Fixed a bug
[libfirm] / ir / be / becopyheur.c
index f215625..bf75e77 100644 (file)
@@ -69,7 +69,7 @@ typedef struct _qnode_t {
        set *changed_nodes;                     /**< contains node_stat_t's. */
 } qnode_t;
 
-pset *pinned_global;                   /**< optimized nodes should not be altered any more */
+static pset *pinned_global;                    /**< optimized nodes should not be altered any more */
 
 static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
        const conflict_t *xx = x;
@@ -208,7 +208,7 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
        ir_node **confl, *cn;
        int i, irn_col;
        const be_chordal_env_t *chordal_env = qn->ou->co->chordal_env;
-       const arch_env_t *arch_env = chordal_env->arch_env;
+       const arch_env_t *arch_env = get_arch_env(qn->ou->co);
        const arch_register_class_t *cls = chordal_env->cls;
 
        DBG((dbg, LEVEL_3, "\t      %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
@@ -225,7 +225,7 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
        }
        if (!arch_reg_is_allocatable(arch_env,
                                                                 irn,
-                                                                arch_pos_make_out(0),
+                                                                -1,
                                                                 arch_register_for_index(cls, col)))
                goto ret_imposs;
 
@@ -243,7 +243,7 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
                        pset *live_ins = put_live_in(irn_bl, pset_new_ptr_default());
                        for (n = pset_first(live_ins); n; n = pset_next(live_ins)) {
                                DBG((dbg, LEVEL_4, "Checking %+F which is live-in at the block\n", n));
-                               if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
+                               if (arch_irn_has_reg_class(arch_env, n, -1, cls)
                                        && n != trigger && qnode_get_new_color(qn, n) == col
                                        && nodes_interfere(chordal_env, irn, n)) {
 
@@ -276,7 +276,7 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
                        for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
                                ir_node *n = get_irn_out(curr_bl, i);
                                DBG((dbg, LEVEL_4, "Checking %+F defined in same block\n", n));
-                               if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
+                               if (arch_irn_has_reg_class(arch_env, n, -1, cls)
                                        && n != trigger && qnode_get_new_color(qn, n) == col
                                        && nodes_interfere(chordal_env, irn, n)) {
                                                DBG((dbg, LEVEL_4, "\t        %+F\ttroubles\n", n));
@@ -377,61 +377,116 @@ static int qnode_try_color(const qnode_t *qn) {
        return 1;
 }
 
+typedef int(*confl_f)(const ir_node *a, const ir_node *b, void *data);
+
+/**
+ * @param result       Gets filled with the computed maximum independent set.
+ * @param count                The size of input arrays / the number of nodes
+ * @param nodes                A set of nodes to copmute the max. ind. set for
+ * @param weights      Weights associated to the nodes in @p nodes
+ * @param confl                Callback function to decide if two values interfere
+ * @param data         Passed into all callbacks
+ * @return The size of the computed set
+ */
+//int max_ind_set(ir_node **result, int count, ir_node **nodes, int *weights, confl_f confl, void *data) {
+//
+//}
+
 /**
  * Determines a maximum weighted independent set with respect to
  * the interference and conflict edges of all nodes in a qnode.
  */
 static INLINE void qnode_max_ind_set(qnode_t *qn, const unit_t *ou) {
-       ir_node **irns;
+       ir_node **safe, **unsafe;
+       int i, o, safe_count, safe_costs, unsafe_count, *unsafe_costs;
+       bitset_t *curr, *best;
        int max, next, pos, curr_weight, best_weight = 0;
-       bitset_t *best, *curr;
-
-       irns = alloca((ou->node_count-1) * sizeof(*irns));
-       best = bitset_alloca(ou->node_count-1);
-       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;
-
-               /* copy the irns */
-               i = 0;
-               bitset_foreach(curr, pos)
-                       irns[i++] = ou->nodes[1+pos];
-               assert(i==max);
-
-               for(i=0; i<max; ++i)
-                       for(o=i; o<max; ++o) /* !!!!! difference to ou_max_ind_set_costs(): NOT o=i+1 */
-                               if (qnode_are_conflicting(qn, irns[i], irns[o])) {
-                                       is_stable_set = 0;
-                                       break;
+
+       /* 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 (qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[o])) {
+                               if (i!=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];
+               }
+       }
 
-               if (is_stable_set) {
-                       /* calc current weigth */
+
+
+       /* now compute the best set out of the unsafe nodes*/
+       best = bitset_alloca(unsafe_count);
+
+       if (unsafe_count > MIS_HEUR_TRIGGER) {
+               /* Heuristik: Greedy trial and error form index 0 to unsafe_count-1 */
+               for (i=0; i<unsafe_count; ++i) {
+                       bitset_set(best, i);
+                       /* check if it is a stable set */
+                       for (o=bitset_next_set(best, 0); o!=-1 && o<=i; o=bitset_next_set(best, o+1))
+                               if (qnode_are_conflicting(qn, unsafe[i], unsafe[o])) {
+                                       bitset_clear(best, i); /* clear the bit and try next one */
+                                       break;
+                               }
+               }
+               /* compute the weight */
+               bitset_foreach(best, pos)
+                       best_weight += unsafe_costs[pos];
+       } else {
+               /* Exact Algorithm: Brute force */
+               curr = bitset_alloca(unsafe_count);
+               bitset_set_all(curr);
+               while ((max = bitset_popcnt(curr)) != 0) {
+                       /* check if curr is a stable set */
+                       for (i=bitset_next_set(curr, 0); i!=-1; i=bitset_next_set(curr, i+1))
+                               for (o=bitset_next_set(curr, i); o!=-1; o=bitset_next_set(curr, o+1)) /* !!!!! difference to ou_max_ind_set_costs(): NOT (curr, i+1) */
+                                               if (qnode_are_conflicting(qn, 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)
-                               curr_weight += ou->costs[1+pos];
+                               curr_weight += unsafe_costs[pos];
 
                        /* any better ? */
                        if (curr_weight > best_weight) {
                                best_weight = curr_weight;
                                bitset_copy(best, curr);
                        }
-               }
 
-               bitset_minus1(curr);
+no_stable_set:
+                       bitset_minus1(curr);
+               }
        }
 
        /* transfer the best set into the qn */
-       qn->mis_size = bitset_popcnt(best)+1;
-       qn->mis_costs = best_weight;
-       next = 0;
-       qn->mis[next++] = ou->nodes[0]; /* the root is alwazs in a max stable set */
+       qn->mis_size = 1+safe_count+bitset_popcnt(best);
+       qn->mis_costs = safe_costs+best_weight;
+       qn->mis[0] = ou->nodes[0]; /* the root is always in a max stable set */
+       next = 1;
+       for (i=0; i<safe_count; ++i)
+               qn->mis[next++] = safe[i];
        bitset_foreach(best, pos)
-               qn->mis[next++] = ou->nodes[1+pos];
+               qn->mis[next++] = unsafe[pos];
 }
 
 /**
@@ -492,7 +547,7 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
  */
 static void ou_optimize(unit_t *ou) {
        int i;
-       qnode_t *curr, *tmp;
+       qnode_t *curr = NULL, *tmp;
        bitset_t *pos_regs = bitset_alloca(ou->co->chordal_env->cls->n_regs);
 
        DBG((dbg, LEVEL_1, "\tOptimizing unit:\n"));
@@ -501,7 +556,7 @@ static void ou_optimize(unit_t *ou) {
 
        /* init queue */
        INIT_LIST_HEAD(&ou->queue);
-       arch_get_allocatable_regs(ou->co->chordal_env->arch_env, ou->nodes[0], arch_pos_make_out(0), ou->co->chordal_env->cls, pos_regs);
+       arch_get_allocatable_regs(get_arch_env(ou->co), ou->nodes[0], -1, ou->co->chordal_env->cls, pos_regs);
        bitset_foreach(pos_regs, i)
                ou_insert_qnode(ou, new_qnode(ou, i));
 
@@ -524,7 +579,7 @@ static void ou_optimize(unit_t *ou) {
        /* apply the best found qnode */
        if (curr->mis_size >= 2) {
                node_stat_t *ns;
-               DBG((dbg, LEVEL_1, "\t  Best color: %d  Costs: %d/%d\n", curr->color, ou->complete_costs - curr->mis_costs, ou->complete_costs));
+               DBG((dbg, LEVEL_1, "\t  Best color: %d  Costs: %d << %d << %d\n", curr->color, ou->min_nodes_costs, ou->all_nodes_costs - curr->mis_costs, ou->all_nodes_costs));
                /* globally pin root and all args which have the same color */
                pset_insert_ptr(pinned_global, ou->nodes[0]);
                for (i=1; i<ou->node_count; ++i) {