X-Git-Url: http://review.tizen.org/git/?a=blobdiff_plain;f=isl_coalesce.c;h=c2c3bb44d7366e169a2337b0f2feb1546011f6f7;hb=de51a9bc4da5dd3f1f9f57c2362da6f9752c44e0;hp=b99c7b53d2b8b1a791f4aeab576e2769720836dc;hpb=d07eed3bf4d0e290118588ca80997a8035eaeb21;p=platform%2Fupstream%2Fisl.git

diff --git a/isl_coalesce.c b/isl_coalesce.c
index b99c7b5..c2c3bb4 100644
--- a/isl_coalesce.c
+++ b/isl_coalesce.c
@@ -1,19 +1,23 @@
 /*
  * Copyright 2008-2009 Katholieke Universiteit Leuven
  * Copyright 2010      INRIA Saclay
+ * Copyright 2012      Ecole Normale Superieure
  *
- * Use of this software is governed by the GNU LGPLv2.1 license
+ * Use of this software is governed by the MIT license
  *
  * Written by Sven Verdoolaege, K.U.Leuven, Departement
  * Computerwetenschappen, Celestijnenlaan 200A, B-3001 Leuven, Belgium
  * and INRIA Saclay - Ile-de-France, Parc Club Orsay Universite,
  * ZAC des vignes, 4 rue Jacques Monod, 91893 Orsay, France 
+ * and Ecole Normale Superieure, 45 rue d’Ulm, 75230 Paris, France
  */
 
 #include "isl_map_private.h"
 #include <isl/seq.h>
+#include <isl/options.h>
 #include "isl_tab.h"
 #include <isl_mat_private.h>
+#include <isl_local_space_private.h>
 
 #define STATUS_ERROR		-1
 #define STATUS_REDUNDANT	 1
@@ -37,24 +41,24 @@ static int status_in(isl_int *ineq, struct isl_tab *tab)
 	}
 }
 
-/* Compute the position of the equalities of basic map "i"
- * with respect to basic map "j".
+/* Compute the position of the equalities of basic map "bmap_i"
+ * with respect to the basic map represented by "tab_j".
  * The resulting array has twice as many entries as the number
  * of equalities corresponding to the two inequalties to which
  * each equality corresponds.
  */
-static int *eq_status_in(struct isl_map *map, int i, int j,
-	struct isl_tab **tabs)
+static int *eq_status_in(__isl_keep isl_basic_map *bmap_i,
+	struct isl_tab *tab_j)
 {
 	int k, l;
-	int *eq = isl_calloc_array(map->ctx, int, 2 * map->p[i]->n_eq);
+	int *eq = isl_calloc_array(bmap_i->ctx, int, 2 * bmap_i->n_eq);
 	unsigned dim;
 
-	dim = isl_basic_map_total_dim(map->p[i]);
-	for (k = 0; k < map->p[i]->n_eq; ++k) {
+	dim = isl_basic_map_total_dim(bmap_i);
+	for (k = 0; k < bmap_i->n_eq; ++k) {
 		for (l = 0; l < 2; ++l) {
-			isl_seq_neg(map->p[i]->eq[k], map->p[i]->eq[k], 1+dim);
-			eq[2 * k + l] = status_in(map->p[i]->eq[k], tabs[j]);
+			isl_seq_neg(bmap_i->eq[k], bmap_i->eq[k], 1+dim);
+			eq[2 * k + l] = status_in(bmap_i->eq[k], tab_j);
 			if (eq[2 * k + l] == STATUS_ERROR)
 				goto error;
 		}
@@ -69,22 +73,23 @@ error:
 	return NULL;
 }
 
-/* Compute the position of the inequalities of basic map "i"
- * with respect to basic map "j".
+/* Compute the position of the inequalities of basic map "bmap_i"
+ * (also represented by "tab_i", if not NULL) with respect to the basic map
+ * represented by "tab_j".
  */
-static int *ineq_status_in(struct isl_map *map, int i, int j,
-	struct isl_tab **tabs)
+static int *ineq_status_in(__isl_keep isl_basic_map *bmap_i,
+	struct isl_tab *tab_i, struct isl_tab *tab_j)
 {
 	int k;
-	unsigned n_eq = map->p[i]->n_eq;
-	int *ineq = isl_calloc_array(map->ctx, int, map->p[i]->n_ineq);
+	unsigned n_eq = bmap_i->n_eq;
+	int *ineq = isl_calloc_array(bmap_i->ctx, int, bmap_i->n_ineq);
 
-	for (k = 0; k < map->p[i]->n_ineq; ++k) {
-		if (isl_tab_is_redundant(tabs[i], n_eq + k)) {
+	for (k = 0; k < bmap_i->n_ineq; ++k) {
+		if (tab_i && isl_tab_is_redundant(tab_i, n_eq + k)) {
 			ineq[k] = STATUS_REDUNDANT;
 			continue;
 		}
-		ineq[k] = status_in(map->p[i]->ineq[k], tabs[j]);
+		ineq[k] = status_in(bmap_i->ineq[k], tab_j);
 		if (ineq[k] == STATUS_ERROR)
 			goto error;
 		if (ineq[k] == STATUS_SEPARATE)
@@ -223,7 +228,7 @@ static int fuse(struct isl_map *map, int i, int j,
 	    ISL_F_ISSET(map->p[j], ISL_BASIC_MAP_RATIONAL))
 		ISL_F_SET(fused, ISL_BASIC_MAP_RATIONAL);
 
-	fused_tab = isl_tab_from_basic_map(fused);
+	fused_tab = isl_tab_from_basic_map(fused, 0);
 	if (isl_tab_detect_redundant(fused_tab) < 0)
 		goto error;
 
@@ -394,6 +399,9 @@ static int is_extension(struct isl_map *map, int i, int j, int k,
 	struct isl_tab_undo *snap, *snap2;
 	unsigned n_eq = map->p[i]->n_eq;
 
+	if (isl_tab_is_equality(tabs[i], n_eq + k))
+		return 0;
+
 	snap = isl_tab_snap(tabs[i]);
 	tabs[i] = isl_tab_relax(tabs[i], n_eq + k);
 	snap2 = isl_tab_snap(tabs[i]);
@@ -417,6 +425,106 @@ static int is_extension(struct isl_map *map, int i, int j, int k,
 	return changed;
 }
 
+/* Data structure that keeps track of the wrapping constraints
+ * and of information to bound the coefficients of those constraints.
+ *
+ * bound is set if we want to apply a bound on the coefficients
+ * mat contains the wrapping constraints
+ * max is the bound on the coefficients (if bound is set)
+ */
+struct isl_wraps {
+	int bound;
+	isl_mat *mat;
+	isl_int max;
+};
+
+/* Update wraps->max to be greater than or equal to the coefficients
+ * in the equalities and inequalities of bmap that can be removed if we end up
+ * applying wrapping.
+ */
+static void wraps_update_max(struct isl_wraps *wraps,
+	__isl_keep isl_basic_map *bmap, int *eq, int *ineq)
+{
+	int k;
+	isl_int max_k;
+	unsigned total = isl_basic_map_total_dim(bmap);
+
+	isl_int_init(max_k);
+
+	for (k = 0; k < bmap->n_eq; ++k) {
+		if (eq[2 * k] == STATUS_VALID &&
+		    eq[2 * k + 1] == STATUS_VALID)
+			continue;
+		isl_seq_abs_max(bmap->eq[k] + 1, total, &max_k);
+		if (isl_int_abs_gt(max_k, wraps->max))
+			isl_int_set(wraps->max, max_k);
+	}
+
+	for (k = 0; k < bmap->n_ineq; ++k) {
+		if (ineq[k] == STATUS_VALID || ineq[k] == STATUS_REDUNDANT)
+			continue;
+		isl_seq_abs_max(bmap->ineq[k] + 1, total, &max_k);
+		if (isl_int_abs_gt(max_k, wraps->max))
+			isl_int_set(wraps->max, max_k);
+	}
+
+	isl_int_clear(max_k);
+}
+
+/* Initialize the isl_wraps data structure.
+ * If we want to bound the coefficients of the wrapping constraints,
+ * we set wraps->max to the largest coefficient
+ * in the equalities and inequalities that can be removed if we end up
+ * applying wrapping.
+ */
+static void wraps_init(struct isl_wraps *wraps, __isl_take isl_mat *mat,
+	__isl_keep isl_map *map, int i, int j,
+	int *eq_i, int *ineq_i, int *eq_j, int *ineq_j)
+{
+	isl_ctx *ctx;
+
+	wraps->bound = 0;
+	wraps->mat = mat;
+	if (!mat)
+		return;
+	ctx = isl_mat_get_ctx(mat);
+	wraps->bound = isl_options_get_coalesce_bounded_wrapping(ctx);
+	if (!wraps->bound)
+		return;
+	isl_int_init(wraps->max);
+	isl_int_set_si(wraps->max, 0);
+	wraps_update_max(wraps, map->p[i], eq_i, ineq_i);
+	wraps_update_max(wraps, map->p[j], eq_j, ineq_j);
+}
+
+/* Free the contents of the isl_wraps data structure.
+ */
+static void wraps_free(struct isl_wraps *wraps)
+{
+	isl_mat_free(wraps->mat);
+	if (wraps->bound)
+		isl_int_clear(wraps->max);
+}
+
+/* Is the wrapping constraint in row "row" allowed?
+ *
+ * If wraps->bound is set, we check that none of the coefficients
+ * is greater than wraps->max.
+ */
+static int allow_wrap(struct isl_wraps *wraps, int row)
+{
+	int i;
+
+	if (!wraps->bound)
+		return 1;
+
+	for (i = 1; i < wraps->mat->n_col; ++i)
+		if (isl_int_abs_gt(wraps->mat->row[row][i], wraps->max))
+			return 0;
+
+	return 1;
+}
+
 /* For each non-redundant constraint in "bmap" (as determined by "tab"),
  * wrap the constraint around "bound" such that it includes the whole
  * set "set" and append the resulting constraint to "wraps".
@@ -427,15 +535,18 @@ static int is_extension(struct isl_map *map, int i, int j, int k,
  * identical to "bound", then this means that "set" is unbounded in such
  * way that no wrapping is possible.  If this happens then wraps->n_row
  * is reset to zero.
+ * Similarly, if we want to bound the coefficients of the wrapping
+ * constraints and a newly added wrapping constraint does not
+ * satisfy the bound, then wraps->n_row is also reset to zero.
  */
-static int add_wraps(__isl_keep isl_mat *wraps, __isl_keep isl_basic_map *bmap,
+static int add_wraps(struct isl_wraps *wraps, __isl_keep isl_basic_map *bmap,
 	struct isl_tab *tab, isl_int *bound, __isl_keep isl_set *set)
 {
 	int l;
 	int w;
 	unsigned total = isl_basic_map_total_dim(bmap);
 
-	w = wraps->n_row;
+	w = wraps->mat->n_row;
 
 	for (l = 0; l < bmap->n_ineq; ++l) {
 		if (isl_seq_is_neg(bound, bmap->ineq[l], 1 + total))
@@ -445,10 +556,12 @@ static int add_wraps(__isl_keep isl_mat *wraps, __isl_keep isl_basic_map *bmap,
 		if (isl_tab_is_redundant(tab, bmap->n_eq + l))
 			continue;
 
-		isl_seq_cpy(wraps->row[w], bound, 1 + total);
-		if (!isl_set_wrap_facet(set, wraps->row[w], bmap->ineq[l]))
+		isl_seq_cpy(wraps->mat->row[w], bound, 1 + total);
+		if (!isl_set_wrap_facet(set, wraps->mat->row[w], bmap->ineq[l]))
 			return -1;
-		if (isl_seq_eq(wraps->row[w], bound, 1 + total))
+		if (isl_seq_eq(wraps->mat->row[w], bound, 1 + total))
+			goto unbounded;
+		if (!allow_wrap(wraps, w))
 			goto unbounded;
 		++w;
 	}
@@ -458,26 +571,31 @@ static int add_wraps(__isl_keep isl_mat *wraps, __isl_keep isl_basic_map *bmap,
 		if (isl_seq_eq(bound, bmap->eq[l], 1 + total))
 			continue;
 
-		isl_seq_cpy(wraps->row[w], bound, 1 + total);
-		isl_seq_neg(wraps->row[w + 1], bmap->eq[l], 1 + total);
-		if (!isl_set_wrap_facet(set, wraps->row[w], wraps->row[w + 1]))
+		isl_seq_cpy(wraps->mat->row[w], bound, 1 + total);
+		isl_seq_neg(wraps->mat->row[w + 1], bmap->eq[l], 1 + total);
+		if (!isl_set_wrap_facet(set, wraps->mat->row[w],
+					wraps->mat->row[w + 1]))
 			return -1;
-		if (isl_seq_eq(wraps->row[w], bound, 1 + total))
+		if (isl_seq_eq(wraps->mat->row[w], bound, 1 + total))
+			goto unbounded;
+		if (!allow_wrap(wraps, w))
 			goto unbounded;
 		++w;
 
-		isl_seq_cpy(wraps->row[w], bound, 1 + total);
-		if (!isl_set_wrap_facet(set, wraps->row[w], bmap->eq[l]))
+		isl_seq_cpy(wraps->mat->row[w], bound, 1 + total);
+		if (!isl_set_wrap_facet(set, wraps->mat->row[w], bmap->eq[l]))
 			return -1;
-		if (isl_seq_eq(wraps->row[w], bound, 1 + total))
+		if (isl_seq_eq(wraps->mat->row[w], bound, 1 + total))
+			goto unbounded;
+		if (!allow_wrap(wraps, w))
 			goto unbounded;
 		++w;
 	}
 
-	wraps->n_row = w;
+	wraps->mat->n_row = w;
 	return 0;
 unbounded:
-	wraps->n_row = 0;
+	wraps->mat->n_row = 0;
 	return 0;
 }
 
@@ -555,7 +673,8 @@ static int can_wrap_in_facet(struct isl_map *map, int i, int j, int k,
 	struct isl_tab **tabs, int *eq_i, int *ineq_i, int *eq_j, int *ineq_j)
 {
 	int changed = 0;
-	struct isl_mat *wraps = NULL;
+	struct isl_wraps wraps;
+	isl_mat *mat;
 	struct isl_set *set_i = NULL;
 	struct isl_set *set_j = NULL;
 	struct isl_vec *bound = NULL;
@@ -565,22 +684,23 @@ static int can_wrap_in_facet(struct isl_map *map, int i, int j, int k,
 
 	set_i = set_from_updated_bmap(map->p[i], tabs[i]);
 	set_j = set_from_updated_bmap(map->p[j], tabs[j]);
-	wraps = isl_mat_alloc(map->ctx, 2 * (map->p[i]->n_eq + map->p[j]->n_eq) +
+	mat = isl_mat_alloc(map->ctx, 2 * (map->p[i]->n_eq + map->p[j]->n_eq) +
 					map->p[i]->n_ineq + map->p[j]->n_ineq,
 					1 + total);
+	wraps_init(&wraps, mat, map, i, j, eq_i, ineq_i, eq_j, ineq_j);
 	bound = isl_vec_alloc(map->ctx, 1 + total);
-	if (!set_i || !set_j || !wraps || !bound)
+	if (!set_i || !set_j || !wraps.mat || !bound)
 		goto error;
 
 	isl_seq_cpy(bound->el, map->p[i]->ineq[k], 1 + total);
 	isl_int_add_ui(bound->el[0], bound->el[0], 1);
 
-	isl_seq_cpy(wraps->row[0], bound->el, 1 + total);
-	wraps->n_row = 1;
+	isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
+	wraps.mat->n_row = 1;
 
-	if (add_wraps(wraps, map->p[j], tabs[j], bound->el, set_i) < 0)
+	if (add_wraps(&wraps, map->p[j], tabs[j], bound->el, set_i) < 0)
 		goto error;
-	if (!wraps->n_row)
+	if (!wraps.mat->n_row)
 		goto unbounded;
 
 	snap = isl_tab_snap(tabs[i]);
@@ -592,21 +712,21 @@ static int can_wrap_in_facet(struct isl_map *map, int i, int j, int k,
 
 	isl_seq_neg(bound->el, map->p[i]->ineq[k], 1 + total);
 
-	n = wraps->n_row;
-	if (add_wraps(wraps, map->p[i], tabs[i], bound->el, set_j) < 0)
+	n = wraps.mat->n_row;
+	if (add_wraps(&wraps, map->p[i], tabs[i], bound->el, set_j) < 0)
 		goto error;
 
 	if (isl_tab_rollback(tabs[i], snap) < 0)
 		goto error;
-	if (check_wraps(wraps, n, tabs[i]) < 0)
+	if (check_wraps(wraps.mat, n, tabs[i]) < 0)
 		goto error;
-	if (!wraps->n_row)
+	if (!wraps.mat->n_row)
 		goto unbounded;
 
-	changed = fuse(map, i, j, tabs, eq_i, ineq_i, eq_j, ineq_j, wraps);
+	changed = fuse(map, i, j, tabs, eq_i, ineq_i, eq_j, ineq_j, wraps.mat);
 
 unbounded:
-	isl_mat_free(wraps);
+	wraps_free(&wraps);
 
 	isl_set_free(set_i);
 	isl_set_free(set_j);
@@ -615,8 +735,8 @@ unbounded:
 
 	return changed;
 error:
+	wraps_free(&wraps);
 	isl_vec_free(bound);
-	isl_mat_free(wraps);
 	isl_set_free(set_i);
 	isl_set_free(set_j);
 	return -1;
@@ -676,7 +796,8 @@ static int wrap_in_facets(struct isl_map *map, int i, int j,
 	int *eq_i, int *ineq_i, int *eq_j, int *ineq_j)
 {
 	int changed = 0;
-	isl_mat *wraps = NULL;
+	struct isl_wraps wraps;
+	isl_mat *mat;
 	isl_set *set = NULL;
 	isl_vec *bound = NULL;
 	unsigned total = isl_basic_map_total_dim(map->p[i]);
@@ -693,15 +814,16 @@ static int wrap_in_facets(struct isl_map *map, int i, int j,
 
 	set = isl_set_union(set_from_updated_bmap(map->p[i], tabs[i]),
 			    set_from_updated_bmap(map->p[j], tabs[j]));
-	wraps = isl_mat_alloc(map->ctx, max_wrap, 1 + total);
+	mat = isl_mat_alloc(map->ctx, max_wrap, 1 + total);
+	wraps_init(&wraps, mat, map, i, j, eq_i, ineq_i, eq_j, ineq_j);
 	bound = isl_vec_alloc(map->ctx, 1 + total);
-	if (!set || !wraps || !bound)
+	if (!set || !wraps.mat || !bound)
 		goto error;
 
 	snap_i = isl_tab_snap(tabs[i]);
 	snap_j = isl_tab_snap(tabs[j]);
 
-	wraps->n_row = 0;
+	wraps.mat->n_row = 0;
 
 	for (k = 0; k < n; ++k) {
 		if (isl_tab_select_facet(tabs[i], map->p[i]->n_eq + cuts[k]) < 0)
@@ -711,14 +833,17 @@ static int wrap_in_facets(struct isl_map *map, int i, int j,
 		set_is_redundant(tabs[i], map->p[i]->n_eq, cuts, n, k, 1);
 
 		isl_seq_neg(bound->el, map->p[i]->ineq[cuts[k]], 1 + total);
-		if (add_wraps(wraps, map->p[i], tabs[i], bound->el, set) < 0)
+		if (!tabs[i]->empty &&
+		    add_wraps(&wraps, map->p[i], tabs[i], bound->el, set) < 0)
 			goto error;
 
 		set_is_redundant(tabs[i], map->p[i]->n_eq, cuts, n, k, 0);
 		if (isl_tab_rollback(tabs[i], snap_i) < 0)
 			goto error;
 
-		if (!wraps->n_row)
+		if (tabs[i]->empty)
+			break;
+		if (!wraps.mat->n_row)
 			break;
 
 		isl_seq_cpy(bound->el, map->p[i]->ineq[cuts[k]], 1 + total);
@@ -729,28 +854,28 @@ static int wrap_in_facets(struct isl_map *map, int i, int j,
 			goto error;
 
 		if (!tabs[j]->empty &&
-		    add_wraps(wraps, map->p[j], tabs[j], bound->el, set) < 0)
+		    add_wraps(&wraps, map->p[j], tabs[j], bound->el, set) < 0)
 			goto error;
 
 		if (isl_tab_rollback(tabs[j], snap_j) < 0)
 			goto error;
 
-		if (!wraps->n_row)
+		if (!wraps.mat->n_row)
 			break;
 	}
 
 	if (k == n)
 		changed = fuse(map, i, j, tabs,
-				eq_i, ineq_i, eq_j, ineq_j, wraps);
+				eq_i, ineq_i, eq_j, ineq_j, wraps.mat);
 
 	isl_vec_free(bound);
-	isl_mat_free(wraps);
+	wraps_free(&wraps);
 	isl_set_free(set);
 
 	return changed;
 error:
 	isl_vec_free(bound);
-	isl_mat_free(wraps);
+	wraps_free(&wraps);
 	isl_set_free(set);
 	return -1;
 }
@@ -971,7 +1096,8 @@ static int check_eq_adj_eq(struct isl_map *map, int i, int j,
 {
 	int k;
 	int changed = 0;
-	struct isl_mat *wraps = NULL;
+	struct isl_wraps wraps;
+	isl_mat *mat;
 	struct isl_set *set_i = NULL;
 	struct isl_set *set_j = NULL;
 	struct isl_vec *bound = NULL;
@@ -986,11 +1112,12 @@ static int check_eq_adj_eq(struct isl_map *map, int i, int j,
 
 	set_i = set_from_updated_bmap(map->p[i], tabs[i]);
 	set_j = set_from_updated_bmap(map->p[j], tabs[j]);
-	wraps = isl_mat_alloc(map->ctx, 2 * (map->p[i]->n_eq + map->p[j]->n_eq) +
+	mat = isl_mat_alloc(map->ctx, 2 * (map->p[i]->n_eq + map->p[j]->n_eq) +
 					map->p[i]->n_ineq + map->p[j]->n_ineq,
 					1 + total);
+	wraps_init(&wraps, mat, map, i, j, eq_i, ineq_i, eq_j, ineq_j);
 	bound = isl_vec_alloc(map->ctx, 1 + total);
-	if (!set_i || !set_j || !wraps || !bound)
+	if (!set_i || !set_j || !wraps.mat || !bound)
 		goto error;
 
 	if (k % 2 == 0)
@@ -999,33 +1126,33 @@ static int check_eq_adj_eq(struct isl_map *map, int i, int j,
 		isl_seq_cpy(bound->el, map->p[i]->eq[k / 2], 1 + total);
 	isl_int_add_ui(bound->el[0], bound->el[0], 1);
 
-	isl_seq_cpy(wraps->row[0], bound->el, 1 + total);
-	wraps->n_row = 1;
+	isl_seq_cpy(wraps.mat->row[0], bound->el, 1 + total);
+	wraps.mat->n_row = 1;
 
-	if (add_wraps(wraps, map->p[j], tabs[j], bound->el, set_i) < 0)
+	if (add_wraps(&wraps, map->p[j], tabs[j], bound->el, set_i) < 0)
 		goto error;
-	if (!wraps->n_row)
+	if (!wraps.mat->n_row)
 		goto unbounded;
 
 	isl_int_sub_ui(bound->el[0], bound->el[0], 1);
 	isl_seq_neg(bound->el, bound->el, 1 + total);
 
-	isl_seq_cpy(wraps->row[wraps->n_row], bound->el, 1 + total);
-	wraps->n_row++;
+	isl_seq_cpy(wraps.mat->row[wraps.mat->n_row], bound->el, 1 + total);
+	wraps.mat->n_row++;
 
-	if (add_wraps(wraps, map->p[i], tabs[i], bound->el, set_j) < 0)
+	if (add_wraps(&wraps, map->p[i], tabs[i], bound->el, set_j) < 0)
 		goto error;
-	if (!wraps->n_row)
+	if (!wraps.mat->n_row)
 		goto unbounded;
 
-	changed = fuse(map, i, j, tabs, eq_i, ineq_i, eq_j, ineq_j, wraps);
+	changed = fuse(map, i, j, tabs, eq_i, ineq_i, eq_j, ineq_j, wraps.mat);
 
 	if (0) {
 error:		changed = -1;
 	}
 unbounded:
 
-	isl_mat_free(wraps);
+	wraps_free(&wraps);
 	isl_set_free(set_i);
 	isl_set_free(set_j);
 	isl_vec_free(bound);
@@ -1037,6 +1164,7 @@ unbounded:
  * can be represented by a single basic map.
  * If so, replace the pair by the single basic map and return 1.
  * Otherwise, return 0;
+ * The two basic maps are assumed to live in the same local space.
  *
  * We first check the effect of each constraint of one basic map
  * on the other basic map.
@@ -1106,7 +1234,7 @@ unbounded:
  * corresponding to the basic maps.  When the basic maps are dropped
  * or combined, the tableaus are modified accordingly.
  */
-static int coalesce_pair(struct isl_map *map, int i, int j,
+static int coalesce_local_pair(__isl_keep isl_map *map, int i, int j,
 	struct isl_tab **tabs)
 {
 	int changed = 0;
@@ -1115,7 +1243,7 @@ static int coalesce_pair(struct isl_map *map, int i, int j,
 	int *ineq_i = NULL;
 	int *ineq_j = NULL;
 
-	eq_i = eq_status_in(map, i, j, tabs);
+	eq_i = eq_status_in(map->p[i], tabs[j]);
 	if (!eq_i)
 		goto error;
 	if (any(eq_i, 2 * map->p[i]->n_eq, STATUS_ERROR))
@@ -1123,7 +1251,7 @@ static int coalesce_pair(struct isl_map *map, int i, int j,
 	if (any(eq_i, 2 * map->p[i]->n_eq, STATUS_SEPARATE))
 		goto done;
 
-	eq_j = eq_status_in(map, j, i, tabs);
+	eq_j = eq_status_in(map->p[j], tabs[i]);
 	if (!eq_j)
 		goto error;
 	if (any(eq_j, 2 * map->p[j]->n_eq, STATUS_ERROR))
@@ -1131,7 +1259,7 @@ static int coalesce_pair(struct isl_map *map, int i, int j,
 	if (any(eq_j, 2 * map->p[j]->n_eq, STATUS_SEPARATE))
 		goto done;
 
-	ineq_i = ineq_status_in(map, i, j, tabs);
+	ineq_i = ineq_status_in(map->p[i], tabs[i], tabs[j]);
 	if (!ineq_i)
 		goto error;
 	if (any(ineq_i, map->p[i]->n_ineq, STATUS_ERROR))
@@ -1139,7 +1267,7 @@ static int coalesce_pair(struct isl_map *map, int i, int j,
 	if (any(ineq_i, map->p[i]->n_ineq, STATUS_SEPARATE))
 		goto done;
 
-	ineq_j = ineq_status_in(map, j, i, tabs);
+	ineq_j = ineq_status_in(map->p[j], tabs[j], tabs[i]);
 	if (!ineq_j)
 		goto error;
 	if (any(ineq_j, map->p[j]->n_ineq, STATUS_ERROR))
@@ -1198,6 +1326,190 @@ error:
 	return -1;
 }
 
+/* Do the two basic maps live in the same local space, i.e.,
+ * do they have the same (known) divs?
+ * If either basic map has any unknown divs, then we can only assume
+ * that they do not live in the same local space.
+ */
+static int same_divs(__isl_keep isl_basic_map *bmap1,
+	__isl_keep isl_basic_map *bmap2)
+{
+	int i;
+	int known;
+	int total;
+
+	if (!bmap1 || !bmap2)
+		return -1;
+	if (bmap1->n_div != bmap2->n_div)
+		return 0;
+
+	if (bmap1->n_div == 0)
+		return 1;
+
+	known = isl_basic_map_divs_known(bmap1);
+	if (known < 0 || !known)
+		return known;
+	known = isl_basic_map_divs_known(bmap2);
+	if (known < 0 || !known)
+		return known;
+
+	total = isl_basic_map_total_dim(bmap1);
+	for (i = 0; i < bmap1->n_div; ++i)
+		if (!isl_seq_eq(bmap1->div[i], bmap2->div[i], 2 + total))
+			return 0;
+
+	return 1;
+}
+
+/* Given two basic maps "i" and "j", where the divs of "i" form a subset
+ * of those of "j", check if basic map "j" is a subset of basic map "i"
+ * and, if so, drop basic map "j".
+ *
+ * We first expand the divs of basic map "i" to match those of basic map "j",
+ * using the divs and expansion computed by the caller.
+ * Then we check if all constraints of the expanded "i" are valid for "j".
+ */
+static int coalesce_subset(__isl_keep isl_map *map, int i, int j,
+	struct isl_tab **tabs, __isl_keep isl_mat *div, int *exp)
+{
+	isl_basic_map *bmap;
+	int changed = 0;
+	int *eq_i = NULL;
+	int *ineq_i = NULL;
+
+	bmap = isl_basic_map_copy(map->p[i]);
+	bmap = isl_basic_set_expand_divs(bmap, isl_mat_copy(div), exp);
+
+	if (!bmap)
+		goto error;
+
+	eq_i = eq_status_in(bmap, tabs[j]);
+	if (!eq_i)
+		goto error;
+	if (any(eq_i, 2 * bmap->n_eq, STATUS_ERROR))
+		goto error;
+	if (any(eq_i, 2 * bmap->n_eq, STATUS_SEPARATE))
+		goto done;
+
+	ineq_i = ineq_status_in(bmap, NULL, tabs[j]);
+	if (!ineq_i)
+		goto error;
+	if (any(ineq_i, bmap->n_ineq, STATUS_ERROR))
+		goto error;
+	if (any(ineq_i, bmap->n_ineq, STATUS_SEPARATE))
+		goto done;
+
+	if (all(eq_i, 2 * map->p[i]->n_eq, STATUS_VALID) &&
+	    all(ineq_i, map->p[i]->n_ineq, STATUS_VALID)) {
+		drop(map, j, tabs);
+		changed = 1;
+	}
+
+done:
+	isl_basic_map_free(bmap);
+	free(eq_i);
+	free(ineq_i);
+	return 0;
+error:
+	isl_basic_map_free(bmap);
+	free(eq_i);
+	free(ineq_i);
+	return -1;
+}
+
+/* Check if the basic map "j" is a subset of basic map "i",
+ * assuming that "i" has fewer divs that "j".
+ * If not, then we change the order.
+ *
+ * If the two basic maps have the same number of divs, then
+ * they must necessarily be different.  Otherwise, we would have
+ * called coalesce_local_pair.  We therefore don't do try anyhing
+ * in this case.
+ *
+ * We first check if the divs of "i" are all known and form a subset
+ * of those of "j".  If so, we pass control over to coalesce_subset.
+ */
+static int check_coalesce_subset(__isl_keep isl_map *map, int i, int j,
+	struct isl_tab **tabs)
+{
+	int known;
+	isl_mat *div_i, *div_j, *div;
+	int *exp1 = NULL;
+	int *exp2 = NULL;
+	isl_ctx *ctx;
+	int subset;
+
+	if (map->p[i]->n_div == map->p[j]->n_div)
+		return 0;
+	if (map->p[j]->n_div < map->p[i]->n_div)
+		return check_coalesce_subset(map, j, i, tabs);
+
+	known = isl_basic_map_divs_known(map->p[i]);
+	if (known < 0 || !known)
+		return known;
+
+	ctx = isl_map_get_ctx(map);
+
+	div_i = isl_basic_map_get_divs(map->p[i]);
+	div_j = isl_basic_map_get_divs(map->p[j]);
+
+	if (!div_i || !div_j)
+		goto error;
+
+	exp1 = isl_alloc_array(ctx, int, div_i->n_row);
+	exp2 = isl_alloc_array(ctx, int, div_j->n_row);
+	if (!exp1 || !exp2)
+		goto error;
+
+	div = isl_merge_divs(div_i, div_j, exp1, exp2);
+	if (!div)
+		goto error;
+
+	if (div->n_row == div_j->n_row)
+		subset = coalesce_subset(map, i, j, tabs, div, exp1);
+	else
+		subset = 0;
+
+	isl_mat_free(div);
+
+	isl_mat_free(div_i);
+	isl_mat_free(div_j);
+
+	free(exp2);
+	free(exp1);
+
+	return subset;
+error:
+	isl_mat_free(div_i);
+	isl_mat_free(div_j);
+	free(exp1);
+	free(exp2);
+	return -1;
+}
+
+/* Check if the union of the given pair of basic maps
+ * can be represented by a single basic map.
+ * If so, replace the pair by the single basic map and return 1.
+ * Otherwise, return 0;
+ *
+ * We first check if the two basic maps live in the same local space.
+ * If so, we do the complete check.  Otherwise, we check if one is
+ * an obvious subset of the other.
+ */
+static int coalesce_pair(__isl_keep isl_map *map, int i, int j,
+	struct isl_tab **tabs)
+{
+	int same;
+
+	same = same_divs(map->p[i], map->p[j]);
+	if (same < 0)
+		return -1;
+	if (same)
+		return coalesce_local_pair(map, i, j, tabs);
+
+	return check_coalesce_subset(map, i, j, tabs);
+}
+
 static struct isl_map *coalesce(struct isl_map *map, struct isl_tab **tabs)
 {
 	int i, j;
@@ -1221,6 +1533,16 @@ error:
 /* For each pair of basic maps in the map, check if the union of the two
  * can be represented by a single basic map.
  * If so, replace the pair by the single basic map and start over.
+ *
+ * Since we are constructing the tableaus of the basic maps anyway,
+ * we exploit them to detect implicit equalities and redundant constraints.
+ * This also helps the coalescing as it can ignore the redundant constraints.
+ * In order to avoid confusion, we make all implicit equalities explicit
+ * in the basic maps.  We don't call isl_basic_map_gauss, though,
+ * as that may affect the number of constraints.
+ * This means that we have to call isl_basic_map_gauss at the end
+ * of the computation to ensure that the basic maps are not left
+ * in an unexpected state.
  */
 struct isl_map *isl_map_coalesce(struct isl_map *map)
 {
@@ -1228,13 +1550,15 @@ struct isl_map *isl_map_coalesce(struct isl_map *map)
 	unsigned n;
 	struct isl_tab **tabs = NULL;
 
+	map = isl_map_remove_empty_parts(map);
 	if (!map)
 		return NULL;
 
 	if (map->n <= 1)
 		return map;
 
-	map = isl_map_align_divs(map);
+	map = isl_map_sort_divs(map);
+	map = isl_map_cow(map);
 
 	tabs = isl_calloc_array(map->ctx, struct isl_tab *, map->n);
 	if (!tabs)
@@ -1242,12 +1566,16 @@ struct isl_map *isl_map_coalesce(struct isl_map *map)
 
 	n = map->n;
 	for (i = 0; i < map->n; ++i) {
-		tabs[i] = isl_tab_from_basic_map(map->p[i]);
+		tabs[i] = isl_tab_from_basic_map(map->p[i], 0);
 		if (!tabs[i])
 			goto error;
 		if (!ISL_F_ISSET(map->p[i], ISL_BASIC_MAP_NO_IMPLICIT))
 			if (isl_tab_detect_implicit_equalities(tabs[i]) < 0)
 				goto error;
+		map->p[i] = isl_tab_make_equalities_explicit(tabs[i],
+								map->p[i]);
+		if (!map->p[i])
+			goto error;
 		if (!ISL_F_ISSET(map->p[i], ISL_BASIC_MAP_NO_REDUNDANT))
 			if (isl_tab_detect_redundant(tabs[i]) < 0)
 				goto error;
@@ -1262,6 +1590,7 @@ struct isl_map *isl_map_coalesce(struct isl_map *map)
 		for (i = 0; i < map->n; ++i) {
 			map->p[i] = isl_basic_map_update_from_tab(map->p[i],
 								    tabs[i]);
+			map->p[i] = isl_basic_map_gauss(map->p[i], NULL);
 			map->p[i] = isl_basic_map_finalize(map->p[i]);
 			if (!map->p[i])
 				goto error;