env;
367 arch_register_req_t req;
369 if (!is_curr_reg_class(co, irn))
371 if (!co_is_optimizable_root(co, irn))
374 /* Init a new unit */
375 unit = xcalloc(1, sizeof(*unit));
377 unit->node_count = 1;
378 INIT_LIST_HEAD(&unit->queue);
380 /* Phi with some/all of its arguments */
381 if (is_Reg_Phi(irn)) {
385 arity = get_irn_arity(irn);
386 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
387 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
388 unit->nodes[0] = irn;
391 for (i=0; i<arity; ++i) {
393 ir_node *arg = get_irn_n(irn, i);
395 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
398 if (nodes_interfere(co->cenv, irn, arg)) {
399 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
403 /* Else insert the argument of the phi to the members of this ou */
404 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
406 /* Check if arg has occurred at a prior position in the arg/list */
408 for (o=0; o<unit->node_count; ++o)
409 if (unit->nodes[o] == arg) {
414 if (!arg_pos) { /* a new argument */
415 /* insert node, set costs */
416 unit->nodes[unit->node_count] = arg;
417 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
419 } else { /* arg has occured before in same phi */
420 /* increase costs for existing arg */
421 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
424 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
425 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
428 /* Proj of a perm with corresponding arg */
429 if (is_Perm_Proj(co->aenv, irn)) {
430 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
431 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
432 unit->costs = xmalloc(2 * sizeof(*unit->costs));
433 unit->node_count = 2;
434 unit->nodes[0] = irn;
435 unit->nodes[1] = get_Perm_src(irn);
436 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
439 /* Src == Tgt of a 2-addr-code instruction */
440 if (is_2addr_code(co->aenv, irn, &req)) {
441 ir_node *other = req.other_same;
442 if (!nodes_interfere(co->cenv, irn, other)) {
443 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
444 unit->costs = xmalloc(2 * sizeof(*unit->costs));
445 unit->node_count = 2;
446 unit->nodes[0] = irn;
447 unit->nodes[1] = other;
448 unit->costs[1] = co->get_costs(co, irn, other, -1);
451 assert(0 && "This is not an optimizable node!");
453 /* Insert the new unit at a position according to its costs */
454 if (unit->node_count > 1) {
456 struct list_head *tmp;
458 /* Determine the maximum costs this unit can cause: all_nodes_cost */
459 for(i=1; i<unit->node_count; ++i) {
460 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
461 unit->all_nodes_costs += unit->costs[i];
464 /* Determine the minimal costs this unit will cause: min_nodes_costs */
465 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
466 /* Insert the new ou according to its sort_key */
468 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
470 list_add(&unit->units, tmp);
476 #ifdef QUICK_AND_DIRTY_HACK
478 static int compare_ous(const void *k1, const void *k2) {
479 const unit_t *u1 = *((const unit_t **) k1);
480 const unit_t *u2 = *((const unit_t **) k2);
481 int i, o, u1_has_constr, u2_has_constr;
482 arch_register_req_t req;
483 const arch_env_t *aenv = u1->co->aenv;
485 /* Units with constraints come first */
487 for (i=0; i<u1->node_count; ++i) {
488 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
489 if (arch_register_req_is(&req, limited)) {
496 for (i=0; i<u2->node_count; ++i) {
497 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
498 if (arch_register_req_is(&req, limited)) {
504 if (u1_has_constr != u2_has_constr)
505 return u2_has_constr - u1_has_constr;
507 /* Now check, whether the two units are connected */
509 for (i=0; i<u1->node_count; ++i)
510 for (o=0; o<u2->node_count; ++o)
511 if (u1->nodes[i] == u2->nodes[o])
515 /* After all, the sort key decides. Greater keys come first. */
516 return u2->sort_key - u1->sort_key;
521 * Sort the ou's according to constraints and their sort_key
523 static void co_sort_units(copy_opt_t *co) {
524 int i, count = 0, costs;
527 /* get the number of ous, remove them form the list and fill the array */
528 list_for_each_entry(unit_t, ou, &co->units, units)
530 ous = alloca(count * sizeof(*ous));
532 costs = co_get_max_copy_costs(co);
535 list_for_each_entry(unit_t, ou, &co->units, units)
538 INIT_LIST_HEAD(&co->units);
540 assert(count == i && list_empty(&co->units));
542 for (i=0; i<count; ++i)
543 ir_printf("%+F\n", ous[i]->nodes[0]);
545 qsort(ous, count, sizeof(*ous), compare_ous);
548 for (i=0; i<count; ++i)
549 ir_printf("%+F\n", ous[i]->nodes[0]);
551 /* reinsert into list in correct order */
552 for (i=0; i<count; ++i)
553 list_add_tail(&ous[i]->units, &co->units);
555 assert(costs == co_get_max_copy_costs(co));
559 void co_build_ou_structure(copy_opt_t *co) {
560 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
561 INIT_LIST_HEAD(&co->units);
562 irg_walk_graph(co->irg, co_collect_units, NULL, co);
563 #ifdef QUICK_AND_DIRTY_HACK
568 void co_free_ou_structure(copy_opt_t *co) {
571 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
576 co->units.next = NULL;
579 /* co_solve_heuristic() is implemented in becopyheur.c */
581 int co_get_max_copy_costs(const copy_opt_t *co) {
587 list_for_each_entry(unit_t, curr, &co->units, units) {
588 res += curr->inevitable_costs;
589 for (i=1; i<curr->node_count; ++i)
590 res += curr->costs[i];
595 int co_get_inevit_copy_costs(const copy_opt_t *co) {
601 list_for_each_entry(unit_t, curr, &co->units, units)
602 res += curr->inevitable_costs;
606 int co_get_copy_costs(const copy_opt_t *co) {
612 list_for_each_entry(unit_t, curr, &co->units, units) {
613 int root_col = get_irn_col(co, curr->nodes[0]);
614 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
615 res += curr->inevitable_costs;
616 for (i=1; i<curr->node_count; ++i) {
617 int arg_col = get_irn_col(co, curr->nodes[i]);
618 if (root_col != arg_col) {
619 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
620 res += curr->costs[i];
627 int co_get_lower_bound(const copy_opt_t *co) {
633 list_for_each_entry(unit_t, curr, &co->units, units)
634 res += curr->inevitable_costs + curr->min_nodes_costs;
638 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
640 bitset_t *seen = bitset_irg_malloc(co->irg);
643 memset(stat, 0, sizeof(stat[0]));
645 /* count affinity edges. */
646 co_gs_foreach_aff_node(co, an) {
648 stat->aff_nodes += 1;
649 bitset_add_irn(seen, an->irn);
650 co_gs_foreach_neighb(an, neigh) {
651 if(!bitset_contains_irn(seen, neigh->irn)) {
652 stat->aff_edges += 1;
653 stat->max_costs += neigh->costs;
655 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
656 stat->costs += neigh->costs;
657 stat->unsatisfied_edges += 1;
660 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
662 stat->inevit_costs += neigh->costs;
672 /******************************************************************************
674 / ____| | | / ____| |
675 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
676 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
677 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
678 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
681 ******************************************************************************/
683 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
684 const affinity_node_t *n1 = k1;
685 const affinity_node_t *n2 = k2;
687 return (n1->irn != n2->irn);
690 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
691 affinity_node_t new_node, *node;
692 neighb_t new_nbr, *nbr;
697 new_node.neighbours = NULL;
698 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
701 for (nbr = node->neighbours; nbr; nbr = nbr->next)
702 if (nbr->irn == n2) {
707 /* if we did not find n2 in n1's neighbourhood insert it */
709 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
710 nbr = obstack_finish(&co->obst);
713 nbr->next = node->neighbours;
714 node->neighbours = nbr;
718 /* now nbr points to n1's neighbour-entry of n2 */
722 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
723 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
724 add_edge(co, n1, n2, costs);
725 add_edge(co, n2, n1, costs);
729 static void build_graph_walker(ir_node *irn, void *env) {
730 copy_opt_t *co = env;
732 arch_register_req_t req;
733 const arch_register_t *reg;
735 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
738 reg = arch_get_irn_register(co->aenv, irn);
739 if (arch_register_type_is(reg, ignore))
744 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
745 ir_node *arg = get_irn_n(irn, pos);
746 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
750 else if (is_Perm_Proj(co->aenv, irn)) {
751 ir_node *arg = get_Perm_src(irn);
752 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
756 else if (is_2addr_code(co->aenv, irn, &req))
757 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
760 void co_build_graph_structure(copy_opt_t *co) {
761 obstack_init(&co->obst);
762 co->nodes = new_set(compare_affinity_node_t, 32);
764 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
767 void co_free_graph_structure(copy_opt_t *co) {
771 obstack_free(&co->obst, NULL);
775 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
777 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
778 affinity_node_t new_node, *n;
783 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
785 return (n->degree > 0);
790 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
792 be_ifg_t *ifg = co->cenv->ifg;
793 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
794 bitset_t *adm = bitset_alloca(co->cls->n_regs);
801 for(i = 0; i < co->cls->n_regs; ++i) {
802 const arch_register_t *reg = &co->cls->regs[i];
803 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
807 * n contains the first node number.
808 * the values below n are the pre-colored register nodes
811 it = be_ifg_nodes_iter_alloca(ifg);
812 nit = be_ifg_neighbours_iter_alloca(ifg);
815 be_ifg_foreach_node(ifg, it, irn) {
816 if(!arch_irn_is(co->aenv, irn, ignore))
817 set_irn_link(irn, INT_TO_PTR(n++));
820 fprintf(f, "%d %d\n", n, n_regs);
822 be_ifg_foreach_node(ifg, it, irn) {
823 if(!arch_irn_is(co->aenv, irn, ignore)) {
824 int idx = PTR_TO_INT(get_irn_link(irn));
825 affinity_node_t *a = get_affinity_info(co, irn);
827 arch_register_req_t req;
830 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
831 if(arch_register_req_is(&req, limited)) {
832 bitset_clear_all(adm);
833 req.limited(req.limited_env, adm);
834 for(i = 0; i < co->cls->n_regs; ++i)
835 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
836 fprintf(f, "%d %d -1\n", color_map[i], idx);
841 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
842 if(!arch_irn_is(co->aenv, adj, ignore)) {
843 int adj_idx = PTR_TO_INT(get_irn_link(adj));
845 fprintf(f, "%d %d -1\n", idx, adj_idx);
852 co_gs_foreach_neighb(a, n) {
853 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
854 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
856 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
864 typedef struct _appel_clique_walker_t {
866 const copy_opt_t *co;
873 } appel_clique_walker_t;
875 typedef struct _appel_block_info_t {
885 } appel_block_info_t;
887 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
890 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
891 int res = (int) freq;
892 return res == 0 ? 1 : res;
894 ir_loop *loop = get_irn_loop(bl);
896 int d = get_loop_depth(loop);
903 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
905 appel_block_info_t *res = NULL;
908 appel_clique_walker_t *d = (void *) phase;
909 res = phase_alloc(phase, sizeof(res[0]));
910 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
911 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
912 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
913 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
914 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
915 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
921 typedef struct _insn_list_t {
923 struct list_head list;
926 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
928 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
931 for(i = 0; i < bli->n_live_end; ++i)
932 if(bli->live_end[i] == irn)
933 return bli->live_end_nr[i];
938 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
940 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
946 foreach_pset(live, irn)
947 live_arr[n_live++] = irn;
949 /* dump the live after clique */
951 for(j = 0; j < n_live; ++j) {
954 for(k = j + 1; k < n_live; ++k) {
955 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
957 fprintf(env->f, "\n");
961 /* dump the affinities */
962 for(j = 0; !env->dumb && j < n_live; ++j) {
963 ir_node *irn = live_arr[j];
964 int old_nr = PTR_TO_INT(get_irn_link(irn));
966 /* if the node was already live in the last insn dump the affinity */
967 if(old_nr > start_nr) {
968 int weight = appel_aff_weight(env, bl);
969 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
973 /* set the current numbers into the link field. */
974 for(j = 0; j < n_live; ++j) {
975 ir_node *irn = live_arr[j];
976 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
979 return curr_nr + n_live;
982 static void appel_walker(ir_node *bl, void *data)
984 appel_clique_walker_t *env = data;
985 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
986 struct obstack *obst = &env->obst;
987 void *base = obstack_base(obst);
988 pset *live = pset_new_ptr_default();
992 int start_nr = env->curr_nr;
993 int curr_nr = start_nr;
995 be_insn_env_t insn_env;
1000 insn_env.aenv = env->co->aenv;
1001 insn_env.cls = env->co->cls;
1002 insn_env.obst = obst;
1003 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1005 /* Guess how many insns will be in this block. */
1006 sched_foreach(bl, irn)
1010 insns = malloc(n_nodes * sizeof(insns[0]));
1012 /* Put all insns in an array. */
1013 irn = sched_first(bl);
1014 while(!sched_is_end(irn)) {
1016 insn = be_scan_insn(&insn_env, irn);
1017 insns[n_insns++] = insn;
1018 irn = insn->next_insn;
1021 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1022 be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
1024 /* Generate the bad and ugly. */
1025 for(i = n_insns - 1; i >= 0; --i) {
1026 be_insn_t *insn = insns[i];
1028 /* The first live set has to be saved in the block border set. */
1029 if(i == n_insns - 1) {
1031 foreach_pset(live, irn) {
1032 bli->live_end[j] = irn;
1033 bli->live_end_nr[j] = curr_nr + j;
1036 bli->n_live_end = j;
1040 for(j = 0; j < insn->use_start; ++j) {
1041 ir_node *op = insn->ops[j].carrier;
1042 bitset_t *adm = insn->ops[j].regs;
1046 if(!insn->ops[j].has_constraints)
1050 foreach_pset(live, irn) {
1058 assert(nr < pset_count(live));
1060 for(k = 0; k < env->co->cls->n_regs; ++k) {
1061 int mapped_col = env->color_map[k];
1062 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1063 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1068 /* dump the clique and update the stuff. */
1069 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1071 /* remove all defs. */
1072 for(j = 0; j < insn->use_start; ++j)
1073 pset_remove_ptr(live, insn->ops[j].carrier);
1075 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1076 bli->phi[bli->n_phi] = insn->irn;
1077 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1083 for(j = insn->use_start; j < insn->n_ops; ++j)
1084 pset_insert_ptr(live, insn->ops[j].carrier);
1087 /* print the start clique. */
1088 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1091 foreach_pset(live, irn) {
1092 bli->live_in[i] = irn;
1093 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1100 obstack_free(obst, base);
1101 env->curr_nr = curr_nr;
1104 static void appel_inter_block_aff(ir_node *bl, void *data)
1106 appel_clique_walker_t *env = data;
1107 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1111 for(i = 0; i < bli->n_live_in; ++i) {
1112 ir_node *irn = bli->live_in[i];
1114 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1115 ir_node *pred = get_Block_cfgpred_block(bl, j);
1117 int nr = appel_get_live_end_nr(env, pred, irn);
1119 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1123 for(i = 0; i < bli->n_phi; ++i) {
1124 ir_node *irn = bli->phi[i];
1126 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1127 ir_node *pred = get_Block_cfgpred_block(bl, j);
1128 ir_node *op = get_irn_n(irn, j);
1130 int nr = appel_get_live_end_nr(env, pred, op);
1132 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1138 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1142 appel_clique_walker_t env;
1143 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1145 be_liveness_recompute(co->cenv->lv);
1146 obstack_init(&env.obst);
1147 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1148 env.curr_nr = co->cls->n_regs;
1152 bitset_copy(adm, co->cenv->ignore_colors);
1153 bitset_flip_all(adm);
1155 /* Make color map. */
1156 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1157 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1158 const arch_register_t *reg = &co->cls->regs[i];
1159 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1163 env.curr_nr = n_colors;
1164 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1165 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1167 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1169 /* make the first k nodes interfere */
1170 for(i = 0; i < n_colors; ++i) {
1172 for(j = i + 1; j < n_colors; ++j)
1173 fprintf(f, "%d %d -1 ", i, j);
1178 env.curr_nr = n_colors;
1179 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1180 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1181 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1182 obstack_free(&env.obst, NULL);
1186 ___ _____ ____ ____ ___ _____ ____ _
1187 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1188 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1189 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1190 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1194 static const char *get_dot_color_name(int col)
1196 static const char *names[] = {
1230 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1233 typedef struct _co_ifg_dump_t {
1234 const copy_opt_t *co;
1238 static void ifg_dump_graph_attr(FILE *f, void *self)
1240 fprintf(f, "overlap=scale");
1243 static int ifg_is_dump_node(void *self, ir_node *irn)
1245 co_ifg_dump_t *cod = self;
1246 return !arch_irn_is(cod->co->aenv, irn, ignore);
1249 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1251 co_ifg_dump_t *env = self;
1252 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1253 arch_register_req_t req;
1256 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1257 limited = arch_register_req_is(&req, limited);
1259 if(env->flags & CO_IFG_DUMP_LABELS) {
1260 ir_fprintf(f, "label=\"%+F", irn);
1262 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1263 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1264 req.limited(req.limited_env, bs);
1265 ir_fprintf(f, "\\n%B", bs);
1267 ir_fprintf(f, "\" ");
1271 fprintf(f, "label=\"\" shape=point " );
1273 if(env->flags & CO_IFG_DUMP_SHAPE)
1274 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1276 if(env->flags & CO_IFG_DUMP_COLORS)
1277 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1280 static void ifg_dump_at_end(FILE *file, void *self)
1282 co_ifg_dump_t *env = self;
1285 co_gs_foreach_aff_node(env->co, a) {
1286 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1287 unsigned aidx = get_irn_idx(a->irn);
1290 co_gs_foreach_neighb(a, n) {
1291 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1292 unsigned nidx = get_irn_idx(n->irn);
1295 const char *color = nr == ar ? "blue" : "red";
1296 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1297 if(env->flags & CO_IFG_DUMP_LABELS)
1298 fprintf(file, "label=\"%d\" ", n->costs);
1299 if(env->flags & CO_IFG_DUMP_COLORS)
1300 fprintf(file, "color=%s ", color);
1302 fprintf(file, "style=dotted");
1303 fprintf(file, "];\n");
1310 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1312 ifg_dump_graph_attr,
1321 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1327 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1331 void co_solve_park_moon(copy_opt_t *opt)
1336 static int void_algo(copy_opt_t *co)
1343 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1344 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1345 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1346 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1350 static co_algo_t *algos[] = {
1353 co_solve_heuristic_new,
1354 co_solve_heuristic_java,
1362 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1363 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1364 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1365 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1369 void co_driver(be_chordal_env_t *cenv)
1371 co_complete_stats_t before, after;
1373 co_algo_t *algo_func;
1374 int was_optimal = 0;
1376 if(algo < 0 || algo >= CO_ALGO_LAST)
1379 co = new_copy_opt(cenv, cost_func);
1380 co_build_ou_structure(co);
1381 co_build_graph_structure(co);
1383 co_complete_stats(co, &before);
1385 be_stat_ev("co_aff_nodes", before.aff_nodes);
1386 be_stat_ev("co_aff_edges", before.aff_edges);
1387 be_stat_ev("co_max_costs", before.max_costs);
1388 be_stat_ev("co_inevit_costs", before.inevit_costs);
1389 be_stat_ev("co_aff_int", before.aff_int);
1391 be_stat_ev("co_init_costs", before.costs);
1392 be_stat_ev("co_init_unsat", before.unsatisfied_edges);
1394 /* Dump the interference graph in Appel's format. */
1395 if(dump_flags & DUMP_APPEL) {
1396 FILE *f = be_chordal_open(cenv, "", ".apl");
1397 co_dump_appel_graph(co, f);
1401 if(dump_flags & DUMP_BEFORE) {
1402 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1403 co_dump_ifg_dot(co, f, style_flags);
1407 algo_func = algos[algo];
1408 was_optimal = algo_func(co);
1409 be_stat_ev("co_optimal", was_optimal);
1411 if(dump_flags & DUMP_AFTER) {
1412 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1413 co_dump_ifg_dot(co, f, style_flags);
1417 co_complete_stats(co, &after);
1420 int optimizable_costs = after.max_costs - after.inevit_costs;
1421 int evitable = after.costs - after.inevit_costs;
1423 ir_printf("%30F %10s %10d%10d%10d", cenv->irg, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1425 if(optimizable_costs > 0)
1426 printf("%10d %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1428 printf("%10d %5s\n", after.costs, "-");
1431 be_stat_ev("co_after_costs", after.costs);
1432 be_stat_ev("co_after_unsat", after.unsatisfied_edges);
1434 co_free_graph_structure(co);
1435 co_free_ou_structure(co);