X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyheur.c;h=bf75e77f5f8a28c361ce19e3a6d9b36e415349d7;hb=f48209ee1e8d81959bd3bd15f456afa4355f8303;hp=55092fea213dee9618b368d99b4605179e57bb94;hpb=9b24fe0ec0f4412c790ee4a7c6fc022fd28064a1;p=libfirm

diff --git a/ir/be/becopyheur.c b/ir/be/becopyheur.c
index 55092fea2..bf75e77f5 100644
--- a/ir/be/becopyheur.c
+++ b/ir/be/becopyheur.c
@@ -26,8 +26,8 @@
 #include "xmalloc.h"
 #include "becopyopt.h"
 #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
@@ -63,12 +63,13 @@ typedef struct _qnode_t {
 	const unit_t *ou;			/**< the opt unit this qnode belongs to */
 	int color;					/**< target color */
 	set *conflicts;				/**< contains conflict_t's. All internal conflicts */
-	int mis_size;				/**< number of nodes in the mis. */
+	int mis_costs;				/**< costs of nodes/copies in the mis. */
+	int mis_size;				/**< size of the array below */
 	ir_node **mis;				/**< the nodes of unit_t->nodes[] being part of the max independent set */
 	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;
@@ -82,7 +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;
@@ -207,22 +208,24 @@ 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      %n \tcaused col(%n) \t%2d --> %2d\n", trigger, irn, qnode_get_new_color(qn, irn), col));
+	DBG((dbg, LEVEL_3, "\t      %+F \tcaused col(%+F) \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);
 
-	if (irn_col == col)
+	if (irn_col == col) {
+		DBG((dbg, LEVEL_4, "\t      Already same color.\n"));
 		goto ret_save;
+	}
 	if (pset_find_ptr(pinned_global, irn) || qnode_is_pinned_local(qn, irn)) {
 		res = irn;
 		goto ret_confl;
 	}
 	if (!arch_reg_is_allocatable(arch_env,
 								 irn,
-								 arch_pos_make_out(0),
+								 -1,
 								 arch_register_for_index(cls, col)))
 		goto ret_imposs;
 
@@ -238,15 +241,17 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
 		{
 			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)) {
+			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, -1, 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));
+					DBG((dbg, LEVEL_4, "\t        %+F\ttroubles\n", n));
 					obstack_ptr_grow(&confl_ob, n);
 					pset_break(live_ins);
 					break;
+				}
 			}
             del_pset(live_ins);
 		}
@@ -270,12 +275,12 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
 			 * 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);
+				DBG((dbg, LEVEL_4, "Checking %+F defined in same block\n", n));
+				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));
+						obstack_ptr_grow(&confl_ob, n);
 				}
 			}
 
@@ -308,18 +313,18 @@ static ir_node *qnode_color_irn(const qnode_t *qn, ir_node *irn, int col, const
 	/* 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));
+	DBG((dbg, LEVEL_3, "\t      %+F save\n", irn));
 	obstack_free(&confl_ob, NULL);
 	qnode_set_new_color(qn, irn, col);
 	return CHANGE_SAVE;
 
 ret_imposs:
-	DBG((dbg, LEVEL_3, "\t      %n impossible\n", irn));
+	DBG((dbg, LEVEL_3, "\t      %+F impossible\n", irn));
 	obstack_free(&confl_ob, NULL);
 	return CHANGE_IMPOSSIBLE;
 
 ret_confl:
-	DBG((dbg, LEVEL_3, "\t      %n conflicting\n", irn));
+	DBG((dbg, LEVEL_3, "\t      %+F conflicting\n", irn));
 	obstack_free(&confl_ob, NULL);
 	return res;
 }
@@ -336,7 +341,7 @@ 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) {
@@ -348,8 +353,16 @@ static int qnode_try_color(const qnode_t *qn) {
 		} else {
 			if (qnode_is_pinned_local(qn, confl_node)) {
 				/* changing test_node would change back a node of current ou */
-				DBG((dbg, LEVEL_3, "\t    Conflicting local --> add conflict\n"));
-				qnode_add_conflict(qn, confl_node, test_node);
+				if (confl_node == qn->ou->nodes[0]) {
+					/* Adding a conflict edge between testnode and conflnode
+					 * would introduce a root -- arg interference.
+					 * So remove the arg of the qn */
+					DBG((dbg, LEVEL_3, "\t    Conflicting local with phi --> remove from qnode\n"));
+					qnode_add_conflict(qn, test_node, test_node);
+				} else {
+					DBG((dbg, LEVEL_3, "\t    Conflicting local --> add conflict\n"));
+					qnode_add_conflict(qn, confl_node, test_node);
+				}
 			}
 			if (pset_find_ptr(pinned_global, confl_node)) {
 				/* changing test_node would change back a node of a prior ou */
@@ -364,72 +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);
+
 /**
- * Determines a maximum independent set with respect to the interference and
- * conflict edges of all nodes in a qnode.
+ * @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) {
-	int all_size, curr_size, i, o;
-	int *which;
-	ir_node **curr, **all = alloca(ou->node_count * sizeof(*all));
+	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;
+
+	/* 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];
+		}
+	}
 
-	/* all contains all nodes not removed in this qn */
-	all_size = 0;
-	for (i=0; i<ou->node_count; ++i)
-		if (!qnode_are_conflicting(qn, ou->nodes[i], ou->nodes[i]))
-			all[all_size++] = ou->nodes[i];
-
-	/* which[i] says which element to take out of all[] and put into curr[i] */
-	which = alloca(all_size*sizeof(*which));
-	for (curr_size=0; curr_size<all_size; ++curr_size)
-		which[curr_size] = curr_size;
-
-	/* stores the currently examined set */
-	curr = alloca(all_size*sizeof(*curr));
-
-	while (1) { /* this loop will terminate because at least a single node will be a max indep. set */
-		/* build current set */
-		for (i=0; i<curr_size; ++i)
-			curr[i] = all[which[all_size-curr_size+i]];
-
-		/* check current set */
-		for (i=0; i<curr_size; ++i)
-			for (o=i+1; o<curr_size; ++o)
-				if (qnode_are_conflicting(qn, curr[i], curr[o]))
-					goto conflict_found;
-
-		/* We had no conflict. This is the max indep. set */
-		qn->mis_size = curr_size;
-		for (i=0; i<curr_size; ++i)
-			qn->mis[i] = curr[i];
-		return;
 
-conflict_found:
-		/* We had a conflict. Generate next set */
-		if (which[all_size-curr_size+1] == all_size-curr_size+1) {
-			curr_size--;
-			for (i=0; i<curr_size; ++i)
-				which[all_size-curr_size+i] = i;
-		} else {
-			int redo = 1;
-			while (redo) {
-				int pos = all_size;
-				do {
-					pos--;
-				} while (!(which[pos] = (which[pos]+1) % all_size));
-
-				for (i=pos+1; i<all_size; ++i)
-					which[i] = MIN(which[i-1]+1, all_size-1);
-
-				redo = 0;
-				for (i=all_size-curr_size; i<all_size-1; ++i)
-					if (which[i]>=which[i+1]) {
-						redo = 1;
-						break;
-					}
+
+	/* 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 += unsafe_costs[pos];
+
+			/* any better ? */
+			if (curr_weight > best_weight) {
+				best_weight = curr_weight;
+				bitset_copy(best, curr);
 			}
+
+no_stable_set:
+			bitset_minus1(curr);
 		}
 	}
+
+	/* transfer the best set into the qn */
+	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++] = unsafe[pos];
 }
 
 /**
@@ -470,11 +527,11 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
 
 	qnode_max_ind_set(qn, ou);
 	/* do the insertion */
-	DBG((dbg, LEVEL_4, "\t  Insert qnode color %d with size %d\n", qn->color, qn->mis_size));
+	DBG((dbg, LEVEL_4, "\t  Insert qnode color %d with cost %d\n", qn->color, qn->mis_costs));
 	lh = &ou->queue;
 	while (lh->next != &ou->queue) {
 		qnode_t *curr = list_entry_queue(lh->next);
-		if (curr->mis_size <= qn->mis_size)
+		if (curr->mis_costs <= qn->mis_costs)
 			break;
 		lh = lh->next;
 	}
@@ -482,24 +539,24 @@ static INLINE void ou_insert_qnode(unit_t *ou, qnode_t *qn) {
 }
 
 /**
- * Tries to re-allocate colors of nodes in this opt unit, to achieve a lower
- * number of copy instructions placed during SSA-destruction and lowering.
+ * Tries to re-allocate colors of nodes in this opt unit, to achieve lower
+ * costs of copy instructions placed during SSA-destruction and lowering.
  * Works only for opt units with exactly 1 root node, which is the
- * case for approximately 80% of all phi classes and all register constrained
+ * case for approximately 80% of all phi classes and 100% of register constrained
  * nodes. (All other phi classes are reduced to this case.)
  */
 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));
 
@@ -508,7 +565,7 @@ static void ou_optimize(unit_t *ou) {
 		/* get head of queue */
 		curr = list_entry_queue(ou->queue.next);
 		list_del(&curr->queue);
-		DBG((dbg, LEVEL_2, "\t  Examine qnode color %d with size %d\n", curr->color, curr->mis_size));
+		DBG((dbg, LEVEL_2, "\t  Examine qnode color %d with cost %d\n", curr->color, curr->mis_costs));
 
 		/* try */
 		if (qnode_try_color(curr))
@@ -522,9 +579,8 @@ 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  Copies: %d/%d\n", curr->color, ou->interf+ou->node_count-curr->mis_size, ou->interf+ou->node_count-1));
-		/* globally pin root and eventually others */
+		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];
@@ -537,10 +593,14 @@ 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);
 			}
 		}
+		/*
+		 * Enable for checking register allocation after each ou
+		 * be_ra_chordal_check(ou->co->chordal_env);
+		 */
 	}
 
 	/* free best qnode (curr) and queue */
@@ -552,11 +612,10 @@ 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);
+		firm_dbg_set_mask(dbg, DEBUG_IRG_LVL_HEUR);
 	else
-		firm_dbg_set_mask(dbg, DEBUG_LVL);
+		firm_dbg_set_mask(dbg, DEBUG_LVL_HEUR);
 
 	pinned_global = pset_new_ptr(SLOTS_PINNED_GLOBAL);
 	list_for_each_entry(unit_t, curr, &co->units, units)