4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
25 #include "iredges_t.h"
28 #include "irphase_t.h"
29 #include "irprintf_t.h"
36 #include "becopyopt_t.h"
37 #include "becopystat.h"
40 #include "besched_t.h"
41 #include "benodesets.h"
42 #include "bejavacoal.h"
43 #include "bestatevent.h"
46 #include <libcore/lc_timing.h>
47 #include <libcore/lc_opts.h>
48 #endif /* WITH_LIBCORE */
53 #define DUMP_ALL 2 * DUMP_APPEL - 1
55 #define COST_FUNC_FREQ 1
56 #define COST_FUNC_LOOP 2
57 #define COST_FUNC_ALL_ONE 3
59 static unsigned dump_flags = 0;
60 static unsigned style_flags = 0;
61 static unsigned do_stats = 0;
62 static cost_fct_t cost_func = co_get_costs_exec_freq;
63 static unsigned algo = CO_ALGO_HEUR2;
64 static int improve = 1;
67 static const lc_opt_enum_mask_items_t dump_items[] = {
68 { "before", DUMP_BEFORE },
69 { "after", DUMP_AFTER },
70 { "appel", DUMP_APPEL },
75 static const lc_opt_enum_mask_items_t style_items[] = {
76 { "color", CO_IFG_DUMP_COLORS },
77 { "labels", CO_IFG_DUMP_LABELS },
78 { "constr", CO_IFG_DUMP_CONSTR },
79 { "shape", CO_IFG_DUMP_SHAPE },
80 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
84 static const lc_opt_enum_mask_items_t algo_items[] = {
85 { "none", CO_ALGO_NONE },
86 { "heur", CO_ALGO_HEUR },
87 { "heur2", CO_ALGO_HEUR2 },
89 { "heur3", CO_ALGO_HEUR3 },
92 { "ilp", CO_ALGO_ILP },
97 typedef int (*opt_funcptr)(void);
99 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
100 { "freq", (opt_funcptr) co_get_costs_exec_freq },
101 { "loop", (opt_funcptr) co_get_costs_loop_depth },
102 { "one", (opt_funcptr) co_get_costs_all_one },
106 static lc_opt_enum_mask_var_t dump_var = {
107 &dump_flags, dump_items
110 static lc_opt_enum_mask_var_t style_var = {
111 &style_flags, style_items
114 static lc_opt_enum_mask_var_t algo_var = {
118 static lc_opt_enum_func_ptr_var_t cost_func_var = {
119 (opt_funcptr*) &cost_func, cost_func_items
122 static const lc_opt_table_entry_t options[] = {
123 LC_OPT_ENT_ENUM_INT ("algo", "select copy optimization algo", &algo_var),
124 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
125 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
126 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
127 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
128 LC_OPT_ENT_BOOL ("improve", "run heur3 before if algo can exploit start solutions", &improve),
132 /* Insert additional options registration functions here. */
133 extern void be_co_ilp_register_options(lc_opt_entry_t *grp);
134 extern void be_co2_register_options(lc_opt_entry_t *grp);
135 extern void be_co3_register_options(lc_opt_entry_t *grp);
137 void co_register_options(lc_opt_entry_t *grp)
139 lc_opt_entry_t *co_grp = lc_opt_get_grp(grp, "co");
140 lc_opt_add_table(co_grp, options);
142 be_co2_register_options(co_grp);
143 be_co3_register_options(co_grp);
145 be_co_ilp_register_options(co_grp);
151 #undef QUICK_AND_DIRTY_HACK
153 /******************************************************************************
156 | | __ ___ _ __ ___ _ __ __ _| |
157 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
158 | |__| | __/ | | | __/ | | (_| | |
159 \_____|\___|_| |_|\___|_| \__,_|_|
161 ******************************************************************************/
163 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
166 void be_copy_opt_init(void) {
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 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 if (is_Reg_Phi(irn) || is_Perm_Proj(co->aenv, irn) || is_2addr_code(co->aenv, irn, &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 arch_register_req_t req;
234 arch_get_register_req(co->aenv, &req, n, -1);
237 is_Perm(co->aenv, n) ||
238 (arch_register_req_is(&req, should_be_same) && req.other_same == irn)
247 int co_get_costs_loop_depth(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
250 ir_node *root_block = get_nodes_block(root);
253 /* for phis the copies are placed in the corresponding pred-block */
254 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
256 /* a perm places the copy in the same block as it resides */
257 loop = get_irn_loop(root_block);
260 int d = get_loop_depth(loop);
266 int co_get_costs_exec_freq(const copy_opt_t *co, ir_node *root, ir_node* arg, int pos) {
268 ir_node *root_bl = get_nodes_block(root);
269 ir_node *copy_bl = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
270 res = get_block_execfreq_ulong(co->cenv->exec_freq, copy_bl);
272 /* don't allow values smaller than one. */
273 return res < 1 ? 1 : res;
277 int co_get_costs_all_one(const copy_opt_t *co, ir_node *root, ir_node *arg, int pos) {
281 /******************************************************************************
282 ____ _ _ _ _ _ _____ _
283 / __ \ | | | | | | (_) | / ____| |
284 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
285 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
286 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
287 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
290 ******************************************************************************/
293 * Determines a maximum weighted independent set with respect to
294 * the interference and conflict edges of all nodes in a qnode.
296 static int ou_max_ind_set_costs(unit_t *ou) {
297 be_chordal_env_t *chordal_env = ou->co->cenv;
298 ir_node **safe, **unsafe;
299 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
301 int max, pos, curr_weight, best_weight = 0;
303 /* assign the nodes into two groups.
304 * safe: node has no interference, hence it is in every max stable set.
305 * unsafe: node has an interference
307 safe = alloca((ou->node_count-1) * sizeof(*safe));
310 unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
311 unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
313 for(i=1; i<ou->node_count; ++i) {
315 for(o=1; o<ou->node_count; ++o) {
318 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
319 unsafe_costs[unsafe_count] = ou->costs[i];
320 unsafe[unsafe_count] = ou->nodes[i];
327 safe_costs += ou->costs[i];
328 safe[safe_count++] = ou->nodes[i];
333 /* now compute the best set out of the unsafe nodes*/
334 if (unsafe_count > MIS_HEUR_TRIGGER) {
335 bitset_t *best = bitset_alloca(unsafe_count);
336 /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
337 for (i=0; i<unsafe_count; ++i) {
339 /* check if it is a stable set */
340 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
341 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
342 bitset_clear(best, i); /* clear the bit and try next one */
346 /* compute the weight */
347 bitset_foreach(best, pos)
348 best_weight += unsafe_costs[pos];
350 /* Exact Algorithm: Brute force */
351 curr = bitset_alloca(unsafe_count);
352 bitset_set_all(curr);
353 while ((max = bitset_popcnt(curr)) != 0) {
354 /* check if curr is a stable set */
355 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
356 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) */
357 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
360 /* if we arrive here, we have a stable set */
361 /* compute the weigth of the stable set*/
363 bitset_foreach(curr, pos)
364 curr_weight += unsafe_costs[pos];
367 if (curr_weight > best_weight) {
368 best_weight = curr_weight;
376 return safe_costs+best_weight;
379 static void co_collect_units(ir_node *irn, void *env) {
380 copy_opt_t *co = env;
382 arch_register_req_t req;
384 if (!is_curr_reg_class(co, irn))
386 if (!co_is_optimizable_root(co, irn))
389 /* Init a new unit */
390 unit = xcalloc(1, sizeof(*unit));
392 unit->node_count = 1;
393 INIT_LIST_HEAD(&unit->queue);
395 /* Phi with some/all of its arguments */
396 if (is_Reg_Phi(irn)) {
400 arity = get_irn_arity(irn);
401 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
402 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
403 unit->nodes[0] = irn;
406 for (i=0; i<arity; ++i) {
408 ir_node *arg = get_irn_n(irn, i);
410 assert(is_curr_reg_class(co, arg) && "Argument not in same register class.");
413 if (nodes_interfere(co->cenv, irn, arg)) {
414 unit->inevitable_costs += co->get_costs(co, irn, arg, i);
418 /* Else insert the argument of the phi to the members of this ou */
419 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
421 /* Check if arg has occurred at a prior position in the arg/list */
423 for (o=0; o<unit->node_count; ++o)
424 if (unit->nodes[o] == arg) {
429 if (!arg_pos) { /* a new argument */
430 /* insert node, set costs */
431 unit->nodes[unit->node_count] = arg;
432 unit->costs[unit->node_count] = co->get_costs(co, irn, arg, i);
434 } else { /* arg has occured before in same phi */
435 /* increase costs for existing arg */
436 unit->costs[arg_pos] += co->get_costs(co, irn, arg, i);
439 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
440 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
443 /* Proj of a perm with corresponding arg */
444 if (is_Perm_Proj(co->aenv, irn)) {
445 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
446 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
447 unit->costs = xmalloc(2 * sizeof(*unit->costs));
448 unit->node_count = 2;
449 unit->nodes[0] = irn;
450 unit->nodes[1] = get_Perm_src(irn);
451 unit->costs[1] = co->get_costs(co, irn, unit->nodes[1], -1);
454 /* Src == Tgt of a 2-addr-code instruction */
455 if (is_2addr_code(co->aenv, irn, &req)) {
456 ir_node *other = req.other_same;
457 if (!nodes_interfere(co->cenv, irn, other)) {
458 unit->nodes = xmalloc(2 * sizeof(*unit->nodes));
459 unit->costs = xmalloc(2 * sizeof(*unit->costs));
460 unit->node_count = 2;
461 unit->nodes[0] = irn;
462 unit->nodes[1] = other;
463 unit->costs[1] = co->get_costs(co, irn, other, -1);
466 assert(0 && "This is not an optimizable node!");
468 /* Insert the new unit at a position according to its costs */
469 if (unit->node_count > 1) {
471 struct list_head *tmp;
473 /* Determine the maximum costs this unit can cause: all_nodes_cost */
474 for(i=1; i<unit->node_count; ++i) {
475 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
476 unit->all_nodes_costs += unit->costs[i];
479 /* Determine the minimal costs this unit will cause: min_nodes_costs */
480 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
481 /* Insert the new ou according to its sort_key */
483 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
485 list_add(&unit->units, tmp);
491 #ifdef QUICK_AND_DIRTY_HACK
493 static int compare_ous(const void *k1, const void *k2) {
494 const unit_t *u1 = *((const unit_t **) k1);
495 const unit_t *u2 = *((const unit_t **) k2);
496 int i, o, u1_has_constr, u2_has_constr;
497 arch_register_req_t req;
498 const arch_env_t *aenv = u1->co->aenv;
500 /* Units with constraints come first */
502 for (i=0; i<u1->node_count; ++i) {
503 arch_get_register_req(aenv, &req, u1->nodes[i], -1);
504 if (arch_register_req_is(&req, limited)) {
511 for (i=0; i<u2->node_count; ++i) {
512 arch_get_register_req(aenv, &req, u2->nodes[i], -1);
513 if (arch_register_req_is(&req, limited)) {
519 if (u1_has_constr != u2_has_constr)
520 return u2_has_constr - u1_has_constr;
522 /* Now check, whether the two units are connected */
524 for (i=0; i<u1->node_count; ++i)
525 for (o=0; o<u2->node_count; ++o)
526 if (u1->nodes[i] == u2->nodes[o])
530 /* After all, the sort key decides. Greater keys come first. */
531 return u2->sort_key - u1->sort_key;
536 * Sort the ou's according to constraints and their sort_key
538 static void co_sort_units(copy_opt_t *co) {
539 int i, count = 0, costs;
542 /* get the number of ous, remove them form the list and fill the array */
543 list_for_each_entry(unit_t, ou, &co->units, units)
545 ous = alloca(count * sizeof(*ous));
547 costs = co_get_max_copy_costs(co);
550 list_for_each_entry(unit_t, ou, &co->units, units)
553 INIT_LIST_HEAD(&co->units);
555 assert(count == i && list_empty(&co->units));
557 for (i=0; i<count; ++i)
558 ir_printf("%+F\n", ous[i]->nodes[0]);
560 qsort(ous, count, sizeof(*ous), compare_ous);
563 for (i=0; i<count; ++i)
564 ir_printf("%+F\n", ous[i]->nodes[0]);
566 /* reinsert into list in correct order */
567 for (i=0; i<count; ++i)
568 list_add_tail(&ous[i]->units, &co->units);
570 assert(costs == co_get_max_copy_costs(co));
574 void co_build_ou_structure(copy_opt_t *co) {
575 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
576 INIT_LIST_HEAD(&co->units);
577 irg_walk_graph(co->irg, co_collect_units, NULL, co);
578 #ifdef QUICK_AND_DIRTY_HACK
583 void co_free_ou_structure(copy_opt_t *co) {
586 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
591 co->units.next = NULL;
594 /* co_solve_heuristic() is implemented in becopyheur.c */
596 int co_get_max_copy_costs(const copy_opt_t *co) {
602 list_for_each_entry(unit_t, curr, &co->units, units) {
603 res += curr->inevitable_costs;
604 for (i=1; i<curr->node_count; ++i)
605 res += curr->costs[i];
610 int co_get_inevit_copy_costs(const copy_opt_t *co) {
616 list_for_each_entry(unit_t, curr, &co->units, units)
617 res += curr->inevitable_costs;
621 int co_get_copy_costs(const copy_opt_t *co) {
627 list_for_each_entry(unit_t, curr, &co->units, units) {
628 int root_col = get_irn_col(co, curr->nodes[0]);
629 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
630 res += curr->inevitable_costs;
631 for (i=1; i<curr->node_count; ++i) {
632 int arg_col = get_irn_col(co, curr->nodes[i]);
633 if (root_col != arg_col) {
634 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
635 res += curr->costs[i];
642 int co_get_lower_bound(const copy_opt_t *co) {
648 list_for_each_entry(unit_t, curr, &co->units, units)
649 res += curr->inevitable_costs + curr->min_nodes_costs;
653 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
655 bitset_t *seen = bitset_irg_malloc(co->irg);
658 memset(stat, 0, sizeof(stat[0]));
660 /* count affinity edges. */
661 co_gs_foreach_aff_node(co, an) {
663 stat->aff_nodes += 1;
664 bitset_add_irn(seen, an->irn);
665 co_gs_foreach_neighb(an, neigh) {
666 if(!bitset_contains_irn(seen, neigh->irn)) {
667 stat->aff_edges += 1;
668 stat->max_costs += neigh->costs;
670 if(get_irn_col(co, an->irn) != get_irn_col(co, neigh->irn)) {
671 stat->costs += neigh->costs;
672 stat->unsatisfied_edges += 1;
675 if(nodes_interfere(co->cenv, an->irn, neigh->irn)) {
677 stat->inevit_costs += neigh->costs;
687 /******************************************************************************
689 / ____| | | / ____| |
690 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
691 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
692 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
693 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
696 ******************************************************************************/
698 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size) {
699 const affinity_node_t *n1 = k1;
700 const affinity_node_t *n2 = k2;
702 return (n1->irn != n2->irn);
705 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
706 affinity_node_t new_node, *node;
707 neighb_t new_nbr, *nbr;
712 new_node.neighbours = NULL;
713 node = set_insert(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
716 for (nbr = node->neighbours; nbr; nbr = nbr->next)
717 if (nbr->irn == n2) {
722 /* if we did not find n2 in n1's neighbourhood insert it */
724 obstack_grow(&co->obst, &new_nbr, sizeof(new_nbr));
725 nbr = obstack_finish(&co->obst);
728 nbr->next = node->neighbours;
729 node->neighbours = nbr;
733 /* now nbr points to n1's neighbour-entry of n2 */
737 static INLINE void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs) {
738 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
739 add_edge(co, n1, n2, costs);
740 add_edge(co, n2, n1, costs);
744 static void build_graph_walker(ir_node *irn, void *env) {
745 copy_opt_t *co = env;
747 arch_register_req_t req;
748 const arch_register_t *reg;
750 if (!is_curr_reg_class(co, irn) || arch_irn_is(co->aenv, irn, ignore))
753 reg = arch_get_irn_register(co->aenv, irn);
754 if (arch_register_type_is(reg, ignore))
759 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
760 ir_node *arg = get_irn_n(irn, pos);
761 add_edges(co, irn, arg, co->get_costs(co, irn, arg, pos));
765 else if (is_Perm_Proj(co->aenv, irn)) {
766 ir_node *arg = get_Perm_src(irn);
767 add_edges(co, irn, arg, co->get_costs(co, irn, arg, 0));
771 else if (is_2addr_code(co->aenv, irn, &req))
772 add_edges(co, irn, req.other_same, co->get_costs(co, irn, req.other_same, 0));
775 void co_build_graph_structure(copy_opt_t *co) {
776 obstack_init(&co->obst);
777 co->nodes = new_set(compare_affinity_node_t, 32);
779 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
782 void co_free_graph_structure(copy_opt_t *co) {
786 obstack_free(&co->obst, NULL);
790 /* co_solve_ilp1() co_solve_ilp2() are implemented in becopyilpX.c */
792 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn) {
793 affinity_node_t new_node, *n;
798 n = set_find(co->nodes, &new_node, sizeof(new_node), nodeset_hash(new_node.irn));
800 return (n->degree > 0);
805 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
807 be_ifg_t *ifg = co->cenv->ifg;
808 int *color_map = alloca(co->cls->n_regs * sizeof(color_map[0]));
809 bitset_t *adm = bitset_alloca(co->cls->n_regs);
816 for(i = 0; i < co->cls->n_regs; ++i) {
817 const arch_register_t *reg = &co->cls->regs[i];
818 color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_regs++;
822 * n contains the first node number.
823 * the values below n are the pre-colored register nodes
826 it = be_ifg_nodes_iter_alloca(ifg);
827 nit = be_ifg_neighbours_iter_alloca(ifg);
830 be_ifg_foreach_node(ifg, it, irn) {
831 if(!arch_irn_is(co->aenv, irn, ignore))
832 set_irn_link(irn, INT_TO_PTR(n++));
835 fprintf(f, "%d %d\n", n, n_regs);
837 be_ifg_foreach_node(ifg, it, irn) {
838 if(!arch_irn_is(co->aenv, irn, ignore)) {
839 int idx = PTR_TO_INT(get_irn_link(irn));
840 affinity_node_t *a = get_affinity_info(co, irn);
842 arch_register_req_t req;
845 arch_get_register_req(co->aenv, &req, irn, BE_OUT_POS(0));
846 if(arch_register_req_is(&req, limited)) {
847 bitset_clear_all(adm);
848 req.limited(req.limited_env, adm);
849 for(i = 0; i < co->cls->n_regs; ++i)
850 if(!bitset_is_set(adm, i) && color_map[i] >= 0)
851 fprintf(f, "%d %d -1\n", color_map[i], idx);
856 be_ifg_foreach_neighbour(ifg, nit, irn, adj) {
857 if(!arch_irn_is(co->aenv, adj, ignore)) {
858 int adj_idx = PTR_TO_INT(get_irn_link(adj));
860 fprintf(f, "%d %d -1\n", idx, adj_idx);
867 co_gs_foreach_neighb(a, n) {
868 if(!arch_irn_is(co->aenv, n->irn, ignore)) {
869 int n_idx = PTR_TO_INT(get_irn_link(n->irn));
871 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
879 typedef struct _appel_clique_walker_t {
881 const copy_opt_t *co;
888 } appel_clique_walker_t;
890 typedef struct _appel_block_info_t {
900 } appel_block_info_t;
902 static int appel_aff_weight(const appel_clique_walker_t *env, ir_node *bl)
905 double freq = get_block_execfreq(env->co->cenv->execfreq, bl);
906 int res = (int) freq;
907 return res == 0 ? 1 : res;
909 ir_loop *loop = get_irn_loop(bl);
911 int d = get_loop_depth(loop);
918 static void *appel_clique_walker_irn_init(phase_t *phase, ir_node *irn, void *old)
920 appel_block_info_t *res = NULL;
923 appel_clique_walker_t *d = (void *) phase;
924 res = phase_alloc(phase, sizeof(res[0]));
925 res->phi_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
926 res->live_end_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end_nr));
927 res->live_in_nr = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in_nr));
928 res->live_end = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_end));
929 res->live_in = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
930 res->phi = phase_alloc(phase, d->co->cls->n_regs * sizeof(res->live_in));
936 typedef struct _insn_list_t {
938 struct list_head list;
941 static int appel_get_live_end_nr(appel_clique_walker_t *env, ir_node *bl, ir_node *irn)
943 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
946 for(i = 0; i < bli->n_live_end; ++i)
947 if(bli->live_end[i] == irn)
948 return bli->live_end_nr[i];
953 static int appel_dump_clique(appel_clique_walker_t *env, pset *live, ir_node *bl, int curr_nr, int start_nr)
955 ir_node **live_arr = alloca(env->co->cls->n_regs * sizeof(live_arr[0]));
961 foreach_pset(live, irn)
962 live_arr[n_live++] = irn;
964 /* dump the live after clique */
966 for(j = 0; j < n_live; ++j) {
969 for(k = j + 1; k < n_live; ++k) {
970 fprintf(env->f, "%d %d -1 ", curr_nr + j, curr_nr + k);
972 fprintf(env->f, "\n");
976 /* dump the affinities */
977 for(j = 0; !env->dumb && j < n_live; ++j) {
978 ir_node *irn = live_arr[j];
979 int old_nr = PTR_TO_INT(get_irn_link(irn));
981 /* if the node was already live in the last insn dump the affinity */
982 if(old_nr > start_nr) {
983 int weight = appel_aff_weight(env, bl);
984 fprintf(env->f, "%d %d %d\n", old_nr, curr_nr + j, weight);
988 /* set the current numbers into the link field. */
989 for(j = 0; j < n_live; ++j) {
990 ir_node *irn = live_arr[j];
991 set_irn_link(irn, INT_TO_PTR(curr_nr + j));
994 return curr_nr + n_live;
997 static void appel_walker(ir_node *bl, void *data)
999 appel_clique_walker_t *env = data;
1000 appel_block_info_t *bli = phase_get_or_set_irn_data(&env->ph, bl);
1001 struct obstack *obst = &env->obst;
1002 void *base = obstack_base(obst);
1003 pset *live = pset_new_ptr_default();
1007 int start_nr = env->curr_nr;
1008 int curr_nr = start_nr;
1010 be_insn_env_t insn_env;
1015 insn_env.aenv = env->co->aenv;
1016 insn_env.cls = env->co->cls;
1017 insn_env.obst = obst;
1018 insn_env.ignore_colors = env->co->cenv->ignore_colors;
1020 /* Guess how many insns will be in this block. */
1021 sched_foreach(bl, irn)
1025 insns = malloc(n_nodes * sizeof(insns[0]));
1027 /* Put all insns in an array. */
1028 irn = sched_first(bl);
1029 while(!sched_is_end(irn)) {
1031 insn = be_scan_insn(&insn_env, irn);
1032 insns[n_insns++] = insn;
1033 irn = insn->next_insn;
1036 DBG((env->co->cenv->dbg, LEVEL_2, "%+F\n", bl));
1037 be_liveness_end_of_block(env->co->cenv->lv, env->co->aenv, env->co->cls, bl, live);
1039 /* Generate the bad and ugly. */
1040 for(i = n_insns - 1; i >= 0; --i) {
1041 be_insn_t *insn = insns[i];
1043 /* The first live set has to be saved in the block border set. */
1044 if(i == n_insns - 1) {
1046 foreach_pset(live, irn) {
1047 bli->live_end[j] = irn;
1048 bli->live_end_nr[j] = curr_nr + j;
1051 bli->n_live_end = j;
1055 for(j = 0; j < insn->use_start; ++j) {
1056 ir_node *op = insn->ops[j].carrier;
1057 bitset_t *adm = insn->ops[j].regs;
1061 if(!insn->ops[j].has_constraints)
1065 foreach_pset(live, irn) {
1073 assert(nr < pset_count(live));
1075 for(k = 0; k < env->co->cls->n_regs; ++k) {
1076 int mapped_col = env->color_map[k];
1077 if(mapped_col >= 0 && !bitset_is_set(adm, k) && !bitset_is_set(env->co->cenv->ignore_colors, k))
1078 fprintf(env->f, "%d %d -1\n", curr_nr + nr, mapped_col);
1083 /* dump the clique and update the stuff. */
1084 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1086 /* remove all defs. */
1087 for(j = 0; j < insn->use_start; ++j)
1088 pset_remove_ptr(live, insn->ops[j].carrier);
1090 if(is_Phi(insn->irn) && arch_irn_consider_in_reg_alloc(env->co->aenv, env->co->cls, insn->irn)) {
1091 bli->phi[bli->n_phi] = insn->irn;
1092 bli->phi_nr[bli->n_phi] = PTR_TO_INT(get_irn_link(insn->irn));
1098 for(j = insn->use_start; j < insn->n_ops; ++j)
1099 pset_insert_ptr(live, insn->ops[j].carrier);
1102 /* print the start clique. */
1103 curr_nr = appel_dump_clique(env, live, bl, curr_nr, start_nr);
1106 foreach_pset(live, irn) {
1107 bli->live_in[i] = irn;
1108 bli->live_in_nr[i] = PTR_TO_INT(get_irn_link(irn));
1115 obstack_free(obst, base);
1116 env->curr_nr = curr_nr;
1119 static void appel_inter_block_aff(ir_node *bl, void *data)
1121 appel_clique_walker_t *env = data;
1122 appel_block_info_t *bli = phase_get_irn_data(&env->ph, bl);
1126 for(i = 0; i < bli->n_live_in; ++i) {
1127 ir_node *irn = bli->live_in[i];
1129 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1130 ir_node *pred = get_Block_cfgpred_block(bl, j);
1132 int nr = appel_get_live_end_nr(env, pred, irn);
1134 fprintf(env->f, "%d %d 1\n", bli->live_in_nr[i], nr);
1138 for(i = 0; i < bli->n_phi; ++i) {
1139 ir_node *irn = bli->phi[i];
1141 for(j = 0, n = get_Block_n_cfgpreds(bl); j < n; ++j) {
1142 ir_node *pred = get_Block_cfgpred_block(bl, j);
1143 ir_node *op = get_irn_n(irn, j);
1145 int nr = appel_get_live_end_nr(env, pred, op);
1147 fprintf(env->f, "%d %d 1\n", bli->phi_nr[i], nr);
1153 void co_dump_appel_graph_cliques(const copy_opt_t *co, FILE *f)
1157 appel_clique_walker_t env;
1158 bitset_t *adm = bitset_alloca(co->cls->n_regs);
1160 be_liveness_recompute(co->cenv->lv);
1161 obstack_init(&env.obst);
1162 phase_init(&env.ph, "appel_clique_dumper", co->irg, PHASE_DEFAULT_GROWTH, appel_clique_walker_irn_init);
1163 env.curr_nr = co->cls->n_regs;
1167 bitset_copy(adm, co->cenv->ignore_colors);
1168 bitset_flip_all(adm);
1170 /* Make color map. */
1171 env.color_map = alloca(co->cls->n_regs * sizeof(env.color_map[0]));
1172 for(i = 0, n_colors = 0; i < co->cls->n_regs; ++i) {
1173 const arch_register_t *reg = &co->cls->regs[i];
1174 env.color_map[i] = arch_register_type_is(reg, ignore) ? -1 : n_colors++;
1178 env.curr_nr = n_colors;
1179 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1180 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1182 fprintf(f, "%d %d\n", env.curr_nr, n_colors);
1184 /* make the first k nodes interfere */
1185 for(i = 0; i < n_colors; ++i) {
1187 for(j = i + 1; j < n_colors; ++j)
1188 fprintf(f, "%d %d -1 ", i, j);
1193 env.curr_nr = n_colors;
1194 irg_block_walk_graph(co->irg, firm_clear_link, NULL, NULL);
1195 irg_block_walk_graph(co->irg, appel_walker, NULL, &env);
1196 irg_block_walk_graph(co->irg, appel_inter_block_aff, NULL, &env);
1197 obstack_free(&env.obst, NULL);
1201 ___ _____ ____ ____ ___ _____ ____ _
1202 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
1203 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
1204 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1205 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1209 static const char *get_dot_color_name(int col)
1211 static const char *names[] = {
1245 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1248 typedef struct _co_ifg_dump_t {
1249 const copy_opt_t *co;
1253 static void ifg_dump_graph_attr(FILE *f, void *self)
1255 fprintf(f, "overlap=scale");
1258 static int ifg_is_dump_node(void *self, ir_node *irn)
1260 co_ifg_dump_t *cod = self;
1261 return !arch_irn_is(cod->co->aenv, irn, ignore);
1264 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1266 co_ifg_dump_t *env = self;
1267 const arch_register_t *reg = arch_get_irn_register(env->co->aenv, irn);
1268 arch_register_req_t req;
1271 arch_get_register_req(env->co->aenv, &req, irn, BE_OUT_POS(0));
1272 limited = arch_register_req_is(&req, limited);
1274 if(env->flags & CO_IFG_DUMP_LABELS) {
1275 ir_fprintf(f, "label=\"%+F", irn);
1277 if((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1278 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1279 req.limited(req.limited_env, bs);
1280 ir_fprintf(f, "\\n%B", bs);
1282 ir_fprintf(f, "\" ");
1286 fprintf(f, "label=\"\" shape=point " );
1288 if(env->flags & CO_IFG_DUMP_SHAPE)
1289 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1291 if(env->flags & CO_IFG_DUMP_COLORS)
1292 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1295 static void ifg_dump_at_end(FILE *file, void *self)
1297 co_ifg_dump_t *env = self;
1300 co_gs_foreach_aff_node(env->co, a) {
1301 const arch_register_t *ar = arch_get_irn_register(env->co->aenv, a->irn);
1302 unsigned aidx = get_irn_idx(a->irn);
1305 co_gs_foreach_neighb(a, n) {
1306 const arch_register_t *nr = arch_get_irn_register(env->co->aenv, n->irn);
1307 unsigned nidx = get_irn_idx(n->irn);
1310 const char *color = nr == ar ? "blue" : "red";
1311 fprintf(file, "\tn%d -- n%d [weight=0.01 ", aidx, nidx);
1312 if(env->flags & CO_IFG_DUMP_LABELS)
1313 fprintf(file, "label=\"%d\" ", n->costs);
1314 if(env->flags & CO_IFG_DUMP_COLORS)
1315 fprintf(file, "color=%s ", color);
1317 fprintf(file, "style=dotted");
1318 fprintf(file, "];\n");
1325 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1327 ifg_dump_graph_attr,
1336 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1342 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1346 void co_solve_park_moon(copy_opt_t *opt)
1351 static int void_algo(copy_opt_t *co)
1358 / \ | | __ _ ___ _ __(_) |_| |__ _ __ ___ ___
1359 / _ \ | |/ _` |/ _ \| '__| | __| '_ \| '_ ` _ \/ __|
1360 / ___ \| | (_| | (_) | | | | |_| | | | | | | | \__ \
1361 /_/ \_\_|\__, |\___/|_| |_|\__|_| |_|_| |_| |_|___/
1368 int can_improve_existing;
1371 static co_algo_info_t algos[] = {
1372 { void_algo, "none", 0 },
1373 { co_solve_heuristic, "heur1", 0 },
1374 { co_solve_heuristic_new, "heur2", 0 },
1375 { co_solve_heuristic_java, "heur3", 0 },
1377 { co_solve_ilp2, "ilp", 1 },
1384 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1385 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1386 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1387 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1391 void co_driver(be_chordal_env_t *cenv)
1394 lc_timer_t *timer = lc_timer_register("firm.be.copyopt", "runtime");
1396 co_complete_stats_t before, after;
1398 co_algo_t *algo_func;
1399 int was_optimal = 0;
1401 if(algo < 0 || algo >= CO_ALGO_LAST)
1404 co = new_copy_opt(cenv, cost_func);
1405 co_build_ou_structure(co);
1406 co_build_graph_structure(co);
1408 co_complete_stats(co, &before);
1410 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1411 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1412 be_stat_ev_ull("co_max_costs", before.max_costs);
1413 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1414 be_stat_ev_ull("co_aff_int", before.aff_int);
1416 be_stat_ev_ull("co_init_costs", before.costs);
1417 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1419 /* Dump the interference graph in Appel's format. */
1420 if(dump_flags & DUMP_APPEL) {
1421 FILE *f = be_chordal_open(cenv, "", ".apl");
1422 co_dump_appel_graph(co, f);
1426 if(dump_flags & DUMP_BEFORE) {
1427 FILE *f = be_chordal_open(cenv, "", "-before.dot");
1428 co_dump_ifg_dot(co, f, style_flags);
1432 /* if the algo can improve results, provide an initial solution with heur3 */
1433 if(improve && algos[algo].can_improve_existing) {
1434 co_complete_stats_t stats;
1436 /* produce a heuristical solution */
1437 co_solve_heuristic_java(co);
1439 /* do the stats and provide the current costs */
1440 co_complete_stats(co, &stats);
1441 be_stat_ev_ull("co_prepare_costs", stats.costs);
1445 /* start the JVM here so that it does not tamper the timing. */
1446 if(algo == CO_ALGO_HEUR3)
1447 be_java_coal_start_jvm();
1450 algo_func = algos[algo].algo;
1453 lc_timer_reset_and_start(timer);
1456 was_optimal = algo_func(co);
1459 lc_timer_stop(timer);
1460 be_stat_ev("co_time", lc_timer_elapsed_msec(timer));
1463 be_stat_ev_ull("co_optimal", was_optimal);
1465 if(dump_flags & DUMP_AFTER) {
1466 FILE *f = be_chordal_open(cenv, "", "-after.dot");
1467 co_dump_ifg_dot(co, f, style_flags);
1471 co_complete_stats(co, &after);
1474 ulong64 optimizable_costs = after.max_costs - after.inevit_costs;
1475 ulong64 evitable = after.costs - after.inevit_costs;
1477 ir_printf("%30F ", cenv->irg);
1478 printf("%10s %10" ULL_FMT "%10" ULL_FMT "%10" ULL_FMT, cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1480 if(optimizable_costs > 0)
1481 printf("%10" ULL_FMT " %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1483 printf("%10" ULL_FMT " %5s\n", after.costs, "-");
1486 be_stat_ev_ull("co_after_costs", after.costs);
1487 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1489 co_free_graph_structure(co);
1490 co_free_ou_structure(co);