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"
46 #define DUMP_ALL 2 * DUMP_APPEL - 1
48 #define COST_FUNC_FREQ 1
49 #define COST_FUNC_LOOP 2
50 #define COST_FUNC_ALL_ONE 3
52 static int dump_flags = 0;
53 static int style_flags = 0;
54 static int do_stats = 0;
55 static cost_fct_t cost_func = co_get_costs_exec_freq;
56 static int algo = CO_ALGO_HEUR2;
59 static const lc_opt_enum_mask_items_t dump_items[] = {
60 { "before", DUMP_BEFORE },
61 { "after", DUMP_AFTER },
62 { "appel", DUMP_APPEL },
67 static const lc_opt_enum_mask_items_t style_items[] = {
68 { "color", CO_IFG_DUMP_COLORS },
69 { "labels", CO_IFG_DUMP_LABELS },
70 { "constr", CO_IFG_DUMP_CONSTR },
71 { "shape", CO_IFG_DUMP_SHAPE },
72 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
76 static const lc_opt_enum_mask_items_t algo_items[] = {
77 { "none", CO_ALGO_NONE },
78 { "heur", CO_ALGO_HEUR },
79 { "heur2", CO_ALGO_HEUR2 },
80 { "heur3", CO_ALGO_HEUR3 },
81 { "ilp", CO_ALGO_ILP },
85 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
86 { "freq", co_get_costs_exec_freq },
87 { "loop", co_get_costs_loop_depth },
88 { "one", co_get_costs_all_one },
92 static lc_opt_enum_mask_var_t dump_var = {
93 &dump_flags, dump_items
96 static lc_opt_enum_mask_var_t style_var = {
97 &style_flags, style_items
100 static lc_opt_enum_mask_var_t algo_var = {
104 static lc_opt_enum_func_ptr_var_t cost_func_var = {
105 &cost_func, cost_func_items
108 static const lc_opt_table_entry_t options[] = {
109 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo (heur, heur2, heur3, ilp)", &algo_var),
110 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function (freq, loop, one)", &cost_func_var),
111 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
112 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
113 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
117 /* Insert additional options registration functions here. */
118 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
119 extern void be_co2_register_options(lc_opt_entry_t *grp);
120 extern void be_co3_register_options(lc_opt_entry_t *grp);
122 void co_register_options(lc_opt_entry_t *grp)
124 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
125 lc_opt_add_table(co_grp, options);
127 be_co2_register_options(co_grp);
128 be_co3_register_options(co_grp);
130 be_co_ilp_register_options(co_grp);
136 #undef QUICK_AND_DIRTY_HACK
138 /******************************************************************************
141 | | __ ___ _ __ ___ _ __ __ _| |
142 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
143 | |__| | __/ | | | __/ | | (_| | |
144 \_____|\___|_| |_|\___|_| \__,_|_|
146 ******************************************************************************/
148 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
150 void be_copy_opt_init(void) {
153 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
155 const char *s1, *s2, *s3;
159 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
161 co = xcalloc(1, sizeof(*co));
162 co->cenv = chordal_env;
163 co->aenv = chordal_env->birg->main_env->arch_env;
164 co->irg = chordal_env->irg;
165 co->cls = chordal_env->cls;
166 co->get_costs = get_costs;
168 s1 = get_irp_prog_name();
169 s2 = get_entity_name(get_irg_entity(co->irg));
170 s3 = chordal_env->cls->name;
171 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
172 co->name = xmalloc(len);
173 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
178 void free_copy_opt(copy_opt_t *co) {
183 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
184 arch_register_req_t req;
185 const arch_register_t *reg;
187 if (arch_irn_is(co->aenv, irn, ignore))
190 reg = arch_get_irn_register(co->aenv, irn);
191 if (arch_register_type_is(reg, ignore))
194 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
200 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
201 const ir_edge_t *edge;
202 const arch_register_t *reg;
204 assert(0 && "Is buggy and obsolete. Do not use");
206 if (arch_irn_is(co->aenv, irn, ignore))
209 reg = arch_get_irn_register(co->aenv, irn);
210 if (arch_register_type_is(reg, ignore))
213 foreach_out_edge(irn, edge) {
214 ir_node *n = edge->src;
216 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
217 arch_register_req_t req;
218 arch_get_register_req(co->aenv, &req, n, -1);
221 is_Perm(co->aenv, n) ||
222 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
231 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
234 ir_node *root_block = get_nodes_block(root);
237 /* for phis the copies are placed in the corresponding pred-block */
238 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
240 /* a perm places the copy in the same block as it resides */
241 loop = get_irn_loop(root_block);
244 int d = get_loop_depth(loop);
250 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
251 ir_node *root_bl = get_nodes_block(root);
252 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
253 unsigned long freq = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
254 return freq > 0 ? (int) freq : 1;
258 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
262 /******************************************************************************
263 ____ _ _ _ _ _ _____ _
264 / __ \ | | | | | | (_) | / ____| |
265 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
266 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
267 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
268 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
271 ******************************************************************************/
274 * Determines a maximum weighted independent set with respect to
275 * the interference and conflict edges of all nodes in a qnode.
277 static int ou_max_ind_set_costs(unit_t *ou) {
278 be_chordal_env_t *chordal_env = ou->co->cenv;
279 ir_node **safe, **unsafe;
280 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
282 int max, pos, curr_weight, best_weight = 0;
284 /* assign the nodes into two groups.
285 * safe: node has no interference, hence it is in every max stable set.
286 * unsafe: node has an interference
288 safe = alloca((ou->node_count-1) * sizeof(*safe));
291 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
292 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
294 for(i=1; i<ou->node_count; ++i) {
296 for(o=1; o<ou->node_count; ++o) {
299 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
300 unsafe_costs[unsafe_count] = ou->costs[i];
301 unsafe[unsafe_count] = ou->nodes[i];
308 safe_costs += ou->costs[i];
309 safe[safe_count++] = ou->nodes[i];
314 /* now compute the best set out of the unsafe nodes*/
315 if (unsafe_count > MIS_HEUR_TRIGGER) {
316 bitset_t *best = bitset_alloca(unsafe_count);
317 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
318 for (i=0; i<unsafe_count; ++i) {
320 /* check if it is a stable set */
321 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
322 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
323 bitset_clear(best, i); /* clear the bit and try next one */
327 /* compute the weight */
328 bitset_foreach(best, pos)
329 best_weight += unsafe_costs[pos];
331 /* Exact Algorithm: Brute force */
332 curr = bitset_alloca(unsafe_count);
333 bitset_set_all(curr);
334 while ((max = bitset_popcnt(curr)) != 0) {
335 /* check if curr is a stable set */
336 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
337 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) */
338 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
341 /* if we arrive here, we have a stable set */
342 /* compute the weigth of the stable set*/
344 bitset_foreach(curr, pos)
345 curr_weight += unsafe_costs[pos];
348 if (curr_weight > best_weight) {
349 best_weight = curr_weight;
357 return safe_costs+best_weight;
360 static void co_collect_units(ir_node *irn, void *env) {
361 copy_opt_t *co = env;
363 arch_register_req_t req;
365 if (!is_curr_reg_class(co, irn))
367 if (!co_is_optimizable_root(co, irn))
370 /* Init a new unit */
371 unit = xcalloc(1, sizeof(*unit));
373 unit->node_count = 1;
374 INIT_LIST_HEAD(&unit->queue);
376 /* Phi with some/all of its arguments */
377 if (is_Reg_Phi(irn)) {
381 arity = get_irn_arity(irn);
382 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
383 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
384 unit->nodes[0] = irn;
387 for (i=0; i<arity; ++i) {
389 ir_node *arg = get_irn_n(irn, i);
391 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
394 if (nodes_interfere(co->cenv, irn, arg)) {
395 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
399 /* Else insert the argument of the phi to the members of this ou */
400 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
402 /* Check if arg has occurred at a prior position in the arg/list */
404 for (o=0; o<unit->node_count; ++o)
405 if (unit->nodes[o] == arg) {
410 if (!arg_pos) { /* a new argument */
411 /* insert node, set costs */
412 unit->nodes[unit->node_count] = arg;
413 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
415 } else { /* arg has occured before in same phi */
416 /* increase costs for existing arg */
417 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
420 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
421 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
424 /* Proj of a perm with corresponding arg */
425 if (is_Perm_Proj(co->aenv, irn)) {
426 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
427 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
428 unit->costs = xmalloc(2 * sizeof(*unit->costs));
429 unit->node_count = 2;
430 unit->nodes[0] = irn;
431 unit->nodes[1] = get_Perm_src(irn);
432 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
435 /* Src == Tgt of a 2-addr-code instruction */
436 if (is_2addr_code(co->aenv, irn, &req)) {
437 ir_node *other = req.other_same;
438 if (!nodes_interfere(co->cenv, irn, other)) {
439 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
440 unit->costs = xmalloc(2 * sizeof(*unit->costs));
441 unit->node_count = 2;
442 unit->nodes[0] = irn;
443 unit->nodes[1] = other;
444 unit->costs[1] = co->get_costs(co, irn, other, -1);
447 assert(0 && "This is not an optimizable node!");
449 /* Insert the new unit at a position according to its costs */
450 if (unit->node_count > 1) {
452 struct list_head *tmp;
454 /* Determine the maximum costs this unit can cause: all_nodes_cost */
455 for(i=1; i<unit->node_count; ++i) {
456 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
457 unit->all_nodes_costs += unit->costs[i];
460 /* Determine the minimal costs this unit will cause: min_nodes_costs */
461 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
462 /* Insert the new ou according to its sort_key */
464 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
466 list_add(&unit->units, tmp);
472 #ifdef QUICK_AND_DIRTY_HACK
474 static int compare_ous(const void *k1, const void *k2) {
475 const unit_t *u1 = *((const unit_t **) k1);
476 const unit_t *u2 = *((const unit_t **) k2);
477 int i, o, u1_has_constr, u2_has_constr;
478 arch_register_req_t req;
479 const arch_env_t *aenv = u1->co->aenv;
481 /* Units with constraints come first */
483 for (i=0; i<u1->node_count; ++i) {
484 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
485 if (arch_register_req_is(&req, limited)) {
492 for (i=0; i<u2->node_count; ++i) {
493 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
494 if (arch_register_req_is(&req, limited)) {
500 if (u1_has_constr != u2_has_constr)
501 return u2_has_constr - u1_has_constr;
503 /* Now check, whether the two units are connected */
505 for (i=0; i<u1->node_count; ++i)
506 for (o=0; o<u2->node_count; ++o)
507 if (u1->nodes[i] == u2->nodes[o])
511 /* After all, the sort key decides. Greater keys come first. */
512 return u2->sort_key - u1->sort_key;
517 * Sort the ou's according to constraints and their sort_key
519 static void co_sort_units(copy_opt_t *co) {
520 int i, count = 0, costs;
523 /* get the number of ous, remove them form the list and fill the array */
524 list_for_each_entry(unit_t, ou, &co->units, units)
526 ous = alloca(count * sizeof(*ous));
528 costs = co_get_max_copy_costs(co);
531 list_for_each_entry(unit_t, ou, &co->units, units)
534 INIT_LIST_HEAD(&co->units);
536 assert(count == i && list_empty(&co->units));
538 for (i=0; i<count; ++i)
539 ir_printf("%+F\n", ous[i]->nodes[0]);
541 qsort(ous, count, sizeof(*ous), compare_ous);
544 for (i=0; i<count; ++i)
545 ir_printf("%+F\n", ous[i]->nodes[0]);
547 /* reinsert into list in correct order */
548 for (i=0; i<count; ++i)
549 list_add_tail(&ous[i]->units, &co->units);
551 assert(costs == co_get_max_copy_costs(co));
555 void co_build_ou_structure(copy_opt_t *co) {
556 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
557 INIT_LIST_HEAD(&co->units);
558 irg_walk_graph(co->irg, co_collect_units, NULL, co);
559 #ifdef QUICK_AND_DIRTY_HACK
564 void co_free_ou_structure(copy_opt_t *co) {
567 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
572 co->units.next = NULL;
575 /* co_solve_heuristic() is implemented in becopyheur.c */
577 int co_get_max_copy_costs(const copy_opt_t *co) {
583 list_for_each_entry(unit_t, curr, &co->units, units) {
584 res += curr->inevitable_costs;
585 for (i=1; i<curr->node_count; ++i)
586 res += curr->costs[i];
591 int co_get_inevit_copy_costs(const copy_opt_t *co) {
597 list_for_each_entry(unit_t, curr, &co->units, units)
598 res += curr->inevitable_costs;
602 int co_get_copy_costs(const copy_opt_t *co) {
608 list_for_each_entry(unit_t, curr, &co->units, units) {
609 int root_col = get_irn_col(co, curr->nodes[0]);
610 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
611 res += curr->inevitable_costs;
612 for (i=1; i<curr->node_count; ++i) {
613 int arg_col = get_irn_col(co, curr->nodes[i]);
614 if (root_col != arg_col) {
615 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
616 res += curr->costs[i];
623 int co_get_lower_bound(const copy_opt_t *co) {
629 list_for_each_entry(unit_t, curr, &co->units, units)
630 res += curr->inevitable_costs + curr->min_nodes_costs;
634 /******************************************************************************
636 / ____| | | / ____| |
637 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
638 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
639 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
640 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
643 ******************************************************************************/
645 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
646 const affinity_node_t *n1 = k1;
647 const affinity_node_t *n2 = k2;
649 return (n1->irn != n2->irn);
652 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
653 affinity_node_t new_node, *node;
654 neighb_t new_nbr, *nbr;
659 new_node.neighbours = NULL;
660 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
663 for (nbr = node->neighbours; nbr; nbr = nbr->next)
664 if (nbr->irn == n2) {
669 /* if we did not find n2 in n1's neighbourhood insert it */
671 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
672 nbr = obstack_finish(&co->obst);
675 nbr->next = node->neighbours;
676 node->neighbours = nbr;
680 /* now nbr points to n1's neighbour-entry of n2 */
684 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
685 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
686 add_edge(co, n1, n2, costs);
687 add_edge(co, n2, n1, costs);
691 static void build_graph_walker(ir_node *irn, void *env) {
692 copy_opt_t *co = env;
694 arch_register_req_t req;
695 const arch_register_t *reg;
697 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
700 reg = arch_get_irn_register(co->aenv, irn);
701 if (arch_register_type_is(reg, ignore))
706 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
707 ir_node *arg = get_irn_n(irn, pos);
708 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
712 else if (is_Perm_Proj(co->aenv, irn)) {
713 ir_node *arg = get_Perm_src(irn);
714 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
718 else if (is_2addr_code(co->aenv, irn, &req))
719 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
722 void co_build_graph_structure(copy_opt_t *co) {
723 obstack_init(&co->obst);
724 co->nodes = new_set(compare_affinity_node_t, 32);
726 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
729 void co_free_graph_structure(copy_opt_t *co) {
733 obstack_free(&co->obst, NULL);
737 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
739 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
740 affinity_node_t new_node, *n;
745 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
747 return (n->degree > 0);
752 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
754 be_ifg_t *ifg = co->cenv->ifg;
755 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
756 bitset_t *adm = bitset_alloca(co->cls->n_regs);
763 for(i = 0; i < co->cls->n_regs; ++i) {
764 const arch_register_t *reg = &co->cls->regs[i];
765 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
769 * n contains the first node number.
770 * the values below n are the pre-colored register nodes
773 it = be_ifg_nodes_iter_alloca(ifg);
774 nit = be_ifg_neighbours_iter_alloca(ifg);
777 be_ifg_foreach_node(ifg, it, irn) {
778 if(!arch_irn_is(co->aenv, irn, ignore))
779 set_irn_link(irn, INT_TO_PTR(n++));
782 fprintf(f, "%d %d\n", n, n_regs);
784 be_ifg_foreach_node(ifg, it, irn) {
785 if(!arch_irn_is(co->aenv, irn, ignore)) {
786 int idx = PTR_TO_INT(get_irn_link(irn));
787 affinity_node_t *a = get_affinity_info(co, irn);
789 arch_register_req_t req;
792 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
793 if(arch_register_req_is(&req, limited)) {
794 bitset_clear_all(adm);
795 req.limited(req.limited_env, adm);
796 for(i = 0; i < co->cls->n_regs; ++i)
797 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
798 fprintf(f, "%d %d -1\n", color_map[i], idx);
803 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
804 if(!arch_irn_is(co->aenv, adj, ignore)) {
805 int adj_idx = PTR_TO_INT(get_irn_link(adj));
807 fprintf(f, "%d %d -1\n", idx, adj_idx);
814 co_gs_foreach_neighb(a, n) {
815 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
816 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
818 fprintf(f, "%d %d %d\n", idx, n_idx, n->costs);
826 typedef struct _appel_clique_walker_t {
828 const copy_opt_t *co;
835 } appel_clique_walker_t;
837 typedef struct _appel_block_info_t {
847 } appel_block_info_t;
849 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
852 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
853 int res = (int) freq;
854 return res == 0 ? 1 : res;
856 ir_loop *loop = get_irn_loop(bl);
858 int d = get_loop_depth(loop);
865 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
867 appel_block_info_t *res = NULL;
870 appel_clique_walker_t *d = (void *) phase;
871 res = phase_alloc(phase, sizeof(res[0]));
872 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
873 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
874 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
875 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
876 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
877 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
883 typedef struct _insn_list_t {
885 struct list_head list;
888 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
890 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
893 for(i = 0; i < bli->n_live_end; ++i)
894 if(bli->live_end[i] == irn)
895 return bli->live_end_nr[i];
900 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
902 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
908 foreach_pset(live, irn)
909 live_arr[n_live++] = irn;
911 /* dump the live after clique */
913 for(j = 0; j < n_live; ++j) {
916 for(k = j + 1; k < n_live; ++k) {
917 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
919 fprintf(env->f, "\n");
923 /* dump the affinities */
924 for(j = 0; !env->dumb && j < n_live; ++j) {
925 ir_node *irn = live_arr[j];
926 int old_nr = PTR_TO_INT(get_irn_link(irn));
928 /* if the node was already live in the last insn dump the affinity */
929 if(old_nr > start_nr) {
930 int weight = appel_aff_weight(env, bl);
931 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
935 /* set the current numbers into the link field. */
936 for(j = 0; j < n_live; ++j) {
937 ir_node *irn = live_arr[j];
938 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
941 return curr_nr + n_live;
944 static void appel_walker(ir_node *bl, void *data)
946 appel_clique_walker_t *env = data;
947 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
948 struct obstack *obst = &env->obst;
949 void *base = obstack_base(obst);
950 pset *live = pset_new_ptr_default();
954 int start_nr = env->curr_nr;
955 int curr_nr = start_nr;
957 be_insn_env_t insn_env;
962 insn_env.aenv = env->co->aenv;
963 insn_env.cls = env->co->cls;
964 insn_env.obst = obst;
965 insn_env.ignore_colors = env->co->cenv->ignore_colors;
967 /* Guess how many insns will be in this block. */
968 sched_foreach(bl, irn)
972 insns = malloc(n_nodes * sizeof(insns[0]));
974 /* Put all insns in an array. */
975 irn = sched_first(bl);
976 while(!sched_is_end(irn)) {
978 insn = be_scan_insn(&insn_env, irn);
979 insns[n_insns++] = insn;
980 irn = insn->next_insn;
983 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
984 be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
986 /* Generate the bad and ugly. */
987 for(i = n_insns - 1; i >= 0; --i) {
988 be_insn_t *insn = insns[i];
990 /* The first live set has to be saved in the block border set. */
991 if(i == n_insns - 1) {
993 foreach_pset(live, irn) {
994 bli->live_end[j] = irn;
995 bli->live_end_nr[j] = curr_nr + j;
1002 for(j = 0; j < insn->use_start; ++j) {
1003 ir_node *op = insn->ops[j].carrier;
1004 bitset_t *adm = insn->ops[j].regs;
1008 if(!insn->ops[j].has_constraints)
1012 foreach_pset(live, irn) {
1020 assert(nr < pset_count(live));
1022 for(k = 0; k < env->co->cls->n_regs; ++k) {
1023 int mapped_col = env->color_map[k];
1024 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1025 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1030 /* dump the clique and update the stuff. */
1031 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1033 /* remove all defs. */
1034 for(j = 0; j < insn->use_start; ++j)
1035 pset_remove_ptr(live, insn->ops[j].carrier);
1037 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1038 bli->phi[bli->n_phi] = insn->irn;
1039 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1045 for(j = insn->use_start; j < insn->n_ops; ++j)
1046 pset_insert_ptr(live, insn->ops[j].carrier);
1049 /* print the start clique. */
1050 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1053 foreach_pset(live, irn) {
1054 bli->live_in[i] = irn;
1055 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1062 obstack_free(obst, base);
1063 env->curr_nr = curr_nr;
1066 static void appel_inter_block_aff(ir_node *bl, void *data)
1068 appel_clique_walker_t *env = data;
1069 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1073 for(i = 0; i < bli->n_live_in; ++i) {
1074 ir_node *irn = bli->live_in[i];
1076 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1077 ir_node *pred = get_Block_cfgpred_block(bl, j);
1079 int nr = appel_get_live_end_nr(env, pred, irn);
1081 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1085 for(i = 0; i < bli->n_phi; ++i) {
1086 ir_node *irn = bli->phi[i];
1088 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1089 ir_node *pred = get_Block_cfgpred_block(bl, j);
1090 ir_node *op = get_irn_n(irn, j);
1092 int nr = appel_get_live_end_nr(env, pred, op);
1094 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1100 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1104 appel_clique_walker_t env;
1105 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1107 be_liveness_recompute(co->cenv->lv);
1108 obstack_init(&env.obst);
1109 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1110 env.curr_nr = co->cls->n_regs;
1114 bitset_copy(adm, co->cenv->ignore_colors);
1115 bitset_flip_all(adm);
1117 /* Make color map. */
1118 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1119 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1120 const arch_register_t *reg = &co->cls->regs[i];
1121 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1125 env.curr_nr = n_colors;
1126 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1127 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1129 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1131 /* make the first k nodes interfere */
1132 for(i = 0; i < n_colors; ++i) {
1134 for(j = i + 1; j < n_colors; ++j)
1135 fprintf(f, "%d %d -1 ", i, j);
1140 env.curr_nr = n_colors;
1141 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1142 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1143 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1144 obstack_free(&env.obst, NULL);
1148 ___ _____ ____ ____ ___ _____ ____ _
1149 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1150 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1151 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1152 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1156 static const char *get_dot_color_name(int col)
1158 static const char *names[] = {
1192 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1195 typedef struct _co_ifg_dump_t {
1196 const copy_opt_t *co;
1200 static void ifg_dump_graph_attr(FILE *f, void *self)
1202 fprintf(f, "overlap=scale");
1205 static int ifg_is_dump_node(void *self, ir_node *irn)
1207 co_ifg_dump_t *cod = self;
1208 return !arch_irn_is(cod->co->aenv, irn, ignore);
1211 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1213 co_ifg_dump_t *env = self;
1214 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1215 arch_register_req_t req;
1218 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1219 limited = arch_register_req_is(&req, limited);
1221 if(env->flags & CO_IFG_DUMP_LABELS) {
1222 ir_fprintf(f, "label=\"%+F", irn);
1224 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1225 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1226 req.limited(req.limited_env, bs);
1227 ir_fprintf(f, "\\n%B", bs);
1229 ir_fprintf(f, "\" ");
1233 fprintf(f, "label=\"\" shape=point " );
1235 if(env->flags & CO_IFG_DUMP_SHAPE)
1236 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1238 if(env->flags & CO_IFG_DUMP_COLORS)
1239 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1242 static void ifg_dump_at_end(FILE *file, void *self)
1244 co_ifg_dump_t *env = self;
1247 co_gs_foreach_aff_node(env->co, a) {
1248 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1249 unsigned aidx = get_irn_idx(a->irn);
1252 co_gs_foreach_neighb(a, n) {
1253 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1254 unsigned nidx = get_irn_idx(n->irn);
1257 const char *color = nr == ar ? "blue" : "red";
1258 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1259 if(env->flags & CO_IFG_DUMP_LABELS)
1260 fprintf(file, "label=\"%d\" ", n->costs);
1261 if(env->flags & CO_IFG_DUMP_COLORS)
1262 fprintf(file, "color=%s ", color);
1264 fprintf(file, "style=dashed");
1265 fprintf(file, "];\n");
1272 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1274 ifg_dump_graph_attr,
1283 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1289 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1293 void co_solve_park_moon(copy_opt_t *opt)
1298 static int void_algo(copy_opt_t *co)
1305 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1306 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1307 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1308 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1312 static co_algo_t *algos[] = {
1315 co_solve_heuristic_new,
1316 co_solve_heuristic_java,
1324 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1325 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1326 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1327 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1331 void co_driver(be_chordal_env_t *cenv)
1334 co_algo_t *algo_func;
1337 if(algo < 0 || algo >= CO_ALGO_LAST)
1340 co = new_copy_opt(cenv, cost_func);
1341 co_build_ou_structure(co);
1342 co_build_graph_structure(co);
1343 init_costs = co_get_copy_costs(co);
1345 /* Dump the interference graph in Appel's format. */
1346 if(dump_flags & DUMP_APPEL) {
1347 FILE *f = be_chordal_open(cenv, "", ".apl");
1348 co_dump_appel_graph(co, f);
1352 if(dump_flags & DUMP_BEFORE) {
1353 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1354 co_dump_ifg_dot(co, f, style_flags);
1358 algo_func = algos[algo];
1361 if(dump_flags & DUMP_AFTER) {
1362 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1363 co_dump_ifg_dot(co, f, style_flags);
1368 int optimizable_costs = co_get_max_copy_costs(co) - co_get_lower_bound(co);
1369 int remaining = co_get_copy_costs(co);
1370 int evitable = remaining - co_get_lower_bound(co);
1372 ir_printf("%30F %10s %10d%10d%10d%10d", cenv->irg, cenv->cls->name,
1373 co_get_max_copy_costs(co), init_costs,
1374 co_get_inevit_copy_costs(co), co_get_lower_bound(co));
1376 if(optimizable_costs > 0)
1377 printf("%10d %5.2f\n", remaining, (evitable * 100.0) / optimizable_costs);
1379 printf("%10d %5s\n", remaining, "-");
1382 co_free_graph_structure(co);
1383 co_free_ou_structure(co);