2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief Copy minimization driver.
23 * @author Daniel Grund
26 * Main file for the optimization reducing the copies needed for:
28 * - Register-constrained nodes
29 * - Two-address code instructions
37 #include "raw_bitset.h"
43 #include "iredges_t.h"
44 #include "irprintf_t.h"
53 #include "beintlive_t.h"
54 #include "becopyopt_t.h"
55 #include "becopystat.h"
59 #include "bestatevent.h"
64 #include "lc_opts_enum.h"
69 #define DUMP_ALL 2 * DUMP_APPEL - 1
71 #define COST_FUNC_FREQ 1
72 #define COST_FUNC_LOOP 2
73 #define COST_FUNC_ALL_ONE 3
76 * Flags for dumping the IFG.
79 CO_IFG_DUMP_COLORS = 1 << 0, /**< Dump the graph colored. */
80 CO_IFG_DUMP_LABELS = 1 << 1, /**< Dump node/edge labels. */
81 CO_IFG_DUMP_SHAPE = 1 << 2, /**< Give constrained nodes special shapes. */
82 CO_IFG_DUMP_CONSTR = 1 << 3, /**< Dump the node constraints in the label. */
85 static int co_get_costs_loop_depth(const ir_node *root, int pos);
86 static int co_get_costs_exec_freq(const ir_node *root, int pos);
87 static int co_get_costs_all_one(const ir_node *root, int pos);
89 static unsigned dump_flags = 0;
90 static unsigned style_flags = 0;
91 static int do_stats = 0;
92 static cost_fct_t cost_func = co_get_costs_exec_freq;
93 static int improve = 1;
95 static const lc_opt_enum_mask_items_t dump_items[] = {
96 { "before", DUMP_BEFORE },
97 { "after", DUMP_AFTER },
98 { "appel", DUMP_APPEL },
104 * Flags for dumping the IFG.
107 CO_IFG_DUMP_COLORS = 1 << 0, /**< Dump the graph colored. */
108 CO_IFG_DUMP_LABELS = 1 << 1, /**< Dump node/edge labels. */
109 CO_IFG_DUMP_SHAPE = 1 << 2, /**< Give constrained nodes special shapes. */
110 CO_IFG_DUMP_CONSTR = 1 << 3, /**< Dump the node constraints in the label. */
113 static const lc_opt_enum_mask_items_t style_items[] = {
114 { "color", CO_IFG_DUMP_COLORS },
115 { "labels", CO_IFG_DUMP_LABELS },
116 { "constr", CO_IFG_DUMP_CONSTR },
117 { "shape", CO_IFG_DUMP_SHAPE },
118 { "full", 2 * CO_IFG_DUMP_SHAPE - 1 },
122 typedef int (*opt_funcptr)(void);
123 static const lc_opt_enum_func_ptr_items_t cost_func_items[] = {
124 { "freq", (opt_funcptr) co_get_costs_exec_freq },
125 { "loop", (opt_funcptr) co_get_costs_loop_depth },
126 { "one", (opt_funcptr) co_get_costs_all_one },
130 static lc_opt_enum_mask_var_t dump_var = {
131 &dump_flags, dump_items
134 static lc_opt_enum_mask_var_t style_var = {
135 &style_flags, style_items
138 static lc_opt_enum_func_ptr_var_t cost_func_var = {
139 (opt_funcptr*) &cost_func, cost_func_items
142 static const lc_opt_table_entry_t options[] = {
143 LC_OPT_ENT_ENUM_FUNC_PTR ("cost", "select a cost function", &cost_func_var),
144 LC_OPT_ENT_ENUM_MASK ("dump", "dump ifg before or after copy optimization", &dump_var),
145 LC_OPT_ENT_ENUM_MASK ("style", "dump style for ifg dumping", &style_var),
146 LC_OPT_ENT_BOOL ("stats", "dump statistics after each optimization", &do_stats),
147 LC_OPT_ENT_BOOL ("improve", "run heur1 before if algo can exploit start solutions", &improve),
151 static be_module_list_entry_t *copyopts = NULL;
152 static const co_algo_info *selected_copyopt = NULL;
154 void be_register_copyopt(const char *name, co_algo_info *copyopt)
156 if (selected_copyopt == NULL)
157 selected_copyopt = copyopt;
158 be_add_module_to_list(©opts, name, copyopt);
161 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copyopt)
162 void be_init_copyopt(void)
164 lc_opt_entry_t *be_grp = lc_opt_get_grp(firm_opt_get_root(), "be");
165 lc_opt_entry_t *ra_grp = lc_opt_get_grp(be_grp, "ra");
166 lc_opt_entry_t *chordal_grp = lc_opt_get_grp(ra_grp, "chordal");
167 lc_opt_entry_t *co_grp = lc_opt_get_grp(chordal_grp, "co");
169 lc_opt_add_table(co_grp, options);
170 be_add_module_list_opt(co_grp, "algo", "select copy optimization algo",
171 ©opts, (void**) &selected_copyopt);
174 static int void_algo(copy_opt_t *co)
180 BE_REGISTER_MODULE_CONSTRUCTOR(be_init_copynone)
181 void be_init_copynone(void)
183 static co_algo_info copyheur = {
187 be_register_copyopt("none", ©heur);
190 #undef QUICK_AND_DIRTY_HACK
192 static int nodes_interfere(const be_chordal_env_t *env, const ir_node *a, const ir_node *b)
195 return be_ifg_connected(env->ifg, a, b);
197 be_lv_t *lv = be_get_irg_liveness(env->irg);
198 return be_values_interfere(lv, a, b);
203 /******************************************************************************
206 | | __ ___ _ __ ___ _ __ __ _| |
207 | | |_ |/ _ \ '_ \ / _ \ '__/ _` | |
208 | |__| | __/ | | | __/ | | (_| | |
209 \_____|\___|_| |_|\___|_| \__,_|_|
211 ******************************************************************************/
213 DEBUG_ONLY(static firm_dbg_module_t *dbg = NULL;)
216 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, cost_fct_t get_costs)
218 const char *s1, *s2, *s3;
222 FIRM_DBG_REGISTER(dbg, "ir.be.copyopt");
224 co = XMALLOCZ(copy_opt_t);
225 co->cenv = chordal_env;
226 co->irg = chordal_env->irg;
227 co->cls = chordal_env->cls;
228 co->get_costs = get_costs;
231 s2 = get_entity_name(get_irg_entity(co->irg));
232 s3 = chordal_env->cls->name;
233 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
234 co->name = XMALLOCN(char, len);
235 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
240 void free_copy_opt(copy_opt_t *co)
247 * Checks if a node is optimizable, viz. has something to do with coalescing
248 * @param irn The irn to check
250 static int co_is_optimizable_root(ir_node *irn)
252 const arch_register_req_t *req;
254 if (arch_irn_is_ignore(irn))
257 if (is_Reg_Phi(irn) || is_Perm_Proj(irn))
260 req = arch_get_irn_register_req(irn);
261 if (is_2addr_code(req))
268 * Computes the costs of a copy according to loop depth
269 * @param pos the argument position of arg in the root arguments
270 * @return Must be >= 0 in all cases.
272 static int co_get_costs_loop_depth(const ir_node *root, int pos)
274 ir_node *block = get_nodes_block(root);
279 block = get_Block_cfgpred_block(block, pos);
281 loop = get_irn_loop(block);
283 int d = get_loop_depth(loop);
292 * Computes the costs of a copy according to execution frequency
293 * @param pos the argument position of arg in the root arguments
294 * @return Must be >= 0 in all cases.
296 static int co_get_costs_exec_freq(const ir_node *root, int pos)
298 ir_graph *irg = get_irn_irg(root);
299 ir_node *root_bl = get_nodes_block(root);
301 = is_Phi(root) ? get_Block_cfgpred_block(root_bl, pos) : root_bl;
302 ir_exec_freq *exec_freq = be_get_irg_exec_freq(irg);
303 int res = get_block_execfreq_ulong(exec_freq, copy_bl);
305 /* don't allow values smaller than one. */
306 return res < 1 ? 1 : res;
310 * All costs equal 1. Using this will reduce the _number_ of copies.
311 * @param co The copy opt object.
312 * @return Must be >= 0 in all cases.
314 static int co_get_costs_all_one(const ir_node *root, int pos)
321 /******************************************************************************
322 ____ _ _ _ _ _ _____ _
323 / __ \ | | | | | | (_) | / ____| |
324 | | | |_ __ | |_| | | |_ __ _| |_ ___ | (___ | |_ ___ _ __ __ _ __ _ ___
325 | | | | '_ \| __| | | | '_ \| | __/ __| \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
326 | |__| | |_) | |_| |__| | | | | | |_\__ \ ____) | || (_) | | | (_| | (_| | __/
327 \____/| .__/ \__|\____/|_| |_|_|\__|___/ |_____/ \__\___/|_| \__,_|\__, |\___|
330 ******************************************************************************/
333 * Determines a maximum weighted independent set with respect to
334 * the interference and conflict edges of all nodes in a qnode.
336 static int ou_max_ind_set_costs(unit_t *ou)
338 be_chordal_env_t *chordal_env = ou->co->cenv;
339 ir_node **safe, **unsafe;
340 int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
343 int curr_weight, best_weight = 0;
345 /* assign the nodes into two groups.
346 * safe: node has no interference, hence it is in every max stable set.
347 * unsafe: node has an interference
349 safe = ALLOCAN(ir_node*, ou->node_count - 1);
352 unsafe = ALLOCAN(ir_node*, ou->node_count - 1);
353 unsafe_costs = ALLOCAN(int, ou->node_count - 1);
355 for (i=1; i<ou->node_count; ++i) {
357 for (o=1; o<ou->node_count; ++o) {
360 if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
361 unsafe_costs[unsafe_count] = ou->costs[i];
362 unsafe[unsafe_count] = ou->nodes[i];
369 safe_costs += ou->costs[i];
370 safe[safe_count++] = ou->nodes[i];
375 /* now compute the best set out of the unsafe nodes*/
376 if (unsafe_count > MIS_HEUR_TRIGGER) {
377 bitset_t *best = bitset_alloca(unsafe_count);
378 /* Heuristic: Greedy trial and error form index 0 to unsafe_count-1 */
379 for (i=0; i<unsafe_count; ++i) {
381 /* check if it is a stable set */
382 for (o=bitset_next_set(best, 0); o!=-1 && o<i; o=bitset_next_set(best, o+1))
383 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o])) {
384 bitset_clear(best, i); /* clear the bit and try next one */
388 /* compute the weight */
389 bitset_foreach(best, pos)
390 best_weight += unsafe_costs[pos];
392 /* Exact Algorithm: Brute force */
393 curr = bitset_alloca(unsafe_count);
394 bitset_set_all(curr);
395 while (bitset_popcount(curr) != 0) {
396 /* check if curr is a stable set */
397 for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
398 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) */
399 if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
402 /* if we arrive here, we have a stable set */
403 /* compute the weight of the stable set*/
405 bitset_foreach(curr, pos)
406 curr_weight += unsafe_costs[pos];
409 if (curr_weight > best_weight) {
410 best_weight = curr_weight;
418 return safe_costs+best_weight;
421 static void co_collect_units(ir_node *irn, void *env)
423 const arch_register_req_t *req;
424 copy_opt_t *co = (copy_opt_t*)env;
427 if (get_irn_mode(irn) == mode_T)
429 req = arch_get_irn_register_req(irn);
430 if (req->cls != co->cls)
432 if (!co_is_optimizable_root(irn))
435 /* Init a new unit */
436 unit = XMALLOCZ(unit_t);
438 unit->node_count = 1;
439 INIT_LIST_HEAD(&unit->queue);
441 /* Phi with some/all of its arguments */
442 if (is_Reg_Phi(irn)) {
446 arity = get_irn_arity(irn);
447 unit->nodes = XMALLOCN(ir_node*, arity + 1);
448 unit->costs = XMALLOCN(int, arity + 1);
449 unit->nodes[0] = irn;
452 for (i=0; i<arity; ++i) {
454 ir_node *arg = get_irn_n(irn, i);
456 assert(arch_get_irn_reg_class(arg) == co->cls && "Argument not in same register class.");
459 if (nodes_interfere(co->cenv, irn, arg)) {
460 unit->inevitable_costs += co->get_costs(irn, i);
464 /* Else insert the argument of the phi to the members of this ou */
465 DBG((dbg, LEVEL_1, "\t Member: %+F\n", arg));
467 if (arch_irn_is_ignore(arg))
470 /* Check if arg has occurred at a prior position in the arg/list */
472 for (o=1; o<unit->node_count; ++o) {
473 if (unit->nodes[o] == arg) {
479 if (!arg_pos) { /* a new argument */
480 /* insert node, set costs */
481 unit->nodes[unit->node_count] = arg;
482 unit->costs[unit->node_count] = co->get_costs(irn, i);
484 } else { /* arg has occurred before in same phi */
485 /* increase costs for existing arg */
486 unit->costs[arg_pos] += co->get_costs(irn, i);
489 unit->nodes = XREALLOC(unit->nodes, ir_node*, unit->node_count);
490 unit->costs = XREALLOC(unit->costs, int, unit->node_count);
491 } else if (is_Perm_Proj(irn)) {
492 /* Proj of a perm with corresponding arg */
493 assert(!nodes_interfere(co->cenv, irn, get_Perm_src(irn)));
494 unit->nodes = XMALLOCN(ir_node*, 2);
495 unit->costs = XMALLOCN(int, 2);
496 unit->node_count = 2;
497 unit->nodes[0] = irn;
498 unit->nodes[1] = get_Perm_src(irn);
499 unit->costs[1] = co->get_costs(irn, -1);
501 /* Src == Tgt of a 2-addr-code instruction */
502 if (is_2addr_code(req)) {
503 const unsigned other = req->other_same;
507 for (i = 0; (1U << i) <= other; ++i) {
508 if (other & (1U << i)) {
509 ir_node *o = get_irn_n(skip_Proj(irn), i);
510 if (arch_irn_is_ignore(o))
512 if (nodes_interfere(co->cenv, irn, o))
521 unit->nodes = XMALLOCN(ir_node*, count);
522 unit->costs = XMALLOCN(int, count);
523 unit->node_count = count;
524 unit->nodes[k++] = irn;
526 for (i = 0; 1U << i <= other; ++i) {
527 if (other & (1U << i)) {
528 ir_node *o = get_irn_n(skip_Proj(irn), i);
529 if (!arch_irn_is_ignore(o) &&
530 !nodes_interfere(co->cenv, irn, o)) {
532 unit->costs[k] = co->get_costs(irn, -1);
539 assert(0 && "This is not an optimizable node!");
543 /* Insert the new unit at a position according to its costs */
544 if (unit->node_count > 1) {
546 struct list_head *tmp;
548 /* Determine the maximum costs this unit can cause: all_nodes_cost */
549 for (i=1; i<unit->node_count; ++i) {
550 unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
551 unit->all_nodes_costs += unit->costs[i];
554 /* Determine the minimal costs this unit will cause: min_nodes_costs */
555 unit->min_nodes_costs += unit->all_nodes_costs - ou_max_ind_set_costs(unit);
556 /* Insert the new ou according to its sort_key */
558 while (tmp->next != &co->units && list_entry_units(tmp->next)->sort_key > unit->sort_key)
560 list_add(&unit->units, tmp);
566 #ifdef QUICK_AND_DIRTY_HACK
568 static int compare_ous(const void *k1, const void *k2)
570 const unit_t *u1 = *((const unit_t **) k1);
571 const unit_t *u2 = *((const unit_t **) k2);
572 int i, o, u1_has_constr, u2_has_constr;
573 arch_register_req_t req;
575 /* Units with constraints come first */
577 for (i=0; i<u1->node_count; ++i) {
578 arch_get_irn_register_req(&req, u1->nodes[i]);
579 if (arch_register_req_is(&req, limited)) {
586 for (i=0; i<u2->node_count; ++i) {
587 arch_get_irn_register_req(&req, u2->nodes[i]);
588 if (arch_register_req_is(&req, limited)) {
594 if (u1_has_constr != u2_has_constr)
595 return u2_has_constr - u1_has_constr;
597 /* Now check, whether the two units are connected */
599 for (i=0; i<u1->node_count; ++i)
600 for (o=0; o<u2->node_count; ++o)
601 if (u1->nodes[i] == u2->nodes[o])
605 /* After all, the sort key decides. Greater keys come first. */
606 return u2->sort_key - u1->sort_key;
611 * Sort the ou's according to constraints and their sort_key
613 static void co_sort_units(copy_opt_t *co)
615 int i, count = 0, costs;
618 /* get the number of ous, remove them form the list and fill the array */
619 list_for_each_entry(unit_t, ou, &co->units, units)
621 ous = ALLOCAN(unit_t, count);
623 costs = co_get_max_copy_costs(co);
626 list_for_each_entry(unit_t, ou, &co->units, units)
629 INIT_LIST_HEAD(&co->units);
631 assert(count == i && list_empty(&co->units));
633 for (i=0; i<count; ++i)
634 ir_printf("%+F\n", ous[i]->nodes[0]);
636 qsort(ous, count, sizeof(*ous), compare_ous);
639 for (i=0; i<count; ++i)
640 ir_printf("%+F\n", ous[i]->nodes[0]);
642 /* reinsert into list in correct order */
643 for (i=0; i<count; ++i)
644 list_add_tail(&ous[i]->units, &co->units);
646 assert(costs == co_get_max_copy_costs(co));
650 void co_build_ou_structure(copy_opt_t *co)
652 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
653 INIT_LIST_HEAD(&co->units);
654 irg_walk_graph(co->irg, co_collect_units, NULL, co);
655 #ifdef QUICK_AND_DIRTY_HACK
660 void co_free_ou_structure(copy_opt_t *co)
664 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
669 co->units.next = NULL;
672 /* co_solve_heuristic() is implemented in becopyheur.c */
674 int co_get_max_copy_costs(const copy_opt_t *co)
681 list_for_each_entry(unit_t, curr, &co->units, units) {
682 res += curr->inevitable_costs;
683 for (i=1; i<curr->node_count; ++i)
684 res += curr->costs[i];
689 int co_get_inevit_copy_costs(const copy_opt_t *co)
696 list_for_each_entry(unit_t, curr, &co->units, units)
697 res += curr->inevitable_costs;
701 int co_get_copy_costs(const copy_opt_t *co)
708 list_for_each_entry(unit_t, curr, &co->units, units) {
709 int root_col = get_irn_col(curr->nodes[0]);
710 DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
711 res += curr->inevitable_costs;
712 for (i=1; i<curr->node_count; ++i) {
713 int arg_col = get_irn_col(curr->nodes[i]);
714 if (root_col != arg_col) {
715 DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
716 res += curr->costs[i];
723 int co_get_lower_bound(const copy_opt_t *co)
730 list_for_each_entry(unit_t, curr, &co->units, units)
731 res += curr->inevitable_costs + curr->min_nodes_costs;
735 void co_complete_stats(const copy_opt_t *co, co_complete_stats_t *stat)
737 bitset_t *seen = bitset_malloc(get_irg_last_idx(co->irg));
740 memset(stat, 0, sizeof(stat[0]));
742 /* count affinity edges. */
743 co_gs_foreach_aff_node(co, an) {
745 stat->aff_nodes += 1;
746 bitset_set(seen, get_irn_idx(an->irn));
747 co_gs_foreach_neighb(an, neigh) {
748 if (!bitset_is_set(seen, get_irn_idx(neigh->irn))) {
749 stat->aff_edges += 1;
750 stat->max_costs += neigh->costs;
752 if (get_irn_col(an->irn) != get_irn_col(neigh->irn)) {
753 stat->costs += neigh->costs;
754 stat->unsatisfied_edges += 1;
757 if (nodes_interfere(co->cenv, an->irn, neigh->irn)) {
759 stat->inevit_costs += neigh->costs;
769 /******************************************************************************
771 / ____| | | / ____| |
772 | | __ _ __ __ _ _ __ | |__ | (___ | |_ ___ _ __ __ _ __ _ ___
773 | | |_ | '__/ _` | '_ \| '_ \ \___ \| __/ _ \| '__/ _` |/ _` |/ _ \
774 | |__| | | | (_| | |_) | | | | ____) | || (_) | | | (_| | (_| | __/
775 \_____|_| \__,_| .__/|_| |_| |_____/ \__\___/|_| \__,_|\__, |\___|
778 ******************************************************************************/
780 static int compare_affinity_node_t(const void *k1, const void *k2, size_t size)
782 const affinity_node_t *n1 = (const affinity_node_t*)k1;
783 const affinity_node_t *n2 = (const affinity_node_t*)k2;
786 return (n1->irn != n2->irn);
789 static void add_edge(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
791 affinity_node_t new_node, *node;
797 new_node.neighbours = NULL;
798 node = (affinity_node_t*)set_insert(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
800 for (nbr = node->neighbours; nbr; nbr = nbr->next)
801 if (nbr->irn == n2) {
806 /* if we did not find n2 in n1's neighbourhood insert it */
808 nbr = OALLOC(&co->obst, neighb_t);
811 nbr->next = node->neighbours;
813 node->neighbours = nbr;
817 /* now nbr points to n1's neighbour-entry of n2 */
821 static inline void add_edges(copy_opt_t *co, ir_node *n1, ir_node *n2, int costs)
823 if (! be_ifg_connected(co->cenv->ifg, n1, n2)) {
824 add_edge(co, n1, n2, costs);
825 add_edge(co, n2, n1, costs);
829 static void build_graph_walker(ir_node *irn, void *env)
831 const arch_register_req_t *req;
832 copy_opt_t *co = (copy_opt_t*)env;
835 if (get_irn_mode(irn) == mode_T)
837 req = arch_get_irn_register_req(irn);
838 if (req->cls != co->cls || arch_irn_is_ignore(irn))
841 if (is_Reg_Phi(irn)) { /* Phis */
842 for (pos=0, max=get_irn_arity(irn); pos<max; ++pos) {
843 ir_node *arg = get_irn_n(irn, pos);
844 add_edges(co, irn, arg, co->get_costs(irn, pos));
846 } else if (is_Perm_Proj(irn)) { /* Perms */
847 ir_node *arg = get_Perm_src(irn);
848 add_edges(co, irn, arg, co->get_costs(irn, -1));
849 } else { /* 2-address code */
850 if (is_2addr_code(req)) {
851 const unsigned other = req->other_same;
854 for (i = 0; 1U << i <= other; ++i) {
855 if (other & (1U << i)) {
856 ir_node *other = get_irn_n(skip_Proj(irn), i);
857 if (!arch_irn_is_ignore(other))
858 add_edges(co, irn, other, co->get_costs(irn, -1));
865 void co_build_graph_structure(copy_opt_t *co)
867 obstack_init(&co->obst);
868 co->nodes = new_set(compare_affinity_node_t, 32);
870 irg_walk_graph(co->irg, build_graph_walker, NULL, co);
873 void co_free_graph_structure(copy_opt_t *co)
878 obstack_free(&co->obst, NULL);
882 int co_gs_is_optimizable(copy_opt_t *co, ir_node *irn)
884 affinity_node_t new_node, *n;
889 n = (affinity_node_t*)set_find(co->nodes, &new_node, sizeof(new_node), hash_irn(new_node.irn));
891 return (n->degree > 0);
896 static int co_dump_appel_disjoint_constraints(const copy_opt_t *co, ir_node *a, ir_node *b)
898 ir_node *nodes[] = { a, b };
899 bitset_t *constr[] = { NULL, NULL };
902 constr[0] = bitset_alloca(co->cls->n_regs);
903 constr[1] = bitset_alloca(co->cls->n_regs);
905 for (j = 0; j < 2; ++j) {
906 const arch_register_req_t *req = arch_get_irn_register_req(nodes[j]);
907 if (arch_register_req_is(req, limited))
908 rbitset_copy_to_bitset(req->limited, constr[j]);
910 bitset_set_all(constr[j]);
914 return !bitset_intersect(constr[0], constr[1]);
917 void co_dump_appel_graph(const copy_opt_t *co, FILE *f)
919 be_ifg_t *ifg = co->cenv->ifg;
920 int *color_map = ALLOCAN(int, co->cls->n_regs);
921 int *node_map = XMALLOCN(int, get_irg_last_idx(co->irg) + 1);
922 ir_graph *irg = co->irg;
923 be_irg_t *birg = be_birg_from_irg(irg);
927 neighbours_iter_t nit;
932 for (i = 0; i < co->cls->n_regs; ++i) {
933 const arch_register_t *reg = &co->cls->regs[i];
934 if (rbitset_is_set(birg->allocatable_regs, reg->global_index)) {
935 color_map[i] = n_regs++;
942 * n contains the first node number.
943 * the values below n are the pre-colored register nodes
947 be_ifg_foreach_node(ifg, &it, irn) {
948 if (arch_irn_is_ignore(irn))
950 node_map[get_irn_idx(irn)] = n++;
953 fprintf(f, "%d %d\n", n, n_regs);
955 be_ifg_foreach_node(ifg, &it, irn) {
956 if (!arch_irn_is_ignore(irn)) {
957 int idx = node_map[get_irn_idx(irn)];
958 affinity_node_t *a = get_affinity_info(co, irn);
959 const arch_register_req_t *req = arch_get_irn_register_req(irn);
962 if (arch_register_req_is(req, limited)) {
963 for (i = 0; i < co->cls->n_regs; ++i) {
964 if (!rbitset_is_set(req->limited, i) && color_map[i] >= 0)
965 fprintf(f, "%d %d -1\n", color_map[i], idx);
969 be_ifg_foreach_neighbour(ifg, &nit, irn, adj) {
970 if (!arch_irn_is_ignore(adj) &&
971 !co_dump_appel_disjoint_constraints(co, irn, adj)) {
972 int adj_idx = node_map[get_irn_idx(adj)];
974 fprintf(f, "%d %d -1\n", idx, adj_idx);
981 co_gs_foreach_neighb(a, n) {
982 if (!arch_irn_is_ignore(n->irn)) {
983 int n_idx = node_map[get_irn_idx(n->irn)];
985 fprintf(f, "%d %d %d\n", idx, n_idx, (int) n->costs);
996 ___ _____ ____ ____ ___ _____ ____ _
997 |_ _| ___/ ___| | _ \ / _ \_ _| | _ \ _ _ _ __ ___ _ __ (_)_ __ __ _
998 | || |_ | | _ | | | | | | || | | | | | | | | '_ ` _ \| '_ \| | '_ \ / _` |
999 | || _|| |_| | | |_| | |_| || | | |_| | |_| | | | | | | |_) | | | | | (_| |
1000 |___|_| \____| |____/ \___/ |_| |____/ \__,_|_| |_| |_| .__/|_|_| |_|\__, |
1004 static const char *get_dot_color_name(size_t col)
1006 static const char *names[] = {
1040 return col < sizeof(names)/sizeof(names[0]) ? names[col] : "white";
1043 typedef struct co_ifg_dump_t {
1044 const copy_opt_t *co;
1048 static void ifg_dump_graph_attr(FILE *f, void *self)
1051 fprintf(f, "overlap=scale");
1054 static int ifg_is_dump_node(void *self, ir_node *irn)
1057 return !arch_irn_is_ignore(irn);
1060 static void ifg_dump_node_attr(FILE *f, void *self, ir_node *irn)
1062 co_ifg_dump_t *env = (co_ifg_dump_t*)self;
1063 const arch_register_t *reg = arch_get_irn_register(irn);
1064 const arch_register_req_t *req = arch_get_irn_register_req(irn);
1065 int limited = arch_register_req_is(req, limited);
1067 if (env->flags & CO_IFG_DUMP_LABELS) {
1068 ir_fprintf(f, "label=\"%+F", irn);
1070 if ((env->flags & CO_IFG_DUMP_CONSTR) && limited) {
1071 bitset_t *bs = bitset_alloca(env->co->cls->n_regs);
1072 rbitset_copy_to_bitset(req->limited, bs);
1073 ir_fprintf(f, "\\n%B", bs);
1075 ir_fprintf(f, "\" ");
1077 fprintf(f, "label=\"\" shape=point " );
1080 if (env->flags & CO_IFG_DUMP_SHAPE)
1081 fprintf(f, "shape=%s ", limited ? "diamond" : "ellipse");
1083 if (env->flags & CO_IFG_DUMP_COLORS)
1084 fprintf(f, "style=filled color=%s ", get_dot_color_name(reg->index));
1087 static void ifg_dump_at_end(FILE *file, void *self)
1089 co_ifg_dump_t *env = (co_ifg_dump_t*)self;
1092 co_gs_foreach_aff_node(env->co, a) {
1093 const arch_register_t *ar = arch_get_irn_register(a->irn);
1094 unsigned aidx = get_irn_idx(a->irn);
1097 co_gs_foreach_neighb(a, n) {
1098 const arch_register_t *nr = arch_get_irn_register(n->irn);
1099 unsigned nidx = get_irn_idx(n->irn);
1102 const char *color = nr == ar ? "blue" : "red";
1103 fprintf(file, "\tn%u -- n%u [weight=0.01 ", aidx, nidx);
1104 if (env->flags & CO_IFG_DUMP_LABELS)
1105 fprintf(file, "label=\"%d\" ", n->costs);
1106 if (env->flags & CO_IFG_DUMP_COLORS)
1107 fprintf(file, "color=%s ", color);
1109 fprintf(file, "style=dotted");
1110 fprintf(file, "];\n");
1117 static be_ifg_dump_dot_cb_t ifg_dot_cb = {
1119 ifg_dump_graph_attr,
1128 void co_dump_ifg_dot(const copy_opt_t *co, FILE *f, unsigned flags)
1134 be_ifg_dump_dot(co->cenv->ifg, co->irg, f, &ifg_dot_cb, &cod);
1139 | \/ | __ _(_)_ __ | _ \ _ __(_)_ _____ _ __
1140 | |\/| |/ _` | | '_ \ | | | | '__| \ \ / / _ \ '__|
1141 | | | | (_| | | | | | | |_| | | | |\ V / __/ |
1142 |_| |_|\__,_|_|_| |_| |____/|_| |_| \_/ \___|_|
1146 static FILE *my_open(const be_chordal_env_t *env, const char *prefix, const char *suffix)
1152 const char *cup_name = be_get_irg_main_env(env->irg)->cup_name;
1154 n = strlen(cup_name);
1155 tu_name = XMALLOCN(char, n + 1);
1156 strcpy(tu_name, cup_name);
1157 for (i = 0; i < n; ++i)
1158 if (tu_name[i] == '.')
1162 ir_snprintf(buf, sizeof(buf), "%s%s_%F_%s%s", prefix, tu_name, env->irg, env->cls->name, suffix);
1164 result = fopen(buf, "wt");
1165 if (result == NULL) {
1166 panic("Couldn't open '%s' for writing.", buf);
1172 void co_driver(be_chordal_env_t *cenv)
1174 ir_timer_t *timer = ir_timer_new();
1175 co_complete_stats_t before, after;
1177 int was_optimal = 0;
1179 assert(selected_copyopt);
1181 /* skip copymin if algo is 'none' */
1182 if (selected_copyopt->copyopt == void_algo)
1185 be_liveness_assure_chk(be_get_irg_liveness(cenv->irg));
1187 co = new_copy_opt(cenv, cost_func);
1188 co_build_ou_structure(co);
1189 co_build_graph_structure(co);
1191 co_complete_stats(co, &before);
1193 be_stat_ev_ull("co_aff_nodes", before.aff_nodes);
1194 be_stat_ev_ull("co_aff_edges", before.aff_edges);
1195 be_stat_ev_ull("co_max_costs", before.max_costs);
1196 be_stat_ev_ull("co_inevit_costs", before.inevit_costs);
1197 be_stat_ev_ull("co_aff_int", before.aff_int);
1199 be_stat_ev_ull("co_init_costs", before.costs);
1200 be_stat_ev_ull("co_init_unsat", before.unsatisfied_edges);
1202 if (dump_flags & DUMP_BEFORE) {
1203 FILE *f = my_open(cenv, "", "-before.dot");
1204 co_dump_ifg_dot(co, f, style_flags);
1208 /* if the algo can improve results, provide an initial solution with heur1 */
1209 if (improve && selected_copyopt->can_improve_existing) {
1210 co_complete_stats_t stats;
1212 /* produce a heuristic solution */
1213 co_solve_heuristic(co);
1215 /* do the stats and provide the current costs */
1216 co_complete_stats(co, &stats);
1217 be_stat_ev_ull("co_prepare_costs", stats.costs);
1220 /* perform actual copy minimization */
1221 ir_timer_reset_and_start(timer);
1222 was_optimal = selected_copyopt->copyopt(co);
1223 ir_timer_stop(timer);
1225 be_stat_ev("co_time", ir_timer_elapsed_msec(timer));
1226 be_stat_ev_ull("co_optimal", was_optimal);
1227 ir_timer_free(timer);
1229 if (dump_flags & DUMP_AFTER) {
1230 FILE *f = my_open(cenv, "", "-after.dot");
1231 co_dump_ifg_dot(co, f, style_flags);
1235 co_complete_stats(co, &after);
1238 unsigned long long optimizable_costs = after.max_costs - after.inevit_costs;
1239 unsigned long long evitable = after.costs - after.inevit_costs;
1241 ir_printf("%30F ", cenv->irg);
1242 printf("%10s %10llu%10llu%10llu", cenv->cls->name, after.max_costs, before.costs, after.inevit_costs);
1244 if (optimizable_costs > 0)
1245 printf("%10llu %5.2f\n", after.costs, (evitable * 100.0) / optimizable_costs);
1247 printf("%10llu %5s\n", after.costs, "-");
1250 /* Dump the interference graph in Appel's format. */
1251 if (dump_flags & DUMP_APPEL) {
1252 FILE *f = my_open(cenv, "", ".apl");
1253 fprintf(f, "# %llu %llu\n", after.costs, after.unsatisfied_edges);
1254 co_dump_appel_graph(co, f);
1258 be_stat_ev_ull("co_after_costs", after.costs);
1259 be_stat_ev_ull("co_after_unsat", after.unsatisfied_edges);
1261 co_free_graph_structure(co);
1262 co_free_ou_structure(co);