4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
26 #include "iredges_t.h"
29 #include "irphase_t.h"
30 #include "irprintf_t.h"
37 #include "becopyopt_t.h"
38 #include "becopystat.h"
41 #include "besched_t.h"
42 #include "benodesets.h"
43 #include "bejavacoal.h"
44 #include "bestatevent.h"
48 #include <libcore/lc_timing.h>
49 #include <libcore/lc_opts.h>
50 #endif /* WITH_LIBCORE */
55 #define DUMP_ALL 2 * DUMP_APPEL - 1
57 #define COST_FUNC_FREQ 1
58 #define COST_FUNC_LOOP 2
59 #define COST_FUNC_ALL_ONE 3
61 static unsigned dump_flags = 0;
62 static unsigned style_flags = 0;
63 static unsigned do_stats = 0;
64 static cost_fct_t cost_func = co_get_costs_exec_freq;
65 static unsigned algo = CO_ALGO_HEUR2;
66 static int improve = 1;
69 static const lc_opt_enum_mask_items_t dump_items[] = {
70 { "before", DUMP_BEFORE },
71 { "after", DUMP_AFTER },
72 { "appel", DUMP_APPEL },
77 static const lc_opt_enum_mask_items_t style_items[] = {
78 { "color", CO_IFG_DUMP_COLORS },
79 { "labels", CO_IFG_DUMP_LABELS },
80 { "constr", CO_IFG_DUMP_CONSTR },
81 { "shape", CO_IFG_DUMP_SHAPE },
82 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
86 static const lc_opt_enum_mask_items_t algo_items[] = {
87 { "none", CO_ALGO_NONE },
88 { "heur", CO_ALGO_HEUR },
89 { "heur2", CO_ALGO_HEUR2 },
91 { "heur3", CO_ALGO_HEUR3 },
94 { "ilp", CO_ALGO_ILP },
99 typedef int (*opt_funcptr)(void);
101 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
102 { "freq", (opt_funcptr) co_get_costs_exec_freq },
103 { "loop", (opt_funcptr) co_get_costs_loop_depth },
104 { "one", (opt_funcptr) co_get_costs_all_one },
108 static lc_opt_enum_mask_var_t dump_var = {
109 &dump_flags, dump_items
112 static lc_opt_enum_mask_var_t style_var = {
113 &style_flags, style_items
116 static lc_opt_enum_mask_var_t algo_var = {
120 static lc_opt_enum_func_ptr_var_t cost_func_var = {
121 (opt_funcptr*) &cost_func, cost_func_items
124 static const lc_opt_table_entry_t options[] = {
125 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var),
126 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
127 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
128 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
129 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
130 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
134 /* Insert additional options registration functions here. */
135 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
136 extern void be_co2_register_options(lc_opt_entry_t *grp);
137 extern void be_co3_register_options(lc_opt_entry_t *grp);
139 void co_register_options(lc_opt_entry_t *grp)
141 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
142 lc_opt_add_table(co_grp, options);
144 be_co2_register_options(co_grp);
145 be_co3_register_options(co_grp);
147 be_co_ilp_register_options(co_grp);
153 #undef QUICK_AND_DIRTY_HACK
155 /******************************************************************************
158 | | __ ___ _ __ ___ _ __ __ _| |
159 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
160 | |__| | __/ | | | __/ | | (_| | |
161 \_____|\___|_| |_|\___|_| \__,_|_|
163 ******************************************************************************/
165 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
168 void be_copy_opt_init(void) {
171 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
173 const char *s1, *s2, *s3;
177 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
179 co = xcalloc(1, sizeof(*co));
180 co->cenv = chordal_env;
181 co->aenv = chordal_env->birg->main_env->arch_env;
182 co->irg = chordal_env->irg;
183 co->cls = chordal_env->cls;
184 co->get_costs = get_costs;
186 s1 = get_irp_prog_name();
187 s2 = get_entity_name(get_irg_entity(co->irg));
188 s3 = chordal_env->cls->name;
189 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
190 co->name = xmalloc(len);
191 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
196 void free_copy_opt(copy_opt_t *co) {
201 int co_is_optimizable_root(const copy_opt_t *co, ir_node *irn) {
202 arch_register_req_t req;
203 const arch_register_t *reg;
205 if (arch_irn_is(co->aenv, irn, ignore))
208 reg = arch_get_irn_register(co->aenv, irn);
209 if (arch_register_type_is(reg, ignore))
212 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &req))
218 int co_is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
219 const ir_edge_t *edge;
220 const arch_register_t *reg;
222 assert(0 && "Is buggy and obsolete. Do not use");
224 if (arch_irn_is(co->aenv, irn, ignore))
227 reg = arch_get_irn_register(co->aenv, irn);
228 if (arch_register_type_is(reg, ignore))
231 foreach_out_edge(irn, edge) {
232 ir_node *n = edge->src;
234 if (!nodes_interfere(co->cenv, irn, n) || irn == n) {
235 arch_register_req_t req;
236 arch_get_register_req(co->aenv, &req, n, -1);
239 is_Perm(co->aenv, n) ||
240 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
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;
384 arch_register_req_t req;
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);
456 /* Src == Tgt of a 2-addr-code instruction */
457 if (is_2addr_code(co->aenv, irn, &req)) {
458 ir_node *other = req.other_same;
459 if (!nodes_interfere(co->cenv, irn, other)) {
460 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
461 unit->costs = xmalloc(2 * sizeof(*unit->costs));
462 unit->node_count = 2;
463 unit->nodes[0] = irn;
464 unit->nodes[1] = other;
465 unit->costs[1] = co->get_costs(co, irn, other, -1);
468 assert(0 && "This is not an optimizable node!");
470 /* Insert the new unit at a position according to its costs */
471 if (unit->node_count > 1) {
473 struct list_head *tmp;
475 /* Determine the maximum costs this unit can cause: all_nodes_cost */
476 for(i=1; i<unit->node_count; ++i) {
477 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
478 unit->all_nodes_costs += unit->costs[i];
481 /* Determine the minimal costs this unit will cause: min_nodes_costs */
482 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
483 /* Insert the new ou according to its sort_key */
485 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
487 list_add(&unit->units, tmp);
493 #ifdef QUICK_AND_DIRTY_HACK
495 static int compare_ous(const void *k1, const void *k2) {
496 const unit_t *u1 = *((const unit_t **) k1);
497 const unit_t *u2 = *((const unit_t **) k2);
498 int i, o, u1_has_constr, u2_has_constr;
499 arch_register_req_t req;
500 const arch_env_t *aenv = u1->co->aenv;
502 /* Units with constraints come first */
504 for (i=0; i<u1->node_count; ++i) {
505 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
506 if (arch_register_req_is(&req, limited)) {
513 for (i=0; i<u2->node_count; ++i) {
514 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
515 if (arch_register_req_is(&req, limited)) {
521 if (u1_has_constr != u2_has_constr)
522 return u2_has_constr - u1_has_constr;
524 /* Now check, whether the two units are connected */
526 for (i=0; i<u1->node_count; ++i)
527 for (o=0; o<u2->node_count; ++o)
528 if (u1->nodes[i] == u2->nodes[o])
532 /* After all, the sort key decides. Greater keys come first. */
533 return u2->sort_key - u1->sort_key;
538 * Sort the ou's according to constraints and their sort_key
540 static void co_sort_units(copy_opt_t *co) {
541 int i, count = 0, costs;
544 /* get the number of ous, remove them form the list and fill the array */
545 list_for_each_entry(unit_t, ou, &co->units, units)
547 ous = alloca(count * sizeof(*ous));
549 costs = co_get_max_copy_costs(co);
552 list_for_each_entry(unit_t, ou, &co->units, units)
555 INIT_LIST_HEAD(&co->units);
557 assert(count == i && list_empty(&co->units));
559 for (i=0; i<count; ++i)
560 ir_printf("%+F\n", ous[i]->nodes[0]);
562 qsort(ous, count, sizeof(*ous), compare_ous);
565 for (i=0; i<count; ++i)
566 ir_printf("%+F\n", ous[i]->nodes[0]);
568 /* reinsert into list in correct order */
569 for (i=0; i<count; ++i)
570 list_add_tail(&ous[i]->units, &co->units);
572 assert(costs == co_get_max_copy_costs(co));
576 void co_build_ou_structure(copy_opt_t *co) {
577 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
578 INIT_LIST_HEAD(&co->units);
579 irg_walk_graph(co->irg, co_collect_units, NULL, co);
580 #ifdef QUICK_AND_DIRTY_HACK
585 void co_free_ou_structure(copy_opt_t *co) {
588 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
593 co->units.next = NULL;
596 /* co_solve_heuristic() is implemented in becopyheur.c */
598 int co_get_max_copy_costs(const copy_opt_t *co) {
604 list_for_each_entry(unit_t, curr, &co->units, units) {
605 res += curr->inevitable_costs;
606 for (i=1; i<curr->node_count; ++i)
607 res += curr->costs[i];
612 int co_get_inevit_copy_costs(const copy_opt_t *co) {
618 list_for_each_entry(unit_t, curr, &co->units, units)
619 res += curr->inevitable_costs;
623 int co_get_copy_costs(const copy_opt_t *co) {
629 list_for_each_entry(unit_t, curr, &co->units, units) {
630 int root_col = get_irn_col(co, curr->nodes[0]);
631 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
632 res += curr->inevitable_costs;
633 for (i=1; i<curr->node_count; ++i) {
634 int arg_col = get_irn_col(co, curr->nodes[i]);
635 if (root_col != arg_col) {
636 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
637 res += curr->costs[i];
644 int co_get_lower_bound(const copy_opt_t *co) {
650 list_for_each_entry(unit_t, curr, &co->units, units)
651 res += curr->inevitable_costs + curr->min_nodes_costs;
655 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
657 bitset_t *seen = bitset_irg_malloc(co->irg);
660 memset(stat, 0, sizeof(stat[0]));
662 /* count affinity edges. */
663 co_gs_foreach_aff_node(co, an) {
665 stat->aff_nodes += 1;
666 bitset_add_irn(seen, an->irn);
667 co_gs_foreach_neighb(an, neigh) {
668 if(!bitset_contains_irn(seen, neigh->irn)) {
669 stat->aff_edges += 1;
670 stat->max_costs += neigh->costs;
672 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
673 stat->costs += neigh->costs;
674 stat->unsatisfied_edges += 1;
677 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
679 stat->inevit_costs += neigh->costs;
689 /******************************************************************************
691 / ____| | | / ____| |
692 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
693 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
694 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
695 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
698 ******************************************************************************/
700 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
701 const affinity_node_t *n1 = k1;
702 const affinity_node_t *n2 = k2;
704 return (n1->irn != n2->irn);
707 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
708 affinity_node_t new_node, *node;
709 neighb_t new_nbr, *nbr;
714 new_node.neighbours = NULL;
715 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
718 for (nbr = node->neighbours; nbr; nbr = nbr->next)
719 if (nbr->irn == n2) {
724 /* if we did not find n2 in n1's neighbourhood insert it */
726 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
727 nbr = obstack_finish(&co->obst);
730 nbr->next = node->neighbours;
731 node->neighbours = nbr;
735 /* now nbr points to n1's neighbour-entry of n2 */
739 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
740 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
741 add_edge(co, n1, n2, costs);
742 add_edge(co, n2, n1, costs);
746 static void build_graph_walker(ir_node *irn, void *env) {
747 copy_opt_t *co = env;
749 arch_register_req_t req;
750 const arch_register_t *reg;
752 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
755 reg = arch_get_irn_register(co->aenv, irn);
756 if (arch_register_type_is(reg, ignore))
761 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
762 ir_node *arg = get_irn_n(irn, pos);
763 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
767 else if (is_Perm_Proj(co->aenv, irn)) {
768 ir_node *arg = get_Perm_src(irn);
769 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
773 else if (is_2addr_code(co->aenv, irn, &req))
774 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
777 void co_build_graph_structure(copy_opt_t *co) {
778 obstack_init(&co->obst);
779 co->nodes = new_set(compare_affinity_node_t, 32);
781 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
784 void co_free_graph_structure(copy_opt_t *co) {
788 obstack_free(&co->obst, NULL);
792 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
794 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
795 affinity_node_t new_node, *n;
800 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
802 return (n->degree > 0);
807 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
809 be_ifg_t *ifg = co->cenv->ifg;
810 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
811 bitset_t *adm = bitset_alloca(co->cls->n_regs);
818 for(i = 0; i < co->cls->n_regs; ++i) {
819 const arch_register_t *reg = &co->cls->regs[i];
820 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
824 * n contains the first node number.
825 * the values below n are the pre-colored register nodes
828 it = be_ifg_nodes_iter_alloca(ifg);
829 nit = be_ifg_neighbours_iter_alloca(ifg);
832 be_ifg_foreach_node(ifg, it, irn) {
833 if(!arch_irn_is(co->aenv, irn, ignore))
834 set_irn_link(irn, INT_TO_PTR(n++));
837 fprintf(f, "%d %d\n", n, n_regs);
839 be_ifg_foreach_node(ifg, it, irn) {
840 if(!arch_irn_is(co->aenv, irn, ignore)) {
841 int idx = PTR_TO_INT(get_irn_link(irn));
842 affinity_node_t *a = get_affinity_info(co, irn);
844 arch_register_req_t req;
847 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
848 if(arch_register_req_is(&req, limited)) {
849 bitset_clear_all(adm);
850 req.limited(req.limited_env, adm);
851 for(i = 0; i < co->cls->n_regs; ++i)
852 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
853 fprintf(f, "%d %d -1\n", color_map[i], idx);
858 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
859 if(!arch_irn_is(co->aenv, adj, ignore)) {
860 int adj_idx = PTR_TO_INT(get_irn_link(adj));
862 fprintf(f, "%d %d -1\n", idx, adj_idx);
869 co_gs_foreach_neighb(a, n) {
870 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
871 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
873 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
881 typedef struct _appel_clique_walker_t {
883 const copy_opt_t *co;
890 } appel_clique_walker_t;
892 typedef struct _appel_block_info_t {
902 } appel_block_info_t;
904 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
907 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
908 int res = (int) freq;
909 return res == 0 ? 1 : res;
911 ir_loop *loop = get_irn_loop(bl);
913 int d = get_loop_depth(loop);
920 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
922 appel_block_info_t *res = NULL;
925 appel_clique_walker_t *d = (void *) phase;
926 res = phase_alloc(phase, sizeof(res[0]));
927 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
928 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
929 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
930 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
931 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
932 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
938 typedef struct _insn_list_t {
940 struct list_head list;
943 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
945 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
948 for(i = 0; i < bli->n_live_end; ++i)
949 if(bli->live_end[i] == irn)
950 return bli->live_end_nr[i];
955 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
957 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
963 foreach_pset(live, irn)
964 live_arr[n_live++] = irn;
966 /* dump the live after clique */
968 for(j = 0; j < n_live; ++j) {
971 for(k = j + 1; k < n_live; ++k) {
972 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
974 fprintf(env->f, "\n");
978 /* dump the affinities */
979 for(j = 0; !env->dumb && j < n_live; ++j) {
980 ir_node *irn = live_arr[j];
981 int old_nr = PTR_TO_INT(get_irn_link(irn));
983 /* if the node was already live in the last insn dump the affinity */
984 if(old_nr > start_nr) {
985 int weight = appel_aff_weight(env, bl);
986 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
990 /* set the current numbers into the link field. */
991 for(j = 0; j < n_live; ++j) {
992 ir_node *irn = live_arr[j];
993 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
996 return curr_nr + n_live;
999 static void appel_walker(ir_node *bl, void *data)
1001 appel_clique_walker_t *env = data;
1002 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
1003 struct obstack *obst = &env->obst;
1004 void *base = obstack_base(obst);
1005 pset *live = pset_new_ptr_default();
1006 be_lv_t *lv = env->co->cenv->birg->lv;
1010 int start_nr = env->curr_nr;
1011 int curr_nr = start_nr;
1013 be_insn_env_t insn_env;
1018 insn_env.aenv = env->co->aenv;
1019 insn_env.cls = env->co->cls;
1020 insn_env.obst = obst;
1021 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1023 /* Guess how many insns will be in this block. */
1024 sched_foreach(bl, irn)
1028 insns = malloc(n_nodes * sizeof(insns[0]));
1030 /* Put all insns in an array. */
1031 irn = sched_first(bl);
1032 while(!sched_is_end(irn)) {
1034 insn = be_scan_insn(&insn_env, irn);
1035 insns[n_insns++] = insn;
1036 irn = insn->next_insn;
1039 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1040 be_liveness_end_of_block(lv, env->co->aenv, env->co->cls, bl, live);
1042 /* Generate the bad and ugly. */
1043 for(i = n_insns - 1; i >= 0; --i) {
1044 be_insn_t *insn = insns[i];
1046 /* The first live set has to be saved in the block border set. */
1047 if(i == n_insns - 1) {
1049 foreach_pset(live, irn) {
1050 bli->live_end[j] = irn;
1051 bli->live_end_nr[j] = curr_nr + j;
1054 bli->n_live_end = j;
1058 for(j = 0; j < insn->use_start; ++j) {
1059 ir_node *op = insn->ops[j].carrier;
1060 bitset_t *adm = insn->ops[j].regs;
1064 if(!insn->ops[j].has_constraints)
1068 foreach_pset(live, irn) {
1076 assert(nr < pset_count(live));
1078 for(k = 0; k < env->co->cls->n_regs; ++k) {
1079 int mapped_col = env->color_map[k];
1080 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1081 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1086 /* dump the clique and update the stuff. */
1087 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1089 /* remove all defs. */
1090 for(j = 0; j < insn->use_start; ++j)
1091 pset_remove_ptr(live, insn->ops[j].carrier);
1093 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1094 bli->phi[bli->n_phi] = insn->irn;
1095 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1101 for(j = insn->use_start; j < insn->n_ops; ++j)
1102 pset_insert_ptr(live, insn->ops[j].carrier);
1105 /* print the start clique. */
1106 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1109 foreach_pset(live, irn) {
1110 bli->live_in[i] = irn;
1111 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1118 obstack_free(obst, base);
1119 env->curr_nr = curr_nr;
1122 static void appel_inter_block_aff(ir_node *bl, void *data)
1124 appel_clique_walker_t *env = data;
1125 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1129 for(i = 0; i < bli->n_live_in; ++i) {
1130 ir_node *irn = bli->live_in[i];
1132 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1133 ir_node *pred = get_Block_cfgpred_block(bl, j);
1135 int nr = appel_get_live_end_nr(env, pred, irn);
1137 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1141 for(i = 0; i < bli->n_phi; ++i) {
1142 ir_node *irn = bli->phi[i];
1144 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1145 ir_node *pred = get_Block_cfgpred_block(bl, j);
1146 ir_node *op = get_irn_n(irn, j);
1148 int nr = appel_get_live_end_nr(env, pred, op);
1150 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1156 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1160 appel_clique_walker_t env;
1161 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1162 be_lv_t *lv = co->cenv->birg->lv;
1164 be_liveness_recompute(lv);
1165 obstack_init(&env.obst);
1166 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1167 env.curr_nr = co->cls->n_regs;
1171 bitset_copy(adm, co->cenv->ignore_colors);
1172 bitset_flip_all(adm);
1174 /* Make color map. */
1175 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1176 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1177 const arch_register_t *reg = &co->cls->regs[i];
1178 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1182 env.curr_nr = n_colors;
1183 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1184 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1186 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1188 /* make the first k nodes interfere */
1189 for(i = 0; i < n_colors; ++i) {
1191 for(j = i + 1; j < n_colors; ++j)
1192 fprintf(f, "%d %d -1 ", i, j);
1197 env.curr_nr = n_colors;
1198 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1199 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1200 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1201 obstack_free(&env.obst, NULL);
1205 ___ _____ ____ ____ ___ _____ ____ _
1206 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1207 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1208 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1209 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1213 static const char *get_dot_color_name(int col)
1215 static const char *names[] = {
1249 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1252 typedef struct _co_ifg_dump_t {
1253 const copy_opt_t *co;
1257 static void ifg_dump_graph_attr(FILE *f, void *self)
1259 fprintf(f, "overlap=scale");
1262 static int ifg_is_dump_node(void *self, ir_node *irn)
1264 co_ifg_dump_t *cod = self;
1265 return !arch_irn_is(cod->co->aenv, irn, ignore);
1268 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1270 co_ifg_dump_t *env = self;
1271 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1272 arch_register_req_t req;
1275 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1276 limited = arch_register_req_is(&req, limited);
1278 if(env->flags & CO_IFG_DUMP_LABELS) {
1279 ir_fprintf(f, "label=\"%+F", irn);
1281 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1282 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1283 req.limited(req.limited_env, bs);
1284 ir_fprintf(f, "\\n%B", bs);
1286 ir_fprintf(f, "\" ");
1290 fprintf(f, "label=\"\" shape=point " );
1292 if(env->flags & CO_IFG_DUMP_SHAPE)
1293 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1295 if(env->flags & CO_IFG_DUMP_COLORS)
1296 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1299 static void ifg_dump_at_end(FILE *file, void *self)
1301 co_ifg_dump_t *env = self;
1304 co_gs_foreach_aff_node(env->co, a) {
1305 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1306 unsigned aidx = get_irn_idx(a->irn);
1309 co_gs_foreach_neighb(a, n) {
1310 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1311 unsigned nidx = get_irn_idx(n->irn);
1314 const char *color = nr == ar ? "blue" : "red";
1315 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1316 if(env->flags & CO_IFG_DUMP_LABELS)
1317 fprintf(file, "label=\"%d\" ", n->costs);
1318 if(env->flags & CO_IFG_DUMP_COLORS)
1319 fprintf(file, "color=%s ", color);
1321 fprintf(file, "style=dotted");
1322 fprintf(file, "];\n");
1329 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1331 ifg_dump_graph_attr,
1340 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1346 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1350 void co_solve_park_moon(copy_opt_t *opt)
1355 static int void_algo(copy_opt_t *co)
1362 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1363 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1364 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1365 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1372 int can_improve_existing;
1375 static co_algo_info_t algos[] = {
1376 { void_algo, "none", 0 },
1377 { co_solve_heuristic, "heur1", 0 },
1378 { co_solve_heuristic_new, "heur2", 0 },
1379 { co_solve_heuristic_java, "heur3", 0 },
1381 { co_solve_ilp2, "ilp", 1 },
1388 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1389 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1390 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1391 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1395 void co_driver(be_chordal_env_t *cenv)
1398 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1400 co_complete_stats_t before, after;
1402 co_algo_t *algo_func;
1403 int was_optimal = 0;
1405 if(algo < 0 || algo >= CO_ALGO_LAST)
1408 co = new_copy_opt(cenv, cost_func);
1409 co_build_ou_structure(co);
1410 co_build_graph_structure(co);
1412 co_complete_stats(co, &before);
1414 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1415 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1416 be_stat_ev_ull("co_max_costs", before.max_costs);
1417 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1418 be_stat_ev_ull("co_aff_int", before.aff_int);
1420 be_stat_ev_ull("co_init_costs", before.costs);
1421 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1423 /* Dump the interference graph in Appel's format. */
1424 if(dump_flags & DUMP_APPEL) {
1425 FILE *f = be_chordal_open(cenv, "", ".apl");
1426 co_dump_appel_graph(co, f);
1430 if(dump_flags & DUMP_BEFORE) {
1431 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1432 co_dump_ifg_dot(co, f, style_flags);
1436 /* if the algo can improve results, provide an initial solution with heur3 */
1437 if(improve && algos[algo].can_improve_existing) {
1438 co_complete_stats_t stats;
1440 /* produce a heuristical solution */
1441 co_solve_heuristic_java(co);
1443 /* do the stats and provide the current costs */
1444 co_complete_stats(co, &stats);
1445 be_stat_ev_ull("co_prepare_costs", stats.costs);
1449 /* start the JVM here so that it does not tamper the timing. */
1450 if(algo == CO_ALGO_HEUR3)
1451 be_java_coal_start_jvm();
1454 algo_func = algos[algo].algo;
1457 lc_timer_reset_and_start(timer);
1460 was_optimal = algo_func(co);
1463 lc_timer_stop(timer);
1464 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1467 be_stat_ev_ull("co_optimal", was_optimal);
1469 if(dump_flags & DUMP_AFTER) {
1470 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1471 co_dump_ifg_dot(co, f, style_flags);
1475 co_complete_stats(co, &after);
1478 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1479 ulong64 evitable = after.costs - after.inevit_costs;
1481 ir_printf("%30F ", cenv->irg);
1482 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1484 if(optimizable_costs > 0)
1485 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1487 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1490 be_stat_ev_ull("co_after_costs", after.costs);
1491 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1493 co_free_graph_structure(co);
1494 co_free_ou_structure(co);