X-Git-Url: http://nsz.repo.hu/git/?a=blobdiff_plain;f=ir%2Fbe%2Fbecopyilp2.c;h=06cb43e6105fc500498d78c4367e81fdc2f9ec76;hb=04906225f6fe1bc22f73b5d6a98eb9c0992a9b54;hp=93cb741318df81bd97da62db8778d995a09056d8;hpb=76c0c4f215716a69cebe49fbcba24c17a9595f32;p=libfirm

diff --git a/ir/be/becopyilp2.c b/ir/be/becopyilp2.c
index 93cb74131..06cb43e61 100644
--- a/ir/be/becopyilp2.c
+++ b/ir/be/becopyilp2.c
@@ -1,9 +1,28 @@
-/**
- * Author:      Daniel Grund
- * Date:		28.02.2006
- * Copyright:   (c) Universitaet Karlsruhe
- * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
+/*
+ * Copyright (C) 1995-2008 University of Karlsruhe.  All right reserved.
+ *
+ * This file is part of libFirm.
+ *
+ * This file may be distributed and/or modified under the terms of the
+ * GNU General Public License version 2 as published by the Free Software
+ * Foundation and appearing in the file LICENSE.GPL included in the
+ * packaging of this file.
  *
+ * Licensees holding valid libFirm Professional Edition licenses may use
+ * this file in accordance with the libFirm Commercial License.
+ * Agreement provided with the Software.
+ *
+ * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
+ * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE.
+ */
+
+/**
+ * @file
+ * @brief       ILP based copy minimization.
+ * @author      Daniel Grund
+ * @date        28.02.2006
+ * @version     $Id$
  *
  * ILP formalization using G=(V, E, Q):
  *  - 2 class of variables: coloring vars x_ic   and   equal color vars y_ij
@@ -25,13 +44,14 @@
  *
  *		x_nc, y_ij \in N,   w_ij \in R^+
  */
-
-#ifdef HAVE_CONFIG_H
 #include "config.h"
-#endif /* HAVE_CONFIG_H */
 
 #ifdef WITH_ILP
 
+#include "bitset.h"
+#include "raw_bitset.h"
+#include "pdeq.h"
+
 #include "irtools.h"
 #include "irgwalk.h"
 #include "becopyilp_t.h"
@@ -41,15 +61,17 @@
 #define DEBUG_LVL 1
 
 typedef struct _local_env_t {
-	firm_dbg_module_t *dbg;
 	double time_limit;
 	int first_x_var, last_x_var;
+	int n_colors;
+	bitset_t *normal_colors;
 	pmap *nr_2_irn;
+	DEBUG_ONLY(firm_dbg_module_t *dbg;)
 } local_env_t;
 
 static void build_coloring_cstr(ilp_env_t *ienv) {
-	be_ifg_t *ifg = ienv->co->cenv->ifg;
-	void *iter = be_ifg_nodes_iter_alloca(ifg);
+	be_ifg_t *ifg     = ienv->co->cenv->ifg;
+	void *iter        = be_ifg_nodes_iter_alloca(ifg);
 	bitset_t *colors;
 	ir_node *irn;
 	char buf[16];
@@ -58,28 +80,32 @@ static void build_coloring_cstr(ilp_env_t *ienv) {
 
 	be_ifg_foreach_node(ifg, iter, irn)
 		if (!sr_is_removed(ienv->sr, irn)) {
-			int col, cst_idx;
-			arch_register_req_t req;
-			int curr_node_color = get_irn_col(ienv->co, irn);
-			int node_nr = (int)get_irn_node_nr(irn);
+			bitset_pos_t col;
+			int cst_idx;
+			const arch_register_req_t *req;
+			int curr_node_color = get_irn_col(irn);
+			int node_nr = (int)get_irn_idx(irn);
 			local_env_t *lenv = ienv->env;
 
 			pmap_insert(lenv->nr_2_irn, INT_TO_PTR(node_nr), irn);
 
-			arch_get_register_req(ienv->co->aenv, &req, irn, -1);
+			req = arch_get_register_req(irn, -1);
+
+			bitset_clear_all(colors);
 
 			/* get assignable colors */
-			if (arch_register_req_is(&req, limited))
-				req.limited(req.limited_env, colors);
-			else
-				arch_put_non_ignore_regs(ienv->co->aenv, req.cls, colors);
+			if (arch_register_req_is(req, limited)) {
+				rbitset_copy_to_bitset(req->limited, colors);
+			} else {
+				bitset_copy(colors, lenv->normal_colors);
+			}
 
 			/* add the coloring constraint */
 			cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_equal, 1.0);
 
 			bitset_foreach(colors, col) {
 				int var_idx = lpp_add_var(ienv->lp, name_cdd(buf, 'x', node_nr, col), lpp_binary, 0.0);
-				lpp_set_start_value(ienv->lp, var_idx, (col == curr_node_color) ? 1.0 : 0.0);
+				lpp_set_start_value(ienv->lp, var_idx, (col == (unsigned) curr_node_color) ? 1.0 : 0.0);
 				lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1);
 
 				lenv->last_x_var = var_idx;
@@ -100,9 +126,10 @@ static void build_coloring_cstr(ilp_env_t *ienv) {
 }
 
 static void build_interference_cstr(ilp_env_t *ienv) {
-	lpp_t *lpp = ienv->lp;
-	be_ifg_t *ifg = ienv->co->cenv->ifg;
-	int n_colors = arch_register_class_n_regs(ienv->co->cls);
+	lpp_t *lpp        = ienv->lp;
+	local_env_t *lenv = ienv->env;
+	be_ifg_t *ifg     = ienv->co->cenv->ifg;
+	int n_colors      = lenv->n_colors;
 	int i, col;
 
 	void *iter = be_ifg_cliques_iter_alloca(ifg);
@@ -113,8 +140,13 @@ static void build_interference_cstr(ilp_env_t *ienv) {
 
 	/* for each maximal clique */
 	be_ifg_foreach_clique(ifg, iter, clique, &size) {
+		int realsize = 0;
+
+		for (i=0; i<size; ++i)
+			if (!sr_is_removed(ienv->sr, clique[i]))
+				++realsize;
 
-		if (size < 2)
+		if (realsize < 2)
 			continue;
 
 		/* for all colors */
@@ -126,7 +158,7 @@ static void build_interference_cstr(ilp_env_t *ienv) {
 				ir_node *irn = clique[i];
 
 				if (!sr_is_removed(ienv->sr, irn)) {
-					int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_node_nr(irn), col));
+					int var_idx = lpp_get_var_idx(lpp, name_cdd(buf, 'x', (int)get_irn_idx(irn), col));
 					lpp_set_factor_fast(lpp, cst_idx, var_idx, 1);
 				}
 			}
@@ -134,9 +166,16 @@ static void build_interference_cstr(ilp_env_t *ienv) {
 	}
 }
 
+
+/**
+ * TODO: Remove the dependency of the opt-units data structure
+ *       by walking over all affinity edges. Graph structure
+ *       does not provide this walker, yet.
+ */
 static void build_affinity_cstr(ilp_env_t *ienv) {
+	local_env_t *lenv = ienv->env;
+	int n_colors      = lenv->n_colors;
 	unit_t *curr;
-	int n_colors = arch_register_class_n_regs(ienv->co->cls);
 
 	/* for all optimization units */
 	list_for_each_entry(unit_t, curr, &ienv->co->units, units) {
@@ -146,13 +185,13 @@ static void build_affinity_cstr(ilp_env_t *ienv) {
 		int root_col, arg_col;
 
 		root = curr->nodes[0];
-		root_nr = (int) get_irn_node_nr(root);
-		root_col = get_irn_col(ienv->co, root);
+		root_nr = (int) get_irn_idx(root);
+		root_col = get_irn_col(root);
 
 		for (i = 1; i < curr->node_count; ++i) {
 			arg = curr->nodes[i];
-			arg_nr = (int) get_irn_node_nr(arg);
-			arg_col = get_irn_col(ienv->co, arg);
+			arg_nr = (int) get_irn_idx(arg);
+			arg_col = get_irn_col(arg);
 
 			/* add a new affinity variable */
 			y_idx = lpp_add_var(ienv->lp, name_cdd_sorted(buf, 'y', root_nr, arg_nr), lpp_binary, curr->costs[i]);
@@ -166,24 +205,277 @@ static void build_affinity_cstr(ilp_env_t *ienv) {
 
 				lpp_set_factor_fast(ienv->lp, cst_idx, root_idx,  1.0);
 				lpp_set_factor_fast(ienv->lp, cst_idx, arg_idx,  -1.0);
-				lpp_set_factor_fast(ienv->lp, cst_idx, root_idx, -1.0);
+				lpp_set_factor_fast(ienv->lp, cst_idx, y_idx, -1.0);
 			}
 		}
 	}
 }
 
 /**
- *
+ * Helping stuff for build_clique_star_cstr
+ */
+typedef struct _edge_t {
+	ir_node *n1, *n2;
+} edge_t;
+
+static int compare_edge_t(const void *k1, const void *k2, size_t size) {
+	const edge_t *e1 = k1;
+	const edge_t *e2 = k2;
+	(void) size;
+
+	return ! (e1->n1 == e2->n1   &&   e1->n2 == e2->n2);
+}
+
+#define HASH_EDGE(e) (hash_irn((e)->n1) ^ hash_irn((e)->n2))
+
+static INLINE edge_t *add_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
+	edge_t new_edge;
+
+	if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
+		new_edge.n1 = n1;
+		new_edge.n2 = n2;
+	} else {
+		new_edge.n1 = n2;
+		new_edge.n2 = n1;
+	}
+	(*counter)++;
+	return set_insert(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
+}
+
+static INLINE edge_t *find_edge(set *edges, ir_node *n1, ir_node *n2) {
+	edge_t new_edge;
+
+	if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
+		new_edge.n1 = n1;
+		new_edge.n2 = n2;
+	} else {
+		new_edge.n1 = n2;
+		new_edge.n2 = n1;
+	}
+	return set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
+}
+
+static INLINE void remove_edge(set *edges, ir_node *n1, ir_node *n2, int *counter) {
+	edge_t new_edge, *e;
+
+	if (PTR_TO_INT(n1) < PTR_TO_INT(n2)) {
+		new_edge.n1 = n1;
+		new_edge.n2 = n2;
+	} else {
+		new_edge.n1 = n2;
+		new_edge.n2 = n1;
+	}
+	e = set_find(edges, &new_edge, sizeof(new_edge), HASH_EDGE(&new_edge));
+	if (e) {
+		e->n1 = NULL;
+		e->n2 = NULL;
+		(*counter)--;
+	}
+}
+
+#define pset_foreach(pset, irn)  for(irn=pset_first(pset); irn; irn=pset_next(pset))
+
+/**
+ * Search for an interference clique and an external node
+ * with affinity edges to all nodes of the clique.
+ * At most 1 node of the clique can be colored equally with the external node.
  */
 static void build_clique_star_cstr(ilp_env_t *ienv) {
+	affinity_node_t *aff;
+
+	/* for each node with affinity edges */
+	co_gs_foreach_aff_node(ienv->co, aff) {
+		struct obstack ob;
+		neighb_t *nbr;
+		const ir_node *center = aff->irn;
+		ir_node **nodes;
+		set *edges;
+		int i, o, n_nodes, n_edges;
+
+		if (arch_irn_is(aff->irn, ignore))
+			continue;
+
+		obstack_init(&ob);
+		edges = new_set(compare_edge_t, 8);
+
+		/* get all affinity neighbours */
+		n_nodes = 0;
+		co_gs_foreach_neighb(aff, nbr) {
+			if (!arch_irn_is(nbr->irn, ignore)) {
+				obstack_ptr_grow(&ob, nbr->irn);
+				++n_nodes;
+			}
+		}
+		nodes = obstack_finish(&ob);
+
+		/* get all interference edges between these */
+		n_edges = 0;
+		for (i=0; i<n_nodes; ++i)
+			for (o=0; o<i; ++o)
+				if (be_ifg_connected(ienv->co->cenv->ifg, nodes[i], nodes[o]))
+					add_edge(edges, nodes[i], nodes[o], &n_edges);
+
+		/* cover all these interference edges with maximal cliques */
+		while (n_edges) {
+			edge_t *e;
+			pset *clique = pset_new_ptr(8);
+			int growed;
+
+			/* get 2 starting nodes to form a clique */
+			for (e=set_first(edges); !e->n1; e=set_next(edges))
+				/*nothing*/ ;
+
+			/* we could be stepped out of the loop before the set iterated to the end */
+			set_break(edges);
+
+			pset_insert_ptr(clique, e->n1);
+			pset_insert_ptr(clique, e->n2);
+			remove_edge(edges, e->n1, e->n2, &n_edges);
+
+			/* while the clique is growing */
+			do {
+				growed = 0;
+
+				/* search for a candidate to extend the clique */
+				for (i=0; i<n_nodes; ++i) {
+					ir_node *member, *cand = nodes[i];
+					int is_cand;
+
+					/* if its already in the clique try the next */
+					if (pset_find_ptr(clique, cand))
+						continue;
+
+					/* are there all necessary interferences? */
+					is_cand = 1;
+					pset_foreach(clique, member) {
+						if (!find_edge(edges, cand, member)) {
+							is_cand = 0;
+							pset_break(clique);
+							break;
+						}
+					}
+
+					/* now we know if we have a clique extender */
+					if (is_cand) {
+						/* first remove all covered edges */
+						pset_foreach(clique, member)
+							remove_edge(edges, cand, member, &n_edges);
+
+						/* insert into clique */
+						pset_insert_ptr(clique, cand);
+						growed = 1;
+						break;
+					}
+				}
+			} while (growed);
+
+			/* now the clique is maximal. Finally add the constraint */
+			{
+				ir_node *member;
+				int var_idx, cst_idx, center_nr, member_nr;
+				char buf[16];
+
+				cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, pset_count(clique)-1);
+				center_nr = get_irn_idx(center);
+
+				pset_foreach(clique, member) {
+					member_nr = get_irn_idx(member);
+					var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', center_nr, member_nr));
+					lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
+				}
+			}
+
+			del_pset(clique);
+		}
+
+		del_set(edges);
+		obstack_free(&ob, NULL);
+	}
+}
+
+
+static void extend_path(ilp_env_t *ienv, pdeq *path, const ir_node *irn) {
+	be_ifg_t *ifg = ienv->co->cenv->ifg;
+	int i, len;
+	ir_node **curr_path;
+	affinity_node_t *aff;
+	neighb_t *nbr;
+
+	/* do not walk backwards or in circles */
+	if (pdeq_contains(path, irn))
+		return;
+
+	if (arch_irn_is(irn, ignore))
+		return;
+
+	/* insert the new irn */
+	pdeq_putr(path, irn);
+
+
+
+	/* check for forbidden interferences */
+	len = pdeq_len(path);
+	curr_path = alloca(len * sizeof(*curr_path));
+	pdeq_copyl(path, (const void **)curr_path);
+
+	for (i=1; i<len; ++i)
+		if (be_ifg_connected(ifg, irn, curr_path[i]))
+			goto end;
+
+
+
+	/* check for terminating interference */
+	if (be_ifg_connected(ifg, irn, curr_path[0])) {
+
+		/* One node is not a path. */
+		/* And a path of length 2 is covered by a clique star constraint. */
+		if (len > 2) {
+			/* finally build the constraint */
+			int cst_idx = lpp_add_cst(ienv->lp, NULL, lpp_greater, 1.0);
+			for (i=1; i<len; ++i) {
+				char buf[16];
+				int nr_1    = get_irn_idx(curr_path[i-1]);
+				int nr_2    = get_irn_idx(curr_path[i]);
+				int var_idx = lpp_get_var_idx(ienv->lp, name_cdd_sorted(buf, 'y', nr_1, nr_2));
+				lpp_set_factor_fast(ienv->lp, cst_idx, var_idx, 1.0);
+			}
+		}
+
+		/* this path cannot be extended anymore */
+		goto end;
+	}
+
+
+
+	/* recursively extend the path */
+	aff = get_affinity_info(ienv->co, irn);
+	co_gs_foreach_neighb(aff, nbr)
+		extend_path(ienv, path, nbr->irn);
+
+
+end:
+	/* remove the irn */
+	pdeq_getr(path);
 
 }
 
 /**
- *
+ *  Search a path of affinity edges, whose ends are connected
+ *  by an interference edge and there are no other interference
+ *  edges in between.
+ *  Then at least one of these affinity edges must break.
  */
 static void build_path_cstr(ilp_env_t *ienv) {
+	affinity_node_t *aff_info;
 
+	/* for each node with affinity edges */
+	co_gs_foreach_aff_node(ienv->co, aff_info) {
+		pdeq *path = new_pdeq();
+
+		extend_path(ienv, path, aff_info->irn);
+
+		del_pdeq(path);
+	}
 }
 
 static void ilp2_build(ilp_env_t *ienv) {
@@ -204,33 +496,37 @@ static void ilp2_build(ilp_env_t *ienv) {
 
 static void ilp2_apply(ilp_env_t *ienv) {
 	local_env_t *lenv = ienv->env;
-	double *sol;
-	lpp_sol_state_t state;
-	int i, count;
-
-	count = lenv->last_x_var - lenv->first_x_var + 1;
-	sol = xmalloc(count * sizeof(sol[0]));
-	state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
-	if (state != lpp_optimal) {
-		printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
-		assert(state >= lpp_feasible && "The solution should at least be feasible!");
-	}
+	int i;
+
+	/* first check if there was sth. to optimize */
+	if (lenv->first_x_var >= 0) {
+		int              count = lenv->last_x_var - lenv->first_x_var + 1;
+		double          *sol   = XMALLOCN(double, count);
+		lpp_sol_state_t  state = lpp_get_solution(ienv->lp, sol, lenv->first_x_var, lenv->last_x_var);
 
-	for (i=0; i<count; ++i) {
-		int nodenr, color;
-		char var_name[16];
+		if (state != lpp_optimal) {
+			printf("WARNING %s: Solution state is not 'optimal': %d\n", ienv->co->name, state);
+			assert(state >= lpp_feasible && "The solution should at least be feasible!");
+		}
 
-		if (sol[i] > 1-EPSILON) { /* split variable name into components */
-			lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
+		for (i=0; i<count; ++i) {
+			int nodenr, color;
+			char var_name[16];
 
-			if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
-				ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
-				assert(irn && "This node number must be present in the map");
+			if (sol[i] > 1-EPSILON) { /* split variable name into components */
+				lpp_get_var_name(ienv->lp, lenv->first_x_var+i, var_name, sizeof(var_name));
 
-				set_irn_col(ienv->co, irn, color);
-			} else
-				assert(0 && "This should be a x-var");
+				if (sscanf(var_name, "x_%d_%d", &nodenr, &color) == 2) {
+					ir_node *irn = pmap_get(lenv->nr_2_irn, INT_TO_PTR(nodenr));
+					assert(irn && "This node number must be present in the map");
+
+					set_irn_col(ienv->co, irn, color);
+				} else
+					assert(0 && "This should be a x-var");
+			}
 		}
+
+		xfree(sol);
 	}
 
 #ifdef COPYOPT_STAT
@@ -242,17 +538,24 @@ static void ilp2_apply(ilp_env_t *ienv) {
 #endif
 }
 
-int co_solve_ilp2(copy_opt_t *co, double time_limit) {
+int co_solve_ilp2(copy_opt_t *co) {
 	lpp_sol_state_t sol_state;
 	ilp_env_t *ienv;
 	local_env_t my;
 
-	my.time_limit  = time_limit;
+	ASSERT_OU_AVAIL(co); //See build_clique_st
+	ASSERT_GS_AVAIL(co);
+
+	my.time_limit  = 0;
 	my.first_x_var = -1;
 	my.last_x_var  = -1;
 	my.nr_2_irn    = pmap_create();
-	my.dbg         = firm_dbg_register("ir.be.coilp2");
-	firm_dbg_set_mask(my.dbg, DEBUG_LVL);
+	FIRM_DBG_REGISTER(my.dbg, "firm.be.coilp2");
+
+	my.normal_colors = bitset_alloca(arch_register_class_n_regs(co->cls));
+	bitset_clear_all(my.normal_colors);
+	arch_put_non_ignore_regs(co->cls, my.normal_colors);
+	my.n_colors = bitset_popcnt(my.normal_colors);
 
 	ienv = new_ilp_env(co, ilp2_build, ilp2_apply, &my);
 
@@ -266,7 +569,7 @@ int co_solve_ilp2(copy_opt_t *co, double time_limit) {
 
 #else /* WITH_ILP */
 
-static void only_that_you_can_compile_without_WITH_ILP_defined(void) {
+static INLINE void only_that_you_can_compile_without_WITH_ILP_defined(void) {
 }
 
 #endif /* WITH_ILP */