static firm_dbg_module_t *dbg = NULL;
-#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)
+#define is_curr_reg_class(irn) \
+ (arch_get_irn_reg_class(get_arch_env(co), \
+ irn, arch_pos_make_out(0)) == co->chordal_env->cls)
#define MIN(a,b) ((a<b)?(a):(b))
#define MAX(a,b) ((a<b)?(b):(a))
+
/**
- * Computes the weight of a 'max independent set' wrt. ifg-edges only
- * (no coloring conflicts, no register constraints)
- * @return The costs of such a mis
- * NOTE: Code adapted from becopyheur
- * BETTER: Here we can be sure having a chordal graph to work on,
- * so, for 'larger' opt-units we could use a special algorithm.
+ * Determines a maximum weighted independent set with respect to
+ * the interference and conflict edges of all nodes in a qnode.
+ * TODO: This runs in n! in worst case. Use a heuristic iff n>???
*/
static int ou_max_ind_set_costs(unit_t *ou) {
- ir_node **irns;
- int max, pos, curr_weight, best_weight = 0;
+ be_chordal_env_t *chordal_env = ou->co->chordal_env;
+ ir_node **safe, **unsafe;
+ int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
bitset_t *curr;
+ int max, pos, curr_weight, best_weight = 0;
+
+ /* assign the nodes into two groups.
+ * safe: node has no interference, hence it is in every max stable set.
+ * unsafe: node has an interference
+ */
+ safe = alloca((ou->node_count-1) * sizeof(*safe));
+ safe_costs = 0;
+ safe_count = 0;
+ unsafe = alloca((ou->node_count-1) * sizeof(*unsafe));
+ unsafe_costs = alloca((ou->node_count-1) * sizeof(*unsafe_costs));
+ unsafe_count = 0;
+ for(i=1; i<ou->node_count; ++i) {
+ int is_safe = 1;
+ for(o=1; o<ou->node_count; ++o) {
+ if (i==o)
+ continue;
+ if (nodes_interfere(chordal_env, ou->nodes[i], ou->nodes[o])) {
+ unsafe_costs[unsafe_count] = ou->costs[i];
+ unsafe[unsafe_count] = ou->nodes[i];
+ ++unsafe_count;
+ is_safe = 0;
+ break;
+ }
+ }
+ if (is_safe) {
+ safe_costs += ou->costs[i];
+ safe[safe_count++] = ou->nodes[i];
+ }
+ }
- irns = alloca((ou->node_count-1) * sizeof(*irns));
- curr = bitset_alloca(ou->node_count-1);
- /* brute force the best set */
+
+ /* now brute force the best set out of the unsafe nodes*/
+ curr = bitset_alloca(unsafe_count);
+
bitset_set_all(curr);
while ((max = bitset_popcnt(curr)) != 0) {
/* check if curr is a stable set */
- int i, o, is_stable_set = 1;
-
- /* copy the irns */
- i = 0;
+ for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
+ 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) */
+ if (nodes_interfere(chordal_env, unsafe[i], unsafe[o]))
+ goto no_stable_set;
+
+ /* if we arrive here, we have a stable set */
+ /* compute the weigth of the stable set*/
+ curr_weight = 0;
bitset_foreach(curr, pos)
- irns[i++] = ou->nodes[1+pos];
- assert(i==max);
+ curr_weight += unsafe_costs[pos];
- for(i=0; i<max; ++i)
- for(o=i+1; o<max; ++o) /* !!!!! difference to qnode_max_ind_set(): NOT o=i */
- if (nodes_interfere(ou->co->chordal_env, irns[i], irns[o])) {
- is_stable_set = 0;
- break;
- }
-
- if (is_stable_set) {
- /* calc current weigth */
- curr_weight = 0;
- bitset_foreach(curr, pos)
- curr_weight += ou->costs[1+pos];
-
- /* any better ? */
- if (curr_weight > best_weight)
- best_weight = curr_weight;
+ /* any better ? */
+ if (curr_weight > best_weight) {
+ best_weight = curr_weight;
}
+no_stable_set:
bitset_minus1(curr);
}
- return best_weight;
+
+ return safe_costs+best_weight;
}
/**
* to remember all roots.
*/
static void co_append_unit(copy_opt_t *co, ir_node *root) {
- int i, arity, inevitable_costs = 0;
+ int i, arity;
unit_t *unit;
struct list_head *tmp;
INIT_LIST_HEAD(&unit->queue);
/* check all args */
- if (is_Phi(root)) {
+ if (is_Phi(root) && mode_is_datab(get_irn_mode(root))) {
for (i=0; i<arity; ++i) {
int o, arg_pos = 0;
ir_node *arg = get_irn_n(root, i);
if (arg == root)
continue;
if (nodes_interfere(co->chordal_env, root, arg)) {
- inevitable_costs += co->get_costs(root, arg, i);
+ unit->inevitable_costs += co->get_costs(root, arg, i);
continue;
}
}
unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
unit->costs = xrealloc(unit->costs, unit->node_count * sizeof(*unit->costs));
- } else if (is_Copy(co->chordal_env->arch_env, root)) {
+ } else if (is_Copy(get_arch_env(co), root)) {
assert(!nodes_interfere(co->chordal_env, root, get_Copy_src(root)));
unit->nodes[1] = get_Copy_src(root);
unit->costs[1] = co->get_costs(root, unit->nodes[1], -1);
/* TODO add ou's for 2-addr-code instructions */
- /* Init costs with inevitable_costs */
- unit->all_nodes_costs = inevitable_costs;
- unit->min_nodes_costs = inevitable_costs;
-
/* Determine the maximum costs this unit can cause: all_nodes_cost */
for(i=1; i<unit->node_count; ++i) {
unit->sort_key = MAX(unit->sort_key, unit->costs[i]);
border_t *curr;
list_for_each_entry_reverse(border_t, curr, head, list)
- if (curr->is_def && curr->is_real && is_optimizable(co->chordal_env->arch_env, curr->irn))
+ if (curr->is_def && curr->is_real && is_optimizable(get_arch_env(co), curr->irn))
co_append_unit(co, curr->irn);
}
static void co_collect_units(copy_opt_t *co) {
DBG((dbg, LEVEL_1, "\tCollecting optimization units\n"));
co->roots = pset_new_ptr(64);
- dom_tree_walk_irg(co->chordal_env->irg, co_collect_in_block, NULL, co);
+ dom_tree_walk_irg(get_irg(co), co_collect_in_block, NULL, co);
del_pset(co->roots);
}
co->get_costs = get_costs;
s1 = get_irp_prog_name();
- s2 = get_entity_name(get_irg_entity(co->chordal_env->irg));
+ s2 = get_entity_name(get_irg_entity(get_irg(co)));
s3 = chordal_env->cls->name;
len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
co->name = xmalloc(len);
int i, max;
for(i=0, max=get_irn_n_outs(irn); i<max; ++i) {
ir_node *n = get_irn_out(irn, i);
- if ((is_Phi(n) || is_Perm(co->chordal_env->arch_env, n)) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
+ if (((is_Phi(n) && mode_is_datab(get_irn_mode(n))) ||
+ is_Perm(get_arch_env(co), n)) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
return 1;
}
return 0;
/* for phis the copies are placed in the corresponding pred-block */
loop = get_irn_loop(get_Block_cfgpred_block(root_block, pos));
}
- if (loop)
- cost = 2*get_loop_depth(loop);
+ if (loop) {
+ int d = get_loop_depth(loop);
+ cost = d*d;
+ }
return cost+1;
}
return 1;
}
+int co_get_max_copy_costs(const copy_opt_t *co) {
+ int i, res = 0;
+ unit_t *curr;
+
+ list_for_each_entry(unit_t, curr, &co->units, units) {
+ res += curr->inevitable_costs;
+ for (i=1; i<curr->node_count; ++i)
+ res += curr->costs[i];
+ }
+ return res;
+}
+
+int co_get_inevit_copy_costs(const copy_opt_t *co) {
+ int res = 0;
+ unit_t *curr;
+
+ list_for_each_entry(unit_t, curr, &co->units, units)
+ res += curr->inevitable_costs;
+ return res;
+}
+
int co_get_copy_costs(const copy_opt_t *co) {
int i, res = 0;
unit_t *curr;
+
list_for_each_entry(unit_t, curr, &co->units, units) {
int root_col = get_irn_col(co, curr->nodes[0]);
- DBG((dbg, LEVEL_1, " Adding costs for root %+F color %d\n", curr->nodes[0], root_col));
+ DBG((dbg, LEVEL_1, " %3d costs for root %+F color %d\n", curr->inevitable_costs, curr->nodes[0], root_col));
+ res += curr->inevitable_costs;
for (i=1; i<curr->node_count; ++i) {
int arg_col = get_irn_col(co, curr->nodes[i]);
if (root_col != arg_col) {
- DBG((dbg, LEVEL_1, " Arg %+F color %d costs %d\n", curr->nodes[i], arg_col, curr->costs[i]));
+ DBG((dbg, LEVEL_1, " %3d for arg %+F color %d\n", curr->costs[i], curr->nodes[i], arg_col));
res += curr->costs[i];
}
}
int res = 0;
unit_t *curr;
list_for_each_entry(unit_t, curr, &co->units, units)
- res += curr->min_nodes_costs;
+ res += curr->inevitable_costs + curr->min_nodes_costs;
return res;
}