4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
15 #include "raw_bitset.h"
20 #include "iredges_t.h"
23 #include "irphase_t.h"
24 #include "irprintf_t.h"
31 #include "becopyopt_t.h"
32 #include "becopystat.h"
35 #include "besched_t.h"
36 #include "benodesets.h"
37 #include "bejavacoal.h"
38 #include "bestatevent.h"
42 #include <libcore/lc_timing.h>
43 #include <libcore/lc_opts.h>
44 #include <libcore/lc_opts_enum.h>
49 #define DUMP_ALL 2 * DUMP_APPEL - 1
51 #define COST_FUNC_FREQ 1
52 #define COST_FUNC_LOOP 2
53 #define COST_FUNC_ALL_ONE 3
55 static unsigned dump_flags = 0;
56 static unsigned style_flags = 0;
57 static unsigned do_stats = 0;
58 static cost_fct_t cost_func = co_get_costs_exec_freq;
59 static unsigned algo = CO_ALGO_HEUR;
60 static int improve = 1;
62 static const lc_opt_enum_mask_items_t dump_items[] = {
63 { "before", DUMP_BEFORE },
64 { "after", DUMP_AFTER },
65 { "appel", DUMP_APPEL },
70 static const lc_opt_enum_mask_items_t style_items[] = {
71 { "color", CO_IFG_DUMP_COLORS },
72 { "labels", CO_IFG_DUMP_LABELS },
73 { "constr", CO_IFG_DUMP_CONSTR },
74 { "shape", CO_IFG_DUMP_SHAPE },
75 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
79 static const lc_opt_enum_mask_items_t algo_items[] = {
80 { "none", CO_ALGO_NONE },
81 { "heur", CO_ALGO_HEUR },
82 { "heur2", CO_ALGO_HEUR2 },
84 { "heur3", CO_ALGO_HEUR3 },
86 { "heur4", CO_ALGO_HEUR4 },
88 { "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", &algo_var),
120 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &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 be_init_copycoal(void)
135 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
136 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
137 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
138 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
140 lc_opt_add_table(co_grp, options);
143 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
145 #undef QUICK_AND_DIRTY_HACK
147 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
150 return be_ifg_connected(env->ifg, a, b);
152 return values_interfere(env->birg->lv, a, b);
156 /******************************************************************************
159 | | __ ___ _ __ ___ _ __ __ _| |
160 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
161 | |__| | __/ | | | __/ | | (_| | |
162 \_____|\___|_| |_|\___|_| \__,_|_|
164 ******************************************************************************/
166 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
169 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
171 const char *s1, *s2, *s3;
175 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
177 co = xcalloc(1, sizeof(*co));
178 co->cenv = chordal_env;
179 co->aenv = chordal_env->birg->main_env->arch_env;
180 co->irg = chordal_env->irg;
181 co->cls = chordal_env->cls;
182 co->get_costs = get_costs;
184 s1 = get_irp_prog_name();
185 s2 = get_entity_name(get_irg_entity(co->irg));
186 s3 = chordal_env->cls->name;
187 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
188 co->name = xmalloc(len);
189 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
194 void free_copy_opt(copy_opt_t *co) {
199 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
200 const arch_register_req_t *req;
201 const arch_register_t *reg;
203 if (arch_irn_is(co->aenv, irn, ignore))
206 reg = arch_get_irn_register(co->aenv, irn);
207 if (arch_register_type_is(reg, ignore))
210 req = arch_get_register_req(co->aenv, irn, -1);
211 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req))
217 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
218 const ir_edge_t *edge;
219 const arch_register_t *reg;
221 assert(0 && "Is buggy and obsolete. Do not use");
223 if (arch_irn_is(co->aenv, irn, ignore))
226 reg = arch_get_irn_register(co->aenv, irn);
227 if (arch_register_type_is(reg, ignore))
230 foreach_out_edge(irn, edge) {
231 ir_node *n = edge->src;
233 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
234 const arch_register_req_t *req;
235 req = arch_get_register_req(co->aenv, n, -1);
238 is_Perm(co->aenv, n) ||
239 (arch_register_req_is(req, should_be_same))) {
240 ir_node *other = get_irn_n(irn, req->other_same);
250 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
253 ir_node *root_block = get_nodes_block(root);
256 /* for phis the copies are placed in the corresponding pred-block */
257 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
259 /* a perm places the copy in the same block as it resides */
260 loop = get_irn_loop(root_block);
263 int d = get_loop_depth(loop);
269 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
271 ir_node *root_bl = get_nodes_block(root);
272 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
273 res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
275 /* don't allow values smaller than one. */
276 return res < 1 ? 1 : res;
280 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
284 /******************************************************************************
285 ____ _ _ _ _ _ _____ _
286 / __ \ | | | | | | (_) | / ____| |
287 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
288 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
289 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
290 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
293 ******************************************************************************/
296 * Determines a maximum weighted independent set with respect to
297 * the interference and conflict edges of all nodes in a qnode.
299 static int ou_max_ind_set_costs(unit_t *ou) {
300 be_chordal_env_t *chordal_env = ou->co->cenv;
301 ir_node **safe, **unsafe;
302 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
304 int max, pos, curr_weight, best_weight = 0;
306 /* assign the nodes into two groups.
307 * safe: node has no interference, hence it is in every max stable set.
308 * unsafe: node has an interference
310 safe = alloca((ou->node_count-1) * sizeof(*safe));
313 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
314 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
316 for(i=1; i<ou->node_count; ++i) {
318 for(o=1; o<ou->node_count; ++o) {
321 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
322 unsafe_costs[unsafe_count] = ou->costs[i];
323 unsafe[unsafe_count] = ou->nodes[i];
330 safe_costs += ou->costs[i];
331 safe[safe_count++] = ou->nodes[i];
336 /* now compute the best set out of the unsafe nodes*/
337 if (unsafe_count > MIS_HEUR_TRIGGER) {
338 bitset_t *best = bitset_alloca(unsafe_count);
339 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
340 for (i=0; i<unsafe_count; ++i) {
342 /* check if it is a stable set */
343 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
344 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
345 bitset_clear(best, i); /* clear the bit and try next one */
349 /* compute the weight */
350 bitset_foreach(best, pos)
351 best_weight += unsafe_costs[pos];
353 /* Exact Algorithm: Brute force */
354 curr = bitset_alloca(unsafe_count);
355 bitset_set_all(curr);
356 while ((max = bitset_popcnt(curr)) != 0) {
357 /* check if curr is a stable set */
358 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
359 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) */
360 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
363 /* if we arrive here, we have a stable set */
364 /* compute the weigth of the stable set*/
366 bitset_foreach(curr, pos)
367 curr_weight += unsafe_costs[pos];
370 if (curr_weight > best_weight) {
371 best_weight = curr_weight;
379 return safe_costs+best_weight;
382 static void co_collect_units(ir_node *irn, void *env) {
383 copy_opt_t *co = env;
386 if (!is_curr_reg_class(co, irn))
388 if (!co_is_optimizable_root(co, irn))
391 /* Init a new unit */
392 unit = xcalloc(1, sizeof(*unit));
394 unit->node_count = 1;
395 INIT_LIST_HEAD(&unit->queue);
397 /* Phi with some/all of its arguments */
398 if (is_Reg_Phi(irn)) {
402 arity = get_irn_arity(irn);
403 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
404 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
405 unit->nodes[0] = irn;
408 for (i=0; i<arity; ++i) {
410 ir_node *arg = get_irn_n(irn, i);
412 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
415 if (nodes_interfere(co->cenv, irn, arg)) {
416 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
420 /* Else insert the argument of the phi to the members of this ou */
421 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
423 /* Check if arg has occurred at a prior position in the arg/list */
425 for (o=0; o<unit->node_count; ++o)
426 if (unit->nodes[o] == arg) {
431 if (!arg_pos) { /* a new argument */
432 /* insert node, set costs */
433 unit->nodes[unit->node_count] = arg;
434 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
436 } else { /* arg has occured before in same phi */
437 /* increase costs for existing arg */
438 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
441 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
442 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
445 /* Proj of a perm with corresponding arg */
446 if (is_Perm_Proj(co->aenv, irn)) {
447 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
448 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
449 unit->costs = xmalloc(2 * sizeof(*unit->costs));
450 unit->node_count = 2;
451 unit->nodes[0] = irn;
452 unit->nodes[1] = get_Perm_src(irn);
453 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
455 const arch_register_req_t *req =
456 arch_get_register_req(co->aenv, irn, -1);
458 /* Src == Tgt of a 2-addr-code instruction */
459 if (is_2addr_code(req)) {
460 ir_node *other = get_irn_n(irn, req->other_same);
461 if (!arch_irn_is(co->aenv, other, ignore) &&
462 !nodes_interfere(co->cenv, irn, other)) {
463 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
464 unit->costs = xmalloc(2 * sizeof(*unit->costs));
465 unit->node_count = 2;
466 unit->nodes[0] = irn;
467 unit->nodes[1] = other;
468 unit->costs[1] = co->get_costs(co, irn, other, -1);
471 assert(0 && "This is not an optimizable node!");
475 /* Insert the new unit at a position according to its costs */
476 if (unit->node_count > 1) {
478 struct list_head *tmp;
480 /* Determine the maximum costs this unit can cause: all_nodes_cost */
481 for(i=1; i<unit->node_count; ++i) {
482 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
483 unit->all_nodes_costs += unit->costs[i];
486 /* Determine the minimal costs this unit will cause: min_nodes_costs */
487 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
488 /* Insert the new ou according to its sort_key */
490 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
492 list_add(&unit->units, tmp);
498 #ifdef QUICK_AND_DIRTY_HACK
500 static int compare_ous(const void *k1, const void *k2) {
501 const unit_t *u1 = *((const unit_t **) k1);
502 const unit_t *u2 = *((const unit_t **) k2);
503 int i, o, u1_has_constr, u2_has_constr;
504 arch_register_req_t req;
505 const arch_env_t *aenv = u1->co->aenv;
507 /* Units with constraints come first */
509 for (i=0; i<u1->node_count; ++i) {
510 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
511 if (arch_register_req_is(&req, limited)) {
518 for (i=0; i<u2->node_count; ++i) {
519 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
520 if (arch_register_req_is(&req, limited)) {
526 if (u1_has_constr != u2_has_constr)
527 return u2_has_constr - u1_has_constr;
529 /* Now check, whether the two units are connected */
531 for (i=0; i<u1->node_count; ++i)
532 for (o=0; o<u2->node_count; ++o)
533 if (u1->nodes[i] == u2->nodes[o])
537 /* After all, the sort key decides. Greater keys come first. */
538 return u2->sort_key - u1->sort_key;
543 * Sort the ou's according to constraints and their sort_key
545 static void co_sort_units(copy_opt_t *co) {
546 int i, count = 0, costs;
549 /* get the number of ous, remove them form the list and fill the array */
550 list_for_each_entry(unit_t, ou, &co->units, units)
552 ous = alloca(count * sizeof(*ous));
554 costs = co_get_max_copy_costs(co);
557 list_for_each_entry(unit_t, ou, &co->units, units)
560 INIT_LIST_HEAD(&co->units);
562 assert(count == i && list_empty(&co->units));
564 for (i=0; i<count; ++i)
565 ir_printf("%+F\n", ous[i]->nodes[0]);
567 qsort(ous, count, sizeof(*ous), compare_ous);
570 for (i=0; i<count; ++i)
571 ir_printf("%+F\n", ous[i]->nodes[0]);
573 /* reinsert into list in correct order */
574 for (i=0; i<count; ++i)
575 list_add_tail(&ous[i]->units, &co->units);
577 assert(costs == co_get_max_copy_costs(co));
581 void co_build_ou_structure(copy_opt_t *co) {
582 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
583 INIT_LIST_HEAD(&co->units);
584 irg_walk_graph(co->irg, co_collect_units, NULL, co);
585 #ifdef QUICK_AND_DIRTY_HACK
590 void co_free_ou_structure(copy_opt_t *co) {
593 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
598 co->units.next = NULL;
601 /* co_solve_heuristic() is implemented in becopyheur.c */
603 int co_get_max_copy_costs(const copy_opt_t *co) {
609 list_for_each_entry(unit_t, curr, &co->units, units) {
610 res += curr->inevitable_costs;
611 for (i=1; i<curr->node_count; ++i)
612 res += curr->costs[i];
617 int co_get_inevit_copy_costs(const copy_opt_t *co) {
623 list_for_each_entry(unit_t, curr, &co->units, units)
624 res += curr->inevitable_costs;
628 int co_get_copy_costs(const copy_opt_t *co) {
634 list_for_each_entry(unit_t, curr, &co->units, units) {
635 int root_col = get_irn_col(co, curr->nodes[0]);
636 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
637 res += curr->inevitable_costs;
638 for (i=1; i<curr->node_count; ++i) {
639 int arg_col = get_irn_col(co, curr->nodes[i]);
640 if (root_col != arg_col) {
641 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
642 res += curr->costs[i];
649 int co_get_lower_bound(const copy_opt_t *co) {
655 list_for_each_entry(unit_t, curr, &co->units, units)
656 res += curr->inevitable_costs + curr->min_nodes_costs;
660 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
662 bitset_t *seen = bitset_irg_malloc(co->irg);
665 memset(stat, 0, sizeof(stat[0]));
667 /* count affinity edges. */
668 co_gs_foreach_aff_node(co, an) {
670 stat->aff_nodes += 1;
671 bitset_add_irn(seen, an->irn);
672 co_gs_foreach_neighb(an, neigh) {
673 if(!bitset_contains_irn(seen, neigh->irn)) {
674 stat->aff_edges += 1;
675 stat->max_costs += neigh->costs;
677 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
678 stat->costs += neigh->costs;
679 stat->unsatisfied_edges += 1;
682 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
684 stat->inevit_costs += neigh->costs;
694 /******************************************************************************
696 / ____| | | / ____| |
697 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
698 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
699 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
700 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
703 ******************************************************************************/
705 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
706 const affinity_node_t *n1 = k1;
707 const affinity_node_t *n2 = k2;
709 return (n1->irn != n2->irn);
712 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
713 affinity_node_t new_node, *node;
714 neighb_t new_nbr, *nbr;
719 new_node.neighbours = NULL;
720 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
723 for (nbr = node->neighbours; nbr; nbr = nbr->next)
724 if (nbr->irn == n2) {
729 /* if we did not find n2 in n1's neighbourhood insert it */
731 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
732 nbr = obstack_finish(&co->obst);
735 nbr->next = node->neighbours;
736 node->neighbours = nbr;
740 /* now nbr points to n1's neighbour-entry of n2 */
744 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
745 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
746 add_edge(co, n1, n2, costs);
747 add_edge(co, n2, n1, costs);
751 static void build_graph_walker(ir_node *irn, void *env) {
752 copy_opt_t *co = env;
754 const arch_register_t *reg;
756 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
759 reg = arch_get_irn_register(co->aenv, irn);
760 if (arch_register_type_is(reg, ignore))
765 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
766 ir_node *arg = get_irn_n(irn, pos);
767 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
771 else if (is_Perm_Proj(co->aenv, irn)) {
772 ir_node *arg = get_Perm_src(irn);
773 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
778 const arch_register_req_t *req =
779 arch_get_register_req(co->aenv, irn, -1);
780 if (is_2addr_code(req)) {
781 ir_node *other = get_irn_n(irn, req->other_same);
782 if(!arch_irn_is(co->aenv, other, ignore))
783 add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
788 void co_build_graph_structure(copy_opt_t *co) {
789 obstack_init(&co->obst);
790 co->nodes = new_set(compare_affinity_node_t, 32);
792 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
795 void co_free_graph_structure(copy_opt_t *co) {
799 obstack_free(&co->obst, NULL);
803 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
805 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
806 affinity_node_t new_node, *n;
811 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
813 return (n->degree > 0);
818 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
820 be_ifg_t *ifg = co->cenv->ifg;
821 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
828 for(i = 0; i < co->cls->n_regs; ++i) {
829 const arch_register_t *reg = &co->cls->regs[i];
830 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
834 * n contains the first node number.
835 * the values below n are the pre-colored register nodes
838 it = be_ifg_nodes_iter_alloca(ifg);
839 nit = be_ifg_neighbours_iter_alloca(ifg);
842 be_ifg_foreach_node(ifg, it, irn) {
843 if(!arch_irn_is(co->aenv, irn, ignore))
844 set_irn_link(irn, INT_TO_PTR(n++));
847 fprintf(f, "%d %d\n", n, n_regs);
849 be_ifg_foreach_node(ifg, it, irn) {
850 if(!arch_irn_is(co->aenv, irn, ignore)) {
851 int idx = PTR_TO_INT(get_irn_link(irn));
852 affinity_node_t *a = get_affinity_info(co, irn);
854 const arch_register_req_t *req;
857 req = arch_get_register_req(co->aenv, irn, BE_OUT_POS(0));
858 if(arch_register_req_is(req, limited)) {
859 for(i = 0; i < co->cls->n_regs; ++i) {
860 if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
861 fprintf(f, "%d %d -1\n", color_map[i], idx);
866 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
867 if(!arch_irn_is(co->aenv, adj, ignore)) {
868 int adj_idx = PTR_TO_INT(get_irn_link(adj));
870 fprintf(f, "%d %d -1\n", idx, adj_idx);
877 co_gs_foreach_neighb(a, n) {
878 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
879 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
881 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
889 typedef struct _appel_clique_walker_t {
891 const copy_opt_t *co;
898 } appel_clique_walker_t;
900 typedef struct _appel_block_info_t {
910 } appel_block_info_t;
912 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
915 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
916 int res = (int) freq;
917 return res == 0 ? 1 : res;
919 ir_loop *loop = get_irn_loop(bl);
921 int d = get_loop_depth(loop);
928 static void *appel_clique_walker_irn_init(ir_phase *phase, ir_node *irn, void *old)
930 appel_block_info_t *res = NULL;
933 appel_clique_walker_t *d = (void *) phase;
934 res = phase_alloc(phase, sizeof(res[0]));
935 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
936 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
937 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
938 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
939 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
940 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
946 typedef struct _insn_list_t {
948 struct list_head list;
951 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
953 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
956 for(i = 0; i < bli->n_live_end; ++i)
957 if(bli->live_end[i] == irn)
958 return bli->live_end_nr[i];
963 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
965 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
971 foreach_pset(live, irn)
972 live_arr[n_live++] = irn;
974 /* dump the live after clique */
976 for(j = 0; j < n_live; ++j) {
979 for(k = j + 1; k < n_live; ++k) {
980 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
982 fprintf(env->f, "\n");
986 /* dump the affinities */
987 for(j = 0; !env->dumb && j < n_live; ++j) {
988 ir_node *irn = live_arr[j];
989 int old_nr = PTR_TO_INT(get_irn_link(irn));
991 /* if the node was already live in the last insn dump the affinity */
992 if(old_nr > start_nr) {
993 int weight = appel_aff_weight(env, bl);
994 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
998 /* set the current numbers into the link field. */
999 for(j = 0; j < n_live; ++j) {
1000 ir_node *irn = live_arr[j];
1001 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
1004 return curr_nr + n_live;
1007 static void appel_walker(ir_node *bl, void *data)
1009 appel_clique_walker_t *env = data;
1010 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
1011 struct obstack *obst = &env->obst;
1012 void *base = obstack_base(obst);
1013 pset *live = pset_new_ptr_default();
1014 be_lv_t *lv = env->co->cenv->birg->lv;
1018 int start_nr = env->curr_nr;
1019 int curr_nr = start_nr;
1021 be_insn_env_t insn_env;
1026 insn_env.aenv = env->co->aenv;
1027 insn_env.cls = env->co->cls;
1028 insn_env.obst = obst;
1029 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1031 /* Guess how many insns will be in this block. */
1032 sched_foreach(bl, irn)
1036 insns = malloc(n_nodes * sizeof(insns[0]));
1038 /* Put all insns in an array. */
1039 irn = sched_first(bl);
1040 while(!sched_is_end(irn)) {
1042 insn = be_scan_insn(&insn_env, irn);
1043 insns[n_insns++] = insn;
1044 irn = insn->next_insn;
1047 DBG((dbg, LEVEL_2, "%+F\n", bl));
1048 be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
1050 /* Generate the bad and ugly. */
1051 for(i = n_insns - 1; i >= 0; --i) {
1052 be_insn_t *insn = insns[i];
1054 /* The first live set has to be saved in the block border set. */
1055 if(i == n_insns - 1) {
1057 foreach_pset(live, irn) {
1058 bli->live_end[j] = irn;
1059 bli->live_end_nr[j] = curr_nr + j;
1062 bli->n_live_end = j;
1066 for(j = 0; j < insn->use_start; ++j) {
1067 ir_node *op = insn->ops[j].carrier;
1068 bitset_t *adm = insn->ops[j].regs;
1072 if(!insn->ops[j].has_constraints)
1076 foreach_pset(live, irn) {
1084 assert(nr < pset_count(live));
1086 for(k = 0; k < env->co->cls->n_regs; ++k) {
1087 int mapped_col = env->color_map[k];
1088 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1089 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1094 /* dump the clique and update the stuff. */
1095 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1097 /* remove all defs. */
1098 for(j = 0; j < insn->use_start; ++j)
1099 pset_remove_ptr(live, insn->ops[j].carrier);
1101 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1102 bli->phi[bli->n_phi] = insn->irn;
1103 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1109 for(j = insn->use_start; j < insn->n_ops; ++j)
1110 pset_insert_ptr(live, insn->ops[j].carrier);
1113 /* print the start clique. */
1114 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1117 foreach_pset(live, irn) {
1118 bli->live_in[i] = irn;
1119 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1126 obstack_free(obst, base);
1127 env->curr_nr = curr_nr;
1130 static void appel_inter_block_aff(ir_node *bl, void *data)
1132 appel_clique_walker_t *env = data;
1133 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1137 for(i = 0; i < bli->n_live_in; ++i) {
1138 ir_node *irn = bli->live_in[i];
1140 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1141 ir_node *pred = get_Block_cfgpred_block(bl, j);
1143 int nr = appel_get_live_end_nr(env, pred, irn);
1145 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1149 for(i = 0; i < bli->n_phi; ++i) {
1150 ir_node *irn = bli->phi[i];
1152 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1153 ir_node *pred = get_Block_cfgpred_block(bl, j);
1154 ir_node *op = get_irn_n(irn, j);
1156 int nr = appel_get_live_end_nr(env, pred, op);
1158 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1164 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1168 appel_clique_walker_t env;
1169 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1170 be_lv_t *lv = co->cenv->birg->lv;
1172 be_liveness_recompute(lv);
1173 obstack_init(&env.obst);
1174 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init, NULL);
1175 env.curr_nr = co->cls->n_regs;
1179 bitset_copy(adm, co->cenv->ignore_colors);
1180 bitset_flip_all(adm);
1182 /* Make color map. */
1183 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1184 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1185 const arch_register_t *reg = &co->cls->regs[i];
1186 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1190 env.curr_nr = n_colors;
1191 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1192 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1194 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1196 /* make the first k nodes interfere */
1197 for(i = 0; i < n_colors; ++i) {
1199 for(j = i + 1; j < n_colors; ++j)
1200 fprintf(f, "%d %d -1 ", i, j);
1205 env.curr_nr = n_colors;
1206 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1207 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1208 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1209 obstack_free(&env.obst, NULL);
1213 ___ _____ ____ ____ ___ _____ ____ _
1214 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1215 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1216 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1217 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1221 static const char *get_dot_color_name(int col)
1223 static const char *names[] = {
1257 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1260 typedef struct _co_ifg_dump_t {
1261 const copy_opt_t *co;
1265 static void ifg_dump_graph_attr(FILE *f, void *self)
1267 fprintf(f, "overlap=scale");
1270 static int ifg_is_dump_node(void *self, ir_node *irn)
1272 co_ifg_dump_t *cod = self;
1273 return !arch_irn_is(cod->co->aenv, irn, ignore);
1276 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1278 co_ifg_dump_t *env = self;
1279 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1280 const arch_register_req_t *req;
1283 req = arch_get_register_req(env->co->aenv, irn, BE_OUT_POS(0));
1284 limited = arch_register_req_is(req, limited);
1286 if(env->flags & CO_IFG_DUMP_LABELS) {
1287 ir_fprintf(f, "label=\"%+F", irn);
1291 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1292 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1293 req.limited(req.limited_env, bs);
1294 ir_fprintf(f, "\\n%B", bs);
1297 ir_fprintf(f, "\" ");
1299 fprintf(f, "label=\"\" shape=point " );
1302 if(env->flags & CO_IFG_DUMP_SHAPE)
1303 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1305 if(env->flags & CO_IFG_DUMP_COLORS)
1306 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1309 static void ifg_dump_at_end(FILE *file, void *self)
1311 co_ifg_dump_t *env = self;
1314 co_gs_foreach_aff_node(env->co, a) {
1315 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1316 unsigned aidx = get_irn_idx(a->irn);
1319 co_gs_foreach_neighb(a, n) {
1320 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1321 unsigned nidx = get_irn_idx(n->irn);
1324 const char *color = nr == ar ? "blue" : "red";
1325 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1326 if(env->flags & CO_IFG_DUMP_LABELS)
1327 fprintf(file, "label=\"%d\" ", n->costs);
1328 if(env->flags & CO_IFG_DUMP_COLORS)
1329 fprintf(file, "color=%s ", color);
1331 fprintf(file, "style=dotted");
1332 fprintf(file, "];\n");
1339 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1341 ifg_dump_graph_attr,
1350 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1356 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1360 void co_solve_park_moon(copy_opt_t *opt)
1365 static int void_algo(copy_opt_t *co)
1372 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1373 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1374 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1375 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1382 int can_improve_existing;
1385 static co_algo_info_t algos[] = {
1386 { void_algo, "none", 0 },
1387 { co_solve_heuristic, "heur1", 0 },
1388 { co_solve_heuristic_new, "heur2", 0 },
1389 { co_solve_heuristic_java, "heur3", 0 },
1390 { co_solve_heuristic_mst, "heur4", 0 },
1392 { co_solve_ilp2, "ilp", 1 },
1399 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1400 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1401 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1402 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1406 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
1411 ir_snprintf(buf, sizeof(buf), "%s%F_%s%s", prefix, env->irg, env->cls->name, suffix);
1412 result = fopen(buf, "wt");
1413 if(result == NULL) {
1414 panic("Couldn't open '%s' for writing.", buf);
1420 void co_driver(be_chordal_env_t *cenv)
1422 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1423 co_complete_stats_t before, after;
1425 co_algo_t *algo_func;
1426 int was_optimal = 0;
1428 if (algo < 0 || algo >= CO_ALGO_LAST)
1431 co = new_copy_opt(cenv, cost_func);
1432 co_build_ou_structure(co);
1433 co_build_graph_structure(co);
1435 co_complete_stats(co, &before);
1437 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1438 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1439 be_stat_ev_ull("co_max_costs", before.max_costs);
1440 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1441 be_stat_ev_ull("co_aff_int", before.aff_int);
1443 be_stat_ev_ull("co_init_costs", before.costs);
1444 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1446 /* Dump the interference graph in Appel's format. */
1447 if (dump_flags & DUMP_APPEL) {
1448 FILE *f = my_open(cenv, "", ".apl");
1449 co_dump_appel_graph(co, f);
1453 if (dump_flags & DUMP_BEFORE) {
1454 FILE *f = my_open(cenv, "", "-before.dot");
1455 co_dump_ifg_dot(co, f, style_flags);
1459 /* if the algo can improve results, provide an initial solution with heur3 */
1460 if (improve && algos[algo].can_improve_existing) {
1461 co_complete_stats_t stats;
1463 /* produce a heuristic solution */
1465 co_solve_heuristic_java(co);
1467 co_solve_heuristic(co);
1468 #endif /* WITH_JVM */
1470 /* do the stats and provide the current costs */
1471 co_complete_stats(co, &stats);
1472 be_stat_ev_ull("co_prepare_costs", stats.costs);
1476 /* start the JVM here so that it does not tamper the timing. */
1477 if (algo == CO_ALGO_HEUR3)
1478 be_java_coal_start_jvm();
1479 #endif /* WITH_JVM */
1481 algo_func = algos[algo].algo;
1483 /* perform actual copy minimization */
1484 lc_timer_reset_and_start(timer);
1485 was_optimal = algo_func(co);
1486 lc_timer_stop(timer);
1488 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1489 be_stat_ev_ull("co_optimal", was_optimal);
1491 if (dump_flags & DUMP_AFTER) {
1492 FILE *f = my_open(cenv, "", "-after.dot");
1493 co_dump_ifg_dot(co, f, style_flags);
1497 co_complete_stats(co, &after);
1500 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1501 ulong64 evitable = after.costs - after.inevit_costs;
1503 ir_printf("%30F ", cenv->irg);
1504 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1506 if(optimizable_costs > 0)
1507 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1509 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1512 be_stat_ev_ull("co_after_costs", after.costs);
1513 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1515 co_free_graph_structure(co);
1516 co_free_ou_structure(co);