4 * Copyright: (c) Universitaet Karlsruhe
5 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
21 #include "bechordal_t.h"
22 #include "becopyopt.h"
23 #include "becopystat.h"
25 #define DEBUG_LVL 0 //SET_LEVEL_1
26 static firm_dbg_module_t *dbg = NULL;
28 #define is_curr_reg_class(irn) (arch_get_irn_reg_class(co->chordal_env->arch_env, irn, arch_pos_make_out(0)) == co->chordal_env->cls)
30 #define MIN(a,b) ((a<b)?(a):(b))
33 * Computes the weight of a 'max independent set' wrt. ifg-edges only
34 * (no coloring conflicts, no register constraints)
35 * @return The costs of such a mis
36 * NOTE: Code adapted from becopyheur
37 * BETTER: Here we can be sure having a chordal graph to work on,
38 * so, for 'larger' opt-units we could use a special algorithm.
40 static int ou_max_ind_set_costs(unit_t *ou) {
42 int max, pos, curr_weight, best_weight = 0;
45 irns = alloca((ou->node_count-1) * sizeof(*irns));
46 curr = bitset_alloca(ou->node_count-1);
48 /* brute force the best set */
50 while ((max = bitset_popcnt(curr)) != 0) {
51 /* check if curr is a stable set */
52 int i, o, is_stable_set = 1;
53 bitset_foreach(curr, pos)
54 irns[pos] = ou->nodes[1+pos];
56 for(o=i+1; o<max; ++o) /* !!!!! difference to qnode_max_ind_set(): NOT o=i */
57 if (nodes_interfere(ou->co->chordal_env, irns[i], irns[o])) {
63 /* calc current weigth */
65 bitset_foreach(curr, pos)
66 curr_weight += ou->costs[1+pos];
69 if (curr_weight > best_weight)
70 best_weight = curr_weight;
79 * Builds an optimization unit for a given optimizable irn (root).
80 * This opt-unit is inserted in the main structure co.
81 * If an arg of root itself is optimizable process this arg before with a
82 * recursive call. For handling this situation and loops co->root is used
83 * to remember all roots.
85 static void co_append_unit(copy_opt_t *co, ir_node *root) {
88 struct list_head *tmp;
90 DBG((dbg, LEVEL_1, "\t Root: %n %N\n", root, root));
91 /* check if we encountered this root earlier */
92 if (pset_find_ptr(co->roots, root))
94 pset_insert_ptr(co->roots, root);
96 assert(is_curr_reg_class(root) && "node is in wrong register class!");
99 arity = get_irn_arity(root);
100 unit = xcalloc(1, sizeof(*unit));
102 unit->node_count = 1;
103 unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
104 unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
105 unit->nodes[0] = root;
106 unit->complete_costs = 0;
108 INIT_LIST_HEAD(&unit->queue);
112 for (i=0; i<arity; ++i) {
113 ir_node *arg = get_irn_n(root, i);
114 assert(is_curr_reg_class(arg) && "Argument not in same register class.");
117 if (nodes_interfere(co->chordal_env, root, arg))
118 assert(0 && "root and arg interfere");
119 //TODO do not insert duplicate args
120 DBG((dbg, LEVEL_1, "\t Member: %n %N\n", arg, arg));
122 /* check if arg has occurred at a prior position in the arg/list */
123 for (o=0; o<unit->node_count; ++o)
124 if (unit->nodes[o] == arg) {
129 if (!arg_pos) { /* a new argument */
130 /* TODO Think about the next 2 lines. (inserting in arg-order) */
131 if (is_optimizable(co->chordal_env->arch_env, arg))
132 co_append_unit(co, arg);
133 /* insert node, set costs */
134 unit->nodes[unit->node_count] = arg;
135 unit->costs[unit->node_count] = co->get_costs(root, arg, i);
137 } else { /* arg has occured before in same phi */
138 /* increase costs for existing arg */
139 unit->costs[arg_pos] = co->get_costs(root, arg, i);
143 unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
144 unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
145 } else if (is_Copy(co->chordal_env->arch_env, root)) {
146 assert(!nodes_interfere(co->chordal_env, root, get_Copy_src(root)));
147 unit->nodes[1] = get_Copy_src(root);
148 unit->costs[1] = co->get_costs(root, unit->nodes[1], -1);
149 unit->node_count = 2;
150 unit->nodes = xrealloc(unit->nodes, 2 * sizeof(*unit->nodes));
151 unit->costs = xrealloc(unit->costs, 2 * sizeof(*unit->costs));
153 assert(0 && "This is not an optimizable node!");
154 /* TODO add ou's for 2-addr-code instructions */
157 for(i=1; i<unit->node_count; ++i)
158 unit->complete_costs += unit->costs[i];
160 assert(unit->node_count > 1);
161 unit->avg_costs = (100 * unit->complete_costs) / (unit->node_count-1);
163 /* insert according to average costs */
165 while (tmp->next != &co->units && list_entry_units(tmp->next)->avg_costs > unit->avg_costs)
167 list_add(&unit->units, tmp);
169 /* Init ifg_mis_size to node_count. So get_lower_bound returns correct results. */
170 unit->minimal_costs = unit->complete_costs - ou_max_ind_set_costs(unit);
173 static void co_collect_in_block(ir_node *block, void *env) {
174 copy_opt_t *co = env;
175 struct list_head *head = get_block_border_head(co->chordal_env, block);
178 list_for_each_entry_reverse(border_t, curr, head, list)
179 if (curr->is_def && curr->is_real && is_optimizable(co->chordal_env->arch_env, curr->irn))
180 co_append_unit(co, curr->irn);
183 static void co_collect_units(copy_opt_t *co) {
184 DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
185 co->roots = pset_new_ptr(64);
186 dom_tree_walk_irg(co->chordal_env->irg, co_collect_in_block, NULL, co);
190 copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)) {
191 const char *s1, *s2, *s3;
195 dbg = firm_dbg_register("ir.be.copyopt");
196 firm_dbg_set_mask(dbg, DEBUG_LVL);
198 co = xcalloc(1, sizeof(*co));
199 co->chordal_env = chordal_env;
200 co->get_costs = get_costs;
202 s1 = get_irp_prog_name();
203 s2 = get_entity_name(get_irg_entity(co->chordal_env->irg));
204 s3 = chordal_env->cls->name;
205 len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
206 co->name = xmalloc(len);
207 snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
208 if (!strcmp(co->name, DEBUG_IRG))
209 firm_dbg_set_mask(dbg, DEBUG_LVL_CO);
211 firm_dbg_set_mask(dbg, DEBUG_LVL);
213 INIT_LIST_HEAD(&co->units);
214 co_collect_units(co);
218 void free_copy_opt(copy_opt_t *co) {
221 list_for_each_entry_safe(unit_t, curr, tmp, &co->units, units) {
227 int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
229 for(i=0, max=get_irn_n_outs(irn); i<max; ++i) {
230 ir_node *n = get_irn_out(irn, i);
231 if ((is_Phi(n) || is_Perm(co->chordal_env->arch_env, n)) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
237 int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos) {
240 ir_node *root_block = get_nodes_block(root);
242 assert(pos==-1 || is_Phi(root));
244 /* a perm places the copy in the same block as it resides */
245 loop = get_irn_loop(root_block);
247 /* for phis the copies are placed in the corresponding pred-block */
248 loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
251 cost = 2*get_loop_depth(loop);
255 int get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
259 int co_get_copy_costs(const copy_opt_t *co) {
262 list_for_each_entry(unit_t, curr, &co->units, units) {
263 int root_col = get_irn_col(co, curr->nodes[0]);
264 for (i=1; i<curr->node_count; ++i)
265 if (root_col != get_irn_col(co, curr->nodes[i])) {
266 DBG((dbg, LEVEL_1, " %n %N\n", curr->nodes[i], curr->nodes[i]));
267 res += curr->costs[i];
273 int co_get_lower_bound(const copy_opt_t *co) {
276 list_for_each_entry(unit_t, curr, &co->units, units)
277 res += curr->minimal_costs;