X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur.c;h=e8bafa6b05b88e10994f288390778a0aacef23c7;hb=e4766594e9640777c92da56f37dbdce7b41385a0;hp=34e2e83def76e9980207fcc799596d2dd9d9fe70;hpb=757e9691e224edc699466891995898e806ca4b25;p=libfirm

diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c
index 34e2e83de..e8bafa6b0 100644
--- a/ir/be/becopyheur.c
+++ b/ir/be/becopyheur.c
@@ -28,7 +28,6 @@
 #include "becopystat.h"
 #include "bitset.h"
 
-#define DEBUG_LVL 0 //SET_LEVEL_1
 static firm_dbg_module_t *dbg = NULL;
 
 #define SLOTS_PINNED_GLOBAL 256
@@ -70,7 +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;
@@ -84,7 +83,7 @@ static int set_cmp_conflict_t(const void *x, const void *y, size_t size) {
  */
 static INLINE void qnode_add_conflict(const qnode_t *qn, const ir_node *n1, const ir_node *n2) {
 	conflict_t c;
-	DBG((dbg, LEVEL_4, "\t      %n -- %n\n", n1, n2));
+	DBG((dbg, LEVEL_4, "\t      %+F -- %+F\n", n1, n2));
 
 	if ((int)n1 < (int)n2) {
 		c.n1 = n1;
@@ -157,6 +156,7 @@ static INLINE int qnode_get_new_color(const qnode_t *qn, ir_node *irn) {
 static INLINE void qnode_set_new_color(const qnode_t *qn, ir_node *irn, int color) {
 	node_stat_t *found = qnode_find_or_insert_node(qn, irn);
 	found->new_color = color;
+	DBG((dbg, LEVEL_3, "\t      col(%+F) := %d\n", irn, color));
 }
 
 /**
@@ -179,6 +179,8 @@ static INLINE int qnode_is_pinned_local(const qnode_t *qn, ir_node *irn) {
 static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
 	node_stat_t *found = qnode_find_or_insert_node(qn, irn);
 	found->pinned_local = 1;
+	if (found->new_color == NO_COLOR)
+		found->new_color = get_irn_col(qn->ou->co, irn);
 }
 
 /**
@@ -204,126 +206,56 @@ static INLINE void qnode_pin_local(const qnode_t *qn, ir_node *irn) {
  *             This should be true in SSA.
  */
 static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const ir_node *trigger) {
-	ir_node *res;
-	struct obstack confl_ob;
-	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_register_class_t *cls = chordal_env->cls;
+	const arch_env_t *arch_env = chordal_env->main_env->arch_env;
+	int irn_col = qnode_get_new_color(qn, irn);
 
-	DBG((dbg, LEVEL_3, "\t      %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
-	obstack_init(&confl_ob);
-	irn_col = qnode_get_new_color(qn, irn);
+	DBG((dbg, LEVEL_3, "\t    %+F \tcaused col(%+F) \t%2d --> %2d\n", trigger, irn, irn_col, col));
 
-	if (irn_col == col)
-		goto ret_save;
+	if (irn_col == col) {
+		DBG((dbg, LEVEL_3, "\t      %+F same color\n", irn));
+		return CHANGE_SAVE;
+	}
 	if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
-		res = irn;
-		goto ret_confl;
+		DBG((dbg, LEVEL_3, "\t      %+F conflicting\n", irn));
+		return irn;
+	}
+	if (!arch_reg_is_allocatable(arch_env, irn, -1, arch_register_for_index(cls, col))) {
+		DBG((dbg, LEVEL_3, "\t      %+F impossible\n", irn));
+		return CHANGE_IMPOSSIBLE;
 	}
-	if (!arch_reg_is_allocatable(arch_env,
-								 irn,
-								 arch_pos_make_out(0),
-								 arch_register_for_index(cls, col)))
-		goto ret_imposs;
 
-	/* get all nodes which would conflict with this change */
+	/*
+	 * Process all nodes which would conflict with this change
+	 */
 	{
-		struct obstack q;
-		int in, out;
-		ir_node *irn_bl;
-
-		irn_bl = get_nodes_block(irn);
-
-		/* first check for a conflicting node which is 'living in' the irns block */
-		{
-			ir_node *n;
-			pset *live_ins = put_live_in(irn_bl, pset_new_ptr_default());
-			for (n = pset_first(live_ins); n; n = pset_next(live_ins))
-				if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
-            && n != trigger && qnode_get_new_color(qn, n) == col
-            && nodes_interfere(chordal_env, irn, n)) {
-
-					DBG((dbg, LEVEL_4, "\t        %n\ttroubles\n", n));
-					obstack_ptr_grow(&confl_ob, n);
-					pset_break(live_ins);
-					break;
-			}
-            del_pset(live_ins);
-		}
-
-		/* setup the queue of blocks. */
-		obstack_init(&q);
-		obstack_ptr_grow(&q, irn_bl);
-		in = 1;
-		out = 0;
-
-		/* process the queue. The code below checks for every block dominated
-		 * by the irns one, and in which the irn is live, if there are
-		 * conflicting nodes */
-		while (out < in) {
-			ir_node *curr_bl, *sub_bl;
-			int i, max;
-
-			curr_bl = ((ir_node **)obstack_base(&q))[out++];
-
-			/* Add to the result all nodes in the block, which have
-			 * the target color and interfere with the irn */
-			for (i = 0, max = get_irn_n_outs(curr_bl); i < max; ++i) {
-				ir_node *n = get_irn_out(curr_bl, i);
-				if (arch_irn_has_reg_class(arch_env, n, arch_pos_make_out(0), cls)
-            && n != trigger && qnode_get_new_color(qn, n) == col
-            && nodes_interfere(chordal_env, irn, n)) {
-
-					DBG((dbg, LEVEL_4, "\t        %n\ttroubles\n", n));
-					obstack_ptr_grow(&confl_ob, n);
+		be_ifg_t *ifg = chordal_env->ifg;
+		void *iter = be_ifg_neighbours_iter_alloca(ifg);
+		ir_node *sub_res, *curr;
+
+		/*
+		 * Try to color all conflicting nodes 'curr'
+		 * with the color of the irn itself.
+		 */
+		be_ifg_foreach_neighbour(ifg, iter, irn, curr) {
+			DBG((dbg, LEVEL_3, "\t      Confl %+F(%d)\n", curr, qnode_get_new_color(qn, curr)));
+			if (qnode_get_new_color(qn, curr) == col && curr != trigger) {
+				sub_res = qnode_color_irn(qn, curr, irn_col, irn);
+				if (sub_res != CHANGE_SAVE) {
+					be_ifg_neighbours_break(ifg, iter);
+					return sub_res;
 				}
 			}
-
-			/* If irn lives out check i-dominated blocks where the irn lives in */
-			/* Fill the queue */
-			if (is_live_out(curr_bl, irn)) {
-				dominates_for_each(curr_bl, sub_bl)
-					if (is_live_in(sub_bl, irn)) {
-						obstack_ptr_grow(&q, sub_bl);
-						in++;
-					}
-			}
 		}
-		obstack_free(&q, NULL);
-		obstack_ptr_grow(&confl_ob, NULL);
-		confl = (ir_node **) obstack_finish(&confl_ob);
 	}
 
-	/* process all nodes which would conflict with this change */
-	for (i = 0, cn = confl[0]; cn; cn = confl[++i]) {
-		ir_node *sub_res;
-
-		/* try to color the conflicting node cn with the color of the irn itself */
-		sub_res = qnode_color_irn(qn, cn, irn_col, irn);
-		if (sub_res != CHANGE_SAVE) {
-			res = sub_res;
-			goto ret_confl;
-		}
-	}
-	/* if we arrive here all sub changes can be applied, so it's save to change this irn */
-
-ret_save:
-	DBG((dbg, LEVEL_3, "\t      %n save\n", irn));
-	obstack_free(&confl_ob, NULL);
+	/*
+	 * If we arrive here, all sub changes have been applied.
+	 * So it's save to change this irn
+	 */
 	qnode_set_new_color(qn, irn, col);
 	return CHANGE_SAVE;
-
-ret_imposs:
-	DBG((dbg, LEVEL_3, "\t      %n impossible\n", irn));
-	obstack_free(&confl_ob, NULL);
-	return CHANGE_IMPOSSIBLE;
-
-ret_confl:
-	DBG((dbg, LEVEL_3, "\t      %n conflicting\n", irn));
-	obstack_free(&confl_ob, NULL);
-	return res;
 }
 
 /**
@@ -338,7 +270,7 @@ static int qnode_try_color(const qnode_t *qn) {
 		ir_node *test_node, *confl_node;
 
 		test_node = qn->mis[i];
-		DBG((dbg, LEVEL_3, "\t    Testing %n\n", test_node));
+		DBG((dbg, LEVEL_3, "\t    Testing %+F\n", test_node));
 		confl_node = qnode_color_irn(qn, test_node, qn->color, test_node);
 
 		if (confl_node == CHANGE_SAVE) {
@@ -379,55 +311,96 @@ static int qnode_try_color(const qnode_t *qn) {
  * 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) {
+		/* Heuristic: 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);
-	qn->mis_costs = best_weight;
-	next = 0;
+	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];
 }
 
 /**
@@ -488,16 +461,16 @@ 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"));
 	for (i=0; i<ou->node_count; ++i)
-		DBG((dbg, LEVEL_1, "\t %n\n", ou->nodes[i]));
+		DBG((dbg, LEVEL_1, "\t %+F\n", ou->nodes[i]));
 
 	/* 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));
 
@@ -511,6 +484,7 @@ static void ou_optimize(unit_t *ou) {
 		/* try */
 		if (qnode_try_color(curr))
 			break;
+
 		/* no success, so re-insert */
 		del_set(curr->changed_nodes);
 		curr->changed_nodes = new_set(set_cmp_node_stat_t, SLOTS_CHANGED_NODES);
@@ -520,14 +494,14 @@ 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));
-		/* globally pin root and eventually others */
+		int root_col = qnode_get_new_color(curr, ou->nodes[0]);
+		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) {
 			ir_node *irn = ou->nodes[i];
 			int nc = qnode_get_new_color(curr, irn);
-			if (nc != NO_COLOR && nc == qnode_get_new_color(curr, ou->nodes[0]))
+			if (nc != NO_COLOR && nc == root_col)
 				pset_insert_ptr(pinned_global, irn);
 		}
 
@@ -535,7 +509,7 @@ static void ou_optimize(unit_t *ou) {
 		for (ns = set_first(curr->changed_nodes); ns; ns = set_next(curr->changed_nodes)) {
 			/* NO_COLOR is possible, if we had an undo */
 			if (ns->new_color != NO_COLOR) {
-				DBG((dbg, LEVEL_2, "\t    color(%n) := %d\n", ns->irn, ns->new_color));
+				DBG((dbg, LEVEL_1, "\t    color(%+F) := %d\n", ns->irn, ns->new_color));
 				set_irn_col(ou->co, ns->irn, ns->new_color);
 			}
 		}
@@ -550,11 +524,6 @@ static void ou_optimize(unit_t *ou) {
 void co_heur_opt(copy_opt_t *co) {
 	unit_t *curr;
 	dbg = firm_dbg_register("ir.be.copyoptheur");
-	firm_dbg_set_mask(dbg, DEBUG_LVL);
-	if (!strcmp(co->name, DEBUG_IRG))
-		firm_dbg_set_mask(dbg, DEBUG_LVL_HEUR);
-	else
-		firm_dbg_set_mask(dbg, DEBUG_LVL);
 
 	pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
 	list_for_each_entry(unit_t, curr, &co->units, units)