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_HEUR2;
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 },
87 { "ilp", CO_ALGO_ILP },
92 typedef int (*opt_funcptr)(void);
94 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
95 { "freq", (opt_funcptr) co_get_costs_exec_freq },
96 { "loop", (opt_funcptr) co_get_costs_loop_depth },
97 { "one", (opt_funcptr) co_get_costs_all_one },
101 static lc_opt_enum_mask_var_t dump_var = {
102 &dump_flags, dump_items
105 static lc_opt_enum_mask_var_t style_var = {
106 &style_flags, style_items
109 static lc_opt_enum_mask_var_t algo_var = {
113 static lc_opt_enum_func_ptr_var_t cost_func_var = {
114 (opt_funcptr*) &cost_func, cost_func_items
117 static const lc_opt_table_entry_t options[] = {
118 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var),
119 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
120 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
121 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
122 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
123 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
127 /* Insert additional options registration functions here. */
128 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
129 extern void be_co2_register_options(lc_opt_entry_t *grp);
130 extern void be_co3_register_options(lc_opt_entry_t *grp);
132 void be_init_copycoal(void)
134 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
135 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
136 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
137 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
139 lc_opt_add_table(co_grp, options);
142 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copycoal);
144 #undef QUICK_AND_DIRTY_HACK
146 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
149 return be_ifg_connected(env->ifg, a, b);
151 return values_interfere(env->birg->lv, a, b);
155 /******************************************************************************
158 | | __ ___ _ __ ___ _ __ __ _| |
159 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
160 | |__| | __/ | | | __/ | | (_| | |
161 \_____|\___|_| |_|\___|_| \__,_|_|
163 ******************************************************************************/
165 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
168 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
170 const char *s1, *s2, *s3;
174 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
176 co = xcalloc(1, sizeof(*co));
177 co->cenv = chordal_env;
178 co->aenv = chordal_env->birg->main_env->arch_env;
179 co->irg = chordal_env->irg;
180 co->cls = chordal_env->cls;
181 co->get_costs = get_costs;
183 s1 = get_irp_prog_name();
184 s2 = get_entity_name(get_irg_entity(co->irg));
185 s3 = chordal_env->cls->name;
186 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
187 co->name = xmalloc(len);
188 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
193 void free_copy_opt(copy_opt_t *co) {
198 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
199 const arch_register_req_t *req;
200 const arch_register_t *reg;
202 if (arch_irn_is(co->aenv, irn, ignore))
205 reg = arch_get_irn_register(co->aenv, irn);
206 if (arch_register_type_is(reg, ignore))
209 req = arch_get_register_req(co->aenv, irn, -1);
210 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(req))
216 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
217 const ir_edge_t *edge;
218 const arch_register_t *reg;
220 assert(0 && "Is buggy and obsolete. Do not use");
222 if (arch_irn_is(co->aenv, irn, ignore))
225 reg = arch_get_irn_register(co->aenv, irn);
226 if (arch_register_type_is(reg, ignore))
229 foreach_out_edge(irn, edge) {
230 ir_node *n = edge->src;
232 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
233 const arch_register_req_t *req;
234 req = arch_get_register_req(co->aenv, n, -1);
237 is_Perm(co->aenv, n) ||
238 (arch_register_req_is(req, should_be_same))) {
239 ir_node *other = get_irn_n(irn, req->other_same);
249 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
252 ir_node *root_block = get_nodes_block(root);
255 /* for phis the copies are placed in the corresponding pred-block */
256 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
258 /* a perm places the copy in the same block as it resides */
259 loop = get_irn_loop(root_block);
262 int d = get_loop_depth(loop);
268 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
270 ir_node *root_bl = get_nodes_block(root);
271 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
272 res = get_block_execfreq_ulong(co->cenv->birg->exec_freq, copy_bl);
274 /* don't allow values smaller than one. */
275 return res < 1 ? 1 : res;
279 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
283 /******************************************************************************
284 ____ _ _ _ _ _ _____ _
285 / __ \ | | | | | | (_) | / ____| |
286 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
287 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
288 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
289 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
292 ******************************************************************************/
295 * Determines a maximum weighted independent set with respect to
296 * the interference and conflict edges of all nodes in a qnode.
298 static int ou_max_ind_set_costs(unit_t *ou) {
299 be_chordal_env_t *chordal_env = ou->co->cenv;
300 ir_node **safe, **unsafe;
301 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
303 int max, pos, curr_weight, best_weight = 0;
305 /* assign the nodes into two groups.
306 * safe: node has no interference, hence it is in every max stable set.
307 * unsafe: node has an interference
309 safe = alloca((ou->node_count-1) * sizeof(*safe));
312 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
313 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
315 for(i=1; i<ou->node_count; ++i) {
317 for(o=1; o<ou->node_count; ++o) {
320 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
321 unsafe_costs[unsafe_count] = ou->costs[i];
322 unsafe[unsafe_count] = ou->nodes[i];
329 safe_costs += ou->costs[i];
330 safe[safe_count++] = ou->nodes[i];
335 /* now compute the best set out of the unsafe nodes*/
336 if (unsafe_count > MIS_HEUR_TRIGGER) {
337 bitset_t *best = bitset_alloca(unsafe_count);
338 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
339 for (i=0; i<unsafe_count; ++i) {
341 /* check if it is a stable set */
342 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
343 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
344 bitset_clear(best, i); /* clear the bit and try next one */
348 /* compute the weight */
349 bitset_foreach(best, pos)
350 best_weight += unsafe_costs[pos];
352 /* Exact Algorithm: Brute force */
353 curr = bitset_alloca(unsafe_count);
354 bitset_set_all(curr);
355 while ((max = bitset_popcnt(curr)) != 0) {
356 /* check if curr is a stable set */
357 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
358 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) */
359 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
362 /* if we arrive here, we have a stable set */
363 /* compute the weigth of the stable set*/
365 bitset_foreach(curr, pos)
366 curr_weight += unsafe_costs[pos];
369 if (curr_weight > best_weight) {
370 best_weight = curr_weight;
378 return safe_costs+best_weight;
381 static void co_collect_units(ir_node *irn, void *env) {
382 copy_opt_t *co = env;
385 if (!is_curr_reg_class(co, irn))
387 if (!co_is_optimizable_root(co, irn))
390 /* Init a new unit */
391 unit = xcalloc(1, sizeof(*unit));
393 unit->node_count = 1;
394 INIT_LIST_HEAD(&unit->queue);
396 /* Phi with some/all of its arguments */
397 if (is_Reg_Phi(irn)) {
401 arity = get_irn_arity(irn);
402 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
403 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
404 unit->nodes[0] = irn;
407 for (i=0; i<arity; ++i) {
409 ir_node *arg = get_irn_n(irn, i);
411 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
414 if (nodes_interfere(co->cenv, irn, arg)) {
415 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
419 /* Else insert the argument of the phi to the members of this ou */
420 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
422 /* Check if arg has occurred at a prior position in the arg/list */
424 for (o=0; o<unit->node_count; ++o)
425 if (unit->nodes[o] == arg) {
430 if (!arg_pos) { /* a new argument */
431 /* insert node, set costs */
432 unit->nodes[unit->node_count] = arg;
433 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
435 } else { /* arg has occured before in same phi */
436 /* increase costs for existing arg */
437 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
440 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
441 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
444 /* Proj of a perm with corresponding arg */
445 if (is_Perm_Proj(co->aenv, irn)) {
446 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
447 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
448 unit->costs = xmalloc(2 * sizeof(*unit->costs));
449 unit->node_count = 2;
450 unit->nodes[0] = irn;
451 unit->nodes[1] = get_Perm_src(irn);
452 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
454 const arch_register_req_t *req =
455 arch_get_register_req(co->aenv, irn, -1);
457 /* Src == Tgt of a 2-addr-code instruction */
458 if (is_2addr_code(req)) {
459 ir_node *other = get_irn_n(irn, req->other_same);
460 if (!arch_irn_is(co->aenv, other, ignore) &&
461 !nodes_interfere(co->cenv, irn, other)) {
462 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
463 unit->costs = xmalloc(2 * sizeof(*unit->costs));
464 unit->node_count = 2;
465 unit->nodes[0] = irn;
466 unit->nodes[1] = other;
467 unit->costs[1] = co->get_costs(co, irn, other, -1);
470 assert(0 && "This is not an optimizable node!");
474 /* Insert the new unit at a position according to its costs */
475 if (unit->node_count > 1) {
477 struct list_head *tmp;
479 /* Determine the maximum costs this unit can cause: all_nodes_cost */
480 for(i=1; i<unit->node_count; ++i) {
481 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
482 unit->all_nodes_costs += unit->costs[i];
485 /* Determine the minimal costs this unit will cause: min_nodes_costs */
486 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
487 /* Insert the new ou according to its sort_key */
489 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
491 list_add(&unit->units, tmp);
497 #ifdef QUICK_AND_DIRTY_HACK
499 static int compare_ous(const void *k1, const void *k2) {
500 const unit_t *u1 = *((const unit_t **) k1);
501 const unit_t *u2 = *((const unit_t **) k2);
502 int i, o, u1_has_constr, u2_has_constr;
503 arch_register_req_t req;
504 const arch_env_t *aenv = u1->co->aenv;
506 /* Units with constraints come first */
508 for (i=0; i<u1->node_count; ++i) {
509 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
510 if (arch_register_req_is(&req, limited)) {
517 for (i=0; i<u2->node_count; ++i) {
518 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
519 if (arch_register_req_is(&req, limited)) {
525 if (u1_has_constr != u2_has_constr)
526 return u2_has_constr - u1_has_constr;
528 /* Now check, whether the two units are connected */
530 for (i=0; i<u1->node_count; ++i)
531 for (o=0; o<u2->node_count; ++o)
532 if (u1->nodes[i] == u2->nodes[o])
536 /* After all, the sort key decides. Greater keys come first. */
537 return u2->sort_key - u1->sort_key;
542 * Sort the ou's according to constraints and their sort_key
544 static void co_sort_units(copy_opt_t *co) {
545 int i, count = 0, costs;
548 /* get the number of ous, remove them form the list and fill the array */
549 list_for_each_entry(unit_t, ou, &co->units, units)
551 ous = alloca(count * sizeof(*ous));
553 costs = co_get_max_copy_costs(co);
556 list_for_each_entry(unit_t, ou, &co->units, units)
559 INIT_LIST_HEAD(&co->units);
561 assert(count == i && list_empty(&co->units));
563 for (i=0; i<count; ++i)
564 ir_printf("%+F\n", ous[i]->nodes[0]);
566 qsort(ous, count, sizeof(*ous), compare_ous);
569 for (i=0; i<count; ++i)
570 ir_printf("%+F\n", ous[i]->nodes[0]);
572 /* reinsert into list in correct order */
573 for (i=0; i<count; ++i)
574 list_add_tail(&ous[i]->units, &co->units);
576 assert(costs == co_get_max_copy_costs(co));
580 void co_build_ou_structure(copy_opt_t *co) {
581 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
582 INIT_LIST_HEAD(&co->units);
583 irg_walk_graph(co->irg, co_collect_units, NULL, co);
584 #ifdef QUICK_AND_DIRTY_HACK
589 void co_free_ou_structure(copy_opt_t *co) {
592 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
597 co->units.next = NULL;
600 /* co_solve_heuristic() is implemented in becopyheur.c */
602 int co_get_max_copy_costs(const copy_opt_t *co) {
608 list_for_each_entry(unit_t, curr, &co->units, units) {
609 res += curr->inevitable_costs;
610 for (i=1; i<curr->node_count; ++i)
611 res += curr->costs[i];
616 int co_get_inevit_copy_costs(const copy_opt_t *co) {
622 list_for_each_entry(unit_t, curr, &co->units, units)
623 res += curr->inevitable_costs;
627 int co_get_copy_costs(const copy_opt_t *co) {
633 list_for_each_entry(unit_t, curr, &co->units, units) {
634 int root_col = get_irn_col(co, curr->nodes[0]);
635 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
636 res += curr->inevitable_costs;
637 for (i=1; i<curr->node_count; ++i) {
638 int arg_col = get_irn_col(co, curr->nodes[i]);
639 if (root_col != arg_col) {
640 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
641 res += curr->costs[i];
648 int co_get_lower_bound(const copy_opt_t *co) {
654 list_for_each_entry(unit_t, curr, &co->units, units)
655 res += curr->inevitable_costs + curr->min_nodes_costs;
659 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
661 bitset_t *seen = bitset_irg_malloc(co->irg);
664 memset(stat, 0, sizeof(stat[0]));
666 /* count affinity edges. */
667 co_gs_foreach_aff_node(co, an) {
669 stat->aff_nodes += 1;
670 bitset_add_irn(seen, an->irn);
671 co_gs_foreach_neighb(an, neigh) {
672 if(!bitset_contains_irn(seen, neigh->irn)) {
673 stat->aff_edges += 1;
674 stat->max_costs += neigh->costs;
676 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
677 stat->costs += neigh->costs;
678 stat->unsatisfied_edges += 1;
681 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
683 stat->inevit_costs += neigh->costs;
693 /******************************************************************************
695 / ____| | | / ____| |
696 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
697 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
698 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
699 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
702 ******************************************************************************/
704 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
705 const affinity_node_t *n1 = k1;
706 const affinity_node_t *n2 = k2;
708 return (n1->irn != n2->irn);
711 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
712 affinity_node_t new_node, *node;
713 neighb_t new_nbr, *nbr;
718 new_node.neighbours = NULL;
719 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
722 for (nbr = node->neighbours; nbr; nbr = nbr->next)
723 if (nbr->irn == n2) {
728 /* if we did not find n2 in n1's neighbourhood insert it */
730 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
731 nbr = obstack_finish(&co->obst);
734 nbr->next = node->neighbours;
735 node->neighbours = nbr;
739 /* now nbr points to n1's neighbour-entry of n2 */
743 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
744 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
745 add_edge(co, n1, n2, costs);
746 add_edge(co, n2, n1, costs);
750 static void build_graph_walker(ir_node *irn, void *env) {
751 copy_opt_t *co = env;
753 const arch_register_t *reg;
755 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
758 reg = arch_get_irn_register(co->aenv, irn);
759 if (arch_register_type_is(reg, ignore))
764 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
765 ir_node *arg = get_irn_n(irn, pos);
766 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
770 else if (is_Perm_Proj(co->aenv, irn)) {
771 ir_node *arg = get_Perm_src(irn);
772 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
777 const arch_register_req_t *req =
778 arch_get_register_req(co->aenv, irn, -1);
779 if (is_2addr_code(req)) {
780 ir_node *other = get_irn_n(irn, req->other_same);
781 if(!arch_irn_is(co->aenv, other, ignore))
782 add_edges(co, irn, other, co->get_costs(co, irn, other, 0));
787 void co_build_graph_structure(copy_opt_t *co) {
788 obstack_init(&co->obst);
789 co->nodes = new_set(compare_affinity_node_t, 32);
791 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
794 void co_free_graph_structure(copy_opt_t *co) {
798 obstack_free(&co->obst, NULL);
802 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
804 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
805 affinity_node_t new_node, *n;
810 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
812 return (n->degree > 0);
817 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
819 be_ifg_t *ifg = co->cenv->ifg;
820 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
827 for(i = 0; i < co->cls->n_regs; ++i) {
828 const arch_register_t *reg = &co->cls->regs[i];
829 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
833 * n contains the first node number.
834 * the values below n are the pre-colored register nodes
837 it = be_ifg_nodes_iter_alloca(ifg);
838 nit = be_ifg_neighbours_iter_alloca(ifg);
841 be_ifg_foreach_node(ifg, it, irn) {
842 if(!arch_irn_is(co->aenv, irn, ignore))
843 set_irn_link(irn, INT_TO_PTR(n++));
846 fprintf(f, "%d %d\n", n, n_regs);
848 be_ifg_foreach_node(ifg, it, irn) {
849 if(!arch_irn_is(co->aenv, irn, ignore)) {
850 int idx = PTR_TO_INT(get_irn_link(irn));
851 affinity_node_t *a = get_affinity_info(co, irn);
853 const arch_register_req_t *req;
856 req = arch_get_register_req(co->aenv, irn, BE_OUT_POS(0));
857 if(arch_register_req_is(req, limited)) {
858 for(i = 0; i < co->cls->n_regs; ++i) {
859 if(!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
860 fprintf(f, "%d %d -1\n", color_map[i], idx);
865 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
866 if(!arch_irn_is(co->aenv, adj, ignore)) {
867 int adj_idx = PTR_TO_INT(get_irn_link(adj));
869 fprintf(f, "%d %d -1\n", idx, adj_idx);
876 co_gs_foreach_neighb(a, n) {
877 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
878 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
880 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
888 typedef struct _appel_clique_walker_t {
890 const copy_opt_t *co;
897 } appel_clique_walker_t;
899 typedef struct _appel_block_info_t {
909 } appel_block_info_t;
911 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
914 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
915 int res = (int) freq;
916 return res == 0 ? 1 : res;
918 ir_loop *loop = get_irn_loop(bl);
920 int d = get_loop_depth(loop);
927 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
929 appel_block_info_t *res = NULL;
932 appel_clique_walker_t *d = (void *) phase;
933 res = phase_alloc(phase, sizeof(res[0]));
934 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
935 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
936 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
937 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
938 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
939 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
945 typedef struct _insn_list_t {
947 struct list_head list;
950 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
952 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
955 for(i = 0; i < bli->n_live_end; ++i)
956 if(bli->live_end[i] == irn)
957 return bli->live_end_nr[i];
962 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
964 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
970 foreach_pset(live, irn)
971 live_arr[n_live++] = irn;
973 /* dump the live after clique */
975 for(j = 0; j < n_live; ++j) {
978 for(k = j + 1; k < n_live; ++k) {
979 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
981 fprintf(env->f, "\n");
985 /* dump the affinities */
986 for(j = 0; !env->dumb && j < n_live; ++j) {
987 ir_node *irn = live_arr[j];
988 int old_nr = PTR_TO_INT(get_irn_link(irn));
990 /* if the node was already live in the last insn dump the affinity */
991 if(old_nr > start_nr) {
992 int weight = appel_aff_weight(env, bl);
993 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
997 /* set the current numbers into the link field. */
998 for(j = 0; j < n_live; ++j) {
999 ir_node *irn = live_arr[j];
1000 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
1003 return curr_nr + n_live;
1006 static void appel_walker(ir_node *bl, void *data)
1008 appel_clique_walker_t *env = data;
1009 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
1010 struct obstack *obst = &env->obst;
1011 void *base = obstack_base(obst);
1012 pset *live = pset_new_ptr_default();
1013 be_lv_t *lv = env->co->cenv->birg->lv;
1017 int start_nr = env->curr_nr;
1018 int curr_nr = start_nr;
1020 be_insn_env_t insn_env;
1025 insn_env.aenv = env->co->aenv;
1026 insn_env.cls = env->co->cls;
1027 insn_env.obst = obst;
1028 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1030 /* Guess how many insns will be in this block. */
1031 sched_foreach(bl, irn)
1035 insns = malloc(n_nodes * sizeof(insns[0]));
1037 /* Put all insns in an array. */
1038 irn = sched_first(bl);
1039 while(!sched_is_end(irn)) {
1041 insn = be_scan_insn(&insn_env, irn);
1042 insns[n_insns++] = insn;
1043 irn = insn->next_insn;
1046 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1047 be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
1049 /* Generate the bad and ugly. */
1050 for(i = n_insns - 1; i >= 0; --i) {
1051 be_insn_t *insn = insns[i];
1053 /* The first live set has to be saved in the block border set. */
1054 if(i == n_insns - 1) {
1056 foreach_pset(live, irn) {
1057 bli->live_end[j] = irn;
1058 bli->live_end_nr[j] = curr_nr + j;
1061 bli->n_live_end = j;
1065 for(j = 0; j < insn->use_start; ++j) {
1066 ir_node *op = insn->ops[j].carrier;
1067 bitset_t *adm = insn->ops[j].regs;
1071 if(!insn->ops[j].has_constraints)
1075 foreach_pset(live, irn) {
1083 assert(nr < pset_count(live));
1085 for(k = 0; k < env->co->cls->n_regs; ++k) {
1086 int mapped_col = env->color_map[k];
1087 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1088 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1093 /* dump the clique and update the stuff. */
1094 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1096 /* remove all defs. */
1097 for(j = 0; j < insn->use_start; ++j)
1098 pset_remove_ptr(live, insn->ops[j].carrier);
1100 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1101 bli->phi[bli->n_phi] = insn->irn;
1102 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1108 for(j = insn->use_start; j < insn->n_ops; ++j)
1109 pset_insert_ptr(live, insn->ops[j].carrier);
1112 /* print the start clique. */
1113 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1116 foreach_pset(live, irn) {
1117 bli->live_in[i] = irn;
1118 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1125 obstack_free(obst, base);
1126 env->curr_nr = curr_nr;
1129 static void appel_inter_block_aff(ir_node *bl, void *data)
1131 appel_clique_walker_t *env = data;
1132 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1136 for(i = 0; i < bli->n_live_in; ++i) {
1137 ir_node *irn = bli->live_in[i];
1139 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1140 ir_node *pred = get_Block_cfgpred_block(bl, j);
1142 int nr = appel_get_live_end_nr(env, pred, irn);
1144 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1148 for(i = 0; i < bli->n_phi; ++i) {
1149 ir_node *irn = bli->phi[i];
1151 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1152 ir_node *pred = get_Block_cfgpred_block(bl, j);
1153 ir_node *op = get_irn_n(irn, j);
1155 int nr = appel_get_live_end_nr(env, pred, op);
1157 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1163 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1167 appel_clique_walker_t env;
1168 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1169 be_lv_t *lv = co->cenv->birg->lv;
1171 be_liveness_recompute(lv);
1172 obstack_init(&env.obst);
1173 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1174 env.curr_nr = co->cls->n_regs;
1178 bitset_copy(adm, co->cenv->ignore_colors);
1179 bitset_flip_all(adm);
1181 /* Make color map. */
1182 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1183 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1184 const arch_register_t *reg = &co->cls->regs[i];
1185 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
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);
1193 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1195 /* make the first k nodes interfere */
1196 for(i = 0; i < n_colors; ++i) {
1198 for(j = i + 1; j < n_colors; ++j)
1199 fprintf(f, "%d %d -1 ", i, j);
1204 env.curr_nr = n_colors;
1205 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1206 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1207 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1208 obstack_free(&env.obst, NULL);
1212 ___ _____ ____ ____ ___ _____ ____ _
1213 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1214 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1215 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1216 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1220 static const char *get_dot_color_name(int col)
1222 static const char *names[] = {
1256 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1259 typedef struct _co_ifg_dump_t {
1260 const copy_opt_t *co;
1264 static void ifg_dump_graph_attr(FILE *f, void *self)
1266 fprintf(f, "overlap=scale");
1269 static int ifg_is_dump_node(void *self, ir_node *irn)
1271 co_ifg_dump_t *cod = self;
1272 return !arch_irn_is(cod->co->aenv, irn, ignore);
1275 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1277 co_ifg_dump_t *env = self;
1278 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1279 const arch_register_req_t *req;
1282 req = arch_get_register_req(env->co->aenv, irn, BE_OUT_POS(0));
1283 limited = arch_register_req_is(req, limited);
1285 if(env->flags & CO_IFG_DUMP_LABELS) {
1286 ir_fprintf(f, "label=\"%+F", irn);
1290 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1291 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1292 req.limited(req.limited_env, bs);
1293 ir_fprintf(f, "\\n%B", bs);
1296 ir_fprintf(f, "\" ");
1298 fprintf(f, "label=\"\" shape=point " );
1301 if(env->flags & CO_IFG_DUMP_SHAPE)
1302 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1304 if(env->flags & CO_IFG_DUMP_COLORS)
1305 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1308 static void ifg_dump_at_end(FILE *file, void *self)
1310 co_ifg_dump_t *env = self;
1313 co_gs_foreach_aff_node(env->co, a) {
1314 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1315 unsigned aidx = get_irn_idx(a->irn);
1318 co_gs_foreach_neighb(a, n) {
1319 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1320 unsigned nidx = get_irn_idx(n->irn);
1323 const char *color = nr == ar ? "blue" : "red";
1324 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1325 if(env->flags & CO_IFG_DUMP_LABELS)
1326 fprintf(file, "label=\"%d\" ", n->costs);
1327 if(env->flags & CO_IFG_DUMP_COLORS)
1328 fprintf(file, "color=%s ", color);
1330 fprintf(file, "style=dotted");
1331 fprintf(file, "];\n");
1338 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1340 ifg_dump_graph_attr,
1349 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1355 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1359 void co_solve_park_moon(copy_opt_t *opt)
1364 static int void_algo(copy_opt_t *co)
1371 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1372 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1373 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1374 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1381 int can_improve_existing;
1384 static co_algo_info_t algos[] = {
1385 { void_algo, "none", 0 },
1386 { co_solve_heuristic, "heur1", 0 },
1387 { co_solve_heuristic_new, "heur2", 0 },
1388 { co_solve_heuristic_java, "heur3", 0 },
1390 { co_solve_ilp2, "ilp", 1 },
1397 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1398 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1399 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1400 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1404 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
1409 ir_snprintf(buf, sizeof(buf), "%s%F_%s%s", prefix, env->irg, env->cls->name, suffix);
1410 result = fopen(buf, "wt");
1411 if(result == NULL) {
1412 panic("Couldn't open '%s' for writing.", buf);
1418 void co_driver(be_chordal_env_t *cenv)
1420 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1421 co_complete_stats_t before, after;
1423 co_algo_t *algo_func;
1424 int was_optimal = 0;
1426 if(algo < 0 || algo >= CO_ALGO_LAST)
1429 co = new_copy_opt(cenv, cost_func);
1430 co_build_ou_structure(co);
1431 co_build_graph_structure(co);
1433 co_complete_stats(co, &before);
1435 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1436 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1437 be_stat_ev_ull("co_max_costs", before.max_costs);
1438 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1439 be_stat_ev_ull("co_aff_int", before.aff_int);
1441 be_stat_ev_ull("co_init_costs", before.costs);
1442 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1444 /* Dump the interference graph in Appel's format. */
1445 if(dump_flags & DUMP_APPEL) {
1446 FILE *f = my_open(cenv, "", ".apl");
1447 co_dump_appel_graph(co, f);
1451 if(dump_flags & DUMP_BEFORE) {
1452 FILE *f = my_open(cenv, "", "-before.dot");
1453 co_dump_ifg_dot(co, f, style_flags);
1457 /* if the algo can improve results, provide an initial solution with heur3 */
1458 if(improve && algos[algo].can_improve_existing) {
1459 co_complete_stats_t stats;
1461 /* produce a heuristical solution */
1462 co_solve_heuristic_java(co);
1464 /* do the stats and provide the current costs */
1465 co_complete_stats(co, &stats);
1466 be_stat_ev_ull("co_prepare_costs", stats.costs);
1470 /* start the JVM here so that it does not tamper the timing. */
1471 if(algo == CO_ALGO_HEUR3)
1472 be_java_coal_start_jvm();
1475 algo_func = algos[algo].algo;
1477 lc_timer_reset_and_start(timer);
1479 was_optimal = algo_func(co);
1481 lc_timer_stop(timer);
1482 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1484 be_stat_ev_ull("co_optimal", was_optimal);
1486 if(dump_flags & DUMP_AFTER) {
1487 FILE *f = my_open(cenv, "", "-after.dot");
1488 co_dump_ifg_dot(co, f, style_flags);
1492 co_complete_stats(co, &after);
1495 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1496 ulong64 evitable = after.costs - after.inevit_costs;
1498 ir_printf("%30F ", cenv->irg);
1499 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1501 if(optimizable_costs > 0)
1502 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1504 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1507 be_stat_ev_ull("co_after_costs", after.costs);
1508 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1510 co_free_graph_structure(co);
1511 co_free_ou_structure(co);