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 typedef int (*opt_funcptr)(void);
95 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
96 { "freq", (opt_funcptr) co_get_costs_exec_freq },
97 { "loop", (opt_funcptr) co_get_costs_loop_depth },
98 { "one", (opt_funcptr) co_get_costs_all_one },
102 static lc_opt_enum_mask_var_t dump_var = {
103 &dump_flags, dump_items
106 static lc_opt_enum_mask_var_t style_var = {
107 &style_flags, style_items
110 static lc_opt_enum_mask_var_t algo_var = {
114 static lc_opt_enum_func_ptr_var_t cost_func_var = {
115 (opt_funcptr*) &cost_func, cost_func_items
118 static const lc_opt_table_entry_t options[] = {
119 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo (heur, heur2, heur3, ilp)", &algo_var),
120 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function (freq, loop, one)", &cost_func_var),
121 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
122 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
123 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
124 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
128 /* Insert additional options registration functions here. */
129 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
130 extern void be_co2_register_options(lc_opt_entry_t *grp);
131 extern void be_co3_register_options(lc_opt_entry_t *grp);
133 void co_register_options(lc_opt_entry_t *grp)
135 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
136 lc_opt_add_table(co_grp, options);
138 be_co2_register_options(co_grp);
139 be_co3_register_options(co_grp);
141 be_co_ilp_register_options(co_grp);
147 #undef QUICK_AND_DIRTY_HACK
149 /******************************************************************************
152 | | __ ___ _ __ ___ _ __ __ _| |
153 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
154 | |__| | __/ | | | __/ | | (_| | |
155 \_____|\___|_| |_|\___|_| \__,_|_|
157 ******************************************************************************/
159 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
162 void be_copy_opt_init(void) {
165 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
167 const char *s1, *s2, *s3;
171 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
173 co = xcalloc(1, sizeof(*co));
174 co->cenv = chordal_env;
175 co->aenv = chordal_env->birg->main_env->arch_env;
176 co->irg = chordal_env->irg;
177 co->cls = chordal_env->cls;
178 co->get_costs = get_costs;
180 s1 = get_irp_prog_name();
181 s2 = get_entity_name(get_irg_entity(co->irg));
182 s3 = chordal_env->cls->name;
183 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
184 co->name = xmalloc(len);
185 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
190 void free_copy_opt(copy_opt_t *co) {
195 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
196 arch_register_req_t req;
197 const arch_register_t *reg;
199 if (arch_irn_is(co->aenv, irn, ignore))
202 reg = arch_get_irn_register(co->aenv, irn);
203 if (arch_register_type_is(reg, ignore))
206 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
212 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
213 const ir_edge_t *edge;
214 const arch_register_t *reg;
216 assert(0 && "Is buggy and obsolete. Do not use");
218 if (arch_irn_is(co->aenv, irn, ignore))
221 reg = arch_get_irn_register(co->aenv, irn);
222 if (arch_register_type_is(reg, ignore))
225 foreach_out_edge(irn, edge) {
226 ir_node *n = edge->src;
228 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
229 arch_register_req_t req;
230 arch_get_register_req(co->aenv, &req, n, -1);
233 is_Perm(co->aenv, n) ||
234 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
243 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
246 ir_node *root_block = get_nodes_block(root);
249 /* for phis the copies are placed in the corresponding pred-block */
250 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
252 /* a perm places the copy in the same block as it resides */
253 loop = get_irn_loop(root_block);
256 int d = get_loop_depth(loop);
262 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
264 ir_node *root_bl = get_nodes_block(root);
265 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
266 res = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
268 /* don't allow values smaller than one. */
269 return res < 1 ? 1 : res;
273 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
277 /******************************************************************************
278 ____ _ _ _ _ _ _____ _
279 / __ \ | | | | | | (_) | / ____| |
280 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
281 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
282 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
283 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
286 ******************************************************************************/
289 * Determines a maximum weighted independent set with respect to
290 * the interference and conflict edges of all nodes in a qnode.
292 static int ou_max_ind_set_costs(unit_t *ou) {
293 be_chordal_env_t *chordal_env = ou->co->cenv;
294 ir_node **safe, **unsafe;
295 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
297 int max, pos, curr_weight, best_weight = 0;
299 /* assign the nodes into two groups.
300 * safe: node has no interference, hence it is in every max stable set.
301 * unsafe: node has an interference
303 safe = alloca((ou->node_count-1) * sizeof(*safe));
306 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
307 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
309 for(i=1; i<ou->node_count; ++i) {
311 for(o=1; o<ou->node_count; ++o) {
314 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
315 unsafe_costs[unsafe_count] = ou->costs[i];
316 unsafe[unsafe_count] = ou->nodes[i];
323 safe_costs += ou->costs[i];
324 safe[safe_count++] = ou->nodes[i];
329 /* now compute the best set out of the unsafe nodes*/
330 if (unsafe_count > MIS_HEUR_TRIGGER) {
331 bitset_t *best = bitset_alloca(unsafe_count);
332 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
333 for (i=0; i<unsafe_count; ++i) {
335 /* check if it is a stable set */
336 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
337 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
338 bitset_clear(best, i); /* clear the bit and try next one */
342 /* compute the weight */
343 bitset_foreach(best, pos)
344 best_weight += unsafe_costs[pos];
346 /* Exact Algorithm: Brute force */
347 curr = bitset_alloca(unsafe_count);
348 bitset_set_all(curr);
349 while ((max = bitset_popcnt(curr)) != 0) {
350 /* check if curr is a stable set */
351 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
352 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) */
353 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
356 /* if we arrive here, we have a stable set */
357 /* compute the weigth of the stable set*/
359 bitset_foreach(curr, pos)
360 curr_weight += unsafe_costs[pos];
363 if (curr_weight > best_weight) {
364 best_weight = curr_weight;
372 return safe_costs+best_weight;
375 static void co_collect_units(ir_node *irn, void *env) {
376 copy_opt_t *co = env;
378 arch_register_req_t req;
380 if (!is_curr_reg_class(co, irn))
382 if (!co_is_optimizable_root(co, irn))
385 /* Init a new unit */
386 unit = xcalloc(1, sizeof(*unit));
388 unit->node_count = 1;
389 INIT_LIST_HEAD(&unit->queue);
391 /* Phi with some/all of its arguments */
392 if (is_Reg_Phi(irn)) {
396 arity = get_irn_arity(irn);
397 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
398 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
399 unit->nodes[0] = irn;
402 for (i=0; i<arity; ++i) {
404 ir_node *arg = get_irn_n(irn, i);
406 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
409 if (nodes_interfere(co->cenv, irn, arg)) {
410 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
414 /* Else insert the argument of the phi to the members of this ou */
415 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
417 /* Check if arg has occurred at a prior position in the arg/list */
419 for (o=0; o<unit->node_count; ++o)
420 if (unit->nodes[o] == arg) {
425 if (!arg_pos) { /* a new argument */
426 /* insert node, set costs */
427 unit->nodes[unit->node_count] = arg;
428 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
430 } else { /* arg has occured before in same phi */
431 /* increase costs for existing arg */
432 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
435 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
436 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
439 /* Proj of a perm with corresponding arg */
440 if (is_Perm_Proj(co->aenv, irn)) {
441 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
442 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
443 unit->costs = xmalloc(2 * sizeof(*unit->costs));
444 unit->node_count = 2;
445 unit->nodes[0] = irn;
446 unit->nodes[1] = get_Perm_src(irn);
447 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
450 /* Src == Tgt of a 2-addr-code instruction */
451 if (is_2addr_code(co->aenv, irn, &req)) {
452 ir_node *other = req.other_same;
453 if (!nodes_interfere(co->cenv, irn, other)) {
454 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
455 unit->costs = xmalloc(2 * sizeof(*unit->costs));
456 unit->node_count = 2;
457 unit->nodes[0] = irn;
458 unit->nodes[1] = other;
459 unit->costs[1] = co->get_costs(co, irn, other, -1);
462 assert(0 && "This is not an optimizable node!");
464 /* Insert the new unit at a position according to its costs */
465 if (unit->node_count > 1) {
467 struct list_head *tmp;
469 /* Determine the maximum costs this unit can cause: all_nodes_cost */
470 for(i=1; i<unit->node_count; ++i) {
471 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
472 unit->all_nodes_costs += unit->costs[i];
475 /* Determine the minimal costs this unit will cause: min_nodes_costs */
476 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
477 /* Insert the new ou according to its sort_key */
479 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
481 list_add(&unit->units, tmp);
487 #ifdef QUICK_AND_DIRTY_HACK
489 static int compare_ous(const void *k1, const void *k2) {
490 const unit_t *u1 = *((const unit_t **) k1);
491 const unit_t *u2 = *((const unit_t **) k2);
492 int i, o, u1_has_constr, u2_has_constr;
493 arch_register_req_t req;
494 const arch_env_t *aenv = u1->co->aenv;
496 /* Units with constraints come first */
498 for (i=0; i<u1->node_count; ++i) {
499 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
500 if (arch_register_req_is(&req, limited)) {
507 for (i=0; i<u2->node_count; ++i) {
508 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
509 if (arch_register_req_is(&req, limited)) {
515 if (u1_has_constr != u2_has_constr)
516 return u2_has_constr - u1_has_constr;
518 /* Now check, whether the two units are connected */
520 for (i=0; i<u1->node_count; ++i)
521 for (o=0; o<u2->node_count; ++o)
522 if (u1->nodes[i] == u2->nodes[o])
526 /* After all, the sort key decides. Greater keys come first. */
527 return u2->sort_key - u1->sort_key;
532 * Sort the ou's according to constraints and their sort_key
534 static void co_sort_units(copy_opt_t *co) {
535 int i, count = 0, costs;
538 /* get the number of ous, remove them form the list and fill the array */
539 list_for_each_entry(unit_t, ou, &co->units, units)
541 ous = alloca(count * sizeof(*ous));
543 costs = co_get_max_copy_costs(co);
546 list_for_each_entry(unit_t, ou, &co->units, units)
549 INIT_LIST_HEAD(&co->units);
551 assert(count == i && list_empty(&co->units));
553 for (i=0; i<count; ++i)
554 ir_printf("%+F\n", ous[i]->nodes[0]);
556 qsort(ous, count, sizeof(*ous), compare_ous);
559 for (i=0; i<count; ++i)
560 ir_printf("%+F\n", ous[i]->nodes[0]);
562 /* reinsert into list in correct order */
563 for (i=0; i<count; ++i)
564 list_add_tail(&ous[i]->units, &co->units);
566 assert(costs == co_get_max_copy_costs(co));
570 void co_build_ou_structure(copy_opt_t *co) {
571 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
572 INIT_LIST_HEAD(&co->units);
573 irg_walk_graph(co->irg, co_collect_units, NULL, co);
574 #ifdef QUICK_AND_DIRTY_HACK
579 void co_free_ou_structure(copy_opt_t *co) {
582 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
587 co->units.next = NULL;
590 /* co_solve_heuristic() is implemented in becopyheur.c */
592 int co_get_max_copy_costs(const copy_opt_t *co) {
598 list_for_each_entry(unit_t, curr, &co->units, units) {
599 res += curr->inevitable_costs;
600 for (i=1; i<curr->node_count; ++i)
601 res += curr->costs[i];
606 int co_get_inevit_copy_costs(const copy_opt_t *co) {
612 list_for_each_entry(unit_t, curr, &co->units, units)
613 res += curr->inevitable_costs;
617 int co_get_copy_costs(const copy_opt_t *co) {
623 list_for_each_entry(unit_t, curr, &co->units, units) {
624 int root_col = get_irn_col(co, curr->nodes[0]);
625 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
626 res += curr->inevitable_costs;
627 for (i=1; i<curr->node_count; ++i) {
628 int arg_col = get_irn_col(co, curr->nodes[i]);
629 if (root_col != arg_col) {
630 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
631 res += curr->costs[i];
638 int co_get_lower_bound(const copy_opt_t *co) {
644 list_for_each_entry(unit_t, curr, &co->units, units)
645 res += curr->inevitable_costs + curr->min_nodes_costs;
649 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
651 bitset_t *seen = bitset_irg_malloc(co->irg);
654 memset(stat, 0, sizeof(stat[0]));
656 /* count affinity edges. */
657 co_gs_foreach_aff_node(co, an) {
659 stat->aff_nodes += 1;
660 bitset_add_irn(seen, an->irn);
661 co_gs_foreach_neighb(an, neigh) {
662 if(!bitset_contains_irn(seen, neigh->irn)) {
663 stat->aff_edges += 1;
664 stat->max_costs += neigh->costs;
666 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
667 stat->costs += neigh->costs;
668 stat->unsatisfied_edges += 1;
671 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
673 stat->inevit_costs += neigh->costs;
683 /******************************************************************************
685 / ____| | | / ____| |
686 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
687 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
688 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
689 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
692 ******************************************************************************/
694 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
695 const affinity_node_t *n1 = k1;
696 const affinity_node_t *n2 = k2;
698 return (n1->irn != n2->irn);
701 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
702 affinity_node_t new_node, *node;
703 neighb_t new_nbr, *nbr;
708 new_node.neighbours = NULL;
709 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
712 for (nbr = node->neighbours; nbr; nbr = nbr->next)
713 if (nbr->irn == n2) {
718 /* if we did not find n2 in n1's neighbourhood insert it */
720 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
721 nbr = obstack_finish(&co->obst);
724 nbr->next = node->neighbours;
725 node->neighbours = nbr;
729 /* now nbr points to n1's neighbour-entry of n2 */
733 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
734 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
735 add_edge(co, n1, n2, costs);
736 add_edge(co, n2, n1, costs);
740 static void build_graph_walker(ir_node *irn, void *env) {
741 copy_opt_t *co = env;
743 arch_register_req_t req;
744 const arch_register_t *reg;
746 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
749 reg = arch_get_irn_register(co->aenv, irn);
750 if (arch_register_type_is(reg, ignore))
755 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
756 ir_node *arg = get_irn_n(irn, pos);
757 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
761 else if (is_Perm_Proj(co->aenv, irn)) {
762 ir_node *arg = get_Perm_src(irn);
763 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
767 else if (is_2addr_code(co->aenv, irn, &req))
768 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
771 void co_build_graph_structure(copy_opt_t *co) {
772 obstack_init(&co->obst);
773 co->nodes = new_set(compare_affinity_node_t, 32);
775 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
778 void co_free_graph_structure(copy_opt_t *co) {
782 obstack_free(&co->obst, NULL);
786 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
788 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
789 affinity_node_t new_node, *n;
794 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
796 return (n->degree > 0);
801 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
803 be_ifg_t *ifg = co->cenv->ifg;
804 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
805 bitset_t *adm = bitset_alloca(co->cls->n_regs);
812 for(i = 0; i < co->cls->n_regs; ++i) {
813 const arch_register_t *reg = &co->cls->regs[i];
814 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
818 * n contains the first node number.
819 * the values below n are the pre-colored register nodes
822 it = be_ifg_nodes_iter_alloca(ifg);
823 nit = be_ifg_neighbours_iter_alloca(ifg);
826 be_ifg_foreach_node(ifg, it, irn) {
827 if(!arch_irn_is(co->aenv, irn, ignore))
828 set_irn_link(irn, INT_TO_PTR(n++));
831 fprintf(f, "%d %d\n", n, n_regs);
833 be_ifg_foreach_node(ifg, it, irn) {
834 if(!arch_irn_is(co->aenv, irn, ignore)) {
835 int idx = PTR_TO_INT(get_irn_link(irn));
836 affinity_node_t *a = get_affinity_info(co, irn);
838 arch_register_req_t req;
841 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
842 if(arch_register_req_is(&req, limited)) {
843 bitset_clear_all(adm);
844 req.limited(req.limited_env, adm);
845 for(i = 0; i < co->cls->n_regs; ++i)
846 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
847 fprintf(f, "%d %d -1\n", color_map[i], idx);
852 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
853 if(!arch_irn_is(co->aenv, adj, ignore)) {
854 int adj_idx = PTR_TO_INT(get_irn_link(adj));
856 fprintf(f, "%d %d -1\n", idx, adj_idx);
863 co_gs_foreach_neighb(a, n) {
864 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
865 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
867 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
875 typedef struct _appel_clique_walker_t {
877 const copy_opt_t *co;
884 } appel_clique_walker_t;
886 typedef struct _appel_block_info_t {
896 } appel_block_info_t;
898 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
901 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
902 int res = (int) freq;
903 return res == 0 ? 1 : res;
905 ir_loop *loop = get_irn_loop(bl);
907 int d = get_loop_depth(loop);
914 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
916 appel_block_info_t *res = NULL;
919 appel_clique_walker_t *d = (void *) phase;
920 res = phase_alloc(phase, sizeof(res[0]));
921 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
922 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
923 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
924 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
925 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
926 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
932 typedef struct _insn_list_t {
934 struct list_head list;
937 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
939 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
942 for(i = 0; i < bli->n_live_end; ++i)
943 if(bli->live_end[i] == irn)
944 return bli->live_end_nr[i];
949 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
951 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
957 foreach_pset(live, irn)
958 live_arr[n_live++] = irn;
960 /* dump the live after clique */
962 for(j = 0; j < n_live; ++j) {
965 for(k = j + 1; k < n_live; ++k) {
966 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
968 fprintf(env->f, "\n");
972 /* dump the affinities */
973 for(j = 0; !env->dumb && j < n_live; ++j) {
974 ir_node *irn = live_arr[j];
975 int old_nr = PTR_TO_INT(get_irn_link(irn));
977 /* if the node was already live in the last insn dump the affinity */
978 if(old_nr > start_nr) {
979 int weight = appel_aff_weight(env, bl);
980 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
984 /* set the current numbers into the link field. */
985 for(j = 0; j < n_live; ++j) {
986 ir_node *irn = live_arr[j];
987 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
990 return curr_nr + n_live;
993 static void appel_walker(ir_node *bl, void *data)
995 appel_clique_walker_t *env = data;
996 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
997 struct obstack *obst = &env->obst;
998 void *base = obstack_base(obst);
999 pset *live = pset_new_ptr_default();
1003 int start_nr = env->curr_nr;
1004 int curr_nr = start_nr;
1006 be_insn_env_t insn_env;
1011 insn_env.aenv = env->co->aenv;
1012 insn_env.cls = env->co->cls;
1013 insn_env.obst = obst;
1014 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1016 /* Guess how many insns will be in this block. */
1017 sched_foreach(bl, irn)
1021 insns = malloc(n_nodes * sizeof(insns[0]));
1023 /* Put all insns in an array. */
1024 irn = sched_first(bl);
1025 while(!sched_is_end(irn)) {
1027 insn = be_scan_insn(&insn_env, irn);
1028 insns[n_insns++] = insn;
1029 irn = insn->next_insn;
1032 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1033 be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
1035 /* Generate the bad and ugly. */
1036 for(i = n_insns - 1; i >= 0; --i) {
1037 be_insn_t *insn = insns[i];
1039 /* The first live set has to be saved in the block border set. */
1040 if(i == n_insns - 1) {
1042 foreach_pset(live, irn) {
1043 bli->live_end[j] = irn;
1044 bli->live_end_nr[j] = curr_nr + j;
1047 bli->n_live_end = j;
1051 for(j = 0; j < insn->use_start; ++j) {
1052 ir_node *op = insn->ops[j].carrier;
1053 bitset_t *adm = insn->ops[j].regs;
1057 if(!insn->ops[j].has_constraints)
1061 foreach_pset(live, irn) {
1069 assert(nr < pset_count(live));
1071 for(k = 0; k < env->co->cls->n_regs; ++k) {
1072 int mapped_col = env->color_map[k];
1073 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1074 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1079 /* dump the clique and update the stuff. */
1080 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1082 /* remove all defs. */
1083 for(j = 0; j < insn->use_start; ++j)
1084 pset_remove_ptr(live, insn->ops[j].carrier);
1086 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1087 bli->phi[bli->n_phi] = insn->irn;
1088 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1094 for(j = insn->use_start; j < insn->n_ops; ++j)
1095 pset_insert_ptr(live, insn->ops[j].carrier);
1098 /* print the start clique. */
1099 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1102 foreach_pset(live, irn) {
1103 bli->live_in[i] = irn;
1104 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1111 obstack_free(obst, base);
1112 env->curr_nr = curr_nr;
1115 static void appel_inter_block_aff(ir_node *bl, void *data)
1117 appel_clique_walker_t *env = data;
1118 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1122 for(i = 0; i < bli->n_live_in; ++i) {
1123 ir_node *irn = bli->live_in[i];
1125 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1126 ir_node *pred = get_Block_cfgpred_block(bl, j);
1128 int nr = appel_get_live_end_nr(env, pred, irn);
1130 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1134 for(i = 0; i < bli->n_phi; ++i) {
1135 ir_node *irn = bli->phi[i];
1137 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1138 ir_node *pred = get_Block_cfgpred_block(bl, j);
1139 ir_node *op = get_irn_n(irn, j);
1141 int nr = appel_get_live_end_nr(env, pred, op);
1143 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1149 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1153 appel_clique_walker_t env;
1154 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1156 be_liveness_recompute(co->cenv->lv);
1157 obstack_init(&env.obst);
1158 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1159 env.curr_nr = co->cls->n_regs;
1163 bitset_copy(adm, co->cenv->ignore_colors);
1164 bitset_flip_all(adm);
1166 /* Make color map. */
1167 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1168 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1169 const arch_register_t *reg = &co->cls->regs[i];
1170 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1174 env.curr_nr = n_colors;
1175 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1176 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1178 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1180 /* make the first k nodes interfere */
1181 for(i = 0; i < n_colors; ++i) {
1183 for(j = i + 1; j < n_colors; ++j)
1184 fprintf(f, "%d %d -1 ", i, j);
1189 env.curr_nr = n_colors;
1190 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1191 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1192 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1193 obstack_free(&env.obst, NULL);
1197 ___ _____ ____ ____ ___ _____ ____ _
1198 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1199 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1200 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1201 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1205 static const char *get_dot_color_name(int col)
1207 static const char *names[] = {
1241 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1244 typedef struct _co_ifg_dump_t {
1245 const copy_opt_t *co;
1249 static void ifg_dump_graph_attr(FILE *f, void *self)
1251 fprintf(f, "overlap=scale");
1254 static int ifg_is_dump_node(void *self, ir_node *irn)
1256 co_ifg_dump_t *cod = self;
1257 return !arch_irn_is(cod->co->aenv, irn, ignore);
1260 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1262 co_ifg_dump_t *env = self;
1263 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1264 arch_register_req_t req;
1267 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1268 limited = arch_register_req_is(&req, limited);
1270 if(env->flags & CO_IFG_DUMP_LABELS) {
1271 ir_fprintf(f, "label=\"%+F", irn);
1273 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1274 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1275 req.limited(req.limited_env, bs);
1276 ir_fprintf(f, "\\n%B", bs);
1278 ir_fprintf(f, "\" ");
1282 fprintf(f, "label=\"\" shape=point " );
1284 if(env->flags & CO_IFG_DUMP_SHAPE)
1285 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1287 if(env->flags & CO_IFG_DUMP_COLORS)
1288 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1291 static void ifg_dump_at_end(FILE *file, void *self)
1293 co_ifg_dump_t *env = self;
1296 co_gs_foreach_aff_node(env->co, a) {
1297 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1298 unsigned aidx = get_irn_idx(a->irn);
1301 co_gs_foreach_neighb(a, n) {
1302 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1303 unsigned nidx = get_irn_idx(n->irn);
1306 const char *color = nr == ar ? "blue" : "red";
1307 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1308 if(env->flags & CO_IFG_DUMP_LABELS)
1309 fprintf(file, "label=\"%d\" ", n->costs);
1310 if(env->flags & CO_IFG_DUMP_COLORS)
1311 fprintf(file, "color=%s ", color);
1313 fprintf(file, "style=dotted");
1314 fprintf(file, "];\n");
1321 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1323 ifg_dump_graph_attr,
1332 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1338 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1342 void co_solve_park_moon(copy_opt_t *opt)
1347 static int void_algo(copy_opt_t *co)
1354 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1355 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1356 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1357 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1364 int can_improve_existing;
1367 static co_algo_info_t algos[] = {
1368 { void_algo, "none", 0 },
1369 { co_solve_heuristic, "heur1", 0 },
1370 { co_solve_heuristic_new, "heur2", 0 },
1371 { co_solve_heuristic_java, "heur3", 0 },
1373 { co_solve_ilp2, "ilp", 1 },
1380 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1381 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1382 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1383 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1387 void co_driver(be_chordal_env_t *cenv)
1390 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1392 co_complete_stats_t before, after;
1394 co_algo_t *algo_func;
1395 int was_optimal = 0;
1397 if(algo < 0 || algo >= CO_ALGO_LAST)
1400 co = new_copy_opt(cenv, cost_func);
1401 co_build_ou_structure(co);
1402 co_build_graph_structure(co);
1404 co_complete_stats(co, &before);
1406 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1407 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1408 be_stat_ev_ull("co_max_costs", before.max_costs);
1409 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1410 be_stat_ev_ull("co_aff_int", before.aff_int);
1412 be_stat_ev_ull("co_init_costs", before.costs);
1413 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1415 /* Dump the interference graph in Appel's format. */
1416 if(dump_flags & DUMP_APPEL) {
1417 FILE *f = be_chordal_open(cenv, "", ".apl");
1418 co_dump_appel_graph(co, f);
1422 if(dump_flags & DUMP_BEFORE) {
1423 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1424 co_dump_ifg_dot(co, f, style_flags);
1428 /* if the algo can improve results, provide an initial solution with heur3 */
1429 if(improve && algos[algo].can_improve_existing) {
1430 co_complete_stats_t stats;
1432 /* produce a heuristical solution */
1433 co_solve_heuristic_java(co);
1435 /* do the stats and provide the current costs */
1436 co_complete_stats(co, &stats);
1437 be_stat_ev_ull("co_prepare_costs", stats.costs);
1440 /* start the JVM here so that it does not tamper the timing. */
1441 if(algo == CO_ALGO_HEUR3)
1442 be_java_coal_start_jvm();
1444 algo_func = algos[algo].algo;
1447 lc_timer_reset_and_start(timer);
1450 was_optimal = algo_func(co);
1453 lc_timer_stop(timer);
1454 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1457 be_stat_ev_ull("co_optimal", was_optimal);
1459 if(dump_flags & DUMP_AFTER) {
1460 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1461 co_dump_ifg_dot(co, f, style_flags);
1465 co_complete_stats(co, &after);
1468 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1469 ulong64 evitable = after.costs - after.inevit_costs;
1471 ir_printf("%30F ", cenv->irg);
1472 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1474 if(optimizable_costs > 0)
1475 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1477 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1480 be_stat_ev_ull("co_after_costs", after.costs);
1481 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1483 co_free_graph_structure(co);
1484 co_free_ou_structure(co);