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 "bejavacoal.h"
43 #include "bestatevent.h"
46 #include <libcore/lc_timing.h>
47 #include <libcore/lc_opts.h>
48 #endif /* WITH_LIBCORE */
53 #define DUMP_ALL 2 * DUMP_APPEL - 1
55 #define COST_FUNC_FREQ 1
56 #define COST_FUNC_LOOP 2
57 #define COST_FUNC_ALL_ONE 3
59 static int dump_flags = 0;
60 static int style_flags = 0;
61 static int do_stats = 0;
62 static cost_fct_t cost_func = co_get_costs_exec_freq;
63 static int algo = CO_ALGO_HEUR2;
64 static int improve = 1;
67 static const lc_opt_enum_mask_items_t dump_items[] = {
68 { "before", DUMP_BEFORE },
69 { "after", DUMP_AFTER },
70 { "appel", DUMP_APPEL },
75 static const lc_opt_enum_mask_items_t style_items[] = {
76 { "color", CO_IFG_DUMP_COLORS },
77 { "labels", CO_IFG_DUMP_LABELS },
78 { "constr", CO_IFG_DUMP_CONSTR },
79 { "shape", CO_IFG_DUMP_SHAPE },
80 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
84 static const lc_opt_enum_mask_items_t algo_items[] = {
85 { "none", CO_ALGO_NONE },
86 { "heur", CO_ALGO_HEUR },
87 { "heur2", CO_ALGO_HEUR2 },
88 { "heur3", CO_ALGO_HEUR3 },
89 { "ilp", CO_ALGO_ILP },
93 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
94 { "freq", co_get_costs_exec_freq },
95 { "loop", co_get_costs_loop_depth },
96 { "one", co_get_costs_all_one },
100 static lc_opt_enum_mask_var_t dump_var = {
101 &dump_flags, dump_items
104 static lc_opt_enum_mask_var_t style_var = {
105 &style_flags, style_items
108 static lc_opt_enum_mask_var_t algo_var = {
112 static lc_opt_enum_func_ptr_var_t cost_func_var = {
113 &cost_func, cost_func_items
116 static const lc_opt_table_entry_t options[] = {
117 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo (heur, heur2, heur3, ilp)", &algo_var),
118 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function (freq, loop, one)", &cost_func_var),
119 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
120 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
121 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
122 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
126 /* Insert additional options registration functions here. */
127 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
128 extern void be_co2_register_options(lc_opt_entry_t *grp);
129 extern void be_co3_register_options(lc_opt_entry_t *grp);
131 void co_register_options(lc_opt_entry_t *grp)
133 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
134 lc_opt_add_table(co_grp, options);
136 be_co2_register_options(co_grp);
137 be_co3_register_options(co_grp);
139 be_co_ilp_register_options(co_grp);
145 #undef QUICK_AND_DIRTY_HACK
147 /******************************************************************************
150 | | __ ___ _ __ ___ _ __ __ _| |
151 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
152 | |__| | __/ | | | __/ | | (_| | |
153 \_____|\___|_| |_|\___|_| \__,_|_|
155 ******************************************************************************/
157 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
160 void be_copy_opt_init(void) {
163 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
165 const char *s1, *s2, *s3;
169 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
171 co = xcalloc(1, sizeof(*co));
172 co->cenv = chordal_env;
173 co->aenv = chordal_env->birg->main_env->arch_env;
174 co->irg = chordal_env->irg;
175 co->cls = chordal_env->cls;
176 co->get_costs = get_costs;
178 s1 = get_irp_prog_name();
179 s2 = get_entity_name(get_irg_entity(co->irg));
180 s3 = chordal_env->cls->name;
181 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
182 co->name = xmalloc(len);
183 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
188 void free_copy_opt(copy_opt_t *co) {
193 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
194 arch_register_req_t req;
195 const arch_register_t *reg;
197 if (arch_irn_is(co->aenv, irn, ignore))
200 reg = arch_get_irn_register(co->aenv, irn);
201 if (arch_register_type_is(reg, ignore))
204 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
210 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
211 const ir_edge_t *edge;
212 const arch_register_t *reg;
214 assert(0 && "Is buggy and obsolete. Do not use");
216 if (arch_irn_is(co->aenv, irn, ignore))
219 reg = arch_get_irn_register(co->aenv, irn);
220 if (arch_register_type_is(reg, ignore))
223 foreach_out_edge(irn, edge) {
224 ir_node *n = edge->src;
226 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
227 arch_register_req_t req;
228 arch_get_register_req(co->aenv, &req, n, -1);
231 is_Perm(co->aenv, n) ||
232 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
241 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
244 ir_node *root_block = get_nodes_block(root);
247 /* for phis the copies are placed in the corresponding pred-block */
248 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
250 /* a perm places the copy in the same block as it resides */
251 loop = get_irn_loop(root_block);
254 int d = get_loop_depth(loop);
260 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
262 ir_node *root_bl = get_nodes_block(root);
263 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
264 res = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
266 /* don't allow values smaller than one. */
267 return res < 1 ? 1 : res;
271 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
275 /******************************************************************************
276 ____ _ _ _ _ _ _____ _
277 / __ \ | | | | | | (_) | / ____| |
278 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
279 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
280 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
281 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
284 ******************************************************************************/
287 * Determines a maximum weighted independent set with respect to
288 * the interference and conflict edges of all nodes in a qnode.
290 static int ou_max_ind_set_costs(unit_t *ou) {
291 be_chordal_env_t *chordal_env = ou->co->cenv;
292 ir_node **safe, **unsafe;
293 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
295 int max, pos, curr_weight, best_weight = 0;
297 /* assign the nodes into two groups.
298 * safe: node has no interference, hence it is in every max stable set.
299 * unsafe: node has an interference
301 safe = alloca((ou->node_count-1) * sizeof(*safe));
304 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
305 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
307 for(i=1; i<ou->node_count; ++i) {
309 for(o=1; o<ou->node_count; ++o) {
312 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
313 unsafe_costs[unsafe_count] = ou->costs[i];
314 unsafe[unsafe_count] = ou->nodes[i];
321 safe_costs += ou->costs[i];
322 safe[safe_count++] = ou->nodes[i];
327 /* now compute the best set out of the unsafe nodes*/
328 if (unsafe_count > MIS_HEUR_TRIGGER) {
329 bitset_t *best = bitset_alloca(unsafe_count);
330 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
331 for (i=0; i<unsafe_count; ++i) {
333 /* check if it is a stable set */
334 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
335 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
336 bitset_clear(best, i); /* clear the bit and try next one */
340 /* compute the weight */
341 bitset_foreach(best, pos)
342 best_weight += unsafe_costs[pos];
344 /* Exact Algorithm: Brute force */
345 curr = bitset_alloca(unsafe_count);
346 bitset_set_all(curr);
347 while ((max = bitset_popcnt(curr)) != 0) {
348 /* check if curr is a stable set */
349 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
350 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) */
351 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
354 /* if we arrive here, we have a stable set */
355 /* compute the weigth of the stable set*/
357 bitset_foreach(curr, pos)
358 curr_weight += unsafe_costs[pos];
361 if (curr_weight > best_weight) {
362 best_weight = curr_weight;
370 return safe_costs+best_weight;
373 static void co_collect_units(ir_node *irn, void *env) {
374 copy_opt_t *co = env;
376 arch_register_req_t req;
378 if (!is_curr_reg_class(co, irn))
380 if (!co_is_optimizable_root(co, irn))
383 /* Init a new unit */
384 unit = xcalloc(1, sizeof(*unit));
386 unit->node_count = 1;
387 INIT_LIST_HEAD(&unit->queue);
389 /* Phi with some/all of its arguments */
390 if (is_Reg_Phi(irn)) {
394 arity = get_irn_arity(irn);
395 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
396 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
397 unit->nodes[0] = irn;
400 for (i=0; i<arity; ++i) {
402 ir_node *arg = get_irn_n(irn, i);
404 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
407 if (nodes_interfere(co->cenv, irn, arg)) {
408 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
412 /* Else insert the argument of the phi to the members of this ou */
413 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
415 /* Check if arg has occurred at a prior position in the arg/list */
417 for (o=0; o<unit->node_count; ++o)
418 if (unit->nodes[o] == arg) {
423 if (!arg_pos) { /* a new argument */
424 /* insert node, set costs */
425 unit->nodes[unit->node_count] = arg;
426 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
428 } else { /* arg has occured before in same phi */
429 /* increase costs for existing arg */
430 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
433 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
434 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
437 /* Proj of a perm with corresponding arg */
438 if (is_Perm_Proj(co->aenv, irn)) {
439 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
440 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
441 unit->costs = xmalloc(2 * sizeof(*unit->costs));
442 unit->node_count = 2;
443 unit->nodes[0] = irn;
444 unit->nodes[1] = get_Perm_src(irn);
445 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
448 /* Src == Tgt of a 2-addr-code instruction */
449 if (is_2addr_code(co->aenv, irn, &req)) {
450 ir_node *other = req.other_same;
451 if (!nodes_interfere(co->cenv, irn, other)) {
452 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
453 unit->costs = xmalloc(2 * sizeof(*unit->costs));
454 unit->node_count = 2;
455 unit->nodes[0] = irn;
456 unit->nodes[1] = other;
457 unit->costs[1] = co->get_costs(co, irn, other, -1);
460 assert(0 && "This is not an optimizable node!");
462 /* Insert the new unit at a position according to its costs */
463 if (unit->node_count > 1) {
465 struct list_head *tmp;
467 /* Determine the maximum costs this unit can cause: all_nodes_cost */
468 for(i=1; i<unit->node_count; ++i) {
469 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
470 unit->all_nodes_costs += unit->costs[i];
473 /* Determine the minimal costs this unit will cause: min_nodes_costs */
474 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
475 /* Insert the new ou according to its sort_key */
477 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
479 list_add(&unit->units, tmp);
485 #ifdef QUICK_AND_DIRTY_HACK
487 static int compare_ous(const void *k1, const void *k2) {
488 const unit_t *u1 = *((const unit_t **) k1);
489 const unit_t *u2 = *((const unit_t **) k2);
490 int i, o, u1_has_constr, u2_has_constr;
491 arch_register_req_t req;
492 const arch_env_t *aenv = u1->co->aenv;
494 /* Units with constraints come first */
496 for (i=0; i<u1->node_count; ++i) {
497 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
498 if (arch_register_req_is(&req, limited)) {
505 for (i=0; i<u2->node_count; ++i) {
506 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
507 if (arch_register_req_is(&req, limited)) {
513 if (u1_has_constr != u2_has_constr)
514 return u2_has_constr - u1_has_constr;
516 /* Now check, whether the two units are connected */
518 for (i=0; i<u1->node_count; ++i)
519 for (o=0; o<u2->node_count; ++o)
520 if (u1->nodes[i] == u2->nodes[o])
524 /* After all, the sort key decides. Greater keys come first. */
525 return u2->sort_key - u1->sort_key;
530 * Sort the ou's according to constraints and their sort_key
532 static void co_sort_units(copy_opt_t *co) {
533 int i, count = 0, costs;
536 /* get the number of ous, remove them form the list and fill the array */
537 list_for_each_entry(unit_t, ou, &co->units, units)
539 ous = alloca(count * sizeof(*ous));
541 costs = co_get_max_copy_costs(co);
544 list_for_each_entry(unit_t, ou, &co->units, units)
547 INIT_LIST_HEAD(&co->units);
549 assert(count == i && list_empty(&co->units));
551 for (i=0; i<count; ++i)
552 ir_printf("%+F\n", ous[i]->nodes[0]);
554 qsort(ous, count, sizeof(*ous), compare_ous);
557 for (i=0; i<count; ++i)
558 ir_printf("%+F\n", ous[i]->nodes[0]);
560 /* reinsert into list in correct order */
561 for (i=0; i<count; ++i)
562 list_add_tail(&ous[i]->units, &co->units);
564 assert(costs == co_get_max_copy_costs(co));
568 void co_build_ou_structure(copy_opt_t *co) {
569 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
570 INIT_LIST_HEAD(&co->units);
571 irg_walk_graph(co->irg, co_collect_units, NULL, co);
572 #ifdef QUICK_AND_DIRTY_HACK
577 void co_free_ou_structure(copy_opt_t *co) {
580 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
585 co->units.next = NULL;
588 /* co_solve_heuristic() is implemented in becopyheur.c */
590 int co_get_max_copy_costs(const copy_opt_t *co) {
596 list_for_each_entry(unit_t, curr, &co->units, units) {
597 res += curr->inevitable_costs;
598 for (i=1; i<curr->node_count; ++i)
599 res += curr->costs[i];
604 int co_get_inevit_copy_costs(const copy_opt_t *co) {
610 list_for_each_entry(unit_t, curr, &co->units, units)
611 res += curr->inevitable_costs;
615 int co_get_copy_costs(const copy_opt_t *co) {
621 list_for_each_entry(unit_t, curr, &co->units, units) {
622 int root_col = get_irn_col(co, curr->nodes[0]);
623 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
624 res += curr->inevitable_costs;
625 for (i=1; i<curr->node_count; ++i) {
626 int arg_col = get_irn_col(co, curr->nodes[i]);
627 if (root_col != arg_col) {
628 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
629 res += curr->costs[i];
636 int co_get_lower_bound(const copy_opt_t *co) {
642 list_for_each_entry(unit_t, curr, &co->units, units)
643 res += curr->inevitable_costs + curr->min_nodes_costs;
647 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
649 bitset_t *seen = bitset_irg_malloc(co->irg);
652 memset(stat, 0, sizeof(stat[0]));
654 /* count affinity edges. */
655 co_gs_foreach_aff_node(co, an) {
657 stat->aff_nodes += 1;
658 bitset_add_irn(seen, an->irn);
659 co_gs_foreach_neighb(an, neigh) {
660 if(!bitset_contains_irn(seen, neigh->irn)) {
661 stat->aff_edges += 1;
662 stat->max_costs += neigh->costs;
664 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
665 stat->costs += neigh->costs;
666 stat->unsatisfied_edges += 1;
669 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
671 stat->inevit_costs += neigh->costs;
681 /******************************************************************************
683 / ____| | | / ____| |
684 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
685 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
686 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
687 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
690 ******************************************************************************/
692 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
693 const affinity_node_t *n1 = k1;
694 const affinity_node_t *n2 = k2;
696 return (n1->irn != n2->irn);
699 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
700 affinity_node_t new_node, *node;
701 neighb_t new_nbr, *nbr;
706 new_node.neighbours = NULL;
707 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
710 for (nbr = node->neighbours; nbr; nbr = nbr->next)
711 if (nbr->irn == n2) {
716 /* if we did not find n2 in n1's neighbourhood insert it */
718 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
719 nbr = obstack_finish(&co->obst);
722 nbr->next = node->neighbours;
723 node->neighbours = nbr;
727 /* now nbr points to n1's neighbour-entry of n2 */
731 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
732 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
733 add_edge(co, n1, n2, costs);
734 add_edge(co, n2, n1, costs);
738 static void build_graph_walker(ir_node *irn, void *env) {
739 copy_opt_t *co = env;
741 arch_register_req_t req;
742 const arch_register_t *reg;
744 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
747 reg = arch_get_irn_register(co->aenv, irn);
748 if (arch_register_type_is(reg, ignore))
753 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
754 ir_node *arg = get_irn_n(irn, pos);
755 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
759 else if (is_Perm_Proj(co->aenv, irn)) {
760 ir_node *arg = get_Perm_src(irn);
761 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
765 else if (is_2addr_code(co->aenv, irn, &req))
766 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
769 void co_build_graph_structure(copy_opt_t *co) {
770 obstack_init(&co->obst);
771 co->nodes = new_set(compare_affinity_node_t, 32);
773 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
776 void co_free_graph_structure(copy_opt_t *co) {
780 obstack_free(&co->obst, NULL);
784 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
786 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
787 affinity_node_t new_node, *n;
792 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
794 return (n->degree > 0);
799 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
801 be_ifg_t *ifg = co->cenv->ifg;
802 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
803 bitset_t *adm = bitset_alloca(co->cls->n_regs);
810 for(i = 0; i < co->cls->n_regs; ++i) {
811 const arch_register_t *reg = &co->cls->regs[i];
812 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
816 * n contains the first node number.
817 * the values below n are the pre-colored register nodes
820 it = be_ifg_nodes_iter_alloca(ifg);
821 nit = be_ifg_neighbours_iter_alloca(ifg);
824 be_ifg_foreach_node(ifg, it, irn) {
825 if(!arch_irn_is(co->aenv, irn, ignore))
826 set_irn_link(irn, INT_TO_PTR(n++));
829 fprintf(f, "%d %d\n", n, n_regs);
831 be_ifg_foreach_node(ifg, it, irn) {
832 if(!arch_irn_is(co->aenv, irn, ignore)) {
833 int idx = PTR_TO_INT(get_irn_link(irn));
834 affinity_node_t *a = get_affinity_info(co, irn);
836 arch_register_req_t req;
839 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
840 if(arch_register_req_is(&req, limited)) {
841 bitset_clear_all(adm);
842 req.limited(req.limited_env, adm);
843 for(i = 0; i < co->cls->n_regs; ++i)
844 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
845 fprintf(f, "%d %d -1\n", color_map[i], idx);
850 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
851 if(!arch_irn_is(co->aenv, adj, ignore)) {
852 int adj_idx = PTR_TO_INT(get_irn_link(adj));
854 fprintf(f, "%d %d -1\n", idx, adj_idx);
861 co_gs_foreach_neighb(a, n) {
862 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
863 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
865 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
873 typedef struct _appel_clique_walker_t {
875 const copy_opt_t *co;
882 } appel_clique_walker_t;
884 typedef struct _appel_block_info_t {
894 } appel_block_info_t;
896 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
899 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
900 int res = (int) freq;
901 return res == 0 ? 1 : res;
903 ir_loop *loop = get_irn_loop(bl);
905 int d = get_loop_depth(loop);
912 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
914 appel_block_info_t *res = NULL;
917 appel_clique_walker_t *d = (void *) phase;
918 res = phase_alloc(phase, sizeof(res[0]));
919 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
920 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
921 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
922 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
923 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
924 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
930 typedef struct _insn_list_t {
932 struct list_head list;
935 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
937 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
940 for(i = 0; i < bli->n_live_end; ++i)
941 if(bli->live_end[i] == irn)
942 return bli->live_end_nr[i];
947 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
949 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
955 foreach_pset(live, irn)
956 live_arr[n_live++] = irn;
958 /* dump the live after clique */
960 for(j = 0; j < n_live; ++j) {
963 for(k = j + 1; k < n_live; ++k) {
964 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
966 fprintf(env->f, "\n");
970 /* dump the affinities */
971 for(j = 0; !env->dumb && j < n_live; ++j) {
972 ir_node *irn = live_arr[j];
973 int old_nr = PTR_TO_INT(get_irn_link(irn));
975 /* if the node was already live in the last insn dump the affinity */
976 if(old_nr > start_nr) {
977 int weight = appel_aff_weight(env, bl);
978 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
982 /* set the current numbers into the link field. */
983 for(j = 0; j < n_live; ++j) {
984 ir_node *irn = live_arr[j];
985 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
988 return curr_nr + n_live;
991 static void appel_walker(ir_node *bl, void *data)
993 appel_clique_walker_t *env = data;
994 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
995 struct obstack *obst = &env->obst;
996 void *base = obstack_base(obst);
997 pset *live = pset_new_ptr_default();
1001 int start_nr = env->curr_nr;
1002 int curr_nr = start_nr;
1004 be_insn_env_t insn_env;
1009 insn_env.aenv = env->co->aenv;
1010 insn_env.cls = env->co->cls;
1011 insn_env.obst = obst;
1012 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1014 /* Guess how many insns will be in this block. */
1015 sched_foreach(bl, irn)
1019 insns = malloc(n_nodes * sizeof(insns[0]));
1021 /* Put all insns in an array. */
1022 irn = sched_first(bl);
1023 while(!sched_is_end(irn)) {
1025 insn = be_scan_insn(&insn_env, irn);
1026 insns[n_insns++] = insn;
1027 irn = insn->next_insn;
1030 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1031 be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
1033 /* Generate the bad and ugly. */
1034 for(i = n_insns - 1; i >= 0; --i) {
1035 be_insn_t *insn = insns[i];
1037 /* The first live set has to be saved in the block border set. */
1038 if(i == n_insns - 1) {
1040 foreach_pset(live, irn) {
1041 bli->live_end[j] = irn;
1042 bli->live_end_nr[j] = curr_nr + j;
1045 bli->n_live_end = j;
1049 for(j = 0; j < insn->use_start; ++j) {
1050 ir_node *op = insn->ops[j].carrier;
1051 bitset_t *adm = insn->ops[j].regs;
1055 if(!insn->ops[j].has_constraints)
1059 foreach_pset(live, irn) {
1067 assert(nr < pset_count(live));
1069 for(k = 0; k < env->co->cls->n_regs; ++k) {
1070 int mapped_col = env->color_map[k];
1071 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1072 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1077 /* dump the clique and update the stuff. */
1078 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1080 /* remove all defs. */
1081 for(j = 0; j < insn->use_start; ++j)
1082 pset_remove_ptr(live, insn->ops[j].carrier);
1084 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1085 bli->phi[bli->n_phi] = insn->irn;
1086 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1092 for(j = insn->use_start; j < insn->n_ops; ++j)
1093 pset_insert_ptr(live, insn->ops[j].carrier);
1096 /* print the start clique. */
1097 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1100 foreach_pset(live, irn) {
1101 bli->live_in[i] = irn;
1102 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1109 obstack_free(obst, base);
1110 env->curr_nr = curr_nr;
1113 static void appel_inter_block_aff(ir_node *bl, void *data)
1115 appel_clique_walker_t *env = data;
1116 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1120 for(i = 0; i < bli->n_live_in; ++i) {
1121 ir_node *irn = bli->live_in[i];
1123 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1124 ir_node *pred = get_Block_cfgpred_block(bl, j);
1126 int nr = appel_get_live_end_nr(env, pred, irn);
1128 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1132 for(i = 0; i < bli->n_phi; ++i) {
1133 ir_node *irn = bli->phi[i];
1135 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1136 ir_node *pred = get_Block_cfgpred_block(bl, j);
1137 ir_node *op = get_irn_n(irn, j);
1139 int nr = appel_get_live_end_nr(env, pred, op);
1141 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1147 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1151 appel_clique_walker_t env;
1152 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1154 be_liveness_recompute(co->cenv->lv);
1155 obstack_init(&env.obst);
1156 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1157 env.curr_nr = co->cls->n_regs;
1161 bitset_copy(adm, co->cenv->ignore_colors);
1162 bitset_flip_all(adm);
1164 /* Make color map. */
1165 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1166 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1167 const arch_register_t *reg = &co->cls->regs[i];
1168 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1172 env.curr_nr = n_colors;
1173 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1174 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1176 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1178 /* make the first k nodes interfere */
1179 for(i = 0; i < n_colors; ++i) {
1181 for(j = i + 1; j < n_colors; ++j)
1182 fprintf(f, "%d %d -1 ", i, j);
1187 env.curr_nr = n_colors;
1188 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1189 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1190 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1191 obstack_free(&env.obst, NULL);
1195 ___ _____ ____ ____ ___ _____ ____ _
1196 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1197 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1198 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1199 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1203 static const char *get_dot_color_name(int col)
1205 static const char *names[] = {
1239 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1242 typedef struct _co_ifg_dump_t {
1243 const copy_opt_t *co;
1247 static void ifg_dump_graph_attr(FILE *f, void *self)
1249 fprintf(f, "overlap=scale");
1252 static int ifg_is_dump_node(void *self, ir_node *irn)
1254 co_ifg_dump_t *cod = self;
1255 return !arch_irn_is(cod->co->aenv, irn, ignore);
1258 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1260 co_ifg_dump_t *env = self;
1261 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1262 arch_register_req_t req;
1265 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1266 limited = arch_register_req_is(&req, limited);
1268 if(env->flags & CO_IFG_DUMP_LABELS) {
1269 ir_fprintf(f, "label=\"%+F", irn);
1271 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1272 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1273 req.limited(req.limited_env, bs);
1274 ir_fprintf(f, "\\n%B", bs);
1276 ir_fprintf(f, "\" ");
1280 fprintf(f, "label=\"\" shape=point " );
1282 if(env->flags & CO_IFG_DUMP_SHAPE)
1283 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1285 if(env->flags & CO_IFG_DUMP_COLORS)
1286 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1289 static void ifg_dump_at_end(FILE *file, void *self)
1291 co_ifg_dump_t *env = self;
1294 co_gs_foreach_aff_node(env->co, a) {
1295 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1296 unsigned aidx = get_irn_idx(a->irn);
1299 co_gs_foreach_neighb(a, n) {
1300 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1301 unsigned nidx = get_irn_idx(n->irn);
1304 const char *color = nr == ar ? "blue" : "red";
1305 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1306 if(env->flags & CO_IFG_DUMP_LABELS)
1307 fprintf(file, "label=\"%d\" ", n->costs);
1308 if(env->flags & CO_IFG_DUMP_COLORS)
1309 fprintf(file, "color=%s ", color);
1311 fprintf(file, "style=dotted");
1312 fprintf(file, "];\n");
1319 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1321 ifg_dump_graph_attr,
1330 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1336 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1340 void co_solve_park_moon(copy_opt_t *opt)
1345 static int void_algo(copy_opt_t *co)
1352 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1353 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1354 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1355 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1362 int can_improve_existing;
1365 static co_algo_info_t algos[] = {
1366 { void_algo, "none", 0 },
1367 { co_solve_heuristic, "heur1", 0 },
1368 { co_solve_heuristic_new, "heur2", 0 },
1369 { co_solve_heuristic_java, "heur3", 0 },
1371 { co_solve_ilp2, "ilp", 1 },
1378 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1379 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1380 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1381 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1385 void co_driver(be_chordal_env_t *cenv)
1388 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1390 co_complete_stats_t before, after;
1392 co_algo_t *algo_func;
1393 int was_optimal = 0;
1395 if(algo < 0 || algo >= CO_ALGO_LAST)
1398 co = new_copy_opt(cenv, cost_func);
1399 co_build_ou_structure(co);
1400 co_build_graph_structure(co);
1402 co_complete_stats(co, &before);
1404 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1405 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1406 be_stat_ev_ull("co_max_costs", before.max_costs);
1407 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1408 be_stat_ev_ull("co_aff_int", before.aff_int);
1410 be_stat_ev_ull("co_init_costs", before.costs);
1411 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1413 /* Dump the interference graph in Appel's format. */
1414 if(dump_flags & DUMP_APPEL) {
1415 FILE *f = be_chordal_open(cenv, "", ".apl");
1416 co_dump_appel_graph(co, f);
1420 if(dump_flags & DUMP_BEFORE) {
1421 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1422 co_dump_ifg_dot(co, f, style_flags);
1426 /* if the algo can improve results, provide an initial solution with heur3 */
1427 if(improve && algos[algo].can_improve_existing) {
1428 co_complete_stats_t stats;
1430 /* produce a heuristical solution */
1431 co_solve_heuristic_java(co);
1433 /* do the stats and provide the current costs */
1434 co_complete_stats(co, &stats);
1435 be_stat_ev_ull("co_prepare_costs", stats.costs);
1438 /* start the JVM here so that it does not tamper the timing. */
1439 if(algo == CO_ALGO_HEUR3)
1440 be_java_coal_start_jvm();
1442 algo_func = algos[algo].algo;
1445 lc_timer_reset_and_start(timer);
1448 was_optimal = algo_func(co);
1451 lc_timer_stop(timer);
1452 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1455 be_stat_ev_ull("co_optimal", was_optimal);
1457 if(dump_flags & DUMP_AFTER) {
1458 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1459 co_dump_ifg_dot(co, f, style_flags);
1463 co_complete_stats(co, &after);
1466 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1467 ulong64 evitable = after.costs - after.inevit_costs;
1469 ir_printf("%30F ", cenv->irg);
1470 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1472 if(optimizable_costs > 0)
1473 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1475 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1478 be_stat_ev_ull("co_after_costs", after.costs);
1479 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1481 co_free_graph_structure(co);
1482 co_free_ou_structure(co);