#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
+#ifdef HAVE_ALLOCA_H
+#include <alloca.h>
+#endif
+#ifdef HAVE_MALLOC_H
+#include <malloc.h>
+#endif
+
+#include "irprog.h"
+#include "irloop_t.h"
#include "xmalloc.h"
+#include "bechordal_t.h"
#include "becopyopt.h"
#include "becopystat.h"
#define DEBUG_LVL 0 //SET_LEVEL_1
static firm_dbg_module_t *dbg = NULL;
-#define is_curr_reg_class(irn) (arch_get_irn_reg_class(co->env, irn, arch_pos_make_out(0)) == co->cls)
+#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 MIN(a,b) ((a<b)?(a):(b))
+
+/**
+ * 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.
+ */
+static int ou_max_ind_set_costs(unit_t *ou) {
+ ir_node **irns;
+ int max, pos, curr_weight, best_weight = 0;
+ bitset_t *curr;
+
+ irns = alloca((ou->node_count-1) * sizeof(*irns));
+ curr = bitset_alloca(ou->node_count-1);
+
+ /* brute force the best set */
+ bitset_set_all(curr);
+ while ((max = bitset_popcnt(curr)) != 0) {
+ /* check if curr is a stable set */
+ int i, o, is_stable_set = 1;
+ bitset_foreach(curr, pos)
+ irns[pos] = ou->nodes[1+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;
+ }
+
+ bitset_minus1(curr);
+ }
+ return best_weight;
+}
/**
* Builds an optimization unit for a given optimizable irn (root).
static void co_append_unit(copy_opt_t *co, ir_node *root) {
int i, arity;
unit_t *unit;
- DBG((dbg, LEVEL_1, "\t Root: %n\n", root));
+ struct list_head *tmp;
+
+ DBG((dbg, LEVEL_1, "\t Root: %n %N\n", root, root));
/* check if we encountered this root earlier */
if (pset_find_ptr(co->roots, root))
return;
arity = get_irn_arity(root);
unit = xcalloc(1, sizeof(*unit));
unit->co = co;
- unit->interf = 0;
unit->node_count = 1;
unit->nodes = xmalloc((arity+1) * sizeof(*unit->nodes));
+ unit->costs = xmalloc((arity+1) * sizeof(*unit->costs));
unit->nodes[0] = root;
+ unit->complete_costs = 0;
+ unit->avg_costs = 0;
INIT_LIST_HEAD(&unit->queue);
/* check all args */
- for (i=0; i<arity; ++i) {
- ir_node *arg = get_irn_n(root, i);
- assert(is_curr_reg_class(arg) && "Argument not in same register class.");
- if (arg != root) {
- if (!values_interfere(root, arg)) {
- DBG((dbg, LEVEL_1, "\t Member: %n\n", arg));
- if (is_optimizable(arg))
- co_append_unit(co, arg);
- unit->nodes[unit->node_count++] = arg;
- } else
- unit->interf++;
+ if (is_Phi(root)) {
+ for (i=0; i<arity; ++i) {
+ ir_node *arg = get_irn_n(root, i);
+ assert(is_curr_reg_class(arg) && "Argument not in same register class.");
+ if (arg != root) {
+ int o, arg_pos = 0;
+ if (nodes_interfere(co->chordal_env, root, arg))
+ assert(0 && "root and arg interfere");
+ //TODO do not insert duplicate args
+ DBG((dbg, LEVEL_1, "\t Member: %n %N\n", arg, arg));
+
+ /* check if arg has occurred at a prior position in the arg/list */
+ for (o=0; o<unit->node_count; ++o)
+ if (unit->nodes[o] == arg) {
+ arg_pos = o;
+ break;
+ }
+
+ if (!arg_pos) { /* a new argument */
+ /* TODO Think about the next 2 lines. (inserting in arg-order) */
+ if (is_optimizable(co->chordal_env->arch_env, arg))
+ co_append_unit(co, arg);
+ /* insert node, set costs */
+ unit->nodes[unit->node_count] = arg;
+ unit->costs[unit->node_count] = co->get_costs(root, arg, i);
+ unit->node_count++;
+ } else { /* arg has occured before in same phi */
+ /* increase costs for existing arg */
+ unit->costs[arg_pos] = co->get_costs(root, arg, i);
+ }
+ }
}
- }
- unit->nodes = xrealloc(unit->nodes, unit->node_count * sizeof(*unit->nodes));
- list_add_tail(&unit->units, &co->units);
- /* Init mis_size to node_count. So get_lower_bound returns correct results.
- * - Now it can be called even before the heuristic has run.
- * - And it will return correct results for units with nodecount 1 which are
- * not optimized during the heuristic and have therefor delivered wrong results for get_lower_bound
- */
- unit->mis_size = unit->node_count;
+ 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)) {
+ 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);
+ unit->node_count = 2;
+ unit->nodes = xrealloc(unit->nodes, 2 * sizeof(*unit->nodes));
+ unit->costs = xrealloc(unit->costs, 2 * sizeof(*unit->costs));
+ } else
+ assert(0 && "This is not an optimizable node!");
+ /* TODO add ou's for 2-addr-code instructions */
+
+ for(i=1; i<unit->node_count; ++i)
+ unit->complete_costs += unit->costs[i];
+
+ assert(unit->node_count > 1);
+ unit->avg_costs = (100 * unit->complete_costs) / (unit->node_count-1);
+
+ /* insert according to average costs */
+ tmp = &co->units;
+ while (tmp->next != &co->units && list_entry_units(tmp->next)->avg_costs > unit->avg_costs)
+ tmp = tmp->next;
+ list_add(&unit->units, tmp);
+
+ /* Init ifg_mis_size to node_count. So get_lower_bound returns correct results. */
+ unit->minimal_costs = unit->complete_costs - ou_max_ind_set_costs(unit);
}
static void co_collect_in_block(ir_node *block, void *env) {
copy_opt_t *co = env;
- struct list_head *head = &get_ra_block_info(block)->border_head;
+ struct list_head *head = get_block_border_head(co->chordal_env, block);
border_t *curr;
list_for_each_entry_reverse(border_t, curr, head, list)
- if (curr->is_def && curr->is_real && is_optimizable(curr->irn))
+ if (curr->is_def && curr->is_real && is_optimizable(co->chordal_env->arch_env, 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->irg, co_collect_in_block, NULL, co);
+ dom_tree_walk_irg(co->chordal_env->irg, co_collect_in_block, NULL, co);
del_pset(co->roots);
}
-copy_opt_t *new_copy_opt(ir_graph *irg, const arch_env_t *env, const arch_register_class_t *cls) {
+copy_opt_t *new_copy_opt(be_chordal_env_t *chordal_env, int (*get_costs)(ir_node*, ir_node*, int)) {
const char *s1, *s2, *s3;
int len;
- copy_opt_t *co;
+ copy_opt_t *co;
dbg = firm_dbg_register("ir.be.copyopt");
firm_dbg_set_mask(dbg, DEBUG_LVL);
co = xcalloc(1, sizeof(*co));
- co->irg = irg;
- co->env = env;
-// co->isa = env->isa;
- co->cls = cls;
+ co->chordal_env = chordal_env;
+ co->get_costs = get_costs;
s1 = get_irp_prog_name();
- s2 = get_entity_name(get_irg_entity(co->irg));
- s3 = cls->name;
+ s2 = get_entity_name(get_irg_entity(co->chordal_env->irg));
+ s3 = chordal_env->cls->name;
len = strlen(s1) + strlen(s2) + strlen(s3) + 5;
co->name = xmalloc(len);
- if (!strcmp(co->name, DEBUG_IRG))
- firm_dbg_set_mask(dbg, -1);
snprintf(co->name, len, "%s__%s__%s", s1, s2, s3);
+ if (!strcmp(co->name, DEBUG_IRG))
+ firm_dbg_set_mask(dbg, DEBUG_LVL_CO);
+ else
+ firm_dbg_set_mask(dbg, DEBUG_LVL);
INIT_LIST_HEAD(&co->units);
co_collect_units(co);
}
}
-int is_optimizable_arg(ir_node *irn) {
+int is_optimizable_arg(const copy_opt_t *co, ir_node *irn) {
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_optimizable(n) && (irn == n || !values_interfere(irn, n)))
+ if ((is_Phi(n) || is_Perm(co->chordal_env->arch_env, n)) && (irn == n || !nodes_interfere(co->chordal_env, irn, n)))
return 1;
}
return 0;
}
+int get_costs_loop_depth(ir_node *root, ir_node* arg, int pos) {
+ int cost = 0;
+ ir_loop *loop;
+ ir_node *root_block = get_nodes_block(root);
-int co_get_copy_count(copy_opt_t *co) {
+ assert(pos==-1 || is_Phi(root));
+ if (pos == -1) {
+ /* a perm places the copy in the same block as it resides */
+ loop = get_irn_loop(root_block);
+ } else {
+ /* 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);
+ return cost+1;
+}
+
+int get_costs_all_one(ir_node *root, ir_node* arg, int pos) {
+ return 1;
+}
+
+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]);
- res += curr->interf;
for (i=1; i<curr->node_count; ++i)
- if (root_col != get_irn_col(co, curr->nodes[i]))
- res++;
+ if (root_col != get_irn_col(co, curr->nodes[i])) {
+ DBG((dbg, LEVEL_1, " %n %N\n", curr->nodes[i], curr->nodes[i]));
+ res += curr->costs[i];
+ }
}
return res;
}
-int co_get_lower_bound(copy_opt_t *co) {
+int co_get_lower_bound(const copy_opt_t *co) {
int res = 0;
unit_t *curr;
list_for_each_entry(unit_t, curr, &co->units, units)
- res += curr->interf + curr->node_count - curr->mis_size;
+ res += curr->minimal_costs;
return res;
}
-
-int co_get_interferer_count(copy_opt_t *co) {
- int res = 0;
- unit_t *curr;
- list_for_each_entry(unit_t, curr, &co->units, units)
- res += curr->interf;
- return res;
-}
-
-/**
- * Needed for result checking
- */
-static void co_collect_for_checker(ir_node *block, void *env) {
- copy_opt_t *co = env;
- struct list_head *head = &get_ra_block_info(block)->border_head;
- border_t *curr;
-
- list_for_each_entry_reverse(border_t, curr, head, list)
- if (curr->is_def && curr->is_real && is_curr_reg_class(curr->irn))
- obstack_ptr_grow(&co->ob, curr->irn);
-}
-
-/**
- * This O(n^2) checker checks, if two ifg-connected nodes have the same color.
- */
-void co_check_allocation(copy_opt_t *co) {
- ir_node **nodes, *n1, *n2;
- int i, o;
-
- obstack_init(&co->ob);
- dom_tree_walk_irg(co->irg, co_collect_for_checker, NULL, co);
- obstack_ptr_grow(&co->ob, NULL);
-
- nodes = (ir_node **) obstack_finish(&co->ob);
- for (i = 0, n1 = nodes[i]; n1; n1 = nodes[++i]) {
- assert(! (is_allocatable_irn(n1) && get_irn_col(co, n1) == NO_COLOR));
- for (o = i+1, n2 = nodes[o]; n2; n2 = nodes[++o])
- if (phi_ops_interfere(n1, n2) && get_irn_col(co, n1) == get_irn_col(co, n2)) {
- DBG((dbg, 0, "Error: %n in %n and %n in %n have the same color.\n", n1, get_nodes_block(n1), n2, get_nodes_block(n2)));
- assert(0 && "Interfering values have the same color!");
- }
- }
- obstack_free(&co->ob, NULL);
- DBG((dbg, 2, "The checker seems to be happy!\n"));
-}